1 /*
2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP
26 #define SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP
27
28 #include "classfile/javaClasses.hpp"
29 #include "gc/g1/g1RegionToSpaceMapper.hpp"
30 #include "gc/g1/heapRegionSet.hpp"
31 #include "gc/shared/taskqueue.hpp"
32
33 class G1CollectedHeap;
34 class G1CMBitMap;
35 class G1CMTask;
36 class G1ConcurrentMark;
37 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue;
38 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
39
40 // Closure used by CM during concurrent reference discovery
41 // and reference processing (during remarking) to determine
42 // if a particular object is alive. It is primarily used
43 // to determine if referents of discovered reference objects
44 // are alive. An instance is also embedded into the
45 // reference processor as the _is_alive_non_header field
46 class G1CMIsAliveClosure: public BoolObjectClosure {
47 G1CollectedHeap* _g1;
48 public:
49 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
50
51 bool do_object_b(oop obj);
52 };
53
54 // A generic CM bit map. This is essentially a wrapper around the BitMap
55 // class, with one bit per (1<<_shifter) HeapWords.
56
57 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
58 protected:
59 HeapWord* _bmStartWord; // base address of range covered by map
60 size_t _bmWordSize; // map size (in #HeapWords covered)
61 const int _shifter; // map to char or bit
62 BitMap _bm; // the bit map itself
63
64 public:
65 // constructor
66 G1CMBitMapRO(int shifter);
67
68 // inquiries
69 HeapWord* startWord() const { return _bmStartWord; }
70 // the following is one past the last word in space
71 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
72
73 // read marks
74
75 bool isMarked(HeapWord* addr) const {
76 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
77 "outside underlying space?");
78 return _bm.at(heapWordToOffset(addr));
79 }
80
81 // iteration
82 inline bool iterate(BitMapClosure* cl, MemRegion mr);
83
84 // Return the address corresponding to the next marked bit at or after
85 // "addr", and before "limit", if "limit" is non-NULL. If there is no
86 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
87 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
88 const HeapWord* limit = NULL) const;
89
90 // conversion utilities
91 HeapWord* offsetToHeapWord(size_t offset) const {
92 return _bmStartWord + (offset << _shifter);
93 }
94 size_t heapWordToOffset(const HeapWord* addr) const {
95 return pointer_delta(addr, _bmStartWord) >> _shifter;
96 }
97
98 // The argument addr should be the start address of a valid object
99 inline HeapWord* nextObject(HeapWord* addr);
100
101 void print_on_error(outputStream* st, const char* prefix) const;
102
103 // debugging
104 NOT_PRODUCT(bool covers(MemRegion rs) const;)
105 };
106
107 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
108 private:
109 G1CMBitMap* _bm;
110 public:
111 G1CMBitMapMappingChangedListener() : _bm(NULL) {}
112
113 void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
114
115 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
116 };
117
118 class G1CMBitMap : public G1CMBitMapRO {
119 private:
120 G1CMBitMapMappingChangedListener _listener;
121
122 public:
123 static size_t compute_size(size_t heap_size);
124 // Returns the amount of bytes on the heap between two marks in the bitmap.
125 static size_t mark_distance();
126 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
127 // mark bitmap corresponds to. This is the same as the mark distance above.
128 static size_t heap_map_factor() {
129 return mark_distance();
130 }
131
132 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
133
134 // Initializes the underlying BitMap to cover the given area.
135 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
136
137 // Write marks.
138 inline void mark(HeapWord* addr);
139 inline void clear(HeapWord* addr);
140 inline bool parMark(HeapWord* addr);
141
142 void clearRange(MemRegion mr);
143
144 // Clear the whole mark bitmap.
145 void clearAll();
146 };
147
148 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
149 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
150 VirtualSpace _virtual_space; // Underlying backing store for actual stack
151 G1ConcurrentMark* _cm;
152 oop* _base; // bottom of stack
153 jint _index; // one more than last occupied index
154 jint _capacity; // max #elements
155 jint _saved_index; // value of _index saved at start of GC
156
157 bool _overflow;
158 bool _should_expand;
159
160 public:
161 G1CMMarkStack(G1ConcurrentMark* cm);
162 ~G1CMMarkStack();
163
164 bool allocate(size_t capacity);
165
166 // Pushes the first "n" elements of "ptr_arr" on the stack.
167 // Locking impl: concurrency is allowed only with
168 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
169 // locking strategy.
170 void par_push_arr(oop* ptr_arr, int n);
171
172 // If returns false, the array was empty. Otherwise, removes up to "max"
173 // elements from the stack, and transfers them to "ptr_arr" in an
174 // unspecified order. The actual number transferred is given in "n" ("n
175 // == 0" is deliberately redundant with the return value.) Locking impl:
176 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
177 // operations, which use the same locking strategy.
178 bool par_pop_arr(oop* ptr_arr, int max, int* n);
179
180 bool isEmpty() { return _index == 0; }
181 int maxElems() { return _capacity; }
182
183 bool overflow() { return _overflow; }
184 void clear_overflow() { _overflow = false; }
185
186 bool should_expand() const { return _should_expand; }
187 void set_should_expand();
188
189 // Expand the stack, typically in response to an overflow condition
190 void expand();
191
192 int size() { return _index; }
193
194 void setEmpty() { _index = 0; clear_overflow(); }
195
196 // Record the current index.
197 void note_start_of_gc();
198
199 // Make sure that we have not added any entries to the stack during GC.
200 void note_end_of_gc();
201
202 // Apply fn to each oop in the mark stack, up to the bound recorded
203 // via one of the above "note" functions. The mark stack must not
204 // be modified while iterating.
205 template<typename Fn> void iterate(Fn fn);
206 };
207
208 class YoungList;
209
210 // Root Regions are regions that are not empty at the beginning of a
211 // marking cycle and which we might collect during an evacuation pause
212 // while the cycle is active. Given that, during evacuation pauses, we
213 // do not copy objects that are explicitly marked, what we have to do
214 // for the root regions is to scan them and mark all objects reachable
215 // from them. According to the SATB assumptions, we only need to visit
216 // each object once during marking. So, as long as we finish this scan
217 // before the next evacuation pause, we can copy the objects from the
218 // root regions without having to mark them or do anything else to them.
219 //
220 // Currently, we only support root region scanning once (at the start
221 // of the marking cycle) and the root regions are all the survivor
222 // regions populated during the initial-mark pause.
223 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
224 private:
225 YoungList* _young_list;
226 G1ConcurrentMark* _cm;
227
228 volatile bool _scan_in_progress;
229 volatile bool _should_abort;
230 HeapRegion* volatile _next_survivor;
231
232 public:
233 G1CMRootRegions();
234 // We actually do most of the initialization in this method.
235 void init(G1CollectedHeap* g1h, G1ConcurrentMark* cm);
236
237 // Reset the claiming / scanning of the root regions.
238 void prepare_for_scan();
239
240 // Forces get_next() to return NULL so that the iteration aborts early.
241 void abort() { _should_abort = true; }
242
243 // Return true if the CM thread are actively scanning root regions,
244 // false otherwise.
245 bool scan_in_progress() { return _scan_in_progress; }
246
247 // Claim the next root region to scan atomically, or return NULL if
248 // all have been claimed.
249 HeapRegion* claim_next();
250
251 void cancel_scan();
252
253 // Flag that we're done with root region scanning and notify anyone
254 // who's waiting on it. If aborted is false, assume that all regions
255 // have been claimed.
256 void scan_finished();
257
258 // If CM threads are still scanning root regions, wait until they
259 // are done. Return true if we had to wait, false otherwise.
260 bool wait_until_scan_finished();
261 };
262
263 class ConcurrentMarkThread;
264
265 class G1ConcurrentMark: public CHeapObj<mtGC> {
266 friend class ConcurrentMarkThread;
267 friend class G1ParNoteEndTask;
268 friend class CalcLiveObjectsClosure;
269 friend class G1CMRefProcTaskProxy;
270 friend class G1CMRefProcTaskExecutor;
271 friend class G1CMKeepAliveAndDrainClosure;
272 friend class G1CMDrainMarkingStackClosure;
273 friend class G1CMBitMapClosure;
274 friend class G1CMConcurrentMarkingTask;
275 friend class G1CMMarkStack;
276 friend class G1CMRemarkTask;
277 friend class G1CMTask;
278
279 protected:
280 ConcurrentMarkThread* _cmThread; // The thread doing the work
281 G1CollectedHeap* _g1h; // The heap
282 uint _parallel_marking_threads; // The number of marking
283 // threads we're using
284 uint _max_parallel_marking_threads; // Max number of marking
285 // threads we'll ever use
286 double _sleep_factor; // How much we have to sleep, with
287 // respect to the work we just did, to
288 // meet the marking overhead goal
289 double _marking_task_overhead; // Marking target overhead for
290 // a single task
291
292 FreeRegionList _cleanup_list;
293
294 // Concurrent marking support structures
295 G1CMBitMap _markBitMap1;
296 G1CMBitMap _markBitMap2;
297 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
298 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
299
300 BitMap _region_bm;
301 BitMap _card_bm;
302
303 // Heap bounds
304 HeapWord* _heap_start;
305 HeapWord* _heap_end;
306
307 // Root region tracking and claiming
308 G1CMRootRegions _root_regions;
309
310 // For gray objects
311 G1CMMarkStack _markStack; // Grey objects behind global finger
312 HeapWord* volatile _finger; // The global finger, region aligned,
313 // always points to the end of the
314 // last claimed region
315
316 // Marking tasks
317 uint _max_worker_id;// Maximum worker id
318 uint _active_tasks; // Task num currently active
319 G1CMTask** _tasks; // Task queue array (max_worker_id len)
320 G1CMTaskQueueSet* _task_queues; // Task queue set
321 ParallelTaskTerminator _terminator; // For termination
322
323 // Two sync barriers that are used to synchronize tasks when an
324 // overflow occurs. The algorithm is the following. All tasks enter
325 // the first one to ensure that they have all stopped manipulating
326 // the global data structures. After they exit it, they re-initialize
327 // their data structures and task 0 re-initializes the global data
328 // structures. Then, they enter the second sync barrier. This
329 // ensure, that no task starts doing work before all data
330 // structures (local and global) have been re-initialized. When they
331 // exit it, they are free to start working again.
332 WorkGangBarrierSync _first_overflow_barrier_sync;
333 WorkGangBarrierSync _second_overflow_barrier_sync;
334
335 // This is set by any task, when an overflow on the global data
336 // structures is detected
337 volatile bool _has_overflown;
338 // True: marking is concurrent, false: we're in remark
339 volatile bool _concurrent;
340 // Set at the end of a Full GC so that marking aborts
341 volatile bool _has_aborted;
342
343 // Used when remark aborts due to an overflow to indicate that
344 // another concurrent marking phase should start
345 volatile bool _restart_for_overflow;
346
347 // This is true from the very start of concurrent marking until the
348 // point when all the tasks complete their work. It is really used
349 // to determine the points between the end of concurrent marking and
350 // time of remark.
351 volatile bool _concurrent_marking_in_progress;
352
353 // Keep track of whether we have started concurrent phase or not.
354 bool _concurrent_phase_started;
355
356 // All of these times are in ms
357 NumberSeq _init_times;
358 NumberSeq _remark_times;
359 NumberSeq _remark_mark_times;
360 NumberSeq _remark_weak_ref_times;
361 NumberSeq _cleanup_times;
362 double _total_counting_time;
363 double _total_rs_scrub_time;
364
365 double* _accum_task_vtime; // Accumulated task vtime
366
367 WorkGang* _parallel_workers;
368
369 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
370 void weakRefsWork(bool clear_all_soft_refs);
371
372 void swapMarkBitMaps();
373
374 // It resets the global marking data structures, as well as the
375 // task local ones; should be called during initial mark.
376 void reset();
377
378 // Resets all the marking data structures. Called when we have to restart
379 // marking or when marking completes (via set_non_marking_state below).
380 void reset_marking_state(bool clear_overflow = true);
381
382 // We do this after we're done with marking so that the marking data
383 // structures are initialized to a sensible and predictable state.
384 void set_non_marking_state();
385
386 // Called to indicate how many threads are currently active.
387 void set_concurrency(uint active_tasks);
388
389 // It should be called to indicate which phase we're in (concurrent
390 // mark or remark) and how many threads are currently active.
391 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
392
393 // Prints all gathered CM-related statistics
394 void print_stats();
395
396 bool cleanup_list_is_empty() {
397 return _cleanup_list.is_empty();
398 }
399
400 // Accessor methods
401 uint parallel_marking_threads() const { return _parallel_marking_threads; }
402 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
403 double sleep_factor() { return _sleep_factor; }
404 double marking_task_overhead() { return _marking_task_overhead;}
405
406 HeapWord* finger() { return _finger; }
407 bool concurrent() { return _concurrent; }
408 uint active_tasks() { return _active_tasks; }
409 ParallelTaskTerminator* terminator() { return &_terminator; }
410
411 // It claims the next available region to be scanned by a marking
412 // task/thread. It might return NULL if the next region is empty or
413 // we have run out of regions. In the latter case, out_of_regions()
414 // determines whether we've really run out of regions or the task
415 // should call claim_region() again. This might seem a bit
416 // awkward. Originally, the code was written so that claim_region()
417 // either successfully returned with a non-empty region or there
418 // were no more regions to be claimed. The problem with this was
419 // that, in certain circumstances, it iterated over large chunks of
420 // the heap finding only empty regions and, while it was working, it
421 // was preventing the calling task to call its regular clock
422 // method. So, this way, each task will spend very little time in
423 // claim_region() and is allowed to call the regular clock method
424 // frequently.
425 HeapRegion* claim_region(uint worker_id);
426
427 // It determines whether we've run out of regions to scan. Note that
428 // the finger can point past the heap end in case the heap was expanded
429 // to satisfy an allocation without doing a GC. This is fine, because all
430 // objects in those regions will be considered live anyway because of
431 // SATB guarantees (i.e. their TAMS will be equal to bottom).
432 bool out_of_regions() { return _finger >= _heap_end; }
433
434 // Returns the task with the given id
435 G1CMTask* task(int id) {
436 assert(0 <= id && id < (int) _active_tasks,
437 "task id not within active bounds");
438 return _tasks[id];
439 }
440
441 // Returns the task queue with the given id
442 G1CMTaskQueue* task_queue(int id) {
443 assert(0 <= id && id < (int) _active_tasks,
444 "task queue id not within active bounds");
445 return (G1CMTaskQueue*) _task_queues->queue(id);
446 }
447
448 // Returns the task queue set
449 G1CMTaskQueueSet* task_queues() { return _task_queues; }
450
451 // Access / manipulation of the overflow flag which is set to
452 // indicate that the global stack has overflown
453 bool has_overflown() { return _has_overflown; }
454 void set_has_overflown() { _has_overflown = true; }
455 void clear_has_overflown() { _has_overflown = false; }
456 bool restart_for_overflow() { return _restart_for_overflow; }
457
458 // Methods to enter the two overflow sync barriers
459 void enter_first_sync_barrier(uint worker_id);
460 void enter_second_sync_barrier(uint worker_id);
461
462 // Live Data Counting data structures...
463 // These data structures are initialized at the start of
464 // marking. They are written to while marking is active.
465 // They are aggregated during remark; the aggregated values
466 // are then used to populate the _region_bm, _card_bm, and
467 // the total live bytes, which are then subsequently updated
468 // during cleanup.
469
470 // An array of bitmaps (one bit map per task). Each bitmap
471 // is used to record the cards spanned by the live objects
472 // marked by that task/worker.
473 BitMap* _count_card_bitmaps;
474
475 // Used to record the number of marked live bytes
476 // (for each region, by worker thread).
477 size_t** _count_marked_bytes;
478
479 // Card index of the bottom of the G1 heap. Used for biasing indices into
480 // the card bitmaps.
481 intptr_t _heap_bottom_card_num;
482
483 // Set to true when initialization is complete
484 bool _completed_initialization;
485
486 public:
487 // Manipulation of the global mark stack.
488 // The push and pop operations are used by tasks for transfers
489 // between task-local queues and the global mark stack, and use
490 // locking for concurrency safety.
491 bool mark_stack_push(oop* arr, int n) {
492 _markStack.par_push_arr(arr, n);
493 if (_markStack.overflow()) {
494 set_has_overflown();
495 return false;
496 }
497 return true;
498 }
499 void mark_stack_pop(oop* arr, int max, int* n) {
500 _markStack.par_pop_arr(arr, max, n);
501 }
502 size_t mark_stack_size() { return _markStack.size(); }
503 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
504 bool mark_stack_overflow() { return _markStack.overflow(); }
505 bool mark_stack_empty() { return _markStack.isEmpty(); }
506
507 G1CMRootRegions* root_regions() { return &_root_regions; }
508
509 bool concurrent_marking_in_progress() {
510 return _concurrent_marking_in_progress;
511 }
512 void set_concurrent_marking_in_progress() {
513 _concurrent_marking_in_progress = true;
514 }
515 void clear_concurrent_marking_in_progress() {
516 _concurrent_marking_in_progress = false;
517 }
518
519 void register_concurrent_phase_start(const char* title);
520 void register_concurrent_phase_end();
521
522 void update_accum_task_vtime(int i, double vtime) {
523 _accum_task_vtime[i] += vtime;
524 }
525
526 double all_task_accum_vtime() {
527 double ret = 0.0;
528 for (uint i = 0; i < _max_worker_id; ++i)
529 ret += _accum_task_vtime[i];
530 return ret;
531 }
532
533 // Attempts to steal an object from the task queues of other tasks
534 bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
535
536 G1ConcurrentMark(G1CollectedHeap* g1h,
537 G1RegionToSpaceMapper* prev_bitmap_storage,
538 G1RegionToSpaceMapper* next_bitmap_storage);
539 ~G1ConcurrentMark();
540
541 ConcurrentMarkThread* cmThread() { return _cmThread; }
542
543 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
544 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
545
546 // Returns the number of GC threads to be used in a concurrent
547 // phase based on the number of GC threads being used in a STW
548 // phase.
549 uint scale_parallel_threads(uint n_par_threads);
550
551 // Calculates the number of GC threads to be used in a concurrent phase.
552 uint calc_parallel_marking_threads();
553
554 // The following three are interaction between CM and
555 // G1CollectedHeap
556
557 // This notifies CM that a root during initial-mark needs to be
558 // grayed. It is MT-safe. word_size is the size of the object in
559 // words. It is passed explicitly as sometimes we cannot calculate
560 // it from the given object because it might be in an inconsistent
561 // state (e.g., in to-space and being copied). So the caller is
562 // responsible for dealing with this issue (e.g., get the size from
563 // the from-space image when the to-space image might be
564 // inconsistent) and always passing the size. hr is the region that
565 // contains the object and it's passed optionally from callers who
566 // might already have it (no point in recalculating it).
567 inline void grayRoot(oop obj,
568 size_t word_size,
569 uint worker_id,
570 HeapRegion* hr = NULL);
571
572 // Clear the next marking bitmap (will be called concurrently).
573 void clearNextBitmap();
574
575 // Return whether the next mark bitmap has no marks set. To be used for assertions
576 // only. Will not yield to pause requests.
577 bool nextMarkBitmapIsClear();
578
579 // These two do the work that needs to be done before and after the
580 // initial root checkpoint. Since this checkpoint can be done at two
581 // different points (i.e. an explicit pause or piggy-backed on a
582 // young collection), then it's nice to be able to easily share the
583 // pre/post code. It might be the case that we can put everything in
584 // the post method. TP
585 void checkpointRootsInitialPre();
586 void checkpointRootsInitialPost();
587
588 // Scan all the root regions and mark everything reachable from
589 // them.
590 void scanRootRegions();
591
592 // Scan a single root region and mark everything reachable from it.
593 void scanRootRegion(HeapRegion* hr, uint worker_id);
594
595 // Do concurrent phase of marking, to a tentative transitive closure.
596 void markFromRoots();
597
598 void checkpointRootsFinal(bool clear_all_soft_refs);
599 void checkpointRootsFinalWork();
600 void cleanup();
601 void completeCleanup();
602
603 // Mark in the previous bitmap. NB: this is usually read-only, so use
604 // this carefully!
605 inline void markPrev(oop p);
606
607 // Clears marks for all objects in the given range, for the prev or
608 // next bitmaps. NB: the previous bitmap is usually
609 // read-only, so use this carefully!
610 void clearRangePrevBitmap(MemRegion mr);
611
612 // Notify data structures that a GC has started.
613 void note_start_of_gc() {
614 _markStack.note_start_of_gc();
615 }
616
617 // Notify data structures that a GC is finished.
618 void note_end_of_gc() {
619 _markStack.note_end_of_gc();
620 }
621
622 // Verify that there are no CSet oops on the stacks (taskqueues /
623 // global mark stack) and fingers (global / per-task).
624 // If marking is not in progress, it's a no-op.
625 void verify_no_cset_oops() PRODUCT_RETURN;
626
627 inline bool isPrevMarked(oop p) const;
628
629 inline bool do_yield_check(uint worker_i = 0);
630
631 // Called to abort the marking cycle after a Full GC takes place.
632 void abort();
633
634 bool has_aborted() { return _has_aborted; }
635
636 void print_summary_info();
637
638 void print_worker_threads_on(outputStream* st) const;
639
640 void print_on_error(outputStream* st) const;
641
642 // Liveness counting
643
644 // Utility routine to set an exclusive range of cards on the given
645 // card liveness bitmap
646 inline void set_card_bitmap_range(BitMap* card_bm,
647 BitMap::idx_t start_idx,
648 BitMap::idx_t end_idx,
649 bool is_par);
650
651 // Returns the card number of the bottom of the G1 heap.
652 // Used in biasing indices into accounting card bitmaps.
653 intptr_t heap_bottom_card_num() const {
654 return _heap_bottom_card_num;
655 }
656
657 // Returns the card bitmap for a given task or worker id.
658 BitMap* count_card_bitmap_for(uint worker_id) {
659 assert(worker_id < _max_worker_id, "oob");
660 assert(_count_card_bitmaps != NULL, "uninitialized");
661 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
662 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
663 return task_card_bm;
664 }
665
666 // Returns the array containing the marked bytes for each region,
667 // for the given worker or task id.
668 size_t* count_marked_bytes_array_for(uint worker_id) {
669 assert(worker_id < _max_worker_id, "oob");
670 assert(_count_marked_bytes != NULL, "uninitialized");
671 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
672 assert(marked_bytes_array != NULL, "uninitialized");
673 return marked_bytes_array;
674 }
675
676 // Returns the index in the liveness accounting card table bitmap
677 // for the given address
678 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
679
680 // Counts the size of the given memory region in the the given
681 // marked_bytes array slot for the given HeapRegion.
682 // Sets the bits in the given card bitmap that are associated with the
683 // cards that are spanned by the memory region.
684 inline void count_region(MemRegion mr,
685 HeapRegion* hr,
686 size_t* marked_bytes_array,
687 BitMap* task_card_bm);
688
689 // Counts the given object in the given task/worker counting
690 // data structures.
691 inline void count_object(oop obj,
692 HeapRegion* hr,
693 size_t* marked_bytes_array,
694 BitMap* task_card_bm,
695 size_t word_size);
696
697 // Attempts to mark the given object and, if successful, counts
698 // the object in the given task/worker counting structures.
699 inline bool par_mark_and_count(oop obj,
700 HeapRegion* hr,
701 size_t* marked_bytes_array,
702 BitMap* task_card_bm);
703
704 // Attempts to mark the given object and, if successful, counts
705 // the object in the task/worker counting structures for the
706 // given worker id.
707 inline bool par_mark_and_count(oop obj,
708 size_t word_size,
709 HeapRegion* hr,
710 uint worker_id);
711
712 // Returns true if initialization was successfully completed.
713 bool completed_initialization() const {
714 return _completed_initialization;
715 }
716
717 protected:
718 // Clear all the per-task bitmaps and arrays used to store the
719 // counting data.
720 void clear_all_count_data();
721
722 // Aggregates the counting data for each worker/task
723 // that was constructed while marking. Also sets
724 // the amount of marked bytes for each region and
725 // the top at concurrent mark count.
726 void aggregate_count_data();
727
728 // Verification routine
729 void verify_count_data();
730 };
731
732 // A class representing a marking task.
733 class G1CMTask : public TerminatorTerminator {
734 private:
735 enum PrivateConstants {
736 // the regular clock call is called once the scanned words reaches
737 // this limit
738 words_scanned_period = 12*1024,
739 // the regular clock call is called once the number of visited
740 // references reaches this limit
741 refs_reached_period = 384,
742 // initial value for the hash seed, used in the work stealing code
743 init_hash_seed = 17,
744 // how many entries will be transferred between global stack and
745 // local queues
746 global_stack_transfer_size = 16
747 };
748
749 uint _worker_id;
750 G1CollectedHeap* _g1h;
751 G1ConcurrentMark* _cm;
752 G1CMBitMap* _nextMarkBitMap;
753 // the task queue of this task
754 G1CMTaskQueue* _task_queue;
755 private:
756 // the task queue set---needed for stealing
757 G1CMTaskQueueSet* _task_queues;
758 // indicates whether the task has been claimed---this is only for
759 // debugging purposes
760 bool _claimed;
761
762 // number of calls to this task
763 int _calls;
764
765 // when the virtual timer reaches this time, the marking step should
766 // exit
767 double _time_target_ms;
768 // the start time of the current marking step
769 double _start_time_ms;
770
771 // the oop closure used for iterations over oops
772 G1CMOopClosure* _cm_oop_closure;
773
774 // the region this task is scanning, NULL if we're not scanning any
775 HeapRegion* _curr_region;
776 // the local finger of this task, NULL if we're not scanning a region
777 HeapWord* _finger;
778 // limit of the region this task is scanning, NULL if we're not scanning one
779 HeapWord* _region_limit;
780
781 // the number of words this task has scanned
782 size_t _words_scanned;
783 // When _words_scanned reaches this limit, the regular clock is
784 // called. Notice that this might be decreased under certain
785 // circumstances (i.e. when we believe that we did an expensive
786 // operation).
787 size_t _words_scanned_limit;
788 // the initial value of _words_scanned_limit (i.e. what it was
789 // before it was decreased).
790 size_t _real_words_scanned_limit;
791
792 // the number of references this task has visited
793 size_t _refs_reached;
794 // When _refs_reached reaches this limit, the regular clock is
795 // called. Notice this this might be decreased under certain
796 // circumstances (i.e. when we believe that we did an expensive
797 // operation).
798 size_t _refs_reached_limit;
799 // the initial value of _refs_reached_limit (i.e. what it was before
800 // it was decreased).
801 size_t _real_refs_reached_limit;
802
803 // used by the work stealing stuff
804 int _hash_seed;
805 // if this is true, then the task has aborted for some reason
806 bool _has_aborted;
807 // set when the task aborts because it has met its time quota
808 bool _has_timed_out;
809 // true when we're draining SATB buffers; this avoids the task
810 // aborting due to SATB buffers being available (as we're already
811 // dealing with them)
812 bool _draining_satb_buffers;
813
814 // number sequence of past step times
815 NumberSeq _step_times_ms;
816 // elapsed time of this task
817 double _elapsed_time_ms;
818 // termination time of this task
819 double _termination_time_ms;
820 // when this task got into the termination protocol
821 double _termination_start_time_ms;
822
823 // true when the task is during a concurrent phase, false when it is
824 // in the remark phase (so, in the latter case, we do not have to
825 // check all the things that we have to check during the concurrent
826 // phase, i.e. SATB buffer availability...)
827 bool _concurrent;
828
829 TruncatedSeq _marking_step_diffs_ms;
830
831 // Counting data structures. Embedding the task's marked_bytes_array
832 // and card bitmap into the actual task saves having to go through
833 // the ConcurrentMark object.
834 size_t* _marked_bytes_array;
835 BitMap* _card_bm;
836
837 // it updates the local fields after this task has claimed
838 // a new region to scan
839 void setup_for_region(HeapRegion* hr);
840 // it brings up-to-date the limit of the region
841 void update_region_limit();
842
843 // called when either the words scanned or the refs visited limit
844 // has been reached
845 void reached_limit();
846 // recalculates the words scanned and refs visited limits
847 void recalculate_limits();
848 // decreases the words scanned and refs visited limits when we reach
849 // an expensive operation
850 void decrease_limits();
851 // it checks whether the words scanned or refs visited reached their
852 // respective limit and calls reached_limit() if they have
853 void check_limits() {
854 if (_words_scanned >= _words_scanned_limit ||
855 _refs_reached >= _refs_reached_limit) {
856 reached_limit();
857 }
858 }
859 // this is supposed to be called regularly during a marking step as
860 // it checks a bunch of conditions that might cause the marking step
861 // to abort
862 void regular_clock_call();
863 bool concurrent() { return _concurrent; }
864
865 // Test whether obj might have already been passed over by the
866 // mark bitmap scan, and so needs to be pushed onto the mark stack.
867 bool is_below_finger(oop obj, HeapWord* global_finger) const;
868
869 template<bool scan> void process_grey_object(oop obj);
870
871 public:
872 // It resets the task; it should be called right at the beginning of
873 // a marking phase.
874 void reset(G1CMBitMap* _nextMarkBitMap);
875 // it clears all the fields that correspond to a claimed region.
876 void clear_region_fields();
877
878 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
879
880 // The main method of this class which performs a marking step
881 // trying not to exceed the given duration. However, it might exit
882 // prematurely, according to some conditions (i.e. SATB buffers are
883 // available for processing).
884 void do_marking_step(double target_ms,
885 bool do_termination,
886 bool is_serial);
887
888 // These two calls start and stop the timer
889 void record_start_time() {
890 _elapsed_time_ms = os::elapsedTime() * 1000.0;
891 }
892 void record_end_time() {
893 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
894 }
895
896 // returns the worker ID associated with this task.
897 uint worker_id() { return _worker_id; }
898
899 // From TerminatorTerminator. It determines whether this task should
900 // exit the termination protocol after it's entered it.
901 virtual bool should_exit_termination();
902
903 // Resets the local region fields after a task has finished scanning a
904 // region; or when they have become stale as a result of the region
905 // being evacuated.
906 void giveup_current_region();
907
908 HeapWord* finger() { return _finger; }
909
910 bool has_aborted() { return _has_aborted; }
911 void set_has_aborted() { _has_aborted = true; }
912 void clear_has_aborted() { _has_aborted = false; }
913 bool has_timed_out() { return _has_timed_out; }
914 bool claimed() { return _claimed; }
915
916 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
917
918 // Increment the number of references this task has visited.
919 void increment_refs_reached() { ++_refs_reached; }
920
921 // Grey the object by marking it. If not already marked, push it on
922 // the local queue if below the finger.
923 // Precondition: obj is in region.
924 // Precondition: obj is below region's NTAMS.
925 inline void make_reference_grey(oop obj, HeapRegion* region);
926
927 // Grey the object (by calling make_grey_reference) if required,
928 // e.g. obj is below its containing region's NTAMS.
929 // Precondition: obj is a valid heap object.
930 inline void deal_with_reference(oop obj);
931
932 // It scans an object and visits its children.
933 inline void scan_object(oop obj);
934
935 // It pushes an object on the local queue.
936 inline void push(oop obj);
937
938 // These two move entries to/from the global stack.
939 void move_entries_to_global_stack();
940 void get_entries_from_global_stack();
941
942 // It pops and scans objects from the local queue. If partially is
943 // true, then it stops when the queue size is of a given limit. If
944 // partially is false, then it stops when the queue is empty.
945 void drain_local_queue(bool partially);
946 // It moves entries from the global stack to the local queue and
947 // drains the local queue. If partially is true, then it stops when
948 // both the global stack and the local queue reach a given size. If
949 // partially if false, it tries to empty them totally.
950 void drain_global_stack(bool partially);
951 // It keeps picking SATB buffers and processing them until no SATB
952 // buffers are available.
953 void drain_satb_buffers();
954
955 // moves the local finger to a new location
956 inline void move_finger_to(HeapWord* new_finger) {
957 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
958 _finger = new_finger;
959 }
960
961 G1CMTask(uint worker_id,
962 G1ConcurrentMark *cm,
963 size_t* marked_bytes,
964 BitMap* card_bm,
965 G1CMTaskQueue* task_queue,
966 G1CMTaskQueueSet* task_queues);
967
968 // it prints statistics associated with this task
969 void print_stats();
970 };
971
972 // Class that's used to to print out per-region liveness
973 // information. It's currently used at the end of marking and also
974 // after we sort the old regions at the end of the cleanup operation.
975 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
976 private:
977 // Accumulators for these values.
978 size_t _total_used_bytes;
979 size_t _total_capacity_bytes;
980 size_t _total_prev_live_bytes;
981 size_t _total_next_live_bytes;
982
983 // These are set up when we come across a "stars humongous" region
984 // (as this is where most of this information is stored, not in the
985 // subsequent "continues humongous" regions). After that, for every
986 // region in a given humongous region series we deduce the right
987 // values for it by simply subtracting the appropriate amount from
988 // these fields. All these values should reach 0 after we've visited
989 // the last region in the series.
990 size_t _hum_used_bytes;
991 size_t _hum_capacity_bytes;
992 size_t _hum_prev_live_bytes;
993 size_t _hum_next_live_bytes;
994
995 // Accumulator for the remembered set size
996 size_t _total_remset_bytes;
997
998 // Accumulator for strong code roots memory size
999 size_t _total_strong_code_roots_bytes;
1000
1001 static double perc(size_t val, size_t total) {
1002 if (total == 0) {
1003 return 0.0;
1004 } else {
1005 return 100.0 * ((double) val / (double) total);
1006 }
1007 }
1008
1009 static double bytes_to_mb(size_t val) {
1010 return (double) val / (double) M;
1011 }
1012
1013 // See the .cpp file.
1014 size_t get_hum_bytes(size_t* hum_bytes);
1015 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1016 size_t* prev_live_bytes, size_t* next_live_bytes);
1017
1018 public:
1019 // The header and footer are printed in the constructor and
1020 // destructor respectively.
1021 G1PrintRegionLivenessInfoClosure(const char* phase_name);
1022 virtual bool doHeapRegion(HeapRegion* r);
1023 ~G1PrintRegionLivenessInfoClosure();
1024 };
1025
1026 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP