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