/*

 * Copyright (c) 2001, 2017, 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/g1ConcurrentMarkObjArrayProcessor.hpp"

#include "gc/g1/g1RegionToSpaceMapper.hpp"

#include "gc/g1/heapRegionSet.hpp"

#include "gc/shared/taskqueue.hpp"

class G1CollectedHeap;

class G1CMBitMap;

class G1CMTask;

class G1ConcurrentMark;

class ConcurrentGCTimer;

class G1OldTracer;

class G1SurvivorRegions;

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

  BitMapView _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 the overflow mark stack used by concurrent marking.

//

// Stores oops in a huge buffer in virtual memory that is always fully committed.

// Resizing may only happen during a STW pause when the stack is empty.

//

// Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark

// stack memory is split into evenly sized chunks of oops. Users can only

// add or remove entries on that basis.

// Chunks are filled in increasing address order. Not completely filled chunks

// have a NULL element as a terminating element.

//

// Every chunk has a header containing a single pointer element used for memory

// management. This wastes some space, but is negligible (< .1% with current sizing).

//

// Memory management is done using a mix of tracking a high water-mark indicating

// that all chunks at a lower address are valid chunks, and a singly linked free

// list connecting all empty chunks.

class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {

public:

private:

  };

  char _pad0[DEFAULT_CACHE_LINE_SIZE];

  volatile size_t _chunks_in_chunk_list;

  volatile size_t _hwm;          // High water mark within the reserved space.

  char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];

  // Allocate a new chunk from the reserved memory, using the high water mark. Returns

  // NULL if out of memory.

  // Atomically add the given chunk to the list.

  // Atomically remove and return a chunk from the given list. Returns NULL if the

  // list is empty.

  bool  _should_expand;

  // Resizes the mark stack to the given new capacity. Releases any previous

  // memory if successful.

  bool resize(size_t new_capacity);

 public:

  G1CMMarkStack();

  ~G1CMMarkStack();

  // Alignment and minimum capacity of this mark stack in number of oops.

  static size_t capacity_alignment();

  // Allocate and initialize the mark stack with the given number of oops.

  bool initialize(size_t initial_capacity, size_t max_capacity);

  // be terminated with a NULL.

  // Returns whether the buffer contents were successfully pushed to the global mark

  // stack.

  // Pops a chunk from this mark stack, copying them into the given buffer. This

  // element in the array is a NULL pointer.

  // Return whether the chunk list is empty. Racy due to unsynchronized access to

  // _chunk_list.

  bool is_empty() const { return _chunk_list == NULL; }

  size_t capacity() const  { return _chunk_capacity; }

  bool should_expand() const { return _should_expand; }

  void set_should_expand(bool value) { _should_expand = value; }

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

  void expand();

  // Return the approximate number of oops on this mark stack. Racy due to

  // unsynchronized access to _chunks_in_chunk_list.

  void set_empty();

  // Apply Fn to every oop on the mark stack. The mark stack must not

  // be modified while iterating.

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

};

// 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:

  const G1SurvivorRegions* _survivors;

  G1ConcurrentMark*        _cm;

  volatile bool            _scan_in_progress;

  volatile bool            _should_abort;

  volatile int             _claimed_survivor_index;

  void notify_scan_done();

public:

  G1CMRootRegions();

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

  void init(const G1SurvivorRegions* survivors, 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();

  // The number of root regions to scan.

  uint num_root_regions() const;

  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 G1VerifyLiveDataClosure;

  friend class G1CMRefProcTaskProxy;

  friend class G1CMRefProcTaskExecutor;

  friend class G1CMKeepAliveAndDrainClosure;

  friend class G1CMDrainMarkingStackClosure;

  friend class G1CMBitMapClosure;

  friend class G1CMConcurrentMarkingTask;

  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

  // Heap bounds

  HeapWord*               _heap_start;

  HeapWord*               _heap_end;

  // Root region tracking and claiming

  G1CMRootRegions         _root_regions;

  // For gray objects

  G1CMMarkStack           _global_mark_stack; // 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;

  ConcurrentGCTimer*      _gc_timer_cm;

  G1OldTracer*            _gc_tracer_cm;

  // 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();

  // 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.