1 /*
2 * Copyright (c) 2001, 2020, 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_GC_G1_G1CONCURRENTMARK_HPP
26 #define SHARE_GC_G1_G1CONCURRENTMARK_HPP
27
28 #include "gc/g1/g1ConcurrentMarkBitMap.hpp"
29 #include "gc/g1/g1ConcurrentMarkObjArrayProcessor.hpp"
30 #include "gc/g1/g1HeapVerifier.hpp"
31 #include "gc/g1/g1RegionMarkStatsCache.hpp"
32 #include "gc/g1/heapRegionSet.hpp"
33 #include "gc/shared/gcCause.hpp"
34 #include "gc/shared/taskTerminator.hpp"
35 #include "gc/shared/taskqueue.hpp"
36 #include "gc/shared/verifyOption.hpp"
37 #include "gc/shared/workgroup.hpp"
38 #include "memory/allocation.hpp"
39 #include "utilities/compilerWarnings.hpp"
40 #include "utilities/numberSeq.hpp"
41
42 class ConcurrentGCTimer;
43 class G1ConcurrentMarkThread;
44 class G1CollectedHeap;
45 class G1CMOopClosure;
46 class G1CMTask;
47 class G1ConcurrentMark;
48 class G1OldTracer;
49 class G1RegionToSpaceMapper;
50 class G1SurvivorRegions;
51 class ThreadClosure;
52
53 PRAGMA_DIAG_PUSH
54 // warning C4522: multiple assignment operators specified
55 PRAGMA_DISABLE_MSVC_WARNING(4522)
56
57 // This is a container class for either an oop or a continuation address for
58 // mark stack entries. Both are pushed onto the mark stack.
59 class G1TaskQueueEntry {
60 private:
61 void* _holder;
62
63 static const uintptr_t ArraySliceBit = 1;
64
65 G1TaskQueueEntry(oop obj) : _holder(obj) {
66 assert(_holder != NULL, "Not allowed to set NULL task queue element");
67 }
68 G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
69 public:
70
71 G1TaskQueueEntry() : _holder(NULL) { }
72 // Trivially copyable, for use in GenericTaskQueue.
73
74 static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
75 static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
76
77 oop obj() const {
78 assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
79 return (oop)_holder;
80 }
81
82 HeapWord* slice() const {
83 assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
84 return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
85 }
86
87 bool is_oop() const { return !is_array_slice(); }
88 bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
89 bool is_null() const { return _holder == NULL; }
90 };
91
92 PRAGMA_DIAG_POP
93
94 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
95 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
96
97 // Closure used by CM during concurrent reference discovery
98 // and reference processing (during remarking) to determine
99 // if a particular object is alive. It is primarily used
100 // to determine if referents of discovered reference objects
101 // are alive. An instance is also embedded into the
102 // reference processor as the _is_alive_non_header field
103 class G1CMIsAliveClosure : public BoolObjectClosure {
104 G1CollectedHeap* _g1h;
105 public:
106 G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
107 bool do_object_b(oop obj);
108 };
109
110 class G1CMSubjectToDiscoveryClosure : public BoolObjectClosure {
111 G1CollectedHeap* _g1h;
112 public:
113 G1CMSubjectToDiscoveryClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
114 bool do_object_b(oop obj);
115 };
116
117 // Represents the overflow mark stack used by concurrent marking.
118 //
119 // Stores oops in a huge buffer in virtual memory that is always fully committed.
120 // Resizing may only happen during a STW pause when the stack is empty.
121 //
122 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
123 // stack memory is split into evenly sized chunks of oops. Users can only
124 // add or remove entries on that basis.
125 // Chunks are filled in increasing address order. Not completely filled chunks
126 // have a NULL element as a terminating element.
127 //
128 // Every chunk has a header containing a single pointer element used for memory
129 // management. This wastes some space, but is negligible (< .1% with current sizing).
130 //
131 // Memory management is done using a mix of tracking a high water-mark indicating
132 // that all chunks at a lower address are valid chunks, and a singly linked free
133 // list connecting all empty chunks.
134 class G1CMMarkStack {
135 public:
136 // Number of TaskQueueEntries that can fit in a single chunk.
137 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
138 private:
139 struct TaskQueueEntryChunk {
140 TaskQueueEntryChunk* next;
141 G1TaskQueueEntry data[EntriesPerChunk];
142 };
143
144 size_t _max_chunk_capacity; // Maximum number of TaskQueueEntryChunk elements on the stack.
145
146 TaskQueueEntryChunk* _base; // Bottom address of allocated memory area.
147 size_t _chunk_capacity; // Current maximum number of TaskQueueEntryChunk elements.
148
149 char _pad0[DEFAULT_CACHE_LINE_SIZE];
150 TaskQueueEntryChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users.
151 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
152 TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
153 volatile size_t _chunks_in_chunk_list;
154 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
155
156 volatile size_t _hwm; // High water mark within the reserved space.
157 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
158
159 // Allocate a new chunk from the reserved memory, using the high water mark. Returns
160 // NULL if out of memory.
161 TaskQueueEntryChunk* allocate_new_chunk();
162
163 // Atomically add the given chunk to the list.
164 void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
165 // Atomically remove and return a chunk from the given list. Returns NULL if the
166 // list is empty.
167 TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
168
169 void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
170 void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
171
172 TaskQueueEntryChunk* remove_chunk_from_chunk_list();
173 TaskQueueEntryChunk* remove_chunk_from_free_list();
174
175 // Resizes the mark stack to the given new capacity. Releases any previous
176 // memory if successful.
177 bool resize(size_t new_capacity);
178
179 public:
180 G1CMMarkStack();
181 ~G1CMMarkStack();
182
183 // Alignment and minimum capacity of this mark stack in number of oops.
184 static size_t capacity_alignment();
185
186 // Allocate and initialize the mark stack with the given number of oops.
187 bool initialize(size_t initial_capacity, size_t max_capacity);
188
189 // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
190 // stack. If less than EntriesPerChunk elements are to be pushed, the array must
191 // be terminated with a NULL.
192 // Returns whether the buffer contents were successfully pushed to the global mark
193 // stack.
194 bool par_push_chunk(G1TaskQueueEntry* buffer);
195
196 // Pops a chunk from this mark stack, copying them into the given buffer. This
197 // chunk may contain up to EntriesPerChunk elements. If there are less, the last
198 // element in the array is a NULL pointer.
199 bool par_pop_chunk(G1TaskQueueEntry* buffer);
200
201 // Return whether the chunk list is empty. Racy due to unsynchronized access to
202 // _chunk_list.
203 bool is_empty() const { return _chunk_list == NULL; }
204
205 size_t capacity() const { return _chunk_capacity; }
206
207 // Expand the stack, typically in response to an overflow condition
208 void expand();
209
210 // Return the approximate number of oops on this mark stack. Racy due to
211 // unsynchronized access to _chunks_in_chunk_list.
212 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
213
214 void set_empty();
215
216 // Apply Fn to every oop on the mark stack. The mark stack must not
217 // be modified while iterating.
218 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
219 };
220
221 // Root MemRegions are memory areas that contain objects which references are
222 // roots wrt to the marking. They must be scanned before marking to maintain the
223 // SATB invariant.
224 // Typically they contain the areas from nTAMS to top of the regions.
225 // We could scan and mark through these objects during the concurrent start pause,
226 // but for pause time reasons we move this work to the concurrent phase.
227 // We need to complete this procedure before the next GC because it might determine
228 // that some of these "root objects" are dead, potentially dropping some required
229 // references.
230 // Root MemRegions comprise of the contents of survivor regions at the end
231 // of the GC, and any objects copied into the old gen during GC.
232 class G1CMRootMemRegions {
233 // The set of root MemRegions.
234 MemRegion* _root_regions;
235 size_t const _max_regions;
236
237 volatile size_t _num_root_regions; // Actual number of root regions.
238
239 volatile size_t _claimed_root_regions; // Number of root regions currently claimed.
240
241 volatile bool _scan_in_progress;
242 volatile bool _should_abort;
243
244 void notify_scan_done();
245
246 public:
247 G1CMRootMemRegions(uint const max_regions);
248 ~G1CMRootMemRegions();
249
250 // Reset the data structure to allow addition of new root regions.
251 void reset();
252
253 void add(HeapWord* start, HeapWord* end);
254
255 // Reset the claiming / scanning of the root regions.
256 void prepare_for_scan();
257
258 // Forces get_next() to return NULL so that the iteration aborts early.
259 void abort() { _should_abort = true; }
260
261 // Return true if the CM thread are actively scanning root regions,
262 // false otherwise.
263 bool scan_in_progress() { return _scan_in_progress; }
264
265 // Claim the next root MemRegion to scan atomically, or return NULL if
266 // all have been claimed.
267 const MemRegion* claim_next();
268
269 // The number of root regions to scan.
270 uint num_root_regions() const;
271
272 void cancel_scan();
273
274 // Flag that we're done with root region scanning and notify anyone
275 // who's waiting on it. If aborted is false, assume that all regions
276 // have been claimed.
277 void scan_finished();
278
279 // If CM threads are still scanning root regions, wait until they
280 // are done. Return true if we had to wait, false otherwise.
281 bool wait_until_scan_finished();
282 };
283
284 // This class manages data structures and methods for doing liveness analysis in
285 // G1's concurrent cycle.
286 class G1ConcurrentMark : public CHeapObj<mtGC> {
287 friend class G1ConcurrentMarkThread;
288 friend class G1CMRefProcTaskProxy;
289 friend class G1CMRefProcTaskExecutor;
290 friend class G1CMKeepAliveAndDrainClosure;
291 friend class G1CMDrainMarkingStackClosure;
292 friend class G1CMBitMapClosure;
293 friend class G1CMConcurrentMarkingTask;
294 friend class G1CMRemarkTask;
295 friend class G1CMTask;
296
297 G1ConcurrentMarkThread* _cm_thread; // The thread doing the work
298 G1CollectedHeap* _g1h; // The heap
299
300 // Concurrent marking support structures
301 G1CMBitMap _mark_bitmap_1;
302 G1CMBitMap _mark_bitmap_2;
303 G1CMBitMap* _prev_mark_bitmap; // Completed mark bitmap
304 G1CMBitMap* _next_mark_bitmap; // Under-construction mark bitmap
305
306 // Heap bounds
307 MemRegion const _heap;
308
309 // Root region tracking and claiming
310 G1CMRootMemRegions _root_regions;
311
312 // For grey objects
313 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger
314 HeapWord* volatile _finger; // The global finger, region aligned,
315 // always pointing to the end of the
316 // last claimed region
317
318 uint _worker_id_offset;
319 uint _max_num_tasks; // Maximum number of marking tasks
320 uint _num_active_tasks; // Number of tasks currently active
321 G1CMTask** _tasks; // Task queue array (max_worker_id length)
322
323 G1CMTaskQueueSet* _task_queues; // Task queue set
324 TaskTerminator _terminator; // For termination
325
326 // Two sync barriers that are used to synchronize tasks when an
327 // overflow occurs. The algorithm is the following. All tasks enter
328 // the first one to ensure that they have all stopped manipulating
329 // the global data structures. After they exit it, they re-initialize
330 // their data structures and task 0 re-initializes the global data
331 // structures. Then, they enter the second sync barrier. This
332 // ensure, that no task starts doing work before all data
333 // structures (local and global) have been re-initialized. When they
334 // exit it, they are free to start working again.
335 WorkGangBarrierSync _first_overflow_barrier_sync;
336 WorkGangBarrierSync _second_overflow_barrier_sync;
337
338 // This is set by any task, when an overflow on the global data
339 // structures is detected
340 volatile bool _has_overflown;
341 // True: marking is concurrent, false: we're in remark
342 volatile bool _concurrent;
343 // Set at the end of a Full GC so that marking aborts
344 volatile bool _has_aborted;
345
346 // Used when remark aborts due to an overflow to indicate that
347 // another concurrent marking phase should start
348 volatile bool _restart_for_overflow;
349
350 ConcurrentGCTimer* _gc_timer_cm;
351
352 G1OldTracer* _gc_tracer_cm;
353
354 // Timing statistics. All of them are in ms
355 NumberSeq _init_times;
356 NumberSeq _remark_times;
357 NumberSeq _remark_mark_times;
358 NumberSeq _remark_weak_ref_times;
359 NumberSeq _cleanup_times;
360 double _total_cleanup_time;
361
362 double* _accum_task_vtime; // Accumulated task vtime
363
364 WorkGang* _concurrent_workers;
365 uint _num_concurrent_workers; // The number of marking worker threads we're using
366 uint _max_concurrent_workers; // Maximum number of marking worker threads
367
368 void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller);
369
370 void finalize_marking();
371
372 void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);
373 void weak_refs_work(bool clear_all_soft_refs);
374
375 void report_object_count(bool mark_completed);
376
377 void swap_mark_bitmaps();
378
379 void reclaim_empty_regions();
380
381 // After reclaiming empty regions, update heap sizes.
382 void compute_new_sizes();
383
384 // Clear statistics gathered during the concurrent cycle for the given region after
385 // it has been reclaimed.
386 void clear_statistics(HeapRegion* r);
387
388 // Resets the global marking data structures, as well as the
389 // task local ones; should be called during concurrent start.
390 void reset();
391
392 // Resets all the marking data structures. Called when we have to restart
393 // marking or when marking completes (via set_non_marking_state below).
394 void reset_marking_for_restart();
395
396 // We do this after we're done with marking so that the marking data
397 // structures are initialized to a sensible and predictable state.
398 void reset_at_marking_complete();
399
400 // Called to indicate how many threads are currently active.
401 void set_concurrency(uint active_tasks);
402
403 // Should be called to indicate which phase we're in (concurrent
404 // mark or remark) and how many threads are currently active.
405 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
406
407 // Prints all gathered CM-related statistics
408 void print_stats();
409
410 HeapWord* finger() { return _finger; }
411 bool concurrent() { return _concurrent; }
412 uint active_tasks() { return _num_active_tasks; }
413 TaskTerminator* terminator() { return &_terminator; }
414
415 // Claims the next available region to be scanned by a marking
416 // task/thread. It might return NULL if the next region is empty or
417 // we have run out of regions. In the latter case, out_of_regions()
418 // determines whether we've really run out of regions or the task
419 // should call claim_region() again. This might seem a bit
420 // awkward. Originally, the code was written so that claim_region()
421 // either successfully returned with a non-empty region or there
422 // were no more regions to be claimed. The problem with this was
423 // that, in certain circumstances, it iterated over large chunks of
424 // the heap finding only empty regions and, while it was working, it
425 // was preventing the calling task to call its regular clock
426 // method. So, this way, each task will spend very little time in
427 // claim_region() and is allowed to call the regular clock method
428 // frequently.
429 HeapRegion* claim_region(uint worker_id);
430
431 // Determines whether we've run out of regions to scan. Note that
432 // the finger can point past the heap end in case the heap was expanded
433 // to satisfy an allocation without doing a GC. This is fine, because all
434 // objects in those regions will be considered live anyway because of
435 // SATB guarantees (i.e. their TAMS will be equal to bottom).
436 bool out_of_regions() { return _finger >= _heap.end(); }
437
438 // Returns the task with the given id
439 G1CMTask* task(uint id) {
440 // During concurrent start we use the parallel gc threads to do some work, so
441 // we can only compare against _max_num_tasks.
442 assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks);
443 return _tasks[id];
444 }
445
446 // Access / manipulation of the overflow flag which is set to
447 // indicate that the global stack has overflown
448 bool has_overflown() { return _has_overflown; }
449 void set_has_overflown() { _has_overflown = true; }
450 void clear_has_overflown() { _has_overflown = false; }
451 bool restart_for_overflow() { return _restart_for_overflow; }
452
453 // Methods to enter the two overflow sync barriers
454 void enter_first_sync_barrier(uint worker_id);
455 void enter_second_sync_barrier(uint worker_id);
456
457 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
458 // true, periodically insert checks to see if this method should exit prematurely.
459 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
460
461 // Region statistics gathered during marking.
462 G1RegionMarkStats* _region_mark_stats;
463 // Top pointer for each region at the start of the rebuild remembered set process
464 // for regions which remembered sets need to be rebuilt. A NULL for a given region
465 // means that this region does not be scanned during the rebuilding remembered
466 // set phase at all.
467 HeapWord* volatile* _top_at_rebuild_starts;
468 public:
469 void add_to_liveness(uint worker_id, oop const obj, size_t size);
470 // Liveness of the given region as determined by concurrent marking, i.e. the amount of
471 // live words between bottom and nTAMS.
472 size_t liveness(uint region) const { return _region_mark_stats[region]._live_words; }
473
474 // Sets the internal top_at_region_start for the given region to current top of the region.
475 inline void update_top_at_rebuild_start(HeapRegion* r);
476 // TARS for the given region during remembered set rebuilding.
477 inline HeapWord* top_at_rebuild_start(uint region) const;
478
479 // Clear statistics gathered during the concurrent cycle for the given region after
480 // it has been reclaimed.
481 void clear_statistics_in_region(uint region_idx);
482 // Notification for eagerly reclaimed regions to clean up.
483 void humongous_object_eagerly_reclaimed(HeapRegion* r);
484 // Manipulation of the global mark stack.
485 // The push and pop operations are used by tasks for transfers
486 // between task-local queues and the global mark stack.
487 bool mark_stack_push(G1TaskQueueEntry* arr) {
488 if (!_global_mark_stack.par_push_chunk(arr)) {
489 set_has_overflown();
490 return false;
491 }
492 return true;
493 }
494 bool mark_stack_pop(G1TaskQueueEntry* arr) {
495 return _global_mark_stack.par_pop_chunk(arr);
496 }
497 size_t mark_stack_size() const { return _global_mark_stack.size(); }
498 size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; }
499 bool mark_stack_empty() const { return _global_mark_stack.is_empty(); }
500
501 G1CMRootMemRegions* root_regions() { return &_root_regions; }
502
503 void concurrent_cycle_start();
504 // Abandon current marking iteration due to a Full GC.
505 void concurrent_cycle_abort();
506 void concurrent_cycle_end();
507
508 void update_accum_task_vtime(int i, double vtime) {
509 _accum_task_vtime[i] += vtime;
510 }
511
512 double all_task_accum_vtime() {
513 double ret = 0.0;
514 for (uint i = 0; i < _max_num_tasks; ++i)
515 ret += _accum_task_vtime[i];
516 return ret;
517 }
518
519 // Attempts to steal an object from the task queues of other tasks
520 bool try_stealing(uint worker_id, G1TaskQueueEntry& task_entry);
521
522 G1ConcurrentMark(G1CollectedHeap* g1h,
523 G1RegionToSpaceMapper* prev_bitmap_storage,
524 G1RegionToSpaceMapper* next_bitmap_storage);
525 ~G1ConcurrentMark();
526
527 G1ConcurrentMarkThread* cm_thread() { return _cm_thread; }
528
529 const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; }
530 G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; }
531
532 // Calculates the number of concurrent GC threads to be used in the marking phase.
533 uint calc_active_marking_workers();
534
535 // Moves all per-task cached data into global state.
536 void flush_all_task_caches();
537 // Prepare internal data structures for the next mark cycle. This includes clearing
538 // the next mark bitmap and some internal data structures. This method is intended
539 // to be called concurrently to the mutator. It will yield to safepoint requests.
540 void cleanup_for_next_mark();
541
542 // Clear the previous marking bitmap during safepoint.
543 void clear_prev_bitmap(WorkGang* workers);
544
545 // These two methods do the work that needs to be done at the start and end of the
546 // concurrent start pause.
547 void pre_concurrent_start(GCCause::Cause cause);
548 void post_concurrent_start(bool is_mark_cycle);
549
550 // Scan all the root regions and mark everything reachable from
551 // them.
552 void scan_root_regions();
553
554 // Scan a single root MemRegion to mark everything reachable from it.
555 void scan_root_region(const MemRegion* region, uint worker_id);
556
557 // Do concurrent phase of marking, to a tentative transitive closure.
558 void mark_from_roots();
559
560 // Do concurrent preclean work.
561 void preclean();
562
563 void remark();
564
565 void cleanup();
566 // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use
567 // this carefully.
568 inline void mark_in_prev_bitmap(oop p);
569
570 // Clears marks for all objects in the given range, for the prev or
571 // next bitmaps. Caution: the previous bitmap is usually
572 // read-only, so use this carefully!
573 void clear_range_in_prev_bitmap(MemRegion mr);
574
575 inline bool is_marked_in_prev_bitmap(oop p) const;
576
577 // Verify that there are no collection set oops on the stacks (taskqueues /
578 // global mark stack) and fingers (global / per-task).
579 // If marking is not in progress, it's a no-op.
580 void verify_no_collection_set_oops() PRODUCT_RETURN;
581
582 inline bool do_yield_check();
583
584 bool has_aborted() { return _has_aborted; }
585
586 void print_summary_info();
587
588 void threads_do(ThreadClosure* tc) const;
589
590 void print_on_error(outputStream* st) const;
591
592 // Mark the given object on the next bitmap if it is below nTAMS.
593 inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj);
594 inline bool mark_in_next_bitmap(uint worker_id, oop const obj);
595
596 inline bool is_marked_in_next_bitmap(oop p) const;
597
598 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
599
600 private:
601 // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application.
602 void rebuild_rem_set_concurrently();
603 };
604
605 // A class representing a marking task.
606 class G1CMTask : public TerminatorTerminator {
607 private:
608 enum PrivateConstants {
609 // The regular clock call is called once the scanned words reaches
610 // this limit
611 words_scanned_period = 12*1024,
612 // The regular clock call is called once the number of visited
613 // references reaches this limit
614 refs_reached_period = 1024,
615 // Initial value for the hash seed, used in the work stealing code
616 init_hash_seed = 17
617 };
618
619 // Number of entries in the per-task stats entry. This seems enough to have a very
620 // low cache miss rate.
621 static const uint RegionMarkStatsCacheSize = 1024;
622
623 G1CMObjArrayProcessor _objArray_processor;
624
625 uint _worker_id;
626 G1CollectedHeap* _g1h;
627 G1ConcurrentMark* _cm;
628 G1CMBitMap* _next_mark_bitmap;
629 // the task queue of this task
630 G1CMTaskQueue* _task_queue;
631
632 G1RegionMarkStatsCache _mark_stats_cache;
633 // Number of calls to this task
634 uint _calls;
635
636 // When the virtual timer reaches this time, the marking step should exit
637 double _time_target_ms;
638 // Start time of the current marking step
639 double _start_time_ms;
640
641 // Oop closure used for iterations over oops
642 G1CMOopClosure* _cm_oop_closure;
643
644 // Region this task is scanning, NULL if we're not scanning any
645 HeapRegion* _curr_region;
646 // Local finger of this task, NULL if we're not scanning a region
647 HeapWord* _finger;
648 // Limit of the region this task is scanning, NULL if we're not scanning one
649 HeapWord* _region_limit;
650
651 // Number of words this task has scanned
652 size_t _words_scanned;
653 // When _words_scanned reaches this limit, the regular clock is
654 // called. Notice that this might be decreased under certain
655 // circumstances (i.e. when we believe that we did an expensive
656 // operation).
657 size_t _words_scanned_limit;
658 // Initial value of _words_scanned_limit (i.e. what it was
659 // before it was decreased).
660 size_t _real_words_scanned_limit;
661
662 // Number of references this task has visited
663 size_t _refs_reached;
664 // When _refs_reached reaches this limit, the regular clock is
665 // called. Notice this this might be decreased under certain
666 // circumstances (i.e. when we believe that we did an expensive
667 // operation).
668 size_t _refs_reached_limit;
669 // Initial value of _refs_reached_limit (i.e. what it was before
670 // it was decreased).
671 size_t _real_refs_reached_limit;
672
673 // If true, then the task has aborted for some reason
674 bool _has_aborted;
675 // Set when the task aborts because it has met its time quota
676 bool _has_timed_out;
677 // True when we're draining SATB buffers; this avoids the task
678 // aborting due to SATB buffers being available (as we're already
679 // dealing with them)
680 bool _draining_satb_buffers;
681
682 // Number sequence of past step times
683 NumberSeq _step_times_ms;
684 // Elapsed time of this task
685 double _elapsed_time_ms;
686 // Termination time of this task
687 double _termination_time_ms;
688 // When this task got into the termination protocol
689 double _termination_start_time_ms;
690
691 TruncatedSeq _marking_step_diff_ms;
692
693 // Updates the local fields after this task has claimed
694 // a new region to scan
695 void setup_for_region(HeapRegion* hr);
696 // Makes the limit of the region up-to-date
697 void update_region_limit();
698
699 // Called when either the words scanned or the refs visited limit
700 // has been reached
701 void reached_limit();
702 // Recalculates the words scanned and refs visited limits
703 void recalculate_limits();
704 // Decreases the words scanned and refs visited limits when we reach
705 // an expensive operation
706 void decrease_limits();
707 // Checks whether the words scanned or refs visited reached their
708 // respective limit and calls reached_limit() if they have
709 void check_limits() {
710 if (_words_scanned >= _words_scanned_limit ||
711 _refs_reached >= _refs_reached_limit) {
712 reached_limit();
713 }
714 }
715 // Supposed to be called regularly during a marking step as
716 // it checks a bunch of conditions that might cause the marking step
717 // to abort
718 // Return true if the marking step should continue. Otherwise, return false to abort
719 bool regular_clock_call();
720
721 // Set abort flag if regular_clock_call() check fails
722 inline void abort_marking_if_regular_check_fail();
723
724 // Test whether obj might have already been passed over by the
725 // mark bitmap scan, and so needs to be pushed onto the mark stack.
726 bool is_below_finger(oop obj, HeapWord* global_finger) const;
727
728 template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
729 public:
730 // Apply the closure on the given area of the objArray. Return the number of words
731 // scanned.
732 inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
733 // Resets the task; should be called right at the beginning of a marking phase.
734 void reset(G1CMBitMap* next_mark_bitmap);
735 // Clears all the fields that correspond to a claimed region.
736 void clear_region_fields();
737
738 // The main method of this class which performs a marking step
739 // trying not to exceed the given duration. However, it might exit
740 // prematurely, according to some conditions (i.e. SATB buffers are
741 // available for processing).
742 void do_marking_step(double target_ms,
743 bool do_termination,
744 bool is_serial);
745
746 // These two calls start and stop the timer
747 void record_start_time() {
748 _elapsed_time_ms = os::elapsedTime() * 1000.0;
749 }
750 void record_end_time() {
751 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
752 }
753
754 // Returns the worker ID associated with this task.
755 uint worker_id() { return _worker_id; }
756
757 // From TerminatorTerminator. It determines whether this task should
758 // exit the termination protocol after it's entered it.
759 virtual bool should_exit_termination();
760
761 // Resets the local region fields after a task has finished scanning a
762 // region; or when they have become stale as a result of the region
763 // being evacuated.
764 void giveup_current_region();
765
766 HeapWord* finger() { return _finger; }
767
768 bool has_aborted() { return _has_aborted; }
769 void set_has_aborted() { _has_aborted = true; }
770 void clear_has_aborted() { _has_aborted = false; }
771
772 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
773
774 // Increment the number of references this task has visited.
775 void increment_refs_reached() { ++_refs_reached; }
776
777 // Grey the object by marking it. If not already marked, push it on
778 // the local queue if below the finger. obj is required to be below its region's NTAMS.
779 // Returns whether there has been a mark to the bitmap.
780 inline bool make_reference_grey(oop obj);
781
782 // Grey the object (by calling make_grey_reference) if required,
783 // e.g. obj is below its containing region's NTAMS.
784 // Precondition: obj is a valid heap object.
785 // Returns true if the reference caused a mark to be set in the next bitmap.
786 template <class T>
787 inline bool deal_with_reference(T* p);
788
789 // Scans an object and visits its children.
790 inline void scan_task_entry(G1TaskQueueEntry task_entry);
791
792 // Pushes an object on the local queue.
793 inline void push(G1TaskQueueEntry task_entry);
794
795 // Move entries to the global stack.
796 void move_entries_to_global_stack();
797 // Move entries from the global stack, return true if we were successful to do so.
798 bool get_entries_from_global_stack();
799
800 // Pops and scans objects from the local queue. If partially is
801 // true, then it stops when the queue size is of a given limit. If
802 // partially is false, then it stops when the queue is empty.
803 void drain_local_queue(bool partially);
804 // Moves entries from the global stack to the local queue and
805 // drains the local queue. If partially is true, then it stops when
806 // both the global stack and the local queue reach a given size. If
807 // partially if false, it tries to empty them totally.
808 void drain_global_stack(bool partially);
809 // Keeps picking SATB buffers and processing them until no SATB
810 // buffers are available.
811 void drain_satb_buffers();
812
813 // Moves the local finger to a new location
814 inline void move_finger_to(HeapWord* new_finger) {
815 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
816 _finger = new_finger;
817 }
818
819 G1CMTask(uint worker_id,
820 G1ConcurrentMark *cm,
821 G1CMTaskQueue* task_queue,
822 G1RegionMarkStats* mark_stats,
823 uint max_regions);
824
825 inline void update_liveness(oop const obj, size_t const obj_size);
826
827 // Clear (without flushing) the mark cache entry for the given region.
828 void clear_mark_stats_cache(uint region_idx);
829 // Evict the whole statistics cache into the global statistics. Returns the
830 // number of cache hits and misses so far.
831 Pair<size_t, size_t> flush_mark_stats_cache();
832 // Prints statistics associated with this task
833 void print_stats();
834 };
835
836 // Class that's used to to print out per-region liveness
837 // information. It's currently used at the end of marking and also
838 // after we sort the old regions at the end of the cleanup operation.
839 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure {
840 // Accumulators for these values.
841 size_t _total_used_bytes;
842 size_t _total_capacity_bytes;
843 size_t _total_prev_live_bytes;
844 size_t _total_next_live_bytes;
845
846 // Accumulator for the remembered set size
847 size_t _total_remset_bytes;
848
849 // Accumulator for strong code roots memory size
850 size_t _total_strong_code_roots_bytes;
851
852 static double bytes_to_mb(size_t val) {
853 return (double) val / (double) M;
854 }
855
856 public:
857 // The header and footer are printed in the constructor and
858 // destructor respectively.
859 G1PrintRegionLivenessInfoClosure(const char* phase_name);
860 virtual bool do_heap_region(HeapRegion* r);
861 ~G1PrintRegionLivenessInfoClosure();
862 };
863
864 #endif // SHARE_GC_G1_G1CONCURRENTMARK_HPP