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   G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }
  69   G1TaskQueueEntry() : _holder(NULL) { }
  70 
  71   static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
  72   static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
  73 
  74   G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {
  75     _holder = t._holder;
  76     return *this;
  77   }
  78 
  79   volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
  80     _holder = t._holder;
  81     return *this;
  82   }
  83 
  84   oop obj() const {
  85     assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
  86     return (oop)_holder;
  87   }
  88 
  89   HeapWord* slice() const {
  90     assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
  91     return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
  92   }
  93 
  94   bool is_oop() const { return !is_array_slice(); }
  95   bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
  96   bool is_null() const { return _holder == NULL; }
  97 };
  98 
  99 PRAGMA_DIAG_POP
 100 
 101 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
 102 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
 103 
 104 // Closure used by CM during concurrent reference discovery
 105 // and reference processing (during remarking) to determine
 106 // if a particular object is alive. It is primarily used
 107 // to determine if referents of discovered reference objects
 108 // are alive. An instance is also embedded into the
 109 // reference processor as the _is_alive_non_header field
 110 class G1CMIsAliveClosure : public BoolObjectClosure {
 111   G1CollectedHeap* _g1h;
 112 public:
 113   G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
 114   bool do_object_b(oop obj);
 115 };
 116 
 117 class G1CMSubjectToDiscoveryClosure : public BoolObjectClosure {
 118   G1CollectedHeap* _g1h;
 119 public:
 120   G1CMSubjectToDiscoveryClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
 121   bool do_object_b(oop obj);
 122 };
 123 
 124 // Represents the overflow mark stack used by concurrent marking.
 125 //
 126 // Stores oops in a huge buffer in virtual memory that is always fully committed.
 127 // Resizing may only happen during a STW pause when the stack is empty.
 128 //
 129 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
 130 // stack memory is split into evenly sized chunks of oops. Users can only
 131 // add or remove entries on that basis.
 132 // Chunks are filled in increasing address order. Not completely filled chunks
 133 // have a NULL element as a terminating element.
 134 //
 135 // Every chunk has a header containing a single pointer element used for memory
 136 // management. This wastes some space, but is negligible (< .1% with current sizing).
 137 //
 138 // Memory management is done using a mix of tracking a high water-mark indicating
 139 // that all chunks at a lower address are valid chunks, and a singly linked free
 140 // list connecting all empty chunks.
 141 class G1CMMarkStack {
 142 public:
 143   // Number of TaskQueueEntries that can fit in a single chunk.
 144   static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
 145 private:
 146   struct TaskQueueEntryChunk {
 147     TaskQueueEntryChunk* next;
 148     G1TaskQueueEntry data[EntriesPerChunk];
 149   };
 150 
 151   size_t _max_chunk_capacity;    // Maximum number of TaskQueueEntryChunk elements on the stack.
 152 
 153   TaskQueueEntryChunk* _base;    // Bottom address of allocated memory area.
 154   size_t _chunk_capacity;        // Current maximum number of TaskQueueEntryChunk elements.
 155 
 156   char _pad0[DEFAULT_CACHE_LINE_SIZE];
 157   TaskQueueEntryChunk* volatile _free_list;  // Linked list of free chunks that can be allocated by users.
 158   char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
 159   TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
 160   volatile size_t _chunks_in_chunk_list;
 161   char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
 162 
 163   volatile size_t _hwm;          // High water mark within the reserved space.
 164   char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
 165 
 166   // Allocate a new chunk from the reserved memory, using the high water mark. Returns
 167   // NULL if out of memory.
 168   TaskQueueEntryChunk* allocate_new_chunk();
 169 
 170   // Atomically add the given chunk to the list.
 171   void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
 172   // Atomically remove and return a chunk from the given list. Returns NULL if the
 173   // list is empty.
 174   TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
 175 
 176   void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
 177   void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
 178 
 179   TaskQueueEntryChunk* remove_chunk_from_chunk_list();
 180   TaskQueueEntryChunk* remove_chunk_from_free_list();
 181 
 182   // Resizes the mark stack to the given new capacity. Releases any previous
 183   // memory if successful.
 184   bool resize(size_t new_capacity);
 185 
 186  public:
 187   G1CMMarkStack();
 188   ~G1CMMarkStack();
 189 
 190   // Alignment and minimum capacity of this mark stack in number of oops.
 191   static size_t capacity_alignment();
 192 
 193   // Allocate and initialize the mark stack with the given number of oops.
 194   bool initialize(size_t initial_capacity, size_t max_capacity);
 195 
 196   // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
 197   // stack. If less than EntriesPerChunk elements are to be pushed, the array must
 198   // be terminated with a NULL.
 199   // Returns whether the buffer contents were successfully pushed to the global mark
 200   // stack.
 201   bool par_push_chunk(G1TaskQueueEntry* buffer);
 202 
 203   // Pops a chunk from this mark stack, copying them into the given buffer. This
 204   // chunk may contain up to EntriesPerChunk elements. If there are less, the last
 205   // element in the array is a NULL pointer.
 206   bool par_pop_chunk(G1TaskQueueEntry* buffer);
 207 
 208   // Return whether the chunk list is empty. Racy due to unsynchronized access to
 209   // _chunk_list.
 210   bool is_empty() const { return _chunk_list == NULL; }
 211 
 212   size_t capacity() const  { return _chunk_capacity; }
 213 
 214   // Expand the stack, typically in response to an overflow condition
 215   void expand();
 216 
 217   // Return the approximate number of oops on this mark stack. Racy due to
 218   // unsynchronized access to _chunks_in_chunk_list.
 219   size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
 220 
 221   void set_empty();
 222 
 223   // Apply Fn to every oop on the mark stack. The mark stack must not
 224   // be modified while iterating.
 225   template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
 226 };
 227 
 228 // Root MemRegions are memory areas that contain objects which references are
 229 // roots wrt to the marking. They must be scanned before marking to maintain the
 230 // SATB invariant.
 231 // Typically they contain the areas from nTAMS to top of the regions.
 232 // We could scan and mark through these objects during the initial-mark pause, but for
 233 // pause time reasons we move this work to the concurrent phase.
 234 // We need to complete this procedure before the next GC because it might determine
 235 // that some of these "root objects" are dead, potentially dropping some required
 236 // references.
 237 // Root MemRegions comprise of the contents of survivor regions at the end
 238 // of the GC, and any objects copied into the old gen during GC.
 239 class G1CMRootMemRegions {
 240   // The set of root MemRegions.
 241   MemRegion*   _root_regions;
 242   size_t const _max_regions;
 243 
 244   volatile size_t _num_root_regions; // Actual number of root regions.
 245 
 246   volatile size_t _claimed_root_regions; // Number of root regions currently claimed.
 247 
 248   volatile bool _scan_in_progress;
 249   volatile bool _should_abort;
 250 
 251   void notify_scan_done();
 252 
 253 public:
 254   G1CMRootMemRegions(uint const max_regions);
 255   ~G1CMRootMemRegions();
 256 
 257   // Reset the data structure to allow addition of new root regions.
 258   void reset();
 259 
 260   void add(HeapWord* start, HeapWord* end);
 261 
 262   // Reset the claiming / scanning of the root regions.
 263   void prepare_for_scan();
 264 
 265   // Forces get_next() to return NULL so that the iteration aborts early.
 266   void abort() { _should_abort = true; }
 267 
 268   // Return true if the CM thread are actively scanning root regions,
 269   // false otherwise.
 270   bool scan_in_progress() { return _scan_in_progress; }
 271 
 272   // Claim the next root MemRegion to scan atomically, or return NULL if
 273   // all have been claimed.
 274   const MemRegion* claim_next();
 275 
 276   // The number of root regions to scan.
 277   uint num_root_regions() const;
 278 
 279   void cancel_scan();
 280 
 281   // Flag that we're done with root region scanning and notify anyone
 282   // who's waiting on it. If aborted is false, assume that all regions
 283   // have been claimed.
 284   void scan_finished();
 285 
 286   // If CM threads are still scanning root regions, wait until they
 287   // are done. Return true if we had to wait, false otherwise.
 288   bool wait_until_scan_finished();
 289 };
 290 
 291 // This class manages data structures and methods for doing liveness analysis in
 292 // G1's concurrent cycle.
 293 class G1ConcurrentMark : public CHeapObj<mtGC> {
 294   friend class G1ConcurrentMarkThread;
 295   friend class G1CMRefProcTaskProxy;
 296   friend class G1CMRefProcTaskExecutor;
 297   friend class G1CMKeepAliveAndDrainClosure;
 298   friend class G1CMDrainMarkingStackClosure;
 299   friend class G1CMBitMapClosure;
 300   friend class G1CMConcurrentMarkingTask;
 301   friend class G1CMRemarkTask;
 302   friend class G1CMTask;
 303 
 304   G1ConcurrentMarkThread* _cm_thread;     // The thread doing the work
 305   G1CollectedHeap*        _g1h;           // The heap
 306 
 307   // Concurrent marking support structures
 308   G1CMBitMap              _mark_bitmap_1;
 309   G1CMBitMap              _mark_bitmap_2;
 310   G1CMBitMap*             _prev_mark_bitmap; // Completed mark bitmap
 311   G1CMBitMap*             _next_mark_bitmap; // Under-construction mark bitmap
 312 
 313   // Heap bounds
 314   MemRegion const         _heap;
 315 
 316   // Root region tracking and claiming
 317   G1CMRootMemRegions         _root_regions;
 318 
 319   // For grey objects
 320   G1CMMarkStack           _global_mark_stack; // Grey objects behind global finger
 321   HeapWord* volatile      _finger;            // The global finger, region aligned,
 322                                               // always pointing to the end of the
 323                                               // last claimed region
 324 
 325   uint                    _worker_id_offset;
 326   uint                    _max_num_tasks;    // Maximum number of marking tasks
 327   uint                    _num_active_tasks; // Number of tasks currently active
 328   G1CMTask**              _tasks;            // Task queue array (max_worker_id length)
 329 
 330   G1CMTaskQueueSet*       _task_queues; // Task queue set
 331   TaskTerminator          _terminator;  // For termination
 332 
 333   // Two sync barriers that are used to synchronize tasks when an
 334   // overflow occurs. The algorithm is the following. All tasks enter
 335   // the first one to ensure that they have all stopped manipulating
 336   // the global data structures. After they exit it, they re-initialize
 337   // their data structures and task 0 re-initializes the global data
 338   // structures. Then, they enter the second sync barrier. This
 339   // ensure, that no task starts doing work before all data
 340   // structures (local and global) have been re-initialized. When they
 341   // exit it, they are free to start working again.
 342   WorkGangBarrierSync     _first_overflow_barrier_sync;
 343   WorkGangBarrierSync     _second_overflow_barrier_sync;
 344 
 345   // This is set by any task, when an overflow on the global data
 346   // structures is detected
 347   volatile bool           _has_overflown;
 348   // True: marking is concurrent, false: we're in remark
 349   volatile bool           _concurrent;
 350   // Set at the end of a Full GC so that marking aborts
 351   volatile bool           _aborted_by_fullgc;
 352   // Set at the end of a inital mark young GC if doesn't need to do concurrent mark
 353   // so that marking aborts
 354   volatile bool           _aborted_by_initial_mark;
 355 
 356   // Used when remark aborts due to an overflow to indicate that
 357   // another concurrent marking phase should start
 358   volatile bool           _restart_for_overflow;
 359 
 360   ConcurrentGCTimer*      _gc_timer_cm;
 361 
 362   G1OldTracer*            _gc_tracer_cm;
 363 
 364   // Timing statistics. All of them are in ms
 365   NumberSeq _init_times;
 366   NumberSeq _remark_times;
 367   NumberSeq _remark_mark_times;
 368   NumberSeq _remark_weak_ref_times;
 369   NumberSeq _cleanup_times;
 370   double    _total_cleanup_time;
 371 
 372   double*   _accum_task_vtime;   // Accumulated task vtime
 373 
 374   WorkGang* _concurrent_workers;
 375   uint      _num_concurrent_workers; // The number of marking worker threads we're using
 376   uint      _max_concurrent_workers; // Maximum number of marking worker threads
 377 
 378   void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller);
 379 
 380   void finalize_marking();
 381 
 382   void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);
 383   void weak_refs_work(bool clear_all_soft_refs);
 384 
 385   void report_object_count(bool mark_completed);
 386 
 387   void swap_mark_bitmaps();
 388 
 389   void reclaim_empty_regions();
 390 
 391   // After reclaiming empty regions, update heap sizes.
 392   void compute_new_sizes();
 393 
 394   // Clear statistics gathered during the concurrent cycle for the given region after
 395   // it has been reclaimed.
 396   void clear_statistics(HeapRegion* r);
 397 
 398   // Resets the global marking data structures, as well as the
 399   // task local ones; should be called during initial mark.
 400   void reset();
 401 
 402   // Resets all the marking data structures. Called when we have to restart
 403   // marking or when marking completes (via set_non_marking_state below).
 404   void reset_marking_for_restart();
 405 
 406   // We do this after we're done with marking so that the marking data
 407   // structures are initialized to a sensible and predictable state.
 408   void reset_at_marking_complete();
 409 
 410   // Called to indicate how many threads are currently active.
 411   void set_concurrency(uint active_tasks);
 412 
 413   // Should be called to indicate which phase we're in (concurrent
 414   // mark or remark) and how many threads are currently active.
 415   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 416 
 417   // Prints all gathered CM-related statistics
 418   void print_stats();
 419 
 420   HeapWord*           finger()       { return _finger;   }
 421   bool                concurrent()   { return _concurrent; }
 422   uint                active_tasks() { return _num_active_tasks; }
 423   TaskTerminator*     terminator()   { return &_terminator; }
 424 
 425   // Claims the next available region to be scanned by a marking
 426   // task/thread. It might return NULL if the next region is empty or
 427   // we have run out of regions. In the latter case, out_of_regions()
 428   // determines whether we've really run out of regions or the task
 429   // should call claim_region() again. This might seem a bit
 430   // awkward. Originally, the code was written so that claim_region()
 431   // either successfully returned with a non-empty region or there
 432   // were no more regions to be claimed. The problem with this was
 433   // that, in certain circumstances, it iterated over large chunks of
 434   // the heap finding only empty regions and, while it was working, it
 435   // was preventing the calling task to call its regular clock
 436   // method. So, this way, each task will spend very little time in
 437   // claim_region() and is allowed to call the regular clock method
 438   // frequently.
 439   HeapRegion* claim_region(uint worker_id);
 440 
 441   // Determines whether we've run out of regions to scan. Note that
 442   // the finger can point past the heap end in case the heap was expanded
 443   // to satisfy an allocation without doing a GC. This is fine, because all
 444   // objects in those regions will be considered live anyway because of
 445   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 446   bool out_of_regions() { return _finger >= _heap.end(); }
 447 
 448   // Returns the task with the given id
 449   G1CMTask* task(uint id) {
 450     // During initial mark we use the parallel gc threads to do some work, so
 451     // we can only compare against _max_num_tasks.
 452     assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks);
 453     return _tasks[id];
 454   }
 455 
 456   // Access / manipulation of the overflow flag which is set to
 457   // indicate that the global stack has overflown
 458   bool has_overflown()           { return _has_overflown; }
 459   void set_has_overflown()       { _has_overflown = true; }
 460   void clear_has_overflown()     { _has_overflown = false; }
 461   bool restart_for_overflow()    { return _restart_for_overflow; }
 462 
 463   // Methods to enter the two overflow sync barriers
 464   void enter_first_sync_barrier(uint worker_id);
 465   void enter_second_sync_barrier(uint worker_id);
 466 
 467   // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
 468   // true, periodically insert checks to see if this method should exit prematurely.
 469   void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
 470 
 471   // Region statistics gathered during marking.
 472   G1RegionMarkStats* _region_mark_stats;
 473   // Top pointer for each region at the start of the rebuild remembered set process
 474   // for regions which remembered sets need to be rebuilt. A NULL for a given region
 475   // means that this region does not be scanned during the rebuilding remembered
 476   // set phase at all.
 477   HeapWord* volatile* _top_at_rebuild_starts;
 478 public:
 479   void add_to_liveness(uint worker_id, oop const obj, size_t size);
 480   // Liveness of the given region as determined by concurrent marking, i.e. the amount of
 481   // live words between bottom and nTAMS.
 482   size_t liveness(uint region) const { return _region_mark_stats[region]._live_words; }
 483 
 484   // Sets the internal top_at_region_start for the given region to current top of the region.
 485   inline void update_top_at_rebuild_start(HeapRegion* r);
 486   // TARS for the given region during remembered set rebuilding.
 487   inline HeapWord* top_at_rebuild_start(uint region) const;
 488 
 489   // Clear statistics gathered during the concurrent cycle for the given region after
 490   // it has been reclaimed.
 491   void clear_statistics_in_region(uint region_idx);
 492   // Notification for eagerly reclaimed regions to clean up.
 493   void humongous_object_eagerly_reclaimed(HeapRegion* r);
 494   // Manipulation of the global mark stack.
 495   // The push and pop operations are used by tasks for transfers
 496   // between task-local queues and the global mark stack.
 497   bool mark_stack_push(G1TaskQueueEntry* arr) {
 498     if (!_global_mark_stack.par_push_chunk(arr)) {
 499       set_has_overflown();
 500       return false;
 501     }
 502     return true;
 503   }
 504   bool mark_stack_pop(G1TaskQueueEntry* arr) {
 505     return _global_mark_stack.par_pop_chunk(arr);
 506   }
 507   size_t mark_stack_size() const                { return _global_mark_stack.size(); }
 508   size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; }
 509   bool mark_stack_empty() const                 { return _global_mark_stack.is_empty(); }
 510 
 511   G1CMRootMemRegions* root_regions() { return &_root_regions; }
 512 
 513   void concurrent_cycle_start();
 514   // Abandon current marking iteration due to a Full GC.
 515   void concurrent_cycle_abort_by_fullgc();
 516   void concurrent_cycle_end();
 517 
 518   void concurrent_cycle_abort_by_initial_mark();
 519 
 520   void update_accum_task_vtime(int i, double vtime) {
 521     _accum_task_vtime[i] += vtime;
 522   }
 523 
 524   double all_task_accum_vtime() {
 525     double ret = 0.0;
 526     for (uint i = 0; i < _max_num_tasks; ++i)
 527       ret += _accum_task_vtime[i];
 528     return ret;
 529   }
 530 
 531   // Attempts to steal an object from the task queues of other tasks
 532   bool try_stealing(uint worker_id, G1TaskQueueEntry& task_entry);
 533 
 534   G1ConcurrentMark(G1CollectedHeap* g1h,
 535                    G1RegionToSpaceMapper* prev_bitmap_storage,
 536                    G1RegionToSpaceMapper* next_bitmap_storage);
 537   ~G1ConcurrentMark();
 538 
 539   G1ConcurrentMarkThread* cm_thread() { return _cm_thread; }
 540 
 541   const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; }
 542   G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; }
 543 
 544   // Calculates the number of concurrent GC threads to be used in the marking phase.
 545   uint calc_active_marking_workers();
 546 
 547   // Moves all per-task cached data into global state.
 548   void flush_all_task_caches();
 549   // Prepare internal data structures for the next mark cycle. This includes clearing
 550   // the next mark bitmap and some internal data structures. This method is intended
 551   // to be called concurrently to the mutator. It will yield to safepoint requests.
 552   void cleanup_for_next_mark();
 553 
 554   // Clear the previous marking bitmap during safepoint.
 555   void clear_prev_bitmap(WorkGang* workers);
 556 
 557   // These two methods do the work that needs to be done at the start and end of the
 558   // initial mark pause.
 559   void pre_initial_mark();
 560   void post_initial_mark();
 561 
 562   // Scan all the root regions and mark everything reachable from
 563   // them.
 564   void scan_root_regions();
 565 
 566   // Scan a single root MemRegion to mark everything reachable from it.
 567   void scan_root_region(const MemRegion* region, uint worker_id);
 568 
 569   // Do concurrent phase of marking, to a tentative transitive closure.
 570   void mark_from_roots();
 571 
 572   // Do concurrent preclean work.
 573   void preclean();
 574 
 575   void remark();
 576 
 577   void cleanup();
 578   // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use
 579   // this carefully.
 580   inline void mark_in_prev_bitmap(oop p);
 581 
 582   // Clears marks for all objects in the given range, for the prev or
 583   // next bitmaps.  Caution: the previous bitmap is usually
 584   // read-only, so use this carefully!
 585   void clear_range_in_prev_bitmap(MemRegion mr);
 586 
 587   inline bool is_marked_in_prev_bitmap(oop p) const;
 588 
 589   // Verify that there are no collection set oops on the stacks (taskqueues /
 590   // global mark stack) and fingers (global / per-task).
 591   // If marking is not in progress, it's a no-op.
 592   void verify_no_collection_set_oops() PRODUCT_RETURN;
 593 
 594   inline bool do_yield_check();
 595 
 596   bool aborted_by_fullgc()             { return _aborted_by_fullgc; }
 597   bool aborted_by_initial_mark()       { return _aborted_by_initial_mark; }
 598   bool has_aborted()                   { return aborted_by_fullgc() || aborted_by_initial_mark(); }
 599 
 600   void print_summary_info();
 601 
 602   void print_worker_threads_on(outputStream* st) const;
 603   void threads_do(ThreadClosure* tc) const;
 604 
 605   void print_on_error(outputStream* st) const;
 606 
 607   // Mark the given object on the next bitmap if it is below nTAMS.
 608   inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj);
 609   inline bool mark_in_next_bitmap(uint worker_id, oop const obj);
 610 
 611   inline bool is_marked_in_next_bitmap(oop p) const;
 612 
 613   ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
 614   G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
 615 
 616 private:
 617   // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application.
 618   void rebuild_rem_set_concurrently();
 619 };
 620 
 621 // A class representing a marking task.
 622 class G1CMTask : public TerminatorTerminator {
 623 private:
 624   enum PrivateConstants {
 625     // The regular clock call is called once the scanned words reaches
 626     // this limit
 627     words_scanned_period          = 12*1024,
 628     // The regular clock call is called once the number of visited
 629     // references reaches this limit
 630     refs_reached_period           = 1024,
 631     // Initial value for the hash seed, used in the work stealing code
 632     init_hash_seed                = 17
 633   };
 634 
 635   // Number of entries in the per-task stats entry. This seems enough to have a very
 636   // low cache miss rate.
 637   static const uint RegionMarkStatsCacheSize = 1024;
 638 
 639   G1CMObjArrayProcessor       _objArray_processor;
 640 
 641   uint                        _worker_id;
 642   G1CollectedHeap*            _g1h;
 643   G1ConcurrentMark*           _cm;
 644   G1CMBitMap*                 _next_mark_bitmap;
 645   // the task queue of this task
 646   G1CMTaskQueue*              _task_queue;
 647 
 648   G1RegionMarkStatsCache      _mark_stats_cache;
 649   // Number of calls to this task
 650   uint                        _calls;
 651 
 652   // When the virtual timer reaches this time, the marking step should exit
 653   double                      _time_target_ms;
 654   // Start time of the current marking step
 655   double                      _start_time_ms;
 656 
 657   // Oop closure used for iterations over oops
 658   G1CMOopClosure*             _cm_oop_closure;
 659 
 660   // Region this task is scanning, NULL if we're not scanning any
 661   HeapRegion*                 _curr_region;
 662   // Local finger of this task, NULL if we're not scanning a region
 663   HeapWord*                   _finger;
 664   // Limit of the region this task is scanning, NULL if we're not scanning one
 665   HeapWord*                   _region_limit;
 666 
 667   // Number of words this task has scanned
 668   size_t                      _words_scanned;
 669   // When _words_scanned reaches this limit, the regular clock is
 670   // called. Notice that this might be decreased under certain
 671   // circumstances (i.e. when we believe that we did an expensive
 672   // operation).
 673   size_t                      _words_scanned_limit;
 674   // Initial value of _words_scanned_limit (i.e. what it was
 675   // before it was decreased).
 676   size_t                      _real_words_scanned_limit;
 677 
 678   // Number of references this task has visited
 679   size_t                      _refs_reached;
 680   // When _refs_reached reaches this limit, the regular clock is
 681   // called. Notice this this might be decreased under certain
 682   // circumstances (i.e. when we believe that we did an expensive
 683   // operation).
 684   size_t                      _refs_reached_limit;
 685   // Initial value of _refs_reached_limit (i.e. what it was before
 686   // it was decreased).
 687   size_t                      _real_refs_reached_limit;
 688 
 689   // If true, then the task has aborted for some reason
 690   bool                        _has_aborted;
 691   // Set when the task aborts because it has met its time quota
 692   bool                        _has_timed_out;
 693   // True when we're draining SATB buffers; this avoids the task
 694   // aborting due to SATB buffers being available (as we're already
 695   // dealing with them)
 696   bool                        _draining_satb_buffers;
 697 
 698   // Number sequence of past step times
 699   NumberSeq                   _step_times_ms;
 700   // Elapsed time of this task
 701   double                      _elapsed_time_ms;
 702   // Termination time of this task
 703   double                      _termination_time_ms;
 704   // When this task got into the termination protocol
 705   double                      _termination_start_time_ms;
 706 
 707   TruncatedSeq                _marking_step_diff_ms;
 708 
 709   // Updates the local fields after this task has claimed
 710   // a new region to scan
 711   void setup_for_region(HeapRegion* hr);
 712   // Makes the limit of the region up-to-date
 713   void update_region_limit();
 714 
 715   // Called when either the words scanned or the refs visited limit
 716   // has been reached
 717   void reached_limit();
 718   // Recalculates the words scanned and refs visited limits
 719   void recalculate_limits();
 720   // Decreases the words scanned and refs visited limits when we reach
 721   // an expensive operation
 722   void decrease_limits();
 723   // Checks whether the words scanned or refs visited reached their
 724   // respective limit and calls reached_limit() if they have
 725   void check_limits() {
 726     if (_words_scanned >= _words_scanned_limit ||
 727         _refs_reached >= _refs_reached_limit) {
 728       reached_limit();
 729     }
 730   }
 731   // Supposed to be called regularly during a marking step as
 732   // it checks a bunch of conditions that might cause the marking step
 733   // to abort
 734   // Return true if the marking step should continue. Otherwise, return false to abort
 735   bool regular_clock_call();
 736 
 737   // Set abort flag if regular_clock_call() check fails
 738   inline void abort_marking_if_regular_check_fail();
 739 
 740   // Test whether obj might have already been passed over by the
 741   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 742   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 743 
 744   template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
 745 public:
 746   // Apply the closure on the given area of the objArray. Return the number of words
 747   // scanned.
 748   inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
 749   // Resets the task; should be called right at the beginning of a marking phase.
 750   void reset(G1CMBitMap* next_mark_bitmap);
 751   // Clears all the fields that correspond to a claimed region.
 752   void clear_region_fields();
 753 
 754   // The main method of this class which performs a marking step
 755   // trying not to exceed the given duration. However, it might exit
 756   // prematurely, according to some conditions (i.e. SATB buffers are
 757   // available for processing).
 758   void do_marking_step(double target_ms,
 759                        bool do_termination,
 760                        bool is_serial);
 761 
 762   // These two calls start and stop the timer
 763   void record_start_time() {
 764     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 765   }
 766   void record_end_time() {
 767     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 768   }
 769 
 770   // Returns the worker ID associated with this task.
 771   uint worker_id() { return _worker_id; }
 772 
 773   // From TerminatorTerminator. It determines whether this task should
 774   // exit the termination protocol after it's entered it.
 775   virtual bool should_exit_termination();
 776 
 777   // Resets the local region fields after a task has finished scanning a
 778   // region; or when they have become stale as a result of the region
 779   // being evacuated.
 780   void giveup_current_region();
 781 
 782   HeapWord* finger()            { return _finger; }
 783 
 784   bool has_aborted()            { return _has_aborted; }
 785   void set_has_aborted()        { _has_aborted = true; }
 786   void clear_has_aborted()      { _has_aborted = false; }
 787 
 788   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
 789 
 790   // Increment the number of references this task has visited.
 791   void increment_refs_reached() { ++_refs_reached; }
 792 
 793   // Grey the object by marking it.  If not already marked, push it on
 794   // the local queue if below the finger. obj is required to be below its region's NTAMS.
 795   // Returns whether there has been a mark to the bitmap.
 796   inline bool make_reference_grey(oop obj);
 797 
 798   // Grey the object (by calling make_grey_reference) if required,
 799   // e.g. obj is below its containing region's NTAMS.
 800   // Precondition: obj is a valid heap object.
 801   // Returns true if the reference caused a mark to be set in the next bitmap.
 802   template <class T>
 803   inline bool deal_with_reference(T* p);
 804 
 805   // Scans an object and visits its children.
 806   inline void scan_task_entry(G1TaskQueueEntry task_entry);
 807 
 808   // Pushes an object on the local queue.
 809   inline void push(G1TaskQueueEntry task_entry);
 810 
 811   // Move entries to the global stack.
 812   void move_entries_to_global_stack();
 813   // Move entries from the global stack, return true if we were successful to do so.
 814   bool get_entries_from_global_stack();
 815 
 816   // Pops and scans objects from the local queue. If partially is
 817   // true, then it stops when the queue size is of a given limit. If
 818   // partially is false, then it stops when the queue is empty.
 819   void drain_local_queue(bool partially);
 820   // Moves entries from the global stack to the local queue and
 821   // drains the local queue. If partially is true, then it stops when
 822   // both the global stack and the local queue reach a given size. If
 823   // partially if false, it tries to empty them totally.
 824   void drain_global_stack(bool partially);
 825   // Keeps picking SATB buffers and processing them until no SATB
 826   // buffers are available.
 827   void drain_satb_buffers();
 828 
 829   // Moves the local finger to a new location
 830   inline void move_finger_to(HeapWord* new_finger) {
 831     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
 832     _finger = new_finger;
 833   }
 834 
 835   G1CMTask(uint worker_id,
 836            G1ConcurrentMark *cm,
 837            G1CMTaskQueue* task_queue,
 838            G1RegionMarkStats* mark_stats,
 839            uint max_regions);
 840 
 841   inline void update_liveness(oop const obj, size_t const obj_size);
 842 
 843   // Clear (without flushing) the mark cache entry for the given region.
 844   void clear_mark_stats_cache(uint region_idx);
 845   // Evict the whole statistics cache into the global statistics. Returns the
 846   // number of cache hits and misses so far.
 847   Pair<size_t, size_t> flush_mark_stats_cache();
 848   // Prints statistics associated with this task
 849   void print_stats();
 850 };
 851 
 852 // Class that's used to to print out per-region liveness
 853 // information. It's currently used at the end of marking and also
 854 // after we sort the old regions at the end of the cleanup operation.
 855 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure {
 856   // Accumulators for these values.
 857   size_t _total_used_bytes;
 858   size_t _total_capacity_bytes;
 859   size_t _total_prev_live_bytes;
 860   size_t _total_next_live_bytes;
 861 
 862   // Accumulator for the remembered set size
 863   size_t _total_remset_bytes;
 864 
 865   // Accumulator for strong code roots memory size
 866   size_t _total_strong_code_roots_bytes;
 867 
 868   static double bytes_to_mb(size_t val) {
 869     return (double) val / (double) M;
 870   }
 871 
 872 public:
 873   // The header and footer are printed in the constructor and
 874   // destructor respectively.
 875   G1PrintRegionLivenessInfoClosure(const char* phase_name);
 876   virtual bool do_heap_region(HeapRegion* r);
 877   ~G1PrintRegionLivenessInfoClosure();
 878 };
 879 
 880 #endif // SHARE_GC_G1_G1CONCURRENTMARK_HPP