/*

 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.

 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.

 *

 * This code is free software; you can redistribute it and/or modify it

 * under the terms of the GNU General Public License version 2 only, as

 * published by the Free Software Foundation.

 *

 * This code is distributed in the hope that it will be useful, but WITHOUT

 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or

 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License

 * version 2 for more details (a copy is included in the LICENSE file that

 * accompanied this code).

 *

 * You should have received a copy of the GNU General Public License version

 * 2 along with this work; if not, write to the Free Software Foundation,

 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.

 *

 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA

 * or visit www.oracle.com if you need additional information or have any

 * questions.

 *

 */

#ifndef SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP

#define SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP

#include "classfile/javaClasses.hpp"

#include "gc/g1/g1RegionToSpaceMapper.hpp"

#include "gc/g1/heapRegionSet.hpp"

#include "gc/shared/taskqueue.hpp"

class G1CollectedHeap;

class G1CMBitMap;

class G1CMTask;

class G1ConcurrentMark;

typedef GenericTaskQueue<oop, mtGC>              G1CMTaskQueue;

typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;

// Closure used by CM during concurrent reference discovery

// and reference processing (during remarking) to determine

// if a particular object is alive. It is primarily used

// to determine if referents of discovered reference objects

// are alive. An instance is also embedded into the

// reference processor as the _is_alive_non_header field

class G1CMIsAliveClosure: public BoolObjectClosure {

  G1CollectedHeap* _g1;

 public:

  G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }

  bool do_object_b(oop obj);

};

// A generic CM bit map.  This is essentially a wrapper around the BitMap

// class, with one bit per (1<<_shifter) HeapWords.

class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {

 protected:

  HeapWord* _bmStartWord;      // base address of range covered by map

  size_t    _bmWordSize;       // map size (in #HeapWords covered)

  const int _shifter;          // map to char or bit

  BitMap    _bm;               // the bit map itself

 public:

  // constructor

  G1CMBitMapRO(int shifter);

  // inquiries

  HeapWord* startWord()   const { return _bmStartWord; }

  // the following is one past the last word in space

  HeapWord* endWord()     const { return _bmStartWord + _bmWordSize; }

  // read marks

  bool isMarked(HeapWord* addr) const {

    assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),

           "outside underlying space?");

    return _bm.at(heapWordToOffset(addr));

  }

  // iteration

  inline bool iterate(BitMapClosure* cl, MemRegion mr);

  // Return the address corresponding to the next marked bit at or after

  // "addr", and before "limit", if "limit" is non-NULL.  If there is no

  // such bit, returns "limit" if that is non-NULL, or else "endWord()".

  HeapWord* getNextMarkedWordAddress(const HeapWord* addr,

                                     const HeapWord* limit = NULL) const;

  // conversion utilities

  HeapWord* offsetToHeapWord(size_t offset) const {

    return _bmStartWord + (offset << _shifter);

  }

  size_t heapWordToOffset(const HeapWord* addr) const {

    return pointer_delta(addr, _bmStartWord) >> _shifter;

  }

  // The argument addr should be the start address of a valid object

  inline HeapWord* nextObject(HeapWord* addr);

  void print_on_error(outputStream* st, const char* prefix) const;

  // debugging

  NOT_PRODUCT(bool covers(MemRegion rs) const;)

};

class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {

 private:

  G1CMBitMap* _bm;

 public:

  G1CMBitMapMappingChangedListener() : _bm(NULL) {}

  void set_bitmap(G1CMBitMap* bm) { _bm = bm; }

  virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);

};

class G1CMBitMap : public G1CMBitMapRO {

 private:

  G1CMBitMapMappingChangedListener _listener;

 public:

  static size_t compute_size(size_t heap_size);

  // Returns the amount of bytes on the heap between two marks in the bitmap.

  static size_t mark_distance();

  // Returns how many bytes (or bits) of the heap a single byte (or bit) of the

  // mark bitmap corresponds to. This is the same as the mark distance above.

  static size_t heap_map_factor() {

    return mark_distance();

  }

  G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }

  // Initializes the underlying BitMap to cover the given area.

  void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);

  // Write marks.

  inline void mark(HeapWord* addr);

  inline void clear(HeapWord* addr);

  inline bool parMark(HeapWord* addr);

  void clear_range(MemRegion mr);

};

// Represents a marking stack used by ConcurrentMarking in the G1 collector.

class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {

  VirtualSpace _virtual_space;   // Underlying backing store for actual stack

  G1ConcurrentMark* _cm;

  oop* _base;        // bottom of stack

  jint _index;       // one more than last occupied index

  jint _capacity;    // max #elements

  jint _saved_index; // value of _index saved at start of GC

  bool  _overflow;

  bool  _should_expand;

 public:

  G1CMMarkStack(G1ConcurrentMark* cm);

  ~G1CMMarkStack();

  bool allocate(size_t capacity);

  // Pushes the first "n" elements of "ptr_arr" on the stack.

  // Locking impl: concurrency is allowed only with

  // "par_push_arr" and/or "par_pop_arr" operations, which use the same

  // locking strategy.

  void par_push_arr(oop* ptr_arr, int n);

  // If returns false, the array was empty.  Otherwise, removes up to "max"

  // elements from the stack, and transfers them to "ptr_arr" in an

  // unspecified order.  The actual number transferred is given in "n" ("n

  // == 0" is deliberately redundant with the return value.)  Locking impl:

  // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"

  // operations, which use the same locking strategy.

  bool par_pop_arr(oop* ptr_arr, int max, int* n);

  bool isEmpty()    { return _index == 0; }

  int  maxElems()   { return _capacity; }

  bool overflow() { return _overflow; }

  void clear_overflow() { _overflow = false; }

  bool should_expand() const { return _should_expand; }

  void set_should_expand();

  // Expand the stack, typically in response to an overflow condition

  void expand();

  int  size() { return _index; }

  void setEmpty()   { _index = 0; clear_overflow(); }

  // Record the current index.

  void note_start_of_gc();

  // Make sure that we have not added any entries to the stack during GC.

  void note_end_of_gc();

  // Apply fn to each oop in the mark stack, up to the bound recorded

  // via one of the above "note" functions.  The mark stack must not

  // be modified while iterating.

  template<typename Fn> void iterate(Fn fn);

};

class YoungList;

// Root Regions are regions that are not empty at the beginning of a

// marking cycle and which we might collect during an evacuation pause

// while the cycle is active. Given that, during evacuation pauses, we

// do not copy objects that are explicitly marked, what we have to do

// for the root regions is to scan them and mark all objects reachable

// from them. According to the SATB assumptions, we only need to visit

// each object once during marking. So, as long as we finish this scan

// before the next evacuation pause, we can copy the objects from the

// root regions without having to mark them or do anything else to them.

//

// Currently, we only support root region scanning once (at the start

// of the marking cycle) and the root regions are all the survivor

// regions populated during the initial-mark pause.

class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {

private:

  YoungList*           _young_list;

  G1ConcurrentMark*    _cm;

  volatile bool        _scan_in_progress;

  volatile bool        _should_abort;

  HeapRegion* volatile _next_survivor;

  void notify_scan_done();

public:

  G1CMRootRegions();

  // We actually do most of the initialization in this method.

  void init(G1CollectedHeap* g1h, G1ConcurrentMark* cm);

  // Reset the claiming / scanning of the root regions.

  void prepare_for_scan();

  // Forces get_next() to return NULL so that the iteration aborts early.

  void abort() { _should_abort = true; }

  // Return true if the CM thread are actively scanning root regions,

  // false otherwise.

  bool scan_in_progress() { return _scan_in_progress; }

  // Claim the next root region to scan atomically, or return NULL if

  // all have been claimed.

  HeapRegion* claim_next();

  void cancel_scan();

  // Flag that we're done with root region scanning and notify anyone

  // who's waiting on it. If aborted is false, assume that all regions

  // have been claimed.

  void scan_finished();

  // If CM threads are still scanning root regions, wait until they

  // are done. Return true if we had to wait, false otherwise.

  bool wait_until_scan_finished();

};

class ConcurrentMarkThread;

class G1ConcurrentMark: public CHeapObj<mtGC> {

  friend class ConcurrentMarkThread;

  friend class G1ParNoteEndTask;

  friend class CalcLiveObjectsClosure;

  friend class G1CMRefProcTaskProxy;

  friend class G1CMRefProcTaskExecutor;

  friend class G1CMKeepAliveAndDrainClosure;

  friend class G1CMDrainMarkingStackClosure;

  friend class G1CMBitMapClosure;

  friend class G1CMConcurrentMarkingTask;

  friend class G1CMMarkStack;

  friend class G1CMRemarkTask;

  friend class G1CMTask;

protected:

  ConcurrentMarkThread* _cmThread;   // The thread doing the work

  G1CollectedHeap*      _g1h;        // The heap

  uint                  _parallel_marking_threads; // The number of marking

                                                   // threads we're using

  uint                  _max_parallel_marking_threads; // Max number of marking

                                                       // threads we'll ever use

  double                _sleep_factor; // How much we have to sleep, with

                                       // respect to the work we just did, to

                                       // meet the marking overhead goal

  double                _marking_task_overhead; // Marking target overhead for

                                                // a single task

  FreeRegionList        _cleanup_list;

  // Concurrent marking support structures

  G1CMBitMap              _markBitMap1;

  G1CMBitMap              _markBitMap2;

  G1CMBitMapRO*           _prevMarkBitMap; // Completed mark bitmap

  G1CMBitMap*             _nextMarkBitMap; // Under-construction mark bitmap

  BitMap                  _region_bm;

  BitMap                  _card_bm;

  // Heap bounds

  HeapWord*               _heap_start;

  HeapWord*               _heap_end;

  // Root region tracking and claiming

  G1CMRootRegions         _root_regions;

  // For gray objects

  G1CMMarkStack           _markStack; // Grey objects behind global finger

  HeapWord* volatile      _finger;  // The global finger, region aligned,

                                    // always points to the end of the

                                    // last claimed region

  // Marking tasks

  uint                    _max_worker_id;// Maximum worker id

  uint                    _active_tasks; // Task num currently active

  G1CMTask**              _tasks;        // Task queue array (max_worker_id len)

  G1CMTaskQueueSet*       _task_queues;  // Task queue set

  ParallelTaskTerminator  _terminator;   // For termination

  // Two sync barriers that are used to synchronize tasks when an

  // overflow occurs. The algorithm is the following. All tasks enter

  // the first one to ensure that they have all stopped manipulating

  // the global data structures. After they exit it, they re-initialize

  // their data structures and task 0 re-initializes the global data

  // structures. Then, they enter the second sync barrier. This

  // ensure, that no task starts doing work before all data

  // structures (local and global) have been re-initialized. When they

  // exit it, they are free to start working again.

  WorkGangBarrierSync     _first_overflow_barrier_sync;

  WorkGangBarrierSync     _second_overflow_barrier_sync;

  // This is set by any task, when an overflow on the global data

  // structures is detected

  volatile bool           _has_overflown;

  // True: marking is concurrent, false: we're in remark

  volatile bool           _concurrent;

  // Set at the end of a Full GC so that marking aborts

  volatile bool           _has_aborted;

  // Used when remark aborts due to an overflow to indicate that

  // another concurrent marking phase should start

  volatile bool           _restart_for_overflow;

  // This is true from the very start of concurrent marking until the

  // point when all the tasks complete their work. It is really used

  // to determine the points between the end of concurrent marking and

  // time of remark.

  volatile bool           _concurrent_marking_in_progress;

  // There would be a race between ConcurrentMarkThread and VMThread(ConcurrentMark::abort())

  // to call ConcurrentGCTimer::register_gc_concurrent_end().

  // And this variable is used to keep track of concurrent phase.

  volatile uint           _concurrent_phase_status;

  // Concurrent phase is not yet started.

  static const uint       ConcPhaseNotStarted = 0;

  // Concurrent phase is started.

  static const uint       ConcPhaseStarted = 1;

  // Caller thread of ConcurrentGCTimer::register_gc_concurrent_end() is ending concurrent phase.

  // So other thread should wait until the status to be changed to ConcPhaseNotStarted.

  static const uint       ConcPhaseStopping = 2;

  // All of these times are in ms

  NumberSeq _init_times;

  NumberSeq _remark_times;

  NumberSeq _remark_mark_times;

  NumberSeq _remark_weak_ref_times;

  NumberSeq _cleanup_times;

  double    _total_counting_time;

  double    _total_rs_scrub_time;

  double*   _accum_task_vtime;   // Accumulated task vtime

  WorkGang* _parallel_workers;

  void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);

  void weakRefsWork(bool clear_all_soft_refs);

  void swapMarkBitMaps();

  // It resets the global marking data structures, as well as the

  // task local ones; should be called during initial mark.

  void reset();

  // Resets all the marking data structures. Called when we have to restart

  // marking or when marking completes (via set_non_marking_state below).

  void reset_marking_state(bool clear_overflow = true);

  // We do this after we're done with marking so that the marking data

  // structures are initialized to a sensible and predictable state.

  void set_non_marking_state();

  // Called to indicate how many threads are currently active.

  void set_concurrency(uint active_tasks);

  // It should be called to indicate which phase we're in (concurrent

  // mark or remark) and how many threads are currently active.

  void set_concurrency_and_phase(uint active_tasks, bool concurrent);

  // Prints all gathered CM-related statistics

  void print_stats();

  bool cleanup_list_is_empty() {

    return _cleanup_list.is_empty();

  }

  // Accessor methods

  uint parallel_marking_threads() const     { return _parallel_marking_threads; }

  uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}

  double sleep_factor()                     { return _sleep_factor; }

  double marking_task_overhead()            { return _marking_task_overhead;}

  HeapWord*               finger()          { return _finger;   }

  bool                    concurrent()      { return _concurrent; }

  uint                    active_tasks()    { return _active_tasks; }

  ParallelTaskTerminator* terminator()      { return &_terminator; }

  // It claims the next available region to be scanned by a marking

  // task/thread. It might return NULL if the next region is empty or

  // we have run out of regions. In the latter case, out_of_regions()

  // determines whether we've really run out of regions or the task

  // should call claim_region() again. This might seem a bit

  // awkward. Originally, the code was written so that claim_region()

  // either successfully returned with a non-empty region or there

  // were no more regions to be claimed. The problem with this was

  // that, in certain circumstances, it iterated over large chunks of

  // the heap finding only empty regions and, while it was working, it

  // was preventing the calling task to call its regular clock

  // method. So, this way, each task will spend very little time in

  // claim_region() and is allowed to call the regular clock method

  // frequently.

  HeapRegion* claim_region(uint worker_id);

  // It determines whether we've run out of regions to scan. Note that

  // the finger can point past the heap end in case the heap was expanded

  // to satisfy an allocation without doing a GC. This is fine, because all

  // objects in those regions will be considered live anyway because of

  // SATB guarantees (i.e. their TAMS will be equal to bottom).

  bool        out_of_regions() { return _finger >= _heap_end; }

  // Returns the task with the given id

  G1CMTask* task(int id) {

    assert(0 <= id && id < (int) _active_tasks,

           "task id not within active bounds");

    return _tasks[id];

  }

  // Returns the task queue with the given id

  G1CMTaskQueue* task_queue(int id) {

    assert(0 <= id && id < (int) _active_tasks,

           "task queue id not within active bounds");

    return (G1CMTaskQueue*) _task_queues->queue(id);

  }

  // Returns the task queue set

  G1CMTaskQueueSet* task_queues()  { return _task_queues; }

  // Access / manipulation of the overflow flag which is set to

  // indicate that the global stack has overflown

  bool has_overflown()           { return _has_overflown; }

  void set_has_overflown()       { _has_overflown = true; }

  void clear_has_overflown()     { _has_overflown = false; }

  bool restart_for_overflow()    { return _restart_for_overflow; }

  // Methods to enter the two overflow sync barriers

  void enter_first_sync_barrier(uint worker_id);

  void enter_second_sync_barrier(uint worker_id);

  // Live Data Counting data structures...

  // These data structures are initialized at the start of

  // marking. They are written to while marking is active.

  // They are aggregated during remark; the aggregated values

  // are then used to populate the _region_bm, _card_bm, and

  // the total live bytes, which are then subsequently updated

  // during cleanup.

  // An array of bitmaps (one bit map per task). Each bitmap

  // is used to record the cards spanned by the live objects

  // marked by that task/worker.

  BitMap*  _count_card_bitmaps;

  // Used to record the number of marked live bytes

  // (for each region, by worker thread).

  size_t** _count_marked_bytes;

  // Card index of the bottom of the G1 heap. Used for biasing indices into

  // the card bitmaps.

  intptr_t _heap_bottom_card_num;

  // Set to true when initialization is complete

  bool _completed_initialization;

  // end_timer, true to end gc timer after ending concurrent phase.

  void register_concurrent_phase_end_common(bool end_timer);