1 /*
2 * Copyright (c) 2001, 2019, 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/taskqueue.hpp"
34 #include "gc/shared/verifyOption.hpp"
35 #include "memory/allocation.hpp"
36 #include "utilities/compilerWarnings.hpp"
37
38 class ConcurrentGCTimer;
39 class G1ConcurrentMarkThread;
40 class G1CollectedHeap;
41 class G1CMOopClosure;
42 class G1CMTask;
43 class G1ConcurrentMark;
44 class G1OldTracer;
45 class G1RegionToSpaceMapper;
46 class G1SurvivorRegions;
47
48 PRAGMA_DIAG_PUSH
49 // warning C4522: multiple assignment operators specified
50 PRAGMA_DISABLE_MSVC_WARNING(4522)
51
52 // This is a container class for either an oop or a continuation address for
53 // mark stack entries. Both are pushed onto the mark stack.
54 class G1TaskQueueEntry {
55 private:
56 void* _holder;
57
58 static const uintptr_t ArraySliceBit = 1;
59
60 G1TaskQueueEntry(oop obj) : _holder(obj) {
61 assert(_holder != NULL, "Not allowed to set NULL task queue element");
62 }
63 G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
64 public:
65 G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }
66 G1TaskQueueEntry() : _holder(NULL) { }
67
68 static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
69 static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
70
71 G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {
72 _holder = t._holder;
73 return *this;
74 }
75
76 volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
77 _holder = t._holder;
78 return *this;
79 }
80
81 oop obj() const {
82 assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
83 return (oop)_holder;
84 }
85
86 HeapWord* slice() const {
87 assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
88 return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
89 }
90
91 bool is_oop() const { return !is_array_slice(); }
92 bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
93 bool is_null() const { return _holder == NULL; }
94 };
95
96 PRAGMA_DIAG_POP
97
98 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
99 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
100
101 // Closure used by CM during concurrent reference discovery
102 // and reference processing (during remarking) to determine
103 // if a particular object is alive. It is primarily used
104 // to determine if referents of discovered reference objects
105 // are alive. An instance is also embedded into the
106 // reference processor as the _is_alive_non_header field
107 class G1CMIsAliveClosure : public BoolObjectClosure {
108 G1CollectedHeap* _g1h;
109 public:
110 G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
111 bool do_object_b(oop obj);
112 };
113
114 class G1CMSubjectToDiscoveryClosure : public BoolObjectClosure {
115 G1CollectedHeap* _g1h;
116 public:
117 G1CMSubjectToDiscoveryClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
118 bool do_object_b(oop obj);
119 };
120
121 // Represents the overflow mark stack used by concurrent marking.
122 //
123 // Stores oops in a huge buffer in virtual memory that is always fully committed.
124 // Resizing may only happen during a STW pause when the stack is empty.
125 //
126 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
127 // stack memory is split into evenly sized chunks of oops. Users can only
128 // add or remove entries on that basis.
129 // Chunks are filled in increasing address order. Not completely filled chunks
130 // have a NULL element as a terminating element.
131 //
132 // Every chunk has a header containing a single pointer element used for memory
133 // management. This wastes some space, but is negligible (< .1% with current sizing).
134 //
135 // Memory management is done using a mix of tracking a high water-mark indicating
136 // that all chunks at a lower address are valid chunks, and a singly linked free
137 // list connecting all empty chunks.
138 class G1CMMarkStack {
139 public:
140 // Number of TaskQueueEntries that can fit in a single chunk.
141 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
142 private:
143 struct TaskQueueEntryChunk {
144 TaskQueueEntryChunk* next;
145 G1TaskQueueEntry data[EntriesPerChunk];
146 };
147
148 size_t _max_chunk_capacity; // Maximum number of TaskQueueEntryChunk elements on the stack.
149
150 TaskQueueEntryChunk* _base; // Bottom address of allocated memory area.
151 size_t _chunk_capacity; // Current maximum number of TaskQueueEntryChunk elements.
152
153 char _pad0[DEFAULT_CACHE_LINE_SIZE];
154 TaskQueueEntryChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users.
155 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
156 TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
157 volatile size_t _chunks_in_chunk_list;
158 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
159
160 volatile size_t _hwm; // High water mark within the reserved space.
161 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
162
163 // Allocate a new chunk from the reserved memory, using the high water mark. Returns
164 // NULL if out of memory.
165 TaskQueueEntryChunk* allocate_new_chunk();
166
167 // Atomically add the given chunk to the list.
168 void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
169 // Atomically remove and return a chunk from the given list. Returns NULL if the
170 // list is empty.
171 TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
172
173 void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
174 void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
175
176 TaskQueueEntryChunk* remove_chunk_from_chunk_list();
177 TaskQueueEntryChunk* remove_chunk_from_free_list();
178
179 // Resizes the mark stack to the given new capacity. Releases any previous
180 // memory if successful.
181 bool resize(size_t new_capacity);
182
183 public:
184 G1CMMarkStack();
185 ~G1CMMarkStack();
186
187 // Alignment and minimum capacity of this mark stack in number of oops.
188 static size_t capacity_alignment();
189
190 // Allocate and initialize the mark stack with the given number of oops.
191 bool initialize(size_t initial_capacity, size_t max_capacity);
192
193 // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
194 // stack. If less than EntriesPerChunk elements are to be pushed, the array must
195 // be terminated with a NULL.
196 // Returns whether the buffer contents were successfully pushed to the global mark
197 // stack.
198 bool par_push_chunk(G1TaskQueueEntry* buffer);
199
200 // Pops a chunk from this mark stack, copying them into the given buffer. This
201 // chunk may contain up to EntriesPerChunk elements. If there are less, the last
202 // element in the array is a NULL pointer.
203 bool par_pop_chunk(G1TaskQueueEntry* buffer);
204
205 // Return whether the chunk list is empty. Racy due to unsynchronized access to
206 // _chunk_list.
207 bool is_empty() const { return _chunk_list == NULL; }
208
209 size_t capacity() const { return _chunk_capacity; }
210
211 // Expand the stack, typically in response to an overflow condition
212 void expand();
213
214 // Return the approximate number of oops on this mark stack. Racy due to
215 // unsynchronized access to _chunks_in_chunk_list.
216 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
217
218 void set_empty();
219
220 // Apply Fn to every oop on the mark stack. The mark stack must not
221 // be modified while iterating.
222 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
223 };
224
225 // Root Regions are regions that contain objects from nTAMS to top. These are roots
226 // for marking, i.e. their referenced objects must be kept alive to maintain the
227 // SATB invariant.
228 // We could scan and mark them through during the initial-mark pause, but for
229 // pause time reasons we move this work to the concurrent phase.
230 // We need to complete this procedure before the next GC because it might determine
231 // that some of these "root objects" are dead, potentially dropping some required
232 // references.
233 // Root regions comprise of the complete contents of survivor regions, and any
234 // objects copied into old gen during GC.
235 class G1CMRootRegions {
236 HeapRegion** _root_regions;
237 size_t const _max_regions;
238
239 volatile size_t _num_root_regions; // Actual number of root regions.
240
241 volatile size_t _claimed_root_regions; // Number of root regions currently claimed.
242
243 volatile bool _scan_in_progress;
244 volatile bool _should_abort;
245
246 void notify_scan_done();
247
248 public:
249 G1CMRootRegions(uint const max_regions);
250 ~G1CMRootRegions();
251
252 // Reset the data structure to allow addition of new root regions.
253 void reset();
254
255 void add(HeapRegion* hr);
256
257 // Reset the claiming / scanning of the root regions.
258 void prepare_for_scan();
259
260 // Forces get_next() to return NULL so that the iteration aborts early.
261 void abort() { _should_abort = true; }
262
263 // Return true if the CM thread are actively scanning root regions,
264 // false otherwise.
265 bool scan_in_progress() { return _scan_in_progress; }
266
267 // Claim the next root region to scan atomically, or return NULL if
268 // all have been claimed.
269 HeapRegion* claim_next();
270
271 // The number of root regions to scan.
272 uint num_root_regions() const;
273
274 void cancel_scan();
275
276 // Flag that we're done with root region scanning and notify anyone
277 // who's waiting on it. If aborted is false, assume that all regions
278 // have been claimed.
279 void scan_finished();
280
281 // If CM threads are still scanning root regions, wait until they
282 // are done. Return true if we had to wait, false otherwise.
283 bool wait_until_scan_finished();
284 };
285
286 // This class manages data structures and methods for doing liveness analysis in
287 // G1's concurrent cycle.
288 class G1ConcurrentMark : public CHeapObj<mtGC> {
289 friend class G1ConcurrentMarkThread;
290 friend class G1CMRefProcTaskProxy;
291 friend class G1CMRefProcTaskExecutor;
292 friend class G1CMKeepAliveAndDrainClosure;
293 friend class G1CMDrainMarkingStackClosure;
294 friend class G1CMBitMapClosure;
295 friend class G1CMConcurrentMarkingTask;
296 friend class G1CMRemarkTask;
297 friend class G1CMTask;
298
299 G1ConcurrentMarkThread* _cm_thread; // The thread doing the work
300 G1CollectedHeap* _g1h; // The heap
301 bool _completed_initialization; // Set to true when initialization is complete
302
303 // Concurrent marking support structures
304 G1CMBitMap _mark_bitmap_1;
305 G1CMBitMap _mark_bitmap_2;
306 G1CMBitMap* _prev_mark_bitmap; // Completed mark bitmap
307 G1CMBitMap* _next_mark_bitmap; // Under-construction mark bitmap
308
309 // Heap bounds
310 MemRegion const _heap;
311
312 // Root region tracking and claiming
313 G1CMRootRegions _root_regions;
314
315 // For grey objects
316 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger
317 HeapWord* volatile _finger; // The global finger, region aligned,
318 // always pointing to the end of the
319 // last claimed region
320
321 uint _worker_id_offset;
322 uint _max_num_tasks; // Maximum number of marking tasks
323 uint _num_active_tasks; // Number of tasks currently active
324 G1CMTask** _tasks; // Task queue array (max_worker_id length)
325
326 G1CMTaskQueueSet* _task_queues; // Task queue set
327 TaskTerminator _terminator; // For termination
328
329 // Two sync barriers that are used to synchronize tasks when an
330 // overflow occurs. The algorithm is the following. All tasks enter
331 // the first one to ensure that they have all stopped manipulating
332 // the global data structures. After they exit it, they re-initialize
333 // their data structures and task 0 re-initializes the global data
334 // structures. Then, they enter the second sync barrier. This
335 // ensure, that no task starts doing work before all data
336 // structures (local and global) have been re-initialized. When they
337 // exit it, they are free to start working again.
338 WorkGangBarrierSync _first_overflow_barrier_sync;
339 WorkGangBarrierSync _second_overflow_barrier_sync;
340
341 // This is set by any task, when an overflow on the global data
342 // structures is detected
343 volatile bool _has_overflown;
344 // True: marking is concurrent, false: we're in remark
345 volatile bool _concurrent;
346 // Set at the end of a Full GC so that marking aborts
347 volatile bool _has_aborted;
348
349 // Used when remark aborts due to an overflow to indicate that
350 // another concurrent marking phase should start
351 volatile bool _restart_for_overflow;
352
353 ConcurrentGCTimer* _gc_timer_cm;
354
355 G1OldTracer* _gc_tracer_cm;
356
357 // Timing statistics. All of them are in ms
358 NumberSeq _init_times;
359 NumberSeq _remark_times;
360 NumberSeq _remark_mark_times;
361 NumberSeq _remark_weak_ref_times;
362 NumberSeq _cleanup_times;
363 double _total_cleanup_time;
364
365 double* _accum_task_vtime; // Accumulated task vtime
366
367 WorkGang* _concurrent_workers;
368 uint _num_concurrent_workers; // The number of marking worker threads we're using
369 uint _max_concurrent_workers; // Maximum number of marking worker threads
370
371 void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller);
372
373 void finalize_marking();
374
375 void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);
376 void weak_refs_work(bool clear_all_soft_refs);
377
378 void report_object_count(bool mark_completed);
379
380 void swap_mark_bitmaps();
381
382 void reclaim_empty_regions();
383
384 // After reclaiming empty regions, update heap sizes.
385 void compute_new_sizes();
386
387 // Clear statistics gathered during the concurrent cycle for the given region after
388 // it has been reclaimed.
389 void clear_statistics(HeapRegion* r);
390
391 // Resets the global marking data structures, as well as the
392 // task local ones; should be called during initial mark.
393 void reset();
394
395 // Resets all the marking data structures. Called when we have to restart
396 // marking or when marking completes (via set_non_marking_state below).
397 void reset_marking_for_restart();
398
399 // We do this after we're done with marking so that the marking data
400 // structures are initialized to a sensible and predictable state.
401 void reset_at_marking_complete();
402
403 // Called to indicate how many threads are currently active.
404 void set_concurrency(uint active_tasks);
405
406 // Should be called to indicate which phase we're in (concurrent
407 // mark or remark) and how many threads are currently active.
408 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
409
410 // Prints all gathered CM-related statistics
411 void print_stats();
412
413 HeapWord* finger() { return _finger; }
414 bool concurrent() { return _concurrent; }
415 uint active_tasks() { return _num_active_tasks; }
416 ParallelTaskTerminator* terminator() const { return _terminator.terminator(); }
417
418 // Claims the next available region to be scanned by a marking
419 // task/thread. It might return NULL if the next region is empty or
420 // we have run out of regions. In the latter case, out_of_regions()
421 // determines whether we've really run out of regions or the task
422 // should call claim_region() again. This might seem a bit
423 // awkward. Originally, the code was written so that claim_region()
424 // either successfully returned with a non-empty region or there
425 // were no more regions to be claimed. The problem with this was
426 // that, in certain circumstances, it iterated over large chunks of
427 // the heap finding only empty regions and, while it was working, it
428 // was preventing the calling task to call its regular clock
429 // method. So, this way, each task will spend very little time in
430 // claim_region() and is allowed to call the regular clock method
431 // frequently.
432 HeapRegion* claim_region(uint worker_id);
433
434 // Determines whether we've run out of regions to scan. Note that
435 // the finger can point past the heap end in case the heap was expanded
436 // to satisfy an allocation without doing a GC. This is fine, because all
437 // objects in those regions will be considered live anyway because of
438 // SATB guarantees (i.e. their TAMS will be equal to bottom).
439 bool out_of_regions() { return _finger >= _heap.end(); }
440
441 // Returns the task with the given id
442 G1CMTask* task(uint id) {
443 // During initial mark we use the parallel gc threads to do some work, so
444 // we can only compare against _max_num_tasks.
445 assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks);
446 return _tasks[id];
447 }
448
449 // Access / manipulation of the overflow flag which is set to
450 // indicate that the global stack has overflown
451 bool has_overflown() { return _has_overflown; }
452 void set_has_overflown() { _has_overflown = true; }
453 void clear_has_overflown() { _has_overflown = false; }
454 bool restart_for_overflow() { return _restart_for_overflow; }
455
456 // Methods to enter the two overflow sync barriers
457 void enter_first_sync_barrier(uint worker_id);
458 void enter_second_sync_barrier(uint worker_id);
459
460 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
461 // true, periodically insert checks to see if this method should exit prematurely.
462 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
463
464 // Region statistics gathered during marking.
465 G1RegionMarkStats* _region_mark_stats;
466 // Top pointer for each region at the start of the rebuild remembered set process
467 // for regions which remembered sets need to be rebuilt. A NULL for a given region
468 // means that this region does not be scanned during the rebuilding remembered
469 // set phase at all.
470 HeapWord* volatile* _top_at_rebuild_starts;
471 public:
472 void add_to_liveness(uint worker_id, oop const obj, size_t size);
473 // Liveness of the given region as determined by concurrent marking, i.e. the amount of
474 // live words between bottom and nTAMS.
475 size_t liveness(uint region) const { return _region_mark_stats[region]._live_words; }
476
477 // Sets the internal top_at_region_start for the given region to current top of the region.
478 inline void update_top_at_rebuild_start(HeapRegion* r);
479 // TARS for the given region during remembered set rebuilding.
480 inline HeapWord* top_at_rebuild_start(uint region) const;
481
482 // Clear statistics gathered during the concurrent cycle for the given region after
483 // it has been reclaimed.
484 void clear_statistics_in_region(uint region_idx);
485 // Notification for eagerly reclaimed regions to clean up.
486 void humongous_object_eagerly_reclaimed(HeapRegion* r);
487 // Manipulation of the global mark stack.
488 // The push and pop operations are used by tasks for transfers
489 // between task-local queues and the global mark stack.
490 bool mark_stack_push(G1TaskQueueEntry* arr) {
491 if (!_global_mark_stack.par_push_chunk(arr)) {
492 set_has_overflown();
493 return false;
494 }
495 return true;
496 }
497 bool mark_stack_pop(G1TaskQueueEntry* arr) {
498 return _global_mark_stack.par_pop_chunk(arr);
499 }
500 size_t mark_stack_size() const { return _global_mark_stack.size(); }
501 size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; }
502 bool mark_stack_empty() const { return _global_mark_stack.is_empty(); }
503
504 G1CMRootRegions* root_regions() { return &_root_regions; }
505
506 void concurrent_cycle_start();
507 // Abandon current marking iteration due to a Full GC.
508 void concurrent_cycle_abort();
509 void concurrent_cycle_end();
510
511 void update_accum_task_vtime(int i, double vtime) {
512 _accum_task_vtime[i] += vtime;
513 }
514
515 double all_task_accum_vtime() {
516 double ret = 0.0;
517 for (uint i = 0; i < _max_num_tasks; ++i)
518 ret += _accum_task_vtime[i];
519 return ret;
520 }
521
522 // Attempts to steal an object from the task queues of other tasks
523 bool try_stealing(uint worker_id, G1TaskQueueEntry& task_entry);
524
525 G1ConcurrentMark(G1CollectedHeap* g1h,
526 G1RegionToSpaceMapper* prev_bitmap_storage,
527 G1RegionToSpaceMapper* next_bitmap_storage);
528 ~G1ConcurrentMark();
529
530 G1ConcurrentMarkThread* cm_thread() { return _cm_thread; }
531
532 const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; }
533 G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; }
534
535 // Calculates the number of concurrent GC threads to be used in the marking phase.
536 uint calc_active_marking_workers();
537
538 // Moves all per-task cached data into global state.
539 void flush_all_task_caches();
540 // Prepare internal data structures for the next mark cycle. This includes clearing
541 // the next mark bitmap and some internal data structures. This method is intended
542 // to be called concurrently to the mutator. It will yield to safepoint requests.
543 void cleanup_for_next_mark();
544
545 // Clear the previous marking bitmap during safepoint.
546 void clear_prev_bitmap(WorkGang* workers);
547
548 // These two methods do the work that needs to be done at the start and end of the
549 // initial mark pause.
550 void pre_initial_mark();
551 void post_initial_mark();
552
553 // Scan all the root regions and mark everything reachable from
554 // them.
555 void scan_root_regions();
556
557 // Scan a single root region from nTAMS to top and mark everything reachable from it.
558 void scan_root_region(HeapRegion* hr, uint worker_id);
559
560 // Do concurrent phase of marking, to a tentative transitive closure.
561 void mark_from_roots();
562
563 // Do concurrent preclean work.
564 void preclean();
565
566 void remark();
567
568 void cleanup();
569 // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use
570 // this carefully.
571 inline void mark_in_prev_bitmap(oop p);
572
573 // Clears marks for all objects in the given range, for the prev or
574 // next bitmaps. Caution: the previous bitmap is usually
575 // read-only, so use this carefully!
576 void clear_range_in_prev_bitmap(MemRegion mr);
577
578 inline bool is_marked_in_prev_bitmap(oop p) const;
579
580 // Verify that there are no collection set oops on the stacks (taskqueues /
581 // global mark stack) and fingers (global / per-task).
582 // If marking is not in progress, it's a no-op.
583 void verify_no_collection_set_oops() PRODUCT_RETURN;
584
585 inline bool do_yield_check();
586
587 bool has_aborted() { return _has_aborted; }
588
589 void print_summary_info();
590
591 void print_worker_threads_on(outputStream* st) const;
592 void threads_do(ThreadClosure* tc) const;
593
594 void print_on_error(outputStream* st) const;
595
596 // Mark the given object on the next bitmap if it is below nTAMS.
597 inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj);
598 inline bool mark_in_next_bitmap(uint worker_id, oop const obj);
599
600 inline bool is_marked_in_next_bitmap(oop p) const;
601
602 // Returns true if initialization was successfully completed.
603 bool completed_initialization() const {
604 return _completed_initialization;
605 }
606
607 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
608 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
609
610 private:
611 // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application.
612 void rebuild_rem_set_concurrently();
613 };
614
615 // A class representing a marking task.
616 class G1CMTask : public TerminatorTerminator {
617 private:
618 enum PrivateConstants {
619 // The regular clock call is called once the scanned words reaches
620 // this limit
621 words_scanned_period = 12*1024,
622 // The regular clock call is called once the number of visited
623 // references reaches this limit
624 refs_reached_period = 1024,
625 // Initial value for the hash seed, used in the work stealing code
626 init_hash_seed = 17
627 };
628
629 // Number of entries in the per-task stats entry. This seems enough to have a very
630 // low cache miss rate.
631 static const uint RegionMarkStatsCacheSize = 1024;
632
633 G1CMObjArrayProcessor _objArray_processor;
634
635 uint _worker_id;
636 G1CollectedHeap* _g1h;
637 G1ConcurrentMark* _cm;
638 G1CMBitMap* _next_mark_bitmap;
639 // the task queue of this task
640 G1CMTaskQueue* _task_queue;
641
642 G1RegionMarkStatsCache _mark_stats_cache;
643 // Number of calls to this task
644 uint _calls;
645
646 // When the virtual timer reaches this time, the marking step should exit
647 double _time_target_ms;
648 // Start time of the current marking step
649 double _start_time_ms;
650
651 // Oop closure used for iterations over oops
652 G1CMOopClosure* _cm_oop_closure;
653
654 // Region this task is scanning, NULL if we're not scanning any
655 HeapRegion* _curr_region;
656 // Local finger of this task, NULL if we're not scanning a region
657 HeapWord* _finger;
658 // Limit of the region this task is scanning, NULL if we're not scanning one
659 HeapWord* _region_limit;
660
661 // Number of words this task has scanned
662 size_t _words_scanned;
663 // When _words_scanned reaches this limit, the regular clock is
664 // called. Notice that this might be decreased under certain
665 // circumstances (i.e. when we believe that we did an expensive
666 // operation).
667 size_t _words_scanned_limit;
668 // Initial value of _words_scanned_limit (i.e. what it was
669 // before it was decreased).
670 size_t _real_words_scanned_limit;
671
672 // Number of references this task has visited
673 size_t _refs_reached;
674 // When _refs_reached reaches this limit, the regular clock is
675 // called. Notice this this might be decreased under certain
676 // circumstances (i.e. when we believe that we did an expensive
677 // operation).
678 size_t _refs_reached_limit;
679 // Initial value of _refs_reached_limit (i.e. what it was before
680 // it was decreased).
681 size_t _real_refs_reached_limit;
682
683 // If true, then the task has aborted for some reason
684 bool _has_aborted;
685 // Set when the task aborts because it has met its time quota
686 bool _has_timed_out;
687 // True when we're draining SATB buffers; this avoids the task
688 // aborting due to SATB buffers being available (as we're already
689 // dealing with them)
690 bool _draining_satb_buffers;
691
692 // Number sequence of past step times
693 NumberSeq _step_times_ms;
694 // Elapsed time of this task
695 double _elapsed_time_ms;
696 // Termination time of this task
697 double _termination_time_ms;
698 // When this task got into the termination protocol
699 double _termination_start_time_ms;
700
701 TruncatedSeq _marking_step_diffs_ms;
702
703 // Updates the local fields after this task has claimed
704 // a new region to scan
705 void setup_for_region(HeapRegion* hr);
706 // Makes the limit of the region up-to-date
707 void update_region_limit();
708
709 // Called when either the words scanned or the refs visited limit
710 // has been reached
711 void reached_limit();
712 // Recalculates the words scanned and refs visited limits
713 void recalculate_limits();
714 // Decreases the words scanned and refs visited limits when we reach
715 // an expensive operation
716 void decrease_limits();
717 // Checks whether the words scanned or refs visited reached their
718 // respective limit and calls reached_limit() if they have
719 void check_limits() {
720 if (_words_scanned >= _words_scanned_limit ||
721 _refs_reached >= _refs_reached_limit) {
722 reached_limit();
723 }
724 }
725 // Supposed to be called regularly during a marking step as
726 // it checks a bunch of conditions that might cause the marking step
727 // to abort
728 // Return true if the marking step should continue. Otherwise, return false to abort
729 bool regular_clock_call();
730
731 // Set abort flag if regular_clock_call() check fails
732 inline void abort_marking_if_regular_check_fail();
733
734 // Test whether obj might have already been passed over by the
735 // mark bitmap scan, and so needs to be pushed onto the mark stack.
736 bool is_below_finger(oop obj, HeapWord* global_finger) const;
737
738 template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
739 public:
740 // Apply the closure on the given area of the objArray. Return the number of words
741 // scanned.
742 inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
743 // Resets the task; should be called right at the beginning of a marking phase.
744 void reset(G1CMBitMap* next_mark_bitmap);
745 // Clears all the fields that correspond to a claimed region.
746 void clear_region_fields();
747
748 // The main method of this class which performs a marking step
749 // trying not to exceed the given duration. However, it might exit
750 // prematurely, according to some conditions (i.e. SATB buffers are
751 // available for processing).
752 void do_marking_step(double target_ms,
753 bool do_termination,
754 bool is_serial);
755
756 // These two calls start and stop the timer
757 void record_start_time() {
758 _elapsed_time_ms = os::elapsedTime() * 1000.0;
759 }
760 void record_end_time() {
761 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
762 }
763
764 // Returns the worker ID associated with this task.
765 uint worker_id() { return _worker_id; }
766
767 // From TerminatorTerminator. It determines whether this task should
768 // exit the termination protocol after it's entered it.
769 virtual bool should_exit_termination();
770
771 // Resets the local region fields after a task has finished scanning a
772 // region; or when they have become stale as a result of the region
773 // being evacuated.
774 void giveup_current_region();
775
776 HeapWord* finger() { return _finger; }
777
778 bool has_aborted() { return _has_aborted; }
779 void set_has_aborted() { _has_aborted = true; }
780 void clear_has_aborted() { _has_aborted = false; }
781
782 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
783
784 // Increment the number of references this task has visited.
785 void increment_refs_reached() { ++_refs_reached; }
786
787 // Grey the object by marking it. If not already marked, push it on
788 // the local queue if below the finger. obj is required to be below its region's NTAMS.
789 // Returns whether there has been a mark to the bitmap.
790 inline bool make_reference_grey(oop obj);
791
792 // Grey the object (by calling make_grey_reference) if required,
793 // e.g. obj is below its containing region's NTAMS.
794 // Precondition: obj is a valid heap object.
795 // Returns true if the reference caused a mark to be set in the next bitmap.
796 template <class T>
797 inline bool deal_with_reference(T* p);
798
799 // Scans an object and visits its children.
800 inline void scan_task_entry(G1TaskQueueEntry task_entry);
801
802 // Pushes an object on the local queue.
803 inline void push(G1TaskQueueEntry task_entry);
804
805 // Move entries to the global stack.
806 void move_entries_to_global_stack();
807 // Move entries from the global stack, return true if we were successful to do so.
808 bool get_entries_from_global_stack();
809
810 // Pops and scans objects from the local queue. If partially is
811 // true, then it stops when the queue size is of a given limit. If
812 // partially is false, then it stops when the queue is empty.
813 void drain_local_queue(bool partially);
814 // Moves entries from the global stack to the local queue and
815 // drains the local queue. If partially is true, then it stops when
816 // both the global stack and the local queue reach a given size. If
817 // partially if false, it tries to empty them totally.
818 void drain_global_stack(bool partially);
819 // Keeps picking SATB buffers and processing them until no SATB
820 // buffers are available.
821 void drain_satb_buffers();
822
823 // Moves the local finger to a new location
824 inline void move_finger_to(HeapWord* new_finger) {
825 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
826 _finger = new_finger;
827 }
828
829 G1CMTask(uint worker_id,
830 G1ConcurrentMark *cm,
831 G1CMTaskQueue* task_queue,
832 G1RegionMarkStats* mark_stats,
833 uint max_regions);
834
835 inline void update_liveness(oop const obj, size_t const obj_size);
836
837 // Clear (without flushing) the mark cache entry for the given region.
838 void clear_mark_stats_cache(uint region_idx);
839 // Evict the whole statistics cache into the global statistics. Returns the
840 // number of cache hits and misses so far.
841 Pair<size_t, size_t> flush_mark_stats_cache();
842 // Prints statistics associated with this task
843 void print_stats();
844 };
845
846 // Class that's used to to print out per-region liveness
847 // information. It's currently used at the end of marking and also
848 // after we sort the old regions at the end of the cleanup operation.
849 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure {
850 // Accumulators for these values.
851 size_t _total_used_bytes;
852 size_t _total_capacity_bytes;
853 size_t _total_prev_live_bytes;
854 size_t _total_next_live_bytes;
855
856 // Accumulator for the remembered set size
857 size_t _total_remset_bytes;
858
859 // Accumulator for strong code roots memory size
860 size_t _total_strong_code_roots_bytes;
861
862 static double bytes_to_mb(size_t val) {
863 return (double) val / (double) M;
864 }
865
866 public:
867 // The header and footer are printed in the constructor and
868 // destructor respectively.
869 G1PrintRegionLivenessInfoClosure(const char* phase_name);
870 virtual bool do_heap_region(HeapRegion* r);
871 ~G1PrintRegionLivenessInfoClosure();
872 };
873
874 #endif // SHARE_GC_G1_G1CONCURRENTMARK_HPP