/*

 * Copyright 2001-2009 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,

 * CA 95054 USA or visit www.sun.com if you need additional information or

 * have any questions.

 *

 */

class G1CollectedHeap;

class CMTask;

typedef GenericTaskQueue<oop> CMTaskQueue;

typedef GenericTaskQueueSet<oop> CMTaskQueueSet;

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

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

class CMBitMapRO 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

  VirtualSpace _virtual_space; // underlying the bit map

  BitMap    _bm;               // the bit map itself

 public:

  // constructor

  CMBitMapRO(ReservedSpace rs, int shifter);

  enum { do_yield = true };

  // inquiries

  HeapWord* startWord()   const { return _bmStartWord; }

  size_t    sizeInWords() const { return _bmWordSize;  }

  // 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

  bool iterate(BitMapClosure* cl) { return _bm.iterate(cl); }

  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(HeapWord* addr,

                                     HeapWord* limit = NULL) const;

  // Return the address corresponding to the next unmarked 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* getNextUnmarkedWordAddress(HeapWord* addr,

                                       HeapWord* limit = NULL) const;

  // conversion utilities

  // XXX Fix these so that offsets are size_t's...

  HeapWord* offsetToHeapWord(size_t offset) const {

    return _bmStartWord + (offset << _shifter);

  }

  size_t heapWordToOffset(HeapWord* addr) const {

    return pointer_delta(addr, _bmStartWord) >> _shifter;

  }

  int heapWordDiffToOffsetDiff(size_t diff) const;

  HeapWord* nextWord(HeapWord* addr) {

    return offsetToHeapWord(heapWordToOffset(addr) + 1);

  }

  void mostly_disjoint_range_union(BitMap*   from_bitmap,

                                   size_t    from_start_index,

                                   HeapWord* to_start_word,

                                   size_t    word_num);

  // debugging

  NOT_PRODUCT(bool covers(ReservedSpace rs) const;)

};

class CMBitMap : public CMBitMapRO {

 public:

  // constructor

  CMBitMap(ReservedSpace rs, int shifter) :

    CMBitMapRO(rs, shifter) {}

  // write marks

  void mark(HeapWord* addr) {

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

           "outside underlying space?");

    _bm.at_put(heapWordToOffset(addr), true);

  }

  void clear(HeapWord* addr) {

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

           "outside underlying space?");

    _bm.at_put(heapWordToOffset(addr), false);

  }

  bool parMark(HeapWord* addr) {

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

           "outside underlying space?");

    return _bm.par_at_put(heapWordToOffset(addr), true);

  }

  bool parClear(HeapWord* addr) {

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

           "outside underlying space?");

    return _bm.par_at_put(heapWordToOffset(addr), false);

  }

  void markRange(MemRegion mr);

  void clearAll();

  void clearRange(MemRegion mr);

  // Starting at the bit corresponding to "addr" (inclusive), find the next

  // "1" bit, if any.  This bit starts some run of consecutive "1"'s; find

  // the end of this run (stopping at "end_addr").  Return the MemRegion

  // covering from the start of the region corresponding to the first bit

  // of the run to the end of the region corresponding to the last bit of

  // the run.  If there is no "1" bit at or after "addr", return an empty

  // MemRegion.

  MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);

};

// Represents a marking stack used by the CM collector.

// Ideally this should be GrowableArray<> just like MSC's marking stack(s).

class CMMarkStack VALUE_OBJ_CLASS_SPEC {

  ConcurrentMark* _cm;

  oop*   _base;      // bottom of stack

  jint   _index;     // one more than last occupied index

  jint   _capacity;  // max #elements

  jint   _oops_do_bound;  // Number of elements to include in next iteration.

  NOT_PRODUCT(jint _max_depth;)  // max depth plumbed during run

  bool   _overflow;

  DEBUG_ONLY(bool _drain_in_progress;)

  DEBUG_ONLY(bool _drain_in_progress_yields;)

 public:

  CMMarkStack(ConcurrentMark* cm);

  ~CMMarkStack();

  void allocate(size_t size);

  oop pop() {

    if (!isEmpty()) {

      return _base[--_index] ;

    }

    return NULL;

  }

  // If overflow happens, don't do the push, and record the overflow.

  // *Requires* that "ptr" is already marked.

  void push(oop ptr) {

    if (isFull()) {

      // Record overflow.

      _overflow = true;

      return;

    } else {

      _base[_index++] = ptr;

      NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));

    }

  }

  // Non-block impl.  Note: concurrency is allowed only with other

  // "par_push" operations, not with "pop" or "drain".  We would need

  // parallel versions of them if such concurrency was desired.

  void par_push(oop ptr);

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

  // Non-block impl.  Note: concurrency is allowed only with other

  // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".

  void par_adjoin_arr(oop* ptr_arr, int n);

  // 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);

  // Drain the mark stack, applying the given closure to all fields of

  // objects on the stack.  (That is, continue until the stack is empty,

  // even if closure applications add entries to the stack.)  The "bm"

  // argument, if non-null, may be used to verify that only marked objects

  // are on the mark stack.  If "yield_after" is "true", then the

  // concurrent marker performing the drain offers to yield after

  // processing each object.  If a yield occurs, stops the drain operation

  // and returns false.  Otherwise, returns true.

  template<class OopClosureClass>

  bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);

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

  bool isFull()     { return _index == _capacity; }

  int maxElems()    { return _capacity; }

  bool overflow() { return _overflow; }

  void clear_overflow() { _overflow = false; }

  int  size() { return _index; }

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

  // Record the current size; a subsequent "oops_do" will iterate only over

  // indices valid at the time of this call.

  void set_oops_do_bound(jint bound = -1) {

    if (bound == -1) {

      _oops_do_bound = _index;

    } else {

      _oops_do_bound = bound;

    }

  }

  jint oops_do_bound() { return _oops_do_bound; }

  // iterate over the oops in the mark stack, up to the bound recorded via

  // the call above.

  void oops_do(OopClosure* f);

};

class CMRegionStack VALUE_OBJ_CLASS_SPEC {

  MemRegion* _base;

  jint _capacity;

  jint _index;

  jint _oops_do_bound;

  bool _overflow;

public:

  CMRegionStack();

  ~CMRegionStack();

  void allocate(size_t size);

  // This is lock-free; assumes that it will only be called in parallel

  // with other "push" operations (no pops).

  void push(MemRegion mr);

  // Lock-free; assumes that it will only be called in parallel

  // with other "pop" operations (no pushes).

  MemRegion pop();

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

  bool isFull()     { return _index == _capacity; }

  bool overflow() { return _overflow; }

  void clear_overflow() { _overflow = false; }

  int  size() { return _index; }

  // It iterates over the entries in the region stack and it

  // invalidates (i.e. assigns MemRegion()) the ones that point to

  // regions in the collection set.

  bool invalidate_entries_into_cset();

  // This gives an upper bound up to which the iteration in

  // invalidate_entries_into_cset() will reach. This prevents

  // newly-added entries to be unnecessarily scanned.

  void set_oops_do_bound() {

    _oops_do_bound = _index;

  }

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

};

// this will enable a variety of different statistics per GC task

#define _MARKING_STATS_       0

// this will enable the higher verbose levels

#define _MARKING_VERBOSE_     0

#if _MARKING_STATS_

#define statsOnly(statement)  \

do {                          \

  statement ;                 \

} while (0)

#else // _MARKING_STATS_

#define statsOnly(statement)  \

do {                          \

} while (0)

#endif // _MARKING_STATS_

// Some extra guarantees that I like to also enable in optimised mode

// when debugging. If you want to enable them, comment out the assert

// macro and uncomment out the guaratee macro

// #define tmp_guarantee_CM(expr, str) guarantee(expr, str)

#define tmp_guarantee_CM(expr, str) assert(expr, str)

typedef enum {

  no_verbose  = 0,   // verbose turned off

  stats_verbose,     // only prints stats at the end of marking

  low_verbose,       // low verbose, mostly per region and per major event

  medium_verbose,    // a bit more detailed than low

  high_verbose       // per object verbose

} CMVerboseLevel;

class ConcurrentMarkThread;

class ConcurrentMark: public CHeapObj {

  friend class ConcurrentMarkThread;

  friend class CMTask;

  friend class CMBitMapClosure;

  friend class CSMarkOopClosure;

  friend class CMGlobalObjectClosure;

  friend class CMRemarkTask;

  friend class CMConcurrentMarkingTask;

  friend class G1ParNoteEndTask;

  friend class CalcLiveObjectsClosure;

protected:

  ConcurrentMarkThread* _cmThread;   // the thread doing the work

  G1CollectedHeap*      _g1h;        // the heap.

  size_t                _parallel_marking_threads; // the number of marking

                                                   // threads we'll 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

  // same as the two above, but for the cleanup task

  double                _cleanup_sleep_factor;

  double                _cleanup_task_overhead;

  // Stuff related to age cohort processing.

  struct ParCleanupThreadState {

    char _pre[64];

    UncleanRegionList list;

    char _post[64];

  };

  ParCleanupThreadState** _par_cleanup_thread_state;

  // CMS marking support structures

  CMBitMap                _markBitMap1;

  CMBitMap                _markBitMap2;

  CMBitMapRO*             _prevMarkBitMap; // completed mark bitmap

  CMBitMap*               _nextMarkBitMap; // under-construction mark bitmap

  bool                    _at_least_one_mark_complete;

  BitMap                  _region_bm;

  BitMap                  _card_bm;

  // Heap bounds

  HeapWord*               _heap_start;

  HeapWord*               _heap_end;

  // For gray objects

  CMMarkStack             _markStack; // Grey objects behind global finger.

  CMRegionStack           _regionStack; // Grey regions behind global finger.

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

                                    // always points to the end of the

                                    // last claimed region

  // marking tasks

  size_t                  _max_task_num; // maximum task number

  size_t                  _active_tasks; // task num currently active

  CMTask**                _tasks;        // task queue array (max_task_num len)

  CMTaskQueueSet*         _task_queues;  // task queue set

  ParallelTaskTerminator  _terminator;   // for termination

  // Two sync barriers that are used to synchronise 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-initialise

  // their data structures and task 0 re-initialises 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-initialised. 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;

  // verbose level

  CMVerboseLevel          _verbose_level;

  // These two fields are used to implement the optimisation that

  // avoids pushing objects on the global/region stack if there are

  // no collection set regions above the lowest finger.

  // This is the lowest finger (among the global and local fingers),

  // which is calculated before a new collection set is chosen.

  HeapWord* _min_finger;

  // If this flag is true, objects/regions that are marked below the

  // finger should be pushed on the stack(s). If this is flag is

  // false, it is safe not to push them on the stack(s).

  bool      _should_gray_objects;

  // 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 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();

  // It resets all the marking data structures.

  void clear_marking_state();

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

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

  void set_phase(size_t active_tasks, bool concurrent);

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

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

  void set_non_marking_state();

  // prints all gathered CM-related statistics

  void print_stats();

  // accessor methods

  size_t parallel_marking_threads() { return _parallel_marking_threads; }

  double sleep_factor()             { return _sleep_factor; }

  double marking_task_overhead()    { return _marking_task_overhead;}

  double cleanup_sleep_factor()     { return _cleanup_sleep_factor; }

  double cleanup_task_overhead()    { return _cleanup_task_overhead;}

  HeapWord*               finger()        { return _finger;   }

  bool                    concurrent()    { return _concurrent; }

  size_t                  active_tasks()  { return _active_tasks; }

  ParallelTaskTerminator* terminator()    { return &_terminator; }

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

  // task. 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(int task);

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

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

  // Returns the task with the given id

  CMTask* task(int id) {

    guarantee( 0 <= id && id < (int) _active_tasks, "task id not within "

               "active bounds" );

    return _tasks[id];

  }

  // Returns the task queue with the given id

  CMTaskQueue* task_queue(int id) {

    guarantee( 0 <= id && id < (int) _active_tasks, "task queue id not within "

               "active bounds" );

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

  }

  // Returns the task queue set

  CMTaskQueueSet* task_queues()  { return _task_queues; }

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

  // indicate that the global stack or region stack has overflown

  bool has_overflown()           { return _has_overflown; }

  void set_has_overflown()       { _has_overflown = true; }

  void clear_has_overflown()     { _has_overflown = false; }

  bool has_aborted()             { return _has_aborted; }

  bool restart_for_overflow()    { return _restart_for_overflow; }

  // Methods to enter the two overflow sync barriers

  void enter_first_sync_barrier(int task_num);

  void enter_second_sync_barrier(int task_num);

public:

  // Manipulation of the global mark stack.

  // Notice that the first mark_stack_push is CAS-based, whereas the

  // two below are Mutex-based. This is OK since the first one is only

  // called during evacuation pauses and doesn't compete with the

  // other two (which are called by the marking tasks during

  // concurrent marking or remark).

  bool mark_stack_push(oop p) {

    _markStack.par_push(p);

    if (_markStack.overflow()) {

      set_has_overflown();

      return false;