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 class ConcurrentGCTimer;
38 class G1OldTracer;
39 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue;
40 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
41
42 // Closure used by CM during concurrent reference discovery
43 // and reference processing (during remarking) to determine
44 // if a particular object is alive. It is primarily used
45 // to determine if referents of discovered reference objects
46 // are alive. An instance is also embedded into the
47 // reference processor as the _is_alive_non_header field
48 class G1CMIsAliveClosure: public BoolObjectClosure {
49 G1CollectedHeap* _g1;
50 public:
51 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
52
53 bool do_object_b(oop obj);
54 };
55
56 // A generic CM bit map. This is essentially a wrapper around the BitMap
57 // class, with one bit per (1<<_shifter) HeapWords.
58
59 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
60 protected:
61 HeapWord* _bmStartWord; // base address of range covered by map
62 size_t _bmWordSize; // map size (in #HeapWords covered)
63 const int _shifter; // map to char or bit
64 BitMap _bm; // the bit map itself
65
66 public:
67 // constructor
68 G1CMBitMapRO(int shifter);
69
70 // inquiries
71 HeapWord* startWord() const { return _bmStartWord; }
72 // the following is one past the last word in space
73 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
74
75 // read marks
76
77 bool isMarked(HeapWord* addr) const {
78 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
79 "outside underlying space?");
80 return _bm.at(heapWordToOffset(addr));
81 }
82
83 // iteration
84 inline bool iterate(BitMapClosure* cl, MemRegion mr);
85
86 // Return the address corresponding to the next marked bit at or after
87 // "addr", and before "limit", if "limit" is non-NULL. If there is no
88 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
89 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
90 const HeapWord* limit = NULL) const;
91
92 // conversion utilities
93 HeapWord* offsetToHeapWord(size_t offset) const {
94 return _bmStartWord + (offset << _shifter);
95 }
96 size_t heapWordToOffset(const HeapWord* addr) const {
97 return pointer_delta(addr, _bmStartWord) >> _shifter;
98 }
99
100 // The argument addr should be the start address of a valid object
101 inline HeapWord* nextObject(HeapWord* addr);
102
103 void print_on_error(outputStream* st, const char* prefix) const;
104
105 // debugging
106 NOT_PRODUCT(bool covers(MemRegion rs) const;)
107 };
108
109 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
110 private:
111 G1CMBitMap* _bm;
112 public:
113 G1CMBitMapMappingChangedListener() : _bm(NULL) {}
114
115 void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
116
117 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
118 };
119
120 class G1CMBitMap : public G1CMBitMapRO {
121 private:
122 G1CMBitMapMappingChangedListener _listener;
123
124 public:
125 static size_t compute_size(size_t heap_size);
126 // Returns the amount of bytes on the heap between two marks in the bitmap.
127 static size_t mark_distance();
128 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
129 // mark bitmap corresponds to. This is the same as the mark distance above.
130 static size_t heap_map_factor() {
131 return mark_distance();
132 }
133
134 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
135
136 // Initializes the underlying BitMap to cover the given area.
137 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
138
139 // Write marks.
140 inline void mark(HeapWord* addr);
141 inline void clear(HeapWord* addr);
142 inline bool parMark(HeapWord* addr);
143
144 void clearRange(MemRegion mr);
145
146 // Clear the whole mark bitmap.
147 void clearAll();
148 };
149
150 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
151 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
152 VirtualSpace _virtual_space; // Underlying backing store for actual stack
153 G1ConcurrentMark* _cm;
154 oop* _base; // bottom of stack
155 jint _index; // one more than last occupied index
156 jint _capacity; // max #elements
157 jint _saved_index; // value of _index saved at start of GC
158
159 bool _overflow;
160 bool _should_expand;
161
162 public:
163 G1CMMarkStack(G1ConcurrentMark* cm);
164 ~G1CMMarkStack();
165
166 bool allocate(size_t capacity);
167
168 // Pushes the first "n" elements of "ptr_arr" on the stack.
169 // Locking impl: concurrency is allowed only with
170 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
171 // locking strategy.
172 void par_push_arr(oop* ptr_arr, int n);
173
174 // If returns false, the array was empty. Otherwise, removes up to "max"
175 // elements from the stack, and transfers them to "ptr_arr" in an
176 // unspecified order. The actual number transferred is given in "n" ("n
177 // == 0" is deliberately redundant with the return value.) Locking impl:
178 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
179 // operations, which use the same locking strategy.
180 bool par_pop_arr(oop* ptr_arr, int max, int* n);
181
182 bool isEmpty() { return _index == 0; }
183 int maxElems() { return _capacity; }
184
185 bool overflow() { return _overflow; }
186 void clear_overflow() { _overflow = false; }
187
188 bool should_expand() const { return _should_expand; }
189 void set_should_expand();
190
191 // Expand the stack, typically in response to an overflow condition
192 void expand();
193
194 int size() { return _index; }
195
196 void setEmpty() { _index = 0; clear_overflow(); }
197
198 // Record the current index.
199 void note_start_of_gc();
200
201 // Make sure that we have not added any entries to the stack during GC.
202 void note_end_of_gc();
203
204 // Apply fn to each oop in the mark stack, up to the bound recorded
205 // via one of the above "note" functions. The mark stack must not
206 // be modified while iterating.
207 template<typename Fn> void iterate(Fn fn);
208 };
209
210 class YoungList;
211
212 // Root Regions are regions that are not empty at the beginning of a
213 // marking cycle and which we might collect during an evacuation pause
214 // while the cycle is active. Given that, during evacuation pauses, we
215 // do not copy objects that are explicitly marked, what we have to do
216 // for the root regions is to scan them and mark all objects reachable
217 // from them. According to the SATB assumptions, we only need to visit
218 // each object once during marking. So, as long as we finish this scan
219 // before the next evacuation pause, we can copy the objects from the
220 // root regions without having to mark them or do anything else to them.
221 //
222 // Currently, we only support root region scanning once (at the start
223 // of the marking cycle) and the root regions are all the survivor
224 // regions populated during the initial-mark pause.
225 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
226 private:
227 YoungList* _young_list;
228 G1ConcurrentMark* _cm;
229
230 volatile bool _scan_in_progress;
231 volatile bool _should_abort;
232 HeapRegion* volatile _next_survivor;
233
234 void notify_scan_done();
235
236 public:
237 G1CMRootRegions();
238 // We actually do most of the initialization in this method.
239 void init(G1CollectedHeap* g1h, G1ConcurrentMark* cm);
240
241 // Reset the claiming / scanning of the root regions.
242 void prepare_for_scan();
243
244 // Forces get_next() to return NULL so that the iteration aborts early.
245 void abort() { _should_abort = true; }
246
247 // Return true if the CM thread are actively scanning root regions,
248 // false otherwise.
249 bool scan_in_progress() { return _scan_in_progress; }
250
251 // Claim the next root region to scan atomically, or return NULL if
252 // all have been claimed.
253 HeapRegion* claim_next();
254
255 void cancel_scan();
256
257 // Flag that we're done with root region scanning and notify anyone
258 // who's waiting on it. If aborted is false, assume that all regions
259 // have been claimed.
260 void scan_finished();
261
262 // If CM threads are still scanning root regions, wait until they
263 // are done. Return true if we had to wait, false otherwise.
264 bool wait_until_scan_finished();
265 };
266
267 class ConcurrentMarkThread;
268
269 class G1ConcurrentMark: public CHeapObj<mtGC> {
270 friend class ConcurrentMarkThread;
271 friend class G1ParNoteEndTask;
272 friend class CalcLiveObjectsClosure;
273 friend class G1CMRefProcTaskProxy;
274 friend class G1CMRefProcTaskExecutor;
275 friend class G1CMKeepAliveAndDrainClosure;
276 friend class G1CMDrainMarkingStackClosure;
277 friend class G1CMBitMapClosure;
278 friend class G1CMConcurrentMarkingTask;
279 friend class G1CMMarkStack;
280 friend class G1CMRemarkTask;
281 friend class G1CMTask;
282
283 protected:
284 ConcurrentMarkThread* _cmThread; // The thread doing the work
285 G1CollectedHeap* _g1h; // The heap
286 uint _parallel_marking_threads; // The number of marking
287 // threads we're using
288 uint _max_parallel_marking_threads; // Max number of marking
289 // threads we'll ever use
290 double _sleep_factor; // How much we have to sleep, with
291 // respect to the work we just did, to
292 // meet the marking overhead goal
293 double _marking_task_overhead; // Marking target overhead for
294 // a single task
295
296 FreeRegionList _cleanup_list;
297
298 // Concurrent marking support structures
299 G1CMBitMap _markBitMap1;
300 G1CMBitMap _markBitMap2;
301 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
302 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
303
304 BitMap _region_bm;
305 BitMap _card_bm;
306
307 // Heap bounds
308 HeapWord* _heap_start;
309 HeapWord* _heap_end;
310
311 // Root region tracking and claiming
312 G1CMRootRegions _root_regions;
313
314 // For gray objects
315 G1CMMarkStack _markStack; // Grey objects behind global finger
316 HeapWord* volatile _finger; // The global finger, region aligned,
317 // always points to the end of the
318 // last claimed region
319
320 // Marking tasks
321 uint _max_worker_id;// Maximum worker id
322 uint _active_tasks; // Task num currently active
323 G1CMTask** _tasks; // Task queue array (max_worker_id len)
324 G1CMTaskQueueSet* _task_queues; // Task queue set
325 ParallelTaskTerminator _terminator; // For termination
326
327 // Two sync barriers that are used to synchronize tasks when an
328 // overflow occurs. The algorithm is the following. All tasks enter
329 // the first one to ensure that they have all stopped manipulating
330 // the global data structures. After they exit it, they re-initialize
331 // their data structures and task 0 re-initializes the global data
332 // structures. Then, they enter the second sync barrier. This
333 // ensure, that no task starts doing work before all data
334 // structures (local and global) have been re-initialized. When they
335 // exit it, they are free to start working again.
336 WorkGangBarrierSync _first_overflow_barrier_sync;
337 WorkGangBarrierSync _second_overflow_barrier_sync;
338
339 // This is set by any task, when an overflow on the global data
340 // structures is detected
341 volatile bool _has_overflown;
342 // True: marking is concurrent, false: we're in remark
343 volatile bool _concurrent;
344 // Set at the end of a Full GC so that marking aborts
345 volatile bool _has_aborted;
346
347 // Used when remark aborts due to an overflow to indicate that
348 // another concurrent marking phase should start
349 volatile bool _restart_for_overflow;
350
351 // This is true from the very start of concurrent marking until the
352 // point when all the tasks complete their work. It is really used
353 // to determine the points between the end of concurrent marking and
354 // time of remark.
355 volatile bool _concurrent_marking_in_progress;
356
357 ConcurrentGCTimer* _gc_timer_cm;
358
359 G1OldTracer* _gc_tracer_cm;
360
361 // All of these times are in ms
362 NumberSeq _init_times;
363 NumberSeq _remark_times;
364 NumberSeq _remark_mark_times;
365 NumberSeq _remark_weak_ref_times;
366 NumberSeq _cleanup_times;
367 double _total_counting_time;
368 double _total_rs_scrub_time;
369
370 double* _accum_task_vtime; // Accumulated task vtime
371
372 WorkGang* _parallel_workers;
373
374 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
375 void weakRefsWork(bool clear_all_soft_refs);
376
377 void swapMarkBitMaps();
378
379 // It resets the global marking data structures, as well as the
380 // task local ones; should be called during initial mark.
381 void reset();
382
383 // Resets all the marking data structures. Called when we have to restart
384 // marking or when marking completes (via set_non_marking_state below).
385 void reset_marking_state(bool clear_overflow = true);
386
387 // We do this after we're done with marking so that the marking data
388 // structures are initialized to a sensible and predictable state.
389 void set_non_marking_state();
390
391 // Called to indicate how many threads are currently active.
392 void set_concurrency(uint active_tasks);
393
394 // It should be called to indicate which phase we're in (concurrent
395 // mark or remark) and how many threads are currently active.
396 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
397
398 // Prints all gathered CM-related statistics
399 void print_stats();
400
401 bool cleanup_list_is_empty() {
402 return _cleanup_list.is_empty();
403 }
404
405 // Accessor methods
406 uint parallel_marking_threads() const { return _parallel_marking_threads; }
407 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
408 double sleep_factor() { return _sleep_factor; }
409 double marking_task_overhead() { return _marking_task_overhead;}
410
411 HeapWord* finger() { return _finger; }
412 bool concurrent() { return _concurrent; }
413 uint active_tasks() { return _active_tasks; }
414 ParallelTaskTerminator* terminator() { return &_terminator; }
415
416 // It claims the next available region to be scanned by a marking
417 // task/thread. It might return NULL if the next region is empty or
418 // we have run out of regions. In the latter case, out_of_regions()
419 // determines whether we've really run out of regions or the task
420 // should call claim_region() again. This might seem a bit
421 // awkward. Originally, the code was written so that claim_region()
422 // either successfully returned with a non-empty region or there
423 // were no more regions to be claimed. The problem with this was
424 // that, in certain circumstances, it iterated over large chunks of
425 // the heap finding only empty regions and, while it was working, it
426 // was preventing the calling task to call its regular clock
427 // method. So, this way, each task will spend very little time in
428 // claim_region() and is allowed to call the regular clock method
429 // frequently.
430 HeapRegion* claim_region(uint worker_id);
431
432 // It determines whether we've run out of regions to scan. Note that
433 // the finger can point past the heap end in case the heap was expanded
434 // to satisfy an allocation without doing a GC. This is fine, because all
435 // objects in those regions will be considered live anyway because of
436 // SATB guarantees (i.e. their TAMS will be equal to bottom).
437 bool out_of_regions() { return _finger >= _heap_end; }
438
439 // Returns the task with the given id
440 G1CMTask* task(int id) {
441 assert(0 <= id && id < (int) _active_tasks,
442 "task id not within active bounds");
443 return _tasks[id];
444 }
445
446 // Returns the task queue with the given id
447 G1CMTaskQueue* task_queue(int id) {
448 assert(0 <= id && id < (int) _active_tasks,
449 "task queue id not within active bounds");
450 return (G1CMTaskQueue*) _task_queues->queue(id);
451 }
452
453 // Returns the task queue set
454 G1CMTaskQueueSet* task_queues() { return _task_queues; }
455
456 // Access / manipulation of the overflow flag which is set to
457 // indicate that the global stack has overflown
458 bool has_overflown() { return _has_overflown; }
459 void set_has_overflown() { _has_overflown = true; }
460 void clear_has_overflown() { _has_overflown = false; }
461 bool restart_for_overflow() { return _restart_for_overflow; }
462
463 // Methods to enter the two overflow sync barriers
464 void enter_first_sync_barrier(uint worker_id);
465 void enter_second_sync_barrier(uint worker_id);
466
467 // Live Data Counting data structures...
468 // These data structures are initialized at the start of
469 // marking. They are written to while marking is active.
470 // They are aggregated during remark; the aggregated values
471 // are then used to populate the _region_bm, _card_bm, and
472 // the total live bytes, which are then subsequently updated
473 // during cleanup.
474
475 // An array of bitmaps (one bit map per task). Each bitmap
476 // is used to record the cards spanned by the live objects
477 // marked by that task/worker.
478 BitMap* _count_card_bitmaps;
479
480 // Used to record the number of marked live bytes
481 // (for each region, by worker thread).
482 size_t** _count_marked_bytes;
483
484 // Card index of the bottom of the G1 heap. Used for biasing indices into
485 // the card bitmaps.
486 intptr_t _heap_bottom_card_num;
487
488 // Set to true when initialization is complete
489 bool _completed_initialization;
490
491 // end_timer, true to end gc timer after ending concurrent phase.
492 void register_concurrent_phase_end_common(bool end_timer);
493
494 public:
495 // Manipulation of the global mark stack.
496 // The push and pop operations are used by tasks for transfers
497 // between task-local queues and the global mark stack, and use
498 // locking for concurrency safety.
499 bool mark_stack_push(oop* arr, int n) {
500 _markStack.par_push_arr(arr, n);
501 if (_markStack.overflow()) {
502 set_has_overflown();
503 return false;
504 }
505 return true;
506 }
507 void mark_stack_pop(oop* arr, int max, int* n) {
508 _markStack.par_pop_arr(arr, max, n);
509 }
510 size_t mark_stack_size() { return _markStack.size(); }
511 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
512 bool mark_stack_overflow() { return _markStack.overflow(); }
513 bool mark_stack_empty() { return _markStack.isEmpty(); }
514
515 G1CMRootRegions* root_regions() { return &_root_regions; }
516
517 bool concurrent_marking_in_progress() {
518 return _concurrent_marking_in_progress;
519 }
520 void set_concurrent_marking_in_progress() {
521 _concurrent_marking_in_progress = true;
522 }
523 void clear_concurrent_marking_in_progress() {
524 _concurrent_marking_in_progress = false;
525 }
526
527 void concurrent_cycle_start();
528 void concurrent_cycle_end();
529
530 void update_accum_task_vtime(int i, double vtime) {
531 _accum_task_vtime[i] += vtime;
532 }
533
534 double all_task_accum_vtime() {
535 double ret = 0.0;
536 for (uint i = 0; i < _max_worker_id; ++i)
537 ret += _accum_task_vtime[i];
538 return ret;
539 }
540
541 // Attempts to steal an object from the task queues of other tasks
542 bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
543
544 G1ConcurrentMark(G1CollectedHeap* g1h,
545 G1RegionToSpaceMapper* prev_bitmap_storage,
546 G1RegionToSpaceMapper* next_bitmap_storage);
547 ~G1ConcurrentMark();
548
549 ConcurrentMarkThread* cmThread() { return _cmThread; }
550
551 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
552 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
553
554 // Returns the number of GC threads to be used in a concurrent
555 // phase based on the number of GC threads being used in a STW
556 // phase.
557 uint scale_parallel_threads(uint n_par_threads);
558
559 // Calculates the number of GC threads to be used in a concurrent phase.
560 uint calc_parallel_marking_threads();
561
562 // The following three are interaction between CM and
563 // G1CollectedHeap
564
565 // This notifies CM that a root during initial-mark needs to be
566 // grayed. It is MT-safe. word_size is the size of the object in
567 // words. It is passed explicitly as sometimes we cannot calculate
568 // it from the given object because it might be in an inconsistent
569 // state (e.g., in to-space and being copied). So the caller is
570 // responsible for dealing with this issue (e.g., get the size from
571 // the from-space image when the to-space image might be
572 // inconsistent) and always passing the size. hr is the region that
573 // contains the object and it's passed optionally from callers who
574 // might already have it (no point in recalculating it).
575 inline void grayRoot(oop obj,
576 size_t word_size,
577 uint worker_id,
578 HeapRegion* hr = NULL);
579
580 // Clear the next marking bitmap (will be called concurrently).
581 void clearNextBitmap();
582
583 // Return whether the next mark bitmap has no marks set. To be used for assertions
584 // only. Will not yield to pause requests.
585 bool nextMarkBitmapIsClear();
586
587 // These two do the work that needs to be done before and after the
588 // initial root checkpoint. Since this checkpoint can be done at two
589 // different points (i.e. an explicit pause or piggy-backed on a
590 // young collection), then it's nice to be able to easily share the
591 // pre/post code. It might be the case that we can put everything in
592 // the post method. TP
593 void checkpointRootsInitialPre();
594 void checkpointRootsInitialPost();
595
596 // Scan all the root regions and mark everything reachable from
597 // them.
598 void scanRootRegions();
599
600 // Scan a single root region and mark everything reachable from it.
601 void scanRootRegion(HeapRegion* hr, uint worker_id);
602
603 // Do concurrent phase of marking, to a tentative transitive closure.
604 void markFromRoots();
605
606 void checkpointRootsFinal(bool clear_all_soft_refs);
607 void checkpointRootsFinalWork();
608 void cleanup();
609 void completeCleanup();
610
611 // Mark in the previous bitmap. NB: this is usually read-only, so use
612 // this carefully!
613 inline void markPrev(oop p);
614
615 // Clears marks for all objects in the given range, for the prev or
616 // next bitmaps. NB: the previous bitmap is usually
617 // read-only, so use this carefully!
618 void clearRangePrevBitmap(MemRegion mr);
619
620 // Notify data structures that a GC has started.
621 void note_start_of_gc() {
622 _markStack.note_start_of_gc();
623 }
624
625 // Notify data structures that a GC is finished.
626 void note_end_of_gc() {
627 _markStack.note_end_of_gc();
628 }
629
630 // Verify that there are no CSet oops on the stacks (taskqueues /
631 // global mark stack) and fingers (global / per-task).
632 // If marking is not in progress, it's a no-op.
633 void verify_no_cset_oops() PRODUCT_RETURN;
634
635 inline bool isPrevMarked(oop p) const;
636
637 inline bool do_yield_check(uint worker_i = 0);
638
639 // Called to abort the marking cycle after a Full GC takes place.
640 void abort();
641
642 bool has_aborted() { return _has_aborted; }
643
644 void print_summary_info();
645
646 void print_worker_threads_on(outputStream* st) const;
647
648 void print_on_error(outputStream* st) const;
649
650 // Liveness counting
651
652 // Utility routine to set an exclusive range of cards on the given
653 // card liveness bitmap
654 inline void set_card_bitmap_range(BitMap* card_bm,
655 BitMap::idx_t start_idx,
656 BitMap::idx_t end_idx,
657 bool is_par);
658
659 // Returns the card number of the bottom of the G1 heap.
660 // Used in biasing indices into accounting card bitmaps.
661 intptr_t heap_bottom_card_num() const {
662 return _heap_bottom_card_num;
663 }
664
665 // Returns the card bitmap for a given task or worker id.
666 BitMap* count_card_bitmap_for(uint worker_id) {
667 assert(worker_id < _max_worker_id, "oob");
668 assert(_count_card_bitmaps != NULL, "uninitialized");
669 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
670 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
671 return task_card_bm;
672 }
673
674 // Returns the array containing the marked bytes for each region,
675 // for the given worker or task id.
676 size_t* count_marked_bytes_array_for(uint worker_id) {
677 assert(worker_id < _max_worker_id, "oob");
678 assert(_count_marked_bytes != NULL, "uninitialized");
679 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
680 assert(marked_bytes_array != NULL, "uninitialized");
681 return marked_bytes_array;
682 }
683
684 // Returns the index in the liveness accounting card table bitmap
685 // for the given address
686 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
687
688 // Counts the size of the given memory region in the the given
689 // marked_bytes array slot for the given HeapRegion.
690 // Sets the bits in the given card bitmap that are associated with the
691 // cards that are spanned by the memory region.
692 inline void count_region(MemRegion mr,
693 HeapRegion* hr,
694 size_t* marked_bytes_array,
695 BitMap* task_card_bm);
696
697 // Counts the given object in the given task/worker counting
698 // data structures.
699 inline void count_object(oop obj,
700 HeapRegion* hr,
701 size_t* marked_bytes_array,
702 BitMap* task_card_bm,
703 size_t word_size);
704
705 // Attempts to mark the given object and, if successful, counts
706 // the object in the given task/worker counting structures.
707 inline bool par_mark_and_count(oop obj,
708 HeapRegion* hr,
709 size_t* marked_bytes_array,
710 BitMap* task_card_bm);
711
712 // Attempts to mark the given object and, if successful, counts
713 // the object in the task/worker counting structures for the
714 // given worker id.
715 inline bool par_mark_and_count(oop obj,
716 size_t word_size,
717 HeapRegion* hr,
718 uint worker_id);
719
720 // Returns true if initialization was successfully completed.
721 bool completed_initialization() const {
722 return _completed_initialization;
723 }
724
725 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
726 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
727
728 protected:
729 // Clear all the per-task bitmaps and arrays used to store the
730 // counting data.
731 void clear_all_count_data();
732
733 // Aggregates the counting data for each worker/task
734 // that was constructed while marking. Also sets
735 // the amount of marked bytes for each region and
736 // the top at concurrent mark count.
737 void aggregate_count_data();
738
739 // Verification routine
740 void verify_count_data();
741 };
742
743 // A class representing a marking task.
744 class G1CMTask : public TerminatorTerminator {
745 private:
746 enum PrivateConstants {
747 // the regular clock call is called once the scanned words reaches
748 // this limit
749 words_scanned_period = 12*1024,
750 // the regular clock call is called once the number of visited
751 // references reaches this limit
752 refs_reached_period = 384,
753 // initial value for the hash seed, used in the work stealing code
754 init_hash_seed = 17,
755 // how many entries will be transferred between global stack and
756 // local queues
757 global_stack_transfer_size = 16
758 };
759
760 uint _worker_id;
761 G1CollectedHeap* _g1h;
762 G1ConcurrentMark* _cm;
763 G1CMBitMap* _nextMarkBitMap;
764 // the task queue of this task
765 G1CMTaskQueue* _task_queue;
766 private:
767 // the task queue set---needed for stealing
768 G1CMTaskQueueSet* _task_queues;
769 // indicates whether the task has been claimed---this is only for
770 // debugging purposes
771 bool _claimed;
772
773 // number of calls to this task
774 int _calls;
775
776 // when the virtual timer reaches this time, the marking step should
777 // exit
778 double _time_target_ms;
779 // the start time of the current marking step
780 double _start_time_ms;
781
782 // the oop closure used for iterations over oops
783 G1CMOopClosure* _cm_oop_closure;
784
785 // the region this task is scanning, NULL if we're not scanning any
786 HeapRegion* _curr_region;
787 // the local finger of this task, NULL if we're not scanning a region
788 HeapWord* _finger;
789 // limit of the region this task is scanning, NULL if we're not scanning one
790 HeapWord* _region_limit;
791
792 // the number of words this task has scanned
793 size_t _words_scanned;
794 // When _words_scanned reaches this limit, the regular clock is
795 // called. Notice that this might be decreased under certain
796 // circumstances (i.e. when we believe that we did an expensive
797 // operation).
798 size_t _words_scanned_limit;
799 // the initial value of _words_scanned_limit (i.e. what it was
800 // before it was decreased).
801 size_t _real_words_scanned_limit;
802
803 // the number of references this task has visited
804 size_t _refs_reached;
805 // When _refs_reached reaches this limit, the regular clock is
806 // called. Notice this this might be decreased under certain
807 // circumstances (i.e. when we believe that we did an expensive
808 // operation).
809 size_t _refs_reached_limit;
810 // the initial value of _refs_reached_limit (i.e. what it was before
811 // it was decreased).
812 size_t _real_refs_reached_limit;
813
814 // used by the work stealing stuff
815 int _hash_seed;
816 // if this is true, then the task has aborted for some reason
817 bool _has_aborted;
818 // set when the task aborts because it has met its time quota
819 bool _has_timed_out;
820 // true when we're draining SATB buffers; this avoids the task
821 // aborting due to SATB buffers being available (as we're already
822 // dealing with them)
823 bool _draining_satb_buffers;
824
825 // number sequence of past step times
826 NumberSeq _step_times_ms;
827 // elapsed time of this task
828 double _elapsed_time_ms;
829 // termination time of this task
830 double _termination_time_ms;
831 // when this task got into the termination protocol
832 double _termination_start_time_ms;
833
834 // true when the task is during a concurrent phase, false when it is
835 // in the remark phase (so, in the latter case, we do not have to
836 // check all the things that we have to check during the concurrent
837 // phase, i.e. SATB buffer availability...)
838 bool _concurrent;
839
840 TruncatedSeq _marking_step_diffs_ms;
841
842 // Counting data structures. Embedding the task's marked_bytes_array
843 // and card bitmap into the actual task saves having to go through
844 // the ConcurrentMark object.
845 size_t* _marked_bytes_array;
846 BitMap* _card_bm;
847
848 // it updates the local fields after this task has claimed
849 // a new region to scan
850 void setup_for_region(HeapRegion* hr);
851 // it brings up-to-date the limit of the region
852 void update_region_limit();
853
854 // called when either the words scanned or the refs visited limit
855 // has been reached
856 void reached_limit();
857 // recalculates the words scanned and refs visited limits
858 void recalculate_limits();
859 // decreases the words scanned and refs visited limits when we reach
860 // an expensive operation
861 void decrease_limits();
862 // it checks whether the words scanned or refs visited reached their
863 // respective limit and calls reached_limit() if they have
864 void check_limits() {
865 if (_words_scanned >= _words_scanned_limit ||
866 _refs_reached >= _refs_reached_limit) {
867 reached_limit();
868 }
869 }
870 // this is supposed to be called regularly during a marking step as
871 // it checks a bunch of conditions that might cause the marking step
872 // to abort
873 void regular_clock_call();
874 bool concurrent() { return _concurrent; }
875
876 // Test whether obj might have already been passed over by the
877 // mark bitmap scan, and so needs to be pushed onto the mark stack.
878 bool is_below_finger(oop obj, HeapWord* global_finger) const;
879
880 template<bool scan> void process_grey_object(oop obj);
881
882 public:
883 // It resets the task; it should be called right at the beginning of
884 // a marking phase.
885 void reset(G1CMBitMap* _nextMarkBitMap);
886 // it clears all the fields that correspond to a claimed region.
887 void clear_region_fields();
888
889 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
890
891 // The main method of this class which performs a marking step
892 // trying not to exceed the given duration. However, it might exit
893 // prematurely, according to some conditions (i.e. SATB buffers are
894 // available for processing).
895 void do_marking_step(double target_ms,
896 bool do_termination,
897 bool is_serial);
898
899 // These two calls start and stop the timer
900 void record_start_time() {
901 _elapsed_time_ms = os::elapsedTime() * 1000.0;
902 }
903 void record_end_time() {
904 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
905 }
906
907 // returns the worker ID associated with this task.
908 uint worker_id() { return _worker_id; }
909
910 // From TerminatorTerminator. It determines whether this task should
911 // exit the termination protocol after it's entered it.
912 virtual bool should_exit_termination();
913
914 // Resets the local region fields after a task has finished scanning a
915 // region; or when they have become stale as a result of the region
916 // being evacuated.
917 void giveup_current_region();
918
919 HeapWord* finger() { return _finger; }
920
921 bool has_aborted() { return _has_aborted; }
922 void set_has_aborted() { _has_aborted = true; }
923 void clear_has_aborted() { _has_aborted = false; }
924 bool has_timed_out() { return _has_timed_out; }
925 bool claimed() { return _claimed; }
926
927 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
928
929 // Increment the number of references this task has visited.
930 void increment_refs_reached() { ++_refs_reached; }
931
932 // Grey the object by marking it. If not already marked, push it on
933 // the local queue if below the finger.
934 // Precondition: obj is in region.
935 // Precondition: obj is below region's NTAMS.
936 inline void make_reference_grey(oop obj, HeapRegion* region);
937
938 // Grey the object (by calling make_grey_reference) if required,
939 // e.g. obj is below its containing region's NTAMS.
940 // Precondition: obj is a valid heap object.
941 inline void deal_with_reference(oop obj);
942
943 // It scans an object and visits its children.
944 inline void scan_object(oop obj);
945
946 // It pushes an object on the local queue.
947 inline void push(oop obj);
948
949 // These two move entries to/from the global stack.
950 void move_entries_to_global_stack();
951 void get_entries_from_global_stack();
952
953 // It pops and scans objects from the local queue. If partially is
954 // true, then it stops when the queue size is of a given limit. If
955 // partially is false, then it stops when the queue is empty.
956 void drain_local_queue(bool partially);
957 // It moves entries from the global stack to the local queue and
958 // drains the local queue. If partially is true, then it stops when
959 // both the global stack and the local queue reach a given size. If
960 // partially if false, it tries to empty them totally.
961 void drain_global_stack(bool partially);
962 // It keeps picking SATB buffers and processing them until no SATB
963 // buffers are available.
964 void drain_satb_buffers();
965
966 // moves the local finger to a new location
967 inline void move_finger_to(HeapWord* new_finger) {
968 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
969 _finger = new_finger;
970 }
971
972 G1CMTask(uint worker_id,
973 G1ConcurrentMark *cm,
974 size_t* marked_bytes,
975 BitMap* card_bm,
976 G1CMTaskQueue* task_queue,
977 G1CMTaskQueueSet* task_queues);
978
979 // it prints statistics associated with this task
980 void print_stats();
981 };
982
983 // Class that's used to to print out per-region liveness
984 // information. It's currently used at the end of marking and also
985 // after we sort the old regions at the end of the cleanup operation.
986 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
987 private:
988 // Accumulators for these values.
989 size_t _total_used_bytes;
990 size_t _total_capacity_bytes;
991 size_t _total_prev_live_bytes;
992 size_t _total_next_live_bytes;
993
994 // These are set up when we come across a "stars humongous" region
995 // (as this is where most of this information is stored, not in the
996 // subsequent "continues humongous" regions). After that, for every
997 // region in a given humongous region series we deduce the right
998 // values for it by simply subtracting the appropriate amount from
999 // these fields. All these values should reach 0 after we've visited
1000 // the last region in the series.
1001 size_t _hum_used_bytes;
1002 size_t _hum_capacity_bytes;
1003 size_t _hum_prev_live_bytes;
1004 size_t _hum_next_live_bytes;
1005
1006 // Accumulator for the remembered set size
1007 size_t _total_remset_bytes;
1008
1009 // Accumulator for strong code roots memory size
1010 size_t _total_strong_code_roots_bytes;
1011
1012 static double perc(size_t val, size_t total) {
1013 if (total == 0) {
1014 return 0.0;
1015 } else {
1016 return 100.0 * ((double) val / (double) total);
1017 }
1018 }
1019
1020 static double bytes_to_mb(size_t val) {
1021 return (double) val / (double) M;
1022 }
1023
1024 // See the .cpp file.
1025 size_t get_hum_bytes(size_t* hum_bytes);
1026 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1027 size_t* prev_live_bytes, size_t* next_live_bytes);
1028
1029 public:
1030 // The header and footer are printed in the constructor and
1031 // destructor respectively.
1032 G1PrintRegionLivenessInfoClosure(const char* phase_name);
1033 virtual bool doHeapRegion(HeapRegion* r);
1034 ~G1PrintRegionLivenessInfoClosure();
1035 };
1036
1037 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP