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 class G1SurvivorRegions;
40 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue;
41 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
42
43 // Closure used by CM during concurrent reference discovery
44 // and reference processing (during remarking) to determine
45 // if a particular object is alive. It is primarily used
46 // to determine if referents of discovered reference objects
47 // are alive. An instance is also embedded into the
48 // reference processor as the _is_alive_non_header field
49 class G1CMIsAliveClosure: public BoolObjectClosure {
50 G1CollectedHeap* _g1;
51 public:
52 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
53
54 bool do_object_b(oop obj);
55 };
56
57 // A generic CM bit map. This is essentially a wrapper around the BitMap
58 // class, with one bit per (1<<_shifter) HeapWords.
59
60 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
61 protected:
62 HeapWord* _bmStartWord; // base address of range covered by map
63 size_t _bmWordSize; // map size (in #HeapWords covered)
64 const int _shifter; // map to char or bit
65 BitMapView _bm; // the bit map itself
66
67 public:
68 // constructor
69 G1CMBitMapRO(int shifter);
70
71 // inquiries
72 HeapWord* startWord() const { return _bmStartWord; }
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
87 // Return the address corresponding to the next marked bit at or after
88 // "addr", and before "limit", if "limit" is non-NULL. If there is no
89 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
90 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
91 const HeapWord* limit = NULL) const;
92
93 // conversion utilities
94 HeapWord* offsetToHeapWord(size_t offset) const {
95 return _bmStartWord + (offset << _shifter);
96 }
97 size_t heapWordToOffset(const HeapWord* addr) const {
98 return pointer_delta(addr, _bmStartWord) >> _shifter;
99 }
100
101 // The argument addr should be the start address of a valid object
102 inline HeapWord* nextObject(HeapWord* addr);
103
104 void print_on_error(outputStream* st, const char* prefix) const;
105
106 // debugging
107 NOT_PRODUCT(bool covers(MemRegion rs) const;)
108 };
109
110 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
111 private:
112 G1CMBitMap* _bm;
113 public:
114 G1CMBitMapMappingChangedListener() : _bm(NULL) {}
115
116 void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
117
118 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
119 };
120
121 class G1CMBitMap : public G1CMBitMapRO {
122 private:
123 G1CMBitMapMappingChangedListener _listener;
124
125 public:
126 static size_t compute_size(size_t heap_size);
127 // Returns the amount of bytes on the heap between two marks in the bitmap.
128 static size_t mark_distance();
129 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
130 // mark bitmap corresponds to. This is the same as the mark distance above.
131 static size_t heap_map_factor() {
132 return mark_distance();
133 }
134
135 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
136
137 // Initializes the underlying BitMap to cover the given area.
138 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
139
140 // Write marks.
141 inline void mark(HeapWord* addr);
142 inline void clear(HeapWord* addr);
143 inline bool parMark(HeapWord* addr);
144
145 void clear_range(MemRegion mr);
146 };
147
148 // Represents the overflow mark stack used by concurrent marking.
149 //
150 // Stores oops in a huge buffer in virtual memory that is always fully committed.
151 // Resizing may only happen during a STW pause when the stack is empty.
152 //
153 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
154 // stack memory is split into evenly sized chunks of oops. Users can only
155 // add or remove entries on that basis.
156 // Chunks are filled in increasing address order. Not completely filled chunks
157 // have a NULL element as a terminating element.
158 //
159 // Every chunk has a header containing a single pointer element used for memory
160 // management. This wastes some space, but is negligible (< .1% with current sizing).
161 //
162 // Memory management is done using a mix of tracking a high water-mark indicating
163 // that all chunks at a lower address are valid chunks, and a singly linked free
164 // list connecting all empty chunks.
165 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
166 public:
167 // Number of oops that can fit in a single chunk.
168 static const size_t OopsPerChunk = 1024 - 1 /* One reference for the next pointer */;
169 private:
170 struct OopChunk {
171 OopChunk* next;
172 oop data[OopsPerChunk];
173 };
174
175 size_t _max_chunk_capacity; // Maximum number of OopChunk elements on the stack.
176
177 OopChunk* _base; // Bottom address of allocated memory area.
178 size_t _chunk_capacity; // Current maximum number of OopChunk elements.
179
180 char _pad0[DEFAULT_CACHE_LINE_SIZE];
181 OopChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users.
182 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(OopChunk*)];
183 OopChunk* volatile _chunk_list; // List of chunks currently containing data.
184 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(OopChunk*)];
185
186 volatile size_t _hwm; // High water mark within the reserved space.
187 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
188
189 // Allocate a new chunk from the reserved memory, using the high water mark. Returns
190 // NULL if out of memory.
191 OopChunk* allocate_new_chunk();
192
193 volatile size_t _chunks_in_chunk_list;
194
195 volatile bool _out_of_memory;
196
197 // Atomically add the given chunk to the list.
198 void add_chunk_to_list(OopChunk* volatile* list, OopChunk* elem);
199 // Atomically remove and return a chunk from the given list. Returns NULL if the
200 // list is empty.
201 OopChunk* remove_chunk_from_list(OopChunk* volatile* list);
202
203 void add_chunk_to_chunk_list(OopChunk* elem);
204 void add_chunk_to_free_list(OopChunk* elem);
205
206 OopChunk* remove_chunk_from_chunk_list();
207 OopChunk* remove_chunk_from_free_list();
208
209 bool _should_expand;
210
211 // Resizes the mark stack to the given new capacity. Releases any previous
212 // memory if successful.
213 bool resize(size_t new_capacity);
214
215 public:
216 G1CMMarkStack();
217 ~G1CMMarkStack();
218
219 // Alignment and minimum capacity of this mark stack in number of oops.
220 static size_t capacity_alignment();
221
222 // Allocate and initialize the mark stack with the given number of oops.
223 bool initialize(size_t initial_capacity, size_t max_capacity);
224
225 // Pushes the given buffer containing at most OopsPerChunk elements on the mark
226 // stack. If less than OopsPerChunk elements are to be pushed, the array must
227 // be terminated with a NULL.
228 // Returns whether the buffer contents were successfully pushed to the global mark
229 // stack.
230 bool par_push_chunk(oop* buffer);
231
232 // Pops a chunk from this mark stack, copying them into the given buffer. This
233 // chunk may contain up to OopsPerChunk elements. If there are less, the last
234 // element in the array is a NULL pointer.
235 bool par_pop_chunk(oop* buffer);
236
237 // Return whether the chunk list is empty. Racy due to unsynchronized access to
238 // _chunk_list.
239 bool is_empty() const { return _chunk_list == NULL; }
240
241 size_t capacity() const { return _chunk_capacity; }
242
243 bool is_out_of_memory() const { return _out_of_memory; }
244 void clear_out_of_memory() { _out_of_memory = false; }
245
246 bool should_expand() const { return _should_expand; }
247 void set_should_expand(bool value) { _should_expand = value; }
248
249 // Expand the stack, typically in response to an overflow condition
250 void expand();
251
252 // Return the approximate number of oops on this mark stack. Racy due to
253 // unsynchronized access to _chunks_in_chunk_list.
254 size_t size() const { return _chunks_in_chunk_list * OopsPerChunk; }
255
256 void set_empty();
257
258 // Apply Fn to every oop on the mark stack. The mark stack must not
259 // be modified while iterating.
260 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
261 };
262
263 // Root Regions are regions that are not empty at the beginning of a
264 // marking cycle and which we might collect during an evacuation pause
265 // while the cycle is active. Given that, during evacuation pauses, we
266 // do not copy objects that are explicitly marked, what we have to do
267 // for the root regions is to scan them and mark all objects reachable
268 // from them. According to the SATB assumptions, we only need to visit
269 // each object once during marking. So, as long as we finish this scan
270 // before the next evacuation pause, we can copy the objects from the
271 // root regions without having to mark them or do anything else to them.
272 //
273 // Currently, we only support root region scanning once (at the start
274 // of the marking cycle) and the root regions are all the survivor
275 // regions populated during the initial-mark pause.
276 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
277 private:
278 const G1SurvivorRegions* _survivors;
279 G1ConcurrentMark* _cm;
280
281 volatile bool _scan_in_progress;
282 volatile bool _should_abort;
283 volatile int _claimed_survivor_index;
284
285 void notify_scan_done();
286
287 public:
288 G1CMRootRegions();
289 // We actually do most of the initialization in this method.
290 void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
291
292 // Reset the claiming / scanning of the root regions.
293 void prepare_for_scan();
294
295 // Forces get_next() to return NULL so that the iteration aborts early.
296 void abort() { _should_abort = true; }
297
298 // Return true if the CM thread are actively scanning root regions,
299 // false otherwise.
300 bool scan_in_progress() { return _scan_in_progress; }
301
302 // Claim the next root region to scan atomically, or return NULL if
303 // all have been claimed.
304 HeapRegion* claim_next();
305
306 // The number of root regions to scan.
307 uint num_root_regions() const;
308
309 void cancel_scan();
310
311 // Flag that we're done with root region scanning and notify anyone
312 // who's waiting on it. If aborted is false, assume that all regions
313 // have been claimed.
314 void scan_finished();
315
316 // If CM threads are still scanning root regions, wait until they
317 // are done. Return true if we had to wait, false otherwise.
318 bool wait_until_scan_finished();
319 };
320
321 class ConcurrentMarkThread;
322
323 class G1ConcurrentMark: public CHeapObj<mtGC> {
324 friend class ConcurrentMarkThread;
325 friend class G1ParNoteEndTask;
326 friend class G1VerifyLiveDataClosure;
327 friend class G1CMRefProcTaskProxy;
328 friend class G1CMRefProcTaskExecutor;
329 friend class G1CMKeepAliveAndDrainClosure;
330 friend class G1CMDrainMarkingStackClosure;
331 friend class G1CMBitMapClosure;
332 friend class G1CMConcurrentMarkingTask;
333 friend class G1CMRemarkTask;
334 friend class G1CMTask;
335
336 protected:
337 ConcurrentMarkThread* _cmThread; // The thread doing the work
338 G1CollectedHeap* _g1h; // The heap
339 uint _parallel_marking_threads; // The number of marking
340 // threads we're using
341 uint _max_parallel_marking_threads; // Max number of marking
342 // threads we'll ever use
343 double _sleep_factor; // How much we have to sleep, with
344 // respect to the work we just did, to
345 // meet the marking overhead goal
346 double _marking_task_overhead; // Marking target overhead for
347 // a single task
348
349 FreeRegionList _cleanup_list;
350
351 // Concurrent marking support structures
352 G1CMBitMap _markBitMap1;
353 G1CMBitMap _markBitMap2;
354 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
355 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
356
357 // Heap bounds
358 HeapWord* _heap_start;
359 HeapWord* _heap_end;
360
361 // Root region tracking and claiming
362 G1CMRootRegions _root_regions;
363
364 // For gray objects
365 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger
366 HeapWord* volatile _finger; // The global finger, region aligned,
367 // always points to the end of the
368 // last claimed region
369
370 // Marking tasks
371 uint _max_worker_id;// Maximum worker id
372 uint _active_tasks; // Task num currently active
373 G1CMTask** _tasks; // Task queue array (max_worker_id len)
374 G1CMTaskQueueSet* _task_queues; // Task queue set
375 ParallelTaskTerminator _terminator; // For termination
376
377 // Two sync barriers that are used to synchronize tasks when an
378 // overflow occurs. The algorithm is the following. All tasks enter
379 // the first one to ensure that they have all stopped manipulating
380 // the global data structures. After they exit it, they re-initialize
381 // their data structures and task 0 re-initializes the global data
382 // structures. Then, they enter the second sync barrier. This
383 // ensure, that no task starts doing work before all data
384 // structures (local and global) have been re-initialized. When they
385 // exit it, they are free to start working again.
386 WorkGangBarrierSync _first_overflow_barrier_sync;
387 WorkGangBarrierSync _second_overflow_barrier_sync;
388
389 // This is set by any task, when an overflow on the global data
390 // structures is detected
391 volatile bool _has_overflown;
392 // True: marking is concurrent, false: we're in remark
393 volatile bool _concurrent;
394 // Set at the end of a Full GC so that marking aborts
395 volatile bool _has_aborted;
396
397 // Used when remark aborts due to an overflow to indicate that
398 // another concurrent marking phase should start
399 volatile bool _restart_for_overflow;
400
401 // This is true from the very start of concurrent marking until the
402 // point when all the tasks complete their work. It is really used
403 // to determine the points between the end of concurrent marking and
404 // time of remark.
405 volatile bool _concurrent_marking_in_progress;
406
407 ConcurrentGCTimer* _gc_timer_cm;
408
409 G1OldTracer* _gc_tracer_cm;
410
411 // All of these times are in ms
412 NumberSeq _init_times;
413 NumberSeq _remark_times;
414 NumberSeq _remark_mark_times;
415 NumberSeq _remark_weak_ref_times;
416 NumberSeq _cleanup_times;
417 double _total_counting_time;
418 double _total_rs_scrub_time;
419
420 double* _accum_task_vtime; // Accumulated task vtime
421
422 WorkGang* _parallel_workers;
423
424 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
425 void weakRefsWork(bool clear_all_soft_refs);
426
427 void swapMarkBitMaps();
428
429 // It resets the global marking data structures, as well as the
430 // task local ones; should be called during initial mark.
431 void reset();
432
433 // Resets all the marking data structures. Called when we have to restart
434 // marking or when marking completes (via set_non_marking_state below).
435 void reset_marking_state(bool clear_overflow = true);
436
437 // We do this after we're done with marking so that the marking data
438 // structures are initialized to a sensible and predictable state.
439 void set_non_marking_state();
440
441 // Called to indicate how many threads are currently active.
442 void set_concurrency(uint active_tasks);
443
444 // It should be called to indicate which phase we're in (concurrent
445 // mark or remark) and how many threads are currently active.
446 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
447
448 // Prints all gathered CM-related statistics
449 void print_stats();
450
451 bool cleanup_list_is_empty() {
452 return _cleanup_list.is_empty();
453 }
454
455 // Accessor methods
456 uint parallel_marking_threads() const { return _parallel_marking_threads; }
457 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
458 double sleep_factor() { return _sleep_factor; }
459 double marking_task_overhead() { return _marking_task_overhead;}
460
461 HeapWord* finger() { return _finger; }
462 bool concurrent() { return _concurrent; }
463 uint active_tasks() { return _active_tasks; }
464 ParallelTaskTerminator* terminator() { return &_terminator; }
465
466 // It claims the next available region to be scanned by a marking
467 // task/thread. It might return NULL if the next region is empty or
468 // we have run out of regions. In the latter case, out_of_regions()
469 // determines whether we've really run out of regions or the task
470 // should call claim_region() again. This might seem a bit
471 // awkward. Originally, the code was written so that claim_region()
472 // either successfully returned with a non-empty region or there
473 // were no more regions to be claimed. The problem with this was
474 // that, in certain circumstances, it iterated over large chunks of
475 // the heap finding only empty regions and, while it was working, it
476 // was preventing the calling task to call its regular clock
477 // method. So, this way, each task will spend very little time in
478 // claim_region() and is allowed to call the regular clock method
479 // frequently.
480 HeapRegion* claim_region(uint worker_id);
481
482 // It determines whether we've run out of regions to scan. Note that
483 // the finger can point past the heap end in case the heap was expanded
484 // to satisfy an allocation without doing a GC. This is fine, because all
485 // objects in those regions will be considered live anyway because of
486 // SATB guarantees (i.e. their TAMS will be equal to bottom).
487 bool out_of_regions() { return _finger >= _heap_end; }
488
489 // Returns the task with the given id
490 G1CMTask* task(int id) {
491 assert(0 <= id && id < (int) _active_tasks,
492 "task id not within active bounds");
493 return _tasks[id];
494 }
495
496 // Returns the task queue with the given id
497 G1CMTaskQueue* task_queue(int id) {
498 assert(0 <= id && id < (int) _active_tasks,
499 "task queue id not within active bounds");
500 return (G1CMTaskQueue*) _task_queues->queue(id);
501 }
502
503 // Returns the task queue set
504 G1CMTaskQueueSet* task_queues() { return _task_queues; }
505
506 // Access / manipulation of the overflow flag which is set to
507 // indicate that the global stack has overflown
508 bool has_overflown() { return _has_overflown; }
509 void set_has_overflown() { _has_overflown = true; }
510 void clear_has_overflown() { _has_overflown = false; }
511 bool restart_for_overflow() { return _restart_for_overflow; }
512
513 // Methods to enter the two overflow sync barriers
514 void enter_first_sync_barrier(uint worker_id);
515 void enter_second_sync_barrier(uint worker_id);
516
517 // Card index of the bottom of the G1 heap. Used for biasing indices into
518 // the card bitmaps.
519 intptr_t _heap_bottom_card_num;
520
521 // Set to true when initialization is complete
522 bool _completed_initialization;
523
524 // end_timer, true to end gc timer after ending concurrent phase.
525 void register_concurrent_phase_end_common(bool end_timer);
526
527 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
528 // true, periodically insert checks to see if this method should exit prematurely.
529 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
530 public:
531 // Manipulation of the global mark stack.
532 // The push and pop operations are used by tasks for transfers
533 // between task-local queues and the global mark stack.
534 bool mark_stack_push(oop* arr) {
535 if (!_global_mark_stack.par_push_chunk(arr)) {
536 set_has_overflown();
537 return false;
538 }
539 return true;
540 }
541 bool mark_stack_pop(oop* arr) {
542 return _global_mark_stack.par_pop_chunk(arr);
543 }
544 size_t mark_stack_size() { return _global_mark_stack.size(); }
545 size_t partial_mark_stack_size_target() { return _global_mark_stack.capacity()/3; }
546 bool mark_stack_overflow() { return _global_mark_stack.is_out_of_memory(); }
547 bool mark_stack_empty() { return _global_mark_stack.is_empty(); }
548
549 G1CMRootRegions* root_regions() { return &_root_regions; }
550
551 bool concurrent_marking_in_progress() {
552 return _concurrent_marking_in_progress;
553 }
554 void set_concurrent_marking_in_progress() {
555 _concurrent_marking_in_progress = true;
556 }
557 void clear_concurrent_marking_in_progress() {
558 _concurrent_marking_in_progress = false;
559 }
560
561 void concurrent_cycle_start();
562 void concurrent_cycle_end();
563
564 void update_accum_task_vtime(int i, double vtime) {
565 _accum_task_vtime[i] += vtime;
566 }
567
568 double all_task_accum_vtime() {
569 double ret = 0.0;
570 for (uint i = 0; i < _max_worker_id; ++i)
571 ret += _accum_task_vtime[i];
572 return ret;
573 }
574
575 // Attempts to steal an object from the task queues of other tasks
576 bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
577
578 G1ConcurrentMark(G1CollectedHeap* g1h,
579 G1RegionToSpaceMapper* prev_bitmap_storage,
580 G1RegionToSpaceMapper* next_bitmap_storage);
581 ~G1ConcurrentMark();
582
583 ConcurrentMarkThread* cmThread() { return _cmThread; }
584
585 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
586 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
587
588 // Returns the number of GC threads to be used in a concurrent
589 // phase based on the number of GC threads being used in a STW
590 // phase.
591 uint scale_parallel_threads(uint n_par_threads);
592
593 // Calculates the number of GC threads to be used in a concurrent phase.
594 uint calc_parallel_marking_threads();
595
596 // The following three are interaction between CM and
597 // G1CollectedHeap
598
599 // This notifies CM that a root during initial-mark needs to be
600 // grayed. It is MT-safe. hr is the region that
601 // contains the object and it's passed optionally from callers who
602 // might already have it (no point in recalculating it).
603 inline void grayRoot(oop obj,
604 HeapRegion* hr = NULL);
605
606 // Prepare internal data structures for the next mark cycle. This includes clearing
607 // the next mark bitmap and some internal data structures. This method is intended
608 // to be called concurrently to the mutator. It will yield to safepoint requests.
609 void cleanup_for_next_mark();
610
611 // Clear the previous marking bitmap during safepoint.
612 void clear_prev_bitmap(WorkGang* workers);
613
614 // Return whether the next mark bitmap has no marks set. To be used for assertions
615 // only. Will not yield to pause requests.
616 bool nextMarkBitmapIsClear();
617
618 // These two do the work that needs to be done before and after the
619 // initial root checkpoint. Since this checkpoint can be done at two
620 // different points (i.e. an explicit pause or piggy-backed on a
621 // young collection), then it's nice to be able to easily share the
622 // pre/post code. It might be the case that we can put everything in
623 // the post method. TP
624 void checkpointRootsInitialPre();
625 void checkpointRootsInitialPost();
626
627 // Scan all the root regions and mark everything reachable from
628 // them.
629 void scan_root_regions();
630
631 // Scan a single root region and mark everything reachable from it.
632 void scanRootRegion(HeapRegion* hr);
633
634 // Do concurrent phase of marking, to a tentative transitive closure.
635 void mark_from_roots();
636
637 void checkpointRootsFinal(bool clear_all_soft_refs);
638 void checkpointRootsFinalWork();
639 void cleanup();
640 void complete_cleanup();
641
642 // Mark in the previous bitmap. NB: this is usually read-only, so use
643 // this carefully!
644 inline void markPrev(oop p);
645
646 // Clears marks for all objects in the given range, for the prev or
647 // next bitmaps. NB: the previous bitmap is usually
648 // read-only, so use this carefully!
649 void clearRangePrevBitmap(MemRegion mr);
650
651 // Verify that there are no CSet oops on the stacks (taskqueues /
652 // global mark stack) and fingers (global / per-task).
653 // If marking is not in progress, it's a no-op.
654 void verify_no_cset_oops() PRODUCT_RETURN;
655
656 inline bool isPrevMarked(oop p) const;
657
658 inline bool do_yield_check();
659
660 // Abandon current marking iteration due to a Full GC.
661 void abort();
662
663 bool has_aborted() { return _has_aborted; }
664
665 void print_summary_info();
666
667 void print_worker_threads_on(outputStream* st) const;
668 void threads_do(ThreadClosure* tc) const;
669
670 void print_on_error(outputStream* st) const;
671
672 // Attempts to mark the given object on the next mark bitmap.
673 inline bool par_mark(oop obj);
674
675 // Returns true if initialization was successfully completed.
676 bool completed_initialization() const {
677 return _completed_initialization;
678 }
679
680 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
681 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
682
683 private:
684 // Clear (Reset) all liveness count data.
685 void clear_live_data(WorkGang* workers);
686
687 #ifdef ASSERT
688 // Verify all of the above data structures that they are in initial state.
689 void verify_live_data_clear();
690 #endif
691
692 // Aggregates the per-card liveness data based on the current marking. Also sets
693 // the amount of marked bytes for each region.
694 void create_live_data();
695
696 void finalize_live_data();
697
698 void verify_live_data();
699 };
700
701 // A class representing a marking task.
702 class G1CMTask : public TerminatorTerminator {
703 private:
704 enum PrivateConstants {
705 // The regular clock call is called once the scanned words reaches
706 // this limit
707 words_scanned_period = 12*1024,
708 // The regular clock call is called once the number of visited
709 // references reaches this limit
710 refs_reached_period = 384,
711 // Initial value for the hash seed, used in the work stealing code
712 init_hash_seed = 17
713 };
714
715 uint _worker_id;
716 G1CollectedHeap* _g1h;
717 G1ConcurrentMark* _cm;
718 G1CMBitMap* _nextMarkBitMap;
719 // the task queue of this task
720 G1CMTaskQueue* _task_queue;
721 private:
722 // the task queue set---needed for stealing
723 G1CMTaskQueueSet* _task_queues;
724 // indicates whether the task has been claimed---this is only for
725 // debugging purposes
726 bool _claimed;
727
728 // number of calls to this task
729 int _calls;
730
731 // when the virtual timer reaches this time, the marking step should
732 // exit
733 double _time_target_ms;
734 // the start time of the current marking step
735 double _start_time_ms;
736
737 // the oop closure used for iterations over oops
738 G1CMOopClosure* _cm_oop_closure;
739
740 // the region this task is scanning, NULL if we're not scanning any
741 HeapRegion* _curr_region;
742 // the local finger of this task, NULL if we're not scanning a region
743 HeapWord* _finger;
744 // limit of the region this task is scanning, NULL if we're not scanning one
745 HeapWord* _region_limit;
746
747 // the number of words this task has scanned
748 size_t _words_scanned;
749 // When _words_scanned reaches this limit, the regular clock is
750 // called. Notice that this might be decreased under certain
751 // circumstances (i.e. when we believe that we did an expensive
752 // operation).
753 size_t _words_scanned_limit;
754 // the initial value of _words_scanned_limit (i.e. what it was
755 // before it was decreased).
756 size_t _real_words_scanned_limit;
757
758 // the number of references this task has visited
759 size_t _refs_reached;
760 // When _refs_reached reaches this limit, the regular clock is
761 // called. Notice this this might be decreased under certain
762 // circumstances (i.e. when we believe that we did an expensive
763 // operation).
764 size_t _refs_reached_limit;
765 // the initial value of _refs_reached_limit (i.e. what it was before
766 // it was decreased).
767 size_t _real_refs_reached_limit;
768
769 // used by the work stealing stuff
770 int _hash_seed;
771 // if this is true, then the task has aborted for some reason
772 bool _has_aborted;
773 // set when the task aborts because it has met its time quota
774 bool _has_timed_out;
775 // true when we're draining SATB buffers; this avoids the task
776 // aborting due to SATB buffers being available (as we're already
777 // dealing with them)
778 bool _draining_satb_buffers;
779
780 // number sequence of past step times
781 NumberSeq _step_times_ms;
782 // elapsed time of this task
783 double _elapsed_time_ms;
784 // termination time of this task
785 double _termination_time_ms;
786 // when this task got into the termination protocol
787 double _termination_start_time_ms;
788
789 // true when the task is during a concurrent phase, false when it is
790 // in the remark phase (so, in the latter case, we do not have to
791 // check all the things that we have to check during the concurrent
792 // phase, i.e. SATB buffer availability...)
793 bool _concurrent;
794
795 TruncatedSeq _marking_step_diffs_ms;
796
797 // it updates the local fields after this task has claimed
798 // a new region to scan
799 void setup_for_region(HeapRegion* hr);
800 // it brings up-to-date the limit of the region
801 void update_region_limit();
802
803 // called when either the words scanned or the refs visited limit
804 // has been reached
805 void reached_limit();
806 // recalculates the words scanned and refs visited limits
807 void recalculate_limits();
808 // decreases the words scanned and refs visited limits when we reach
809 // an expensive operation
810 void decrease_limits();
811 // it checks whether the words scanned or refs visited reached their
812 // respective limit and calls reached_limit() if they have
813 void check_limits() {
814 if (_words_scanned >= _words_scanned_limit ||
815 _refs_reached >= _refs_reached_limit) {
816 reached_limit();
817 }
818 }
819 // this is supposed to be called regularly during a marking step as
820 // it checks a bunch of conditions that might cause the marking step
821 // to abort
822 void regular_clock_call();
823 bool concurrent() { return _concurrent; }
824
825 // Test whether obj might have already been passed over by the
826 // mark bitmap scan, and so needs to be pushed onto the mark stack.
827 bool is_below_finger(oop obj, HeapWord* global_finger) const;
828
829 template<bool scan> void process_grey_object(oop obj);
830
831 public:
832 // It resets the task; it should be called right at the beginning of
833 // a marking phase.
834 void reset(G1CMBitMap* _nextMarkBitMap);
835 // it clears all the fields that correspond to a claimed region.
836 void clear_region_fields();
837
838 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
839
840 // The main method of this class which performs a marking step
841 // trying not to exceed the given duration. However, it might exit
842 // prematurely, according to some conditions (i.e. SATB buffers are
843 // available for processing).
844 void do_marking_step(double target_ms,
845 bool do_termination,
846 bool is_serial);
847
848 // These two calls start and stop the timer
849 void record_start_time() {
850 _elapsed_time_ms = os::elapsedTime() * 1000.0;
851 }
852 void record_end_time() {
853 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
854 }
855
856 // returns the worker ID associated with this task.
857 uint worker_id() { return _worker_id; }
858
859 // From TerminatorTerminator. It determines whether this task should
860 // exit the termination protocol after it's entered it.
861 virtual bool should_exit_termination();
862
863 // Resets the local region fields after a task has finished scanning a
864 // region; or when they have become stale as a result of the region
865 // being evacuated.
866 void giveup_current_region();
867
868 HeapWord* finger() { return _finger; }
869
870 bool has_aborted() { return _has_aborted; }
871 void set_has_aborted() { _has_aborted = true; }
872 void clear_has_aborted() { _has_aborted = false; }
873 bool has_timed_out() { return _has_timed_out; }
874 bool claimed() { return _claimed; }
875
876 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
877
878 // Increment the number of references this task has visited.
879 void increment_refs_reached() { ++_refs_reached; }
880
881 // Grey the object by marking it. If not already marked, push it on
882 // the local queue if below the finger.
883 // obj is below its region's NTAMS.
884 inline void make_reference_grey(oop obj);
885
886 // Grey the object (by calling make_grey_reference) if required,
887 // e.g. obj is below its containing region's NTAMS.
888 // Precondition: obj is a valid heap object.
889 inline void deal_with_reference(oop obj);
890
891 // It scans an object and visits its children.
892 inline void scan_object(oop obj);
893
894 // It pushes an object on the local queue.
895 inline void push(oop obj);
896
897 // Move entries to the global stack.
898 void move_entries_to_global_stack();
899 // Move entries from the global stack, return true if we were successful to do so.
900 bool get_entries_from_global_stack();
901
902 // It pops and scans objects from the local queue. If partially is
903 // true, then it stops when the queue size is of a given limit. If
904 // partially is false, then it stops when the queue is empty.
905 void drain_local_queue(bool partially);
906 // It moves entries from the global stack to the local queue and
907 // drains the local queue. If partially is true, then it stops when
908 // both the global stack and the local queue reach a given size. If
909 // partially if false, it tries to empty them totally.
910 void drain_global_stack(bool partially);
911 // It keeps picking SATB buffers and processing them until no SATB
912 // buffers are available.
913 void drain_satb_buffers();
914
915 // moves the local finger to a new location
916 inline void move_finger_to(HeapWord* new_finger) {
917 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
918 _finger = new_finger;
919 }
920
921 G1CMTask(uint worker_id,
922 G1ConcurrentMark *cm,
923 G1CMTaskQueue* task_queue,
924 G1CMTaskQueueSet* task_queues);
925
926 // it prints statistics associated with this task
927 void print_stats();
928 };
929
930 // Class that's used to to print out per-region liveness
931 // information. It's currently used at the end of marking and also
932 // after we sort the old regions at the end of the cleanup operation.
933 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
934 private:
935 // Accumulators for these values.
936 size_t _total_used_bytes;
937 size_t _total_capacity_bytes;
938 size_t _total_prev_live_bytes;
939 size_t _total_next_live_bytes;
940
941 // Accumulator for the remembered set size
942 size_t _total_remset_bytes;
943
944 // Accumulator for strong code roots memory size
945 size_t _total_strong_code_roots_bytes;
946
947 static double perc(size_t val, size_t total) {
948 if (total == 0) {
949 return 0.0;
950 } else {
951 return 100.0 * ((double) val / (double) total);
952 }
953 }
954
955 static double bytes_to_mb(size_t val) {
956 return (double) val / (double) M;
957 }
958
959 public:
960 // The header and footer are printed in the constructor and
961 // destructor respectively.
962 G1PrintRegionLivenessInfoClosure(const char* phase_name);
963 virtual bool doHeapRegion(HeapRegion* r);
964 ~G1PrintRegionLivenessInfoClosure();
965 };
966
967 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP