1 /*
   2  * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP
  26 #define SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP
  27 
  28 #include "classfile/javaClasses.hpp"
  29 #include "gc/g1/g1ConcurrentMarkObjArrayProcessor.hpp"
  30 #include "gc/g1/g1RegionToSpaceMapper.hpp"
  31 #include "gc/g1/heapRegionSet.hpp"
  32 #include "gc/shared/taskqueue.hpp"
  33 
  34 class G1CollectedHeap;
  35 class G1CMBitMap;
  36 class G1CMTask;
  37 class G1ConcurrentMark;
  38 class ConcurrentGCTimer;
  39 class G1OldTracer;
  40 class G1SurvivorRegions;
  41 
  42 #ifdef _MSC_VER
  43 #pragma warning(push)
  44 // warning C4522: multiple assignment operators specified
  45 #pragma warning(disable:4522)
  46 #endif
  47 
  48 // This is a container class for either an oop or a continuation address for
  49 // mark stack entries. Both are pushed onto the mark stack.
  50 class G1TaskQueueEntry VALUE_OBJ_CLASS_SPEC {
  51 private:
  52   void* _holder;
  53 
  54   static const uintptr_t ArraySliceBit = 1;
  55 
  56   G1TaskQueueEntry(oop obj) : _holder(obj) {
  57     assert(_holder != NULL, "Not allowed to set NULL task queue element");
  58   }
  59   G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
  60 public:
  61   G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }
  62   G1TaskQueueEntry() : _holder(NULL) { }
  63 
  64   static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
  65   static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
  66 
  67   G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {
  68     _holder = t._holder;
  69     return *this;
  70   }
  71 
  72   volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
  73     _holder = t._holder;
  74     return *this;
  75   }
  76 
  77   oop obj() const {
  78     assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
  79     return (oop)_holder;
  80   }
  81 
  82   HeapWord* slice() const {
  83     assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
  84     return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
  85   }
  86 
  87   bool is_oop() const { return !is_array_slice(); }
  88   bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
  89   bool is_null() const { return _holder == NULL; }
  90 };
  91 
  92 #ifdef _MSC_VER
  93 #pragma warning(pop)
  94 #endif
  95 
  96 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
  97 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
  98 
  99 // Closure used by CM during concurrent reference discovery
 100 // and reference processing (during remarking) to determine
 101 // if a particular object is alive. It is primarily used
 102 // to determine if referents of discovered reference objects
 103 // are alive. An instance is also embedded into the
 104 // reference processor as the _is_alive_non_header field
 105 class G1CMIsAliveClosure: public BoolObjectClosure {
 106   G1CollectedHeap* _g1;
 107  public:
 108   G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
 109 
 110   bool do_object_b(oop obj);
 111 };
 112 
 113 // Closure for iteration over bitmaps
 114 class G1CMBitMapClosure VALUE_OBJ_CLASS_SPEC {
 115 private:
 116   G1ConcurrentMark* const _cm;
 117   G1CMTask* const _task; 
 118 public:
 119   G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm) : _task(task), _cm(cm) { }
 120 
 121   bool do_addr(HeapWord* const addr);
 122 };
 123 
 124 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
 125  private:
 126   G1CMBitMap* _bm;
 127  public:
 128   G1CMBitMapMappingChangedListener() : _bm(NULL) {}
 129 
 130   void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
 131 
 132   virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
 133 };
 134 
 135 // A generic mark bitmap for concurrent marking.  This is essentially a wrapper
 136 // around the BitMap class that is based on HeapWords, with one bit per (1 << _shifter) HeapWords.
 137 class G1CMBitMap VALUE_OBJ_CLASS_SPEC {
 138 private:
 139   MemRegion _covered;    // The heap area covered by this bitmap.
 140 
 141   const int _shifter;    // Shift amount from heap index to bit index in the bitmap.
 142 
 143   BitMapView _bm;        // The actual bitmap.
 144  
 145   G1CMBitMapMappingChangedListener _listener;
 146 
 147   inline void check_mark(HeapWord* addr) NOT_DEBUG_RETURN;
 148 
 149   // Convert from bit offset to address.
 150   HeapWord* offset_to_addr(size_t offset) const {
 151     return _covered.start() + (offset << _shifter);
 152   }
 153   // Convert from address to bit offset.
 154   size_t addr_to_offset(const HeapWord* addr) const {
 155     return pointer_delta(addr, _covered.start()) >> _shifter;
 156   }
 157 public:
 158   static size_t compute_size(size_t heap_size);
 159   // Returns the amount of bytes on the heap between two marks in the bitmap.
 160   static size_t mark_distance();
 161   // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
 162   // mark bitmap corresponds to. This is the same as the mark distance above.
 163   static size_t heap_map_factor() {
 164     return mark_distance();
 165   }
 166 
 167   G1CMBitMap() : _covered(), _bm(), _shifter(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
 168 
 169   // Initializes the underlying BitMap to cover the given area.
 170   void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
 171 
 172   // read marks
 173   bool is_marked(HeapWord* addr) const {
 174     assert(_covered.contains(addr),
 175            "Address " PTR_FORMAT " is outside underlying space from " PTR_FORMAT " to " PTR_FORMAT,
 176            p2i(addr), p2i(_covered.start()), p2i(_covered.end()));
 177     return _bm.at(addr_to_offset(addr));
 178   }
 179 
 180   // Apply the closure to the addresses that correspond to marked bits in the bitmap.
 181   inline bool iterate(G1CMBitMapClosure* cl, MemRegion mr);
 182 
 183   // Return the address corresponding to the next marked bit at or after
 184   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
 185   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
 186   inline HeapWord* get_next_marked_addr(const HeapWord* addr,
 187                                         const HeapWord* limit) const;
 188 
 189   // The argument addr should be the start address of a valid object
 190   inline HeapWord* addr_after_obj(HeapWord* addr);
 191 
 192   void print_on_error(outputStream* st, const char* prefix) const;
 193 
 194   // Write marks.
 195   inline void mark(HeapWord* addr);
 196   inline void clear(HeapWord* addr);
 197   inline bool par_mark(HeapWord* addr);
 198 
 199   void clear_range(MemRegion mr);
 200 };
 201 
 202 // Represents the overflow mark stack used by concurrent marking.
 203 //
 204 // Stores oops in a huge buffer in virtual memory that is always fully committed.
 205 // Resizing may only happen during a STW pause when the stack is empty.
 206 //
 207 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
 208 // stack memory is split into evenly sized chunks of oops. Users can only
 209 // add or remove entries on that basis.
 210 // Chunks are filled in increasing address order. Not completely filled chunks
 211 // have a NULL element as a terminating element.
 212 //
 213 // Every chunk has a header containing a single pointer element used for memory
 214 // management. This wastes some space, but is negligible (< .1% with current sizing).
 215 //
 216 // Memory management is done using a mix of tracking a high water-mark indicating
 217 // that all chunks at a lower address are valid chunks, and a singly linked free
 218 // list connecting all empty chunks.
 219 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
 220 public:
 221   // Number of TaskQueueEntries that can fit in a single chunk.
 222   static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
 223 private:
 224   struct TaskQueueEntryChunk {
 225     TaskQueueEntryChunk* next;
 226     G1TaskQueueEntry data[EntriesPerChunk];
 227   };
 228 
 229   size_t _max_chunk_capacity;    // Maximum number of TaskQueueEntryChunk elements on the stack.
 230 
 231   TaskQueueEntryChunk* _base;    // Bottom address of allocated memory area.
 232   size_t _chunk_capacity;        // Current maximum number of TaskQueueEntryChunk elements.
 233 
 234   char _pad0[DEFAULT_CACHE_LINE_SIZE];
 235   TaskQueueEntryChunk* volatile _free_list;  // Linked list of free chunks that can be allocated by users.
 236   char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
 237   TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
 238   volatile size_t _chunks_in_chunk_list;
 239   char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
 240 
 241   volatile size_t _hwm;          // High water mark within the reserved space.
 242   char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
 243 
 244   // Allocate a new chunk from the reserved memory, using the high water mark. Returns
 245   // NULL if out of memory.
 246   TaskQueueEntryChunk* allocate_new_chunk();
 247 
 248   // Atomically add the given chunk to the list.
 249   void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
 250   // Atomically remove and return a chunk from the given list. Returns NULL if the
 251   // list is empty.
 252   TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
 253 
 254   void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
 255   void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
 256 
 257   TaskQueueEntryChunk* remove_chunk_from_chunk_list();
 258   TaskQueueEntryChunk* remove_chunk_from_free_list();
 259 
 260   // Resizes the mark stack to the given new capacity. Releases any previous
 261   // memory if successful.
 262   bool resize(size_t new_capacity);
 263 
 264  public:
 265   G1CMMarkStack();
 266   ~G1CMMarkStack();
 267 
 268   // Alignment and minimum capacity of this mark stack in number of oops.
 269   static size_t capacity_alignment();
 270 
 271   // Allocate and initialize the mark stack with the given number of oops.
 272   bool initialize(size_t initial_capacity, size_t max_capacity);
 273 
 274   // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
 275   // stack. If less than EntriesPerChunk elements are to be pushed, the array must
 276   // be terminated with a NULL.
 277   // Returns whether the buffer contents were successfully pushed to the global mark
 278   // stack.
 279   bool par_push_chunk(G1TaskQueueEntry* buffer);
 280 
 281   // Pops a chunk from this mark stack, copying them into the given buffer. This
 282   // chunk may contain up to EntriesPerChunk elements. If there are less, the last
 283   // element in the array is a NULL pointer.
 284   bool par_pop_chunk(G1TaskQueueEntry* buffer);
 285 
 286   // Return whether the chunk list is empty. Racy due to unsynchronized access to
 287   // _chunk_list.
 288   bool is_empty() const { return _chunk_list == NULL; }
 289 
 290   size_t capacity() const  { return _chunk_capacity; }
 291 
 292   // Expand the stack, typically in response to an overflow condition
 293   void expand();
 294 
 295   // Return the approximate number of oops on this mark stack. Racy due to
 296   // unsynchronized access to _chunks_in_chunk_list.
 297   size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
 298 
 299   void set_empty();
 300 
 301   // Apply Fn to every oop on the mark stack. The mark stack must not
 302   // be modified while iterating.
 303   template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
 304 };
 305 
 306 // Root Regions are regions that are not empty at the beginning of a
 307 // marking cycle and which we might collect during an evacuation pause
 308 // while the cycle is active. Given that, during evacuation pauses, we
 309 // do not copy objects that are explicitly marked, what we have to do
 310 // for the root regions is to scan them and mark all objects reachable
 311 // from them. According to the SATB assumptions, we only need to visit
 312 // each object once during marking. So, as long as we finish this scan
 313 // before the next evacuation pause, we can copy the objects from the
 314 // root regions without having to mark them or do anything else to them.
 315 //
 316 // Currently, we only support root region scanning once (at the start
 317 // of the marking cycle) and the root regions are all the survivor
 318 // regions populated during the initial-mark pause.
 319 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
 320 private:
 321   const G1SurvivorRegions* _survivors;
 322   G1ConcurrentMark*        _cm;
 323 
 324   volatile bool            _scan_in_progress;
 325   volatile bool            _should_abort;
 326   volatile int             _claimed_survivor_index;
 327 
 328   void notify_scan_done();
 329 
 330 public:
 331   G1CMRootRegions();
 332   // We actually do most of the initialization in this method.
 333   void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
 334 
 335   // Reset the claiming / scanning of the root regions.
 336   void prepare_for_scan();
 337 
 338   // Forces get_next() to return NULL so that the iteration aborts early.
 339   void abort() { _should_abort = true; }
 340 
 341   // Return true if the CM thread are actively scanning root regions,
 342   // false otherwise.
 343   bool scan_in_progress() { return _scan_in_progress; }
 344 
 345   // Claim the next root region to scan atomically, or return NULL if
 346   // all have been claimed.
 347   HeapRegion* claim_next();
 348 
 349   // The number of root regions to scan.
 350   uint num_root_regions() const;
 351 
 352   void cancel_scan();
 353 
 354   // Flag that we're done with root region scanning and notify anyone
 355   // who's waiting on it. If aborted is false, assume that all regions
 356   // have been claimed.
 357   void scan_finished();
 358 
 359   // If CM threads are still scanning root regions, wait until they
 360   // are done. Return true if we had to wait, false otherwise.
 361   bool wait_until_scan_finished();
 362 };
 363 
 364 class ConcurrentMarkThread;
 365 
 366 class G1ConcurrentMark: public CHeapObj<mtGC> {
 367   friend class ConcurrentMarkThread;
 368   friend class G1ParNoteEndTask;
 369   friend class G1VerifyLiveDataClosure;
 370   friend class G1CMRefProcTaskProxy;
 371   friend class G1CMRefProcTaskExecutor;
 372   friend class G1CMKeepAliveAndDrainClosure;
 373   friend class G1CMDrainMarkingStackClosure;
 374   friend class G1CMBitMapClosure;
 375   friend class G1CMConcurrentMarkingTask;
 376   friend class G1CMRemarkTask;
 377   friend class G1CMTask;
 378 
 379 protected:
 380   ConcurrentMarkThread* _cmThread;   // The thread doing the work
 381   G1CollectedHeap*      _g1h;        // The heap
 382   uint                  _parallel_marking_threads; // The number of marking
 383                                                    // threads we're using
 384   uint                  _max_parallel_marking_threads; // Max number of marking
 385                                                        // threads we'll ever use
 386   double                _sleep_factor; // How much we have to sleep, with
 387                                        // respect to the work we just did, to
 388                                        // meet the marking overhead goal
 389   double                _marking_task_overhead; // Marking target overhead for
 390                                                 // a single task
 391 
 392   FreeRegionList        _cleanup_list;
 393 
 394   // Concurrent marking support structures
 395   G1CMBitMap              _markBitMap1;
 396   G1CMBitMap              _markBitMap2;
 397   G1CMBitMap*             _prevMarkBitMap; // Completed mark bitmap
 398   G1CMBitMap*             _nextMarkBitMap; // Under-construction mark bitmap
 399 
 400   // Heap bounds
 401   HeapWord*               _heap_start;
 402   HeapWord*               _heap_end;
 403 
 404   // Root region tracking and claiming
 405   G1CMRootRegions         _root_regions;
 406 
 407   // For gray objects
 408   G1CMMarkStack           _global_mark_stack; // Grey objects behind global finger
 409   HeapWord* volatile      _finger;  // The global finger, region aligned,
 410                                     // always points to the end of the
 411                                     // last claimed region
 412 
 413   // Marking tasks
 414   uint                    _max_worker_id;// Maximum worker id
 415   uint                    _active_tasks; // Task num currently active
 416   G1CMTask**              _tasks;        // Task queue array (max_worker_id len)
 417   G1CMTaskQueueSet*       _task_queues;  // Task queue set
 418   ParallelTaskTerminator  _terminator;   // For termination
 419 
 420   // Two sync barriers that are used to synchronize tasks when an
 421   // overflow occurs. The algorithm is the following. All tasks enter
 422   // the first one to ensure that they have all stopped manipulating
 423   // the global data structures. After they exit it, they re-initialize
 424   // their data structures and task 0 re-initializes the global data
 425   // structures. Then, they enter the second sync barrier. This
 426   // ensure, that no task starts doing work before all data
 427   // structures (local and global) have been re-initialized. When they
 428   // exit it, they are free to start working again.
 429   WorkGangBarrierSync     _first_overflow_barrier_sync;
 430   WorkGangBarrierSync     _second_overflow_barrier_sync;
 431 
 432   // This is set by any task, when an overflow on the global data
 433   // structures is detected
 434   volatile bool           _has_overflown;
 435   // True: marking is concurrent, false: we're in remark
 436   volatile bool           _concurrent;
 437   // Set at the end of a Full GC so that marking aborts
 438   volatile bool           _has_aborted;
 439 
 440   // Used when remark aborts due to an overflow to indicate that
 441   // another concurrent marking phase should start
 442   volatile bool           _restart_for_overflow;
 443 
 444   // This is true from the very start of concurrent marking until the
 445   // point when all the tasks complete their work. It is really used
 446   // to determine the points between the end of concurrent marking and
 447   // time of remark.
 448   volatile bool           _concurrent_marking_in_progress;
 449 
 450   ConcurrentGCTimer*      _gc_timer_cm;
 451 
 452   G1OldTracer*            _gc_tracer_cm;
 453 
 454   // All of these times are in ms
 455   NumberSeq _init_times;
 456   NumberSeq _remark_times;
 457   NumberSeq _remark_mark_times;
 458   NumberSeq _remark_weak_ref_times;
 459   NumberSeq _cleanup_times;
 460   double    _total_counting_time;
 461   double    _total_rs_scrub_time;
 462 
 463   double*   _accum_task_vtime;   // Accumulated task vtime
 464 
 465   WorkGang* _parallel_workers;
 466 
 467   void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
 468   void weakRefsWork(bool clear_all_soft_refs);
 469 
 470   void swapMarkBitMaps();
 471 
 472   // It resets the global marking data structures, as well as the
 473   // task local ones; should be called during initial mark.
 474   void reset();
 475 
 476   // Resets all the marking data structures. Called when we have to restart
 477   // marking or when marking completes (via set_non_marking_state below).
 478   void reset_marking_state();
 479 
 480   // We do this after we're done with marking so that the marking data
 481   // structures are initialized to a sensible and predictable state.
 482   void set_non_marking_state();
 483 
 484   // Called to indicate how many threads are currently active.
 485   void set_concurrency(uint active_tasks);
 486 
 487   // It should be called to indicate which phase we're in (concurrent
 488   // mark or remark) and how many threads are currently active.
 489   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 490 
 491   // Prints all gathered CM-related statistics
 492   void print_stats();
 493 
 494   bool cleanup_list_is_empty() {
 495     return _cleanup_list.is_empty();
 496   }
 497 
 498   // Accessor methods
 499   uint parallel_marking_threads() const     { return _parallel_marking_threads; }
 500   uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
 501   double sleep_factor()                     { return _sleep_factor; }
 502   double marking_task_overhead()            { return _marking_task_overhead;}
 503 
 504   HeapWord*               finger()          { return _finger;   }
 505   bool                    concurrent()      { return _concurrent; }
 506   uint                    active_tasks()    { return _active_tasks; }
 507   ParallelTaskTerminator* terminator()      { return &_terminator; }
 508 
 509   // It claims the next available region to be scanned by a marking
 510   // task/thread. It might return NULL if the next region is empty or
 511   // we have run out of regions. In the latter case, out_of_regions()
 512   // determines whether we've really run out of regions or the task
 513   // should call claim_region() again. This might seem a bit
 514   // awkward. Originally, the code was written so that claim_region()
 515   // either successfully returned with a non-empty region or there
 516   // were no more regions to be claimed. The problem with this was
 517   // that, in certain circumstances, it iterated over large chunks of
 518   // the heap finding only empty regions and, while it was working, it
 519   // was preventing the calling task to call its regular clock
 520   // method. So, this way, each task will spend very little time in
 521   // claim_region() and is allowed to call the regular clock method
 522   // frequently.
 523   HeapRegion* claim_region(uint worker_id);
 524 
 525   // It determines whether we've run out of regions to scan. Note that
 526   // the finger can point past the heap end in case the heap was expanded
 527   // to satisfy an allocation without doing a GC. This is fine, because all
 528   // objects in those regions will be considered live anyway because of
 529   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 530   bool        out_of_regions() { return _finger >= _heap_end; }
 531 
 532   // Returns the task with the given id
 533   G1CMTask* task(int id) {
 534     assert(0 <= id && id < (int) _active_tasks,
 535            "task id not within active bounds");
 536     return _tasks[id];
 537   }
 538 
 539   // Returns the task queue with the given id
 540   G1CMTaskQueue* task_queue(int id) {
 541     assert(0 <= id && id < (int) _active_tasks,
 542            "task queue id not within active bounds");
 543     return (G1CMTaskQueue*) _task_queues->queue(id);
 544   }
 545 
 546   // Returns the task queue set
 547   G1CMTaskQueueSet* task_queues()  { return _task_queues; }
 548 
 549   // Access / manipulation of the overflow flag which is set to
 550   // indicate that the global stack has overflown
 551   bool has_overflown()           { return _has_overflown; }
 552   void set_has_overflown()       { _has_overflown = true; }
 553   void clear_has_overflown()     { _has_overflown = false; }
 554   bool restart_for_overflow()    { return _restart_for_overflow; }
 555 
 556   // Methods to enter the two overflow sync barriers
 557   void enter_first_sync_barrier(uint worker_id);
 558   void enter_second_sync_barrier(uint worker_id);
 559 
 560   // Card index of the bottom of the G1 heap. Used for biasing indices into
 561   // the card bitmaps.
 562   intptr_t _heap_bottom_card_num;
 563 
 564   // Set to true when initialization is complete
 565   bool _completed_initialization;
 566 
 567   // end_timer, true to end gc timer after ending concurrent phase.
 568   void register_concurrent_phase_end_common(bool end_timer);
 569 
 570   // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
 571   // true, periodically insert checks to see if this method should exit prematurely.
 572   void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
 573 public:
 574   // Manipulation of the global mark stack.
 575   // The push and pop operations are used by tasks for transfers
 576   // between task-local queues and the global mark stack.
 577   bool mark_stack_push(G1TaskQueueEntry* arr) {
 578     if (!_global_mark_stack.par_push_chunk(arr)) {
 579       set_has_overflown();
 580       return false;
 581     }
 582     return true;
 583   }
 584   bool mark_stack_pop(G1TaskQueueEntry* arr) {
 585     return _global_mark_stack.par_pop_chunk(arr);
 586   }
 587   size_t mark_stack_size()                { return _global_mark_stack.size(); }
 588   size_t partial_mark_stack_size_target() { return _global_mark_stack.capacity()/3; }
 589   bool mark_stack_empty()                 { return _global_mark_stack.is_empty(); }
 590 
 591   G1CMRootRegions* root_regions() { return &_root_regions; }
 592 
 593   bool concurrent_marking_in_progress() {
 594     return _concurrent_marking_in_progress;
 595   }
 596   void set_concurrent_marking_in_progress() {
 597     _concurrent_marking_in_progress = true;
 598   }
 599   void clear_concurrent_marking_in_progress() {
 600     _concurrent_marking_in_progress = false;
 601   }
 602 
 603   void concurrent_cycle_start();
 604   void concurrent_cycle_end();
 605 
 606   void update_accum_task_vtime(int i, double vtime) {
 607     _accum_task_vtime[i] += vtime;
 608   }
 609 
 610   double all_task_accum_vtime() {
 611     double ret = 0.0;
 612     for (uint i = 0; i < _max_worker_id; ++i)
 613       ret += _accum_task_vtime[i];
 614     return ret;
 615   }
 616 
 617   // Attempts to steal an object from the task queues of other tasks
 618   bool try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry);
 619 
 620   G1ConcurrentMark(G1CollectedHeap* g1h,
 621                    G1RegionToSpaceMapper* prev_bitmap_storage,
 622                    G1RegionToSpaceMapper* next_bitmap_storage);
 623   ~G1ConcurrentMark();
 624 
 625   ConcurrentMarkThread* cmThread() { return _cmThread; }
 626 
 627   const G1CMBitMap* const prevMarkBitMap() const { return _prevMarkBitMap; }
 628   G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
 629 
 630   // Returns the number of GC threads to be used in a concurrent
 631   // phase based on the number of GC threads being used in a STW
 632   // phase.
 633   uint scale_parallel_threads(uint n_par_threads);
 634 
 635   // Calculates the number of GC threads to be used in a concurrent phase.
 636   uint calc_parallel_marking_threads();
 637 
 638   // The following three are interaction between CM and
 639   // G1CollectedHeap
 640 
 641   // This notifies CM that a root during initial-mark needs to be
 642   // grayed. It is MT-safe. hr is the region that
 643   // contains the object and it's passed optionally from callers who
 644   // might already have it (no point in recalculating it).
 645   inline void grayRoot(oop obj,
 646                        HeapRegion* hr = NULL);
 647 
 648   // Prepare internal data structures for the next mark cycle. This includes clearing
 649   // the next mark bitmap and some internal data structures. This method is intended
 650   // to be called concurrently to the mutator. It will yield to safepoint requests.
 651   void cleanup_for_next_mark();
 652 
 653   // Clear the previous marking bitmap during safepoint.
 654   void clear_prev_bitmap(WorkGang* workers);
 655 
 656   // Return whether the next mark bitmap has no marks set. To be used for assertions
 657   // only. Will not yield to pause requests.
 658   bool nextMarkBitmapIsClear();
 659 
 660   // These two do the work that needs to be done before and after the
 661   // initial root checkpoint. Since this checkpoint can be done at two
 662   // different points (i.e. an explicit pause or piggy-backed on a
 663   // young collection), then it's nice to be able to easily share the
 664   // pre/post code. It might be the case that we can put everything in
 665   // the post method. TP
 666   void checkpointRootsInitialPre();
 667   void checkpointRootsInitialPost();
 668 
 669   // Scan all the root regions and mark everything reachable from
 670   // them.
 671   void scan_root_regions();
 672 
 673   // Scan a single root region and mark everything reachable from it.
 674   void scanRootRegion(HeapRegion* hr);
 675 
 676   // Do concurrent phase of marking, to a tentative transitive closure.
 677   void mark_from_roots();
 678 
 679   void checkpointRootsFinal(bool clear_all_soft_refs);
 680   void checkpointRootsFinalWork();
 681   void cleanup();
 682   void complete_cleanup();
 683 
 684   // Mark in the previous bitmap.  NB: this is usually read-only, so use
 685   // this carefully!
 686   inline void markPrev(oop p);
 687 
 688   // Clears marks for all objects in the given range, for the prev or
 689   // next bitmaps.  NB: the previous bitmap is usually
 690   // read-only, so use this carefully!
 691   void clearRangePrevBitmap(MemRegion mr);
 692 
 693   // Verify that there are no CSet oops on the stacks (taskqueues /
 694   // global mark stack) and fingers (global / per-task).
 695   // If marking is not in progress, it's a no-op.
 696   void verify_no_cset_oops() PRODUCT_RETURN;
 697 
 698   inline bool isPrevMarked(oop p) const;
 699 
 700   inline bool do_yield_check();
 701 
 702   // Abandon current marking iteration due to a Full GC.
 703   void abort();
 704 
 705   bool has_aborted()      { return _has_aborted; }
 706 
 707   void print_summary_info();
 708 
 709   void print_worker_threads_on(outputStream* st) const;
 710   void threads_do(ThreadClosure* tc) const;
 711 
 712   void print_on_error(outputStream* st) const;
 713 
 714   // Attempts to mark the given object on the next mark bitmap.
 715   inline bool par_mark(oop obj);
 716 
 717   // Returns true if initialization was successfully completed.
 718   bool completed_initialization() const {
 719     return _completed_initialization;
 720   }
 721 
 722   ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
 723   G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
 724 
 725 private:
 726   // Clear (Reset) all liveness count data.
 727   void clear_live_data(WorkGang* workers);
 728 
 729 #ifdef ASSERT
 730   // Verify all of the above data structures that they are in initial state.
 731   void verify_live_data_clear();
 732 #endif
 733 
 734   // Aggregates the per-card liveness data based on the current marking. Also sets
 735   // the amount of marked bytes for each region.
 736   void create_live_data();
 737 
 738   void finalize_live_data();
 739 
 740   void verify_live_data();
 741 };
 742 
 743 // A class representing a marking task.
 744 class G1CMTask : public TerminatorTerminator {
 745 private:
 746   enum PrivateConstants {
 747     // The regular clock call is called once the scanned words reaches
 748     // this limit
 749     words_scanned_period          = 12*1024,
 750     // The regular clock call is called once the number of visited
 751     // references reaches this limit
 752     refs_reached_period           = 1024,
 753     // Initial value for the hash seed, used in the work stealing code
 754     init_hash_seed                = 17
 755   };
 756 
 757   G1CMObjArrayProcessor       _objArray_processor;
 758 
 759   uint                        _worker_id;
 760   G1CollectedHeap*            _g1h;
 761   G1ConcurrentMark*           _cm;
 762   G1CMBitMap*                 _nextMarkBitMap;
 763   // the task queue of this task
 764   G1CMTaskQueue*              _task_queue;
 765 private:
 766   // the task queue set---needed for stealing
 767   G1CMTaskQueueSet*           _task_queues;
 768   // indicates whether the task has been claimed---this is only  for
 769   // debugging purposes
 770   bool                        _claimed;
 771 
 772   // number of calls to this task
 773   int                         _calls;
 774 
 775   // when the virtual timer reaches this time, the marking step should
 776   // exit
 777   double                      _time_target_ms;
 778   // the start time of the current marking step
 779   double                      _start_time_ms;
 780 
 781   // the oop closure used for iterations over oops
 782   G1CMOopClosure*             _cm_oop_closure;
 783 
 784   // the region this task is scanning, NULL if we're not scanning any
 785   HeapRegion*                 _curr_region;
 786   // the local finger of this task, NULL if we're not scanning a region
 787   HeapWord*                   _finger;
 788   // limit of the region this task is scanning, NULL if we're not scanning one
 789   HeapWord*                   _region_limit;
 790 
 791   // the number of words this task has scanned
 792   size_t                      _words_scanned;
 793   // When _words_scanned reaches this limit, the regular clock is
 794   // called. Notice that this might be decreased under certain
 795   // circumstances (i.e. when we believe that we did an expensive
 796   // operation).
 797   size_t                      _words_scanned_limit;
 798   // the initial value of _words_scanned_limit (i.e. what it was
 799   // before it was decreased).
 800   size_t                      _real_words_scanned_limit;
 801 
 802   // the number of references this task has visited
 803   size_t                      _refs_reached;
 804   // When _refs_reached reaches this limit, the regular clock is
 805   // called. Notice this this might be decreased under certain
 806   // circumstances (i.e. when we believe that we did an expensive
 807   // operation).
 808   size_t                      _refs_reached_limit;
 809   // the initial value of _refs_reached_limit (i.e. what it was before
 810   // it was decreased).
 811   size_t                      _real_refs_reached_limit;
 812 
 813   // used by the work stealing stuff
 814   int                         _hash_seed;
 815   // if this is true, then the task has aborted for some reason
 816   bool                        _has_aborted;
 817   // set when the task aborts because it has met its time quota
 818   bool                        _has_timed_out;
 819   // true when we're draining SATB buffers; this avoids the task
 820   // aborting due to SATB buffers being available (as we're already
 821   // dealing with them)
 822   bool                        _draining_satb_buffers;
 823 
 824   // number sequence of past step times
 825   NumberSeq                   _step_times_ms;
 826   // elapsed time of this task
 827   double                      _elapsed_time_ms;
 828   // termination time of this task
 829   double                      _termination_time_ms;
 830   // when this task got into the termination protocol
 831   double                      _termination_start_time_ms;
 832 
 833   // true when the task is during a concurrent phase, false when it is
 834   // in the remark phase (so, in the latter case, we do not have to
 835   // check all the things that we have to check during the concurrent
 836   // phase, i.e. SATB buffer availability...)
 837   bool                        _concurrent;
 838 
 839   TruncatedSeq                _marking_step_diffs_ms;
 840 
 841   // it updates the local fields after this task has claimed
 842   // a new region to scan
 843   void setup_for_region(HeapRegion* hr);
 844   // it brings up-to-date the limit of the region
 845   void update_region_limit();
 846 
 847   // called when either the words scanned or the refs visited limit
 848   // has been reached
 849   void reached_limit();
 850   // recalculates the words scanned and refs visited limits
 851   void recalculate_limits();
 852   // decreases the words scanned and refs visited limits when we reach
 853   // an expensive operation
 854   void decrease_limits();
 855   // it checks whether the words scanned or refs visited reached their
 856   // respective limit and calls reached_limit() if they have
 857   void check_limits() {
 858     if (_words_scanned >= _words_scanned_limit ||
 859         _refs_reached >= _refs_reached_limit) {
 860       reached_limit();
 861     }
 862   }
 863   // this is supposed to be called regularly during a marking step as
 864   // it checks a bunch of conditions that might cause the marking step
 865   // to abort
 866   void regular_clock_call();
 867   bool concurrent() { return _concurrent; }
 868 
 869   // Test whether obj might have already been passed over by the
 870   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 871   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 872 
 873   template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
 874 public:
 875   // Apply the closure on the given area of the objArray. Return the number of words
 876   // scanned.
 877   inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
 878   // It resets the task; it should be called right at the beginning of
 879   // a marking phase.
 880   void reset(G1CMBitMap* _nextMarkBitMap);
 881   // it clears all the fields that correspond to a claimed region.
 882   void clear_region_fields();
 883 
 884   void set_concurrent(bool concurrent) { _concurrent = concurrent; }
 885 
 886   // The main method of this class which performs a marking step
 887   // trying not to exceed the given duration. However, it might exit
 888   // prematurely, according to some conditions (i.e. SATB buffers are
 889   // available for processing).
 890   void do_marking_step(double target_ms,
 891                        bool do_termination,
 892                        bool is_serial);
 893 
 894   // These two calls start and stop the timer
 895   void record_start_time() {
 896     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 897   }
 898   void record_end_time() {
 899     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 900   }
 901 
 902   // returns the worker ID associated with this task.
 903   uint worker_id() { return _worker_id; }
 904 
 905   // From TerminatorTerminator. It determines whether this task should
 906   // exit the termination protocol after it's entered it.
 907   virtual bool should_exit_termination();
 908 
 909   // Resets the local region fields after a task has finished scanning a
 910   // region; or when they have become stale as a result of the region
 911   // being evacuated.
 912   void giveup_current_region();
 913 
 914   HeapWord* finger()            { return _finger; }
 915 
 916   bool has_aborted()            { return _has_aborted; }
 917   void set_has_aborted()        { _has_aborted = true; }
 918   void clear_has_aborted()      { _has_aborted = false; }
 919   bool has_timed_out()          { return _has_timed_out; }
 920   bool claimed()                { return _claimed; }
 921 
 922   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
 923 
 924   // Increment the number of references this task has visited.
 925   void increment_refs_reached() { ++_refs_reached; }
 926 
 927   // Grey the object by marking it.  If not already marked, push it on
 928   // the local queue if below the finger.
 929   // obj is below its region's NTAMS.
 930   inline void make_reference_grey(oop obj);
 931 
 932   // Grey the object (by calling make_grey_reference) if required,
 933   // e.g. obj is below its containing region's NTAMS.
 934   // Precondition: obj is a valid heap object.
 935   inline void deal_with_reference(oop obj);
 936 
 937   // It scans an object and visits its children.
 938   inline void scan_task_entry(G1TaskQueueEntry task_entry);
 939 
 940   // It pushes an object on the local queue.
 941   inline void push(G1TaskQueueEntry task_entry);
 942 
 943   // Move entries to the global stack.
 944   void move_entries_to_global_stack();
 945   // Move entries from the global stack, return true if we were successful to do so.
 946   bool get_entries_from_global_stack();
 947 
 948   // It pops and scans objects from the local queue. If partially is
 949   // true, then it stops when the queue size is of a given limit. If
 950   // partially is false, then it stops when the queue is empty.
 951   void drain_local_queue(bool partially);
 952   // It moves entries from the global stack to the local queue and
 953   // drains the local queue. If partially is true, then it stops when
 954   // both the global stack and the local queue reach a given size. If
 955   // partially if false, it tries to empty them totally.
 956   void drain_global_stack(bool partially);
 957   // It keeps picking SATB buffers and processing them until no SATB
 958   // buffers are available.
 959   void drain_satb_buffers();
 960 
 961   // moves the local finger to a new location
 962   inline void move_finger_to(HeapWord* new_finger) {
 963     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
 964     _finger = new_finger;
 965   }
 966 
 967   G1CMTask(uint worker_id,
 968            G1ConcurrentMark *cm,
 969            G1CMTaskQueue* task_queue,
 970            G1CMTaskQueueSet* task_queues);
 971 
 972   // it prints statistics associated with this task
 973   void print_stats();
 974 };
 975 
 976 // Class that's used to to print out per-region liveness
 977 // information. It's currently used at the end of marking and also
 978 // after we sort the old regions at the end of the cleanup operation.
 979 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
 980 private:
 981   // Accumulators for these values.
 982   size_t _total_used_bytes;
 983   size_t _total_capacity_bytes;
 984   size_t _total_prev_live_bytes;
 985   size_t _total_next_live_bytes;
 986 
 987   // Accumulator for the remembered set size
 988   size_t _total_remset_bytes;
 989 
 990   // Accumulator for strong code roots memory size
 991   size_t _total_strong_code_roots_bytes;
 992 
 993   static double perc(size_t val, size_t total) {
 994     if (total == 0) {
 995       return 0.0;
 996     } else {
 997       return 100.0 * ((double) val / (double) total);
 998     }
 999   }
1000 
1001   static double bytes_to_mb(size_t val) {
1002     return (double) val / (double) M;
1003   }
1004 
1005 public:
1006   // The header and footer are printed in the constructor and
1007   // destructor respectively.
1008   G1PrintRegionLivenessInfoClosure(const char* phase_name);
1009   virtual bool doHeapRegion(HeapRegion* r);
1010   ~G1PrintRegionLivenessInfoClosure();
1011 };
1012 
1013 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP