1 /*
   2  * Copyright (c) 2001, 2020, 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_GC_G1_G1CONCURRENTMARK_HPP
  26 #define SHARE_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/taskTerminator.hpp"
  34 #include "gc/shared/taskqueue.hpp"
  35 #include "gc/shared/verifyOption.hpp"
  36 #include "gc/shared/workgroup.hpp"
  37 #include "memory/allocation.hpp"
  38 #include "utilities/compilerWarnings.hpp"
  39 
  40 class ConcurrentGCTimer;
  41 class G1ConcurrentMarkThread;
  42 class G1CollectedHeap;
  43 class G1CMOopClosure;
  44 class G1CMTask;
  45 class G1ConcurrentMark;
  46 class G1OldTracer;
  47 class G1RegionToSpaceMapper;
  48 class G1SurvivorRegions;
  49 class ThreadClosure;
  50 
  51 PRAGMA_DIAG_PUSH
  52 // warning C4522: multiple assignment operators specified
  53 PRAGMA_DISABLE_MSVC_WARNING(4522)
  54 
  55 // This is a container class for either an oop or a continuation address for
  56 // mark stack entries. Both are pushed onto the mark stack.
  57 class G1TaskQueueEntry {
  58 private:
  59   void* _holder;
  60 
  61   static const uintptr_t ArraySliceBit = 1;
  62 
  63   G1TaskQueueEntry(oop obj) : _holder(obj) {
  64     assert(_holder != NULL, "Not allowed to set NULL task queue element");
  65   }
  66   G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
  67 public:
  68 
  69   G1TaskQueueEntry() : _holder(NULL) { }
  70   // Trivially copyable, for use in GenericTaskQueue.
  71 
  72   static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
  73   static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
  74 
  75   oop obj() const {
  76     assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
  77     return (oop)_holder;
  78   }
  79 
  80   HeapWord* slice() const {
  81     assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
  82     return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
  83   }
  84 
  85   bool is_oop() const { return !is_array_slice(); }
  86   bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
  87   bool is_null() const { return _holder == NULL; }
  88 };
  89 
  90 PRAGMA_DIAG_POP
  91 
  92 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
  93 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
  94 
  95 // Closure used by CM during concurrent reference discovery
  96 // and reference processing (during remarking) to determine
  97 // if a particular object is alive. It is primarily used
  98 // to determine if referents of discovered reference objects
  99 // are alive. An instance is also embedded into the
 100 // reference processor as the _is_alive_non_header field
 101 class G1CMIsAliveClosure : public BoolObjectClosure {
 102   G1CollectedHeap* _g1h;
 103 public:
 104   G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
 105   bool do_object_b(oop obj);
 106 };
 107 
 108 class G1CMSubjectToDiscoveryClosure : public BoolObjectClosure {
 109   G1CollectedHeap* _g1h;
 110 public:
 111   G1CMSubjectToDiscoveryClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
 112   bool do_object_b(oop obj);
 113 };
 114 
 115 // Represents the overflow mark stack used by concurrent marking.
 116 //
 117 // Stores oops in a huge buffer in virtual memory that is always fully committed.
 118 // Resizing may only happen during a STW pause when the stack is empty.
 119 //
 120 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
 121 // stack memory is split into evenly sized chunks of oops. Users can only
 122 // add or remove entries on that basis.
 123 // Chunks are filled in increasing address order. Not completely filled chunks
 124 // have a NULL element as a terminating element.
 125 //
 126 // Every chunk has a header containing a single pointer element used for memory
 127 // management. This wastes some space, but is negligible (< .1% with current sizing).
 128 //
 129 // Memory management is done using a mix of tracking a high water-mark indicating
 130 // that all chunks at a lower address are valid chunks, and a singly linked free
 131 // list connecting all empty chunks.
 132 class G1CMMarkStack {
 133 public:
 134   // Number of TaskQueueEntries that can fit in a single chunk.
 135   static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
 136 private:
 137   struct TaskQueueEntryChunk {
 138     TaskQueueEntryChunk* next;
 139     G1TaskQueueEntry data[EntriesPerChunk];
 140   };
 141 
 142   size_t _max_chunk_capacity;    // Maximum number of TaskQueueEntryChunk elements on the stack.
 143 
 144   TaskQueueEntryChunk* _base;    // Bottom address of allocated memory area.
 145   size_t _chunk_capacity;        // Current maximum number of TaskQueueEntryChunk elements.
 146 
 147   char _pad0[DEFAULT_CACHE_LINE_SIZE];
 148   TaskQueueEntryChunk* volatile _free_list;  // Linked list of free chunks that can be allocated by users.
 149   char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
 150   TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
 151   volatile size_t _chunks_in_chunk_list;
 152   char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
 153 
 154   volatile size_t _hwm;          // High water mark within the reserved space.
 155   char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
 156 
 157   // Allocate a new chunk from the reserved memory, using the high water mark. Returns
 158   // NULL if out of memory.
 159   TaskQueueEntryChunk* allocate_new_chunk();
 160 
 161   // Atomically add the given chunk to the list.
 162   void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
 163   // Atomically remove and return a chunk from the given list. Returns NULL if the
 164   // list is empty.
 165   TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
 166 
 167   void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
 168   void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
 169 
 170   TaskQueueEntryChunk* remove_chunk_from_chunk_list();
 171   TaskQueueEntryChunk* remove_chunk_from_free_list();
 172 
 173   // Resizes the mark stack to the given new capacity. Releases any previous
 174   // memory if successful.
 175   bool resize(size_t new_capacity);
 176 
 177  public:
 178   G1CMMarkStack();
 179   ~G1CMMarkStack();
 180 
 181   // Alignment and minimum capacity of this mark stack in number of oops.
 182   static size_t capacity_alignment();
 183 
 184   // Allocate and initialize the mark stack with the given number of oops.
 185   bool initialize(size_t initial_capacity, size_t max_capacity);
 186 
 187   // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
 188   // stack. If less than EntriesPerChunk elements are to be pushed, the array must
 189   // be terminated with a NULL.
 190   // Returns whether the buffer contents were successfully pushed to the global mark
 191   // stack.
 192   bool par_push_chunk(G1TaskQueueEntry* buffer);
 193 
 194   // Pops a chunk from this mark stack, copying them into the given buffer. This
 195   // chunk may contain up to EntriesPerChunk elements. If there are less, the last
 196   // element in the array is a NULL pointer.
 197   bool par_pop_chunk(G1TaskQueueEntry* buffer);
 198 
 199   // Return whether the chunk list is empty. Racy due to unsynchronized access to
 200   // _chunk_list.
 201   bool is_empty() const { return _chunk_list == NULL; }
 202 
 203   size_t capacity() const  { return _chunk_capacity; }
 204 
 205   // Expand the stack, typically in response to an overflow condition
 206   void expand();
 207 
 208   // Return the approximate number of oops on this mark stack. Racy due to
 209   // unsynchronized access to _chunks_in_chunk_list.
 210   size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
 211 
 212   void set_empty();
 213 
 214   // Apply Fn to every oop on the mark stack. The mark stack must not
 215   // be modified while iterating.
 216   template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
 217 };
 218 
 219 // Root MemRegions are memory areas that contain objects which references are
 220 // roots wrt to the marking. They must be scanned before marking to maintain the
 221 // SATB invariant.
 222 // Typically they contain the areas from nTAMS to top of the regions.
 223 // We could scan and mark through these objects during the concurrent start pause,
 224 // but for pause time reasons we move this work to the concurrent phase.
 225 // We need to complete this procedure before the next GC because it might determine
 226 // that some of these "root objects" are dead, potentially dropping some required
 227 // references.
 228 // Root MemRegions comprise of the contents of survivor regions at the end
 229 // of the GC, and any objects copied into the old gen during GC.
 230 class G1CMRootMemRegions {
 231   // The set of root MemRegions.
 232   MemRegion*   _root_regions;
 233   size_t const _max_regions;
 234 
 235   volatile size_t _num_root_regions; // Actual number of root regions.
 236 
 237   volatile size_t _claimed_root_regions; // Number of root regions currently claimed.
 238 
 239   volatile bool _scan_in_progress;
 240   volatile bool _should_abort;
 241 
 242   void notify_scan_done();
 243 
 244 public:
 245   G1CMRootMemRegions(uint const max_regions);
 246   ~G1CMRootMemRegions();
 247 
 248   // Reset the data structure to allow addition of new root regions.
 249   void reset();
 250 
 251   void add(HeapWord* start, HeapWord* end);
 252 
 253   // Reset the claiming / scanning of the root regions.
 254   void prepare_for_scan();
 255 
 256   // Forces get_next() to return NULL so that the iteration aborts early.
 257   void abort() { _should_abort = true; }
 258 
 259   // Return true if the CM thread are actively scanning root regions,
 260   // false otherwise.
 261   bool scan_in_progress() { return _scan_in_progress; }
 262 
 263   // Claim the next root MemRegion to scan atomically, or return NULL if
 264   // all have been claimed.
 265   const MemRegion* claim_next();
 266 
 267   // The number of root regions to scan.
 268   uint num_root_regions() const;
 269 
 270   void cancel_scan();
 271 
 272   // Flag that we're done with root region scanning and notify anyone
 273   // who's waiting on it. If aborted is false, assume that all regions
 274   // have been claimed.
 275   void scan_finished();
 276 
 277   // If CM threads are still scanning root regions, wait until they
 278   // are done. Return true if we had to wait, false otherwise.
 279   bool wait_until_scan_finished();
 280 };
 281 
 282 // This class manages data structures and methods for doing liveness analysis in
 283 // G1's concurrent cycle.
 284 class G1ConcurrentMark : public CHeapObj<mtGC> {
 285   friend class G1ConcurrentMarkThread;
 286   friend class G1CMRefProcTaskProxy;
 287   friend class G1CMRefProcTaskExecutor;
 288   friend class G1CMKeepAliveAndDrainClosure;
 289   friend class G1CMDrainMarkingStackClosure;
 290   friend class G1CMBitMapClosure;
 291   friend class G1CMConcurrentMarkingTask;
 292   friend class G1CMRemarkTask;
 293   friend class G1CMTask;
 294 
 295   G1ConcurrentMarkThread* _cm_thread;     // The thread doing the work
 296   G1CollectedHeap*        _g1h;           // The heap
 297 
 298   // Concurrent marking support structures
 299   G1CMBitMap              _mark_bitmap_1;
 300   G1CMBitMap              _mark_bitmap_2;
 301   G1CMBitMap*             _prev_mark_bitmap; // Completed mark bitmap
 302   G1CMBitMap*             _next_mark_bitmap; // Under-construction mark bitmap
 303 
 304   // Heap bounds
 305   MemRegion const         _heap;
 306 
 307   // Root region tracking and claiming
 308   G1CMRootMemRegions         _root_regions;
 309 
 310   // For grey objects
 311   G1CMMarkStack           _global_mark_stack; // Grey objects behind global finger
 312   HeapWord* volatile      _finger;            // The global finger, region aligned,
 313                                               // always pointing to the end of the
 314                                               // last claimed region
 315 
 316   uint                    _worker_id_offset;
 317   uint                    _max_num_tasks;    // Maximum number of marking tasks
 318   uint                    _num_active_tasks; // Number of tasks currently active
 319   G1CMTask**              _tasks;            // Task queue array (max_worker_id length)
 320 
 321   G1CMTaskQueueSet*       _task_queues; // Task queue set
 322   TaskTerminator          _terminator;  // For termination
 323 
 324   // Two sync barriers that are used to synchronize tasks when an
 325   // overflow occurs. The algorithm is the following. All tasks enter
 326   // the first one to ensure that they have all stopped manipulating
 327   // the global data structures. After they exit it, they re-initialize
 328   // their data structures and task 0 re-initializes the global data
 329   // structures. Then, they enter the second sync barrier. This
 330   // ensure, that no task starts doing work before all data
 331   // structures (local and global) have been re-initialized. When they
 332   // exit it, they are free to start working again.
 333   WorkGangBarrierSync     _first_overflow_barrier_sync;
 334   WorkGangBarrierSync     _second_overflow_barrier_sync;
 335 
 336   // This is set by any task, when an overflow on the global data
 337   // structures is detected
 338   volatile bool           _has_overflown;
 339   // True: marking is concurrent, false: we're in remark
 340   volatile bool           _concurrent;
 341   // Set at the end of a Full GC so that marking aborts
 342   volatile bool           _has_aborted;
 343 
 344   // Used when remark aborts due to an overflow to indicate that
 345   // another concurrent marking phase should start
 346   volatile bool           _restart_for_overflow;
 347 
 348   ConcurrentGCTimer*      _gc_timer_cm;
 349 
 350   G1OldTracer*            _gc_tracer_cm;
 351 
 352   // Timing statistics. All of them are in ms
 353   NumberSeq _init_times;
 354   NumberSeq _remark_times;
 355   NumberSeq _remark_mark_times;
 356   NumberSeq _remark_weak_ref_times;
 357   NumberSeq _cleanup_times;
 358   double    _total_cleanup_time;
 359 
 360   double*   _accum_task_vtime;   // Accumulated task vtime
 361 
 362   WorkGang* _concurrent_workers;
 363   uint      _num_concurrent_workers; // The number of marking worker threads we're using
 364   uint      _max_concurrent_workers; // Maximum number of marking worker threads
 365 
 366   void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller);
 367 
 368   void finalize_marking();
 369 
 370   void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);
 371   void weak_refs_work(bool clear_all_soft_refs);
 372 
 373   void report_object_count(bool mark_completed);
 374 
 375   void swap_mark_bitmaps();
 376 
 377   void reclaim_empty_regions();
 378 
 379   // After reclaiming empty regions, update heap sizes.
 380   void compute_new_sizes();
 381 
 382   // Clear statistics gathered during the concurrent cycle for the given region after
 383   // it has been reclaimed.
 384   void clear_statistics(HeapRegion* r);
 385 
 386   // Resets the global marking data structures, as well as the
 387   // task local ones; should be called during concurrent start.
 388   void reset();
 389 
 390   // Resets all the marking data structures. Called when we have to restart
 391   // marking or when marking completes (via set_non_marking_state below).
 392   void reset_marking_for_restart();
 393 
 394   // We do this after we're done with marking so that the marking data
 395   // structures are initialized to a sensible and predictable state.
 396   void reset_at_marking_complete();
 397 
 398   // Called to indicate how many threads are currently active.
 399   void set_concurrency(uint active_tasks);
 400 
 401   // Should be called to indicate which phase we're in (concurrent
 402   // mark or remark) and how many threads are currently active.
 403   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 404 
 405   // Prints all gathered CM-related statistics
 406   void print_stats();
 407 
 408   HeapWord*           finger()       { return _finger;   }
 409   bool                concurrent()   { return _concurrent; }
 410   uint                active_tasks() { return _num_active_tasks; }
 411   TaskTerminator*     terminator()   { return &_terminator; }
 412 
 413   // Claims the next available region to be scanned by a marking
 414   // task/thread. It might return NULL if the next region is empty or
 415   // we have run out of regions. In the latter case, out_of_regions()
 416   // determines whether we've really run out of regions or the task
 417   // should call claim_region() again. This might seem a bit
 418   // awkward. Originally, the code was written so that claim_region()
 419   // either successfully returned with a non-empty region or there
 420   // were no more regions to be claimed. The problem with this was
 421   // that, in certain circumstances, it iterated over large chunks of
 422   // the heap finding only empty regions and, while it was working, it
 423   // was preventing the calling task to call its regular clock
 424   // method. So, this way, each task will spend very little time in
 425   // claim_region() and is allowed to call the regular clock method
 426   // frequently.
 427   HeapRegion* claim_region(uint worker_id);
 428 
 429   // Determines whether we've run out of regions to scan. Note that
 430   // the finger can point past the heap end in case the heap was expanded
 431   // to satisfy an allocation without doing a GC. This is fine, because all
 432   // objects in those regions will be considered live anyway because of
 433   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 434   bool out_of_regions() { return _finger >= _heap.end(); }
 435 
 436   // Returns the task with the given id
 437   G1CMTask* task(uint id) {
 438     // During concurrent start we use the parallel gc threads to do some work, so
 439     // we can only compare against _max_num_tasks.
 440     assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks);
 441     return _tasks[id];
 442   }
 443 
 444   // Access / manipulation of the overflow flag which is set to
 445   // indicate that the global stack has overflown
 446   bool has_overflown()           { return _has_overflown; }
 447   void set_has_overflown()       { _has_overflown = true; }
 448   void clear_has_overflown()     { _has_overflown = false; }
 449   bool restart_for_overflow()    { return _restart_for_overflow; }
 450 
 451   // Methods to enter the two overflow sync barriers
 452   void enter_first_sync_barrier(uint worker_id);
 453   void enter_second_sync_barrier(uint worker_id);
 454 
 455   // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
 456   // true, periodically insert checks to see if this method should exit prematurely.
 457   void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
 458 
 459   // Region statistics gathered during marking.
 460   G1RegionMarkStats* _region_mark_stats;
 461   // Top pointer for each region at the start of the rebuild remembered set process
 462   // for regions which remembered sets need to be rebuilt. A NULL for a given region
 463   // means that this region does not be scanned during the rebuilding remembered
 464   // set phase at all.
 465   HeapWord* volatile* _top_at_rebuild_starts;
 466 public:
 467   void add_to_liveness(uint worker_id, oop const obj, size_t size);
 468   // Liveness of the given region as determined by concurrent marking, i.e. the amount of
 469   // live words between bottom and nTAMS.
 470   size_t liveness(uint region) const { return _region_mark_stats[region]._live_words; }
 471 
 472   // Sets the internal top_at_region_start for the given region to current top of the region.
 473   inline void update_top_at_rebuild_start(HeapRegion* r);
 474   // TARS for the given region during remembered set rebuilding.
 475   inline HeapWord* top_at_rebuild_start(uint region) const;
 476 
 477   // Clear statistics gathered during the concurrent cycle for the given region after
 478   // it has been reclaimed.
 479   void clear_statistics_in_region(uint region_idx);
 480   // Notification for eagerly reclaimed regions to clean up.
 481   void humongous_object_eagerly_reclaimed(HeapRegion* r);
 482   // Manipulation of the global mark stack.
 483   // The push and pop operations are used by tasks for transfers
 484   // between task-local queues and the global mark stack.
 485   bool mark_stack_push(G1TaskQueueEntry* arr) {
 486     if (!_global_mark_stack.par_push_chunk(arr)) {
 487       set_has_overflown();
 488       return false;
 489     }
 490     return true;
 491   }
 492   bool mark_stack_pop(G1TaskQueueEntry* arr) {
 493     return _global_mark_stack.par_pop_chunk(arr);
 494   }
 495   size_t mark_stack_size() const                { return _global_mark_stack.size(); }
 496   size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; }
 497   bool mark_stack_empty() const                 { return _global_mark_stack.is_empty(); }
 498 
 499   G1CMRootMemRegions* root_regions() { return &_root_regions; }
 500 
 501   void concurrent_cycle_start();
 502   // Abandon current marking iteration due to a Full GC.
 503   void concurrent_cycle_abort();
 504   void concurrent_cycle_end();
 505 
 506   void update_accum_task_vtime(int i, double vtime) {
 507     _accum_task_vtime[i] += vtime;
 508   }
 509 
 510   double all_task_accum_vtime() {
 511     double ret = 0.0;
 512     for (uint i = 0; i < _max_num_tasks; ++i)
 513       ret += _accum_task_vtime[i];
 514     return ret;
 515   }
 516 
 517   // Attempts to steal an object from the task queues of other tasks
 518   bool try_stealing(uint worker_id, G1TaskQueueEntry& task_entry);
 519 
 520   G1ConcurrentMark(G1CollectedHeap* g1h,
 521                    G1RegionToSpaceMapper* prev_bitmap_storage,
 522                    G1RegionToSpaceMapper* next_bitmap_storage);
 523   ~G1ConcurrentMark();
 524 
 525   G1ConcurrentMarkThread* cm_thread() { return _cm_thread; }
 526 
 527   const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; }
 528   G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; }
 529 
 530   // Calculates the number of concurrent GC threads to be used in the marking phase.
 531   uint calc_active_marking_workers();
 532 
 533   // Moves all per-task cached data into global state.
 534   void flush_all_task_caches();
 535   // Prepare internal data structures for the next mark cycle. This includes clearing
 536   // the next mark bitmap and some internal data structures. This method is intended
 537   // to be called concurrently to the mutator. It will yield to safepoint requests.
 538   void cleanup_for_next_mark();
 539 
 540   // Clear the previous marking bitmap during safepoint.
 541   void clear_prev_bitmap(WorkGang* workers);
 542 
 543   // These two methods do the work that needs to be done at the start and end of the
 544   // concurrent start pause.
 545   void pre_concurrent_start();
 546   void post_concurrent_start();
 547 
 548   // Scan all the root regions and mark everything reachable from
 549   // them.
 550   void scan_root_regions();
 551 
 552   // Scan a single root MemRegion to mark everything reachable from it.
 553   void scan_root_region(const MemRegion* region, uint worker_id);
 554 
 555   // Do concurrent phase of marking, to a tentative transitive closure.
 556   void mark_from_roots();
 557 
 558   // Do concurrent preclean work.
 559   void preclean();
 560 
 561   void remark();
 562 
 563   void cleanup();
 564   // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use
 565   // this carefully.
 566   inline void mark_in_prev_bitmap(oop p);
 567 
 568   // Clears marks for all objects in the given range, for the prev or
 569   // next bitmaps.  Caution: the previous bitmap is usually
 570   // read-only, so use this carefully!
 571   void clear_range_in_prev_bitmap(MemRegion mr);
 572 
 573   inline bool is_marked_in_prev_bitmap(oop p) const;
 574 
 575   // Verify that there are no collection set oops on the stacks (taskqueues /
 576   // global mark stack) and fingers (global / per-task).
 577   // If marking is not in progress, it's a no-op.
 578   void verify_no_collection_set_oops() PRODUCT_RETURN;
 579 
 580   inline bool do_yield_check();
 581 
 582   bool has_aborted()      { return _has_aborted; }
 583 
 584   void print_summary_info();
 585 
 586   void threads_do(ThreadClosure* tc) const;
 587 
 588   void print_on_error(outputStream* st) const;
 589 
 590   // Mark the given object on the next bitmap if it is below nTAMS.
 591   inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj);
 592   inline bool mark_in_next_bitmap(uint worker_id, oop const obj);
 593 
 594   inline bool is_marked_in_next_bitmap(oop p) const;
 595 
 596   ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
 597   G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
 598 
 599 private:
 600   // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application.
 601   void rebuild_rem_set_concurrently();
 602 };
 603 
 604 // A class representing a marking task.
 605 class G1CMTask : public TerminatorTerminator {
 606 private:
 607   enum PrivateConstants {
 608     // The regular clock call is called once the scanned words reaches
 609     // this limit
 610     words_scanned_period          = 12*1024,
 611     // The regular clock call is called once the number of visited
 612     // references reaches this limit
 613     refs_reached_period           = 1024,
 614     // Initial value for the hash seed, used in the work stealing code
 615     init_hash_seed                = 17
 616   };
 617 
 618   // Number of entries in the per-task stats entry. This seems enough to have a very
 619   // low cache miss rate.
 620   static const uint RegionMarkStatsCacheSize = 1024;
 621 
 622   G1CMObjArrayProcessor       _objArray_processor;
 623 
 624   uint                        _worker_id;
 625   G1CollectedHeap*            _g1h;
 626   G1ConcurrentMark*           _cm;
 627   G1CMBitMap*                 _next_mark_bitmap;
 628   // the task queue of this task
 629   G1CMTaskQueue*              _task_queue;
 630 
 631   G1RegionMarkStatsCache      _mark_stats_cache;
 632   // Number of calls to this task
 633   uint                        _calls;
 634 
 635   // When the virtual timer reaches this time, the marking step should exit
 636   double                      _time_target_ms;
 637   // Start time of the current marking step
 638   double                      _start_time_ms;
 639 
 640   // Oop closure used for iterations over oops
 641   G1CMOopClosure*             _cm_oop_closure;
 642 
 643   // Region this task is scanning, NULL if we're not scanning any
 644   HeapRegion*                 _curr_region;
 645   // Local finger of this task, NULL if we're not scanning a region
 646   HeapWord*                   _finger;
 647   // Limit of the region this task is scanning, NULL if we're not scanning one
 648   HeapWord*                   _region_limit;
 649 
 650   // Number of words this task has scanned
 651   size_t                      _words_scanned;
 652   // When _words_scanned reaches this limit, the regular clock is
 653   // called. Notice that this might be decreased under certain
 654   // circumstances (i.e. when we believe that we did an expensive
 655   // operation).
 656   size_t                      _words_scanned_limit;
 657   // Initial value of _words_scanned_limit (i.e. what it was
 658   // before it was decreased).
 659   size_t                      _real_words_scanned_limit;
 660 
 661   // Number of references this task has visited
 662   size_t                      _refs_reached;
 663   // When _refs_reached reaches this limit, the regular clock is
 664   // called. Notice this this might be decreased under certain
 665   // circumstances (i.e. when we believe that we did an expensive
 666   // operation).
 667   size_t                      _refs_reached_limit;
 668   // Initial value of _refs_reached_limit (i.e. what it was before
 669   // it was decreased).
 670   size_t                      _real_refs_reached_limit;
 671 
 672   // If true, then the task has aborted for some reason
 673   bool                        _has_aborted;
 674   // Set when the task aborts because it has met its time quota
 675   bool                        _has_timed_out;
 676   // True when we're draining SATB buffers; this avoids the task
 677   // aborting due to SATB buffers being available (as we're already
 678   // dealing with them)
 679   bool                        _draining_satb_buffers;
 680 
 681   // Number sequence of past step times
 682   NumberSeq                   _step_times_ms;
 683   // Elapsed time of this task
 684   double                      _elapsed_time_ms;
 685   // Termination time of this task
 686   double                      _termination_time_ms;
 687   // When this task got into the termination protocol
 688   double                      _termination_start_time_ms;
 689 
 690   TruncatedSeq                _marking_step_diff_ms;
 691 
 692   // Updates the local fields after this task has claimed
 693   // a new region to scan
 694   void setup_for_region(HeapRegion* hr);
 695   // Makes the limit of the region up-to-date
 696   void update_region_limit();
 697 
 698   // Called when either the words scanned or the refs visited limit
 699   // has been reached
 700   void reached_limit();
 701   // Recalculates the words scanned and refs visited limits
 702   void recalculate_limits();
 703   // Decreases the words scanned and refs visited limits when we reach
 704   // an expensive operation
 705   void decrease_limits();
 706   // Checks whether the words scanned or refs visited reached their
 707   // respective limit and calls reached_limit() if they have
 708   void check_limits() {
 709     if (_words_scanned >= _words_scanned_limit ||
 710         _refs_reached >= _refs_reached_limit) {
 711       reached_limit();
 712     }
 713   }
 714   // Supposed to be called regularly during a marking step as
 715   // it checks a bunch of conditions that might cause the marking step
 716   // to abort
 717   // Return true if the marking step should continue. Otherwise, return false to abort
 718   bool regular_clock_call();
 719 
 720   // Set abort flag if regular_clock_call() check fails
 721   inline void abort_marking_if_regular_check_fail();
 722 
 723   // Test whether obj might have already been passed over by the
 724   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 725   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 726 
 727   template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
 728 public:
 729   // Apply the closure on the given area of the objArray. Return the number of words
 730   // scanned.
 731   inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
 732   // Resets the task; should be called right at the beginning of a marking phase.
 733   void reset(G1CMBitMap* next_mark_bitmap);
 734   // Clears all the fields that correspond to a claimed region.
 735   void clear_region_fields();
 736 
 737   // The main method of this class which performs a marking step
 738   // trying not to exceed the given duration. However, it might exit
 739   // prematurely, according to some conditions (i.e. SATB buffers are
 740   // available for processing).
 741   void do_marking_step(double target_ms,
 742                        bool do_termination,
 743                        bool is_serial);
 744 
 745   // These two calls start and stop the timer
 746   void record_start_time() {
 747     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 748   }
 749   void record_end_time() {
 750     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 751   }
 752 
 753   // Returns the worker ID associated with this task.
 754   uint worker_id() { return _worker_id; }
 755 
 756   // From TerminatorTerminator. It determines whether this task should
 757   // exit the termination protocol after it's entered it.
 758   virtual bool should_exit_termination();
 759 
 760   // Resets the local region fields after a task has finished scanning a
 761   // region; or when they have become stale as a result of the region
 762   // being evacuated.
 763   void giveup_current_region();
 764 
 765   HeapWord* finger()            { return _finger; }
 766 
 767   bool has_aborted()            { return _has_aborted; }
 768   void set_has_aborted()        { _has_aborted = true; }
 769   void clear_has_aborted()      { _has_aborted = false; }
 770 
 771   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
 772 
 773   // Increment the number of references this task has visited.
 774   void increment_refs_reached() { ++_refs_reached; }
 775 
 776   // Grey the object by marking it.  If not already marked, push it on
 777   // the local queue if below the finger. obj is required to be below its region's NTAMS.
 778   // Returns whether there has been a mark to the bitmap.
 779   inline bool make_reference_grey(oop obj);
 780 
 781   // Grey the object (by calling make_grey_reference) if required,
 782   // e.g. obj is below its containing region's NTAMS.
 783   // Precondition: obj is a valid heap object.
 784   // Returns true if the reference caused a mark to be set in the next bitmap.
 785   template <class T>
 786   inline bool deal_with_reference(T* p);
 787 
 788   // Scans an object and visits its children.
 789   inline void scan_task_entry(G1TaskQueueEntry task_entry);
 790 
 791   // Pushes an object on the local queue.
 792   inline void push(G1TaskQueueEntry task_entry);
 793 
 794   // Move entries to the global stack.
 795   void move_entries_to_global_stack();
 796   // Move entries from the global stack, return true if we were successful to do so.
 797   bool get_entries_from_global_stack();
 798 
 799   // Pops and scans objects from the local queue. If partially is
 800   // true, then it stops when the queue size is of a given limit. If
 801   // partially is false, then it stops when the queue is empty.
 802   void drain_local_queue(bool partially);
 803   // Moves entries from the global stack to the local queue and
 804   // drains the local queue. If partially is true, then it stops when
 805   // both the global stack and the local queue reach a given size. If
 806   // partially if false, it tries to empty them totally.
 807   void drain_global_stack(bool partially);
 808   // Keeps picking SATB buffers and processing them until no SATB
 809   // buffers are available.
 810   void drain_satb_buffers();
 811 
 812   // Moves the local finger to a new location
 813   inline void move_finger_to(HeapWord* new_finger) {
 814     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
 815     _finger = new_finger;
 816   }
 817 
 818   G1CMTask(uint worker_id,
 819            G1ConcurrentMark *cm,
 820            G1CMTaskQueue* task_queue,
 821            G1RegionMarkStats* mark_stats,
 822            uint max_regions);
 823 
 824   inline void update_liveness(oop const obj, size_t const obj_size);
 825 
 826   // Clear (without flushing) the mark cache entry for the given region.
 827   void clear_mark_stats_cache(uint region_idx);
 828   // Evict the whole statistics cache into the global statistics. Returns the
 829   // number of cache hits and misses so far.
 830   Pair<size_t, size_t> flush_mark_stats_cache();
 831   // Prints statistics associated with this task
 832   void print_stats();
 833 };
 834 
 835 // Class that's used to to print out per-region liveness
 836 // information. It's currently used at the end of marking and also
 837 // after we sort the old regions at the end of the cleanup operation.
 838 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure {
 839   // Accumulators for these values.
 840   size_t _total_used_bytes;
 841   size_t _total_capacity_bytes;
 842   size_t _total_prev_live_bytes;
 843   size_t _total_next_live_bytes;
 844 
 845   // Accumulator for the remembered set size
 846   size_t _total_remset_bytes;
 847 
 848   // Accumulator for strong code roots memory size
 849   size_t _total_strong_code_roots_bytes;
 850 
 851   static double bytes_to_mb(size_t val) {
 852     return (double) val / (double) M;
 853   }
 854 
 855 public:
 856   // The header and footer are printed in the constructor and
 857   // destructor respectively.
 858   G1PrintRegionLivenessInfoClosure(const char* phase_name);
 859   virtual bool do_heap_region(HeapRegion* r);
 860   ~G1PrintRegionLivenessInfoClosure();
 861 };
 862 
 863 #endif // SHARE_GC_G1_G1CONCURRENTMARK_HPP