1 /*
   2  * Copyright (c) 2001, 2018, 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 "gc/g1/g1ConcurrentMarkBitMap.hpp"
  29 #include "gc/g1/g1ConcurrentMarkObjArrayProcessor.hpp"
  30 #include "gc/g1/g1HeapVerifier.hpp"
  31 #include "gc/g1/g1RegionMarkStatsCache.hpp"
  32 #include "gc/g1/heapRegionSet.hpp"
  33 #include "gc/shared/taskqueue.hpp"
  34 #include "memory/allocation.hpp"
  35 
  36 class ConcurrentGCTimer;
  37 class ConcurrentMarkThread;
  38 class G1CollectedHeap;
  39 class G1CMTask;
  40 class G1ConcurrentMark;
  41 class G1OldTracer;
  42 class G1RegionToSpaceMapper;
  43 class G1SurvivorRegions;
  44 
  45 #ifdef _MSC_VER
  46 #pragma warning(push)
  47 // warning C4522: multiple assignment operators specified
  48 #pragma warning(disable:4522)
  49 #endif
  50 
  51 // This is a container class for either an oop or a continuation address for
  52 // mark stack entries. Both are pushed onto the mark stack.
  53 class G1TaskQueueEntry {
  54 private:
  55   void* _holder;
  56 
  57   static const uintptr_t ArraySliceBit = 1;
  58 
  59   G1TaskQueueEntry(oop obj) : _holder(obj) {
  60     assert(_holder != NULL, "Not allowed to set NULL task queue element");
  61   }
  62   G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
  63 public:
  64   G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }
  65   G1TaskQueueEntry() : _holder(NULL) { }
  66 
  67   static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
  68   static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
  69 
  70   G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {
  71     _holder = t._holder;
  72     return *this;
  73   }
  74 
  75   volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
  76     _holder = t._holder;
  77     return *this;
  78   }
  79 
  80   oop obj() const {
  81     assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
  82     return (oop)_holder;
  83   }
  84 
  85   HeapWord* slice() const {
  86     assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
  87     return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
  88   }
  89 
  90   bool is_oop() const { return !is_array_slice(); }
  91   bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
  92   bool is_null() const { return _holder == NULL; }
  93 };
  94 
  95 #ifdef _MSC_VER
  96 #pragma warning(pop)
  97 #endif
  98 
  99 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
 100 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
 101 
 102 // Closure used by CM during concurrent reference discovery
 103 // and reference processing (during remarking) to determine
 104 // if a particular object is alive. It is primarily used
 105 // to determine if referents of discovered reference objects
 106 // are alive. An instance is also embedded into the
 107 // reference processor as the _is_alive_non_header field
 108 class G1CMIsAliveClosure : public BoolObjectClosure {
 109   G1CollectedHeap* _g1h;
 110  public:
 111   G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1h(g1) { }
 112 
 113   bool do_object_b(oop obj);
 114 };
 115 
 116 // Represents the overflow mark stack used by concurrent marking.
 117 //
 118 // Stores oops in a huge buffer in virtual memory that is always fully committed.
 119 // Resizing may only happen during a STW pause when the stack is empty.
 120 //
 121 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
 122 // stack memory is split into evenly sized chunks of oops. Users can only
 123 // add or remove entries on that basis.
 124 // Chunks are filled in increasing address order. Not completely filled chunks
 125 // have a NULL element as a terminating element.
 126 //
 127 // Every chunk has a header containing a single pointer element used for memory
 128 // management. This wastes some space, but is negligible (< .1% with current sizing).
 129 //
 130 // Memory management is done using a mix of tracking a high water-mark indicating
 131 // that all chunks at a lower address are valid chunks, and a singly linked free
 132 // list connecting all empty chunks.
 133 class G1CMMarkStack {
 134 public:
 135   // Number of TaskQueueEntries that can fit in a single chunk.
 136   static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
 137 private:
 138   struct TaskQueueEntryChunk {
 139     TaskQueueEntryChunk* next;
 140     G1TaskQueueEntry data[EntriesPerChunk];
 141   };
 142 
 143   size_t _max_chunk_capacity;    // Maximum number of TaskQueueEntryChunk elements on the stack.
 144 
 145   TaskQueueEntryChunk* _base;    // Bottom address of allocated memory area.
 146   size_t _chunk_capacity;        // Current maximum number of TaskQueueEntryChunk elements.
 147 
 148   char _pad0[DEFAULT_CACHE_LINE_SIZE];
 149   TaskQueueEntryChunk* volatile _free_list;  // Linked list of free chunks that can be allocated by users.
 150   char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
 151   TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
 152   volatile size_t _chunks_in_chunk_list;
 153   char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
 154 
 155   volatile size_t _hwm;          // High water mark within the reserved space.
 156   char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
 157 
 158   // Allocate a new chunk from the reserved memory, using the high water mark. Returns
 159   // NULL if out of memory.
 160   TaskQueueEntryChunk* allocate_new_chunk();
 161 
 162   // Atomically add the given chunk to the list.
 163   void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
 164   // Atomically remove and return a chunk from the given list. Returns NULL if the
 165   // list is empty.
 166   TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
 167 
 168   void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
 169   void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
 170 
 171   TaskQueueEntryChunk* remove_chunk_from_chunk_list();
 172   TaskQueueEntryChunk* remove_chunk_from_free_list();
 173 
 174   // Resizes the mark stack to the given new capacity. Releases any previous
 175   // memory if successful.
 176   bool resize(size_t new_capacity);
 177 
 178  public:
 179   G1CMMarkStack();
 180   ~G1CMMarkStack();
 181 
 182   // Alignment and minimum capacity of this mark stack in number of oops.
 183   static size_t capacity_alignment();
 184 
 185   // Allocate and initialize the mark stack with the given number of oops.
 186   bool initialize(size_t initial_capacity, size_t max_capacity);
 187 
 188   // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
 189   // stack. If less than EntriesPerChunk elements are to be pushed, the array must
 190   // be terminated with a NULL.
 191   // Returns whether the buffer contents were successfully pushed to the global mark
 192   // stack.
 193   bool par_push_chunk(G1TaskQueueEntry* buffer);
 194 
 195   // Pops a chunk from this mark stack, copying them into the given buffer. This
 196   // chunk may contain up to EntriesPerChunk elements. If there are less, the last
 197   // element in the array is a NULL pointer.
 198   bool par_pop_chunk(G1TaskQueueEntry* buffer);
 199 
 200   // Return whether the chunk list is empty. Racy due to unsynchronized access to
 201   // _chunk_list.
 202   bool is_empty() const { return _chunk_list == NULL; }
 203 
 204   size_t capacity() const  { return _chunk_capacity; }
 205 
 206   // Expand the stack, typically in response to an overflow condition
 207   void expand();
 208 
 209   // Return the approximate number of oops on this mark stack. Racy due to
 210   // unsynchronized access to _chunks_in_chunk_list.
 211   size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
 212 
 213   void set_empty();
 214 
 215   // Apply Fn to every oop on the mark stack. The mark stack must not
 216   // be modified while iterating.
 217   template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
 218 };
 219 
 220 // Root Regions are regions that are not empty at the beginning of a
 221 // marking cycle and which we might collect during an evacuation pause
 222 // while the cycle is active. Given that, during evacuation pauses, we
 223 // do not copy objects that are explicitly marked, what we have to do
 224 // for the root regions is to scan them and mark all objects reachable
 225 // from them. According to the SATB assumptions, we only need to visit
 226 // each object once during marking. So, as long as we finish this scan
 227 // before the next evacuation pause, we can copy the objects from the
 228 // root regions without having to mark them or do anything else to them.
 229 //
 230 // Currently, we only support root region scanning once (at the start
 231 // of the marking cycle) and the root regions are all the survivor
 232 // regions populated during the initial-mark pause.
 233 class G1CMRootRegions {
 234 private:
 235   const G1SurvivorRegions* _survivors;
 236   G1ConcurrentMark*        _cm;
 237 
 238   volatile bool            _scan_in_progress;
 239   volatile bool            _should_abort;
 240   volatile int             _claimed_survivor_index;
 241 
 242   void notify_scan_done();
 243 
 244 public:
 245   G1CMRootRegions();
 246   // We actually do most of the initialization in this method.
 247   void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
 248 
 249   // Reset the claiming / scanning of the root regions.
 250   void prepare_for_scan();
 251 
 252   // Forces get_next() to return NULL so that the iteration aborts early.
 253   void abort() { _should_abort = true; }
 254 
 255   // Return true if the CM thread are actively scanning root regions,
 256   // false otherwise.
 257   bool scan_in_progress() { return _scan_in_progress; }
 258 
 259   // Claim the next root region to scan atomically, or return NULL if
 260   // all have been claimed.
 261   HeapRegion* claim_next();
 262 
 263   // The number of root regions to scan.
 264   uint num_root_regions() const;
 265 
 266   void cancel_scan();
 267 
 268   // Flag that we're done with root region scanning and notify anyone
 269   // who's waiting on it. If aborted is false, assume that all regions
 270   // have been claimed.
 271   void scan_finished();
 272 
 273   // If CM threads are still scanning root regions, wait until they
 274   // are done. Return true if we had to wait, false otherwise.
 275   bool wait_until_scan_finished();
 276 };
 277 
 278 // This class manages data structures and methods for doing liveness analysis in
 279 // G1's concurrent cycle.
 280 class G1ConcurrentMark : public CHeapObj<mtGC> {
 281   friend class ConcurrentMarkThread;
 282   friend class G1CMRefProcTaskProxy;
 283   friend class G1CMRefProcTaskExecutor;
 284   friend class G1CMKeepAliveAndDrainClosure;
 285   friend class G1CMDrainMarkingStackClosure;
 286   friend class G1CMBitMapClosure;
 287   friend class G1CMConcurrentMarkingTask;
 288   friend class G1CMRemarkTask;
 289   friend class G1CMTask;
 290 
 291   ConcurrentMarkThread*  _cm_thread;     // The thread doing the work
 292   G1CollectedHeap*       _g1h;           // The heap
 293   bool                   _completed_initialization; // Set to true when initialization is complete
 294 
 295   // Concurrent marking support structures
 296   G1CMBitMap             _mark_bitmap_1;
 297   G1CMBitMap             _mark_bitmap_2;
 298   G1CMBitMap*            _prev_mark_bitmap; // Completed mark bitmap
 299   G1CMBitMap*            _next_mark_bitmap; // Under-construction mark bitmap
 300 
 301   // Heap bounds
 302   MemRegion const        _heap;
 303 
 304   // Root region tracking and claiming
 305   G1CMRootRegions        _root_regions;
 306 
 307   // For grey objects
 308   G1CMMarkStack          _global_mark_stack; // Grey objects behind global finger
 309   HeapWord* volatile     _finger;            // The global finger, region aligned,
 310                                              // always pointing to the end of the
 311                                              // last claimed region
 312 
 313   uint                   _worker_id_offset;
 314   uint                   _max_num_tasks;    // Maximum number of marking tasks
 315   uint                   _num_active_tasks; // Number of tasks currently active
 316   G1CMTask**             _tasks;            // Task queue array (max_worker_id length)
 317 
 318   G1CMTaskQueueSet*      _task_queues;      // Task queue set
 319   ParallelTaskTerminator _terminator;       // For termination
 320 
 321   // Two sync barriers that are used to synchronize tasks when an
 322   // overflow occurs. The algorithm is the following. All tasks enter
 323   // the first one to ensure that they have all stopped manipulating
 324   // the global data structures. After they exit it, they re-initialize
 325   // their data structures and task 0 re-initializes the global data
 326   // structures. Then, they enter the second sync barrier. This
 327   // ensure, that no task starts doing work before all data
 328   // structures (local and global) have been re-initialized. When they
 329   // exit it, they are free to start working again.
 330   WorkGangBarrierSync    _first_overflow_barrier_sync;
 331   WorkGangBarrierSync    _second_overflow_barrier_sync;
 332 
 333   // This is set by any task, when an overflow on the global data
 334   // structures is detected
 335   volatile bool          _has_overflown;
 336   // True: marking is concurrent, false: we're in remark
 337   volatile bool          _concurrent;
 338   // Set at the end of a Full GC so that marking aborts
 339   volatile bool          _has_aborted;
 340 
 341   // Used when remark aborts due to an overflow to indicate that
 342   // another concurrent marking phase should start
 343   volatile bool          _restart_for_overflow;
 344 
 345   ConcurrentGCTimer*     _gc_timer_cm;
 346 
 347   G1OldTracer*           _gc_tracer_cm;
 348 
 349   // Timing statistics. All of them are in ms
 350   NumberSeq _init_times;
 351   NumberSeq _remark_times;
 352   NumberSeq _remark_mark_times;
 353   NumberSeq _remark_weak_ref_times;
 354   NumberSeq _cleanup_times;
 355   double    _total_cleanup_time;
 356 
 357   double*   _accum_task_vtime;   // Accumulated task vtime
 358 
 359   WorkGang* _concurrent_workers;
 360   uint      _num_concurrent_workers; // The number of marking worker threads we're using
 361   uint      _max_concurrent_workers; // Maximum number of marking worker threads
 362 
 363   void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller);
 364 
 365   void finalize_marking();
 366 
 367   void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);
 368   void weak_refs_work(bool clear_all_soft_refs);
 369 
 370   void report_object_count(bool mark_completed);
 371 
 372   void swap_mark_bitmaps();
 373 
 374   void reclaim_empty_regions();
 375 
 376   // After reclaiming empty regions, update heap sizes.
 377   void compute_new_sizes();
 378 
 379   // Clear statistics gathered during the concurrent cycle for the given region after
 380   // it has been reclaimed.
 381   void clear_statistics(HeapRegion* r);
 382 
 383   // Resets the global marking data structures, as well as the
 384   // task local ones; should be called during initial mark.
 385   void reset();
 386 
 387   // Resets all the marking data structures. Called when we have to restart
 388   // marking or when marking completes (via set_non_marking_state below).
 389   void reset_marking_for_restart();
 390 
 391   // We do this after we're done with marking so that the marking data
 392   // structures are initialized to a sensible and predictable state.
 393   void reset_at_marking_complete();
 394 
 395   // Called to indicate how many threads are currently active.
 396   void set_concurrency(uint active_tasks);
 397 
 398   // Should be called to indicate which phase we're in (concurrent
 399   // mark or remark) and how many threads are currently active.
 400   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 401 
 402   // Prints all gathered CM-related statistics
 403   void print_stats();
 404 
 405   HeapWord*               finger()          { return _finger;   }
 406   bool                    concurrent()      { return _concurrent; }
 407   uint                    active_tasks()    { return _num_active_tasks; }
 408   ParallelTaskTerminator* terminator()      { return &_terminator; }
 409 
 410   // Claims the next available region to be scanned by a marking
 411   // task/thread. It might return NULL if the next region is empty or
 412   // we have run out of regions. In the latter case, out_of_regions()
 413   // determines whether we've really run out of regions or the task
 414   // should call claim_region() again. This might seem a bit
 415   // awkward. Originally, the code was written so that claim_region()
 416   // either successfully returned with a non-empty region or there
 417   // were no more regions to be claimed. The problem with this was
 418   // that, in certain circumstances, it iterated over large chunks of
 419   // the heap finding only empty regions and, while it was working, it
 420   // was preventing the calling task to call its regular clock
 421   // method. So, this way, each task will spend very little time in
 422   // claim_region() and is allowed to call the regular clock method
 423   // frequently.
 424   HeapRegion* claim_region(uint worker_id);
 425 
 426   // Determines whether we've run out of regions to scan. Note that
 427   // the finger can point past the heap end in case the heap was expanded
 428   // to satisfy an allocation without doing a GC. This is fine, because all
 429   // objects in those regions will be considered live anyway because of
 430   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 431   bool out_of_regions() { return _finger >= _heap.end(); }
 432 
 433   // Returns the task with the given id
 434   G1CMTask* task(uint id) {
 435     // During initial mark we use the parallel gc threads to do some work, so
 436     // we can only compare against _max_num_tasks.
 437     assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks);
 438     return _tasks[id];
 439   }
 440 
 441   // Access / manipulation of the overflow flag which is set to
 442   // indicate that the global stack has overflown
 443   bool has_overflown()           { return _has_overflown; }
 444   void set_has_overflown()       { _has_overflown = true; }
 445   void clear_has_overflown()     { _has_overflown = false; }
 446   bool restart_for_overflow()    { return _restart_for_overflow; }
 447 
 448   // Methods to enter the two overflow sync barriers
 449   void enter_first_sync_barrier(uint worker_id);
 450   void enter_second_sync_barrier(uint worker_id);
 451 
 452   // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
 453   // true, periodically insert checks to see if this method should exit prematurely.
 454   void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
 455 
 456   // Region statistics gathered during marking.
 457   G1RegionMarkStats* _region_mark_stats;
 458   // Top pointer for each region at the start of the rebuild remembered set process
 459   // for regions which remembered sets need to be rebuilt. A NULL for a given region
 460   // means that this region does not be scanned during the rebuilding remembered
 461   // set phase at all.
 462   HeapWord* volatile* _top_at_rebuild_starts;
 463 public:
 464   void add_to_liveness(uint worker_id, oop const obj, size_t size);
 465   // Liveness of the given region as determined by concurrent marking, i.e. the amount of
 466   // live words between bottom and nTAMS.
 467   size_t liveness(uint region)  { return _region_mark_stats[region]._live_words; }
 468 
 469   // Sets the internal top_at_region_start for the given region to current top of the region.
 470   inline void update_top_at_rebuild_start(HeapRegion* r);
 471   // TARS for the given region during remembered set rebuilding.
 472   inline HeapWord* top_at_rebuild_start(uint region) const;
 473 
 474   // Clear statistics gathered during the concurrent cycle for the given region after
 475   // it has been reclaimed.
 476   void clear_statistics_in_region(uint region_idx);
 477   // Notification for eagerly reclaimed regions to clean up.
 478   void humongous_object_eagerly_reclaimed(HeapRegion* r);
 479   // Manipulation of the global mark stack.
 480   // The push and pop operations are used by tasks for transfers
 481   // between task-local queues and the global mark stack.
 482   bool mark_stack_push(G1TaskQueueEntry* arr) {
 483     if (!_global_mark_stack.par_push_chunk(arr)) {
 484       set_has_overflown();
 485       return false;
 486     }
 487     return true;
 488   }
 489   bool mark_stack_pop(G1TaskQueueEntry* arr) {
 490     return _global_mark_stack.par_pop_chunk(arr);
 491   }
 492   size_t mark_stack_size() const                { return _global_mark_stack.size(); }
 493   size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; }
 494   bool mark_stack_empty() const                 { return _global_mark_stack.is_empty(); }
 495 
 496   G1CMRootRegions* root_regions() { return &_root_regions; }
 497 
 498   void concurrent_cycle_start();
 499   // Abandon current marking iteration due to a Full GC.
 500   void concurrent_cycle_abort();
 501   void concurrent_cycle_end();
 502 
 503   void update_accum_task_vtime(int i, double vtime) {
 504     _accum_task_vtime[i] += vtime;
 505   }
 506 
 507   double all_task_accum_vtime() {
 508     double ret = 0.0;
 509     for (uint i = 0; i < _max_num_tasks; ++i)
 510       ret += _accum_task_vtime[i];
 511     return ret;
 512   }
 513 
 514   // Attempts to steal an object from the task queues of other tasks
 515   bool try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry);
 516 
 517   G1ConcurrentMark(G1CollectedHeap* g1h,
 518                    G1RegionToSpaceMapper* prev_bitmap_storage,
 519                    G1RegionToSpaceMapper* next_bitmap_storage);
 520   ~G1ConcurrentMark();
 521 
 522   ConcurrentMarkThread* cm_thread() { return _cm_thread; }
 523 
 524   const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; }
 525   G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; }
 526 
 527   // Calculates the number of concurrent GC threads to be used in the marking phase.
 528   uint calc_active_marking_workers();
 529 
 530   // Moves all per-task cached data into global state.
 531   void flush_all_task_caches();
 532   // Prepare internal data structures for the next mark cycle. This includes clearing
 533   // the next mark bitmap and some internal data structures. This method is intended
 534   // to be called concurrently to the mutator. It will yield to safepoint requests.
 535   void cleanup_for_next_mark();
 536 
 537   // Clear the previous marking bitmap during safepoint.
 538   void clear_prev_bitmap(WorkGang* workers);
 539 
 540   // Return whether the next mark bitmap has no marks set. To be used for assertions
 541   // only. Will not yield to pause requests.
 542   bool next_mark_bitmap_is_clear();
 543 
 544   // These two methods do the work that needs to be done at the start and end of the
 545   // initial mark pause.
 546   void pre_initial_mark();
 547   void post_initial_mark();
 548 
 549   // Scan all the root regions and mark everything reachable from
 550   // them.
 551   void scan_root_regions();
 552 
 553   // Scan a single root region and mark everything reachable from it.
 554   void scan_root_region(HeapRegion* hr, uint worker_id);
 555 
 556   // Do concurrent phase of marking, to a tentative transitive closure.
 557   void mark_from_roots();
 558 
 559   void remark();
 560 
 561   void cleanup();
 562   // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use
 563   // this carefully.
 564   inline void mark_in_prev_bitmap(oop p);
 565 
 566   // Clears marks for all objects in the given range, for the prev or
 567   // next bitmaps.  Caution: the previous bitmap is usually
 568   // read-only, so use this carefully!
 569   void clear_range_in_prev_bitmap(MemRegion mr);
 570 
 571   inline bool is_marked_in_prev_bitmap(oop p) const;
 572 
 573   // Verify that there are no collection set oops on the stacks (taskqueues /
 574   // global mark stack) and fingers (global / per-task).
 575   // If marking is not in progress, it's a no-op.
 576   void verify_no_cset_oops() PRODUCT_RETURN;
 577 
 578   inline bool do_yield_check();
 579 
 580   bool has_aborted()      { return _has_aborted; }
 581 
 582   void print_summary_info();
 583 
 584   void print_worker_threads_on(outputStream* st) const;
 585   void threads_do(ThreadClosure* tc) const;
 586 
 587   void print_on_error(outputStream* st) const;
 588 
 589   // Mark the given object on the next bitmap if it is below nTAMS.
 590   // If the passed obj_size is zero, it is recalculated from the given object if
 591   // needed. This is to be as lazy as possible with accessing the object's size.
 592   inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj, size_t const obj_size = 0);
 593   inline bool mark_in_next_bitmap(uint worker_id, oop const obj, size_t const obj_size = 0);
 594 
 595   inline bool is_marked_in_next_bitmap(oop p) const;
 596 
 597   // Returns true if initialization was successfully completed.
 598   bool completed_initialization() const {
 599     return _completed_initialization;
 600   }
 601 
 602   ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
 603   G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
 604 
 605 private:
 606   // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application.
 607   void rebuild_rem_set_concurrently();
 608 };
 609 
 610 // A class representing a marking task.
 611 class G1CMTask : public TerminatorTerminator {
 612 private:
 613   enum PrivateConstants {
 614     // The regular clock call is called once the scanned words reaches
 615     // this limit
 616     words_scanned_period          = 12*1024,
 617     // The regular clock call is called once the number of visited
 618     // references reaches this limit
 619     refs_reached_period           = 1024,
 620     // Initial value for the hash seed, used in the work stealing code
 621     init_hash_seed                = 17
 622   };
 623 
 624   // Number of entries in the per-task stats entry. This seems enough to have a very
 625   // low cache miss rate.
 626   static const uint RegionMarkStatsCacheSize = 1024;
 627 
 628   G1CMObjArrayProcessor       _objArray_processor;
 629 
 630   uint                        _worker_id;
 631   G1CollectedHeap*            _g1h;
 632   G1ConcurrentMark*           _cm;
 633   G1CMBitMap*                 _next_mark_bitmap;
 634   // the task queue of this task
 635   G1CMTaskQueue*              _task_queue;
 636 
 637   G1RegionMarkStatsCache      _mark_stats_cache;
 638   // Number of calls to this task
 639   uint                        _calls;
 640 
 641   // When the virtual timer reaches this time, the marking step should exit
 642   double                      _time_target_ms;
 643   // Start time of the current marking step
 644   double                      _start_time_ms;
 645 
 646   // Oop closure used for iterations over oops
 647   G1CMOopClosure*             _cm_oop_closure;
 648 
 649   // Region this task is scanning, NULL if we're not scanning any
 650   HeapRegion*                 _curr_region;
 651   // Local finger of this task, NULL if we're not scanning a region
 652   HeapWord*                   _finger;
 653   // Limit of the region this task is scanning, NULL if we're not scanning one
 654   HeapWord*                   _region_limit;
 655 
 656   // Number of words this task has scanned
 657   size_t                      _words_scanned;
 658   // When _words_scanned reaches this limit, the regular clock is
 659   // called. Notice that this might be decreased under certain
 660   // circumstances (i.e. when we believe that we did an expensive
 661   // operation).
 662   size_t                      _words_scanned_limit;
 663   // Initial value of _words_scanned_limit (i.e. what it was
 664   // before it was decreased).
 665   size_t                      _real_words_scanned_limit;
 666 
 667   // Number of references this task has visited
 668   size_t                      _refs_reached;
 669   // When _refs_reached reaches this limit, the regular clock is
 670   // called. Notice this this might be decreased under certain
 671   // circumstances (i.e. when we believe that we did an expensive
 672   // operation).
 673   size_t                      _refs_reached_limit;
 674   // Initial value of _refs_reached_limit (i.e. what it was before
 675   // it was decreased).
 676   size_t                      _real_refs_reached_limit;
 677 
 678   // Used by the work stealing
 679   int                         _hash_seed;
 680   // If true, then the task has aborted for some reason
 681   bool                        _has_aborted;
 682   // Set when the task aborts because it has met its time quota
 683   bool                        _has_timed_out;
 684   // True when we're draining SATB buffers; this avoids the task
 685   // aborting due to SATB buffers being available (as we're already
 686   // dealing with them)
 687   bool                        _draining_satb_buffers;
 688 
 689   // Number sequence of past step times
 690   NumberSeq                   _step_times_ms;
 691   // Elapsed time of this task
 692   double                      _elapsed_time_ms;
 693   // Termination time of this task
 694   double                      _termination_time_ms;
 695   // When this task got into the termination protocol
 696   double                      _termination_start_time_ms;
 697 
 698   TruncatedSeq                _marking_step_diffs_ms;
 699 
 700   // Updates the local fields after this task has claimed
 701   // a new region to scan
 702   void setup_for_region(HeapRegion* hr);
 703   // Makes the limit of the region up-to-date
 704   void update_region_limit();
 705 
 706   // Called when either the words scanned or the refs visited limit
 707   // has been reached
 708   void reached_limit();
 709   // Recalculates the words scanned and refs visited limits
 710   void recalculate_limits();
 711   // Decreases the words scanned and refs visited limits when we reach
 712   // an expensive operation
 713   void decrease_limits();
 714   // Checks whether the words scanned or refs visited reached their
 715   // respective limit and calls reached_limit() if they have
 716   void check_limits() {
 717     if (_words_scanned >= _words_scanned_limit ||
 718         _refs_reached >= _refs_reached_limit) {
 719       reached_limit();
 720     }
 721   }
 722   // Supposed to be called regularly during a marking step as
 723   // it checks a bunch of conditions that might cause the marking step
 724   // to abort
 725   void regular_clock_call();
 726 
 727   // Test whether obj might have already been passed over by the
 728   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 729   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 730 
 731   template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
 732 public:
 733   // Apply the closure on the given area of the objArray. Return the number of words
 734   // scanned.
 735   inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
 736   // Resets the task; should be called right at the beginning of a marking phase.
 737   void reset(G1CMBitMap* next_mark_bitmap);
 738   // Clears all the fields that correspond to a claimed region.
 739   void clear_region_fields();
 740 
 741   // The main method of this class which performs a marking step
 742   // trying not to exceed the given duration. However, it might exit
 743   // prematurely, according to some conditions (i.e. SATB buffers are
 744   // available for processing).
 745   void do_marking_step(double target_ms,
 746                        bool do_termination,
 747                        bool is_serial);
 748 
 749   // These two calls start and stop the timer
 750   void record_start_time() {
 751     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 752   }
 753   void record_end_time() {
 754     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 755   }
 756 
 757   // Returns the worker ID associated with this task.
 758   uint worker_id() { return _worker_id; }
 759 
 760   // From TerminatorTerminator. It determines whether this task should
 761   // exit the termination protocol after it's entered it.
 762   virtual bool should_exit_termination();
 763 
 764   // Resets the local region fields after a task has finished scanning a
 765   // region; or when they have become stale as a result of the region
 766   // being evacuated.
 767   void giveup_current_region();
 768 
 769   HeapWord* finger()            { return _finger; }
 770 
 771   bool has_aborted()            { return _has_aborted; }
 772   void set_has_aborted()        { _has_aborted = true; }
 773   void clear_has_aborted()      { _has_aborted = false; }
 774 
 775   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
 776 
 777   // Increment the number of references this task has visited.
 778   void increment_refs_reached() { ++_refs_reached; }
 779 
 780   // Grey the object by marking it.  If not already marked, push it on
 781   // the local queue if below the finger.
 782   // obj is below its region's NTAMS.
 783   inline void make_reference_grey(oop obj);
 784 
 785   // Grey the object (by calling make_grey_reference) if required,
 786   // e.g. obj is below its containing region's NTAMS.
 787   // Precondition: obj is a valid heap object.
 788   template <class T>
 789   inline void deal_with_reference(T* p);
 790 
 791   // Scans an object and visits its children.
 792   inline void scan_task_entry(G1TaskQueueEntry task_entry);
 793 
 794   // Pushes an object on the local queue.
 795   inline void push(G1TaskQueueEntry task_entry);
 796 
 797   // Move entries to the global stack.
 798   void move_entries_to_global_stack();
 799   // Move entries from the global stack, return true if we were successful to do so.
 800   bool get_entries_from_global_stack();
 801 
 802   // Pops and scans objects from the local queue. If partially is
 803   // true, then it stops when the queue size is of a given limit. If
 804   // partially is false, then it stops when the queue is empty.
 805   void drain_local_queue(bool partially);
 806   // Moves entries from the global stack to the local queue and
 807   // drains the local queue. If partially is true, then it stops when
 808   // both the global stack and the local queue reach a given size. If
 809   // partially if false, it tries to empty them totally.
 810   void drain_global_stack(bool partially);
 811   // Keeps picking SATB buffers and processing them until no SATB
 812   // buffers are available.
 813   void drain_satb_buffers();
 814 
 815   // Moves the local finger to a new location
 816   inline void move_finger_to(HeapWord* new_finger) {
 817     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
 818     _finger = new_finger;
 819   }
 820 
 821   G1CMTask(uint worker_id,
 822            G1ConcurrentMark *cm,
 823            G1CMTaskQueue* task_queue,
 824            G1RegionMarkStats* mark_stats,
 825            uint max_regions);
 826 
 827   inline void update_liveness(oop const obj, size_t const obj_size);
 828 
 829   // Clear (without flushing) the mark cache entry for the given region.
 830   void clear_mark_stats_cache(uint region_idx);
 831   // Evict the whole statistics cache into the global statistics. Returns the
 832   // number of cache hits and misses so far.
 833   Pair<size_t, size_t> flush_mark_stats_cache();
 834   // Prints statistics associated with this task
 835   void print_stats();
 836 };
 837 
 838 // Class that's used to to print out per-region liveness
 839 // information. It's currently used at the end of marking and also
 840 // after we sort the old regions at the end of the cleanup operation.
 841 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure {
 842 private:
 843   // Accumulators for these values.
 844   size_t _total_used_bytes;
 845   size_t _total_capacity_bytes;
 846   size_t _total_prev_live_bytes;
 847   size_t _total_next_live_bytes;
 848 
 849   // Accumulator for the remembered set size
 850   size_t _total_remset_bytes;
 851 
 852   // Accumulator for strong code roots memory size
 853   size_t _total_strong_code_roots_bytes;
 854 
 855   static double bytes_to_mb(size_t val) {
 856     return (double) val / (double) M;
 857   }
 858 
 859 public:
 860   // The header and footer are printed in the constructor and
 861   // destructor respectively.
 862   G1PrintRegionLivenessInfoClosure(const char* phase_name);
 863   virtual bool do_heap_region(HeapRegion* r);
 864   ~G1PrintRegionLivenessInfoClosure();
 865 };
 866 
 867 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP