1 /*
2 * Copyright (c) 2001, 2015, 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_CONCURRENTMARK_HPP
26 #define SHARE_VM_GC_G1_CONCURRENTMARK_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 CMBitMap;
35 class CMTask;
36 class ConcurrentMark;
37 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue;
38 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
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 CMBitMapRO 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 CMBitMapRO(int shifter);
67
68 enum { do_yield = true };
69
70 // inquiries
71 HeapWord* startWord() const { return _bmStartWord; }
72 size_t sizeInWords() const { return _bmWordSize; }
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 inline bool iterate(BitMapClosure* cl);
87
88 // Return the address corresponding to the next marked bit at or after
89 // "addr", and before "limit", if "limit" is non-NULL. If there is no
90 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
91 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
92 const HeapWord* limit = NULL) const;
93 // Return the address corresponding to the next unmarked bit at or after
94 // "addr", and before "limit", if "limit" is non-NULL. If there is no
95 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
96 HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr,
97 const HeapWord* limit = NULL) const;
98
99 // conversion utilities
100 HeapWord* offsetToHeapWord(size_t offset) const {
101 return _bmStartWord + (offset << _shifter);
102 }
103 size_t heapWordToOffset(const HeapWord* addr) const {
104 return pointer_delta(addr, _bmStartWord) >> _shifter;
105 }
106 int heapWordDiffToOffsetDiff(size_t diff) const;
107
108 // The argument addr should be the start address of a valid object
109 HeapWord* nextObject(HeapWord* addr) {
110 oop obj = (oop) addr;
111 HeapWord* res = addr + obj->size();
112 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
113 return res;
114 }
115
116 void print_on_error(outputStream* st, const char* prefix) const;
117
118 // debugging
119 NOT_PRODUCT(bool covers(MemRegion rs) const;)
120 };
121
122 class CMBitMapMappingChangedListener : public G1MappingChangedListener {
123 private:
124 CMBitMap* _bm;
125 public:
126 CMBitMapMappingChangedListener() : _bm(NULL) {}
127
128 void set_bitmap(CMBitMap* bm) { _bm = bm; }
129
130 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
131 };
132
133 class CMBitMap : public CMBitMapRO {
134 private:
135 CMBitMapMappingChangedListener _listener;
136
137 public:
138 static size_t compute_size(size_t heap_size);
139 // Returns the amount of bytes on the heap between two marks in the bitmap.
140 static size_t mark_distance();
141 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
142 // mark bitmap corresponds to. This is the same as the mark distance above.
143 static size_t heap_map_factor() {
144 return mark_distance();
145 }
146
147 CMBitMap() : CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
148
149 // Initializes the underlying BitMap to cover the given area.
150 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
151
152 // Write marks.
153 inline void mark(HeapWord* addr);
154 inline void clear(HeapWord* addr);
155 inline bool parMark(HeapWord* addr);
156 inline bool parClear(HeapWord* addr);
157
158 void markRange(MemRegion mr);
159 void clearRange(MemRegion mr);
160
161 // Starting at the bit corresponding to "addr" (inclusive), find the next
162 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
163 // the end of this run (stopping at "end_addr"). Return the MemRegion
164 // covering from the start of the region corresponding to the first bit
165 // of the run to the end of the region corresponding to the last bit of
166 // the run. If there is no "1" bit at or after "addr", return an empty
167 // MemRegion.
168 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
169
170 // Clear the whole mark bitmap.
171 void clearAll();
172 };
173
174 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
175 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
176 VirtualSpace _virtual_space; // Underlying backing store for actual stack
177 ConcurrentMark* _cm;
178 oop* _base; // bottom of stack
179 jint _index; // one more than last occupied index
180 jint _capacity; // max #elements
181 jint _saved_index; // value of _index saved at start of GC
182
183 bool _overflow;
184 bool _should_expand;
185 DEBUG_ONLY(bool _drain_in_progress;)
186 DEBUG_ONLY(bool _drain_in_progress_yields;)
187
188 oop pop() {
189 if (!isEmpty()) {
190 return _base[--_index] ;
191 }
192 return NULL;
193 }
194
195 public:
196 CMMarkStack(ConcurrentMark* cm);
197 ~CMMarkStack();
198
199 bool allocate(size_t capacity);
200
201 // Pushes the first "n" elements of "ptr_arr" on the stack.
202 // Locking impl: concurrency is allowed only with
203 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
204 // locking strategy.
205 void par_push_arr(oop* ptr_arr, int n);
206
207 // If returns false, the array was empty. Otherwise, removes up to "max"
208 // elements from the stack, and transfers them to "ptr_arr" in an
209 // unspecified order. The actual number transferred is given in "n" ("n
210 // == 0" is deliberately redundant with the return value.) Locking impl:
211 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
212 // operations, which use the same locking strategy.
213 bool par_pop_arr(oop* ptr_arr, int max, int* n);
214
215 // Drain the mark stack, applying the given closure to all fields of
216 // objects on the stack. (That is, continue until the stack is empty,
217 // even if closure applications add entries to the stack.) The "bm"
218 // argument, if non-null, may be used to verify that only marked objects
219 // are on the mark stack. If "yield_after" is "true", then the
220 // concurrent marker performing the drain offers to yield after
221 // processing each object. If a yield occurs, stops the drain operation
222 // and returns false. Otherwise, returns true.
223 template<class OopClosureClass>
224 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
225
226 bool isEmpty() { return _index == 0; }
227 int maxElems() { return _capacity; }
228
229 bool overflow() { return _overflow; }
230 void clear_overflow() { _overflow = false; }
231
232 bool should_expand() const { return _should_expand; }
233 void set_should_expand();
234
235 // Expand the stack, typically in response to an overflow condition
236 void expand();
237
238 int size() { return _index; }
239
240 void setEmpty() { _index = 0; clear_overflow(); }
241
242 // Record the current index.
243 void note_start_of_gc();
244
245 // Make sure that we have not added any entries to the stack during GC.
246 void note_end_of_gc();
247
248 // Apply fn to each oop in the mark stack, up to the bound recorded
249 // via one of the above "note" functions. The mark stack must not
250 // be modified while iterating.
251 template<typename Fn> void iterate(Fn fn);
252 };
253
254 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
255 private:
256 #ifndef PRODUCT
257 uintx _num_remaining;
258 bool _force;
259 #endif // !defined(PRODUCT)
260
261 public:
262 void init() PRODUCT_RETURN;
263 void update() PRODUCT_RETURN;
264 bool should_force() PRODUCT_RETURN_( return false; );
265 };
266
267 // this will enable a variety of different statistics per GC task
268 #define _MARKING_STATS_ 0
269 // this will enable the higher verbose levels
270 #define _MARKING_VERBOSE_ 0
271
272 #if _MARKING_STATS_
273 #define statsOnly(statement) \
274 do { \
275 statement ; \
276 } while (0)
277 #else // _MARKING_STATS_
278 #define statsOnly(statement) \
279 do { \
280 } while (0)
281 #endif // _MARKING_STATS_
282
283 typedef enum {
284 no_verbose = 0, // verbose turned off
285 stats_verbose, // only prints stats at the end of marking
286 low_verbose, // low verbose, mostly per region and per major event
287 medium_verbose, // a bit more detailed than low
288 high_verbose // per object verbose
289 } CMVerboseLevel;
290
291 class YoungList;
292
293 // Root Regions are regions that are not empty at the beginning of a
294 // marking cycle and which we might collect during an evacuation pause
295 // while the cycle is active. Given that, during evacuation pauses, we
296 // do not copy objects that are explicitly marked, what we have to do
297 // for the root regions is to scan them and mark all objects reachable
298 // from them. According to the SATB assumptions, we only need to visit
299 // each object once during marking. So, as long as we finish this scan
300 // before the next evacuation pause, we can copy the objects from the
301 // root regions without having to mark them or do anything else to them.
302 //
303 // Currently, we only support root region scanning once (at the start
304 // of the marking cycle) and the root regions are all the survivor
305 // regions populated during the initial-mark pause.
306 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
307 private:
308 YoungList* _young_list;
309 ConcurrentMark* _cm;
310
311 volatile bool _scan_in_progress;
312 volatile bool _should_abort;
313 HeapRegion* volatile _next_survivor;
314
315 public:
316 CMRootRegions();
317 // We actually do most of the initialization in this method.
318 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
319
320 // Reset the claiming / scanning of the root regions.
321 void prepare_for_scan();
322
323 // Forces get_next() to return NULL so that the iteration aborts early.
324 void abort() { _should_abort = true; }
325
326 // Return true if the CM thread are actively scanning root regions,
327 // false otherwise.
328 bool scan_in_progress() { return _scan_in_progress; }
329
330 // Claim the next root region to scan atomically, or return NULL if
331 // all have been claimed.
332 HeapRegion* claim_next();
333
334 // Flag that we're done with root region scanning and notify anyone
335 // who's waiting on it. If aborted is false, assume that all regions
336 // have been claimed.
337 void scan_finished();
338
339 // If CM threads are still scanning root regions, wait until they
340 // are done. Return true if we had to wait, false otherwise.
341 bool wait_until_scan_finished();
342 };
343
344 class ConcurrentMarkThread;
345
346 class ConcurrentMark: public CHeapObj<mtGC> {
347 friend class CMMarkStack;
348 friend class ConcurrentMarkThread;
349 friend class CMTask;
350 friend class CMBitMapClosure;
351 friend class CMRemarkTask;
352 friend class CMConcurrentMarkingTask;
353 friend class G1ParNoteEndTask;
354 friend class CalcLiveObjectsClosure;
355 friend class G1CMRefProcTaskProxy;
356 friend class G1CMRefProcTaskExecutor;
357 friend class G1CMKeepAliveAndDrainClosure;
358 friend class G1CMDrainMarkingStackClosure;
359
360 protected:
361 ConcurrentMarkThread* _cmThread; // The thread doing the work
362 G1CollectedHeap* _g1h; // The heap
363 uint _parallel_marking_threads; // The number of marking
364 // threads we're using
365 uint _max_parallel_marking_threads; // Max number of marking
366 // threads we'll ever use
367 double _sleep_factor; // How much we have to sleep, with
368 // respect to the work we just did, to
369 // meet the marking overhead goal
370 double _marking_task_overhead; // Marking target overhead for
371 // a single task
372
373 // Same as the two above, but for the cleanup task
374 double _cleanup_sleep_factor;
375 double _cleanup_task_overhead;
376
377 FreeRegionList _cleanup_list;
378
379 // Concurrent marking support structures
380 CMBitMap _markBitMap1;
381 CMBitMap _markBitMap2;
382 CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
383 CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
384
385 BitMap _region_bm;
386 BitMap _card_bm;
387
388 // Heap bounds
389 HeapWord* _heap_start;
390 HeapWord* _heap_end;
391
392 // Root region tracking and claiming
393 CMRootRegions _root_regions;
394
395 // For gray objects
396 CMMarkStack _markStack; // Grey objects behind global finger
397 HeapWord* volatile _finger; // The global finger, region aligned,
398 // always points to the end of the
399 // last claimed region
400
401 // Marking tasks
402 uint _max_worker_id;// Maximum worker id
403 uint _active_tasks; // Task num currently active
404 CMTask** _tasks; // Task queue array (max_worker_id len)
405 CMTaskQueueSet* _task_queues; // Task queue set
406 ParallelTaskTerminator _terminator; // For termination
407
408 // Two sync barriers that are used to synchronize tasks when an
409 // overflow occurs. The algorithm is the following. All tasks enter
410 // the first one to ensure that they have all stopped manipulating
411 // the global data structures. After they exit it, they re-initialize
412 // their data structures and task 0 re-initializes the global data
413 // structures. Then, they enter the second sync barrier. This
414 // ensure, that no task starts doing work before all data
415 // structures (local and global) have been re-initialized. When they
416 // exit it, they are free to start working again.
417 WorkGangBarrierSync _first_overflow_barrier_sync;
418 WorkGangBarrierSync _second_overflow_barrier_sync;
419
420 // This is set by any task, when an overflow on the global data
421 // structures is detected
422 volatile bool _has_overflown;
423 // True: marking is concurrent, false: we're in remark
424 volatile bool _concurrent;
425 // Set at the end of a Full GC so that marking aborts
426 volatile bool _has_aborted;
427
428 // Used when remark aborts due to an overflow to indicate that
429 // another concurrent marking phase should start
430 volatile bool _restart_for_overflow;
431
432 // This is true from the very start of concurrent marking until the
433 // point when all the tasks complete their work. It is really used
434 // to determine the points between the end of concurrent marking and
435 // time of remark.
436 volatile bool _concurrent_marking_in_progress;
437
438 // Verbose level
439 CMVerboseLevel _verbose_level;
440
441 // All of these times are in ms
442 NumberSeq _init_times;
443 NumberSeq _remark_times;
444 NumberSeq _remark_mark_times;
445 NumberSeq _remark_weak_ref_times;
446 NumberSeq _cleanup_times;
447 double _total_counting_time;
448 double _total_rs_scrub_time;
449
450 double* _accum_task_vtime; // Accumulated task vtime
451
452 WorkGang* _parallel_workers;
453
454 ForceOverflowSettings _force_overflow_conc;
455 ForceOverflowSettings _force_overflow_stw;
456
457 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
458 void weakRefsWork(bool clear_all_soft_refs);
459
460 void swapMarkBitMaps();
461
462 // It resets the global marking data structures, as well as the
463 // task local ones; should be called during initial mark.
464 void reset();
465
466 // Resets all the marking data structures. Called when we have to restart
467 // marking or when marking completes (via set_non_marking_state below).
468 void reset_marking_state(bool clear_overflow = true);
469
470 // We do this after we're done with marking so that the marking data
471 // structures are initialized to a sensible and predictable state.
472 void set_non_marking_state();
473
474 // Called to indicate how many threads are currently active.
475 void set_concurrency(uint active_tasks);
476
477 // It should be called to indicate which phase we're in (concurrent
478 // mark or remark) and how many threads are currently active.
479 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
480
481 // Prints all gathered CM-related statistics
482 void print_stats();
483
484 bool cleanup_list_is_empty() {
485 return _cleanup_list.is_empty();
486 }
487
488 // Accessor methods
489 uint parallel_marking_threads() const { return _parallel_marking_threads; }
490 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
491 double sleep_factor() { return _sleep_factor; }
492 double marking_task_overhead() { return _marking_task_overhead;}
493 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
494 double cleanup_task_overhead() { return _cleanup_task_overhead;}
495
496 HeapWord* finger() { return _finger; }
497 bool concurrent() { return _concurrent; }
498 uint active_tasks() { return _active_tasks; }
499 ParallelTaskTerminator* terminator() { return &_terminator; }
500
501 // It claims the next available region to be scanned by a marking
502 // task/thread. It might return NULL if the next region is empty or
503 // we have run out of regions. In the latter case, out_of_regions()
504 // determines whether we've really run out of regions or the task
505 // should call claim_region() again. This might seem a bit
506 // awkward. Originally, the code was written so that claim_region()
507 // either successfully returned with a non-empty region or there
508 // were no more regions to be claimed. The problem with this was
509 // that, in certain circumstances, it iterated over large chunks of
510 // the heap finding only empty regions and, while it was working, it
511 // was preventing the calling task to call its regular clock
512 // method. So, this way, each task will spend very little time in
513 // claim_region() and is allowed to call the regular clock method
514 // frequently.
515 HeapRegion* claim_region(uint worker_id);
516
517 // It determines whether we've run out of regions to scan. Note that
518 // the finger can point past the heap end in case the heap was expanded
519 // to satisfy an allocation without doing a GC. This is fine, because all
520 // objects in those regions will be considered live anyway because of
521 // SATB guarantees (i.e. their TAMS will be equal to bottom).
522 bool out_of_regions() { return _finger >= _heap_end; }
523
524 // Returns the task with the given id
525 CMTask* task(int id) {
526 assert(0 <= id && id < (int) _active_tasks,
527 "task id not within active bounds");
528 return _tasks[id];
529 }
530
531 // Returns the task queue with the given id
532 CMTaskQueue* task_queue(int id) {
533 assert(0 <= id && id < (int) _active_tasks,
534 "task queue id not within active bounds");
535 return (CMTaskQueue*) _task_queues->queue(id);
536 }
537
538 // Returns the task queue set
539 CMTaskQueueSet* task_queues() { return _task_queues; }
540
541 // Access / manipulation of the overflow flag which is set to
542 // indicate that the global stack has overflown
543 bool has_overflown() { return _has_overflown; }
544 void set_has_overflown() { _has_overflown = true; }
545 void clear_has_overflown() { _has_overflown = false; }
546 bool restart_for_overflow() { return _restart_for_overflow; }
547
548 // Methods to enter the two overflow sync barriers
549 void enter_first_sync_barrier(uint worker_id);
550 void enter_second_sync_barrier(uint worker_id);
551
552 ForceOverflowSettings* force_overflow_conc() {
553 return &_force_overflow_conc;
554 }
555
556 ForceOverflowSettings* force_overflow_stw() {
557 return &_force_overflow_stw;
558 }
559
560 ForceOverflowSettings* force_overflow() {
561 if (concurrent()) {
562 return force_overflow_conc();
563 } else {
564 return force_overflow_stw();
565 }
566 }
567
568 // Live Data Counting data structures...
569 // These data structures are initialized at the start of
570 // marking. They are written to while marking is active.
571 // They are aggregated during remark; the aggregated values
572 // are then used to populate the _region_bm, _card_bm, and
573 // the total live bytes, which are then subsequently updated
574 // during cleanup.
575
576 // An array of bitmaps (one bit map per task). Each bitmap
577 // is used to record the cards spanned by the live objects
578 // marked by that task/worker.
579 BitMap* _count_card_bitmaps;
580
581 // Used to record the number of marked live bytes
582 // (for each region, by worker thread).
583 size_t** _count_marked_bytes;
584
585 // Card index of the bottom of the G1 heap. Used for biasing indices into
586 // the card bitmaps.
587 intptr_t _heap_bottom_card_num;
588
589 // Set to true when initialization is complete
590 bool _completed_initialization;
591
592 public:
593 // Manipulation of the global mark stack.
594 // The push and pop operations are used by tasks for transfers
595 // between task-local queues and the global mark stack, and use
596 // locking for concurrency safety.
597 bool mark_stack_push(oop* arr, int n) {
598 _markStack.par_push_arr(arr, n);
599 if (_markStack.overflow()) {
600 set_has_overflown();
601 return false;
602 }
603 return true;
604 }
605 void mark_stack_pop(oop* arr, int max, int* n) {
606 _markStack.par_pop_arr(arr, max, n);
607 }
608 size_t mark_stack_size() { return _markStack.size(); }
609 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
610 bool mark_stack_overflow() { return _markStack.overflow(); }
611 bool mark_stack_empty() { return _markStack.isEmpty(); }
612
613 CMRootRegions* root_regions() { return &_root_regions; }
614
615 bool concurrent_marking_in_progress() {
616 return _concurrent_marking_in_progress;
617 }
618 void set_concurrent_marking_in_progress() {
619 _concurrent_marking_in_progress = true;
620 }
621 void clear_concurrent_marking_in_progress() {
622 _concurrent_marking_in_progress = false;
623 }
624
625 void update_accum_task_vtime(int i, double vtime) {
626 _accum_task_vtime[i] += vtime;
627 }
628
629 double all_task_accum_vtime() {
630 double ret = 0.0;
631 for (uint i = 0; i < _max_worker_id; ++i)
632 ret += _accum_task_vtime[i];
633 return ret;
634 }
635
636 // Attempts to steal an object from the task queues of other tasks
637 bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
638
639 ConcurrentMark(G1CollectedHeap* g1h,
640 G1RegionToSpaceMapper* prev_bitmap_storage,
641 G1RegionToSpaceMapper* next_bitmap_storage);
642 ~ConcurrentMark();
643
644 ConcurrentMarkThread* cmThread() { return _cmThread; }
645
646 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
647 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
648
649 // Returns the number of GC threads to be used in a concurrent
650 // phase based on the number of GC threads being used in a STW
651 // phase.
652 uint scale_parallel_threads(uint n_par_threads);
653
654 // Calculates the number of GC threads to be used in a concurrent phase.
655 uint calc_parallel_marking_threads();
656
657 // The following three are interaction between CM and
658 // G1CollectedHeap
659
660 // This notifies CM that a root during initial-mark needs to be
661 // grayed. It is MT-safe. word_size is the size of the object in
662 // words. It is passed explicitly as sometimes we cannot calculate
663 // it from the given object because it might be in an inconsistent
664 // state (e.g., in to-space and being copied). So the caller is
665 // responsible for dealing with this issue (e.g., get the size from
666 // the from-space image when the to-space image might be
667 // inconsistent) and always passing the size. hr is the region that
668 // contains the object and it's passed optionally from callers who
669 // might already have it (no point in recalculating it).
670 inline void grayRoot(oop obj,
671 size_t word_size,
672 uint worker_id,
673 HeapRegion* hr = NULL);
674
675 // It iterates over the heap and for each object it comes across it
676 // will dump the contents of its reference fields, as well as
677 // liveness information for the object and its referents. The dump
678 // will be written to a file with the following name:
679 // G1PrintReachableBaseFile + "." + str.
680 // vo decides whether the prev (vo == UsePrevMarking), the next
681 // (vo == UseNextMarking) marking information, or the mark word
682 // (vo == UseMarkWord) will be used to determine the liveness of
683 // each object / referent.
684 // If all is true, all objects in the heap will be dumped, otherwise
685 // only the live ones. In the dump the following symbols / breviations
686 // are used:
687 // M : an explicitly live object (its bitmap bit is set)
688 // > : an implicitly live object (over tams)
689 // O : an object outside the G1 heap (typically: in the perm gen)
690 // NOT : a reference field whose referent is not live
691 // AND MARKED : indicates that an object is both explicitly and
692 // implicitly live (it should be one or the other, not both)
693 void print_reachable(const char* str,
694 VerifyOption vo,
695 bool all) PRODUCT_RETURN;
696
697 // Clear the next marking bitmap (will be called concurrently).
698 void clearNextBitmap();
699
700 // Return whether the next mark bitmap has no marks set. To be used for assertions
701 // only. Will not yield to pause requests.
702 bool nextMarkBitmapIsClear();
703
704 // These two do the work that needs to be done before and after the
705 // initial root checkpoint. Since this checkpoint can be done at two
706 // different points (i.e. an explicit pause or piggy-backed on a
707 // young collection), then it's nice to be able to easily share the
708 // pre/post code. It might be the case that we can put everything in
709 // the post method. TP
710 void checkpointRootsInitialPre();
711 void checkpointRootsInitialPost();
712
713 // Scan all the root regions and mark everything reachable from
714 // them.
715 void scanRootRegions();
716
717 // Scan a single root region and mark everything reachable from it.
718 void scanRootRegion(HeapRegion* hr, uint worker_id);
719
720 // Do concurrent phase of marking, to a tentative transitive closure.
721 void markFromRoots();
722
723 void checkpointRootsFinal(bool clear_all_soft_refs);
724 void checkpointRootsFinalWork();
725 void cleanup();
726 void completeCleanup();
727
728 // Mark in the previous bitmap. NB: this is usually read-only, so use
729 // this carefully!
730 inline void markPrev(oop p);
731
732 // Clears marks for all objects in the given range, for the prev or
733 // next bitmaps. NB: the previous bitmap is usually
734 // read-only, so use this carefully!
735 void clearRangePrevBitmap(MemRegion mr);
736 void clearRangeNextBitmap(MemRegion mr);
737
738 // Notify data structures that a GC has started.
739 void note_start_of_gc() {
740 _markStack.note_start_of_gc();
741 }
742
743 // Notify data structures that a GC is finished.
744 void note_end_of_gc() {
745 _markStack.note_end_of_gc();
746 }
747
748 // Verify that there are no CSet oops on the stacks (taskqueues /
749 // global mark stack) and fingers (global / per-task).
750 // If marking is not in progress, it's a no-op.
751 void verify_no_cset_oops() PRODUCT_RETURN;
752
753 bool isPrevMarked(oop p) const {
754 assert(p != NULL && p->is_oop(), "expected an oop");
755 HeapWord* addr = (HeapWord*)p;
756 assert(addr >= _prevMarkBitMap->startWord() ||
757 addr < _prevMarkBitMap->endWord(), "in a region");
758
759 return _prevMarkBitMap->isMarked(addr);
760 }
761
762 inline bool do_yield_check(uint worker_i = 0);
763
764 // Called to abort the marking cycle after a Full GC takes place.
765 void abort();
766
767 bool has_aborted() { return _has_aborted; }
768
769 // This prints the global/local fingers. It is used for debugging.
770 NOT_PRODUCT(void print_finger();)
771
772 void print_summary_info();
773
774 void print_worker_threads_on(outputStream* st) const;
775
776 void print_on_error(outputStream* st) const;
777
778 // The following indicate whether a given verbose level has been
779 // set. Notice that anything above stats is conditional to
780 // _MARKING_VERBOSE_ having been set to 1
781 bool verbose_stats() {
782 return _verbose_level >= stats_verbose;
783 }
784 bool verbose_low() {
785 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
786 }
787 bool verbose_medium() {
788 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
789 }
790 bool verbose_high() {
791 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
792 }
793
794 // Liveness counting
795
796 // Utility routine to set an exclusive range of cards on the given
797 // card liveness bitmap
798 inline void set_card_bitmap_range(BitMap* card_bm,
799 BitMap::idx_t start_idx,
800 BitMap::idx_t end_idx,
801 bool is_par);
802
803 // Returns the card number of the bottom of the G1 heap.
804 // Used in biasing indices into accounting card bitmaps.
805 intptr_t heap_bottom_card_num() const {
806 return _heap_bottom_card_num;
807 }
808
809 // Returns the card bitmap for a given task or worker id.
810 BitMap* count_card_bitmap_for(uint worker_id) {
811 assert(worker_id < _max_worker_id, "oob");
812 assert(_count_card_bitmaps != NULL, "uninitialized");
813 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
814 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
815 return task_card_bm;
816 }
817
818 // Returns the array containing the marked bytes for each region,
819 // for the given worker or task id.
820 size_t* count_marked_bytes_array_for(uint worker_id) {
821 assert(worker_id < _max_worker_id, "oob");
822 assert(_count_marked_bytes != NULL, "uninitialized");
823 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
824 assert(marked_bytes_array != NULL, "uninitialized");
825 return marked_bytes_array;
826 }
827
828 // Returns the index in the liveness accounting card table bitmap
829 // for the given address
830 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
831
832 // Counts the size of the given memory region in the the given
833 // marked_bytes array slot for the given HeapRegion.
834 // Sets the bits in the given card bitmap that are associated with the
835 // cards that are spanned by the memory region.
836 inline void count_region(MemRegion mr,
837 HeapRegion* hr,
838 size_t* marked_bytes_array,
839 BitMap* task_card_bm);
840
841 // Counts the given memory region in the task/worker counting
842 // data structures for the given worker id.
843 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
844
845 // Counts the given object in the given task/worker counting
846 // data structures.
847 inline void count_object(oop obj,
848 HeapRegion* hr,
849 size_t* marked_bytes_array,
850 BitMap* task_card_bm);
851
852 // Attempts to mark the given object and, if successful, counts
853 // the object in the given task/worker counting structures.
854 inline bool par_mark_and_count(oop obj,
855 HeapRegion* hr,
856 size_t* marked_bytes_array,
857 BitMap* task_card_bm);
858
859 // Attempts to mark the given object and, if successful, counts
860 // the object in the task/worker counting structures for the
861 // given worker id.
862 inline bool par_mark_and_count(oop obj,
863 size_t word_size,
864 HeapRegion* hr,
865 uint worker_id);
866
867 // Returns true if initialization was successfully completed.
868 bool completed_initialization() const {
869 return _completed_initialization;
870 }
871
872 protected:
873 // Clear all the per-task bitmaps and arrays used to store the
874 // counting data.
875 void clear_all_count_data();
876
877 // Aggregates the counting data for each worker/task
878 // that was constructed while marking. Also sets
879 // the amount of marked bytes for each region and
880 // the top at concurrent mark count.
881 void aggregate_count_data();
882
883 // Verification routine
884 void verify_count_data();
885 };
886
887 // A class representing a marking task.
888 class CMTask : public TerminatorTerminator {
889 private:
890 enum PrivateConstants {
891 // the regular clock call is called once the scanned words reaches
892 // this limit
893 words_scanned_period = 12*1024,
894 // the regular clock call is called once the number of visited
895 // references reaches this limit
896 refs_reached_period = 384,
897 // initial value for the hash seed, used in the work stealing code
898 init_hash_seed = 17,
899 // how many entries will be transferred between global stack and
900 // local queues
901 global_stack_transfer_size = 16
902 };
903
904 uint _worker_id;
905 G1CollectedHeap* _g1h;
906 ConcurrentMark* _cm;
907 CMBitMap* _nextMarkBitMap;
908 // the task queue of this task
909 CMTaskQueue* _task_queue;
910 private:
911 // the task queue set---needed for stealing
912 CMTaskQueueSet* _task_queues;
913 // indicates whether the task has been claimed---this is only for
914 // debugging purposes
915 bool _claimed;
916
917 // number of calls to this task
918 int _calls;
919
920 // when the virtual timer reaches this time, the marking step should
921 // exit
922 double _time_target_ms;
923 // the start time of the current marking step
924 double _start_time_ms;
925
926 // the oop closure used for iterations over oops
927 G1CMOopClosure* _cm_oop_closure;
928
929 // the region this task is scanning, NULL if we're not scanning any
930 HeapRegion* _curr_region;
931 // the local finger of this task, NULL if we're not scanning a region
932 HeapWord* _finger;
933 // limit of the region this task is scanning, NULL if we're not scanning one
934 HeapWord* _region_limit;
935
936 // the number of words this task has scanned
937 size_t _words_scanned;
938 // When _words_scanned reaches this limit, the regular clock is
939 // called. Notice that this might be decreased under certain
940 // circumstances (i.e. when we believe that we did an expensive
941 // operation).
942 size_t _words_scanned_limit;
943 // the initial value of _words_scanned_limit (i.e. what it was
944 // before it was decreased).
945 size_t _real_words_scanned_limit;
946
947 // the number of references this task has visited
948 size_t _refs_reached;
949 // When _refs_reached reaches this limit, the regular clock is
950 // called. Notice this this might be decreased under certain
951 // circumstances (i.e. when we believe that we did an expensive
952 // operation).
953 size_t _refs_reached_limit;
954 // the initial value of _refs_reached_limit (i.e. what it was before
955 // it was decreased).
956 size_t _real_refs_reached_limit;
957
958 // used by the work stealing stuff
959 int _hash_seed;
960 // if this is true, then the task has aborted for some reason
961 bool _has_aborted;
962 // set when the task aborts because it has met its time quota
963 bool _has_timed_out;
964 // true when we're draining SATB buffers; this avoids the task
965 // aborting due to SATB buffers being available (as we're already
966 // dealing with them)
967 bool _draining_satb_buffers;
968
969 // number sequence of past step times
970 NumberSeq _step_times_ms;
971 // elapsed time of this task
972 double _elapsed_time_ms;
973 // termination time of this task
974 double _termination_time_ms;
975 // when this task got into the termination protocol
976 double _termination_start_time_ms;
977
978 // true when the task is during a concurrent phase, false when it is
979 // in the remark phase (so, in the latter case, we do not have to
980 // check all the things that we have to check during the concurrent
981 // phase, i.e. SATB buffer availability...)
982 bool _concurrent;
983
984 TruncatedSeq _marking_step_diffs_ms;
985
986 // Counting data structures. Embedding the task's marked_bytes_array
987 // and card bitmap into the actual task saves having to go through
988 // the ConcurrentMark object.
989 size_t* _marked_bytes_array;
990 BitMap* _card_bm;
991
992 // LOTS of statistics related with this task
993 #if _MARKING_STATS_
994 NumberSeq _all_clock_intervals_ms;
995 double _interval_start_time_ms;
996
997 size_t _aborted;
998 size_t _aborted_overflow;
999 size_t _aborted_cm_aborted;
1000 size_t _aborted_yield;
1001 size_t _aborted_timed_out;
1002 size_t _aborted_satb;
1003 size_t _aborted_termination;
1004
1005 size_t _steal_attempts;
1006 size_t _steals;
1007
1008 size_t _clock_due_to_marking;
1009 size_t _clock_due_to_scanning;
1010
1011 size_t _local_pushes;
1012 size_t _local_pops;
1013 size_t _local_max_size;
1014 size_t _objs_scanned;
1015
1016 size_t _global_pushes;
1017 size_t _global_pops;
1018 size_t _global_max_size;
1019
1020 size_t _global_transfers_to;
1021 size_t _global_transfers_from;
1022
1023 size_t _regions_claimed;
1024 size_t _objs_found_on_bitmap;
1025
1026 size_t _satb_buffers_processed;
1027 #endif // _MARKING_STATS_
1028
1029 // it updates the local fields after this task has claimed
1030 // a new region to scan
1031 void setup_for_region(HeapRegion* hr);
1032 // it brings up-to-date the limit of the region
1033 void update_region_limit();
1034
1035 // called when either the words scanned or the refs visited limit
1036 // has been reached
1037 void reached_limit();
1038 // recalculates the words scanned and refs visited limits
1039 void recalculate_limits();
1040 // decreases the words scanned and refs visited limits when we reach
1041 // an expensive operation
1042 void decrease_limits();
1043 // it checks whether the words scanned or refs visited reached their
1044 // respective limit and calls reached_limit() if they have
1045 void check_limits() {
1046 if (_words_scanned >= _words_scanned_limit ||
1047 _refs_reached >= _refs_reached_limit) {
1048 reached_limit();
1049 }
1050 }
1051 // this is supposed to be called regularly during a marking step as
1052 // it checks a bunch of conditions that might cause the marking step
1053 // to abort
1054 void regular_clock_call();
1055 bool concurrent() { return _concurrent; }
1056
1057 // Test whether obj might have already been passed over by the
1058 // mark bitmap scan, and so needs to be pushed onto the mark stack.
1059 bool is_below_finger(oop obj, HeapWord* global_finger) const;
1060
1061 template<bool scan> void process_grey_object(oop obj);
1062
1063 public:
1064 // It resets the task; it should be called right at the beginning of
1065 // a marking phase.
1066 void reset(CMBitMap* _nextMarkBitMap);
1067 // it clears all the fields that correspond to a claimed region.
1068 void clear_region_fields();
1069
1070 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1071
1072 // The main method of this class which performs a marking step
1073 // trying not to exceed the given duration. However, it might exit
1074 // prematurely, according to some conditions (i.e. SATB buffers are
1075 // available for processing).
1076 void do_marking_step(double target_ms,
1077 bool do_termination,
1078 bool is_serial);
1079
1080 // These two calls start and stop the timer
1081 void record_start_time() {
1082 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1083 }
1084 void record_end_time() {
1085 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1086 }
1087
1088 // returns the worker ID associated with this task.
1089 uint worker_id() { return _worker_id; }
1090
1091 // From TerminatorTerminator. It determines whether this task should
1092 // exit the termination protocol after it's entered it.
1093 virtual bool should_exit_termination();
1094
1095 // Resets the local region fields after a task has finished scanning a
1096 // region; or when they have become stale as a result of the region
1097 // being evacuated.
1098 void giveup_current_region();
1099
1100 HeapWord* finger() { return _finger; }
1101
1102 bool has_aborted() { return _has_aborted; }
1103 void set_has_aborted() { _has_aborted = true; }
1104 void clear_has_aborted() { _has_aborted = false; }
1105 bool has_timed_out() { return _has_timed_out; }
1106 bool claimed() { return _claimed; }
1107
1108 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1109
1110 // Increment the number of references this task has visited.
1111 void increment_refs_reached() { ++_refs_reached; }
1112
1113 // Grey the object by marking it. If not already marked, push it on
1114 // the local queue if below the finger.
1115 // Precondition: obj is in region.
1116 // Precondition: obj is below region's NTAMS.
1117 inline void make_reference_grey(oop obj, HeapRegion* region);
1118
1119 // Grey the object (by calling make_grey_reference) if required,
1120 // e.g. obj is below its containing region's NTAMS.
1121 // Precondition: obj is a valid heap object.
1122 inline void deal_with_reference(oop obj);
1123
1124 // It scans an object and visits its children.
1125 inline void scan_object(oop obj);
1126
1127 // It pushes an object on the local queue.
1128 inline void push(oop obj);
1129
1130 // These two move entries to/from the global stack.
1131 void move_entries_to_global_stack();
1132 void get_entries_from_global_stack();
1133
1134 // It pops and scans objects from the local queue. If partially is
1135 // true, then it stops when the queue size is of a given limit. If
1136 // partially is false, then it stops when the queue is empty.
1137 void drain_local_queue(bool partially);
1138 // It moves entries from the global stack to the local queue and
1139 // drains the local queue. If partially is true, then it stops when
1140 // both the global stack and the local queue reach a given size. If
1141 // partially if false, it tries to empty them totally.
1142 void drain_global_stack(bool partially);
1143 // It keeps picking SATB buffers and processing them until no SATB
1144 // buffers are available.
1145 void drain_satb_buffers();
1146
1147 // moves the local finger to a new location
1148 inline void move_finger_to(HeapWord* new_finger) {
1149 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1150 _finger = new_finger;
1151 }
1152
1153 CMTask(uint worker_id,
1154 ConcurrentMark *cm,
1155 size_t* marked_bytes,
1156 BitMap* card_bm,
1157 CMTaskQueue* task_queue,
1158 CMTaskQueueSet* task_queues);
1159
1160 // it prints statistics associated with this task
1161 void print_stats();
1162
1163 #if _MARKING_STATS_
1164 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1165 #endif // _MARKING_STATS_
1166 };
1167
1168 // Class that's used to to print out per-region liveness
1169 // information. It's currently used at the end of marking and also
1170 // after we sort the old regions at the end of the cleanup operation.
1171 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1172 private:
1173 outputStream* _out;
1174
1175 // Accumulators for these values.
1176 size_t _total_used_bytes;
1177 size_t _total_capacity_bytes;
1178 size_t _total_prev_live_bytes;
1179 size_t _total_next_live_bytes;
1180
1181 // These are set up when we come across a "stars humongous" region
1182 // (as this is where most of this information is stored, not in the
1183 // subsequent "continues humongous" regions). After that, for every
1184 // region in a given humongous region series we deduce the right
1185 // values for it by simply subtracting the appropriate amount from
1186 // these fields. All these values should reach 0 after we've visited
1187 // the last region in the series.
1188 size_t _hum_used_bytes;
1189 size_t _hum_capacity_bytes;
1190 size_t _hum_prev_live_bytes;
1191 size_t _hum_next_live_bytes;
1192
1193 // Accumulator for the remembered set size
1194 size_t _total_remset_bytes;
1195
1196 // Accumulator for strong code roots memory size
1197 size_t _total_strong_code_roots_bytes;
1198
1199 static double perc(size_t val, size_t total) {
1200 if (total == 0) {
1201 return 0.0;
1202 } else {
1203 return 100.0 * ((double) val / (double) total);
1204 }
1205 }
1206
1207 static double bytes_to_mb(size_t val) {
1208 return (double) val / (double) M;
1209 }
1210
1211 // See the .cpp file.
1212 size_t get_hum_bytes(size_t* hum_bytes);
1213 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1214 size_t* prev_live_bytes, size_t* next_live_bytes);
1215
1216 public:
1217 // The header and footer are printed in the constructor and
1218 // destructor respectively.
1219 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1220 virtual bool doHeapRegion(HeapRegion* r);
1221 ~G1PrintRegionLivenessInfoClosure();
1222 };
1223
1224 #endif // SHARE_VM_GC_G1_CONCURRENTMARK_HPP