1 /*
2 * Copyright (c) 2001, 2016, 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/g1RegionToSpaceMapper.hpp"
30 #include "gc/g1/heapRegionSet.hpp"
31 #include "gc/shared/taskqueue.hpp"
32
33 class G1CollectedHeap;
34 class G1CMBitMap;
35 class G1CMTask;
36 class G1ConcurrentMark;
37 class ConcurrentGCTimer;
38 class G1OldTracer;
39 class G1SurvivorRegions;
40 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue;
41 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
42
43 // Closure used by CM during concurrent reference discovery
44 // and reference processing (during remarking) to determine
45 // if a particular object is alive. It is primarily used
46 // to determine if referents of discovered reference objects
47 // are alive. An instance is also embedded into the
48 // reference processor as the _is_alive_non_header field
49 class G1CMIsAliveClosure: public BoolObjectClosure {
50 G1CollectedHeap* _g1;
51 public:
52 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
53
54 bool do_object_b(oop obj);
55 };
56
57 // A generic CM bit map. This is essentially a wrapper around the BitMap
58 // class, with one bit per (1<<_shifter) HeapWords.
59
60 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
61 protected:
62 HeapWord* _bmStartWord; // base address of range covered by map
63 size_t _bmWordSize; // map size (in #HeapWords covered)
64 const int _shifter; // map to char or bit
65 BitMapView _bm; // the bit map itself
66
67 public:
68 // constructor
69 G1CMBitMapRO(int shifter);
70
71 // inquiries
72 HeapWord* startWord() const { return _bmStartWord; }
73 // the following is one past the last word in space
74 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
75
76 // read marks
77
78 bool isMarked(HeapWord* addr) const {
79 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
80 "outside underlying space?");
81 return _bm.at(heapWordToOffset(addr));
82 }
83
84 // iteration
85 inline bool iterate(BitMapClosure* cl, MemRegion mr);
86
87 // Return the address corresponding to the next marked bit at or after
88 // "addr", and before "limit", if "limit" is non-NULL. If there is no
89 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
90 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
91 const HeapWord* limit = NULL) const;
92
93 // conversion utilities
94 HeapWord* offsetToHeapWord(size_t offset) const {
95 return _bmStartWord + (offset << _shifter);
96 }
97 size_t heapWordToOffset(const HeapWord* addr) const {
98 return pointer_delta(addr, _bmStartWord) >> _shifter;
99 }
100
101 // The argument addr should be the start address of a valid object
102 inline HeapWord* nextObject(HeapWord* addr);
103
104 void print_on_error(outputStream* st, const char* prefix) const;
105
106 // debugging
107 NOT_PRODUCT(bool covers(MemRegion rs) const;)
108 };
109
110 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
111 private:
112 G1CMBitMap* _bm;
113 public:
114 G1CMBitMapMappingChangedListener() : _bm(NULL) {}
115
116 void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
117
118 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
119 };
120
121 class G1CMBitMap : public G1CMBitMapRO {
122 private:
123 G1CMBitMapMappingChangedListener _listener;
124
125 public:
126 static size_t compute_size(size_t heap_size);
127 // Returns the amount of bytes on the heap between two marks in the bitmap.
128 static size_t mark_distance();
129 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
130 // mark bitmap corresponds to. This is the same as the mark distance above.
131 static size_t heap_map_factor() {
132 return mark_distance();
133 }
134
135 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
136
137 // Initializes the underlying BitMap to cover the given area.
138 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
139
140 // Write marks.
141 inline void mark(HeapWord* addr);
142 inline void clear(HeapWord* addr);
143 inline bool parMark(HeapWord* addr);
144
145 void clear_range(MemRegion mr);
146 };
147
148 // Represents the overflow mark stack used by concurrent marking.
149 //
150 // Stores oops in a huge buffer in virtual memory that is always fully committed.
151 // Resizing may only happen during a STW pause when the stack is empty.
152 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
153 ReservedSpace _reserved_space; // Space currently reserved for the mark stack.
154
155 oop* _base; // Bottom address of allocated memory area.
156 size_t _capacity; // Maximum number of elements.
157 size_t _index; // One more than last occupied index.
158
159 size_t _saved_index; // Value of _index saved at start of GC to detect mark stack modifications during that time.
160
161 bool _overflow;
162 bool _should_expand;
163
164 // Resizes the mark stack to the given new capacity. Releases any previous
165 // memory if successful.
166 bool resize(size_t new_capacity);
167
168 bool stack_modified() const { return _index != _saved_index; }
169 public:
170 G1CMMarkStack();
171 ~G1CMMarkStack();
172
173 bool allocate(size_t capacity);
174
175 // Pushes the first "n" elements of the given buffer on the stack.
176 void par_push_arr(oop* buffer, size_t n);
177
178 // Moves up to max elements from the stack into the given buffer. Returns
179 // the number of elements pushed, and false if the array has been empty.
180 // Returns true if the buffer contains at least one element.
181 bool par_pop_arr(oop* buffer, size_t max, size_t* n);
182
183 bool is_empty() const { return _index == 0; }
184 size_t capacity() const { return _capacity; }
185
186 bool overflow() const { return _overflow; }
187 void clear_overflow() { _overflow = false; }
188
189 bool should_expand() const { return _should_expand; }
190 void set_should_expand(bool value) { _should_expand = value; }
191
192 // Expand the stack, typically in response to an overflow condition
193 void expand();
194
195 size_t size() const { return _index; }
196
197 void set_empty() { _index = 0; clear_overflow(); }
198
199 // Record the current index.
200 void note_start_of_gc();
201
202 // Make sure that we have not added any entries to the stack during GC.
203 void note_end_of_gc();
204
205 // Apply fn to each oop in the mark stack, up to the bound recorded
206 // via one of the above "note" functions. The mark stack must not
207 // be modified while iterating.
208 template<typename Fn> void iterate(Fn fn);
209 };
210
211 // Root Regions are regions that are not empty at the beginning of a
212 // marking cycle and which we might collect during an evacuation pause
213 // while the cycle is active. Given that, during evacuation pauses, we
214 // do not copy objects that are explicitly marked, what we have to do
215 // for the root regions is to scan them and mark all objects reachable
216 // from them. According to the SATB assumptions, we only need to visit
217 // each object once during marking. So, as long as we finish this scan
218 // before the next evacuation pause, we can copy the objects from the
219 // root regions without having to mark them or do anything else to them.
220 //
221 // Currently, we only support root region scanning once (at the start
222 // of the marking cycle) and the root regions are all the survivor
223 // regions populated during the initial-mark pause.
224 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
225 private:
226 const G1SurvivorRegions* _survivors;
227 G1ConcurrentMark* _cm;
228
229 volatile bool _scan_in_progress;
230 volatile bool _should_abort;
231 volatile int _claimed_survivor_index;
232
233 void notify_scan_done();
234
235 public:
236 G1CMRootRegions();
237 // We actually do most of the initialization in this method.
238 void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
239
240 // Reset the claiming / scanning of the root regions.
241 void prepare_for_scan();
242
243 // Forces get_next() to return NULL so that the iteration aborts early.
244 void abort() { _should_abort = true; }
245
246 // Return true if the CM thread are actively scanning root regions,
247 // false otherwise.
248 bool scan_in_progress() { return _scan_in_progress; }
249
250 // Claim the next root region to scan atomically, or return NULL if
251 // all have been claimed.
252 HeapRegion* claim_next();
253
254 // The number of root regions to scan.
255 uint num_root_regions() const;
256
257 void cancel_scan();
258
259 // Flag that we're done with root region scanning and notify anyone
260 // who's waiting on it. If aborted is false, assume that all regions
261 // have been claimed.
262 void scan_finished();
263
264 // If CM threads are still scanning root regions, wait until they
265 // are done. Return true if we had to wait, false otherwise.
266 bool wait_until_scan_finished();
267 };
268
269 class ConcurrentMarkThread;
270
271 class G1ConcurrentMark: public CHeapObj<mtGC> {
272 friend class ConcurrentMarkThread;
273 friend class G1ParNoteEndTask;
274 friend class G1VerifyLiveDataClosure;
275 friend class G1CMRefProcTaskProxy;
276 friend class G1CMRefProcTaskExecutor;
277 friend class G1CMKeepAliveAndDrainClosure;
278 friend class G1CMDrainMarkingStackClosure;
279 friend class G1CMBitMapClosure;
280 friend class G1CMConcurrentMarkingTask;
281 friend class G1CMMarkStack;
282 friend class G1CMRemarkTask;
283 friend class G1CMTask;
284
285 protected:
286 ConcurrentMarkThread* _cmThread; // The thread doing the work
287 G1CollectedHeap* _g1h; // The heap
288 uint _parallel_marking_threads; // The number of marking
289 // threads we're using
290 uint _max_parallel_marking_threads; // Max number of marking
291 // threads we'll ever use
292 double _sleep_factor; // How much we have to sleep, with
293 // respect to the work we just did, to
294 // meet the marking overhead goal
295 double _marking_task_overhead; // Marking target overhead for
296 // a single task
297
298 FreeRegionList _cleanup_list;
299
300 // Concurrent marking support structures
301 G1CMBitMap _markBitMap1;
302 G1CMBitMap _markBitMap2;
303 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
304 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
305
306 // Heap bounds
307 HeapWord* _heap_start;
308 HeapWord* _heap_end;
309
310 // Root region tracking and claiming
311 G1CMRootRegions _root_regions;
312
313 // For gray objects
314 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger
315 HeapWord* volatile _finger; // The global finger, region aligned,
316 // always points to the end of the
317 // last claimed region
318
319 // Marking tasks
320 uint _max_worker_id;// Maximum worker id
321 uint _active_tasks; // Task num currently active
322 G1CMTask** _tasks; // Task queue array (max_worker_id len)
323 G1CMTaskQueueSet* _task_queues; // Task queue set
324 ParallelTaskTerminator _terminator; // For termination
325
326 // Two sync barriers that are used to synchronize tasks when an
327 // overflow occurs. The algorithm is the following. All tasks enter
328 // the first one to ensure that they have all stopped manipulating
329 // the global data structures. After they exit it, they re-initialize
330 // their data structures and task 0 re-initializes the global data
331 // structures. Then, they enter the second sync barrier. This
332 // ensure, that no task starts doing work before all data
333 // structures (local and global) have been re-initialized. When they
334 // exit it, they are free to start working again.
335 WorkGangBarrierSync _first_overflow_barrier_sync;
336 WorkGangBarrierSync _second_overflow_barrier_sync;
337
338 // This is set by any task, when an overflow on the global data
339 // structures is detected
340 volatile bool _has_overflown;
341 // True: marking is concurrent, false: we're in remark
342 volatile bool _concurrent;
343 // Set at the end of a Full GC so that marking aborts
344 volatile bool _has_aborted;
345
346 // Used when remark aborts due to an overflow to indicate that
347 // another concurrent marking phase should start
348 volatile bool _restart_for_overflow;
349
350 // This is true from the very start of concurrent marking until the
351 // point when all the tasks complete their work. It is really used
352 // to determine the points between the end of concurrent marking and
353 // time of remark.
354 volatile bool _concurrent_marking_in_progress;
355
356 ConcurrentGCTimer* _gc_timer_cm;
357
358 G1OldTracer* _gc_tracer_cm;
359
360 // All of these times are in ms
361 NumberSeq _init_times;
362 NumberSeq _remark_times;
363 NumberSeq _remark_mark_times;
364 NumberSeq _remark_weak_ref_times;
365 NumberSeq _cleanup_times;
366 double _total_counting_time;
367 double _total_rs_scrub_time;
368
369 double* _accum_task_vtime; // Accumulated task vtime
370
371 WorkGang* _parallel_workers;
372
373 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
374 void weakRefsWork(bool clear_all_soft_refs);
375
376 void swapMarkBitMaps();
377
378 // It resets the global marking data structures, as well as the
379 // task local ones; should be called during initial mark.
380 void reset();
381
382 // Resets all the marking data structures. Called when we have to restart
383 // marking or when marking completes (via set_non_marking_state below).
384 void reset_marking_state(bool clear_overflow = true);
385
386 // We do this after we're done with marking so that the marking data
387 // structures are initialized to a sensible and predictable state.
388 void set_non_marking_state();
389
390 // Called to indicate how many threads are currently active.
391 void set_concurrency(uint active_tasks);
392
393 // It should be called to indicate which phase we're in (concurrent
394 // mark or remark) and how many threads are currently active.
395 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
396
397 // Prints all gathered CM-related statistics
398 void print_stats();
399
400 bool cleanup_list_is_empty() {
401 return _cleanup_list.is_empty();
402 }
403
404 // Accessor methods
405 uint parallel_marking_threads() const { return _parallel_marking_threads; }
406 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
407 double sleep_factor() { return _sleep_factor; }
408 double marking_task_overhead() { return _marking_task_overhead;}
409
410 HeapWord* finger() { return _finger; }
411 bool concurrent() { return _concurrent; }
412 uint active_tasks() { return _active_tasks; }
413 ParallelTaskTerminator* terminator() { return &_terminator; }
414
415 // It claims the next available region to be scanned by a marking
416 // task/thread. It might return NULL if the next region is empty or
417 // we have run out of regions. In the latter case, out_of_regions()
418 // determines whether we've really run out of regions or the task
419 // should call claim_region() again. This might seem a bit
420 // awkward. Originally, the code was written so that claim_region()
421 // either successfully returned with a non-empty region or there
422 // were no more regions to be claimed. The problem with this was
423 // that, in certain circumstances, it iterated over large chunks of
424 // the heap finding only empty regions and, while it was working, it
425 // was preventing the calling task to call its regular clock
426 // method. So, this way, each task will spend very little time in
427 // claim_region() and is allowed to call the regular clock method
428 // frequently.
429 HeapRegion* claim_region(uint worker_id);
430
431 // It determines whether we've run out of regions to scan. Note that
432 // the finger can point past the heap end in case the heap was expanded
433 // to satisfy an allocation without doing a GC. This is fine, because all
434 // objects in those regions will be considered live anyway because of
435 // SATB guarantees (i.e. their TAMS will be equal to bottom).
436 bool out_of_regions() { return _finger >= _heap_end; }
437
438 // Returns the task with the given id
439 G1CMTask* task(int id) {
440 assert(0 <= id && id < (int) _active_tasks,
441 "task id not within active bounds");
442 return _tasks[id];
443 }
444
445 // Returns the task queue with the given id
446 G1CMTaskQueue* task_queue(int id) {
447 assert(0 <= id && id < (int) _active_tasks,
448 "task queue id not within active bounds");
449 return (G1CMTaskQueue*) _task_queues->queue(id);
450 }
451
452 // Returns the task queue set
453 G1CMTaskQueueSet* task_queues() { return _task_queues; }
454
455 // Access / manipulation of the overflow flag which is set to
456 // indicate that the global stack has overflown
457 bool has_overflown() { return _has_overflown; }
458 void set_has_overflown() { _has_overflown = true; }
459 void clear_has_overflown() { _has_overflown = false; }
460 bool restart_for_overflow() { return _restart_for_overflow; }
461
462 // Methods to enter the two overflow sync barriers
463 void enter_first_sync_barrier(uint worker_id);
464 void enter_second_sync_barrier(uint worker_id);
465
466 // Card index of the bottom of the G1 heap. Used for biasing indices into
467 // the card bitmaps.
468 intptr_t _heap_bottom_card_num;
469
470 // Set to true when initialization is complete
471 bool _completed_initialization;
472
473 // end_timer, true to end gc timer after ending concurrent phase.
474 void register_concurrent_phase_end_common(bool end_timer);
475
476 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
477 // true, periodically insert checks to see if this method should exit prematurely.
478 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
479 public:
480 // Manipulation of the global mark stack.
481 // The push and pop operations are used by tasks for transfers
482 // between task-local queues and the global mark stack, and use
483 // locking for concurrency safety.
484 bool mark_stack_push(oop* arr, size_t n) {
485 _global_mark_stack.par_push_arr(arr, n);
486 if (_global_mark_stack.overflow()) {
487 set_has_overflown();
488 return false;
489 }
490 return true;
491 }
492 void mark_stack_pop(oop* arr, size_t max, size_t* n) {
493 _global_mark_stack.par_pop_arr(arr, max, n);
494 }
495 size_t mark_stack_size() { return _global_mark_stack.size(); }
496 size_t partial_mark_stack_size_target() { return _global_mark_stack.capacity()/3; }
497 bool mark_stack_overflow() { return _global_mark_stack.overflow(); }
498 bool mark_stack_empty() { return _global_mark_stack.is_empty(); }
499
500 G1CMRootRegions* root_regions() { return &_root_regions; }
501
502 bool concurrent_marking_in_progress() {
503 return _concurrent_marking_in_progress;
504 }
505 void set_concurrent_marking_in_progress() {
506 _concurrent_marking_in_progress = true;
507 }
508 void clear_concurrent_marking_in_progress() {
509 _concurrent_marking_in_progress = false;
510 }
511
512 void concurrent_cycle_start();
513 void concurrent_cycle_end();
514
515 void update_accum_task_vtime(int i, double vtime) {
516 _accum_task_vtime[i] += vtime;
517 }
518
519 double all_task_accum_vtime() {
520 double ret = 0.0;
521 for (uint i = 0; i < _max_worker_id; ++i)
522 ret += _accum_task_vtime[i];
523 return ret;
524 }
525
526 // Attempts to steal an object from the task queues of other tasks
527 bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
528
529 G1ConcurrentMark(G1CollectedHeap* g1h,
530 G1RegionToSpaceMapper* prev_bitmap_storage,
531 G1RegionToSpaceMapper* next_bitmap_storage);
532 ~G1ConcurrentMark();
533
534 ConcurrentMarkThread* cmThread() { return _cmThread; }
535
536 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
537 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
538
539 // Returns the number of GC threads to be used in a concurrent
540 // phase based on the number of GC threads being used in a STW
541 // phase.
542 uint scale_parallel_threads(uint n_par_threads);
543
544 // Calculates the number of GC threads to be used in a concurrent phase.
545 uint calc_parallel_marking_threads();
546
547 // The following three are interaction between CM and
548 // G1CollectedHeap
549
550 // This notifies CM that a root during initial-mark needs to be
551 // grayed. It is MT-safe. hr is the region that
552 // contains the object and it's passed optionally from callers who
553 // might already have it (no point in recalculating it).
554 inline void grayRoot(oop obj,
555 HeapRegion* hr = NULL);
556
557 // Prepare internal data structures for the next mark cycle. This includes clearing
558 // the next mark bitmap and some internal data structures. This method is intended
559 // to be called concurrently to the mutator. It will yield to safepoint requests.
560 void cleanup_for_next_mark();
561
562 // Clear the previous marking bitmap during safepoint.
563 void clear_prev_bitmap(WorkGang* workers);
564
565 // Return whether the next mark bitmap has no marks set. To be used for assertions
566 // only. Will not yield to pause requests.
567 bool nextMarkBitmapIsClear();
568
569 // These two do the work that needs to be done before and after the
570 // initial root checkpoint. Since this checkpoint can be done at two
571 // different points (i.e. an explicit pause or piggy-backed on a
572 // young collection), then it's nice to be able to easily share the
573 // pre/post code. It might be the case that we can put everything in
574 // the post method. TP
575 void checkpointRootsInitialPre();
576 void checkpointRootsInitialPost();
577
578 // Scan all the root regions and mark everything reachable from
579 // them.
580 void scan_root_regions();
581
582 // Scan a single root region and mark everything reachable from it.
583 void scanRootRegion(HeapRegion* hr);
584
585 // Do concurrent phase of marking, to a tentative transitive closure.
586 void mark_from_roots();
587
588 void checkpointRootsFinal(bool clear_all_soft_refs);
589 void checkpointRootsFinalWork();
590 void cleanup();
591 void complete_cleanup();
592
593 // Mark in the previous bitmap. NB: this is usually read-only, so use
594 // this carefully!
595 inline void markPrev(oop p);
596
597 // Clears marks for all objects in the given range, for the prev or
598 // next bitmaps. NB: the previous bitmap is usually
599 // read-only, so use this carefully!
600 void clearRangePrevBitmap(MemRegion mr);
601
602 // Notify data structures that a GC has started.
603 void note_start_of_gc() {
604 _global_mark_stack.note_start_of_gc();
605 }
606
607 // Notify data structures that a GC is finished.
608 void note_end_of_gc() {
609 _global_mark_stack.note_end_of_gc();
610 }
611
612 // Verify that there are no CSet oops on the stacks (taskqueues /
613 // global mark stack) and fingers (global / per-task).
614 // If marking is not in progress, it's a no-op.
615 void verify_no_cset_oops() PRODUCT_RETURN;
616
617 inline bool isPrevMarked(oop p) const;
618
619 inline bool do_yield_check();
620
621 // Abandon current marking iteration due to a Full GC.
622 void abort();
623
624 bool has_aborted() { return _has_aborted; }
625
626 void print_summary_info();
627
628 void print_worker_threads_on(outputStream* st) const;
629 void threads_do(ThreadClosure* tc) const;
630
631 void print_on_error(outputStream* st) const;
632
633 // Attempts to mark the given object on the next mark bitmap.
634 inline bool par_mark(oop obj);
635
636 // Returns true if initialization was successfully completed.
637 bool completed_initialization() const {
638 return _completed_initialization;
639 }
640
641 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
642 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
643
644 private:
645 // Clear (Reset) all liveness count data.
646 void clear_live_data(WorkGang* workers);
647
648 #ifdef ASSERT
649 // Verify all of the above data structures that they are in initial state.
650 void verify_live_data_clear();
651 #endif
652
653 // Aggregates the per-card liveness data based on the current marking. Also sets
654 // the amount of marked bytes for each region.
655 void create_live_data();
656
657 void finalize_live_data();
658
659 void verify_live_data();
660 };
661
662 // A class representing a marking task.
663 class G1CMTask : public TerminatorTerminator {
664 private:
665 enum PrivateConstants {
666 // The regular clock call is called once the scanned words reaches
667 // this limit
668 words_scanned_period = 12*1024,
669 // The regular clock call is called once the number of visited
670 // references reaches this limit
671 refs_reached_period = 384,
672 // Initial value for the hash seed, used in the work stealing code
673 init_hash_seed = 17,
674 // How many entries will be transferred between global stack and
675 // local queues at once.
676 global_stack_transfer_size = 1024
677 };
678
679 uint _worker_id;
680 G1CollectedHeap* _g1h;
681 G1ConcurrentMark* _cm;
682 G1CMBitMap* _nextMarkBitMap;
683 // the task queue of this task
684 G1CMTaskQueue* _task_queue;
685 private:
686 // the task queue set---needed for stealing
687 G1CMTaskQueueSet* _task_queues;
688 // indicates whether the task has been claimed---this is only for
689 // debugging purposes
690 bool _claimed;
691
692 // number of calls to this task
693 int _calls;
694
695 // when the virtual timer reaches this time, the marking step should
696 // exit
697 double _time_target_ms;
698 // the start time of the current marking step
699 double _start_time_ms;
700
701 // the oop closure used for iterations over oops
702 G1CMOopClosure* _cm_oop_closure;
703
704 // the region this task is scanning, NULL if we're not scanning any
705 HeapRegion* _curr_region;
706 // the local finger of this task, NULL if we're not scanning a region
707 HeapWord* _finger;
708 // limit of the region this task is scanning, NULL if we're not scanning one
709 HeapWord* _region_limit;
710
711 // the number of words this task has scanned
712 size_t _words_scanned;
713 // When _words_scanned reaches this limit, the regular clock is
714 // called. Notice that this might be decreased under certain
715 // circumstances (i.e. when we believe that we did an expensive
716 // operation).
717 size_t _words_scanned_limit;
718 // the initial value of _words_scanned_limit (i.e. what it was
719 // before it was decreased).
720 size_t _real_words_scanned_limit;
721
722 // the number of references this task has visited
723 size_t _refs_reached;
724 // When _refs_reached reaches this limit, the regular clock is
725 // called. Notice this this might be decreased under certain
726 // circumstances (i.e. when we believe that we did an expensive
727 // operation).
728 size_t _refs_reached_limit;
729 // the initial value of _refs_reached_limit (i.e. what it was before
730 // it was decreased).
731 size_t _real_refs_reached_limit;
732
733 // used by the work stealing stuff
734 int _hash_seed;
735 // if this is true, then the task has aborted for some reason
736 bool _has_aborted;
737 // set when the task aborts because it has met its time quota
738 bool _has_timed_out;
739 // true when we're draining SATB buffers; this avoids the task
740 // aborting due to SATB buffers being available (as we're already
741 // dealing with them)
742 bool _draining_satb_buffers;
743
744 // number sequence of past step times
745 NumberSeq _step_times_ms;
746 // elapsed time of this task
747 double _elapsed_time_ms;
748 // termination time of this task
749 double _termination_time_ms;
750 // when this task got into the termination protocol
751 double _termination_start_time_ms;
752
753 // true when the task is during a concurrent phase, false when it is
754 // in the remark phase (so, in the latter case, we do not have to
755 // check all the things that we have to check during the concurrent
756 // phase, i.e. SATB buffer availability...)
757 bool _concurrent;
758
759 TruncatedSeq _marking_step_diffs_ms;
760
761 // it updates the local fields after this task has claimed
762 // a new region to scan
763 void setup_for_region(HeapRegion* hr);
764 // it brings up-to-date the limit of the region
765 void update_region_limit();
766
767 // called when either the words scanned or the refs visited limit
768 // has been reached
769 void reached_limit();
770 // recalculates the words scanned and refs visited limits
771 void recalculate_limits();
772 // decreases the words scanned and refs visited limits when we reach
773 // an expensive operation
774 void decrease_limits();
775 // it checks whether the words scanned or refs visited reached their
776 // respective limit and calls reached_limit() if they have
777 void check_limits() {
778 if (_words_scanned >= _words_scanned_limit ||
779 _refs_reached >= _refs_reached_limit) {
780 reached_limit();
781 }
782 }
783 // this is supposed to be called regularly during a marking step as
784 // it checks a bunch of conditions that might cause the marking step
785 // to abort
786 void regular_clock_call();
787 bool concurrent() { return _concurrent; }
788
789 // Test whether obj might have already been passed over by the
790 // mark bitmap scan, and so needs to be pushed onto the mark stack.
791 bool is_below_finger(oop obj, HeapWord* global_finger) const;
792
793 template<bool scan> void process_grey_object(oop obj);
794
795 public:
796 // It resets the task; it should be called right at the beginning of
797 // a marking phase.
798 void reset(G1CMBitMap* _nextMarkBitMap);
799 // it clears all the fields that correspond to a claimed region.
800 void clear_region_fields();
801
802 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
803
804 // The main method of this class which performs a marking step
805 // trying not to exceed the given duration. However, it might exit
806 // prematurely, according to some conditions (i.e. SATB buffers are
807 // available for processing).
808 void do_marking_step(double target_ms,
809 bool do_termination,
810 bool is_serial);
811
812 // These two calls start and stop the timer
813 void record_start_time() {
814 _elapsed_time_ms = os::elapsedTime() * 1000.0;
815 }
816 void record_end_time() {
817 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
818 }
819
820 // returns the worker ID associated with this task.
821 uint worker_id() { return _worker_id; }
822
823 // From TerminatorTerminator. It determines whether this task should
824 // exit the termination protocol after it's entered it.
825 virtual bool should_exit_termination();
826
827 // Resets the local region fields after a task has finished scanning a
828 // region; or when they have become stale as a result of the region
829 // being evacuated.
830 void giveup_current_region();
831
832 HeapWord* finger() { return _finger; }
833
834 bool has_aborted() { return _has_aborted; }
835 void set_has_aborted() { _has_aborted = true; }
836 void clear_has_aborted() { _has_aborted = false; }
837 bool has_timed_out() { return _has_timed_out; }
838 bool claimed() { return _claimed; }
839
840 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
841
842 // Increment the number of references this task has visited.
843 void increment_refs_reached() { ++_refs_reached; }
844
845 // Grey the object by marking it. If not already marked, push it on
846 // the local queue if below the finger.
847 // obj is below its region's NTAMS.
848 inline void make_reference_grey(oop obj);
849
850 // Grey the object (by calling make_grey_reference) if required,
851 // e.g. obj is below its containing region's NTAMS.
852 // Precondition: obj is a valid heap object.
853 inline void deal_with_reference(oop obj);
854
855 // It scans an object and visits its children.
856 inline void scan_object(oop obj);
857
858 // It pushes an object on the local queue.
859 inline void push(oop obj);
860
861 // These two move entries to/from the global stack.
862 void move_entries_to_global_stack();
863 void get_entries_from_global_stack();
864
865 // It pops and scans objects from the local queue. If partially is
866 // true, then it stops when the queue size is of a given limit. If
867 // partially is false, then it stops when the queue is empty.
868 void drain_local_queue(bool partially);
869 // It moves entries from the global stack to the local queue and
870 // drains the local queue. If partially is true, then it stops when
871 // both the global stack and the local queue reach a given size. If
872 // partially if false, it tries to empty them totally.
873 void drain_global_stack(bool partially);
874 // It keeps picking SATB buffers and processing them until no SATB
875 // buffers are available.
876 void drain_satb_buffers();
877
878 // moves the local finger to a new location
879 inline void move_finger_to(HeapWord* new_finger) {
880 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
881 _finger = new_finger;
882 }
883
884 G1CMTask(uint worker_id,
885 G1ConcurrentMark *cm,
886 G1CMTaskQueue* task_queue,
887 G1CMTaskQueueSet* task_queues);
888
889 // it prints statistics associated with this task
890 void print_stats();
891 };
892
893 // Class that's used to to print out per-region liveness
894 // information. It's currently used at the end of marking and also
895 // after we sort the old regions at the end of the cleanup operation.
896 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
897 private:
898 // Accumulators for these values.
899 size_t _total_used_bytes;
900 size_t _total_capacity_bytes;
901 size_t _total_prev_live_bytes;
902 size_t _total_next_live_bytes;
903
904 // Accumulator for the remembered set size
905 size_t _total_remset_bytes;
906
907 // Accumulator for strong code roots memory size
908 size_t _total_strong_code_roots_bytes;
909
910 static double perc(size_t val, size_t total) {
911 if (total == 0) {
912 return 0.0;
913 } else {
914 return 100.0 * ((double) val / (double) total);
915 }
916 }
917
918 static double bytes_to_mb(size_t val) {
919 return (double) val / (double) M;
920 }
921
922 public:
923 // The header and footer are printed in the constructor and
924 // destructor respectively.
925 G1PrintRegionLivenessInfoClosure(const char* phase_name);
926 virtual bool doHeapRegion(HeapRegion* r);
927 ~G1PrintRegionLivenessInfoClosure();
928 };
929
930 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP