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 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue;
38 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
39
40 // Closure used by CM during concurrent reference discovery
41 // and reference processing (during remarking) to determine
42 // if a particular object is alive. It is primarily used
43 // to determine if referents of discovered reference objects
44 // are alive. An instance is also embedded into the
45 // reference processor as the _is_alive_non_header field
46 class G1CMIsAliveClosure: public BoolObjectClosure {
47 G1CollectedHeap* _g1;
48 public:
49 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
50
51 bool do_object_b(oop obj);
52 };
53
54 // A generic CM bit map. This is essentially a wrapper around the BitMap
55 // class, with one bit per (1<<_shifter) HeapWords.
56
57 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
58 protected:
59 HeapWord* _bmStartWord; // base address of range covered by map
60 size_t _bmWordSize; // map size (in #HeapWords covered)
61 const int _shifter; // map to char or bit
62 BitMap _bm; // the bit map itself
63
64 public:
65 // constructor
66 G1CMBitMapRO(int shifter);
67
68 // inquiries
69 HeapWord* startWord() const { return _bmStartWord; }
70 // the following is one past the last word in space
71 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
72
73 // read marks
74
75 bool isMarked(HeapWord* addr) const {
76 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
77 "outside underlying space?");
78 return _bm.at(heapWordToOffset(addr));
79 }
80
81 // iteration
82 inline bool iterate(BitMapClosure* cl, MemRegion mr);
83
84 // Return the address corresponding to the next marked bit at or after
85 // "addr", and before "limit", if "limit" is non-NULL. If there is no
86 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
87 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
88 const HeapWord* limit = NULL) const;
89
90 // conversion utilities
91 HeapWord* offsetToHeapWord(size_t offset) const {
92 return _bmStartWord + (offset << _shifter);
93 }
94 size_t heapWordToOffset(const HeapWord* addr) const {
95 return pointer_delta(addr, _bmStartWord) >> _shifter;
96 }
97
98 // The argument addr should be the start address of a valid object
99 inline HeapWord* nextObject(HeapWord* addr);
100
101 void print_on_error(outputStream* st, const char* prefix) const;
102
103 // debugging
104 NOT_PRODUCT(bool covers(MemRegion rs) const;)
105 };
106
107 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
108 private:
109 G1CMBitMap* _bm;
110 public:
111 G1CMBitMapMappingChangedListener() : _bm(NULL) {}
112
113 void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
114
115 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
116 };
117
118 class G1CMBitMap : public G1CMBitMapRO {
119 private:
120 G1CMBitMapMappingChangedListener _listener;
121
122 public:
123 static size_t compute_size(size_t heap_size);
124 // Returns the amount of bytes on the heap between two marks in the bitmap.
125 static size_t mark_distance();
126 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
127 // mark bitmap corresponds to. This is the same as the mark distance above.
128 static size_t heap_map_factor() {
129 return mark_distance();
130 }
131
132 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
133
134 // Initializes the underlying BitMap to cover the given area.
135 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
136
137 // Write marks.
138 inline void mark(HeapWord* addr);
139 inline void clear(HeapWord* addr);
140 inline bool parMark(HeapWord* addr);
141
142 void clearRange(MemRegion mr);
143
144 // Clear the whole mark bitmap.
145 void clearAll();
146 };
147
148 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
149 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
150 VirtualSpace _virtual_space; // Underlying backing store for actual stack
151 G1ConcurrentMark* _cm;
152 oop* _base; // bottom of stack
153 jint _index; // one more than last occupied index
154 jint _capacity; // max #elements
155 jint _saved_index; // value of _index saved at start of GC
156
157 bool _overflow;
158 bool _should_expand;
159
160 public:
161 G1CMMarkStack(G1ConcurrentMark* cm);
162 ~G1CMMarkStack();
163
164 bool allocate(size_t capacity);
165
166 // Pushes the first "n" elements of "ptr_arr" on the stack.
167 // Locking impl: concurrency is allowed only with
168 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
169 // locking strategy.
170 void par_push_arr(oop* ptr_arr, int n);
171
172 // If returns false, the array was empty. Otherwise, removes up to "max"
173 // elements from the stack, and transfers them to "ptr_arr" in an
174 // unspecified order. The actual number transferred is given in "n" ("n
175 // == 0" is deliberately redundant with the return value.) Locking impl:
176 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
177 // operations, which use the same locking strategy.
178 bool par_pop_arr(oop* ptr_arr, int max, int* n);
179
180 bool isEmpty() { return _index == 0; }
181 int maxElems() { return _capacity; }
182
183 bool overflow() { return _overflow; }
184 void clear_overflow() { _overflow = false; }
185
186 bool should_expand() const { return _should_expand; }
187 void set_should_expand();
188
189 // Expand the stack, typically in response to an overflow condition
190 void expand();
191
192 int size() { return _index; }
193
194 void setEmpty() { _index = 0; clear_overflow(); }
195
196 // Record the current index.
197 void note_start_of_gc();
198
199 // Make sure that we have not added any entries to the stack during GC.
200 void note_end_of_gc();
201
202 // Apply fn to each oop in the mark stack, up to the bound recorded
203 // via one of the above "note" functions. The mark stack must not
204 // be modified while iterating.
205 template<typename Fn> void iterate(Fn fn);
206 };
207
208 class YoungList;
209
210 // Root Regions are regions that are not empty at the beginning of a
211 // marking cycle and which we might collect during an evacuation pause
212 // while the cycle is active. Given that, during evacuation pauses, we
213 // do not copy objects that are explicitly marked, what we have to do
214 // for the root regions is to scan them and mark all objects reachable
215 // from them. According to the SATB assumptions, we only need to visit
216 // each object once during marking. So, as long as we finish this scan
217 // before the next evacuation pause, we can copy the objects from the
218 // root regions without having to mark them or do anything else to them.
219 //
220 // Currently, we only support root region scanning once (at the start
221 // of the marking cycle) and the root regions are all the survivor
222 // regions populated during the initial-mark pause.
223 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
224 private:
225 YoungList* _young_list;
226 G1ConcurrentMark* _cm;
227
228 volatile bool _scan_in_progress;
229 volatile bool _should_abort;
230 HeapRegion* volatile _next_survivor;
231
232 void notify_scan_done();
233
234 public:
235 G1CMRootRegions();
236 // We actually do most of the initialization in this method.
237 void init(G1CollectedHeap* g1h, G1ConcurrentMark* cm);
238
239 // Reset the claiming / scanning of the root regions.
240 void prepare_for_scan();
241
242 // Forces get_next() to return NULL so that the iteration aborts early.
243 void abort() { _should_abort = true; }
244
245 // Return true if the CM thread are actively scanning root regions,
246 // false otherwise.
247 bool scan_in_progress() { return _scan_in_progress; }
248
249 // Claim the next root region to scan atomically, or return NULL if
250 // all have been claimed.
251 HeapRegion* claim_next();
252
253 void cancel_scan();
254
255 // Flag that we're done with root region scanning and notify anyone
256 // who's waiting on it. If aborted is false, assume that all regions
257 // have been claimed.
258 void scan_finished();
259
260 // If CM threads are still scanning root regions, wait until they
261 // are done. Return true if we had to wait, false otherwise.
262 bool wait_until_scan_finished();
263 };
264
265 class ConcurrentMarkThread;
266
267 class G1ConcurrentMark: public CHeapObj<mtGC> {
268 friend class ConcurrentMarkThread;
269 friend class G1ParNoteEndTask;
270 friend class CalcLiveObjectsClosure;
271 friend class G1CMRefProcTaskProxy;
272 friend class G1CMRefProcTaskExecutor;
273 friend class G1CMKeepAliveAndDrainClosure;
274 friend class G1CMDrainMarkingStackClosure;
275 friend class G1CMBitMapClosure;
276 friend class G1CMConcurrentMarkingTask;
277 friend class G1CMMarkStack;
278 friend class G1CMRemarkTask;
279 friend class G1CMTask;
280
281 protected:
282 ConcurrentMarkThread* _cmThread; // The thread doing the work
283 G1CollectedHeap* _g1h; // The heap
284 uint _parallel_marking_threads; // The number of marking
285 // threads we're using
286 uint _max_parallel_marking_threads; // Max number of marking
287 // threads we'll ever use
288 double _sleep_factor; // How much we have to sleep, with
289 // respect to the work we just did, to
290 // meet the marking overhead goal
291 double _marking_task_overhead; // Marking target overhead for
292 // a single task
293
294 FreeRegionList _cleanup_list;
295
296 // Concurrent marking support structures
297 G1CMBitMap _markBitMap1;
298 G1CMBitMap _markBitMap2;
299 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
300 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
301
302 BitMap _region_bm;
303 BitMap _card_bm;
304
305 // Heap bounds
306 HeapWord* _heap_start;
307 HeapWord* _heap_end;
308
309 // Root region tracking and claiming
310 G1CMRootRegions _root_regions;
311
312 // For gray objects
313 G1CMMarkStack _markStack; // Grey objects behind global finger
314 HeapWord* volatile _finger; // The global finger, region aligned,
315 // always points to the end of the
316 // last claimed region
317
318 // Marking tasks
319 uint _max_worker_id;// Maximum worker id
320 uint _active_tasks; // Task num currently active
321 G1CMTask** _tasks; // Task queue array (max_worker_id len)
322 G1CMTaskQueueSet* _task_queues; // Task queue set
323 ParallelTaskTerminator _terminator; // For termination
324
325 // Two sync barriers that are used to synchronize tasks when an
326 // overflow occurs. The algorithm is the following. All tasks enter
327 // the first one to ensure that they have all stopped manipulating
328 // the global data structures. After they exit it, they re-initialize
329 // their data structures and task 0 re-initializes the global data
330 // structures. Then, they enter the second sync barrier. This
331 // ensure, that no task starts doing work before all data
332 // structures (local and global) have been re-initialized. When they
333 // exit it, they are free to start working again.
334 WorkGangBarrierSync _first_overflow_barrier_sync;
335 WorkGangBarrierSync _second_overflow_barrier_sync;
336
337 // This is set by any task, when an overflow on the global data
338 // structures is detected
339 volatile bool _has_overflown;
340 // True: marking is concurrent, false: we're in remark
341 volatile bool _concurrent;
342 // Set at the end of a Full GC so that marking aborts
343 volatile bool _has_aborted;
344
345 // Used when remark aborts due to an overflow to indicate that
346 // another concurrent marking phase should start
347 volatile bool _restart_for_overflow;
348
349 // This is true from the very start of concurrent marking until the
350 // point when all the tasks complete their work. It is really used
351 // to determine the points between the end of concurrent marking and
352 // time of remark.
353 volatile bool _concurrent_marking_in_progress;
354
355 // There would be a race between ConcurrentMarkThread and VMThread(ConcurrentMark::abort())
356 // to call ConcurrentGCTimer::register_gc_concurrent_end().
357 // And this variable is used to keep track of concurrent phase.
358 volatile uint _concurrent_phase_status;
359 // Concurrent phase is not yet started.
360 static const uint ConcPhaseNotStarted = 0;
361 // Concurrent phase is started.
362 static const uint ConcPhaseStarted = 1;
363 // Caller thread of ConcurrentGCTimer::register_gc_concurrent_end() is ending concurrent phase.
364 // So other thread should wait until the status to be changed to ConcPhaseNotStarted.
365 static const uint ConcPhaseStopping = 2;
366
367 // All of these times are in ms
368 NumberSeq _init_times;
369 NumberSeq _remark_times;
370 NumberSeq _remark_mark_times;
371 NumberSeq _remark_weak_ref_times;
372 NumberSeq _cleanup_times;
373 double _total_counting_time;
374 double _total_rs_scrub_time;
375
376 double* _accum_task_vtime; // Accumulated task vtime
377
378 WorkGang* _parallel_workers;
379
380 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
381 void weakRefsWork(bool clear_all_soft_refs);
382
383 void swapMarkBitMaps();
384
385 // It resets the global marking data structures, as well as the
386 // task local ones; should be called during initial mark.
387 void reset();
388
389 // Resets all the marking data structures. Called when we have to restart
390 // marking or when marking completes (via set_non_marking_state below).
391 void reset_marking_state(bool clear_overflow = true);
392
393 // We do this after we're done with marking so that the marking data
394 // structures are initialized to a sensible and predictable state.
395 void set_non_marking_state();
396
397 // Called to indicate how many threads are currently active.
398 void set_concurrency(uint active_tasks);
399
400 // It should be called to indicate which phase we're in (concurrent
401 // mark or remark) and how many threads are currently active.
402 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
403
404 // Prints all gathered CM-related statistics
405 void print_stats();
406
407 bool cleanup_list_is_empty() {
408 return _cleanup_list.is_empty();
409 }
410
411 // Accessor methods
412 uint parallel_marking_threads() const { return _parallel_marking_threads; }
413 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
414 double sleep_factor() { return _sleep_factor; }
415 double marking_task_overhead() { return _marking_task_overhead;}
416
417 HeapWord* finger() { return _finger; }
418 bool concurrent() { return _concurrent; }
419 uint active_tasks() { return _active_tasks; }
420 ParallelTaskTerminator* terminator() { return &_terminator; }
421
422 // It claims the next available region to be scanned by a marking
423 // task/thread. It might return NULL if the next region is empty or
424 // we have run out of regions. In the latter case, out_of_regions()
425 // determines whether we've really run out of regions or the task
426 // should call claim_region() again. This might seem a bit
427 // awkward. Originally, the code was written so that claim_region()
428 // either successfully returned with a non-empty region or there
429 // were no more regions to be claimed. The problem with this was
430 // that, in certain circumstances, it iterated over large chunks of
431 // the heap finding only empty regions and, while it was working, it
432 // was preventing the calling task to call its regular clock
433 // method. So, this way, each task will spend very little time in
434 // claim_region() and is allowed to call the regular clock method
435 // frequently.
436 HeapRegion* claim_region(uint worker_id);
437
438 // It determines whether we've run out of regions to scan. Note that
439 // the finger can point past the heap end in case the heap was expanded
440 // to satisfy an allocation without doing a GC. This is fine, because all
441 // objects in those regions will be considered live anyway because of
442 // SATB guarantees (i.e. their TAMS will be equal to bottom).
443 bool out_of_regions() { return _finger >= _heap_end; }
444
445 // Returns the task with the given id
446 G1CMTask* task(int id) {
447 assert(0 <= id && id < (int) _active_tasks,
448 "task id not within active bounds");
449 return _tasks[id];
450 }
451
452 // Returns the task queue with the given id
453 G1CMTaskQueue* task_queue(int id) {
454 assert(0 <= id && id < (int) _active_tasks,
455 "task queue id not within active bounds");
456 return (G1CMTaskQueue*) _task_queues->queue(id);
457 }
458
459 // Returns the task queue set
460 G1CMTaskQueueSet* task_queues() { return _task_queues; }
461
462 // Access / manipulation of the overflow flag which is set to
463 // indicate that the global stack has overflown
464 bool has_overflown() { return _has_overflown; }
465 void set_has_overflown() { _has_overflown = true; }
466 void clear_has_overflown() { _has_overflown = false; }
467 bool restart_for_overflow() { return _restart_for_overflow; }
468
469 // Methods to enter the two overflow sync barriers
470 void enter_first_sync_barrier(uint worker_id);
471 void enter_second_sync_barrier(uint worker_id);
472
473 // Live Data Counting data structures...
474 // These data structures are initialized at the start of
475 // marking. They are written to while marking is active.
476 // They are aggregated during remark; the aggregated values
477 // are then used to populate the _region_bm, _card_bm, and
478 // the total live bytes, which are then subsequently updated
479 // during cleanup.
480
481 // An array of bitmaps (one bit map per task). Each bitmap
482 // is used to record the cards spanned by the live objects
483 // marked by that task/worker.
484 BitMap* _count_card_bitmaps;
485
486 // Used to record the number of marked live bytes
487 // (for each region, by worker thread).
488 size_t** _count_marked_bytes;
489
490 // Card index of the bottom of the G1 heap. Used for biasing indices into
491 // the card bitmaps.
492 intptr_t _heap_bottom_card_num;
493
494 // Set to true when initialization is complete
495 bool _completed_initialization;
496
497 // end_timer, true to end gc timer after ending concurrent phase.
498 void register_concurrent_phase_end_common(bool end_timer);
499
500 public:
501 // Manipulation of the global mark stack.
502 // The push and pop operations are used by tasks for transfers
503 // between task-local queues and the global mark stack, and use
504 // locking for concurrency safety.
505 bool mark_stack_push(oop* arr, int n) {
506 _markStack.par_push_arr(arr, n);
507 if (_markStack.overflow()) {
508 set_has_overflown();
509 return false;
510 }
511 return true;
512 }
513 void mark_stack_pop(oop* arr, int max, int* n) {
514 _markStack.par_pop_arr(arr, max, n);
515 }
516 size_t mark_stack_size() { return _markStack.size(); }
517 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
518 bool mark_stack_overflow() { return _markStack.overflow(); }
519 bool mark_stack_empty() { return _markStack.isEmpty(); }
520
521 G1CMRootRegions* root_regions() { return &_root_regions; }
522
523 bool concurrent_marking_in_progress() {
524 return _concurrent_marking_in_progress;
525 }
526 void set_concurrent_marking_in_progress() {
527 _concurrent_marking_in_progress = true;
528 }
529 void clear_concurrent_marking_in_progress() {
530 _concurrent_marking_in_progress = false;
531 }
532
533 void register_concurrent_phase_start(const char* title);
534 void register_concurrent_phase_end();
535 // Ends both concurrent phase and timer.
536 void register_concurrent_gc_end_and_stop_timer();
537
538 void update_accum_task_vtime(int i, double vtime) {
539 _accum_task_vtime[i] += vtime;
540 }
541
542 double all_task_accum_vtime() {
543 double ret = 0.0;
544 for (uint i = 0; i < _max_worker_id; ++i)
545 ret += _accum_task_vtime[i];
546 return ret;
547 }
548
549 // Attempts to steal an object from the task queues of other tasks
550 bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
551
552 G1ConcurrentMark(G1CollectedHeap* g1h,
553 G1RegionToSpaceMapper* prev_bitmap_storage,
554 G1RegionToSpaceMapper* next_bitmap_storage);
555 ~G1ConcurrentMark();
556
557 ConcurrentMarkThread* cmThread() { return _cmThread; }
558
559 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
560 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
561
562 // Returns the number of GC threads to be used in a concurrent
563 // phase based on the number of GC threads being used in a STW
564 // phase.
565 uint scale_parallel_threads(uint n_par_threads);
566
567 // Calculates the number of GC threads to be used in a concurrent phase.
568 uint calc_parallel_marking_threads();
569
570 // The following three are interaction between CM and
571 // G1CollectedHeap
572
573 // This notifies CM that a root during initial-mark needs to be
574 // grayed. It is MT-safe. word_size is the size of the object in
575 // words. It is passed explicitly as sometimes we cannot calculate
576 // it from the given object because it might be in an inconsistent
577 // state (e.g., in to-space and being copied). So the caller is
578 // responsible for dealing with this issue (e.g., get the size from
579 // the from-space image when the to-space image might be
580 // inconsistent) and always passing the size. hr is the region that
581 // contains the object and it's passed optionally from callers who
582 // might already have it (no point in recalculating it).
583 inline void grayRoot(oop obj,
584 size_t word_size,
585 uint worker_id,
586 HeapRegion* hr = NULL);
587
588 // Clear the next marking bitmap (will be called concurrently).
589 void clearNextBitmap();
590
591 // Return whether the next mark bitmap has no marks set. To be used for assertions
592 // only. Will not yield to pause requests.
593 bool nextMarkBitmapIsClear();
594
595 // These two do the work that needs to be done before and after the
596 // initial root checkpoint. Since this checkpoint can be done at two
597 // different points (i.e. an explicit pause or piggy-backed on a
598 // young collection), then it's nice to be able to easily share the
599 // pre/post code. It might be the case that we can put everything in
600 // the post method. TP
601 void checkpointRootsInitialPre();
602 void checkpointRootsInitialPost();
603
604 // Scan all the root regions and mark everything reachable from
605 // them.
606 void scanRootRegions();
607
608 // Scan a single root region and mark everything reachable from it.
609 void scanRootRegion(HeapRegion* hr, uint worker_id);
610
611 // Do concurrent phase of marking, to a tentative transitive closure.
612 void markFromRoots();
613
614 void checkpointRootsFinal(bool clear_all_soft_refs);
615 void checkpointRootsFinalWork();
616 void cleanup();
617 void completeCleanup();
618
619 // Mark in the previous bitmap. NB: this is usually read-only, so use
620 // this carefully!
621 inline void markPrev(oop p);
622
623 // Clears marks for all objects in the given range, for the prev or
624 // next bitmaps. NB: the previous bitmap is usually
625 // read-only, so use this carefully!
626 void clearRangePrevBitmap(MemRegion mr);
627
628 // Notify data structures that a GC has started.
629 void note_start_of_gc() {
630 _markStack.note_start_of_gc();
631 }
632
633 // Notify data structures that a GC is finished.
634 void note_end_of_gc() {
635 _markStack.note_end_of_gc();
636 }
637
638 // Verify that there are no CSet oops on the stacks (taskqueues /
639 // global mark stack) and fingers (global / per-task).
640 // If marking is not in progress, it's a no-op.
641 void verify_no_cset_oops() PRODUCT_RETURN;
642
643 inline bool isPrevMarked(oop p) const;
644
645 inline bool do_yield_check(uint worker_i = 0);
646
647 // Called to abort the marking cycle after a Full GC takes place.
648 void abort();
649
650 bool has_aborted() { return _has_aborted; }
651
652 void print_summary_info();
653
654 void print_worker_threads_on(outputStream* st) const;
655
656 void print_on_error(outputStream* st) const;
657
658 // Liveness counting
659
660 // Utility routine to set an exclusive range of cards on the given
661 // card liveness bitmap
662 inline void set_card_bitmap_range(BitMap* card_bm,
663 BitMap::idx_t start_idx,
664 BitMap::idx_t end_idx,
665 bool is_par);
666
667 // Returns the card number of the bottom of the G1 heap.
668 // Used in biasing indices into accounting card bitmaps.
669 intptr_t heap_bottom_card_num() const {
670 return _heap_bottom_card_num;
671 }
672
673 // Returns the card bitmap for a given task or worker id.
674 BitMap* count_card_bitmap_for(uint worker_id) {
675 assert(worker_id < _max_worker_id, "oob");
676 assert(_count_card_bitmaps != NULL, "uninitialized");
677 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
678 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
679 return task_card_bm;
680 }
681
682 // Returns the array containing the marked bytes for each region,
683 // for the given worker or task id.
684 size_t* count_marked_bytes_array_for(uint worker_id) {
685 assert(worker_id < _max_worker_id, "oob");
686 assert(_count_marked_bytes != NULL, "uninitialized");
687 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
688 assert(marked_bytes_array != NULL, "uninitialized");
689 return marked_bytes_array;
690 }
691
692 // Returns the index in the liveness accounting card table bitmap
693 // for the given address
694 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
695
696 // Counts the size of the given memory region in the the given
697 // marked_bytes array slot for the given HeapRegion.
698 // Sets the bits in the given card bitmap that are associated with the
699 // cards that are spanned by the memory region.
700 inline void count_region(MemRegion mr,
701 HeapRegion* hr,
702 size_t* marked_bytes_array,
703 BitMap* task_card_bm);
704
705 // Counts the given object in the given task/worker counting
706 // data structures.
707 inline void count_object(oop obj,
708 HeapRegion* hr,
709 size_t* marked_bytes_array,
710 BitMap* task_card_bm,
711 size_t word_size);
712
713 // Attempts to mark the given object and, if successful, counts
714 // the object in the given task/worker counting structures.
715 inline bool par_mark_and_count(oop obj,
716 HeapRegion* hr,
717 size_t* marked_bytes_array,
718 BitMap* task_card_bm);
719
720 // Attempts to mark the given object and, if successful, counts
721 // the object in the task/worker counting structures for the
722 // given worker id.
723 inline bool par_mark_and_count(oop obj,
724 size_t word_size,
725 HeapRegion* hr,
726 uint worker_id);
727
728 // Returns true if initialization was successfully completed.
729 bool completed_initialization() const {
730 return _completed_initialization;
731 }
732
733 protected:
734 // Clear all the per-task bitmaps and arrays used to store the
735 // counting data.
736 void clear_all_count_data();
737
738 // Aggregates the counting data for each worker/task
739 // that was constructed while marking. Also sets
740 // the amount of marked bytes for each region and
741 // the top at concurrent mark count.
742 void aggregate_count_data();
743
744 // Verification routine
745 void verify_count_data();
746 };
747
748 // A class representing a marking task.
749 class G1CMTask : public TerminatorTerminator {
750 private:
751 enum PrivateConstants {
752 // the regular clock call is called once the scanned words reaches
753 // this limit
754 words_scanned_period = 12*1024,
755 // the regular clock call is called once the number of visited
756 // references reaches this limit
757 refs_reached_period = 384,
758 // initial value for the hash seed, used in the work stealing code
759 init_hash_seed = 17,
760 // how many entries will be transferred between global stack and
761 // local queues
762 global_stack_transfer_size = 16
763 };
764
765 uint _worker_id;
766 G1CollectedHeap* _g1h;
767 G1ConcurrentMark* _cm;
768 G1CMBitMap* _nextMarkBitMap;
769 // the task queue of this task
770 G1CMTaskQueue* _task_queue;
771 private:
772 // the task queue set---needed for stealing
773 G1CMTaskQueueSet* _task_queues;
774 // indicates whether the task has been claimed---this is only for
775 // debugging purposes
776 bool _claimed;
777
778 // number of calls to this task
779 int _calls;
780
781 // when the virtual timer reaches this time, the marking step should
782 // exit
783 double _time_target_ms;
784 // the start time of the current marking step
785 double _start_time_ms;
786
787 // the oop closure used for iterations over oops
788 G1CMOopClosure* _cm_oop_closure;
789
790 // the region this task is scanning, NULL if we're not scanning any
791 HeapRegion* _curr_region;
792 // the local finger of this task, NULL if we're not scanning a region
793 HeapWord* _finger;
794 // limit of the region this task is scanning, NULL if we're not scanning one
795 HeapWord* _region_limit;
796
797 // the number of words this task has scanned
798 size_t _words_scanned;
799 // When _words_scanned reaches this limit, the regular clock is
800 // called. Notice that this might be decreased under certain
801 // circumstances (i.e. when we believe that we did an expensive
802 // operation).
803 size_t _words_scanned_limit;
804 // the initial value of _words_scanned_limit (i.e. what it was
805 // before it was decreased).
806 size_t _real_words_scanned_limit;
807
808 // the number of references this task has visited
809 size_t _refs_reached;
810 // When _refs_reached reaches this limit, the regular clock is
811 // called. Notice this this might be decreased under certain
812 // circumstances (i.e. when we believe that we did an expensive
813 // operation).
814 size_t _refs_reached_limit;
815 // the initial value of _refs_reached_limit (i.e. what it was before
816 // it was decreased).
817 size_t _real_refs_reached_limit;
818
819 // used by the work stealing stuff
820 int _hash_seed;
821 // if this is true, then the task has aborted for some reason
822 bool _has_aborted;
823 // set when the task aborts because it has met its time quota
824 bool _has_timed_out;
825 // true when we're draining SATB buffers; this avoids the task
826 // aborting due to SATB buffers being available (as we're already
827 // dealing with them)
828 bool _draining_satb_buffers;
829
830 // number sequence of past step times
831 NumberSeq _step_times_ms;
832 // elapsed time of this task
833 double _elapsed_time_ms;
834 // termination time of this task
835 double _termination_time_ms;
836 // when this task got into the termination protocol
837 double _termination_start_time_ms;
838
839 // true when the task is during a concurrent phase, false when it is
840 // in the remark phase (so, in the latter case, we do not have to
841 // check all the things that we have to check during the concurrent
842 // phase, i.e. SATB buffer availability...)
843 bool _concurrent;
844
845 TruncatedSeq _marking_step_diffs_ms;
846
847 // Counting data structures. Embedding the task's marked_bytes_array
848 // and card bitmap into the actual task saves having to go through
849 // the ConcurrentMark object.
850 size_t* _marked_bytes_array;
851 BitMap* _card_bm;
852
853 // it updates the local fields after this task has claimed
854 // a new region to scan
855 void setup_for_region(HeapRegion* hr);
856 // it brings up-to-date the limit of the region
857 void update_region_limit();
858
859 // called when either the words scanned or the refs visited limit
860 // has been reached
861 void reached_limit();
862 // recalculates the words scanned and refs visited limits
863 void recalculate_limits();
864 // decreases the words scanned and refs visited limits when we reach
865 // an expensive operation
866 void decrease_limits();
867 // it checks whether the words scanned or refs visited reached their
868 // respective limit and calls reached_limit() if they have
869 void check_limits() {
870 if (_words_scanned >= _words_scanned_limit ||
871 _refs_reached >= _refs_reached_limit) {
872 reached_limit();
873 }
874 }
875 // this is supposed to be called regularly during a marking step as
876 // it checks a bunch of conditions that might cause the marking step
877 // to abort
878 void regular_clock_call();
879 bool concurrent() { return _concurrent; }
880
881 // Test whether obj might have already been passed over by the
882 // mark bitmap scan, and so needs to be pushed onto the mark stack.
883 bool is_below_finger(oop obj, HeapWord* global_finger) const;
884
885 template<bool scan> void process_grey_object(oop obj);
886
887 public:
888 // It resets the task; it should be called right at the beginning of
889 // a marking phase.
890 void reset(G1CMBitMap* _nextMarkBitMap);
891 // it clears all the fields that correspond to a claimed region.
892 void clear_region_fields();
893
894 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
895
896 // The main method of this class which performs a marking step
897 // trying not to exceed the given duration. However, it might exit
898 // prematurely, according to some conditions (i.e. SATB buffers are
899 // available for processing).
900 void do_marking_step(double target_ms,
901 bool do_termination,
902 bool is_serial);
903
904 // These two calls start and stop the timer
905 void record_start_time() {
906 _elapsed_time_ms = os::elapsedTime() * 1000.0;
907 }
908 void record_end_time() {
909 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
910 }
911
912 // returns the worker ID associated with this task.
913 uint worker_id() { return _worker_id; }
914
915 // From TerminatorTerminator. It determines whether this task should
916 // exit the termination protocol after it's entered it.
917 virtual bool should_exit_termination();
918
919 // Resets the local region fields after a task has finished scanning a
920 // region; or when they have become stale as a result of the region
921 // being evacuated.
922 void giveup_current_region();
923
924 HeapWord* finger() { return _finger; }
925
926 bool has_aborted() { return _has_aborted; }
927 void set_has_aborted() { _has_aborted = true; }
928 void clear_has_aborted() { _has_aborted = false; }
929 bool has_timed_out() { return _has_timed_out; }
930 bool claimed() { return _claimed; }
931
932 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
933
934 // Increment the number of references this task has visited.
935 void increment_refs_reached() { ++_refs_reached; }
936
937 // Grey the object by marking it. If not already marked, push it on
938 // the local queue if below the finger.
939 // Precondition: obj is in region.
940 // Precondition: obj is below region's NTAMS.
941 inline void make_reference_grey(oop obj, HeapRegion* region);
942
943 // Grey the object (by calling make_grey_reference) if required,
944 // e.g. obj is below its containing region's NTAMS.
945 // Precondition: obj is a valid heap object.
946 inline void deal_with_reference(oop obj);
947
948 // It scans an object and visits its children.
949 inline void scan_object(oop obj);
950
951 // It pushes an object on the local queue.
952 inline void push(oop obj);
953
954 // These two move entries to/from the global stack.
955 void move_entries_to_global_stack();
956 void get_entries_from_global_stack();
957
958 // It pops and scans objects from the local queue. If partially is
959 // true, then it stops when the queue size is of a given limit. If
960 // partially is false, then it stops when the queue is empty.
961 void drain_local_queue(bool partially);
962 // It moves entries from the global stack to the local queue and
963 // drains the local queue. If partially is true, then it stops when
964 // both the global stack and the local queue reach a given size. If
965 // partially if false, it tries to empty them totally.
966 void drain_global_stack(bool partially);
967 // It keeps picking SATB buffers and processing them until no SATB
968 // buffers are available.
969 void drain_satb_buffers();
970
971 // moves the local finger to a new location
972 inline void move_finger_to(HeapWord* new_finger) {
973 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
974 _finger = new_finger;
975 }
976
977 G1CMTask(uint worker_id,
978 G1ConcurrentMark *cm,
979 size_t* marked_bytes,
980 BitMap* card_bm,
981 G1CMTaskQueue* task_queue,
982 G1CMTaskQueueSet* task_queues);
983
984 // it prints statistics associated with this task
985 void print_stats();
986 };
987
988 // Class that's used to to print out per-region liveness
989 // information. It's currently used at the end of marking and also
990 // after we sort the old regions at the end of the cleanup operation.
991 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
992 private:
993 // Accumulators for these values.
994 size_t _total_used_bytes;
995 size_t _total_capacity_bytes;
996 size_t _total_prev_live_bytes;
997 size_t _total_next_live_bytes;
998
999 // These are set up when we come across a "stars humongous" region
1000 // (as this is where most of this information is stored, not in the
1001 // subsequent "continues humongous" regions). After that, for every
1002 // region in a given humongous region series we deduce the right
1003 // values for it by simply subtracting the appropriate amount from
1004 // these fields. All these values should reach 0 after we've visited
1005 // the last region in the series.
1006 size_t _hum_used_bytes;
1007 size_t _hum_capacity_bytes;
1008 size_t _hum_prev_live_bytes;
1009 size_t _hum_next_live_bytes;
1010
1011 // Accumulator for the remembered set size
1012 size_t _total_remset_bytes;
1013
1014 // Accumulator for strong code roots memory size
1015 size_t _total_strong_code_roots_bytes;
1016
1017 static double perc(size_t val, size_t total) {
1018 if (total == 0) {
1019 return 0.0;
1020 } else {
1021 return 100.0 * ((double) val / (double) total);
1022 }
1023 }
1024
1025 static double bytes_to_mb(size_t val) {
1026 return (double) val / (double) M;
1027 }
1028
1029 // See the .cpp file.
1030 size_t get_hum_bytes(size_t* hum_bytes);
1031 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1032 size_t* prev_live_bytes, size_t* next_live_bytes);
1033
1034 public:
1035 // The header and footer are printed in the constructor and
1036 // destructor respectively.
1037 G1PrintRegionLivenessInfoClosure(const char* phase_name);
1038 virtual void doHeapRegion(HeapRegion* r);
1039 ~G1PrintRegionLivenessInfoClosure();
1040 };
1041
1042 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP