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