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_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
27
28 #include "classfile/javaClasses.hpp"
29 #include "gc_implementation/g1/heapRegionSet.hpp"
30 #include "gc_implementation/g1/g1RegionToSpaceMapper.hpp"
31 #include "gc_implementation/shared/gcId.hpp"
32 #include "utilities/taskqueue.hpp"
33
34 class G1CollectedHeap;
35 class CMBitMap;
36 class CMTask;
37 class ConcurrentMark;
38 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue;
39 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
40
41 // Closure used by CM during concurrent reference discovery
42 // and reference processing (during remarking) to determine
43 // if a particular object is alive. It is primarily used
44 // to determine if referents of discovered reference objects
45 // are alive. An instance is also embedded into the
46 // reference processor as the _is_alive_non_header field
47 class G1CMIsAliveClosure: public BoolObjectClosure {
48 G1CollectedHeap* _g1;
49 public:
50 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
51
52 bool do_object_b(oop obj);
53 };
54
55 // A generic CM bit map. This is essentially a wrapper around the BitMap
56 // class, with one bit per (1<<_shifter) HeapWords.
57
58 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
59 protected:
60 HeapWord* _bmStartWord; // base address of range covered by map
61 size_t _bmWordSize; // map size (in #HeapWords covered)
62 const int _shifter; // map to char or bit
63 BitMap _bm; // the bit map itself
64
65 public:
66 // constructor
67 CMBitMapRO(int shifter);
68
69 enum { do_yield = true };
70
71 // inquiries
72 HeapWord* startWord() const { return _bmStartWord; }
73 size_t sizeInWords() const { return _bmWordSize; }
74 // the following is one past the last word in space
75 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
76
77 // read marks
78
79 bool isMarked(HeapWord* addr) const {
80 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
81 "outside underlying space?");
82 return _bm.at(heapWordToOffset(addr));
83 }
84
85 // iteration
86 inline bool iterate(BitMapClosure* cl, MemRegion mr);
87 inline bool iterate(BitMapClosure* cl);
88
89 // Return the address corresponding to the next marked bit at or after
90 // "addr", and before "limit", if "limit" is non-NULL. If there is no
91 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
92 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
93 const HeapWord* limit = NULL) const;
94 // Return the address corresponding to the next unmarked bit at or after
95 // "addr", and before "limit", if "limit" is non-NULL. If there is no
96 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
97 HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr,
98 const HeapWord* limit = NULL) const;
99
100 // conversion utilities
101 HeapWord* offsetToHeapWord(size_t offset) const {
102 return _bmStartWord + (offset << _shifter);
103 }
104 size_t heapWordToOffset(const HeapWord* addr) const {
105 return pointer_delta(addr, _bmStartWord) >> _shifter;
106 }
107 int heapWordDiffToOffsetDiff(size_t diff) const;
108
109 // The argument addr should be the start address of a valid object
110 HeapWord* nextObject(HeapWord* addr) {
111 oop obj = (oop) addr;
112 HeapWord* res = addr + obj->size();
113 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
114 return res;
115 }
116
117 void print_on_error(outputStream* st, const char* prefix) const;
118
119 // debugging
120 NOT_PRODUCT(bool covers(MemRegion rs) const;)
121 };
122
123 class CMBitMapMappingChangedListener : public G1MappingChangedListener {
124 private:
125 CMBitMap* _bm;
126 public:
127 CMBitMapMappingChangedListener() : _bm(NULL) {}
128
129 void set_bitmap(CMBitMap* bm) { _bm = bm; }
130
131 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
132 };
133
134 class CMBitMap : public CMBitMapRO {
135 private:
136 CMBitMapMappingChangedListener _listener;
137
138 public:
139 static size_t compute_size(size_t heap_size);
140 // Returns the amount of bytes on the heap between two marks in the bitmap.
141 static size_t mark_distance();
142 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
143 // mark bitmap corresponds to. This is the same as the mark distance above.
144 static size_t heap_map_factor() {
145 return mark_distance();
146 }
147
148 CMBitMap() : CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
149
150 // Initializes the underlying BitMap to cover the given area.
151 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
152
153 // Write marks.
154 inline void mark(HeapWord* addr);
155 inline void clear(HeapWord* addr);
156 inline bool parMark(HeapWord* addr);
157 inline bool parClear(HeapWord* addr);
158
159 void markRange(MemRegion mr);
160 void clearRange(MemRegion mr);
161
162 // Starting at the bit corresponding to "addr" (inclusive), find the next
163 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
164 // the end of this run (stopping at "end_addr"). Return the MemRegion
165 // covering from the start of the region corresponding to the first bit
166 // of the run to the end of the region corresponding to the last bit of
167 // the run. If there is no "1" bit at or after "addr", return an empty
168 // MemRegion.
169 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
170
171 // Clear the whole mark bitmap.
172 void clearAll();
173 };
174
175 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
176 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
177 VirtualSpace _virtual_space; // Underlying backing store for actual stack
178 ConcurrentMark* _cm;
179 oop* _base; // bottom of stack
180 jint _index; // one more than last occupied index
181 jint _capacity; // max #elements
182 jint _saved_index; // value of _index saved at start of GC
183 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
184
185 bool _overflow;
186 bool _should_expand;
187 DEBUG_ONLY(bool _drain_in_progress;)
188 DEBUG_ONLY(bool _drain_in_progress_yields;)
189
190 public:
191 CMMarkStack(ConcurrentMark* cm);
192 ~CMMarkStack();
193
194 #ifndef PRODUCT
195 jint max_depth() const {
196 return _max_depth;
197 }
198 #endif
199
200 bool allocate(size_t capacity);
201
202 oop pop() {
203 if (!isEmpty()) {
204 return _base[--_index] ;
205 }
206 return NULL;
207 }
208
209 // If overflow happens, don't do the push, and record the overflow.
210 // *Requires* that "ptr" is already marked.
211 void push(oop ptr) {
212 if (isFull()) {
213 // Record overflow.
214 _overflow = true;
215 return;
216 } else {
217 _base[_index++] = ptr;
218 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
219 }
220 }
221 // Non-block impl. Note: concurrency is allowed only with other
222 // "par_push" operations, not with "pop" or "drain". We would need
223 // parallel versions of them if such concurrency was desired.
224 void par_push(oop ptr);
225
226 // Pushes the first "n" elements of "ptr_arr" on the stack.
227 // Non-block impl. Note: concurrency is allowed only with other
228 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
229 void par_adjoin_arr(oop* ptr_arr, int n);
230
231 // Pushes the first "n" elements of "ptr_arr" on the stack.
232 // Locking impl: concurrency is allowed only with
233 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
234 // locking strategy.
235 void par_push_arr(oop* ptr_arr, int n);
236
237 // If returns false, the array was empty. Otherwise, removes up to "max"
238 // elements from the stack, and transfers them to "ptr_arr" in an
239 // unspecified order. The actual number transferred is given in "n" ("n
240 // == 0" is deliberately redundant with the return value.) Locking impl:
241 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
242 // operations, which use the same locking strategy.
243 bool par_pop_arr(oop* ptr_arr, int max, int* n);
244
245 // Drain the mark stack, applying the given closure to all fields of
246 // objects on the stack. (That is, continue until the stack is empty,
247 // even if closure applications add entries to the stack.) The "bm"
248 // argument, if non-null, may be used to verify that only marked objects
249 // are on the mark stack. If "yield_after" is "true", then the
250 // concurrent marker performing the drain offers to yield after
251 // processing each object. If a yield occurs, stops the drain operation
252 // and returns false. Otherwise, returns true.
253 template<class OopClosureClass>
254 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
255
256 bool isEmpty() { return _index == 0; }
257 bool isFull() { return _index == _capacity; }
258 int maxElems() { return _capacity; }
259
260 bool overflow() { return _overflow; }
261 void clear_overflow() { _overflow = false; }
262
263 bool should_expand() const { return _should_expand; }
264 void set_should_expand();
265
266 // Expand the stack, typically in response to an overflow condition
267 void expand();
268
269 int size() { return _index; }
270
271 void setEmpty() { _index = 0; clear_overflow(); }
272
273 // Record the current index.
274 void note_start_of_gc();
275
276 // Make sure that we have not added any entries to the stack during GC.
277 void note_end_of_gc();
278
279 // iterate over the oops in the mark stack, up to the bound recorded via
280 // the call above.
281 void oops_do(OopClosure* f);
282 };
283
284 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
285 private:
286 #ifndef PRODUCT
287 uintx _num_remaining;
288 bool _force;
289 #endif // !defined(PRODUCT)
290
291 public:
292 void init() PRODUCT_RETURN;
293 void update() PRODUCT_RETURN;
294 bool should_force() PRODUCT_RETURN_( return false; );
295 };
296
297 // this will enable a variety of different statistics per GC task
298 #define _MARKING_STATS_ 0
299 // this will enable the higher verbose levels
300 #define _MARKING_VERBOSE_ 0
301
302 #if _MARKING_STATS_
303 #define statsOnly(statement) \
304 do { \
305 statement ; \
306 } while (0)
307 #else // _MARKING_STATS_
308 #define statsOnly(statement) \
309 do { \
310 } while (0)
311 #endif // _MARKING_STATS_
312
313 typedef enum {
314 no_verbose = 0, // verbose turned off
315 stats_verbose, // only prints stats at the end of marking
316 low_verbose, // low verbose, mostly per region and per major event
317 medium_verbose, // a bit more detailed than low
318 high_verbose // per object verbose
319 } CMVerboseLevel;
320
321 class YoungList;
322
323 // Root Regions are regions that are not empty at the beginning of a
324 // marking cycle and which we might collect during an evacuation pause
325 // while the cycle is active. Given that, during evacuation pauses, we
326 // do not copy objects that are explicitly marked, what we have to do
327 // for the root regions is to scan them and mark all objects reachable
328 // from them. According to the SATB assumptions, we only need to visit
329 // each object once during marking. So, as long as we finish this scan
330 // before the next evacuation pause, we can copy the objects from the
331 // root regions without having to mark them or do anything else to them.
332 //
333 // Currently, we only support root region scanning once (at the start
334 // of the marking cycle) and the root regions are all the survivor
335 // regions populated during the initial-mark pause.
336 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
337 private:
338 YoungList* _young_list;
339 ConcurrentMark* _cm;
340
341 volatile bool _scan_in_progress;
342 volatile bool _should_abort;
343 HeapRegion* volatile _next_survivor;
344
345 public:
346 CMRootRegions();
347 // We actually do most of the initialization in this method.
348 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
349
350 // Reset the claiming / scanning of the root regions.
351 void prepare_for_scan();
352
353 // Forces get_next() to return NULL so that the iteration aborts early.
354 void abort() { _should_abort = true; }
355
356 // Return true if the CM thread are actively scanning root regions,
357 // false otherwise.
358 bool scan_in_progress() { return _scan_in_progress; }
359
360 // Claim the next root region to scan atomically, or return NULL if
361 // all have been claimed.
362 HeapRegion* claim_next();
363
364 // Flag that we're done with root region scanning and notify anyone
365 // who's waiting on it. If aborted is false, assume that all regions
366 // have been claimed.
367 void scan_finished();
368
369 // If CM threads are still scanning root regions, wait until they
370 // are done. Return true if we had to wait, false otherwise.
371 bool wait_until_scan_finished();
372 };
373
374 class ConcurrentMarkThread;
375
376 class ConcurrentMark: public CHeapObj<mtGC> {
377 friend class CMMarkStack;
378 friend class ConcurrentMarkThread;
379 friend class CMTask;
380 friend class CMBitMapClosure;
381 friend class CMGlobalObjectClosure;
382 friend class CMRemarkTask;
383 friend class CMConcurrentMarkingTask;
384 friend class G1ParNoteEndTask;
385 friend class CalcLiveObjectsClosure;
386 friend class G1CMRefProcTaskProxy;
387 friend class G1CMRefProcTaskExecutor;
388 friend class G1CMKeepAliveAndDrainClosure;
389 friend class G1CMDrainMarkingStackClosure;
390
391 protected:
392 ConcurrentMarkThread* _cmThread; // The thread doing the work
393 G1CollectedHeap* _g1h; // The heap
394 uint _parallel_marking_threads; // The number of marking
395 // threads we're using
396 uint _max_parallel_marking_threads; // Max number of marking
397 // threads we'll ever use
398 double _sleep_factor; // How much we have to sleep, with
399 // respect to the work we just did, to
400 // meet the marking overhead goal
401 double _marking_task_overhead; // Marking target overhead for
402 // a single task
403
404 // Same as the two above, but for the cleanup task
405 double _cleanup_sleep_factor;
406 double _cleanup_task_overhead;
407
408 FreeRegionList _cleanup_list;
409
410 // Concurrent marking support structures
411 CMBitMap _markBitMap1;
412 CMBitMap _markBitMap2;
413 CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
414 CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
415
416 BitMap _region_bm;
417 BitMap _card_bm;
418
419 // Heap bounds
420 HeapWord* _heap_start;
421 HeapWord* _heap_end;
422
423 // Root region tracking and claiming
424 CMRootRegions _root_regions;
425
426 // For gray objects
427 CMMarkStack _markStack; // Grey objects behind global finger
428 HeapWord* volatile _finger; // The global finger, region aligned,
429 // always points to the end of the
430 // last claimed region
431
432 // Marking tasks
433 uint _max_worker_id;// Maximum worker id
434 uint _active_tasks; // Task num currently active
435 CMTask** _tasks; // Task queue array (max_worker_id len)
436 CMTaskQueueSet* _task_queues; // Task queue set
437 ParallelTaskTerminator _terminator; // For termination
438
439 // Two sync barriers that are used to synchronize tasks when an
440 // overflow occurs. The algorithm is the following. All tasks enter
441 // the first one to ensure that they have all stopped manipulating
442 // the global data structures. After they exit it, they re-initialize
443 // their data structures and task 0 re-initializes the global data
444 // structures. Then, they enter the second sync barrier. This
445 // ensure, that no task starts doing work before all data
446 // structures (local and global) have been re-initialized. When they
447 // exit it, they are free to start working again.
448 WorkGangBarrierSync _first_overflow_barrier_sync;
449 WorkGangBarrierSync _second_overflow_barrier_sync;
450
451 // This is set by any task, when an overflow on the global data
452 // structures is detected
453 volatile bool _has_overflown;
454 // True: marking is concurrent, false: we're in remark
455 volatile bool _concurrent;
456 // Set at the end of a Full GC so that marking aborts
457 volatile bool _has_aborted;
458 GCId _aborted_gc_id;
459
460 // Used when remark aborts due to an overflow to indicate that
461 // another concurrent marking phase should start
462 volatile bool _restart_for_overflow;
463
464 // This is true from the very start of concurrent marking until the
465 // point when all the tasks complete their work. It is really used
466 // to determine the points between the end of concurrent marking and
467 // time of remark.
468 volatile bool _concurrent_marking_in_progress;
469
470 // Verbose level
471 CMVerboseLevel _verbose_level;
472
473 // All of these times are in ms
474 NumberSeq _init_times;
475 NumberSeq _remark_times;
476 NumberSeq _remark_mark_times;
477 NumberSeq _remark_weak_ref_times;
478 NumberSeq _cleanup_times;
479 double _total_counting_time;
480 double _total_rs_scrub_time;
481
482 double* _accum_task_vtime; // Accumulated task vtime
483
484 FlexibleWorkGang* _parallel_workers;
485
486 ForceOverflowSettings _force_overflow_conc;
487 ForceOverflowSettings _force_overflow_stw;
488
489 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
490 void weakRefsWork(bool clear_all_soft_refs);
491
492 void swapMarkBitMaps();
493
494 // It resets the global marking data structures, as well as the
495 // task local ones; should be called during initial mark.
496 void reset();
497
498 // Resets all the marking data structures. Called when we have to restart
499 // marking or when marking completes (via set_non_marking_state below).
500 void reset_marking_state(bool clear_overflow = true);
501
502 // We do this after we're done with marking so that the marking data
503 // structures are initialized to a sensible and predictable state.
504 void set_non_marking_state();
505
506 // Called to indicate how many threads are currently active.
507 void set_concurrency(uint active_tasks);
508
509 // It should be called to indicate which phase we're in (concurrent
510 // mark or remark) and how many threads are currently active.
511 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
512
513 // Prints all gathered CM-related statistics
514 void print_stats();
515
516 bool cleanup_list_is_empty() {
517 return _cleanup_list.is_empty();
518 }
519
520 // Accessor methods
521 uint parallel_marking_threads() const { return _parallel_marking_threads; }
522 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
523 double sleep_factor() { return _sleep_factor; }
524 double marking_task_overhead() { return _marking_task_overhead;}
525 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
526 double cleanup_task_overhead() { return _cleanup_task_overhead;}
527
528 HeapWord* finger() { return _finger; }
529 bool concurrent() { return _concurrent; }
530 uint active_tasks() { return _active_tasks; }
531 ParallelTaskTerminator* terminator() { return &_terminator; }
532
533 // It claims the next available region to be scanned by a marking
534 // task/thread. It might return NULL if the next region is empty or
535 // we have run out of regions. In the latter case, out_of_regions()
536 // determines whether we've really run out of regions or the task
537 // should call claim_region() again. This might seem a bit
538 // awkward. Originally, the code was written so that claim_region()
539 // either successfully returned with a non-empty region or there
540 // were no more regions to be claimed. The problem with this was
541 // that, in certain circumstances, it iterated over large chunks of
542 // the heap finding only empty regions and, while it was working, it
543 // was preventing the calling task to call its regular clock
544 // method. So, this way, each task will spend very little time in
545 // claim_region() and is allowed to call the regular clock method
546 // frequently.
547 HeapRegion* claim_region(uint worker_id);
548
549 // It determines whether we've run out of regions to scan. Note that
550 // the finger can point past the heap end in case the heap was expanded
551 // to satisfy an allocation without doing a GC. This is fine, because all
552 // objects in those regions will be considered live anyway because of
553 // SATB guarantees (i.e. their TAMS will be equal to bottom).
554 bool out_of_regions() { return _finger >= _heap_end; }
555
556 // Returns the task with the given id
557 CMTask* task(int id) {
558 assert(0 <= id && id < (int) _active_tasks,
559 "task id not within active bounds");
560 return _tasks[id];
561 }
562
563 // Returns the task queue with the given id
564 CMTaskQueue* task_queue(int id) {
565 assert(0 <= id && id < (int) _active_tasks,
566 "task queue id not within active bounds");
567 return (CMTaskQueue*) _task_queues->queue(id);
568 }
569
570 // Returns the task queue set
571 CMTaskQueueSet* task_queues() { return _task_queues; }
572
573 // Access / manipulation of the overflow flag which is set to
574 // indicate that the global stack has overflown
575 bool has_overflown() { return _has_overflown; }
576 void set_has_overflown() { _has_overflown = true; }
577 void clear_has_overflown() { _has_overflown = false; }
578 bool restart_for_overflow() { return _restart_for_overflow; }
579
580 // Methods to enter the two overflow sync barriers
581 void enter_first_sync_barrier(uint worker_id);
582 void enter_second_sync_barrier(uint worker_id);
583
584 ForceOverflowSettings* force_overflow_conc() {
585 return &_force_overflow_conc;
586 }
587
588 ForceOverflowSettings* force_overflow_stw() {
589 return &_force_overflow_stw;
590 }
591
592 ForceOverflowSettings* force_overflow() {
593 if (concurrent()) {
594 return force_overflow_conc();
595 } else {
596 return force_overflow_stw();
597 }
598 }
599
600 // Live Data Counting data structures...
601 // These data structures are initialized at the start of
602 // marking. They are written to while marking is active.
603 // They are aggregated during remark; the aggregated values
604 // are then used to populate the _region_bm, _card_bm, and
605 // the total live bytes, which are then subsequently updated
606 // during cleanup.
607
608 // An array of bitmaps (one bit map per task). Each bitmap
609 // is used to record the cards spanned by the live objects
610 // marked by that task/worker.
611 BitMap* _count_card_bitmaps;
612
613 // Used to record the number of marked live bytes
614 // (for each region, by worker thread).
615 size_t** _count_marked_bytes;
616
617 // Card index of the bottom of the G1 heap. Used for biasing indices into
618 // the card bitmaps.
619 intptr_t _heap_bottom_card_num;
620
621 // Set to true when initialization is complete
622 bool _completed_initialization;
623
624 public:
625 // Manipulation of the global mark stack.
626 // Notice that the first mark_stack_push is CAS-based, whereas the
627 // two below are Mutex-based. This is OK since the first one is only
628 // called during evacuation pauses and doesn't compete with the
629 // other two (which are called by the marking tasks during
630 // concurrent marking or remark).
631 bool mark_stack_push(oop p) {
632 _markStack.par_push(p);
633 if (_markStack.overflow()) {
634 set_has_overflown();
635 return false;
636 }
637 return true;
638 }
639 bool mark_stack_push(oop* arr, int n) {
640 _markStack.par_push_arr(arr, n);
641 if (_markStack.overflow()) {
642 set_has_overflown();
643 return false;
644 }
645 return true;
646 }
647 void mark_stack_pop(oop* arr, int max, int* n) {
648 _markStack.par_pop_arr(arr, max, n);
649 }
650 size_t mark_stack_size() { return _markStack.size(); }
651 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
652 bool mark_stack_overflow() { return _markStack.overflow(); }
653 bool mark_stack_empty() { return _markStack.isEmpty(); }
654
655 CMRootRegions* root_regions() { return &_root_regions; }
656
657 bool concurrent_marking_in_progress() {
658 return _concurrent_marking_in_progress;
659 }
660 void set_concurrent_marking_in_progress() {
661 _concurrent_marking_in_progress = true;
662 }
663 void clear_concurrent_marking_in_progress() {
664 _concurrent_marking_in_progress = false;
665 }
666
667 void update_accum_task_vtime(int i, double vtime) {
668 _accum_task_vtime[i] += vtime;
669 }
670
671 double all_task_accum_vtime() {
672 double ret = 0.0;
673 for (uint i = 0; i < _max_worker_id; ++i)
674 ret += _accum_task_vtime[i];
675 return ret;
676 }
677
678 // Attempts to steal an object from the task queues of other tasks
679 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
680 return _task_queues->steal(worker_id, hash_seed, obj);
681 }
682
683 ConcurrentMark(G1CollectedHeap* g1h,
684 G1RegionToSpaceMapper* prev_bitmap_storage,
685 G1RegionToSpaceMapper* next_bitmap_storage);
686 ~ConcurrentMark();
687
688 ConcurrentMarkThread* cmThread() { return _cmThread; }
689
690 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
691 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
692
693 // Returns the number of GC threads to be used in a concurrent
694 // phase based on the number of GC threads being used in a STW
695 // phase.
696 uint scale_parallel_threads(uint n_par_threads);
697
698 // Calculates the number of GC threads to be used in a concurrent phase.
699 uint calc_parallel_marking_threads();
700
701 // The following three are interaction between CM and
702 // G1CollectedHeap
703
704 // This notifies CM that a root during initial-mark needs to be
705 // grayed. It is MT-safe. word_size is the size of the object in
706 // words. It is passed explicitly as sometimes we cannot calculate
707 // it from the given object because it might be in an inconsistent
708 // state (e.g., in to-space and being copied). So the caller is
709 // responsible for dealing with this issue (e.g., get the size from
710 // the from-space image when the to-space image might be
711 // inconsistent) and always passing the size. hr is the region that
712 // contains the object and it's passed optionally from callers who
713 // might already have it (no point in recalculating it).
714 inline void grayRoot(oop obj,
715 size_t word_size,
716 uint worker_id,
717 HeapRegion* hr = NULL);
718
719 // It iterates over the heap and for each object it comes across it
720 // will dump the contents of its reference fields, as well as
721 // liveness information for the object and its referents. The dump
722 // will be written to a file with the following name:
723 // G1PrintReachableBaseFile + "." + str.
724 // vo decides whether the prev (vo == UsePrevMarking), the next
725 // (vo == UseNextMarking) marking information, or the mark word
726 // (vo == UseMarkWord) will be used to determine the liveness of
727 // each object / referent.
728 // If all is true, all objects in the heap will be dumped, otherwise
729 // only the live ones. In the dump the following symbols / breviations
730 // are used:
731 // M : an explicitly live object (its bitmap bit is set)
732 // > : an implicitly live object (over tams)
733 // O : an object outside the G1 heap (typically: in the perm gen)
734 // NOT : a reference field whose referent is not live
735 // AND MARKED : indicates that an object is both explicitly and
736 // implicitly live (it should be one or the other, not both)
737 void print_reachable(const char* str,
738 VerifyOption vo,
739 bool all) PRODUCT_RETURN;
740
741 // Clear the next marking bitmap (will be called concurrently).
742 void clearNextBitmap();
743
744 // Return whether the next mark bitmap has no marks set. To be used for assertions
745 // only. Will not yield to pause requests.
746 bool nextMarkBitmapIsClear();
747
748 // These two do the work that needs to be done before and after the
749 // initial root checkpoint. Since this checkpoint can be done at two
750 // different points (i.e. an explicit pause or piggy-backed on a
751 // young collection), then it's nice to be able to easily share the
752 // pre/post code. It might be the case that we can put everything in
753 // the post method. TP
754 void checkpointRootsInitialPre();
755 void checkpointRootsInitialPost();
756
757 // Scan all the root regions and mark everything reachable from
758 // them.
759 void scanRootRegions();
760
761 // Scan a single root region and mark everything reachable from it.
762 void scanRootRegion(HeapRegion* hr, uint worker_id);
763
764 // Do concurrent phase of marking, to a tentative transitive closure.
765 void markFromRoots();
766
767 void checkpointRootsFinal(bool clear_all_soft_refs);
768 void checkpointRootsFinalWork();
769 void cleanup();
770 void completeCleanup();
771
772 // Mark in the previous bitmap. NB: this is usually read-only, so use
773 // this carefully!
774 inline void markPrev(oop p);
775
776 // Clears marks for all objects in the given range, for the prev or
777 // next bitmaps. NB: the previous bitmap is usually
778 // read-only, so use this carefully!
779 void clearRangePrevBitmap(MemRegion mr);
780 void clearRangeNextBitmap(MemRegion mr);
781
782 // Notify data structures that a GC has started.
783 void note_start_of_gc() {
784 _markStack.note_start_of_gc();
785 }
786
787 // Notify data structures that a GC is finished.
788 void note_end_of_gc() {
789 _markStack.note_end_of_gc();
790 }
791
792 // Verify that there are no CSet oops on the stacks (taskqueues /
793 // global mark stack), enqueued SATB buffers, per-thread SATB
794 // buffers, and fingers (global / per-task). The boolean parameters
795 // decide which of the above data structures to verify. If marking
796 // is not in progress, it's a no-op.
797 void verify_no_cset_oops(bool verify_stacks,
798 bool verify_enqueued_buffers,
799 bool verify_thread_buffers,
800 bool verify_fingers) PRODUCT_RETURN;
801
802 bool isPrevMarked(oop p) const {
803 assert(p != NULL && p->is_oop(), "expected an oop");
804 HeapWord* addr = (HeapWord*)p;
805 assert(addr >= _prevMarkBitMap->startWord() ||
806 addr < _prevMarkBitMap->endWord(), "in a region");
807
808 return _prevMarkBitMap->isMarked(addr);
809 }
810
811 inline bool do_yield_check(uint worker_i = 0);
812
813 // Called to abort the marking cycle after a Full GC takes place.
814 void abort();
815
816 bool has_aborted() { return _has_aborted; }
817
818 const GCId& concurrent_gc_id();
819
820 // This prints the global/local fingers. It is used for debugging.
821 NOT_PRODUCT(void print_finger();)
822
823 void print_summary_info();
824
825 void print_worker_threads_on(outputStream* st) const;
826
827 void print_on_error(outputStream* st) const;
828
829 // The following indicate whether a given verbose level has been
830 // set. Notice that anything above stats is conditional to
831 // _MARKING_VERBOSE_ having been set to 1
832 bool verbose_stats() {
833 return _verbose_level >= stats_verbose;
834 }
835 bool verbose_low() {
836 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
837 }
838 bool verbose_medium() {
839 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
840 }
841 bool verbose_high() {
842 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
843 }
844
845 // Liveness counting
846
847 // Utility routine to set an exclusive range of cards on the given
848 // card liveness bitmap
849 inline void set_card_bitmap_range(BitMap* card_bm,
850 BitMap::idx_t start_idx,
851 BitMap::idx_t end_idx,
852 bool is_par);
853
854 // Returns the card number of the bottom of the G1 heap.
855 // Used in biasing indices into accounting card bitmaps.
856 intptr_t heap_bottom_card_num() const {
857 return _heap_bottom_card_num;
858 }
859
860 // Returns the card bitmap for a given task or worker id.
861 BitMap* count_card_bitmap_for(uint worker_id) {
862 assert(worker_id < _max_worker_id, "oob");
863 assert(_count_card_bitmaps != NULL, "uninitialized");
864 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
865 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
866 return task_card_bm;
867 }
868
869 // Returns the array containing the marked bytes for each region,
870 // for the given worker or task id.
871 size_t* count_marked_bytes_array_for(uint worker_id) {
872 assert(worker_id < _max_worker_id, "oob");
873 assert(_count_marked_bytes != NULL, "uninitialized");
874 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
875 assert(marked_bytes_array != NULL, "uninitialized");
876 return marked_bytes_array;
877 }
878
879 // Returns the index in the liveness accounting card table bitmap
880 // for the given address
881 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
882
883 // Counts the size of the given memory region in the the given
884 // marked_bytes array slot for the given HeapRegion.
885 // Sets the bits in the given card bitmap that are associated with the
886 // cards that are spanned by the memory region.
887 inline void count_region(MemRegion mr,
888 HeapRegion* hr,
889 size_t* marked_bytes_array,
890 BitMap* task_card_bm);
891
892 // Counts the given memory region in the task/worker counting
893 // data structures for the given worker id.
894 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
895
896 // Counts the given object in the given task/worker counting
897 // data structures.
898 inline void count_object(oop obj,
899 HeapRegion* hr,
900 size_t* marked_bytes_array,
901 BitMap* task_card_bm);
902
903 // Attempts to mark the given object and, if successful, counts
904 // the object in the given task/worker counting structures.
905 inline bool par_mark_and_count(oop obj,
906 HeapRegion* hr,
907 size_t* marked_bytes_array,
908 BitMap* task_card_bm);
909
910 // Attempts to mark the given object and, if successful, counts
911 // the object in the task/worker counting structures for the
912 // given worker id.
913 inline bool par_mark_and_count(oop obj,
914 size_t word_size,
915 HeapRegion* hr,
916 uint worker_id);
917
918 // Returns true if initialization was successfully completed.
919 bool completed_initialization() const {
920 return _completed_initialization;
921 }
922
923 protected:
924 // Clear all the per-task bitmaps and arrays used to store the
925 // counting data.
926 void clear_all_count_data();
927
928 // Aggregates the counting data for each worker/task
929 // that was constructed while marking. Also sets
930 // the amount of marked bytes for each region and
931 // the top at concurrent mark count.
932 void aggregate_count_data();
933
934 // Verification routine
935 void verify_count_data();
936 };
937
938 // A class representing a marking task.
939 class CMTask : public TerminatorTerminator {
940 private:
941 enum PrivateConstants {
942 // the regular clock call is called once the scanned words reaches
943 // this limit
944 words_scanned_period = 12*1024,
945 // the regular clock call is called once the number of visited
946 // references reaches this limit
947 refs_reached_period = 384,
948 // initial value for the hash seed, used in the work stealing code
949 init_hash_seed = 17,
950 // how many entries will be transferred between global stack and
951 // local queues
952 global_stack_transfer_size = 16
953 };
954
955 uint _worker_id;
956 G1CollectedHeap* _g1h;
957 ConcurrentMark* _cm;
958 CMBitMap* _nextMarkBitMap;
959 // the task queue of this task
960 CMTaskQueue* _task_queue;
961 private:
962 // the task queue set---needed for stealing
963 CMTaskQueueSet* _task_queues;
964 // indicates whether the task has been claimed---this is only for
965 // debugging purposes
966 bool _claimed;
967
968 // number of calls to this task
969 int _calls;
970
971 // when the virtual timer reaches this time, the marking step should
972 // exit
973 double _time_target_ms;
974 // the start time of the current marking step
975 double _start_time_ms;
976
977 // the oop closure used for iterations over oops
978 G1CMOopClosure* _cm_oop_closure;
979
980 // the region this task is scanning, NULL if we're not scanning any
981 HeapRegion* _curr_region;
982 // the local finger of this task, NULL if we're not scanning a region
983 HeapWord* _finger;
984 // limit of the region this task is scanning, NULL if we're not scanning one
985 HeapWord* _region_limit;
986
987 // the number of words this task has scanned
988 size_t _words_scanned;
989 // When _words_scanned reaches this limit, the regular clock is
990 // called. Notice that this might be decreased under certain
991 // circumstances (i.e. when we believe that we did an expensive
992 // operation).
993 size_t _words_scanned_limit;
994 // the initial value of _words_scanned_limit (i.e. what it was
995 // before it was decreased).
996 size_t _real_words_scanned_limit;
997
998 // the number of references this task has visited
999 size_t _refs_reached;
1000 // When _refs_reached reaches this limit, the regular clock is
1001 // called. Notice this this might be decreased under certain
1002 // circumstances (i.e. when we believe that we did an expensive
1003 // operation).
1004 size_t _refs_reached_limit;
1005 // the initial value of _refs_reached_limit (i.e. what it was before
1006 // it was decreased).
1007 size_t _real_refs_reached_limit;
1008
1009 // used by the work stealing stuff
1010 int _hash_seed;
1011 // if this is true, then the task has aborted for some reason
1012 bool _has_aborted;
1013 // set when the task aborts because it has met its time quota
1014 bool _has_timed_out;
1015 // true when we're draining SATB buffers; this avoids the task
1016 // aborting due to SATB buffers being available (as we're already
1017 // dealing with them)
1018 bool _draining_satb_buffers;
1019
1020 // number sequence of past step times
1021 NumberSeq _step_times_ms;
1022 // elapsed time of this task
1023 double _elapsed_time_ms;
1024 // termination time of this task
1025 double _termination_time_ms;
1026 // when this task got into the termination protocol
1027 double _termination_start_time_ms;
1028
1029 // true when the task is during a concurrent phase, false when it is
1030 // in the remark phase (so, in the latter case, we do not have to
1031 // check all the things that we have to check during the concurrent
1032 // phase, i.e. SATB buffer availability...)
1033 bool _concurrent;
1034
1035 TruncatedSeq _marking_step_diffs_ms;
1036
1037 // Counting data structures. Embedding the task's marked_bytes_array
1038 // and card bitmap into the actual task saves having to go through
1039 // the ConcurrentMark object.
1040 size_t* _marked_bytes_array;
1041 BitMap* _card_bm;
1042
1043 // LOTS of statistics related with this task
1044 #if _MARKING_STATS_
1045 NumberSeq _all_clock_intervals_ms;
1046 double _interval_start_time_ms;
1047
1048 size_t _aborted;
1049 size_t _aborted_overflow;
1050 size_t _aborted_cm_aborted;
1051 size_t _aborted_yield;
1052 size_t _aborted_timed_out;
1053 size_t _aborted_satb;
1054 size_t _aborted_termination;
1055
1056 size_t _steal_attempts;
1057 size_t _steals;
1058
1059 size_t _clock_due_to_marking;
1060 size_t _clock_due_to_scanning;
1061
1062 size_t _local_pushes;
1063 size_t _local_pops;
1064 size_t _local_max_size;
1065 size_t _objs_scanned;
1066
1067 size_t _global_pushes;
1068 size_t _global_pops;
1069 size_t _global_max_size;
1070
1071 size_t _global_transfers_to;
1072 size_t _global_transfers_from;
1073
1074 size_t _regions_claimed;
1075 size_t _objs_found_on_bitmap;
1076
1077 size_t _satb_buffers_processed;
1078 #endif // _MARKING_STATS_
1079
1080 // it updates the local fields after this task has claimed
1081 // a new region to scan
1082 void setup_for_region(HeapRegion* hr);
1083 // it brings up-to-date the limit of the region
1084 void update_region_limit();
1085
1086 // called when either the words scanned or the refs visited limit
1087 // has been reached
1088 void reached_limit();
1089 // recalculates the words scanned and refs visited limits
1090 void recalculate_limits();
1091 // decreases the words scanned and refs visited limits when we reach
1092 // an expensive operation
1093 void decrease_limits();
1094 // it checks whether the words scanned or refs visited reached their
1095 // respective limit and calls reached_limit() if they have
1096 void check_limits() {
1097 if (_words_scanned >= _words_scanned_limit ||
1098 _refs_reached >= _refs_reached_limit) {
1099 reached_limit();
1100 }
1101 }
1102 // this is supposed to be called regularly during a marking step as
1103 // it checks a bunch of conditions that might cause the marking step
1104 // to abort
1105 void regular_clock_call();
1106 bool concurrent() { return _concurrent; }
1107
1108 // Test whether objAddr might have already been passed over by the
1109 // mark bitmap scan, and so needs to be pushed onto the mark stack.
1110 bool is_below_finger(HeapWord* objAddr, HeapWord* global_finger) const;
1111
1112 template<bool scan> void process_grey_object(oop obj);
1113
1114 public:
1115 // It resets the task; it should be called right at the beginning of
1116 // a marking phase.
1117 void reset(CMBitMap* _nextMarkBitMap);
1118 // it clears all the fields that correspond to a claimed region.
1119 void clear_region_fields();
1120
1121 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1122
1123 // The main method of this class which performs a marking step
1124 // trying not to exceed the given duration. However, it might exit
1125 // prematurely, according to some conditions (i.e. SATB buffers are
1126 // available for processing).
1127 void do_marking_step(double target_ms,
1128 bool do_termination,
1129 bool is_serial);
1130
1131 // These two calls start and stop the timer
1132 void record_start_time() {
1133 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1134 }
1135 void record_end_time() {
1136 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1137 }
1138
1139 // returns the worker ID associated with this task.
1140 uint worker_id() { return _worker_id; }
1141
1142 // From TerminatorTerminator. It determines whether this task should
1143 // exit the termination protocol after it's entered it.
1144 virtual bool should_exit_termination();
1145
1146 // Resets the local region fields after a task has finished scanning a
1147 // region; or when they have become stale as a result of the region
1148 // being evacuated.
1149 void giveup_current_region();
1150
1151 HeapWord* finger() { return _finger; }
1152
1153 bool has_aborted() { return _has_aborted; }
1154 void set_has_aborted() { _has_aborted = true; }
1155 void clear_has_aborted() { _has_aborted = false; }
1156 bool has_timed_out() { return _has_timed_out; }
1157 bool claimed() { return _claimed; }
1158
1159 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1160
1161 // It grays the object by marking it and, if necessary, pushing it
1162 // on the local queue
1163 inline void deal_with_reference(oop obj);
1164
1165 // It scans an object and visits its children.
1166 void scan_object(oop obj) { process_grey_object<true>(obj); }
1167
1168 // It pushes an object on the local queue.
1169 inline void push(oop obj);
1170
1171 // These two move entries to/from the global stack.
1172 void move_entries_to_global_stack();
1173 void get_entries_from_global_stack();
1174
1175 // It pops and scans objects from the local queue. If partially is
1176 // true, then it stops when the queue size is of a given limit. If
1177 // partially is false, then it stops when the queue is empty.
1178 void drain_local_queue(bool partially);
1179 // It moves entries from the global stack to the local queue and
1180 // drains the local queue. If partially is true, then it stops when
1181 // both the global stack and the local queue reach a given size. If
1182 // partially if false, it tries to empty them totally.
1183 void drain_global_stack(bool partially);
1184 // It keeps picking SATB buffers and processing them until no SATB
1185 // buffers are available.
1186 void drain_satb_buffers();
1187
1188 // moves the local finger to a new location
1189 inline void move_finger_to(HeapWord* new_finger) {
1190 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1191 _finger = new_finger;
1192 }
1193
1194 CMTask(uint worker_id,
1195 ConcurrentMark *cm,
1196 size_t* marked_bytes,
1197 BitMap* card_bm,
1198 CMTaskQueue* task_queue,
1199 CMTaskQueueSet* task_queues);
1200
1201 // it prints statistics associated with this task
1202 void print_stats();
1203
1204 #if _MARKING_STATS_
1205 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1206 #endif // _MARKING_STATS_
1207 };
1208
1209 // Class that's used to to print out per-region liveness
1210 // information. It's currently used at the end of marking and also
1211 // after we sort the old regions at the end of the cleanup operation.
1212 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1213 private:
1214 outputStream* _out;
1215
1216 // Accumulators for these values.
1217 size_t _total_used_bytes;
1218 size_t _total_capacity_bytes;
1219 size_t _total_prev_live_bytes;
1220 size_t _total_next_live_bytes;
1221
1222 // These are set up when we come across a "stars humongous" region
1223 // (as this is where most of this information is stored, not in the
1224 // subsequent "continues humongous" regions). After that, for every
1225 // region in a given humongous region series we deduce the right
1226 // values for it by simply subtracting the appropriate amount from
1227 // these fields. All these values should reach 0 after we've visited
1228 // the last region in the series.
1229 size_t _hum_used_bytes;
1230 size_t _hum_capacity_bytes;
1231 size_t _hum_prev_live_bytes;
1232 size_t _hum_next_live_bytes;
1233
1234 // Accumulator for the remembered set size
1235 size_t _total_remset_bytes;
1236
1237 // Accumulator for strong code roots memory size
1238 size_t _total_strong_code_roots_bytes;
1239
1240 static double perc(size_t val, size_t total) {
1241 if (total == 0) {
1242 return 0.0;
1243 } else {
1244 return 100.0 * ((double) val / (double) total);
1245 }
1246 }
1247
1248 static double bytes_to_mb(size_t val) {
1249 return (double) val / (double) M;
1250 }
1251
1252 // See the .cpp file.
1253 size_t get_hum_bytes(size_t* hum_bytes);
1254 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1255 size_t* prev_live_bytes, size_t* next_live_bytes);
1256
1257 public:
1258 // The header and footer are printed in the constructor and
1259 // destructor respectively.
1260 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1261 virtual bool doHeapRegion(HeapRegion* r);
1262 ~G1PrintRegionLivenessInfoClosure();
1263 };
1264
1265 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP