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 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue;
40 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
41
42 // Closure used by CM during concurrent reference discovery
43 // and reference processing (during remarking) to determine
44 // if a particular object is alive. It is primarily used
45 // to determine if referents of discovered reference objects
46 // are alive. An instance is also embedded into the
47 // reference processor as the _is_alive_non_header field
48 class G1CMIsAliveClosure: public BoolObjectClosure {
49 G1CollectedHeap* _g1;
50 public:
51 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
52
53 bool do_object_b(oop obj);
54 };
55
56 // A generic CM bit map. This is essentially a wrapper around the BitMap
57 // class, with one bit per (1<<_shifter) HeapWords.
58
59 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
60 protected:
61 HeapWord* _bmStartWord; // base address of range covered by map
62 size_t _bmWordSize; // map size (in #HeapWords covered)
63 const int _shifter; // map to char or bit
64 BitMap _bm; // the bit map itself
65
66 public:
67 // constructor
68 G1CMBitMapRO(int shifter);
69
70 // inquiries
71 HeapWord* startWord() const { return _bmStartWord; }
72 // the following is one past the last word in space
73 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
74
75 // read marks
76
77 bool isMarked(HeapWord* addr) const {
78 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
79 "outside underlying space?");
80 return _bm.at(heapWordToOffset(addr));
81 }
82
83 // iteration
84 inline bool iterate(BitMapClosure* cl, MemRegion mr);
85
86 // Return the address corresponding to the next marked bit at or after
87 // "addr", and before "limit", if "limit" is non-NULL. If there is no
88 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
89 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
90 const HeapWord* limit = NULL) const;
91
92 // conversion utilities
93 HeapWord* offsetToHeapWord(size_t offset) const {
94 return _bmStartWord + (offset << _shifter);
95 }
96 size_t heapWordToOffset(const HeapWord* addr) const {
97 return pointer_delta(addr, _bmStartWord) >> _shifter;
98 }
99
100 // The argument addr should be the start address of a valid object
101 inline HeapWord* nextObject(HeapWord* addr);
102
103 void print_on_error(outputStream* st, const char* prefix) const;
104
105 // debugging
106 NOT_PRODUCT(bool covers(MemRegion rs) const;)
107 };
108
109 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
110 private:
111 G1CMBitMap* _bm;
112 public:
113 G1CMBitMapMappingChangedListener() : _bm(NULL) {}
114
115 void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
116
117 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
118 };
119
120 class G1CMBitMap : public G1CMBitMapRO {
121 private:
122 G1CMBitMapMappingChangedListener _listener;
123
124 public:
125 static size_t compute_size(size_t heap_size);
126 // Returns the amount of bytes on the heap between two marks in the bitmap.
127 static size_t mark_distance();
128 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
129 // mark bitmap corresponds to. This is the same as the mark distance above.
130 static size_t heap_map_factor() {
131 return mark_distance();
132 }
133
134 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
135
136 // Initializes the underlying BitMap to cover the given area.
137 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
138
139 // Write marks.
140 inline void mark(HeapWord* addr);
141 inline void clear(HeapWord* addr);
142 inline bool parMark(HeapWord* addr);
143
144 void clear_range(MemRegion mr);
145 };
146
147 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
148 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
149 VirtualSpace _virtual_space; // Underlying backing store for actual stack
150 G1ConcurrentMark* _cm;
151 oop* _base; // bottom of stack
152 jint _index; // one more than last occupied index
153 jint _capacity; // max #elements
154 jint _saved_index; // value of _index saved at start of GC
155
156 bool _overflow;
157 bool _should_expand;
158
159 public:
160 G1CMMarkStack(G1ConcurrentMark* cm);
161 ~G1CMMarkStack();
162
163 bool allocate(size_t capacity);
164
165 // Pushes the first "n" elements of "ptr_arr" on the stack.
166 // Locking impl: concurrency is allowed only with
167 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
168 // locking strategy.
169 void par_push_arr(oop* ptr_arr, int n);
170
171 // If returns false, the array was empty. Otherwise, removes up to "max"
172 // elements from the stack, and transfers them to "ptr_arr" in an
173 // unspecified order. The actual number transferred is given in "n" ("n
174 // == 0" is deliberately redundant with the return value.) Locking impl:
175 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
176 // operations, which use the same locking strategy.
177 bool par_pop_arr(oop* ptr_arr, int max, int* n);
178
179 bool isEmpty() { return _index == 0; }
180 int maxElems() { return _capacity; }
181
182 bool overflow() { return _overflow; }
183 void clear_overflow() { _overflow = false; }
184
185 bool should_expand() const { return _should_expand; }
186 void set_should_expand();
187
188 // Expand the stack, typically in response to an overflow condition
189 void expand();
190
191 int size() { return _index; }
192
193 void setEmpty() { _index = 0; clear_overflow(); }
194
195 // Record the current index.
196 void note_start_of_gc();
197
198 // Make sure that we have not added any entries to the stack during GC.
199 void note_end_of_gc();
200
201 // Apply fn to each oop in the mark stack, up to the bound recorded
202 // via one of the above "note" functions. The mark stack must not
203 // be modified while iterating.
204 template<typename Fn> void iterate(Fn fn);
205 };
206
207 class YoungList;
208
209 // Root Regions are regions that are not empty at the beginning of a
210 // marking cycle and which we might collect during an evacuation pause
211 // while the cycle is active. Given that, during evacuation pauses, we
212 // do not copy objects that are explicitly marked, what we have to do
213 // for the root regions is to scan them and mark all objects reachable
214 // from them. According to the SATB assumptions, we only need to visit
215 // each object once during marking. So, as long as we finish this scan
216 // before the next evacuation pause, we can copy the objects from the
217 // root regions without having to mark them or do anything else to them.
218 //
219 // Currently, we only support root region scanning once (at the start
220 // of the marking cycle) and the root regions are all the survivor
221 // regions populated during the initial-mark pause.
222 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
223 private:
224 YoungList* _young_list;
225 G1ConcurrentMark* _cm;
226
227 volatile bool _scan_in_progress;
228 volatile bool _should_abort;
229 HeapRegion* volatile _next_survivor;
230
231 void notify_scan_done();
232
233 public:
234 G1CMRootRegions();
235 // We actually do most of the initialization in this method.
236 void init(G1CollectedHeap* g1h, G1ConcurrentMark* cm);
237
238 // Reset the claiming / scanning of the root regions.
239 void prepare_for_scan();
240
241 // Forces get_next() to return NULL so that the iteration aborts early.
242 void abort() { _should_abort = true; }
243
244 // Return true if the CM thread are actively scanning root regions,
245 // false otherwise.
246 bool scan_in_progress() { return _scan_in_progress; }
247
248 // Claim the next root region to scan atomically, or return NULL if
249 // all have been claimed.
250 HeapRegion* claim_next();
251
252 void cancel_scan();
253
254 // Flag that we're done with root region scanning and notify anyone
255 // who's waiting on it. If aborted is false, assume that all regions
256 // have been claimed.
257 void scan_finished();
258
259 // If CM threads are still scanning root regions, wait until they
260 // are done. Return true if we had to wait, false otherwise.
261 bool wait_until_scan_finished();
262 };
263
264 class ConcurrentMarkThread;
265
266 class G1ConcurrentMark: public CHeapObj<mtGC> {
267 friend class ConcurrentMarkThread;
268 friend class G1ParNoteEndTask;
269 friend class G1VerifyLiveDataClosure;
270 friend class G1CMRefProcTaskProxy;
271 friend class G1CMRefProcTaskExecutor;
272 friend class G1CMKeepAliveAndDrainClosure;
273 friend class G1CMDrainMarkingStackClosure;
274 friend class G1CMBitMapClosure;
275 friend class G1CMConcurrentMarkingTask;
276 friend class G1CMMarkStack;
277 friend class G1CMRemarkTask;
278 friend class G1CMTask;
279
280 protected:
281 ConcurrentMarkThread* _cmThread; // The thread doing the work
282 G1CollectedHeap* _g1h; // The heap
283 uint _parallel_marking_threads; // The number of marking
284 // threads we're using
285 uint _max_parallel_marking_threads; // Max number of marking
286 // threads we'll ever use
287 double _sleep_factor; // How much we have to sleep, with
288 // respect to the work we just did, to
289 // meet the marking overhead goal
290 double _marking_task_overhead; // Marking target overhead for
291 // a single task
292
293 FreeRegionList _cleanup_list;
294
295 // Concurrent marking support structures
296 G1CMBitMap _markBitMap1;
297 G1CMBitMap _markBitMap2;
298 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
299 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
300
301 // Heap bounds
302 HeapWord* _heap_start;
303 HeapWord* _heap_end;
304
305 // Root region tracking and claiming
306 G1CMRootRegions _root_regions;
307
308 // For gray objects
309 G1CMMarkStack _markStack; // Grey objects behind global finger
310 HeapWord* volatile _finger; // The global finger, region aligned,
311 // always points to the end of the
312 // last claimed region
313
314 // Marking tasks
315 uint _max_worker_id;// Maximum worker id
316 uint _active_tasks; // Task num currently active
317 G1CMTask** _tasks; // Task queue array (max_worker_id len)
318 G1CMTaskQueueSet* _task_queues; // Task queue set
319 ParallelTaskTerminator _terminator; // For termination
320
321 // Two sync barriers that are used to synchronize tasks when an
322 // overflow occurs. The algorithm is the following. All tasks enter
323 // the first one to ensure that they have all stopped manipulating
324 // the global data structures. After they exit it, they re-initialize
325 // their data structures and task 0 re-initializes the global data
326 // structures. Then, they enter the second sync barrier. This
327 // ensure, that no task starts doing work before all data
328 // structures (local and global) have been re-initialized. When they
329 // exit it, they are free to start working again.
330 WorkGangBarrierSync _first_overflow_barrier_sync;
331 WorkGangBarrierSync _second_overflow_barrier_sync;
332
333 // This is set by any task, when an overflow on the global data
334 // structures is detected
335 volatile bool _has_overflown;
336 // True: marking is concurrent, false: we're in remark
337 volatile bool _concurrent;
338 // Set at the end of a Full GC so that marking aborts
339 volatile bool _has_aborted;
340
341 // Used when remark aborts due to an overflow to indicate that
342 // another concurrent marking phase should start
343 volatile bool _restart_for_overflow;
344
345 // This is true from the very start of concurrent marking until the
346 // point when all the tasks complete their work. It is really used
347 // to determine the points between the end of concurrent marking and
348 // time of remark.
349 volatile bool _concurrent_marking_in_progress;
350
351 ConcurrentGCTimer* _gc_timer_cm;
352
353 G1OldTracer* _gc_tracer_cm;
354
355 // All of these times are in ms
356 NumberSeq _remark_times;
357 NumberSeq _remark_mark_times;
358 NumberSeq _remark_weak_ref_times;
359 NumberSeq _cleanup_times;
360 double _total_counting_time;
361 double _total_rs_scrub_time;
362
363 double* _accum_task_vtime; // Accumulated task vtime
364
365 WorkGang* _parallel_workers;
366
367 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
368 void weakRefsWork(bool clear_all_soft_refs);
369
370 void swapMarkBitMaps();
371
372 // It resets the global marking data structures, as well as the
373 // task local ones; should be called during initial mark.
374 void reset();
375
376 // Resets all the marking data structures. Called when we have to restart
377 // marking or when marking completes (via set_non_marking_state below).
378 void reset_marking_state(bool clear_overflow = true);
379
380 // We do this after we're done with marking so that the marking data
381 // structures are initialized to a sensible and predictable state.
382 void set_non_marking_state();
383
384 // Called to indicate how many threads are currently active.
385 void set_concurrency(uint active_tasks);
386
387 // It should be called to indicate which phase we're in (concurrent
388 // mark or remark) and how many threads are currently active.
389 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
390
391 // Prints all gathered CM-related statistics
392 void print_stats();
393
394 bool cleanup_list_is_empty() {
395 return _cleanup_list.is_empty();
396 }
397
398 // Accessor methods
399 uint parallel_marking_threads() const { return _parallel_marking_threads; }
400 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
401 double sleep_factor() { return _sleep_factor; }
402 double marking_task_overhead() { return _marking_task_overhead;}
403
404 HeapWord* finger() { return _finger; }
405 bool concurrent() { return _concurrent; }
406 uint active_tasks() { return _active_tasks; }
407 ParallelTaskTerminator* terminator() { return &_terminator; }
408
409 // It claims the next available region to be scanned by a marking
410 // task/thread. It might return NULL if the next region is empty or
411 // we have run out of regions. In the latter case, out_of_regions()
412 // determines whether we've really run out of regions or the task
413 // should call claim_region() again. This might seem a bit
414 // awkward. Originally, the code was written so that claim_region()
415 // either successfully returned with a non-empty region or there
416 // were no more regions to be claimed. The problem with this was
417 // that, in certain circumstances, it iterated over large chunks of
418 // the heap finding only empty regions and, while it was working, it
419 // was preventing the calling task to call its regular clock
420 // method. So, this way, each task will spend very little time in
421 // claim_region() and is allowed to call the regular clock method
422 // frequently.
423 HeapRegion* claim_region(uint worker_id);
424
425 // It determines whether we've run out of regions to scan. Note that
426 // the finger can point past the heap end in case the heap was expanded
427 // to satisfy an allocation without doing a GC. This is fine, because all
428 // objects in those regions will be considered live anyway because of
429 // SATB guarantees (i.e. their TAMS will be equal to bottom).
430 bool out_of_regions() { return _finger >= _heap_end; }
431
432 // Returns the task with the given id
433 G1CMTask* task(int id) {
434 assert(0 <= id && id < (int) _active_tasks,
435 "task id not within active bounds");
436 return _tasks[id];
437 }
438
439 // Returns the task queue with the given id
440 G1CMTaskQueue* task_queue(int id) {
441 assert(0 <= id && id < (int) _active_tasks,
442 "task queue id not within active bounds");
443 return (G1CMTaskQueue*) _task_queues->queue(id);
444 }
445
446 // Returns the task queue set
447 G1CMTaskQueueSet* task_queues() { return _task_queues; }
448
449 // Access / manipulation of the overflow flag which is set to
450 // indicate that the global stack has overflown
451 bool has_overflown() { return _has_overflown; }
452 void set_has_overflown() { _has_overflown = true; }
453 void clear_has_overflown() { _has_overflown = false; }
454 bool restart_for_overflow() { return _restart_for_overflow; }
455
456 // Methods to enter the two overflow sync barriers
457 void enter_first_sync_barrier(uint worker_id);
458 void enter_second_sync_barrier(uint worker_id);
459
460 // Card index of the bottom of the G1 heap. Used for biasing indices into
461 // the card bitmaps.
462 intptr_t _heap_bottom_card_num;
463
464 // Set to true when initialization is complete
465 bool _completed_initialization;
466
467 // end_timer, true to end gc timer after ending concurrent phase.
468 void register_concurrent_phase_end_common(bool end_timer);
469
470 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
471 // true, periodically insert checks to see if this method should exit prematurely.
472 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
473 public:
474 // Manipulation of the global mark stack.
475 // The push and pop operations are used by tasks for transfers
476 // between task-local queues and the global mark stack, and use
477 // locking for concurrency safety.
478 bool mark_stack_push(oop* arr, int n) {
479 _markStack.par_push_arr(arr, n);
480 if (_markStack.overflow()) {
481 set_has_overflown();
482 return false;
483 }
484 return true;
485 }
486 void mark_stack_pop(oop* arr, int max, int* n) {
487 _markStack.par_pop_arr(arr, max, n);
488 }
489 size_t mark_stack_size() { return _markStack.size(); }
490 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
491 bool mark_stack_overflow() { return _markStack.overflow(); }
492 bool mark_stack_empty() { return _markStack.isEmpty(); }
493
494 G1CMRootRegions* root_regions() { return &_root_regions; }
495
496 bool concurrent_marking_in_progress() {
497 return _concurrent_marking_in_progress;
498 }
499 void set_concurrent_marking_in_progress() {
500 _concurrent_marking_in_progress = true;
501 }
502 void clear_concurrent_marking_in_progress() {
503 _concurrent_marking_in_progress = false;
504 }
505
506 void concurrent_cycle_start();
507 void concurrent_cycle_end();
508
509 void update_accum_task_vtime(int i, double vtime) {
510 _accum_task_vtime[i] += vtime;
511 }
512
513 double all_task_accum_vtime() {
514 double ret = 0.0;
515 for (uint i = 0; i < _max_worker_id; ++i)
516 ret += _accum_task_vtime[i];
517 return ret;
518 }
519
520 // Attempts to steal an object from the task queues of other tasks
521 bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
522
523 G1ConcurrentMark(G1CollectedHeap* g1h,
524 G1RegionToSpaceMapper* prev_bitmap_storage,
525 G1RegionToSpaceMapper* next_bitmap_storage);
526 ~G1ConcurrentMark();
527
528 ConcurrentMarkThread* cmThread() { return _cmThread; }
529
530 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
531 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
532
533 // Returns the number of GC threads to be used in a concurrent
534 // phase based on the number of GC threads being used in a STW
535 // phase.
536 uint scale_parallel_threads(uint n_par_threads);
537
538 // Calculates the number of GC threads to be used in a concurrent phase.
539 uint calc_parallel_marking_threads();
540
541 // The following three are interaction between CM and
542 // G1CollectedHeap
543
544 // This notifies CM that a root during initial-mark needs to be
545 // grayed. It is MT-safe. hr is the region that
546 // contains the object and it's passed optionally from callers who
547 // might already have it (no point in recalculating it).
548 inline void grayRoot(oop obj,
549 HeapRegion* hr = NULL);
550
551 // Prepare internal data structures for the next mark cycle. This includes clearing
552 // the next mark bitmap and some internal data structures. This method is intended
553 // to be called concurrently to the mutator. It will yield to safepoint requests.
554 void cleanup_for_next_mark();
555
556 // Clear the previous marking bitmap during safepoint.
557 void clear_prev_bitmap(WorkGang* workers);
558
559 // Return whether the next mark bitmap has no marks set. To be used for assertions
560 // only. Will not yield to pause requests.
561 bool nextMarkBitmapIsClear();
562
563 // These two do the work that needs to be done before and after the
564 // initial root checkpoint. Since this checkpoint can be done at two
565 // different points (i.e. an explicit pause or piggy-backed on a
566 // young collection), then it's nice to be able to easily share the
567 // pre/post code. It might be the case that we can put everything in
568 // the post method. TP
569 void checkpointRootsInitialPre();
570 void checkpointRootsInitialPost();
571
572 // Scan all the root regions and mark everything reachable from
573 // them.
574 void scan_root_regions();
575
576 // Scan a single root region and mark everything reachable from it.
577 void scanRootRegion(HeapRegion* hr);
578
579 // Do concurrent phase of marking, to a tentative transitive closure.
580 void mark_from_roots();
581
582 void checkpointRootsFinal(bool clear_all_soft_refs);
583 void checkpointRootsFinalWork();
584 void cleanup();
585 void complete_cleanup();
586
587 // Mark in the previous bitmap. NB: this is usually read-only, so use
588 // this carefully!
589 inline void markPrev(oop p);
590
591 // Clears marks for all objects in the given range, for the prev or
592 // next bitmaps. NB: the previous bitmap is usually
593 // read-only, so use this carefully!
594 void clearRangePrevBitmap(MemRegion mr);
595
596 // Notify data structures that a GC has started.
597 void note_start_of_gc() {
598 _markStack.note_start_of_gc();
599 }
600
601 // Notify data structures that a GC is finished.
602 void note_end_of_gc() {
603 _markStack.note_end_of_gc();
604 }
605
606 // Verify that there are no CSet oops on the stacks (taskqueues /
607 // global mark stack) and fingers (global / per-task).
608 // If marking is not in progress, it's a no-op.
609 void verify_no_cset_oops() PRODUCT_RETURN;
610
611 inline bool isPrevMarked(oop p) const;
612
613 inline bool do_yield_check();
614
615 // Abandon current marking iteration due to a Full GC.
616 void abort();
617
618 bool has_aborted() { return _has_aborted; }
619
620 void print_summary_info();
621
622 void print_worker_threads_on(outputStream* st) const;
623
624 void print_on_error(outputStream* st) const;
625
626 // Attempts to mark the given object on the next mark bitmap.
627 inline bool par_mark(oop obj);
628
629 // Returns true if initialization was successfully completed.
630 bool completed_initialization() const {
631 return _completed_initialization;
632 }
633
634 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
635 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
636
637 private:
638 // Clear (Reset) all liveness count data.
639 void clear_live_data(WorkGang* workers);
640
641 #ifdef ASSERT
642 // Verify all of the above data structures that they are in initial state.
643 void verify_live_data_clear();
644 #endif
645
646 // Aggregates the per-card liveness data based on the current marking. Also sets
647 // the amount of marked bytes for each region.
648 void create_live_data();
649
650 void finalize_live_data();
651
652 void verify_live_data();
653 };
654
655 // A class representing a marking task.
656 class G1CMTask : public TerminatorTerminator {
657 private:
658 enum PrivateConstants {
659 // the regular clock call is called once the scanned words reaches
660 // this limit
661 words_scanned_period = 12*1024,
662 // the regular clock call is called once the number of visited
663 // references reaches this limit
664 refs_reached_period = 384,
665 // initial value for the hash seed, used in the work stealing code
666 init_hash_seed = 17,
667 // how many entries will be transferred between global stack and
668 // local queues
669 global_stack_transfer_size = 16
670 };
671
672 uint _worker_id;
673 G1CollectedHeap* _g1h;
674 G1ConcurrentMark* _cm;
675 G1CMBitMap* _nextMarkBitMap;
676 // the task queue of this task
677 G1CMTaskQueue* _task_queue;
678 private:
679 // the task queue set---needed for stealing
680 G1CMTaskQueueSet* _task_queues;
681 // indicates whether the task has been claimed---this is only for
682 // debugging purposes
683 bool _claimed;
684
685 // number of calls to this task
686 int _calls;
687
688 // when the virtual timer reaches this time, the marking step should
689 // exit
690 double _time_target_ms;
691 // the start time of the current marking step
692 double _start_time_ms;
693
694 // the oop closure used for iterations over oops
695 G1CMOopClosure* _cm_oop_closure;
696
697 // the region this task is scanning, NULL if we're not scanning any
698 HeapRegion* _curr_region;
699 // the local finger of this task, NULL if we're not scanning a region
700 HeapWord* _finger;
701 // limit of the region this task is scanning, NULL if we're not scanning one
702 HeapWord* _region_limit;
703
704 // the number of words this task has scanned
705 size_t _words_scanned;
706 // When _words_scanned reaches this limit, the regular clock is
707 // called. Notice that this might be decreased under certain
708 // circumstances (i.e. when we believe that we did an expensive
709 // operation).
710 size_t _words_scanned_limit;
711 // the initial value of _words_scanned_limit (i.e. what it was
712 // before it was decreased).
713 size_t _real_words_scanned_limit;
714
715 // the number of references this task has visited
716 size_t _refs_reached;
717 // When _refs_reached reaches this limit, the regular clock is
718 // called. Notice this this might be decreased under certain
719 // circumstances (i.e. when we believe that we did an expensive
720 // operation).
721 size_t _refs_reached_limit;
722 // the initial value of _refs_reached_limit (i.e. what it was before
723 // it was decreased).
724 size_t _real_refs_reached_limit;
725
726 // used by the work stealing stuff
727 int _hash_seed;
728 // if this is true, then the task has aborted for some reason
729 bool _has_aborted;
730 // set when the task aborts because it has met its time quota
731 bool _has_timed_out;
732 // true when we're draining SATB buffers; this avoids the task
733 // aborting due to SATB buffers being available (as we're already
734 // dealing with them)
735 bool _draining_satb_buffers;
736
737 // number sequence of past step times
738 NumberSeq _step_times_ms;
739 // elapsed time of this task
740 double _elapsed_time_ms;
741 // termination time of this task
742 double _termination_time_ms;
743 // when this task got into the termination protocol
744 double _termination_start_time_ms;
745
746 // true when the task is during a concurrent phase, false when it is
747 // in the remark phase (so, in the latter case, we do not have to
748 // check all the things that we have to check during the concurrent
749 // phase, i.e. SATB buffer availability...)
750 bool _concurrent;
751
752 TruncatedSeq _marking_step_diffs_ms;
753
754 // it updates the local fields after this task has claimed
755 // a new region to scan
756 void setup_for_region(HeapRegion* hr);
757 // it brings up-to-date the limit of the region
758 void update_region_limit();
759
760 // called when either the words scanned or the refs visited limit
761 // has been reached
762 void reached_limit();
763 // recalculates the words scanned and refs visited limits
764 void recalculate_limits();
765 // decreases the words scanned and refs visited limits when we reach
766 // an expensive operation
767 void decrease_limits();
768 // it checks whether the words scanned or refs visited reached their
769 // respective limit and calls reached_limit() if they have
770 void check_limits() {
771 if (_words_scanned >= _words_scanned_limit ||
772 _refs_reached >= _refs_reached_limit) {
773 reached_limit();
774 }
775 }
776 // this is supposed to be called regularly during a marking step as
777 // it checks a bunch of conditions that might cause the marking step
778 // to abort
779 void regular_clock_call();
780 bool concurrent() { return _concurrent; }
781
782 // Test whether obj might have already been passed over by the
783 // mark bitmap scan, and so needs to be pushed onto the mark stack.
784 bool is_below_finger(oop obj, HeapWord* global_finger) const;
785
786 template<bool scan> void process_grey_object(oop obj);
787
788 public:
789 // It resets the task; it should be called right at the beginning of
790 // a marking phase.
791 void reset(G1CMBitMap* _nextMarkBitMap);
792 // it clears all the fields that correspond to a claimed region.
793 void clear_region_fields();
794
795 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
796
797 // The main method of this class which performs a marking step
798 // trying not to exceed the given duration. However, it might exit
799 // prematurely, according to some conditions (i.e. SATB buffers are
800 // available for processing).
801 void do_marking_step(double target_ms,
802 bool do_termination,
803 bool is_serial);
804
805 // These two calls start and stop the timer
806 void record_start_time() {
807 _elapsed_time_ms = os::elapsedTime() * 1000.0;
808 }
809 void record_end_time() {
810 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
811 }
812
813 // returns the worker ID associated with this task.
814 uint worker_id() { return _worker_id; }
815
816 // From TerminatorTerminator. It determines whether this task should
817 // exit the termination protocol after it's entered it.
818 virtual bool should_exit_termination();
819
820 // Resets the local region fields after a task has finished scanning a
821 // region; or when they have become stale as a result of the region
822 // being evacuated.
823 void giveup_current_region();
824
825 HeapWord* finger() { return _finger; }
826
827 bool has_aborted() { return _has_aborted; }
828 void set_has_aborted() { _has_aborted = true; }
829 void clear_has_aborted() { _has_aborted = false; }
830 bool has_timed_out() { return _has_timed_out; }
831 bool claimed() { return _claimed; }
832
833 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
834
835 // Increment the number of references this task has visited.
836 void increment_refs_reached() { ++_refs_reached; }
837
838 // Grey the object by marking it. If not already marked, push it on
839 // the local queue if below the finger.
840 // obj is below its region's NTAMS.
841 inline void make_reference_grey(oop obj);
842
843 // Grey the object (by calling make_grey_reference) if required,
844 // e.g. obj is below its containing region's NTAMS.
845 // Precondition: obj is a valid heap object.
846 inline void deal_with_reference(oop obj);
847
848 // It scans an object and visits its children.
849 inline void scan_object(oop obj);
850
851 // It pushes an object on the local queue.
852 inline void push(oop obj);
853
854 // These two move entries to/from the global stack.
855 void move_entries_to_global_stack();
856 void get_entries_from_global_stack();
857
858 // It pops and scans objects from the local queue. If partially is
859 // true, then it stops when the queue size is of a given limit. If
860 // partially is false, then it stops when the queue is empty.
861 void drain_local_queue(bool partially);
862 // It moves entries from the global stack to the local queue and
863 // drains the local queue. If partially is true, then it stops when
864 // both the global stack and the local queue reach a given size. If
865 // partially if false, it tries to empty them totally.
866 void drain_global_stack(bool partially);
867 // It keeps picking SATB buffers and processing them until no SATB
868 // buffers are available.
869 void drain_satb_buffers();
870
871 // moves the local finger to a new location
872 inline void move_finger_to(HeapWord* new_finger) {
873 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
874 _finger = new_finger;
875 }
876
877 G1CMTask(uint worker_id,
878 G1ConcurrentMark *cm,
879 G1CMTaskQueue* task_queue,
880 G1CMTaskQueueSet* task_queues);
881
882 // it prints statistics associated with this task
883 void print_stats();
884 };
885
886 // Class that's used to to print out per-region liveness
887 // information. It's currently used at the end of marking and also
888 // after we sort the old regions at the end of the cleanup operation.
889 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
890 private:
891 // Accumulators for these values.
892 size_t _total_used_bytes;
893 size_t _total_capacity_bytes;
894 size_t _total_prev_live_bytes;
895 size_t _total_next_live_bytes;
896
897 // Accumulator for the remembered set size
898 size_t _total_remset_bytes;
899
900 // Accumulator for strong code roots memory size
901 size_t _total_strong_code_roots_bytes;
902
903 static double perc(size_t val, size_t total) {
904 if (total == 0) {
905 return 0.0;
906 } else {
907 return 100.0 * ((double) val / (double) total);
908 }
909 }
910
911 static double bytes_to_mb(size_t val) {
912 return (double) val / (double) M;
913 }
914
915 public:
916 // The header and footer are printed in the constructor and
917 // destructor respectively.
918 G1PrintRegionLivenessInfoClosure(const char* phase_name);
919 virtual bool doHeapRegion(HeapRegion* r);
920 ~G1PrintRegionLivenessInfoClosure();
921 };
922
923 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP