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 G1ConcurrentMarkThread;
  38 class G1CollectedHeap;
  39 class G1CMOopClosure;
  40 class G1CMTask;
  41 class G1ConcurrentMark;
  42 class G1OldTracer;
  43 class G1RegionToSpaceMapper;
  44 class G1SurvivorRegions;
  45 
  46 #ifdef _MSC_VER
  47 #pragma warning(push)
  48 // warning C4522: multiple assignment operators specified
  49 #pragma warning(disable:4522)
  50 #endif
  51 
  52 // This is a container class for either an oop or a continuation address for
  53 // mark stack entries. Both are pushed onto the mark stack.
  54 class G1TaskQueueEntry {
  55 private:
  56   void* _holder;
  57 
  58   static const uintptr_t ArraySliceBit = 1;
  59 
  60   G1TaskQueueEntry(oop obj) : _holder(obj) {
  61     assert(_holder != NULL, "Not allowed to set NULL task queue element");
  62   }
  63   G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
  64 public:
  65   G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }
  66   G1TaskQueueEntry() : _holder(NULL) { }
  67 
  68   static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
  69   static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
  70 
  71   G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {
  72     _holder = t._holder;
  73     return *this;
  74   }
  75 
  76   volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
  77     _holder = t._holder;
  78     return *this;
  79   }
  80 
  81   oop obj() const {
  82     assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
  83     return (oop)_holder;
  84   }
  85 
  86   HeapWord* slice() const {
  87     assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
  88     return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
  89   }
  90 
  91   bool is_oop() const { return !is_array_slice(); }
  92   bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
  93   bool is_null() const { return _holder == NULL; }
  94 };
  95 
  96 #ifdef _MSC_VER
  97 #pragma warning(pop)
  98 #endif
  99 
 100 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
 101 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
 102 
 103 // Closure used by CM during concurrent reference discovery
 104 // and reference processing (during remarking) to determine
 105 // if a particular object is alive. It is primarily used
 106 // to determine if referents of discovered reference objects
 107 // are alive. An instance is also embedded into the
 108 // reference processor as the _is_alive_non_header field
 109 class G1CMIsAliveClosure : public BoolObjectClosure {
 110   G1CollectedHeap* _g1h;
 111 public:
 112   G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
 113   bool do_object_b(oop obj);
 114 };
 115 
 116 class G1CMSubjectToDiscoveryClosure : public BoolObjectClosure {
 117   G1CollectedHeap* _g1h;
 118 public:
 119   G1CMSubjectToDiscoveryClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
 120   bool do_object_b(oop obj);
 121 };
 122 
 123 // Represents the overflow mark stack used by concurrent marking.
 124 //
 125 // Stores oops in a huge buffer in virtual memory that is always fully committed.
 126 // Resizing may only happen during a STW pause when the stack is empty.
 127 //
 128 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
 129 // stack memory is split into evenly sized chunks of oops. Users can only
 130 // add or remove entries on that basis.
 131 // Chunks are filled in increasing address order. Not completely filled chunks
 132 // have a NULL element as a terminating element.
 133 //
 134 // Every chunk has a header containing a single pointer element used for memory
 135 // management. This wastes some space, but is negligible (< .1% with current sizing).
 136 //
 137 // Memory management is done using a mix of tracking a high water-mark indicating
 138 // that all chunks at a lower address are valid chunks, and a singly linked free
 139 // list connecting all empty chunks.
 140 class G1CMMarkStack {
 141 public:
 142   // Number of TaskQueueEntries that can fit in a single chunk.
 143   static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
 144 private:
 145   struct TaskQueueEntryChunk {
 146     TaskQueueEntryChunk* next;
 147     G1TaskQueueEntry data[EntriesPerChunk];
 148   };
 149 
 150   size_t _max_chunk_capacity;    // Maximum number of TaskQueueEntryChunk elements on the stack.
 151 
 152   TaskQueueEntryChunk* _base;    // Bottom address of allocated memory area.
 153   size_t _chunk_capacity;        // Current maximum number of TaskQueueEntryChunk elements.
 154 
 155   char _pad0[DEFAULT_CACHE_LINE_SIZE];
 156   TaskQueueEntryChunk* volatile _free_list;  // Linked list of free chunks that can be allocated by users.
 157   char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
 158   TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
 159   volatile size_t _chunks_in_chunk_list;
 160   char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
 161 
 162   volatile size_t _hwm;          // High water mark within the reserved space.
 163   char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
 164 
 165   // Allocate a new chunk from the reserved memory, using the high water mark. Returns
 166   // NULL if out of memory.
 167   TaskQueueEntryChunk* allocate_new_chunk();
 168 
 169   // Atomically add the given chunk to the list.
 170   void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
 171   // Atomically remove and return a chunk from the given list. Returns NULL if the
 172   // list is empty.
 173   TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
 174 
 175   void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
 176   void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
 177 
 178   TaskQueueEntryChunk* remove_chunk_from_chunk_list();
 179   TaskQueueEntryChunk* remove_chunk_from_free_list();
 180 
 181   // Resizes the mark stack to the given new capacity. Releases any previous
 182   // memory if successful.
 183   bool resize(size_t new_capacity);
 184 
 185  public:
 186   G1CMMarkStack();
 187   ~G1CMMarkStack();
 188 
 189   // Alignment and minimum capacity of this mark stack in number of oops.
 190   static size_t capacity_alignment();
 191 
 192   // Allocate and initialize the mark stack with the given number of oops.
 193   bool initialize(size_t initial_capacity, size_t max_capacity);
 194 
 195   // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
 196   // stack. If less than EntriesPerChunk elements are to be pushed, the array must
 197   // be terminated with a NULL.
 198   // Returns whether the buffer contents were successfully pushed to the global mark
 199   // stack.
 200   bool par_push_chunk(G1TaskQueueEntry* buffer);
 201 
 202   // Pops a chunk from this mark stack, copying them into the given buffer. This
 203   // chunk may contain up to EntriesPerChunk elements. If there are less, the last
 204   // element in the array is a NULL pointer.
 205   bool par_pop_chunk(G1TaskQueueEntry* buffer);
 206 
 207   // Return whether the chunk list is empty. Racy due to unsynchronized access to
 208   // _chunk_list.
 209   bool is_empty() const { return _chunk_list == NULL; }
 210 
 211   size_t capacity() const  { return _chunk_capacity; }
 212 
 213   // Expand the stack, typically in response to an overflow condition
 214   void expand();
 215 
 216   // Return the approximate number of oops on this mark stack. Racy due to
 217   // unsynchronized access to _chunks_in_chunk_list.
 218   size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
 219 
 220   void set_empty();
 221 
 222   // Apply Fn to every oop on the mark stack. The mark stack must not
 223   // be modified while iterating.
 224   template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
 225 };
 226 
 227 // Root Regions are regions that are not empty at the beginning of a
 228 // marking cycle and which we might collect during an evacuation pause
 229 // while the cycle is active. Given that, during evacuation pauses, we
 230 // do not copy objects that are explicitly marked, what we have to do
 231 // for the root regions is to scan them and mark all objects reachable
 232 // from them. According to the SATB assumptions, we only need to visit
 233 // each object once during marking. So, as long as we finish this scan
 234 // before the next evacuation pause, we can copy the objects from the
 235 // root regions without having to mark them or do anything else to them.
 236 //
 237 // Currently, we only support root region scanning once (at the start
 238 // of the marking cycle) and the root regions are all the survivor
 239 // regions populated during the initial-mark pause.
 240 class G1CMRootRegions {
 241 private:
 242   const G1SurvivorRegions* _survivors;
 243   G1ConcurrentMark*        _cm;
 244 
 245   volatile bool            _scan_in_progress;
 246   volatile bool            _should_abort;
 247   volatile int             _claimed_survivor_index;
 248 
 249   void notify_scan_done();
 250 
 251 public:
 252   G1CMRootRegions();
 253   // We actually do most of the initialization in this method.
 254   void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
 255 
 256   // Reset the claiming / scanning of the root regions.
 257   void prepare_for_scan();
 258 
 259   // Forces get_next() to return NULL so that the iteration aborts early.
 260   void abort() { _should_abort = true; }
 261 
 262   // Return true if the CM thread are actively scanning root regions,
 263   // false otherwise.
 264   bool scan_in_progress() { return _scan_in_progress; }
 265 
 266   // Claim the next root region to scan atomically, or return NULL if
 267   // all have been claimed.
 268   HeapRegion* claim_next();
 269 
 270   // The number of root regions to scan.
 271   uint num_root_regions() const;
 272 
 273   void cancel_scan();
 274 
 275   // Flag that we're done with root region scanning and notify anyone
 276   // who's waiting on it. If aborted is false, assume that all regions
 277   // have been claimed.
 278   void scan_finished();
 279 
 280   // If CM threads are still scanning root regions, wait until they
 281   // are done. Return true if we had to wait, false otherwise.
 282   bool wait_until_scan_finished();
 283 };
 284 
 285 // This class manages data structures and methods for doing liveness analysis in
 286 // G1's concurrent cycle.
 287 class G1ConcurrentMark : public CHeapObj<mtGC> {
 288   friend class G1ConcurrentMarkThread;
 289   friend class G1CMRefProcTaskProxy;
 290   friend class G1CMRefProcTaskExecutor;
 291   friend class G1CMKeepAliveAndDrainClosure;
 292   friend class G1CMDrainMarkingStackClosure;
 293   friend class G1CMBitMapClosure;
 294   friend class G1CMConcurrentMarkingTask;
 295   friend class G1CMRemarkTask;
 296   friend class G1CMTask;
 297 
 298   G1ConcurrentMarkThread* _cm_thread;     // The thread doing the work
 299   G1CollectedHeap*        _g1h;           // The heap
 300   bool                    _completed_initialization; // Set to true when initialization is complete
 301 
 302   // Concurrent marking support structures
 303   G1CMBitMap              _mark_bitmap_1;
 304   G1CMBitMap              _mark_bitmap_2;
 305   G1CMBitMap*             _prev_mark_bitmap; // Completed mark bitmap
 306   G1CMBitMap*             _next_mark_bitmap; // Under-construction mark bitmap
 307 
 308   // Heap bounds
 309   MemRegion const         _heap;
 310 
 311   // Root region tracking and claiming
 312   G1CMRootRegions         _root_regions;
 313 
 314   // For grey objects
 315   G1CMMarkStack           _global_mark_stack; // Grey objects behind global finger
 316   HeapWord* volatile      _finger;            // The global finger, region aligned,
 317                                               // always pointing to the end of the
 318                                               // last claimed region
 319 
 320   uint                    _worker_id_offset;
 321   uint                    _max_num_tasks;    // Maximum number of marking tasks
 322   uint                    _num_active_tasks; // Number of tasks currently active
 323   G1CMTask**              _tasks;            // Task queue array (max_worker_id length)
 324 
 325   G1CMTaskQueueSet*       _task_queues;      // Task queue set
 326   ParallelTaskTerminator  _terminator;       // For termination
 327 
 328   // Two sync barriers that are used to synchronize tasks when an
 329   // overflow occurs. The algorithm is the following. All tasks enter
 330   // the first one to ensure that they have all stopped manipulating
 331   // the global data structures. After they exit it, they re-initialize
 332   // their data structures and task 0 re-initializes the global data
 333   // structures. Then, they enter the second sync barrier. This
 334   // ensure, that no task starts doing work before all data
 335   // structures (local and global) have been re-initialized. When they
 336   // exit it, they are free to start working again.
 337   WorkGangBarrierSync     _first_overflow_barrier_sync;
 338   WorkGangBarrierSync     _second_overflow_barrier_sync;
 339 
 340   // This is set by any task, when an overflow on the global data
 341   // structures is detected
 342   volatile bool           _has_overflown;
 343   // True: marking is concurrent, false: we're in remark
 344   volatile bool           _concurrent;
 345   // Set at the end of a Full GC so that marking aborts
 346   volatile bool           _has_aborted;
 347 
 348   // Used when remark aborts due to an overflow to indicate that
 349   // another concurrent marking phase should start
 350   volatile bool           _restart_for_overflow;
 351 
 352   ConcurrentGCTimer*      _gc_timer_cm;
 353 
 354   G1OldTracer*            _gc_tracer_cm;
 355 
 356   // Timing statistics. All of them are in ms
 357   NumberSeq _init_times;
 358   NumberSeq _remark_times;
 359   NumberSeq _remark_mark_times;
 360   NumberSeq _remark_weak_ref_times;
 361   NumberSeq _cleanup_times;
 362   double    _total_cleanup_time;
 363 
 364   double*   _accum_task_vtime;   // Accumulated task vtime
 365 
 366   WorkGang* _concurrent_workers;
 367   uint      _num_concurrent_workers; // The number of marking worker threads we're using
 368   uint      _max_concurrent_workers; // Maximum number of marking worker threads
 369 
 370   void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller);
 371 
 372   void finalize_marking();
 373 
 374   void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);
 375   void weak_refs_work(bool clear_all_soft_refs);
 376 
 377   void report_object_count(bool mark_completed);
 378 
 379   void swap_mark_bitmaps();
 380 
 381   void reclaim_empty_regions();
 382 
 383   // After reclaiming empty regions, update heap sizes.
 384   void compute_new_sizes();
 385 
 386   // Clear statistics gathered during the concurrent cycle for the given region after
 387   // it has been reclaimed.
 388   void clear_statistics(HeapRegion* r);
 389 
 390   // Resets the global marking data structures, as well as the
 391   // task local ones; should be called during initial mark.
 392   void reset();
 393 
 394   // Resets all the marking data structures. Called when we have to restart
 395   // marking or when marking completes (via set_non_marking_state below).
 396   void reset_marking_for_restart();
 397 
 398   // We do this after we're done with marking so that the marking data
 399   // structures are initialized to a sensible and predictable state.
 400   void reset_at_marking_complete();
 401 
 402   // Called to indicate how many threads are currently active.
 403   void set_concurrency(uint active_tasks);
 404 
 405   // Should be called to indicate which phase we're in (concurrent
 406   // mark or remark) and how many threads are currently active.
 407   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 408 
 409   // Prints all gathered CM-related statistics
 410   void print_stats();
 411 
 412   HeapWord*               finger()          { return _finger;   }
 413   bool                    concurrent()      { return _concurrent; }
 414   uint                    active_tasks()    { return _num_active_tasks; }
 415   ParallelTaskTerminator* terminator()      { return &_terminator; }
 416 
 417   // Claims the next available region to be scanned by a marking
 418   // task/thread. It might return NULL if the next region is empty or
 419   // we have run out of regions. In the latter case, out_of_regions()
 420   // determines whether we've really run out of regions or the task
 421   // should call claim_region() again. This might seem a bit
 422   // awkward. Originally, the code was written so that claim_region()
 423   // either successfully returned with a non-empty region or there
 424   // were no more regions to be claimed. The problem with this was
 425   // that, in certain circumstances, it iterated over large chunks of
 426   // the heap finding only empty regions and, while it was working, it
 427   // was preventing the calling task to call its regular clock
 428   // method. So, this way, each task will spend very little time in
 429   // claim_region() and is allowed to call the regular clock method
 430   // frequently.
 431   HeapRegion* claim_region(uint worker_id);
 432 
 433   // Determines whether we've run out of regions to scan. Note that
 434   // the finger can point past the heap end in case the heap was expanded
 435   // to satisfy an allocation without doing a GC. This is fine, because all
 436   // objects in those regions will be considered live anyway because of
 437   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 438   bool out_of_regions() { return _finger >= _heap.end(); }
 439 
 440   // Returns the task with the given id
 441   G1CMTask* task(uint id) {
 442     // During initial mark we use the parallel gc threads to do some work, so
 443     // we can only compare against _max_num_tasks.
 444     assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks);
 445     return _tasks[id];
 446   }
 447 
 448   // Access / manipulation of the overflow flag which is set to
 449   // indicate that the global stack has overflown
 450   bool has_overflown()           { return _has_overflown; }
 451   void set_has_overflown()       { _has_overflown = true; }
 452   void clear_has_overflown()     { _has_overflown = false; }
 453   bool restart_for_overflow()    { return _restart_for_overflow; }
 454 
 455   // Methods to enter the two overflow sync barriers
 456   void enter_first_sync_barrier(uint worker_id);
 457   void enter_second_sync_barrier(uint worker_id);
 458 
 459   // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
 460   // true, periodically insert checks to see if this method should exit prematurely.
 461   void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
 462 
 463   // Region statistics gathered during marking.
 464   G1RegionMarkStats* _region_mark_stats;
 465   // Top pointer for each region at the start of the rebuild remembered set process
 466   // for regions which remembered sets need to be rebuilt. A NULL for a given region
 467   // means that this region does not be scanned during the rebuilding remembered
 468   // set phase at all.
 469   HeapWord* volatile* _top_at_rebuild_starts;
 470 public:
 471   void add_to_liveness(uint worker_id, oop const obj, size_t size);
 472   // Liveness of the given region as determined by concurrent marking, i.e. the amount of
 473   // live words between bottom and nTAMS.
 474   size_t liveness(uint region) const { return _region_mark_stats[region]._live_words; }
 475 
 476   // Sets the internal top_at_region_start for the given region to current top of the region.
 477   inline void update_top_at_rebuild_start(HeapRegion* r);
 478   // TARS for the given region during remembered set rebuilding.
 479   inline HeapWord* top_at_rebuild_start(uint region) const;
 480 
 481   // Clear statistics gathered during the concurrent cycle for the given region after
 482   // it has been reclaimed.
 483   void clear_statistics_in_region(uint region_idx);
 484   // Notification for eagerly reclaimed regions to clean up.
 485   void humongous_object_eagerly_reclaimed(HeapRegion* r);
 486   // Manipulation of the global mark stack.
 487   // The push and pop operations are used by tasks for transfers
 488   // between task-local queues and the global mark stack.
 489   bool mark_stack_push(G1TaskQueueEntry* arr) {
 490     if (!_global_mark_stack.par_push_chunk(arr)) {
 491       set_has_overflown();
 492       return false;
 493     }
 494     return true;
 495   }
 496   bool mark_stack_pop(G1TaskQueueEntry* arr) {
 497     return _global_mark_stack.par_pop_chunk(arr);
 498   }
 499   size_t mark_stack_size() const                { return _global_mark_stack.size(); }
 500   size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; }
 501   bool mark_stack_empty() const                 { return _global_mark_stack.is_empty(); }
 502 
 503   G1CMRootRegions* root_regions() { return &_root_regions; }
 504 
 505   void concurrent_cycle_start();
 506   // Abandon current marking iteration due to a Full GC.
 507   void concurrent_cycle_abort();
 508   void concurrent_cycle_end();
 509 
 510   void update_accum_task_vtime(int i, double vtime) {
 511     _accum_task_vtime[i] += vtime;
 512   }
 513 
 514   double all_task_accum_vtime() {
 515     double ret = 0.0;
 516     for (uint i = 0; i < _max_num_tasks; ++i)
 517       ret += _accum_task_vtime[i];
 518     return ret;
 519   }
 520 
 521   // Attempts to steal an object from the task queues of other tasks
 522   bool try_stealing(uint worker_id, G1TaskQueueEntry& task_entry);
 523 
 524   G1ConcurrentMark(G1CollectedHeap* g1h,
 525                    G1RegionToSpaceMapper* prev_bitmap_storage,
 526                    G1RegionToSpaceMapper* next_bitmap_storage);
 527   ~G1ConcurrentMark();
 528 
 529   G1ConcurrentMarkThread* cm_thread() { return _cm_thread; }
 530 
 531   const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; }
 532   G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; }
 533 
 534   // Calculates the number of concurrent GC threads to be used in the marking phase.
 535   uint calc_active_marking_workers();
 536 
 537   // Moves all per-task cached data into global state.
 538   void flush_all_task_caches();
 539   // Prepare internal data structures for the next mark cycle. This includes clearing
 540   // the next mark bitmap and some internal data structures. This method is intended
 541   // to be called concurrently to the mutator. It will yield to safepoint requests.
 542   void cleanup_for_next_mark();
 543 
 544   // Clear the previous marking bitmap during safepoint.
 545   void clear_prev_bitmap(WorkGang* workers);
 546 
 547   // These two methods do the work that needs to be done at the start and end of the
 548   // initial mark pause.
 549   void pre_initial_mark();
 550   void post_initial_mark();
 551 
 552   // Scan all the root regions and mark everything reachable from
 553   // them.
 554   void scan_root_regions();
 555 
 556   // Scan a single root region and mark everything reachable from it.
 557   void scan_root_region(HeapRegion* hr, uint worker_id);
 558 
 559   // Do concurrent phase of marking, to a tentative transitive closure.
 560   void mark_from_roots();
 561 
 562   // Do concurrent preclean work.
 563   void preclean();
 564 
 565   void remark();
 566 
 567   void cleanup();
 568   // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use
 569   // this carefully.
 570   inline void mark_in_prev_bitmap(oop p);
 571 
 572   // Clears marks for all objects in the given range, for the prev or
 573   // next bitmaps.  Caution: the previous bitmap is usually
 574   // read-only, so use this carefully!
 575   void clear_range_in_prev_bitmap(MemRegion mr);
 576 
 577   inline bool is_marked_in_prev_bitmap(oop p) const;
 578 
 579   // Verify that there are no collection set oops on the stacks (taskqueues /
 580   // global mark stack) and fingers (global / per-task).
 581   // If marking is not in progress, it's a no-op.
 582   void verify_no_cset_oops() PRODUCT_RETURN;
 583 
 584   inline bool do_yield_check();
 585 
 586   bool has_aborted()      { return _has_aborted; }
 587 
 588   void print_summary_info();
 589 
 590   void print_worker_threads_on(outputStream* st) const;
 591   void threads_do(ThreadClosure* tc) const;
 592 
 593   void print_on_error(outputStream* st) const;
 594 
 595   // Mark the given object on the next bitmap if it is below nTAMS.
 596   // If the passed obj_size is zero, it is recalculated from the given object if
 597   // needed. This is to be as lazy as possible with accessing the object's size.
 598   inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj, size_t const obj_size = 0);
 599   inline bool mark_in_next_bitmap(uint worker_id, oop const obj, size_t const obj_size = 0);
 600 
 601   inline bool is_marked_in_next_bitmap(oop p) const;
 602 
 603   // Returns true if initialization was successfully completed.
 604   bool completed_initialization() const {
 605     return _completed_initialization;
 606   }
 607 
 608   ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
 609   G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
 610 
 611 private:
 612   // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application.
 613   void rebuild_rem_set_concurrently();
 614 };
 615 
 616 // A class representing a marking task.
 617 class G1CMTask : public TerminatorTerminator {
 618 private:
 619   enum PrivateConstants {
 620     // The regular clock call is called once the scanned words reaches
 621     // this limit
 622     words_scanned_period          = 12*1024,
 623     // The regular clock call is called once the number of visited
 624     // references reaches this limit
 625     refs_reached_period           = 1024,
 626     // Initial value for the hash seed, used in the work stealing code
 627     init_hash_seed                = 17
 628   };
 629 
 630   // Number of entries in the per-task stats entry. This seems enough to have a very
 631   // low cache miss rate.
 632   static const uint RegionMarkStatsCacheSize = 1024;
 633 
 634   G1CMObjArrayProcessor       _objArray_processor;
 635 
 636   uint                        _worker_id;
 637   G1CollectedHeap*            _g1h;
 638   G1ConcurrentMark*           _cm;
 639   G1CMBitMap*                 _next_mark_bitmap;
 640   // the task queue of this task
 641   G1CMTaskQueue*              _task_queue;
 642 
 643   G1RegionMarkStatsCache      _mark_stats_cache;
 644   // Number of calls to this task
 645   uint                        _calls;
 646 
 647   // When the virtual timer reaches this time, the marking step should exit
 648   double                      _time_target_ms;
 649   // Start time of the current marking step
 650   double                      _start_time_ms;
 651 
 652   // Oop closure used for iterations over oops
 653   G1CMOopClosure*             _cm_oop_closure;
 654 
 655   // Region this task is scanning, NULL if we're not scanning any
 656   HeapRegion*                 _curr_region;
 657   // Local finger of this task, NULL if we're not scanning a region
 658   HeapWord*                   _finger;
 659   // Limit of the region this task is scanning, NULL if we're not scanning one
 660   HeapWord*                   _region_limit;
 661 
 662   // Number of words this task has scanned
 663   size_t                      _words_scanned;
 664   // When _words_scanned reaches this limit, the regular clock is
 665   // called. Notice that this might be decreased under certain
 666   // circumstances (i.e. when we believe that we did an expensive
 667   // operation).
 668   size_t                      _words_scanned_limit;
 669   // Initial value of _words_scanned_limit (i.e. what it was
 670   // before it was decreased).
 671   size_t                      _real_words_scanned_limit;
 672 
 673   // Number of references this task has visited
 674   size_t                      _refs_reached;
 675   // When _refs_reached reaches this limit, the regular clock is
 676   // called. Notice this this might be decreased under certain
 677   // circumstances (i.e. when we believe that we did an expensive
 678   // operation).
 679   size_t                      _refs_reached_limit;
 680   // Initial value of _refs_reached_limit (i.e. what it was before
 681   // it was decreased).
 682   size_t                      _real_refs_reached_limit;
 683 
 684   // If true, then the task has aborted for some reason
 685   bool                        _has_aborted;
 686   // Set when the task aborts because it has met its time quota
 687   bool                        _has_timed_out;
 688   // True when we're draining SATB buffers; this avoids the task
 689   // aborting due to SATB buffers being available (as we're already
 690   // dealing with them)
 691   bool                        _draining_satb_buffers;
 692 
 693   // Number sequence of past step times
 694   NumberSeq                   _step_times_ms;
 695   // Elapsed time of this task
 696   double                      _elapsed_time_ms;
 697   // Termination time of this task
 698   double                      _termination_time_ms;
 699   // When this task got into the termination protocol
 700   double                      _termination_start_time_ms;
 701 
 702   TruncatedSeq                _marking_step_diffs_ms;
 703 
 704   // Updates the local fields after this task has claimed
 705   // a new region to scan
 706   void setup_for_region(HeapRegion* hr);
 707   // Makes the limit of the region up-to-date
 708   void update_region_limit();
 709 
 710   // Called when either the words scanned or the refs visited limit
 711   // has been reached
 712   void reached_limit();
 713   // Recalculates the words scanned and refs visited limits
 714   void recalculate_limits();
 715   // Decreases the words scanned and refs visited limits when we reach
 716   // an expensive operation
 717   void decrease_limits();
 718   // Checks whether the words scanned or refs visited reached their
 719   // respective limit and calls reached_limit() if they have
 720   void check_limits() {
 721     if (_words_scanned >= _words_scanned_limit ||
 722         _refs_reached >= _refs_reached_limit) {
 723       reached_limit();
 724     }
 725   }
 726   // Supposed to be called regularly during a marking step as
 727   // it checks a bunch of conditions that might cause the marking step
 728   // to abort
 729   void regular_clock_call();
 730 
 731   // Test whether obj might have already been passed over by the
 732   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 733   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 734 
 735   template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
 736 public:
 737   // Apply the closure on the given area of the objArray. Return the number of words
 738   // scanned.
 739   inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
 740   // Resets the task; should be called right at the beginning of a marking phase.
 741   void reset(G1CMBitMap* next_mark_bitmap);
 742   // Clears all the fields that correspond to a claimed region.
 743   void clear_region_fields();
 744 
 745   // The main method of this class which performs a marking step
 746   // trying not to exceed the given duration. However, it might exit
 747   // prematurely, according to some conditions (i.e. SATB buffers are
 748   // available for processing).
 749   void do_marking_step(double target_ms,
 750                        bool do_termination,
 751                        bool is_serial);
 752 
 753   // These two calls start and stop the timer
 754   void record_start_time() {
 755     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 756   }
 757   void record_end_time() {
 758     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 759   }
 760 
 761   // Returns the worker ID associated with this task.
 762   uint worker_id() { return _worker_id; }
 763 
 764   // From TerminatorTerminator. It determines whether this task should
 765   // exit the termination protocol after it's entered it.
 766   virtual bool should_exit_termination();
 767 
 768   // Resets the local region fields after a task has finished scanning a
 769   // region; or when they have become stale as a result of the region
 770   // being evacuated.
 771   void giveup_current_region();
 772 
 773   HeapWord* finger()            { return _finger; }
 774 
 775   bool has_aborted()            { return _has_aborted; }
 776   void set_has_aborted()        { _has_aborted = true; }
 777   void clear_has_aborted()      { _has_aborted = false; }
 778 
 779   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
 780 
 781   // Increment the number of references this task has visited.
 782   void increment_refs_reached() { ++_refs_reached; }
 783 
 784   // Grey the object by marking it.  If not already marked, push it on
 785   // the local queue if below the finger. obj is required to be below its region's NTAMS.
 786   // Returns whether there has been a mark to the bitmap.
 787   inline bool make_reference_grey(oop obj);
 788 
 789   // Grey the object (by calling make_grey_reference) if required,
 790   // e.g. obj is below its containing region's NTAMS.
 791   // Precondition: obj is a valid heap object.
 792   // Returns true if the reference caused a mark to be set in the next bitmap.
 793   template <class T>
 794   inline bool deal_with_reference(T* p);
 795 
 796   // Scans an object and visits its children.
 797   inline void scan_task_entry(G1TaskQueueEntry task_entry);
 798 
 799   // Pushes an object on the local queue.
 800   inline void push(G1TaskQueueEntry task_entry);
 801 
 802   // Move entries to the global stack.
 803   void move_entries_to_global_stack();
 804   // Move entries from the global stack, return true if we were successful to do so.
 805   bool get_entries_from_global_stack();
 806 
 807   // Pops and scans objects from the local queue. If partially is
 808   // true, then it stops when the queue size is of a given limit. If
 809   // partially is false, then it stops when the queue is empty.
 810   void drain_local_queue(bool partially);
 811   // Moves entries from the global stack to the local queue and
 812   // drains the local queue. If partially is true, then it stops when
 813   // both the global stack and the local queue reach a given size. If
 814   // partially if false, it tries to empty them totally.
 815   void drain_global_stack(bool partially);
 816   // Keeps picking SATB buffers and processing them until no SATB
 817   // buffers are available.
 818   void drain_satb_buffers();
 819 
 820   // Moves the local finger to a new location
 821   inline void move_finger_to(HeapWord* new_finger) {
 822     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
 823     _finger = new_finger;
 824   }
 825 
 826   G1CMTask(uint worker_id,
 827            G1ConcurrentMark *cm,
 828            G1CMTaskQueue* task_queue,
 829            G1RegionMarkStats* mark_stats,
 830            uint max_regions);
 831 
 832   inline void update_liveness(oop const obj, size_t const obj_size);
 833 
 834   // Clear (without flushing) the mark cache entry for the given region.
 835   void clear_mark_stats_cache(uint region_idx);
 836   // Evict the whole statistics cache into the global statistics. Returns the
 837   // number of cache hits and misses so far.
 838   Pair<size_t, size_t> flush_mark_stats_cache();
 839   // Prints statistics associated with this task
 840   void print_stats();
 841 };
 842 
 843 // Class that's used to to print out per-region liveness
 844 // information. It's currently used at the end of marking and also
 845 // after we sort the old regions at the end of the cleanup operation.
 846 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure {
 847   // Accumulators for these values.
 848   size_t _total_used_bytes;
 849   size_t _total_capacity_bytes;
 850   size_t _total_prev_live_bytes;
 851   size_t _total_next_live_bytes;
 852 
 853   // Accumulator for the remembered set size
 854   size_t _total_remset_bytes;
 855 
 856   // Accumulator for strong code roots memory size
 857   size_t _total_strong_code_roots_bytes;
 858 
 859   static double bytes_to_mb(size_t val) {
 860     return (double) val / (double) M;
 861   }
 862 
 863 public:
 864   // The header and footer are printed in the constructor and
 865   // destructor respectively.
 866   G1PrintRegionLivenessInfoClosure(const char* phase_name);
 867   virtual bool do_heap_region(HeapRegion* r);
 868   ~G1PrintRegionLivenessInfoClosure();
 869 };
 870 
 871 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP