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 "gc/shared/workgroup.hpp"
36 #include "memory/allocation.hpp"
37 #include "utilities/compilerWarnings.hpp"
38
39 class ConcurrentGCTimer;
40 class G1ConcurrentMarkThread;
41 class G1CollectedHeap;
42 class G1CMOopClosure;
43 class G1CMTask;
44 class G1ConcurrentMark;
45 class G1OldTracer;
46 class G1RegionToSpaceMapper;
47 class G1SurvivorRegions;
48
49 PRAGMA_DIAG_PUSH
50 // warning C4522: multiple assignment operators specified
51 PRAGMA_DISABLE_MSVC_WARNING(4522)
52
53 // This is a container class for either an oop or a continuation address for
54 // mark stack entries. Both are pushed onto the mark stack.
55 class G1TaskQueueEntry {
56 private:
57 void* _holder;
58
59 static const uintptr_t ArraySliceBit = 1;
60
61 G1TaskQueueEntry(oop obj) : _holder(obj) {
62 assert(_holder != NULL, "Not allowed to set NULL task queue element");
63 }
64 G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
65 public:
66 G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }
67 G1TaskQueueEntry() : _holder(NULL) { }
68
69 static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
70 static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
71
72 G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {
73 _holder = t._holder;
74 return *this;
75 }
76
77 volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
78 _holder = t._holder;
79 return *this;
80 }
81
82 oop obj() const {
83 assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
84 return (oop)_holder;
85 }
86
87 HeapWord* slice() const {
88 assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
89 return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
90 }
91
92 bool is_oop() const { return !is_array_slice(); }
93 bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
94 bool is_null() const { return _holder == NULL; }
95 };
96
97 PRAGMA_DIAG_POP
98
99 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
100 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
101
102 // Closure used by CM during concurrent reference discovery
103 // and reference processing (during remarking) to determine
104 // if a particular object is alive. It is primarily used
105 // to determine if referents of discovered reference objects
106 // are alive. An instance is also embedded into the
107 // reference processor as the _is_alive_non_header field
108 class G1CMIsAliveClosure : public BoolObjectClosure {
109 G1CollectedHeap* _g1h;
110 public:
111 G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
112 bool do_object_b(oop obj);
113 };
114
115 class G1CMSubjectToDiscoveryClosure : public BoolObjectClosure {
116 G1CollectedHeap* _g1h;
117 public:
118 G1CMSubjectToDiscoveryClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
119 bool do_object_b(oop obj);
120 };
121
122 // Represents the overflow mark stack used by concurrent marking.
123 //
124 // Stores oops in a huge buffer in virtual memory that is always fully committed.
125 // Resizing may only happen during a STW pause when the stack is empty.
126 //
127 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
128 // stack memory is split into evenly sized chunks of oops. Users can only
129 // add or remove entries on that basis.
130 // Chunks are filled in increasing address order. Not completely filled chunks
131 // have a NULL element as a terminating element.
132 //
133 // Every chunk has a header containing a single pointer element used for memory
134 // management. This wastes some space, but is negligible (< .1% with current sizing).
135 //
136 // Memory management is done using a mix of tracking a high water-mark indicating
137 // that all chunks at a lower address are valid chunks, and a singly linked free
138 // list connecting all empty chunks.
139 class G1CMMarkStack {
140 public:
141 // Number of TaskQueueEntries that can fit in a single chunk.
142 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
143 private:
144 struct TaskQueueEntryChunk {
145 TaskQueueEntryChunk* next;
146 G1TaskQueueEntry data[EntriesPerChunk];
147 };
148
149 size_t _max_chunk_capacity; // Maximum number of TaskQueueEntryChunk elements on the stack.
150
151 TaskQueueEntryChunk* _base; // Bottom address of allocated memory area.
152 size_t _chunk_capacity; // Current maximum number of TaskQueueEntryChunk elements.
153
154 char _pad0[DEFAULT_CACHE_LINE_SIZE];
155 TaskQueueEntryChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users.
156 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
157 TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
158 volatile size_t _chunks_in_chunk_list;
159 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
160
161 volatile size_t _hwm; // High water mark within the reserved space.
162 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
163
164 // Allocate a new chunk from the reserved memory, using the high water mark. Returns
165 // NULL if out of memory.
166 TaskQueueEntryChunk* allocate_new_chunk();
167
168 // Atomically add the given chunk to the list.
169 void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
170 // Atomically remove and return a chunk from the given list. Returns NULL if the
171 // list is empty.
172 TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
173
174 void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
175 void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
176
177 TaskQueueEntryChunk* remove_chunk_from_chunk_list();
178 TaskQueueEntryChunk* remove_chunk_from_free_list();
179
180 // Resizes the mark stack to the given new capacity. Releases any previous
181 // memory if successful.
182 bool resize(size_t new_capacity);
183
184 public:
185 G1CMMarkStack();
186 ~G1CMMarkStack();
187
188 // Alignment and minimum capacity of this mark stack in number of oops.
189 static size_t capacity_alignment();
190
191 // Allocate and initialize the mark stack with the given number of oops.
192 bool initialize(size_t initial_capacity, size_t max_capacity);
193
194 // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
195 // stack. If less than EntriesPerChunk elements are to be pushed, the array must
196 // be terminated with a NULL.
197 // Returns whether the buffer contents were successfully pushed to the global mark
198 // stack.
199 bool par_push_chunk(G1TaskQueueEntry* buffer);
200
201 // Pops a chunk from this mark stack, copying them into the given buffer. This
202 // chunk may contain up to EntriesPerChunk elements. If there are less, the last
203 // element in the array is a NULL pointer.
204 bool par_pop_chunk(G1TaskQueueEntry* buffer);
205
206 // Return whether the chunk list is empty. Racy due to unsynchronized access to
207 // _chunk_list.
208 bool is_empty() const { return _chunk_list == NULL; }
209
210 size_t capacity() const { return _chunk_capacity; }
211
212 // Expand the stack, typically in response to an overflow condition
213 void expand();
214
215 // Return the approximate number of oops on this mark stack. Racy due to
216 // unsynchronized access to _chunks_in_chunk_list.
217 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
218
219 void set_empty();
220
221 // Apply Fn to every oop on the mark stack. The mark stack must not
222 // be modified while iterating.
223 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
224 };
225
226 // Root MemRegions are memory areas that contain objects which references are
227 // roots wrt to the marking. They must be scanned before marking to maintain the
228 // SATB invariant.
229 // Typically they contain the areas from nTAMS to top of the regions.
230 // We could scan and mark through these objects during the initial-mark pause, but for
231 // pause time reasons we move this work to the concurrent phase.
232 // We need to complete this procedure before the next GC because it might determine
233 // that some of these "root objects" are dead, potentially dropping some required
234 // references.
235 // Root MemRegions comprise of the contents of survivor regions at the end
236 // of the GC, and any objects copied into the old gen during GC.
237 class G1CMRootMemRegions {
238 // The set of root MemRegions.
239 MemRegion* _root_regions;
240 size_t const _max_regions;
241
242 volatile size_t _num_root_regions; // Actual number of root regions.
243
244 volatile size_t _claimed_root_regions; // Number of root regions currently claimed.
245
246 volatile bool _scan_in_progress;
247 volatile bool _should_abort;
248
249 void notify_scan_done();
250
251 public:
252 G1CMRootMemRegions(uint const max_regions);
253 ~G1CMRootMemRegions();
254
255 // Reset the data structure to allow addition of new root regions.
256 void reset();
257
258 void add(HeapWord* start, HeapWord* end);
259
260 // Reset the claiming / scanning of the root regions.
261 void prepare_for_scan();
262
263 // Forces get_next() to return NULL so that the iteration aborts early.
264 void abort() { _should_abort = true; }
265
266 // Return true if the CM thread are actively scanning root regions,
267 // false otherwise.
268 bool scan_in_progress() { return _scan_in_progress; }
269
270 // Claim the next root MemRegion to scan atomically, or return NULL if
271 // all have been claimed.
272 const MemRegion* claim_next();
273
274 // The number of root regions to scan.
275 uint num_root_regions() const;
276
277 void cancel_scan();
278
279 // Flag that we're done with root region scanning and notify anyone
280 // who's waiting on it. If aborted is false, assume that all regions
281 // have been claimed.
282 void scan_finished();
283
284 // If CM threads are still scanning root regions, wait until they
285 // are done. Return true if we had to wait, false otherwise.
286 bool wait_until_scan_finished();
287 };
288
289 // This class manages data structures and methods for doing liveness analysis in
290 // G1's concurrent cycle.
291 class G1ConcurrentMark : public CHeapObj<mtGC> {
292 friend class G1ConcurrentMarkThread;
293 friend class G1CMRefProcTaskProxy;
294 friend class G1CMRefProcTaskExecutor;
295 friend class G1CMKeepAliveAndDrainClosure;
296 friend class G1CMDrainMarkingStackClosure;
297 friend class G1CMBitMapClosure;
298 friend class G1CMConcurrentMarkingTask;
299 friend class G1CMRemarkTask;
300 friend class G1CMTask;
301
302 G1ConcurrentMarkThread* _cm_thread; // The thread doing the work
303 G1CollectedHeap* _g1h; // The heap
304 bool _completed_initialization; // Set to true when initialization is complete
305
306 // Concurrent marking support structures
307 G1CMBitMap _mark_bitmap_1;
308 G1CMBitMap _mark_bitmap_2;
309 G1CMBitMap* _prev_mark_bitmap; // Completed mark bitmap
310 G1CMBitMap* _next_mark_bitmap; // Under-construction mark bitmap
311
312 // Heap bounds
313 MemRegion const _heap;
314
315 // Root region tracking and claiming
316 G1CMRootMemRegions _root_regions;
317
318 // For grey objects
319 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger
320 HeapWord* volatile _finger; // The global finger, region aligned,
321 // always pointing to the end of the
322 // last claimed region
323
324 uint _worker_id_offset;
325 uint _max_num_tasks; // Maximum number of marking tasks
326 uint _num_active_tasks; // Number of tasks currently active
327 G1CMTask** _tasks; // Task queue array (max_worker_id length)
328
329 G1CMTaskQueueSet* _task_queues; // Task queue set
330 TaskTerminator _terminator; // For termination
331
332 // Two sync barriers that are used to synchronize tasks when an
333 // overflow occurs. The algorithm is the following. All tasks enter
334 // the first one to ensure that they have all stopped manipulating
335 // the global data structures. After they exit it, they re-initialize
336 // their data structures and task 0 re-initializes the global data
337 // structures. Then, they enter the second sync barrier. This
338 // ensure, that no task starts doing work before all data
339 // structures (local and global) have been re-initialized. When they
340 // exit it, they are free to start working again.
341 WorkGangBarrierSync _first_overflow_barrier_sync;
342 WorkGangBarrierSync _second_overflow_barrier_sync;
343
344 // This is set by any task, when an overflow on the global data
345 // structures is detected
346 volatile bool _has_overflown;
347 // True: marking is concurrent, false: we're in remark
348 volatile bool _concurrent;
349 // Set at the end of a Full GC so that marking aborts
350 volatile bool _has_aborted;
351
352 // Used when remark aborts due to an overflow to indicate that
353 // another concurrent marking phase should start
354 volatile bool _restart_for_overflow;
355
356 ConcurrentGCTimer* _gc_timer_cm;
357
358 G1OldTracer* _gc_tracer_cm;
359
360 // Timing statistics. All of them are in ms
361 NumberSeq _init_times;
362 NumberSeq _remark_times;
363 NumberSeq _remark_mark_times;
364 NumberSeq _remark_weak_ref_times;
365 NumberSeq _cleanup_times;
366 double _total_cleanup_time;
367
368 double* _accum_task_vtime; // Accumulated task vtime
369
370 WorkGang* _concurrent_workers;
371 uint _num_concurrent_workers; // The number of marking worker threads we're using
372 uint _max_concurrent_workers; // Maximum number of marking worker threads
373
374 void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller);
375
376 void finalize_marking();
377
378 void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);
379 void weak_refs_work(bool clear_all_soft_refs);
380
381 void report_object_count(bool mark_completed);
382
383 void swap_mark_bitmaps();
384
385 void reclaim_empty_regions();
386
387 // After reclaiming empty regions, update heap sizes.
388 void compute_new_sizes();
389
390 // Clear statistics gathered during the concurrent cycle for the given region after
391 // it has been reclaimed.
392 void clear_statistics(HeapRegion* r);
393
394 // Resets the global marking data structures, as well as the
395 // task local ones; should be called during initial mark.
396 void reset();
397
398 // Resets all the marking data structures. Called when we have to restart
399 // marking or when marking completes (via set_non_marking_state below).
400 void reset_marking_for_restart();
401
402 // We do this after we're done with marking so that the marking data
403 // structures are initialized to a sensible and predictable state.
404 void reset_at_marking_complete();
405
406 // Called to indicate how many threads are currently active.
407 void set_concurrency(uint active_tasks);
408
409 // Should be called to indicate which phase we're in (concurrent
410 // mark or remark) and how many threads are currently active.
411 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
412
413 // Prints all gathered CM-related statistics
414 void print_stats();
415
416 HeapWord* finger() { return _finger; }
417 bool concurrent() { return _concurrent; }
418 uint active_tasks() { return _num_active_tasks; }
419 ParallelTaskTerminator* terminator() const { return _terminator.terminator(); }
420
421 // Claims the next available region to be scanned by a marking
422 // task/thread. It might return NULL if the next region is empty or
423 // we have run out of regions. In the latter case, out_of_regions()
424 // determines whether we've really run out of regions or the task
425 // should call claim_region() again. This might seem a bit
426 // awkward. Originally, the code was written so that claim_region()
427 // either successfully returned with a non-empty region or there
428 // were no more regions to be claimed. The problem with this was
429 // that, in certain circumstances, it iterated over large chunks of
430 // the heap finding only empty regions and, while it was working, it
431 // was preventing the calling task to call its regular clock
432 // method. So, this way, each task will spend very little time in
433 // claim_region() and is allowed to call the regular clock method
434 // frequently.
435 HeapRegion* claim_region(uint worker_id);
436
437 // Determines whether we've run out of regions to scan. Note that
438 // the finger can point past the heap end in case the heap was expanded
439 // to satisfy an allocation without doing a GC. This is fine, because all
440 // objects in those regions will be considered live anyway because of
441 // SATB guarantees (i.e. their TAMS will be equal to bottom).
442 bool out_of_regions() { return _finger >= _heap.end(); }
443
444 // Returns the task with the given id
445 G1CMTask* task(uint id) {
446 // During initial mark we use the parallel gc threads to do some work, so
447 // we can only compare against _max_num_tasks.
448 assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks);
449 return _tasks[id];
450 }
451
452 // Access / manipulation of the overflow flag which is set to
453 // indicate that the global stack has overflown
454 bool has_overflown() { return _has_overflown; }
455 void set_has_overflown() { _has_overflown = true; }
456 void clear_has_overflown() { _has_overflown = false; }
457 bool restart_for_overflow() { return _restart_for_overflow; }
458
459 // Methods to enter the two overflow sync barriers
460 void enter_first_sync_barrier(uint worker_id);
461 void enter_second_sync_barrier(uint worker_id);
462
463 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
464 // true, periodically insert checks to see if this method should exit prematurely.
465 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
466
467 // Region statistics gathered during marking.
468 G1RegionMarkStats* _region_mark_stats;
469 // Top pointer for each region at the start of the rebuild remembered set process
470 // for regions which remembered sets need to be rebuilt. A NULL for a given region
471 // means that this region does not be scanned during the rebuilding remembered
472 // set phase at all.
473 HeapWord* volatile* _top_at_rebuild_starts;
474 public:
475 void add_to_liveness(uint worker_id, oop const obj, size_t size);
476 // Liveness of the given region as determined by concurrent marking, i.e. the amount of
477 // live words between bottom and nTAMS.
478 size_t liveness(uint region) const { return _region_mark_stats[region]._live_words; }
479
480 // Sets the internal top_at_region_start for the given region to current top of the region.
481 inline void update_top_at_rebuild_start(HeapRegion* r);
482 // TARS for the given region during remembered set rebuilding.
483 inline HeapWord* top_at_rebuild_start(uint region) const;
484
485 // Clear statistics gathered during the concurrent cycle for the given region after
486 // it has been reclaimed.
487 void clear_statistics_in_region(uint region_idx);
488 // Notification for eagerly reclaimed regions to clean up.
489 void humongous_object_eagerly_reclaimed(HeapRegion* r);
490 // Manipulation of the global mark stack.
491 // The push and pop operations are used by tasks for transfers
492 // between task-local queues and the global mark stack.
493 bool mark_stack_push(G1TaskQueueEntry* arr) {
494 if (!_global_mark_stack.par_push_chunk(arr)) {
495 set_has_overflown();
496 return false;
497 }
498 return true;
499 }
500 bool mark_stack_pop(G1TaskQueueEntry* arr) {
501 return _global_mark_stack.par_pop_chunk(arr);
502 }
503 size_t mark_stack_size() const { return _global_mark_stack.size(); }
504 size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; }
505 bool mark_stack_empty() const { return _global_mark_stack.is_empty(); }
506
507 G1CMRootMemRegions* root_regions() { return &_root_regions; }
508
509 void concurrent_cycle_start();
510 // Abandon current marking iteration due to a Full GC.
511 void concurrent_cycle_abort();
512 void concurrent_cycle_end();
513
514 void update_accum_task_vtime(int i, double vtime) {
515 _accum_task_vtime[i] += vtime;
516 }
517
518 double all_task_accum_vtime() {
519 double ret = 0.0;
520 for (uint i = 0; i < _max_num_tasks; ++i)
521 ret += _accum_task_vtime[i];
522 return ret;
523 }
524
525 // Attempts to steal an object from the task queues of other tasks
526 bool try_stealing(uint worker_id, G1TaskQueueEntry& task_entry);
527
528 G1ConcurrentMark(G1CollectedHeap* g1h,
529 G1RegionToSpaceMapper* prev_bitmap_storage,
530 G1RegionToSpaceMapper* next_bitmap_storage);
531 ~G1ConcurrentMark();
532
533 G1ConcurrentMarkThread* cm_thread() { return _cm_thread; }
534
535 const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; }
536 G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; }
537
538 // Calculates the number of concurrent GC threads to be used in the marking phase.
539 uint calc_active_marking_workers();
540
541 // Moves all per-task cached data into global state.
542 void flush_all_task_caches();
543 // Prepare internal data structures for the next mark cycle. This includes clearing
544 // the next mark bitmap and some internal data structures. This method is intended
545 // to be called concurrently to the mutator. It will yield to safepoint requests.
546 void cleanup_for_next_mark();
547
548 // Clear the previous marking bitmap during safepoint.
549 void clear_prev_bitmap(WorkGang* workers);
550
551 // These two methods do the work that needs to be done at the start and end of the
552 // initial mark pause.
553 void pre_initial_mark();
554 void post_initial_mark();
555
556 // Scan all the root regions and mark everything reachable from
557 // them.
558 void scan_root_regions();
559
560 // Scan a single root MemRegion to mark everything reachable from it.
561 void scan_root_region(const MemRegion* region, uint worker_id);
562
563 // Do concurrent phase of marking, to a tentative transitive closure.
564 void mark_from_roots();
565
566 // Do concurrent preclean work.
567 void preclean();
568
569 void remark();
570
571 void cleanup();
572 // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use
573 // this carefully.
574 inline void mark_in_prev_bitmap(oop p);
575
576 // Clears marks for all objects in the given range, for the prev or
577 // next bitmaps. Caution: the previous bitmap is usually
578 // read-only, so use this carefully!
579 void clear_range_in_prev_bitmap(MemRegion mr);
580
581 inline bool is_marked_in_prev_bitmap(oop p) const;
582
583 // Verify that there are no collection set oops on the stacks (taskqueues /
584 // global mark stack) and fingers (global / per-task).
585 // If marking is not in progress, it's a no-op.
586 void verify_no_collection_set_oops() PRODUCT_RETURN;
587
588 inline bool do_yield_check();
589
590 bool has_aborted() { return _has_aborted; }
591
592 void print_summary_info();
593
594 void print_worker_threads_on(outputStream* st) const;
595 void threads_do(ThreadClosure* tc) const;
596
597 void print_on_error(outputStream* st) const;
598
599 // Mark the given object on the next bitmap if it is below nTAMS.
600 inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj);
601 inline bool mark_in_next_bitmap(uint worker_id, oop const obj);
602
603 inline bool is_marked_in_next_bitmap(oop p) const;
604
605 // Returns true if initialization was successfully completed.
606 bool completed_initialization() const {
607 return _completed_initialization;
608 }
609
610 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
611 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
612
613 private:
614 // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application.
615 void rebuild_rem_set_concurrently();
616 };
617
618 // A class representing a marking task.
619 class G1CMTask : public TerminatorTerminator {
620 private:
621 enum PrivateConstants {
622 // The regular clock call is called once the scanned words reaches
623 // this limit
624 words_scanned_period = 12*1024,
625 // The regular clock call is called once the number of visited
626 // references reaches this limit
627 refs_reached_period = 1024,
628 // Initial value for the hash seed, used in the work stealing code
629 init_hash_seed = 17
630 };
631
632 // Number of entries in the per-task stats entry. This seems enough to have a very
633 // low cache miss rate.
634 static const uint RegionMarkStatsCacheSize = 1024;
635
636 G1CMObjArrayProcessor _objArray_processor;
637
638 uint _worker_id;
639 G1CollectedHeap* _g1h;
640 G1ConcurrentMark* _cm;
641 G1CMBitMap* _next_mark_bitmap;
642 // the task queue of this task
643 G1CMTaskQueue* _task_queue;
644
645 G1RegionMarkStatsCache _mark_stats_cache;
646 // Number of calls to this task
647 uint _calls;
648
649 // When the virtual timer reaches this time, the marking step should exit
650 double _time_target_ms;
651 // Start time of the current marking step
652 double _start_time_ms;
653
654 // Oop closure used for iterations over oops
655 G1CMOopClosure* _cm_oop_closure;
656
657 // Region this task is scanning, NULL if we're not scanning any
658 HeapRegion* _curr_region;
659 // Local finger of this task, NULL if we're not scanning a region
660 HeapWord* _finger;
661 // Limit of the region this task is scanning, NULL if we're not scanning one
662 HeapWord* _region_limit;
663
664 // Number of words this task has scanned
665 size_t _words_scanned;
666 // When _words_scanned reaches this limit, the regular clock is
667 // called. Notice that this might be decreased under certain
668 // circumstances (i.e. when we believe that we did an expensive
669 // operation).
670 size_t _words_scanned_limit;
671 // Initial value of _words_scanned_limit (i.e. what it was
672 // before it was decreased).
673 size_t _real_words_scanned_limit;
674
675 // Number of references this task has visited
676 size_t _refs_reached;
677 // When _refs_reached reaches this limit, the regular clock is
678 // called. Notice this this might be decreased under certain
679 // circumstances (i.e. when we believe that we did an expensive
680 // operation).
681 size_t _refs_reached_limit;
682 // Initial value of _refs_reached_limit (i.e. what it was before
683 // it was decreased).
684 size_t _real_refs_reached_limit;
685
686 // If true, then the task has aborted for some reason
687 bool _has_aborted;
688 // Set when the task aborts because it has met its time quota
689 bool _has_timed_out;
690 // True when we're draining SATB buffers; this avoids the task
691 // aborting due to SATB buffers being available (as we're already
692 // dealing with them)
693 bool _draining_satb_buffers;
694
695 // Number sequence of past step times
696 NumberSeq _step_times_ms;
697 // Elapsed time of this task
698 double _elapsed_time_ms;
699 // Termination time of this task
700 double _termination_time_ms;
701 // When this task got into the termination protocol
702 double _termination_start_time_ms;
703
704 TruncatedSeq _marking_step_diff_ms;
705
706 // Updates the local fields after this task has claimed
707 // a new region to scan
708 void setup_for_region(HeapRegion* hr);
709 // Makes the limit of the region up-to-date
710 void update_region_limit();
711
712 // Called when either the words scanned or the refs visited limit
713 // has been reached
714 void reached_limit();
715 // Recalculates the words scanned and refs visited limits
716 void recalculate_limits();
717 // Decreases the words scanned and refs visited limits when we reach
718 // an expensive operation
719 void decrease_limits();
720 // Checks whether the words scanned or refs visited reached their
721 // respective limit and calls reached_limit() if they have
722 void check_limits() {
723 if (_words_scanned >= _words_scanned_limit ||
724 _refs_reached >= _refs_reached_limit) {
725 reached_limit();
726 }
727 }
728 // Supposed to be called regularly during a marking step as
729 // it checks a bunch of conditions that might cause the marking step
730 // to abort
731 // Return true if the marking step should continue. Otherwise, return false to abort
732 bool regular_clock_call();
733
734 // Set abort flag if regular_clock_call() check fails
735 inline void abort_marking_if_regular_check_fail();
736
737 // Test whether obj might have already been passed over by the
738 // mark bitmap scan, and so needs to be pushed onto the mark stack.
739 bool is_below_finger(oop obj, HeapWord* global_finger) const;
740
741 template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
742 public:
743 // Apply the closure on the given area of the objArray. Return the number of words
744 // scanned.
745 inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
746 // Resets the task; should be called right at the beginning of a marking phase.
747 void reset(G1CMBitMap* next_mark_bitmap);
748 // Clears all the fields that correspond to a claimed region.
749 void clear_region_fields();
750
751 // The main method of this class which performs a marking step
752 // trying not to exceed the given duration. However, it might exit
753 // prematurely, according to some conditions (i.e. SATB buffers are
754 // available for processing).
755 void do_marking_step(double target_ms,
756 bool do_termination,
757 bool is_serial);
758
759 // These two calls start and stop the timer
760 void record_start_time() {
761 _elapsed_time_ms = os::elapsedTime() * 1000.0;
762 }
763 void record_end_time() {
764 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
765 }
766
767 // Returns the worker ID associated with this task.
768 uint worker_id() { return _worker_id; }
769
770 // From TerminatorTerminator. It determines whether this task should
771 // exit the termination protocol after it's entered it.
772 virtual bool should_exit_termination();
773
774 // Resets the local region fields after a task has finished scanning a
775 // region; or when they have become stale as a result of the region
776 // being evacuated.
777 void giveup_current_region();
778
779 HeapWord* finger() { return _finger; }
780
781 bool has_aborted() { return _has_aborted; }
782 void set_has_aborted() { _has_aborted = true; }
783 void clear_has_aborted() { _has_aborted = false; }
784
785 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
786
787 // Increment the number of references this task has visited.
788 void increment_refs_reached() { ++_refs_reached; }
789
790 // Grey the object by marking it. If not already marked, push it on
791 // the local queue if below the finger. obj is required to be below its region's NTAMS.
792 // Returns whether there has been a mark to the bitmap.
793 inline bool make_reference_grey(oop obj);
794
795 // Grey the object (by calling make_grey_reference) if required,
796 // e.g. obj is below its containing region's NTAMS.
797 // Precondition: obj is a valid heap object.
798 // Returns true if the reference caused a mark to be set in the next bitmap.
799 template <class T>
800 inline bool deal_with_reference(T* p);
801
802 // Scans an object and visits its children.
803 inline void scan_task_entry(G1TaskQueueEntry task_entry);
804
805 // Pushes an object on the local queue.
806 inline void push(G1TaskQueueEntry task_entry);
807
808 // Move entries to the global stack.
809 void move_entries_to_global_stack();
810 // Move entries from the global stack, return true if we were successful to do so.
811 bool get_entries_from_global_stack();
812
813 // Pops and scans objects from the local queue. If partially is
814 // true, then it stops when the queue size is of a given limit. If
815 // partially is false, then it stops when the queue is empty.
816 void drain_local_queue(bool partially);
817 // Moves entries from the global stack to the local queue and
818 // drains the local queue. If partially is true, then it stops when
819 // both the global stack and the local queue reach a given size. If
820 // partially if false, it tries to empty them totally.
821 void drain_global_stack(bool partially);
822 // Keeps picking SATB buffers and processing them until no SATB
823 // buffers are available.
824 void drain_satb_buffers();
825
826 // Moves the local finger to a new location
827 inline void move_finger_to(HeapWord* new_finger) {
828 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
829 _finger = new_finger;
830 }
831
832 G1CMTask(uint worker_id,
833 G1ConcurrentMark *cm,
834 G1CMTaskQueue* task_queue,
835 G1RegionMarkStats* mark_stats,
836 uint max_regions);
837
838 inline void update_liveness(oop const obj, size_t const obj_size);
839
840 // Clear (without flushing) the mark cache entry for the given region.
841 void clear_mark_stats_cache(uint region_idx);
842 // Evict the whole statistics cache into the global statistics. Returns the
843 // number of cache hits and misses so far.
844 Pair<size_t, size_t> flush_mark_stats_cache();
845 // Prints statistics associated with this task
846 void print_stats();
847 };
848
849 // Class that's used to to print out per-region liveness
850 // information. It's currently used at the end of marking and also
851 // after we sort the old regions at the end of the cleanup operation.
852 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure {
853 // Accumulators for these values.
854 size_t _total_used_bytes;
855 size_t _total_capacity_bytes;
856 size_t _total_prev_live_bytes;
857 size_t _total_next_live_bytes;
858
859 // Accumulator for the remembered set size
860 size_t _total_remset_bytes;
861
862 // Accumulator for strong code roots memory size
863 size_t _total_strong_code_roots_bytes;
864
865 static double bytes_to_mb(size_t val) {
866 return (double) val / (double) M;
867 }
868
869 public:
870 // The header and footer are printed in the constructor and
871 // destructor respectively.
872 G1PrintRegionLivenessInfoClosure(const char* phase_name);
873 virtual bool do_heap_region(HeapRegion* r);
874 ~G1PrintRegionLivenessInfoClosure();
875 };
876
877 #endif // SHARE_GC_G1_G1CONCURRENTMARK_HPP