1 /*
2 * Copyright (c) 2001, 2017, 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/g1ConcurrentMarkObjArrayProcessor.hpp"
30 #include "gc/g1/g1RegionToSpaceMapper.hpp"
31 #include "gc/g1/heapRegionSet.hpp"
32 #include "gc/shared/taskqueue.hpp"
33
34 class G1CollectedHeap;
35 class G1CMBitMap;
36 class G1CMTask;
37 class G1ConcurrentMark;
38 class ConcurrentGCTimer;
39 class G1OldTracer;
40 class G1SurvivorRegions;
41
42 #ifdef _MSC_VER
43 #pragma warning(push)
44 // warning C4522: multiple assignment operators specified
45 #pragma warning(disable:4522)
46 #endif
47
48 // This is a container class for either an oop or a continuation address for
49 // mark stack entries. Both are pushed onto the mark stack.
50 class G1TaskQueueEntry VALUE_OBJ_CLASS_SPEC {
51 private:
52 void* _holder;
53
54 static const uintptr_t ArraySliceBit = 1;
55 public:
56 G1TaskQueueEntry() : _holder(NULL) { }
57 G1TaskQueueEntry(oop obj) : _holder(obj) { }
58 G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
59
60 G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {
61 _holder = t._holder;
62 return *this;
63 }
64
65 volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
66 _holder = t._holder;
67 return *this;
68 }
69
70 oop obj() const {
71 assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
72 return (oop)_holder;
73 }
74
75 HeapWord* slice() const {
76 assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
77 return (HeapWord*)((uintptr_t)_holder &~ ArraySliceBit);
78 }
79
80 bool is_oop() const { return !is_array_slice(); }
81 bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
82 bool is_null() const { return _holder == NULL; }
83 };
84
85 #ifdef _MSC_VER
86 #pragma warning(pop)
87 #endif
88
89 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
90 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
91
92 // Closure used by CM during concurrent reference discovery
93 // and reference processing (during remarking) to determine
94 // if a particular object is alive. It is primarily used
95 // to determine if referents of discovered reference objects
96 // are alive. An instance is also embedded into the
97 // reference processor as the _is_alive_non_header field
98 class G1CMIsAliveClosure: public BoolObjectClosure {
99 G1CollectedHeap* _g1;
100 public:
101 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
102
103 bool do_object_b(oop obj);
104 };
105
106 // A generic CM bit map. This is essentially a wrapper around the BitMap
107 // class, with one bit per (1<<_shifter) HeapWords.
108
109 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
110 protected:
111 HeapWord* _bmStartWord; // base address of range covered by map
112 size_t _bmWordSize; // map size (in #HeapWords covered)
113 const int _shifter; // map to char or bit
114 BitMapView _bm; // the bit map itself
115
116 public:
117 // constructor
118 G1CMBitMapRO(int shifter);
119
120 // inquiries
121 HeapWord* startWord() const { return _bmStartWord; }
122 // the following is one past the last word in space
123 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
124
125 // read marks
126
127 bool isMarked(HeapWord* addr) const {
128 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
129 "outside underlying space?");
130 return _bm.at(heapWordToOffset(addr));
131 }
132
133 // iteration
134 inline bool iterate(BitMapClosure* cl, MemRegion mr);
135
136 // Return the address corresponding to the next marked bit at or after
137 // "addr", and before "limit", if "limit" is non-NULL. If there is no
138 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
139 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
140 const HeapWord* limit = NULL) const;
141
142 // conversion utilities
143 HeapWord* offsetToHeapWord(size_t offset) const {
144 return _bmStartWord + (offset << _shifter);
145 }
146 size_t heapWordToOffset(const HeapWord* addr) const {
147 return pointer_delta(addr, _bmStartWord) >> _shifter;
148 }
149
150 // The argument addr should be the start address of a valid object
151 inline HeapWord* nextObject(HeapWord* addr);
152
153 void print_on_error(outputStream* st, const char* prefix) const;
154
155 // debugging
156 NOT_PRODUCT(bool covers(MemRegion rs) const;)
157 };
158
159 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
160 private:
161 G1CMBitMap* _bm;
162 public:
163 G1CMBitMapMappingChangedListener() : _bm(NULL) {}
164
165 void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
166
167 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
168 };
169
170 class G1CMBitMap : public G1CMBitMapRO {
171 private:
172 G1CMBitMapMappingChangedListener _listener;
173
174 public:
175 static size_t compute_size(size_t heap_size);
176 // Returns the amount of bytes on the heap between two marks in the bitmap.
177 static size_t mark_distance();
178 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
179 // mark bitmap corresponds to. This is the same as the mark distance above.
180 static size_t heap_map_factor() {
181 return mark_distance();
182 }
183
184 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
185
186 // Initializes the underlying BitMap to cover the given area.
187 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
188
189 // Write marks.
190 inline void mark(HeapWord* addr);
191 inline void clear(HeapWord* addr);
192 inline bool parMark(HeapWord* addr);
193
194 void clear_range(MemRegion mr);
195 };
196
197 // Represents the overflow mark stack used by concurrent marking.
198 //
199 // Stores oops in a huge buffer in virtual memory that is always fully committed.
200 // Resizing may only happen during a STW pause when the stack is empty.
201 //
202 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
203 // stack memory is split into evenly sized chunks of oops. Users can only
204 // add or remove entries on that basis.
205 // Chunks are filled in increasing address order. Not completely filled chunks
206 // have a NULL element as a terminating element.
207 //
208 // Every chunk has a header containing a single pointer element used for memory
209 // management. This wastes some space, but is negligible (< .1% with current sizing).
210 //
211 // Memory management is done using a mix of tracking a high water-mark indicating
212 // that all chunks at a lower address are valid chunks, and a singly linked free
213 // list connecting all empty chunks.
214 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
215 public:
216 // Number of oops that can fit in a single chunk.
217 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
218 private:
219 struct OopChunk {
220 OopChunk* next;
221 G1TaskQueueEntry data[EntriesPerChunk];
222 };
223
224 size_t _max_chunk_capacity; // Maximum number of OopChunk elements on the stack.
225
226 OopChunk* _base; // Bottom address of allocated memory area.
227 size_t _chunk_capacity; // Current maximum number of OopChunk elements.
228
229 char _pad0[DEFAULT_CACHE_LINE_SIZE];
230 OopChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users.
231 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(OopChunk*)];
232 OopChunk* volatile _chunk_list; // List of chunks currently containing data.
233 volatile size_t _chunks_in_chunk_list;
234 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(OopChunk*) - sizeof(size_t)];
235
236 volatile size_t _hwm; // High water mark within the reserved space.
237 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
238
239 // Allocate a new chunk from the reserved memory, using the high water mark. Returns
240 // NULL if out of memory.
241 OopChunk* allocate_new_chunk();
242
243 volatile bool _out_of_memory;
244
245 // Atomically add the given chunk to the list.
246 void add_chunk_to_list(OopChunk* volatile* list, OopChunk* elem);
247 // Atomically remove and return a chunk from the given list. Returns NULL if the
248 // list is empty.
249 OopChunk* remove_chunk_from_list(OopChunk* volatile* list);
250
251 void add_chunk_to_chunk_list(OopChunk* elem);
252 void add_chunk_to_free_list(OopChunk* elem);
253
254 OopChunk* remove_chunk_from_chunk_list();
255 OopChunk* remove_chunk_from_free_list();
256
257 bool _should_expand;
258
259 // Resizes the mark stack to the given new capacity. Releases any previous
260 // memory if successful.
261 bool resize(size_t new_capacity);
262
263 public:
264 G1CMMarkStack();
265 ~G1CMMarkStack();
266
267 // Alignment and minimum capacity of this mark stack in number of oops.
268 static size_t capacity_alignment();
269
270 // Allocate and initialize the mark stack with the given number of oops.
271 bool initialize(size_t initial_capacity, size_t max_capacity);
272
273 // Pushes the given buffer containing at most OopsPerChunk elements on the mark
274 // stack. If less than OopsPerChunk elements are to be pushed, the array must
275 // be terminated with a NULL.
276 // Returns whether the buffer contents were successfully pushed to the global mark
277 // stack.
278 bool par_push_chunk(G1TaskQueueEntry* buffer);
279
280 // Pops a chunk from this mark stack, copying them into the given buffer. This
281 // chunk may contain up to OopsPerChunk elements. If there are less, the last
282 // element in the array is a NULL pointer.
283 bool par_pop_chunk(G1TaskQueueEntry* buffer);
284
285 // Return whether the chunk list is empty. Racy due to unsynchronized access to
286 // _chunk_list.
287 bool is_empty() const { return _chunk_list == NULL; }
288
289 size_t capacity() const { return _chunk_capacity; }
290
291 bool is_out_of_memory() const { return _out_of_memory; }
292 void clear_out_of_memory() { _out_of_memory = false; }
293
294 bool should_expand() const { return _should_expand; }
295 void set_should_expand(bool value) { _should_expand = value; }
296
297 // Expand the stack, typically in response to an overflow condition
298 void expand();
299
300 // Return the approximate number of oops on this mark stack. Racy due to
301 // unsynchronized access to _chunks_in_chunk_list.
302 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
303
304 void set_empty();
305
306 // Apply Fn to every oop on the mark stack. The mark stack must not
307 // be modified while iterating.
308 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
309 };
310
311 // Root Regions are regions that are not empty at the beginning of a
312 // marking cycle and which we might collect during an evacuation pause
313 // while the cycle is active. Given that, during evacuation pauses, we
314 // do not copy objects that are explicitly marked, what we have to do
315 // for the root regions is to scan them and mark all objects reachable
316 // from them. According to the SATB assumptions, we only need to visit
317 // each object once during marking. So, as long as we finish this scan
318 // before the next evacuation pause, we can copy the objects from the
319 // root regions without having to mark them or do anything else to them.
320 //
321 // Currently, we only support root region scanning once (at the start
322 // of the marking cycle) and the root regions are all the survivor
323 // regions populated during the initial-mark pause.
324 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
325 private:
326 const G1SurvivorRegions* _survivors;
327 G1ConcurrentMark* _cm;
328
329 volatile bool _scan_in_progress;
330 volatile bool _should_abort;
331 volatile int _claimed_survivor_index;
332
333 void notify_scan_done();
334
335 public:
336 G1CMRootRegions();
337 // We actually do most of the initialization in this method.
338 void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
339
340 // Reset the claiming / scanning of the root regions.
341 void prepare_for_scan();
342
343 // Forces get_next() to return NULL so that the iteration aborts early.
344 void abort() { _should_abort = true; }
345
346 // Return true if the CM thread are actively scanning root regions,
347 // false otherwise.
348 bool scan_in_progress() { return _scan_in_progress; }
349
350 // Claim the next root region to scan atomically, or return NULL if
351 // all have been claimed.
352 HeapRegion* claim_next();
353
354 // The number of root regions to scan.
355 uint num_root_regions() const;
356
357 void cancel_scan();
358
359 // Flag that we're done with root region scanning and notify anyone
360 // who's waiting on it. If aborted is false, assume that all regions
361 // have been claimed.
362 void scan_finished();
363
364 // If CM threads are still scanning root regions, wait until they
365 // are done. Return true if we had to wait, false otherwise.
366 bool wait_until_scan_finished();
367 };
368
369 class ConcurrentMarkThread;
370
371 class G1ConcurrentMark: public CHeapObj<mtGC> {
372 friend class ConcurrentMarkThread;
373 friend class G1ParNoteEndTask;
374 friend class G1VerifyLiveDataClosure;
375 friend class G1CMRefProcTaskProxy;
376 friend class G1CMRefProcTaskExecutor;
377 friend class G1CMKeepAliveAndDrainClosure;
378 friend class G1CMDrainMarkingStackClosure;
379 friend class G1CMBitMapClosure;
380 friend class G1CMConcurrentMarkingTask;
381 friend class G1CMRemarkTask;
382 friend class G1CMTask;
383
384 protected:
385 ConcurrentMarkThread* _cmThread; // The thread doing the work
386 G1CollectedHeap* _g1h; // The heap
387 uint _parallel_marking_threads; // The number of marking
388 // threads we're using
389 uint _max_parallel_marking_threads; // Max number of marking
390 // threads we'll ever use
391 double _sleep_factor; // How much we have to sleep, with
392 // respect to the work we just did, to
393 // meet the marking overhead goal
394 double _marking_task_overhead; // Marking target overhead for
395 // a single task
396
397 FreeRegionList _cleanup_list;
398
399 // Concurrent marking support structures
400 G1CMBitMap _markBitMap1;
401 G1CMBitMap _markBitMap2;
402 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
403 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
404
405 // Heap bounds
406 HeapWord* _heap_start;
407 HeapWord* _heap_end;
408
409 // Root region tracking and claiming
410 G1CMRootRegions _root_regions;
411
412 // For gray objects
413 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger
414 HeapWord* volatile _finger; // The global finger, region aligned,
415 // always points to the end of the
416 // last claimed region
417
418 // Marking tasks
419 uint _max_worker_id;// Maximum worker id
420 uint _active_tasks; // Task num currently active
421 G1CMTask** _tasks; // Task queue array (max_worker_id len)
422 G1CMTaskQueueSet* _task_queues; // Task queue set
423 ParallelTaskTerminator _terminator; // For termination
424
425 // Two sync barriers that are used to synchronize tasks when an
426 // overflow occurs. The algorithm is the following. All tasks enter
427 // the first one to ensure that they have all stopped manipulating
428 // the global data structures. After they exit it, they re-initialize
429 // their data structures and task 0 re-initializes the global data
430 // structures. Then, they enter the second sync barrier. This
431 // ensure, that no task starts doing work before all data
432 // structures (local and global) have been re-initialized. When they
433 // exit it, they are free to start working again.
434 WorkGangBarrierSync _first_overflow_barrier_sync;
435 WorkGangBarrierSync _second_overflow_barrier_sync;
436
437 // This is set by any task, when an overflow on the global data
438 // structures is detected
439 volatile bool _has_overflown;
440 // True: marking is concurrent, false: we're in remark
441 volatile bool _concurrent;
442 // Set at the end of a Full GC so that marking aborts
443 volatile bool _has_aborted;
444
445 // Used when remark aborts due to an overflow to indicate that
446 // another concurrent marking phase should start
447 volatile bool _restart_for_overflow;
448
449 // This is true from the very start of concurrent marking until the
450 // point when all the tasks complete their work. It is really used
451 // to determine the points between the end of concurrent marking and
452 // time of remark.
453 volatile bool _concurrent_marking_in_progress;
454
455 ConcurrentGCTimer* _gc_timer_cm;
456
457 G1OldTracer* _gc_tracer_cm;
458
459 // All of these times are in ms
460 NumberSeq _init_times;
461 NumberSeq _remark_times;
462 NumberSeq _remark_mark_times;
463 NumberSeq _remark_weak_ref_times;
464 NumberSeq _cleanup_times;
465 double _total_counting_time;
466 double _total_rs_scrub_time;
467
468 double* _accum_task_vtime; // Accumulated task vtime
469
470 WorkGang* _parallel_workers;
471
472 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
473 void weakRefsWork(bool clear_all_soft_refs);
474
475 void swapMarkBitMaps();
476
477 // It resets the global marking data structures, as well as the
478 // task local ones; should be called during initial mark.
479 void reset();
480
481 // Resets all the marking data structures. Called when we have to restart
482 // marking or when marking completes (via set_non_marking_state below).
483 void reset_marking_state(bool clear_overflow = true);
484
485 // We do this after we're done with marking so that the marking data
486 // structures are initialized to a sensible and predictable state.
487 void set_non_marking_state();
488
489 // Called to indicate how many threads are currently active.
490 void set_concurrency(uint active_tasks);
491
492 // It should be called to indicate which phase we're in (concurrent
493 // mark or remark) and how many threads are currently active.
494 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
495
496 // Prints all gathered CM-related statistics
497 void print_stats();
498
499 bool cleanup_list_is_empty() {
500 return _cleanup_list.is_empty();
501 }
502
503 // Accessor methods
504 uint parallel_marking_threads() const { return _parallel_marking_threads; }
505 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
506 double sleep_factor() { return _sleep_factor; }
507 double marking_task_overhead() { return _marking_task_overhead;}
508
509 HeapWord* finger() { return _finger; }
510 bool concurrent() { return _concurrent; }
511 uint active_tasks() { return _active_tasks; }
512 ParallelTaskTerminator* terminator() { return &_terminator; }
513
514 // It claims the next available region to be scanned by a marking
515 // task/thread. It might return NULL if the next region is empty or
516 // we have run out of regions. In the latter case, out_of_regions()
517 // determines whether we've really run out of regions or the task
518 // should call claim_region() again. This might seem a bit
519 // awkward. Originally, the code was written so that claim_region()
520 // either successfully returned with a non-empty region or there
521 // were no more regions to be claimed. The problem with this was
522 // that, in certain circumstances, it iterated over large chunks of
523 // the heap finding only empty regions and, while it was working, it
524 // was preventing the calling task to call its regular clock
525 // method. So, this way, each task will spend very little time in
526 // claim_region() and is allowed to call the regular clock method
527 // frequently.
528 HeapRegion* claim_region(uint worker_id);
529
530 // It determines whether we've run out of regions to scan. Note that
531 // the finger can point past the heap end in case the heap was expanded
532 // to satisfy an allocation without doing a GC. This is fine, because all
533 // objects in those regions will be considered live anyway because of
534 // SATB guarantees (i.e. their TAMS will be equal to bottom).
535 bool out_of_regions() { return _finger >= _heap_end; }
536
537 // Returns the task with the given id
538 G1CMTask* task(int id) {
539 assert(0 <= id && id < (int) _active_tasks,
540 "task id not within active bounds");
541 return _tasks[id];
542 }
543
544 // Returns the task queue with the given id
545 G1CMTaskQueue* task_queue(int id) {
546 assert(0 <= id && id < (int) _active_tasks,
547 "task queue id not within active bounds");
548 return (G1CMTaskQueue*) _task_queues->queue(id);
549 }
550
551 // Returns the task queue set
552 G1CMTaskQueueSet* task_queues() { return _task_queues; }
553
554 // Access / manipulation of the overflow flag which is set to
555 // indicate that the global stack has overflown
556 bool has_overflown() { return _has_overflown; }
557 void set_has_overflown() { _has_overflown = true; }
558 void clear_has_overflown() { _has_overflown = false; }
559 bool restart_for_overflow() { return _restart_for_overflow; }
560
561 // Methods to enter the two overflow sync barriers
562 void enter_first_sync_barrier(uint worker_id);
563 void enter_second_sync_barrier(uint worker_id);
564
565 // Card index of the bottom of the G1 heap. Used for biasing indices into
566 // the card bitmaps.
567 intptr_t _heap_bottom_card_num;
568
569 // Set to true when initialization is complete
570 bool _completed_initialization;
571
572 // end_timer, true to end gc timer after ending concurrent phase.
573 void register_concurrent_phase_end_common(bool end_timer);
574
575 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
576 // true, periodically insert checks to see if this method should exit prematurely.
577 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
578 public:
579 // Manipulation of the global mark stack.
580 // The push and pop operations are used by tasks for transfers
581 // between task-local queues and the global mark stack.
582 bool mark_stack_push(G1TaskQueueEntry* arr) {
583 if (!_global_mark_stack.par_push_chunk(arr)) {
584 set_has_overflown();
585 return false;
586 }
587 return true;
588 }
589 bool mark_stack_pop(G1TaskQueueEntry* arr) {
590 return _global_mark_stack.par_pop_chunk(arr);
591 }
592 size_t mark_stack_size() { return _global_mark_stack.size(); }
593 size_t partial_mark_stack_size_target() { return _global_mark_stack.capacity()/3; }
594 bool mark_stack_overflow() { return _global_mark_stack.is_out_of_memory(); }
595 bool mark_stack_empty() { return _global_mark_stack.is_empty(); }
596
597 G1CMRootRegions* root_regions() { return &_root_regions; }
598
599 bool concurrent_marking_in_progress() {
600 return _concurrent_marking_in_progress;
601 }
602 void set_concurrent_marking_in_progress() {
603 _concurrent_marking_in_progress = true;
604 }
605 void clear_concurrent_marking_in_progress() {
606 _concurrent_marking_in_progress = false;
607 }
608
609 void concurrent_cycle_start();
610 void concurrent_cycle_end();
611
612 void update_accum_task_vtime(int i, double vtime) {
613 _accum_task_vtime[i] += vtime;
614 }
615
616 double all_task_accum_vtime() {
617 double ret = 0.0;
618 for (uint i = 0; i < _max_worker_id; ++i)
619 ret += _accum_task_vtime[i];
620 return ret;
621 }
622
623 // Attempts to steal an object from the task queues of other tasks
624 bool try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry);
625
626 G1ConcurrentMark(G1CollectedHeap* g1h,
627 G1RegionToSpaceMapper* prev_bitmap_storage,
628 G1RegionToSpaceMapper* next_bitmap_storage);
629 ~G1ConcurrentMark();
630
631 ConcurrentMarkThread* cmThread() { return _cmThread; }
632
633 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
634 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
635
636 // Returns the number of GC threads to be used in a concurrent
637 // phase based on the number of GC threads being used in a STW
638 // phase.
639 uint scale_parallel_threads(uint n_par_threads);
640
641 // Calculates the number of GC threads to be used in a concurrent phase.
642 uint calc_parallel_marking_threads();
643
644 // The following three are interaction between CM and
645 // G1CollectedHeap
646
647 // This notifies CM that a root during initial-mark needs to be
648 // grayed. It is MT-safe. hr is the region that
649 // contains the object and it's passed optionally from callers who
650 // might already have it (no point in recalculating it).
651 inline void grayRoot(oop obj,
652 HeapRegion* hr = NULL);
653
654 // Prepare internal data structures for the next mark cycle. This includes clearing
655 // the next mark bitmap and some internal data structures. This method is intended
656 // to be called concurrently to the mutator. It will yield to safepoint requests.
657 void cleanup_for_next_mark();
658
659 // Clear the previous marking bitmap during safepoint.
660 void clear_prev_bitmap(WorkGang* workers);
661
662 // Return whether the next mark bitmap has no marks set. To be used for assertions
663 // only. Will not yield to pause requests.
664 bool nextMarkBitmapIsClear();
665
666 // These two do the work that needs to be done before and after the
667 // initial root checkpoint. Since this checkpoint can be done at two
668 // different points (i.e. an explicit pause or piggy-backed on a
669 // young collection), then it's nice to be able to easily share the
670 // pre/post code. It might be the case that we can put everything in
671 // the post method. TP
672 void checkpointRootsInitialPre();
673 void checkpointRootsInitialPost();
674
675 // Scan all the root regions and mark everything reachable from
676 // them.
677 void scan_root_regions();
678
679 // Scan a single root region and mark everything reachable from it.
680 void scanRootRegion(HeapRegion* hr);
681
682 // Do concurrent phase of marking, to a tentative transitive closure.
683 void mark_from_roots();
684
685 void checkpointRootsFinal(bool clear_all_soft_refs);
686 void checkpointRootsFinalWork();
687 void cleanup();
688 void complete_cleanup();
689
690 // Mark in the previous bitmap. NB: this is usually read-only, so use
691 // this carefully!
692 inline void markPrev(oop p);
693
694 // Clears marks for all objects in the given range, for the prev or
695 // next bitmaps. NB: the previous bitmap is usually
696 // read-only, so use this carefully!
697 void clearRangePrevBitmap(MemRegion mr);
698
699 // Verify that there are no CSet oops on the stacks (taskqueues /
700 // global mark stack) and fingers (global / per-task).
701 // If marking is not in progress, it's a no-op.
702 void verify_no_cset_oops() PRODUCT_RETURN;
703
704 inline bool isPrevMarked(oop p) const;
705
706 inline bool do_yield_check();
707
708 // Abandon current marking iteration due to a Full GC.
709 void abort();
710
711 bool has_aborted() { return _has_aborted; }
712
713 void print_summary_info();
714
715 void print_worker_threads_on(outputStream* st) const;
716 void threads_do(ThreadClosure* tc) const;
717
718 void print_on_error(outputStream* st) const;
719
720 // Attempts to mark the given object on the next mark bitmap.
721 inline bool par_mark(oop obj);
722
723 // Returns true if initialization was successfully completed.
724 bool completed_initialization() const {
725 return _completed_initialization;
726 }
727
728 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
729 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
730
731 private:
732 // Clear (Reset) all liveness count data.
733 void clear_live_data(WorkGang* workers);
734
735 #ifdef ASSERT
736 // Verify all of the above data structures that they are in initial state.
737 void verify_live_data_clear();
738 #endif
739
740 // Aggregates the per-card liveness data based on the current marking. Also sets
741 // the amount of marked bytes for each region.
742 void create_live_data();
743
744 void finalize_live_data();
745
746 void verify_live_data();
747 };
748
749 // A class representing a marking task.
750 class G1CMTask : public TerminatorTerminator {
751 private:
752 enum PrivateConstants {
753 // The regular clock call is called once the scanned words reaches
754 // this limit
755 words_scanned_period = 12*1024,
756 // The regular clock call is called once the number of visited
757 // references reaches this limit
758 refs_reached_period = 1024,
759 // Initial value for the hash seed, used in the work stealing code
760 init_hash_seed = 17
761 };
762
763 G1CMObjArrayProcessor _objArray_processor;
764
765 uint _worker_id;
766 G1CollectedHeap* _g1h;
767 G1ConcurrentMark* _cm;
768 G1CMBitMap* _nextMarkBitMap;
769 // the task queue of this task
770 G1CMTaskQueue* _task_queue;
771 private:
772 // the task queue set---needed for stealing
773 G1CMTaskQueueSet* _task_queues;
774 // indicates whether the task has been claimed---this is only for
775 // debugging purposes
776 bool _claimed;
777
778 // number of calls to this task
779 int _calls;
780
781 // when the virtual timer reaches this time, the marking step should
782 // exit
783 double _time_target_ms;
784 // the start time of the current marking step
785 double _start_time_ms;
786
787 // the oop closure used for iterations over oops
788 G1CMOopClosure* _cm_oop_closure;
789
790 // the region this task is scanning, NULL if we're not scanning any
791 HeapRegion* _curr_region;
792 // the local finger of this task, NULL if we're not scanning a region
793 HeapWord* _finger;
794 // limit of the region this task is scanning, NULL if we're not scanning one
795 HeapWord* _region_limit;
796
797 // the number of words this task has scanned
798 size_t _words_scanned;
799 // When _words_scanned reaches this limit, the regular clock is
800 // called. Notice that this might be decreased under certain
801 // circumstances (i.e. when we believe that we did an expensive
802 // operation).
803 size_t _words_scanned_limit;
804 // the initial value of _words_scanned_limit (i.e. what it was
805 // before it was decreased).
806 size_t _real_words_scanned_limit;
807
808 // the number of references this task has visited
809 size_t _refs_reached;
810 // When _refs_reached reaches this limit, the regular clock is
811 // called. Notice this this might be decreased under certain
812 // circumstances (i.e. when we believe that we did an expensive
813 // operation).
814 size_t _refs_reached_limit;
815 // the initial value of _refs_reached_limit (i.e. what it was before
816 // it was decreased).
817 size_t _real_refs_reached_limit;
818
819 // used by the work stealing stuff
820 int _hash_seed;
821 // if this is true, then the task has aborted for some reason
822 bool _has_aborted;
823 // set when the task aborts because it has met its time quota
824 bool _has_timed_out;
825 // true when we're draining SATB buffers; this avoids the task
826 // aborting due to SATB buffers being available (as we're already
827 // dealing with them)
828 bool _draining_satb_buffers;
829
830 // number sequence of past step times
831 NumberSeq _step_times_ms;
832 // elapsed time of this task
833 double _elapsed_time_ms;
834 // termination time of this task
835 double _termination_time_ms;
836 // when this task got into the termination protocol
837 double _termination_start_time_ms;
838
839 // true when the task is during a concurrent phase, false when it is
840 // in the remark phase (so, in the latter case, we do not have to
841 // check all the things that we have to check during the concurrent
842 // phase, i.e. SATB buffer availability...)
843 bool _concurrent;
844
845 TruncatedSeq _marking_step_diffs_ms;
846
847 // it updates the local fields after this task has claimed
848 // a new region to scan
849 void setup_for_region(HeapRegion* hr);
850 // it brings up-to-date the limit of the region
851 void update_region_limit();
852
853 // called when either the words scanned or the refs visited limit
854 // has been reached
855 void reached_limit();
856 // recalculates the words scanned and refs visited limits
857 void recalculate_limits();
858 // decreases the words scanned and refs visited limits when we reach
859 // an expensive operation
860 void decrease_limits();
861 // it checks whether the words scanned or refs visited reached their
862 // respective limit and calls reached_limit() if they have
863 void check_limits() {
864 if (_words_scanned >= _words_scanned_limit ||
865 _refs_reached >= _refs_reached_limit) {
866 reached_limit();
867 }
868 }
869 // this is supposed to be called regularly during a marking step as
870 // it checks a bunch of conditions that might cause the marking step
871 // to abort
872 void regular_clock_call();
873 bool concurrent() { return _concurrent; }
874
875 // Test whether obj might have already been passed over by the
876 // mark bitmap scan, and so needs to be pushed onto the mark stack.
877 bool is_below_finger(oop obj, HeapWord* global_finger) const;
878
879 template<bool scan> void process_grey_object(G1TaskQueueEntry task_entry);
880 public:
881 // Apply the closure on the given area of the objArray. Return the number of words
882 // scanned.
883 inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
884 // It resets the task; it should be called right at the beginning of
885 // a marking phase.
886 void reset(G1CMBitMap* _nextMarkBitMap);
887 // it clears all the fields that correspond to a claimed region.
888 void clear_region_fields();
889
890 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
891
892 // The main method of this class which performs a marking step
893 // trying not to exceed the given duration. However, it might exit
894 // prematurely, according to some conditions (i.e. SATB buffers are
895 // available for processing).
896 void do_marking_step(double target_ms,
897 bool do_termination,
898 bool is_serial);
899
900 // These two calls start and stop the timer
901 void record_start_time() {
902 _elapsed_time_ms = os::elapsedTime() * 1000.0;
903 }
904 void record_end_time() {
905 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
906 }
907
908 // returns the worker ID associated with this task.
909 uint worker_id() { return _worker_id; }
910
911 // From TerminatorTerminator. It determines whether this task should
912 // exit the termination protocol after it's entered it.
913 virtual bool should_exit_termination();
914
915 // Resets the local region fields after a task has finished scanning a
916 // region; or when they have become stale as a result of the region
917 // being evacuated.
918 void giveup_current_region();
919
920 HeapWord* finger() { return _finger; }
921
922 bool has_aborted() { return _has_aborted; }
923 void set_has_aborted() { _has_aborted = true; }
924 void clear_has_aborted() { _has_aborted = false; }
925 bool has_timed_out() { return _has_timed_out; }
926 bool claimed() { return _claimed; }
927
928 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
929
930 // Increment the number of references this task has visited.
931 void increment_refs_reached() { ++_refs_reached; }
932
933 // Grey the object by marking it. If not already marked, push it on
934 // the local queue if below the finger.
935 // obj is below its region's NTAMS.
936 inline void make_reference_grey(oop obj);
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(G1TaskQueueEntry task_entry);
945
946 // It pushes an object on the local queue.
947 inline void push(G1TaskQueueEntry task_entry);
948
949 // Move entries to the global stack.
950 void move_entries_to_global_stack();
951 // Move entries from the global stack, return true if we were successful to do so.
952 bool get_entries_from_global_stack();
953
954 // It pops and scans objects from the local queue. If partially is
955 // true, then it stops when the queue size is of a given limit. If
956 // partially is false, then it stops when the queue is empty.
957 void drain_local_queue(bool partially);
958 // It moves entries from the global stack to the local queue and
959 // drains the local queue. If partially is true, then it stops when
960 // both the global stack and the local queue reach a given size. If
961 // partially if false, it tries to empty them totally.
962 void drain_global_stack(bool partially);
963 // It keeps picking SATB buffers and processing them until no SATB
964 // buffers are available.
965 void drain_satb_buffers();
966
967 // moves the local finger to a new location
968 inline void move_finger_to(HeapWord* new_finger) {
969 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
970 _finger = new_finger;
971 }
972
973 G1CMTask(uint worker_id,
974 G1ConcurrentMark *cm,
975 G1CMTaskQueue* task_queue,
976 G1CMTaskQueueSet* task_queues);
977
978 // it prints statistics associated with this task
979 void print_stats();
980 };
981
982 // Class that's used to to print out per-region liveness
983 // information. It's currently used at the end of marking and also
984 // after we sort the old regions at the end of the cleanup operation.
985 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
986 private:
987 // Accumulators for these values.
988 size_t _total_used_bytes;
989 size_t _total_capacity_bytes;
990 size_t _total_prev_live_bytes;
991 size_t _total_next_live_bytes;
992
993 // Accumulator for the remembered set size
994 size_t _total_remset_bytes;
995
996 // Accumulator for strong code roots memory size
997 size_t _total_strong_code_roots_bytes;
998
999 static double perc(size_t val, size_t total) {
1000 if (total == 0) {
1001 return 0.0;
1002 } else {
1003 return 100.0 * ((double) val / (double) total);
1004 }
1005 }
1006
1007 static double bytes_to_mb(size_t val) {
1008 return (double) val / (double) M;
1009 }
1010
1011 public:
1012 // The header and footer are printed in the constructor and
1013 // destructor respectively.
1014 G1PrintRegionLivenessInfoClosure(const char* phase_name);
1015 virtual bool doHeapRegion(HeapRegion* r);
1016 ~G1PrintRegionLivenessInfoClosure();
1017 };
1018
1019 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP