/*

 * Copyright (c) 2001, 2018, 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 "gc/g1/g1ConcurrentMarkBitMap.hpp"

#include "gc/g1/g1ConcurrentMarkObjArrayProcessor.hpp"

#include "gc/g1/g1HeapVerifier.hpp"

#include "gc/g1/g1RegionMarkStatsCache.hpp"

#include "gc/g1/heapRegionSet.hpp"

#include "gc/shared/taskqueue.hpp"

#include "memory/allocation.hpp"

class ConcurrentGCTimer;

class G1ConcurrentMarkThread;

class G1CollectedHeap;

class G1CMTask;

class G1ConcurrentMark;

class G1OldTracer;

class G1RegionToSpaceMapper;

class G1SurvivorRegions;

#ifdef _MSC_VER

#pragma warning(push)

// warning C4522: multiple assignment operators specified

#pragma warning(disable:4522)

#endif

// This is a container class for either an oop or a continuation address for

// mark stack entries. Both are pushed onto the mark stack.

class G1TaskQueueEntry {

private:

  void* _holder;

  static const uintptr_t ArraySliceBit = 1;

  G1TaskQueueEntry(oop obj) : _holder(obj) {

    assert(_holder != NULL, "Not allowed to set NULL task queue element");

  }

  G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }

public:

  G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }

  G1TaskQueueEntry() : _holder(NULL) { }

  static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }

  static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }

  G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {

    _holder = t._holder;

    return *this;

  }

  volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {

    _holder = t._holder;

    return *this;

  }

  oop obj() const {

    assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));

    return (oop)_holder;

  }

  HeapWord* slice() const {

    assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));

    return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);

  }

  bool is_oop() const { return !is_array_slice(); }

  bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }

  bool is_null() const { return _holder == NULL; }

};

#ifdef _MSC_VER

#pragma warning(pop)

#endif

typedef GenericTaskQueue<G1TaskQueueEntry, 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* _g1h;

public:

  G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }

  bool do_object_b(oop obj);

};

// 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 {

public:

  // Number of TaskQueueEntries that can fit in a single chunk.

  static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;

private:

  struct TaskQueueEntryChunk {

    TaskQueueEntryChunk* next;

    G1TaskQueueEntry data[EntriesPerChunk];

  };

  size_t _max_chunk_capacity;    // Maximum number of TaskQueueEntryChunk elements on the stack.

  TaskQueueEntryChunk* _base;    // Bottom address of allocated memory area.

  size_t _chunk_capacity;        // Current maximum number of TaskQueueEntryChunk elements.

  char _pad0[DEFAULT_CACHE_LINE_SIZE];

  TaskQueueEntryChunk* volatile _free_list;  // Linked list of free chunks that can be allocated by users.

  char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];

  TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.

  volatile size_t _chunks_in_chunk_list;

  char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];

  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.

  TaskQueueEntryChunk* allocate_new_chunk();

  // Atomically add the given chunk to the list.

  void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);

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

  // list is empty.

  TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);

  void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);

  void add_chunk_to_free_list(TaskQueueEntryChunk* elem);

  TaskQueueEntryChunk* remove_chunk_from_chunk_list();

  TaskQueueEntryChunk* remove_chunk_from_free_list();

  // 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);

  // Pushes the given buffer containing at most EntriesPerChunk elements on the mark

  // stack. If less than EntriesPerChunk elements are to be pushed, the array must

  // be terminated with a NULL.

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

  // stack.

  bool par_push_chunk(G1TaskQueueEntry* buffer);

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

  // chunk may contain up to EntriesPerChunk elements. If there are less, the last

  // element in the array is a NULL pointer.

  bool par_pop_chunk(G1TaskQueueEntry* buffer);

  // 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; }

  // 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.

  size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }

  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 {

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();

};

// This class manages data structures and methods for doing liveness analysis in

// G1's concurrent cycle.

class G1ConcurrentMark : public CHeapObj<mtGC> {

  friend class G1ConcurrentMarkThread;

  friend class G1CMRefProcTaskProxy;

  friend class G1CMRefProcTaskExecutor;

  friend class G1CMKeepAliveAndDrainClosure;

  friend class G1CMDrainMarkingStackClosure;

  friend class G1CMBitMapClosure;

  friend class G1CMConcurrentMarkingTask;

  friend class G1CMRemarkTask;

  friend class G1CMTask;

  G1ConcurrentMarkThread* _cm_thread;     // The thread doing the work

  G1CollectedHeap*        _g1h;           // The heap

  bool                    _completed_initialization; // Set to true when initialization is complete

  // Concurrent marking support structures

  G1CMBitMap              _mark_bitmap_1;

  G1CMBitMap              _mark_bitmap_2;

  G1CMBitMap*             _prev_mark_bitmap; // Completed mark bitmap

  G1CMBitMap*             _next_mark_bitmap; // Under-construction mark bitmap

  // Heap bounds

  MemRegion const         _heap;

  // Root region tracking and claiming

  G1CMRootRegions         _root_regions;

  // For grey objects

  G1CMMarkStack           _global_mark_stack; // Grey objects behind global finger

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

                                              // always pointing to the end of the

                                              // last claimed region

  uint                    _worker_id_offset;

  uint                    _max_num_tasks;    // Maximum number of marking tasks

  uint                    _num_active_tasks; // Number of tasks currently active

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

  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;

  ConcurrentGCTimer*      _gc_timer_cm;

  G1OldTracer*            _gc_tracer_cm;

  // Timing statistics. All of them are in ms

  NumberSeq _init_times;

  NumberSeq _remark_times;

  NumberSeq _remark_mark_times;

  NumberSeq _remark_weak_ref_times;

  NumberSeq _cleanup_times;

  double    _total_cleanup_time;

  double*   _accum_task_vtime;   // Accumulated task vtime

  WorkGang* _concurrent_workers;

  uint      _num_concurrent_workers; // The number of marking worker threads we're using

  uint      _max_concurrent_workers; // Maximum number of marking worker threads

  void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller);

  void finalize_marking();

  void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);

  void weak_refs_work(bool clear_all_soft_refs);

  void report_object_count(bool mark_completed);

  void swap_mark_bitmaps();

  void reclaim_empty_regions();

  // After reclaiming empty regions, update heap sizes.

  void compute_new_sizes();

  // Clear statistics gathered during the concurrent cycle for the given region after

  // it has been reclaimed.

  void clear_statistics(HeapRegion* r);

  // 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_for_restart();

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

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

  void reset_at_marking_complete();

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

  void set_concurrency(uint active_tasks);

  // 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();

  HeapWord*               finger()          { return _finger;   }

  bool                    concurrent()      { return _concurrent; }

  uint                    active_tasks()    { return _num_active_tasks; }

  ParallelTaskTerminator* terminator()      { return &_terminator; }

  // 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);

  // 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(uint id) {

    // During initial mark we use the parallel gc threads to do some work, so

    // we can only compare against _max_num_tasks.

    assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks);

    return _tasks[id];

  }

  // 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);

  // Clear the given bitmap in parallel using the given WorkGang. If may_yield is

  // true, periodically insert checks to see if this method should exit prematurely.

  void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);

  // Region statistics gathered during marking.

  G1RegionMarkStats* _region_mark_stats;

  // Top pointer for each region at the start of the rebuild remembered set process

  // for regions which remembered sets need to be rebuilt. A NULL for a given region

  // means that this region does not be scanned during the rebuilding remembered

  // set phase at all.

  HeapWord* volatile* _top_at_rebuild_starts;

public:

  void add_to_liveness(uint worker_id, oop const obj, size_t size);

  // Liveness of the given region as determined by concurrent marking, i.e. the amount of

  // live words between bottom and nTAMS.

  // Sets the internal top_at_region_start for the given region to current top of the region.

  inline void update_top_at_rebuild_start(HeapRegion* r);

  // TARS for the given region during remembered set rebuilding.

  inline HeapWord* top_at_rebuild_start(uint region) const;

  // Clear statistics gathered during the concurrent cycle for the given region after

  // it has been reclaimed.

  void clear_statistics_in_region(uint region_idx);

  // Notification for eagerly reclaimed regions to clean up.

  void humongous_object_eagerly_reclaimed(HeapRegion* r);

  // Manipulation of the global mark stack.

  // The push and pop operations are used by tasks for transfers

  // between task-local queues and the global mark stack.

  bool mark_stack_push(G1TaskQueueEntry* arr) {

    if (!_global_mark_stack.par_push_chunk(arr)) {

      set_has_overflown();

      return false;

    }

    return true;