1 /*
2 * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/metadataOnStackMark.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc/g1/concurrentMarkThread.inline.hpp"
30 #include "gc/g1/g1CollectedHeap.inline.hpp"
31 #include "gc/g1/g1CollectorState.hpp"
32 #include "gc/g1/g1ConcurrentMark.inline.hpp"
33 #include "gc/g1/g1HeapVerifier.hpp"
34 #include "gc/g1/g1OopClosures.inline.hpp"
35 #include "gc/g1/g1CardLiveData.inline.hpp"
36 #include "gc/g1/g1Policy.hpp"
37 #include "gc/g1/g1StringDedup.hpp"
38 #include "gc/g1/heapRegion.inline.hpp"
39 #include "gc/g1/heapRegionRemSet.hpp"
40 #include "gc/g1/heapRegionSet.inline.hpp"
41 #include "gc/g1/suspendibleThreadSet.hpp"
42 #include "gc/shared/gcId.hpp"
43 #include "gc/shared/gcTimer.hpp"
44 #include "gc/shared/gcTrace.hpp"
45 #include "gc/shared/gcTraceTime.inline.hpp"
46 #include "gc/shared/genOopClosures.inline.hpp"
47 #include "gc/shared/referencePolicy.hpp"
48 #include "gc/shared/strongRootsScope.hpp"
49 #include "gc/shared/taskqueue.inline.hpp"
50 #include "gc/shared/vmGCOperations.hpp"
51 #include "logging/log.hpp"
52 #include "memory/allocation.hpp"
53 #include "memory/resourceArea.hpp"
54 #include "oops/oop.inline.hpp"
55 #include "runtime/atomic.hpp"
56 #include "runtime/handles.inline.hpp"
57 #include "runtime/java.hpp"
58 #include "runtime/prefetch.inline.hpp"
59 #include "services/memTracker.hpp"
60 #include "utilities/growableArray.hpp"
61
62 // Concurrent marking bit map wrapper
63
64 G1CMBitMapRO::G1CMBitMapRO(int shifter) :
65 _bm(),
66 _shifter(shifter) {
67 _bmStartWord = 0;
68 _bmWordSize = 0;
69 }
70
71 HeapWord* G1CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
72 const HeapWord* limit) const {
73 // First we must round addr *up* to a possible object boundary.
74 addr = (HeapWord*)align_size_up((intptr_t)addr,
75 HeapWordSize << _shifter);
76 size_t addrOffset = heapWordToOffset(addr);
77 assert(limit != NULL, "limit must not be NULL");
78 size_t limitOffset = heapWordToOffset(limit);
79 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
80 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
81 assert(nextAddr >= addr, "get_next_one postcondition");
82 assert(nextAddr == limit || isMarked(nextAddr),
83 "get_next_one postcondition");
84 return nextAddr;
85 }
86
87 #ifndef PRODUCT
88 bool G1CMBitMapRO::covers(MemRegion heap_rs) const {
89 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
90 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
91 "size inconsistency");
92 return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
93 _bmWordSize == heap_rs.word_size();
94 }
95 #endif
96
97 void G1CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
98 _bm.print_on_error(st, prefix);
99 }
100
101 size_t G1CMBitMap::compute_size(size_t heap_size) {
102 return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
103 }
104
105 size_t G1CMBitMap::mark_distance() {
106 return MinObjAlignmentInBytes * BitsPerByte;
107 }
108
109 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
110 _bmStartWord = heap.start();
111 _bmWordSize = heap.word_size();
112
113 _bm = BitMapView((BitMap::bm_word_t*) storage->reserved().start(), _bmWordSize >> _shifter);
114
115 storage->set_mapping_changed_listener(&_listener);
116 }
117
118 void G1CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
119 if (zero_filled) {
120 return;
121 }
122 // We need to clear the bitmap on commit, removing any existing information.
123 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
124 _bm->clear_range(mr);
125 }
126
127 void G1CMBitMap::clear_range(MemRegion mr) {
128 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
129 assert(!mr.is_empty(), "unexpected empty region");
130 // convert address range into offset range
131 _bm.at_put_range(heapWordToOffset(mr.start()),
132 heapWordToOffset(mr.end()), false);
133 }
134
135 G1CMMarkStack::G1CMMarkStack() :
136 _max_chunk_capacity(0),
137 _base(NULL),
138 _chunk_capacity(0),
139 _should_expand(false) {
140 set_empty();
141 }
142
143 bool G1CMMarkStack::resize(size_t new_capacity) {
144 assert(is_empty(), "Only resize when stack is empty.");
145 assert(new_capacity <= _max_chunk_capacity,
146 "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
147
148 TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk, mtGC>::allocate_or_null(new_capacity);
149
150 if (new_base == NULL) {
151 log_warning(gc)("Failed to reserve memory for new overflow mark stack with " SIZE_FORMAT " chunks and size " SIZE_FORMAT "B.", new_capacity, new_capacity * sizeof(TaskQueueEntryChunk));
152 return false;
153 }
154 // Release old mapping.
155 if (_base != NULL) {
156 MmapArrayAllocator<TaskQueueEntryChunk, mtGC>::free(_base, _chunk_capacity);
157 }
158
159 _base = new_base;
160 _chunk_capacity = new_capacity;
161 set_empty();
162 _should_expand = false;
163
164 return true;
165 }
166
167 size_t G1CMMarkStack::capacity_alignment() {
168 return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
169 }
170
171 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
172 guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
173
174 size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
175
176 _max_chunk_capacity = (size_t)align_size_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
177 size_t initial_chunk_capacity = (size_t)align_size_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
178
179 guarantee(initial_chunk_capacity <= _max_chunk_capacity,
180 "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
181 _max_chunk_capacity,
182 initial_chunk_capacity);
183
184 log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
185 initial_chunk_capacity, _max_chunk_capacity);
186
187 return resize(initial_chunk_capacity);
188 }
189
190 void G1CMMarkStack::expand() {
191 // Clear expansion flag
192 _should_expand = false;
193
194 if (_chunk_capacity == _max_chunk_capacity) {
195 log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
196 return;
197 }
198 size_t old_capacity = _chunk_capacity;
199 // Double capacity if possible
200 size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
201
202 if (resize(new_capacity)) {
203 log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
204 old_capacity, new_capacity);
205 } else {
206 log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
207 old_capacity, new_capacity);
208 }
209 }
210
211 G1CMMarkStack::~G1CMMarkStack() {
212 if (_base != NULL) {
213 MmapArrayAllocator<TaskQueueEntryChunk, mtGC>::free(_base, _chunk_capacity);
214 }
215 }
216
217 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
218 elem->next = *list;
219 *list = elem;
220 }
221
222 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
223 MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
224 add_chunk_to_list(&_chunk_list, elem);
225 _chunks_in_chunk_list++;
226 }
227
228 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
229 MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
230 add_chunk_to_list(&_free_list, elem);
231 }
232
233 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
234 TaskQueueEntryChunk* result = *list;
235 if (result != NULL) {
236 *list = (*list)->next;
237 }
238 return result;
239 }
240
241 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
242 MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
243 TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
244 if (result != NULL) {
245 _chunks_in_chunk_list--;
246 }
247 return result;
248 }
249
250 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
251 MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
252 return remove_chunk_from_list(&_free_list);
253 }
254
255 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
256 // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
257 // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
258 // wraparound of _hwm.
259 if (_hwm >= _chunk_capacity) {
260 return NULL;
261 }
262
263 size_t cur_idx = Atomic::add(1, &_hwm) - 1;
264 if (cur_idx >= _chunk_capacity) {
265 return NULL;
266 }
267
268 TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
269 result->next = NULL;
270 return result;
271 }
272
273 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
274 // Get a new chunk.
275 TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
276
277 if (new_chunk == NULL) {
278 // Did not get a chunk from the free list. Allocate from backing memory.
279 new_chunk = allocate_new_chunk();
280
281 if (new_chunk == NULL) {
282 return false;
283 }
284 }
285
286 Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
287
288 add_chunk_to_chunk_list(new_chunk);
289
290 return true;
291 }
292
293 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
294 TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
295
296 if (cur == NULL) {
297 return false;
298 }
299
300 Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
301
302 add_chunk_to_free_list(cur);
303 return true;
304 }
305
306 void G1CMMarkStack::set_empty() {
307 _chunks_in_chunk_list = 0;
308 _hwm = 0;
309 _chunk_list = NULL;
310 _free_list = NULL;
311 }
312
313 G1CMRootRegions::G1CMRootRegions() :
314 _cm(NULL), _scan_in_progress(false),
315 _should_abort(false), _claimed_survivor_index(0) { }
316
317 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
318 _survivors = survivors;
319 _cm = cm;
320 }
321
322 void G1CMRootRegions::prepare_for_scan() {
323 assert(!scan_in_progress(), "pre-condition");
324
325 // Currently, only survivors can be root regions.
326 _claimed_survivor_index = 0;
327 _scan_in_progress = _survivors->regions()->is_nonempty();
328 _should_abort = false;
329 }
330
331 HeapRegion* G1CMRootRegions::claim_next() {
332 if (_should_abort) {
333 // If someone has set the should_abort flag, we return NULL to
334 // force the caller to bail out of their loop.
335 return NULL;
336 }
337
338 // Currently, only survivors can be root regions.
339 const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
340
341 int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
342 if (claimed_index < survivor_regions->length()) {
343 return survivor_regions->at(claimed_index);
344 }
345 return NULL;
346 }
347
348 uint G1CMRootRegions::num_root_regions() const {
349 return (uint)_survivors->regions()->length();
350 }
351
352 void G1CMRootRegions::notify_scan_done() {
353 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
354 _scan_in_progress = false;
355 RootRegionScan_lock->notify_all();
356 }
357
358 void G1CMRootRegions::cancel_scan() {
359 notify_scan_done();
360 }
361
362 void G1CMRootRegions::scan_finished() {
363 assert(scan_in_progress(), "pre-condition");
364
365 // Currently, only survivors can be root regions.
366 if (!_should_abort) {
367 assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
368 assert((uint)_claimed_survivor_index >= _survivors->length(),
369 "we should have claimed all survivors, claimed index = %u, length = %u",
370 (uint)_claimed_survivor_index, _survivors->length());
371 }
372
373 notify_scan_done();
374 }
375
376 bool G1CMRootRegions::wait_until_scan_finished() {
377 if (!scan_in_progress()) return false;
378
379 {
380 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
381 while (scan_in_progress()) {
382 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
383 }
384 }
385 return true;
386 }
387
388 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
389 return MAX2((n_par_threads + 2) / 4, 1U);
390 }
391
392 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
393 _g1h(g1h),
394 _markBitMap1(),
395 _markBitMap2(),
396 _parallel_marking_threads(0),
397 _max_parallel_marking_threads(0),
398 _sleep_factor(0.0),
399 _marking_task_overhead(1.0),
400 _cleanup_list("Cleanup List"),
401
402 _prevMarkBitMap(&_markBitMap1),
403 _nextMarkBitMap(&_markBitMap2),
404
405 _global_mark_stack(),
406 // _finger set in set_non_marking_state
407
408 _max_worker_id(ParallelGCThreads),
409 // _active_tasks set in set_non_marking_state
410 // _tasks set inside the constructor
411 _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)),
412 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
413
414 _has_overflown(false),
415 _concurrent(false),
416 _has_aborted(false),
417 _restart_for_overflow(false),
418 _concurrent_marking_in_progress(false),
419 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
420 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
421
422 // _verbose_level set below
423
424 _init_times(),
425 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
426 _cleanup_times(),
427 _total_counting_time(0.0),
428 _total_rs_scrub_time(0.0),
429
430 _parallel_workers(NULL),
431
432 _completed_initialization(false) {
433
434 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
435 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
436
437 // Create & start a ConcurrentMark thread.
438 _cmThread = new ConcurrentMarkThread(this);
439 assert(cmThread() != NULL, "CM Thread should have been created");
440 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
441 if (_cmThread->osthread() == NULL) {
442 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
443 }
444
445 assert(CGC_lock != NULL, "Where's the CGC_lock?");
446 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
447 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
448
449 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
450 satb_qs.set_buffer_size(G1SATBBufferSize);
451
452 _root_regions.init(_g1h->survivor(), this);
453
454 if (ConcGCThreads > ParallelGCThreads) {
455 log_warning(gc)("Can't have more ConcGCThreads (%u) than ParallelGCThreads (%u).",
456 ConcGCThreads, ParallelGCThreads);
457 return;
458 }
459 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
460 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
461 // if both are set
462 _sleep_factor = 0.0;
463 _marking_task_overhead = 1.0;
464 } else if (G1MarkingOverheadPercent > 0) {
465 // We will calculate the number of parallel marking threads based
466 // on a target overhead with respect to the soft real-time goal
467 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
468 double overall_cm_overhead =
469 (double) MaxGCPauseMillis * marking_overhead /
470 (double) GCPauseIntervalMillis;
471 double cpu_ratio = 1.0 / os::initial_active_processor_count();
472 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
473 double marking_task_overhead =
474 overall_cm_overhead / marking_thread_num * os::initial_active_processor_count();
475 double sleep_factor =
476 (1.0 - marking_task_overhead) / marking_task_overhead;
477
478 FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num);
479 _sleep_factor = sleep_factor;
480 _marking_task_overhead = marking_task_overhead;
481 } else {
482 // Calculate the number of parallel marking threads by scaling
483 // the number of parallel GC threads.
484 uint marking_thread_num = scale_parallel_threads(ParallelGCThreads);
485 FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
486 _sleep_factor = 0.0;
487 _marking_task_overhead = 1.0;
488 }
489
490 assert(ConcGCThreads > 0, "Should have been set");
491 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
492 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
493 _parallel_marking_threads = ConcGCThreads;
494 _max_parallel_marking_threads = _parallel_marking_threads;
495
496 _parallel_workers = new WorkGang("G1 Marker",
497 _max_parallel_marking_threads, false, true);
498 if (_parallel_workers == NULL) {
499 vm_exit_during_initialization("Failed necessary allocation.");
500 } else {
501 _parallel_workers->initialize_workers();
502 }
503
504 if (FLAG_IS_DEFAULT(MarkStackSize)) {
505 size_t mark_stack_size =
506 MIN2(MarkStackSizeMax,
507 MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE)));
508 // Verify that the calculated value for MarkStackSize is in range.
509 // It would be nice to use the private utility routine from Arguments.
510 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
511 log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
512 "must be between 1 and " SIZE_FORMAT,
513 mark_stack_size, MarkStackSizeMax);
514 return;
515 }
516 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
517 } else {
518 // Verify MarkStackSize is in range.
519 if (FLAG_IS_CMDLINE(MarkStackSize)) {
520 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
521 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
522 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
523 "must be between 1 and " SIZE_FORMAT,
524 MarkStackSize, MarkStackSizeMax);
525 return;
526 }
527 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
528 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
529 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
530 " or for MarkStackSizeMax (" SIZE_FORMAT ")",
531 MarkStackSize, MarkStackSizeMax);
532 return;
533 }
534 }
535 }
536 }
537
538 if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
539 vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
540 }
541
542 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC);
543 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
544
545 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
546 _active_tasks = _max_worker_id;
547
548 for (uint i = 0; i < _max_worker_id; ++i) {
549 G1CMTaskQueue* task_queue = new G1CMTaskQueue();
550 task_queue->initialize();
551 _task_queues->register_queue(i, task_queue);
552
553 _tasks[i] = new G1CMTask(i, this, task_queue, _task_queues);
554
555 _accum_task_vtime[i] = 0.0;
556 }
557
558 // so that the call below can read a sensible value
559 _heap_start = g1h->reserved_region().start();
560 set_non_marking_state();
561 _completed_initialization = true;
562 }
563
564 void G1ConcurrentMark::reset() {
565 // Starting values for these two. This should be called in a STW
566 // phase.
567 MemRegion reserved = _g1h->g1_reserved();
568 _heap_start = reserved.start();
569 _heap_end = reserved.end();
570
571 // Separated the asserts so that we know which one fires.
572 assert(_heap_start != NULL, "heap bounds should look ok");
573 assert(_heap_end != NULL, "heap bounds should look ok");
574 assert(_heap_start < _heap_end, "heap bounds should look ok");
575
576 // Reset all the marking data structures and any necessary flags
577 reset_marking_state();
578
579 // We do reset all of them, since different phases will use
580 // different number of active threads. So, it's easiest to have all
581 // of them ready.
582 for (uint i = 0; i < _max_worker_id; ++i) {
583 _tasks[i]->reset(_nextMarkBitMap);
584 }
585
586 // we need this to make sure that the flag is on during the evac
587 // pause with initial mark piggy-backed
588 set_concurrent_marking_in_progress();
589 }
590
591
592 void G1ConcurrentMark::reset_marking_state() {
593 _global_mark_stack.set_should_expand(has_overflown());
594 _global_mark_stack.set_empty();
595 clear_has_overflown();
596 _finger = _heap_start;
597
598 for (uint i = 0; i < _max_worker_id; ++i) {
599 G1CMTaskQueue* queue = _task_queues->queue(i);
600 queue->set_empty();
601 }
602 }
603
604 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
605 assert(active_tasks <= _max_worker_id, "we should not have more");
606
607 _active_tasks = active_tasks;
608 // Need to update the three data structures below according to the
609 // number of active threads for this phase.
610 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
611 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
612 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
613 }
614
615 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
616 set_concurrency(active_tasks);
617
618 _concurrent = concurrent;
619 // We propagate this to all tasks, not just the active ones.
620 for (uint i = 0; i < _max_worker_id; ++i)
621 _tasks[i]->set_concurrent(concurrent);
622
623 if (concurrent) {
624 set_concurrent_marking_in_progress();
625 } else {
626 // We currently assume that the concurrent flag has been set to
627 // false before we start remark. At this point we should also be
628 // in a STW phase.
629 assert(!concurrent_marking_in_progress(), "invariant");
630 assert(out_of_regions(),
631 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
632 p2i(_finger), p2i(_heap_end));
633 }
634 }
635
636 void G1ConcurrentMark::set_non_marking_state() {
637 // We set the global marking state to some default values when we're
638 // not doing marking.
639 reset_marking_state();
640 _active_tasks = 0;
641 clear_concurrent_marking_in_progress();
642 }
643
644 G1ConcurrentMark::~G1ConcurrentMark() {
645 // The G1ConcurrentMark instance is never freed.
646 ShouldNotReachHere();
647 }
648
649 class G1ClearBitMapTask : public AbstractGangTask {
650 public:
651 static size_t chunk_size() { return M; }
652
653 private:
654 // Heap region closure used for clearing the given mark bitmap.
655 class G1ClearBitmapHRClosure : public HeapRegionClosure {
656 private:
657 G1CMBitMap* _bitmap;
658 G1ConcurrentMark* _cm;
659 public:
660 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
661 }
662
663 virtual bool doHeapRegion(HeapRegion* r) {
664 size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
665
666 HeapWord* cur = r->bottom();
667 HeapWord* const end = r->end();
668
669 while (cur < end) {
670 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
671 _bitmap->clear_range(mr);
672
673 cur += chunk_size_in_words;
674
675 // Abort iteration if after yielding the marking has been aborted.
676 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
677 return true;
678 }
679 // Repeat the asserts from before the start of the closure. We will do them
680 // as asserts here to minimize their overhead on the product. However, we
681 // will have them as guarantees at the beginning / end of the bitmap
682 // clearing to get some checking in the product.
683 assert(_cm == NULL || _cm->cmThread()->during_cycle(), "invariant");
684 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant");
685 }
686 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
687
688 return false;
689 }
690 };
691
692 G1ClearBitmapHRClosure _cl;
693 HeapRegionClaimer _hr_claimer;
694 bool _suspendible; // If the task is suspendible, workers must join the STS.
695
696 public:
697 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
698 AbstractGangTask("G1 Clear Bitmap"),
699 _cl(bitmap, suspendible ? cm : NULL),
700 _hr_claimer(n_workers),
701 _suspendible(suspendible)
702 { }
703
704 void work(uint worker_id) {
705 SuspendibleThreadSetJoiner sts_join(_suspendible);
706 G1CollectedHeap::heap()->heap_region_par_iterate(&_cl, worker_id, &_hr_claimer, true);
707 }
708
709 bool is_complete() {
710 return _cl.complete();
711 }
712 };
713
714 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
715 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
716
717 size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
718 size_t const num_chunks = align_size_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
719
720 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
721
722 G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
723
724 log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
725 workers->run_task(&cl, num_workers);
726 guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
727 }
728
729 void G1ConcurrentMark::cleanup_for_next_mark() {
730 // Make sure that the concurrent mark thread looks to still be in
731 // the current cycle.
732 guarantee(cmThread()->during_cycle(), "invariant");
733
734 // We are finishing up the current cycle by clearing the next
735 // marking bitmap and getting it ready for the next cycle. During
736 // this time no other cycle can start. So, let's make sure that this
737 // is the case.
738 guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
739
740 clear_bitmap(_nextMarkBitMap, _parallel_workers, true);
741
742 // Clear the live count data. If the marking has been aborted, the abort()
743 // call already did that.
744 if (!has_aborted()) {
745 clear_live_data(_parallel_workers);
746 DEBUG_ONLY(verify_live_data_clear());
747 }
748
749 // Repeat the asserts from above.
750 guarantee(cmThread()->during_cycle(), "invariant");
751 guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
752 }
753
754 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
755 assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint.");
756 clear_bitmap((G1CMBitMap*)_prevMarkBitMap, workers, false);
757 }
758
759 class CheckBitmapClearHRClosure : public HeapRegionClosure {
760 G1CMBitMap* _bitmap;
761 bool _error;
762 public:
763 CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
764 }
765
766 virtual bool doHeapRegion(HeapRegion* r) {
767 // This closure can be called concurrently to the mutator, so we must make sure
768 // that the result of the getNextMarkedWordAddress() call is compared to the
769 // value passed to it as limit to detect any found bits.
770 // end never changes in G1.
771 HeapWord* end = r->end();
772 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
773 }
774 };
775
776 bool G1ConcurrentMark::nextMarkBitmapIsClear() {
777 CheckBitmapClearHRClosure cl(_nextMarkBitMap);
778 _g1h->heap_region_iterate(&cl);
779 return cl.complete();
780 }
781
782 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
783 public:
784 bool doHeapRegion(HeapRegion* r) {
785 r->note_start_of_marking();
786 return false;
787 }
788 };
789
790 void G1ConcurrentMark::checkpointRootsInitialPre() {
791 G1CollectedHeap* g1h = G1CollectedHeap::heap();
792 G1Policy* g1p = g1h->g1_policy();
793
794 _has_aborted = false;
795
796 // Initialize marking structures. This has to be done in a STW phase.
797 reset();
798
799 // For each region note start of marking.
800 NoteStartOfMarkHRClosure startcl;
801 g1h->heap_region_iterate(&startcl);
802 }
803
804
805 void G1ConcurrentMark::checkpointRootsInitialPost() {
806 G1CollectedHeap* g1h = G1CollectedHeap::heap();
807
808 // Start Concurrent Marking weak-reference discovery.
809 ReferenceProcessor* rp = g1h->ref_processor_cm();
810 // enable ("weak") refs discovery
811 rp->enable_discovery();
812 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
813
814 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
815 // This is the start of the marking cycle, we're expected all
816 // threads to have SATB queues with active set to false.
817 satb_mq_set.set_active_all_threads(true, /* new active value */
818 false /* expected_active */);
819
820 _root_regions.prepare_for_scan();
821
822 // update_g1_committed() will be called at the end of an evac pause
823 // when marking is on. So, it's also called at the end of the
824 // initial-mark pause to update the heap end, if the heap expands
825 // during it. No need to call it here.
826 }
827
828 /*
829 * Notice that in the next two methods, we actually leave the STS
830 * during the barrier sync and join it immediately afterwards. If we
831 * do not do this, the following deadlock can occur: one thread could
832 * be in the barrier sync code, waiting for the other thread to also
833 * sync up, whereas another one could be trying to yield, while also
834 * waiting for the other threads to sync up too.
835 *
836 * Note, however, that this code is also used during remark and in
837 * this case we should not attempt to leave / enter the STS, otherwise
838 * we'll either hit an assert (debug / fastdebug) or deadlock
839 * (product). So we should only leave / enter the STS if we are
840 * operating concurrently.
841 *
842 * Because the thread that does the sync barrier has left the STS, it
843 * is possible to be suspended for a Full GC or an evacuation pause
844 * could occur. This is actually safe, since the entering the sync
845 * barrier is one of the last things do_marking_step() does, and it
846 * doesn't manipulate any data structures afterwards.
847 */
848
849 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
850 bool barrier_aborted;
851 {
852 SuspendibleThreadSetLeaver sts_leave(concurrent());
853 barrier_aborted = !_first_overflow_barrier_sync.enter();
854 }
855
856 // at this point everyone should have synced up and not be doing any
857 // more work
858
859 if (barrier_aborted) {
860 // If the barrier aborted we ignore the overflow condition and
861 // just abort the whole marking phase as quickly as possible.
862 return;
863 }
864
865 // If we're executing the concurrent phase of marking, reset the marking
866 // state; otherwise the marking state is reset after reference processing,
867 // during the remark pause.
868 // If we reset here as a result of an overflow during the remark we will
869 // see assertion failures from any subsequent set_concurrency_and_phase()
870 // calls.
871 if (concurrent()) {
872 // let the task associated with with worker 0 do this
873 if (worker_id == 0) {
874 // task 0 is responsible for clearing the global data structures
875 // We should be here because of an overflow. During STW we should
876 // not clear the overflow flag since we rely on it being true when
877 // we exit this method to abort the pause and restart concurrent
878 // marking.
879 reset_marking_state();
880
881 log_info(gc, marking)("Concurrent Mark reset for overflow");
882 }
883 }
884
885 // after this, each task should reset its own data structures then
886 // then go into the second barrier
887 }
888
889 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
890 SuspendibleThreadSetLeaver sts_leave(concurrent());
891 _second_overflow_barrier_sync.enter();
892
893 // at this point everything should be re-initialized and ready to go
894 }
895
896 class G1CMConcurrentMarkingTask: public AbstractGangTask {
897 private:
898 G1ConcurrentMark* _cm;
899 ConcurrentMarkThread* _cmt;
900
901 public:
902 void work(uint worker_id) {
903 assert(Thread::current()->is_ConcurrentGC_thread(),
904 "this should only be done by a conc GC thread");
905 ResourceMark rm;
906
907 double start_vtime = os::elapsedVTime();
908
909 {
910 SuspendibleThreadSetJoiner sts_join;
911
912 assert(worker_id < _cm->active_tasks(), "invariant");
913 G1CMTask* the_task = _cm->task(worker_id);
914 the_task->record_start_time();
915 if (!_cm->has_aborted()) {
916 do {
917 double start_vtime_sec = os::elapsedVTime();
918 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
919
920 the_task->do_marking_step(mark_step_duration_ms,
921 true /* do_termination */,
922 false /* is_serial*/);
923
924 double end_vtime_sec = os::elapsedVTime();
925 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
926 _cm->do_yield_check();
927
928 jlong sleep_time_ms;
929 if (!_cm->has_aborted() && the_task->has_aborted()) {
930 sleep_time_ms =
931 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
932 {
933 SuspendibleThreadSetLeaver sts_leave;
934 os::sleep(Thread::current(), sleep_time_ms, false);
935 }
936 }
937 } while (!_cm->has_aborted() && the_task->has_aborted());
938 }
939 the_task->record_end_time();
940 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
941 }
942
943 double end_vtime = os::elapsedVTime();
944 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
945 }
946
947 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm,
948 ConcurrentMarkThread* cmt) :
949 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
950
951 ~G1CMConcurrentMarkingTask() { }
952 };
953
954 // Calculates the number of active workers for a concurrent
955 // phase.
956 uint G1ConcurrentMark::calc_parallel_marking_threads() {
957 uint n_conc_workers = 0;
958 if (!UseDynamicNumberOfGCThreads ||
959 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
960 !ForceDynamicNumberOfGCThreads)) {
961 n_conc_workers = max_parallel_marking_threads();
962 } else {
963 n_conc_workers =
964 AdaptiveSizePolicy::calc_default_active_workers(max_parallel_marking_threads(),
965 1, /* Minimum workers */
966 parallel_marking_threads(),
967 Threads::number_of_non_daemon_threads());
968 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
969 // that scaling has already gone into "_max_parallel_marking_threads".
970 }
971 assert(n_conc_workers > 0 && n_conc_workers <= max_parallel_marking_threads(),
972 "Calculated number of workers must be larger than zero and at most the maximum %u, but is %u",
973 max_parallel_marking_threads(), n_conc_workers);
974 return n_conc_workers;
975 }
976
977 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr) {
978 // Currently, only survivors can be root regions.
979 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
980 G1RootRegionScanClosure cl(_g1h, this);
981
982 const uintx interval = PrefetchScanIntervalInBytes;
983 HeapWord* curr = hr->bottom();
984 const HeapWord* end = hr->top();
985 while (curr < end) {
986 Prefetch::read(curr, interval);
987 oop obj = oop(curr);
988 int size = obj->oop_iterate_size(&cl);
989 assert(size == obj->size(), "sanity");
990 curr += size;
991 }
992 }
993
994 class G1CMRootRegionScanTask : public AbstractGangTask {
995 private:
996 G1ConcurrentMark* _cm;
997
998 public:
999 G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
1000 AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
1001
1002 void work(uint worker_id) {
1003 assert(Thread::current()->is_ConcurrentGC_thread(),
1004 "this should only be done by a conc GC thread");
1005
1006 G1CMRootRegions* root_regions = _cm->root_regions();
1007 HeapRegion* hr = root_regions->claim_next();
1008 while (hr != NULL) {
1009 _cm->scanRootRegion(hr);
1010 hr = root_regions->claim_next();
1011 }
1012 }
1013 };
1014
1015 void G1ConcurrentMark::scan_root_regions() {
1016 // scan_in_progress() will have been set to true only if there was
1017 // at least one root region to scan. So, if it's false, we
1018 // should not attempt to do any further work.
1019 if (root_regions()->scan_in_progress()) {
1020 assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
1021
1022 _parallel_marking_threads = MIN2(calc_parallel_marking_threads(),
1023 // We distribute work on a per-region basis, so starting
1024 // more threads than that is useless.
1025 root_regions()->num_root_regions());
1026 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1027 "Maximum number of marking threads exceeded");
1028
1029 G1CMRootRegionScanTask task(this);
1030 log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
1031 task.name(), _parallel_marking_threads, root_regions()->num_root_regions());
1032 _parallel_workers->run_task(&task, _parallel_marking_threads);
1033
1034 // It's possible that has_aborted() is true here without actually
1035 // aborting the survivor scan earlier. This is OK as it's
1036 // mainly used for sanity checking.
1037 root_regions()->scan_finished();
1038 }
1039 }
1040
1041 void G1ConcurrentMark::concurrent_cycle_start() {
1042 _gc_timer_cm->register_gc_start();
1043
1044 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
1045
1046 _g1h->trace_heap_before_gc(_gc_tracer_cm);
1047 }
1048
1049 void G1ConcurrentMark::concurrent_cycle_end() {
1050 _g1h->trace_heap_after_gc(_gc_tracer_cm);
1051
1052 if (has_aborted()) {
1053 _gc_tracer_cm->report_concurrent_mode_failure();
1054 }
1055
1056 _gc_timer_cm->register_gc_end();
1057
1058 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1059 }
1060
1061 void G1ConcurrentMark::mark_from_roots() {
1062 // we might be tempted to assert that:
1063 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1064 // "inconsistent argument?");
1065 // However that wouldn't be right, because it's possible that
1066 // a safepoint is indeed in progress as a younger generation
1067 // stop-the-world GC happens even as we mark in this generation.
1068
1069 _restart_for_overflow = false;
1070
1071 // _g1h has _n_par_threads
1072 _parallel_marking_threads = calc_parallel_marking_threads();
1073 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1074 "Maximum number of marking threads exceeded");
1075
1076 uint active_workers = MAX2(1U, parallel_marking_threads());
1077 assert(active_workers > 0, "Should have been set");
1078
1079 // Setting active workers is not guaranteed since fewer
1080 // worker threads may currently exist and more may not be
1081 // available.
1082 active_workers = _parallel_workers->update_active_workers(active_workers);
1083 log_info(gc, task)("Using %u workers of %u for marking", active_workers, _parallel_workers->total_workers());
1084
1085 // Parallel task terminator is set in "set_concurrency_and_phase()"
1086 set_concurrency_and_phase(active_workers, true /* concurrent */);
1087
1088 G1CMConcurrentMarkingTask markingTask(this, cmThread());
1089 _parallel_workers->run_task(&markingTask);
1090 print_stats();
1091 }
1092
1093 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1094 // world is stopped at this checkpoint
1095 assert(SafepointSynchronize::is_at_safepoint(),
1096 "world should be stopped");
1097
1098 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1099
1100 // If a full collection has happened, we shouldn't do this.
1101 if (has_aborted()) {
1102 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1103 return;
1104 }
1105
1106 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1107
1108 if (VerifyDuringGC) {
1109 HandleMark hm; // handle scope
1110 g1h->prepare_for_verify();
1111 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1112 }
1113 g1h->verifier()->check_bitmaps("Remark Start");
1114
1115 G1Policy* g1p = g1h->g1_policy();
1116 g1p->record_concurrent_mark_remark_start();
1117
1118 double start = os::elapsedTime();
1119
1120 checkpointRootsFinalWork();
1121
1122 double mark_work_end = os::elapsedTime();
1123
1124 weakRefsWork(clear_all_soft_refs);
1125
1126 if (has_overflown()) {
1127 // We overflowed. Restart concurrent marking.
1128 _restart_for_overflow = true;
1129
1130 // Verify the heap w.r.t. the previous marking bitmap.
1131 if (VerifyDuringGC) {
1132 HandleMark hm; // handle scope
1133 g1h->prepare_for_verify();
1134 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)");
1135 }
1136
1137 // Clear the marking state because we will be restarting
1138 // marking due to overflowing the global mark stack.
1139 reset_marking_state();
1140 } else {
1141 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1142 // We're done with marking.
1143 // This is the end of the marking cycle, we're expected all
1144 // threads to have SATB queues with active set to true.
1145 satb_mq_set.set_active_all_threads(false, /* new active value */
1146 true /* expected_active */);
1147
1148 if (VerifyDuringGC) {
1149 HandleMark hm; // handle scope
1150 g1h->prepare_for_verify();
1151 Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)");
1152 }
1153 g1h->verifier()->check_bitmaps("Remark End");
1154 assert(!restart_for_overflow(), "sanity");
1155 // Completely reset the marking state since marking completed
1156 set_non_marking_state();
1157 }
1158
1159 // Expand the marking stack, if we have to and if we can.
1160 if (_global_mark_stack.should_expand()) {
1161 _global_mark_stack.expand();
1162 }
1163
1164 // Statistics
1165 double now = os::elapsedTime();
1166 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1167 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1168 _remark_times.add((now - start) * 1000.0);
1169
1170 g1p->record_concurrent_mark_remark_end();
1171
1172 G1CMIsAliveClosure is_alive(g1h);
1173 _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1174 }
1175
1176 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1177 G1CollectedHeap* _g1;
1178 size_t _freed_bytes;
1179 FreeRegionList* _local_cleanup_list;
1180 uint _old_regions_removed;
1181 uint _humongous_regions_removed;
1182 HRRSCleanupTask* _hrrs_cleanup_task;
1183
1184 public:
1185 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1186 FreeRegionList* local_cleanup_list,
1187 HRRSCleanupTask* hrrs_cleanup_task) :
1188 _g1(g1),
1189 _freed_bytes(0),
1190 _local_cleanup_list(local_cleanup_list),
1191 _old_regions_removed(0),
1192 _humongous_regions_removed(0),
1193 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1194
1195 size_t freed_bytes() { return _freed_bytes; }
1196 const uint old_regions_removed() { return _old_regions_removed; }
1197 const uint humongous_regions_removed() { return _humongous_regions_removed; }
1198
1199 bool doHeapRegion(HeapRegion *hr) {
1200 if (hr->is_archive()) {
1201 return false;
1202 }
1203 _g1->reset_gc_time_stamps(hr);
1204 hr->note_end_of_marking();
1205
1206 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1207 _freed_bytes += hr->used();
1208 hr->set_containing_set(NULL);
1209 if (hr->is_humongous()) {
1210 _humongous_regions_removed++;
1211 _g1->free_humongous_region(hr, _local_cleanup_list, true /* skip_remset */);
1212 } else {
1213 _old_regions_removed++;
1214 _g1->free_region(hr, _local_cleanup_list, true /* skip_remset */);
1215 }
1216 } else {
1217 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1218 }
1219
1220 return false;
1221 }
1222 };
1223
1224 class G1ParNoteEndTask: public AbstractGangTask {
1225 friend class G1NoteEndOfConcMarkClosure;
1226
1227 protected:
1228 G1CollectedHeap* _g1h;
1229 FreeRegionList* _cleanup_list;
1230 HeapRegionClaimer _hrclaimer;
1231
1232 public:
1233 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1234 AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1235 }
1236
1237 void work(uint worker_id) {
1238 FreeRegionList local_cleanup_list("Local Cleanup List");
1239 HRRSCleanupTask hrrs_cleanup_task;
1240 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1241 &hrrs_cleanup_task);
1242 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1243 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1244
1245 // Now update the lists
1246 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1247 {
1248 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1249 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1250
1251 // If we iterate over the global cleanup list at the end of
1252 // cleanup to do this printing we will not guarantee to only
1253 // generate output for the newly-reclaimed regions (the list
1254 // might not be empty at the beginning of cleanup; we might
1255 // still be working on its previous contents). So we do the
1256 // printing here, before we append the new regions to the global
1257 // cleanup list.
1258
1259 G1HRPrinter* hr_printer = _g1h->hr_printer();
1260 if (hr_printer->is_active()) {
1261 FreeRegionListIterator iter(&local_cleanup_list);
1262 while (iter.more_available()) {
1263 HeapRegion* hr = iter.get_next();
1264 hr_printer->cleanup(hr);
1265 }
1266 }
1267
1268 _cleanup_list->add_ordered(&local_cleanup_list);
1269 assert(local_cleanup_list.is_empty(), "post-condition");
1270
1271 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1272 }
1273 }
1274 };
1275
1276 void G1ConcurrentMark::cleanup() {
1277 // world is stopped at this checkpoint
1278 assert(SafepointSynchronize::is_at_safepoint(),
1279 "world should be stopped");
1280 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1281
1282 // If a full collection has happened, we shouldn't do this.
1283 if (has_aborted()) {
1284 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1285 return;
1286 }
1287
1288 g1h->verifier()->verify_region_sets_optional();
1289
1290 if (VerifyDuringGC) {
1291 HandleMark hm; // handle scope
1292 g1h->prepare_for_verify();
1293 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1294 }
1295 g1h->verifier()->check_bitmaps("Cleanup Start");
1296
1297 G1Policy* g1p = g1h->g1_policy();
1298 g1p->record_concurrent_mark_cleanup_start();
1299
1300 double start = os::elapsedTime();
1301
1302 HeapRegionRemSet::reset_for_cleanup_tasks();
1303
1304 {
1305 GCTraceTime(Debug, gc)("Finalize Live Data");
1306 finalize_live_data();
1307 }
1308
1309 if (VerifyDuringGC) {
1310 GCTraceTime(Debug, gc)("Verify Live Data");
1311 verify_live_data();
1312 }
1313
1314 g1h->collector_state()->set_mark_in_progress(false);
1315
1316 double count_end = os::elapsedTime();
1317 double this_final_counting_time = (count_end - start);
1318 _total_counting_time += this_final_counting_time;
1319
1320 if (log_is_enabled(Trace, gc, liveness)) {
1321 G1PrintRegionLivenessInfoClosure cl("Post-Marking");
1322 _g1h->heap_region_iterate(&cl);
1323 }
1324
1325 // Install newly created mark bitMap as "prev".
1326 swapMarkBitMaps();
1327
1328 g1h->reset_gc_time_stamp();
1329
1330 uint n_workers = _g1h->workers()->active_workers();
1331
1332 // Note end of marking in all heap regions.
1333 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
1334 g1h->workers()->run_task(&g1_par_note_end_task);
1335 g1h->check_gc_time_stamps();
1336
1337 if (!cleanup_list_is_empty()) {
1338 // The cleanup list is not empty, so we'll have to process it
1339 // concurrently. Notify anyone else that might be wanting free
1340 // regions that there will be more free regions coming soon.
1341 g1h->set_free_regions_coming();
1342 }
1343
1344 // call below, since it affects the metric by which we sort the heap
1345 // regions.
1346 if (G1ScrubRemSets) {
1347 double rs_scrub_start = os::elapsedTime();
1348 g1h->scrub_rem_set();
1349 _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start);
1350 }
1351
1352 // this will also free any regions totally full of garbage objects,
1353 // and sort the regions.
1354 g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1355
1356 // Statistics.
1357 double end = os::elapsedTime();
1358 _cleanup_times.add((end - start) * 1000.0);
1359
1360 // Clean up will have freed any regions completely full of garbage.
1361 // Update the soft reference policy with the new heap occupancy.
1362 Universe::update_heap_info_at_gc();
1363
1364 if (VerifyDuringGC) {
1365 HandleMark hm; // handle scope
1366 g1h->prepare_for_verify();
1367 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)");
1368 }
1369
1370 g1h->verifier()->check_bitmaps("Cleanup End");
1371
1372 g1h->verifier()->verify_region_sets_optional();
1373
1374 // We need to make this be a "collection" so any collection pause that
1375 // races with it goes around and waits for completeCleanup to finish.
1376 g1h->increment_total_collections();
1377
1378 // Clean out dead classes and update Metaspace sizes.
1379 if (ClassUnloadingWithConcurrentMark) {
1380 ClassLoaderDataGraph::purge();
1381 }
1382 MetaspaceGC::compute_new_size();
1383
1384 // We reclaimed old regions so we should calculate the sizes to make
1385 // sure we update the old gen/space data.
1386 g1h->g1mm()->update_sizes();
1387 g1h->allocation_context_stats().update_after_mark();
1388 }
1389
1390 void G1ConcurrentMark::complete_cleanup() {
1391 if (has_aborted()) return;
1392
1393 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1394
1395 _cleanup_list.verify_optional();
1396 FreeRegionList tmp_free_list("Tmp Free List");
1397
1398 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1399 "cleanup list has %u entries",
1400 _cleanup_list.length());
1401
1402 // No one else should be accessing the _cleanup_list at this point,
1403 // so it is not necessary to take any locks
1404 while (!_cleanup_list.is_empty()) {
1405 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
1406 assert(hr != NULL, "Got NULL from a non-empty list");
1407 hr->par_clear();
1408 tmp_free_list.add_ordered(hr);
1409
1410 // Instead of adding one region at a time to the secondary_free_list,
1411 // we accumulate them in the local list and move them a few at a
1412 // time. This also cuts down on the number of notify_all() calls
1413 // we do during this process. We'll also append the local list when
1414 // _cleanup_list is empty (which means we just removed the last
1415 // region from the _cleanup_list).
1416 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
1417 _cleanup_list.is_empty()) {
1418 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1419 "appending %u entries to the secondary_free_list, "
1420 "cleanup list still has %u entries",
1421 tmp_free_list.length(),
1422 _cleanup_list.length());
1423
1424 {
1425 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
1426 g1h->secondary_free_list_add(&tmp_free_list);
1427 SecondaryFreeList_lock->notify_all();
1428 }
1429 #ifndef PRODUCT
1430 if (G1StressConcRegionFreeing) {
1431 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
1432 os::sleep(Thread::current(), (jlong) 1, false);
1433 }
1434 }
1435 #endif
1436 }
1437 }
1438 assert(tmp_free_list.is_empty(), "post-condition");
1439 }
1440
1441 // Supporting Object and Oop closures for reference discovery
1442 // and processing in during marking
1443
1444 bool G1CMIsAliveClosure::do_object_b(oop obj) {
1445 HeapWord* addr = (HeapWord*)obj;
1446 return addr != NULL &&
1447 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
1448 }
1449
1450 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1451 // Uses the G1CMTask associated with a worker thread (for serial reference
1452 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1453 // trace referent objects.
1454 //
1455 // Using the G1CMTask and embedded local queues avoids having the worker
1456 // threads operating on the global mark stack. This reduces the risk
1457 // of overflowing the stack - which we would rather avoid at this late
1458 // state. Also using the tasks' local queues removes the potential
1459 // of the workers interfering with each other that could occur if
1460 // operating on the global stack.
1461
1462 class G1CMKeepAliveAndDrainClosure: public OopClosure {
1463 G1ConcurrentMark* _cm;
1464 G1CMTask* _task;
1465 int _ref_counter_limit;
1466 int _ref_counter;
1467 bool _is_serial;
1468 public:
1469 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1470 _cm(cm), _task(task), _is_serial(is_serial),
1471 _ref_counter_limit(G1RefProcDrainInterval) {
1472 assert(_ref_counter_limit > 0, "sanity");
1473 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1474 _ref_counter = _ref_counter_limit;
1475 }
1476
1477 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1478 virtual void do_oop( oop* p) { do_oop_work(p); }
1479
1480 template <class T> void do_oop_work(T* p) {
1481 if (!_cm->has_overflown()) {
1482 oop obj = oopDesc::load_decode_heap_oop(p);
1483 _task->deal_with_reference(obj);
1484 _ref_counter--;
1485
1486 if (_ref_counter == 0) {
1487 // We have dealt with _ref_counter_limit references, pushing them
1488 // and objects reachable from them on to the local stack (and
1489 // possibly the global stack). Call G1CMTask::do_marking_step() to
1490 // process these entries.
1491 //
1492 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1493 // there's nothing more to do (i.e. we're done with the entries that
1494 // were pushed as a result of the G1CMTask::deal_with_reference() calls
1495 // above) or we overflow.
1496 //
1497 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1498 // flag while there may still be some work to do. (See the comment at
1499 // the beginning of G1CMTask::do_marking_step() for those conditions -
1500 // one of which is reaching the specified time target.) It is only
1501 // when G1CMTask::do_marking_step() returns without setting the
1502 // has_aborted() flag that the marking step has completed.
1503 do {
1504 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1505 _task->do_marking_step(mark_step_duration_ms,
1506 false /* do_termination */,
1507 _is_serial);
1508 } while (_task->has_aborted() && !_cm->has_overflown());
1509 _ref_counter = _ref_counter_limit;
1510 }
1511 }
1512 }
1513 };
1514
1515 // 'Drain' oop closure used by both serial and parallel reference processing.
1516 // Uses the G1CMTask associated with a given worker thread (for serial
1517 // reference processing the G1CMtask for worker 0 is used). Calls the
1518 // do_marking_step routine, with an unbelievably large timeout value,
1519 // to drain the marking data structures of the remaining entries
1520 // added by the 'keep alive' oop closure above.
1521
1522 class G1CMDrainMarkingStackClosure: public VoidClosure {
1523 G1ConcurrentMark* _cm;
1524 G1CMTask* _task;
1525 bool _is_serial;
1526 public:
1527 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1528 _cm(cm), _task(task), _is_serial(is_serial) {
1529 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1530 }
1531
1532 void do_void() {
1533 do {
1534 // We call G1CMTask::do_marking_step() to completely drain the local
1535 // and global marking stacks of entries pushed by the 'keep alive'
1536 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1537 //
1538 // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1539 // if there's nothing more to do (i.e. we've completely drained the
1540 // entries that were pushed as a a result of applying the 'keep alive'
1541 // closure to the entries on the discovered ref lists) or we overflow
1542 // the global marking stack.
1543 //
1544 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1545 // flag while there may still be some work to do. (See the comment at
1546 // the beginning of G1CMTask::do_marking_step() for those conditions -
1547 // one of which is reaching the specified time target.) It is only
1548 // when G1CMTask::do_marking_step() returns without setting the
1549 // has_aborted() flag that the marking step has completed.
1550
1551 _task->do_marking_step(1000000000.0 /* something very large */,
1552 true /* do_termination */,
1553 _is_serial);
1554 } while (_task->has_aborted() && !_cm->has_overflown());
1555 }
1556 };
1557
1558 // Implementation of AbstractRefProcTaskExecutor for parallel
1559 // reference processing at the end of G1 concurrent marking
1560
1561 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
1562 private:
1563 G1CollectedHeap* _g1h;
1564 G1ConcurrentMark* _cm;
1565 WorkGang* _workers;
1566 uint _active_workers;
1567
1568 public:
1569 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1570 G1ConcurrentMark* cm,
1571 WorkGang* workers,
1572 uint n_workers) :
1573 _g1h(g1h), _cm(cm),
1574 _workers(workers), _active_workers(n_workers) { }
1575
1576 // Executes the given task using concurrent marking worker threads.
1577 virtual void execute(ProcessTask& task);
1578 virtual void execute(EnqueueTask& task);
1579 };
1580
1581 class G1CMRefProcTaskProxy: public AbstractGangTask {
1582 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1583 ProcessTask& _proc_task;
1584 G1CollectedHeap* _g1h;
1585 G1ConcurrentMark* _cm;
1586
1587 public:
1588 G1CMRefProcTaskProxy(ProcessTask& proc_task,
1589 G1CollectedHeap* g1h,
1590 G1ConcurrentMark* cm) :
1591 AbstractGangTask("Process reference objects in parallel"),
1592 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1593 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1594 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1595 }
1596
1597 virtual void work(uint worker_id) {
1598 ResourceMark rm;
1599 HandleMark hm;
1600 G1CMTask* task = _cm->task(worker_id);
1601 G1CMIsAliveClosure g1_is_alive(_g1h);
1602 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1603 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1604
1605 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1606 }
1607 };
1608
1609 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1610 assert(_workers != NULL, "Need parallel worker threads.");
1611 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1612
1613 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1614
1615 // We need to reset the concurrency level before each
1616 // proxy task execution, so that the termination protocol
1617 // and overflow handling in G1CMTask::do_marking_step() knows
1618 // how many workers to wait for.
1619 _cm->set_concurrency(_active_workers);
1620 _workers->run_task(&proc_task_proxy);
1621 }
1622
1623 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
1624 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1625 EnqueueTask& _enq_task;
1626
1627 public:
1628 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1629 AbstractGangTask("Enqueue reference objects in parallel"),
1630 _enq_task(enq_task) { }
1631
1632 virtual void work(uint worker_id) {
1633 _enq_task.work(worker_id);
1634 }
1635 };
1636
1637 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1638 assert(_workers != NULL, "Need parallel worker threads.");
1639 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1640
1641 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1642
1643 // Not strictly necessary but...
1644 //
1645 // We need to reset the concurrency level before each
1646 // proxy task execution, so that the termination protocol
1647 // and overflow handling in G1CMTask::do_marking_step() knows
1648 // how many workers to wait for.
1649 _cm->set_concurrency(_active_workers);
1650 _workers->run_task(&enq_task_proxy);
1651 }
1652
1653 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
1654 if (has_overflown()) {
1655 // Skip processing the discovered references if we have
1656 // overflown the global marking stack. Reference objects
1657 // only get discovered once so it is OK to not
1658 // de-populate the discovered reference lists. We could have,
1659 // but the only benefit would be that, when marking restarts,
1660 // less reference objects are discovered.
1661 return;
1662 }
1663
1664 ResourceMark rm;
1665 HandleMark hm;
1666
1667 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1668
1669 // Is alive closure.
1670 G1CMIsAliveClosure g1_is_alive(g1h);
1671
1672 // Inner scope to exclude the cleaning of the string and symbol
1673 // tables from the displayed time.
1674 {
1675 GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm);
1676
1677 ReferenceProcessor* rp = g1h->ref_processor_cm();
1678
1679 // See the comment in G1CollectedHeap::ref_processing_init()
1680 // about how reference processing currently works in G1.
1681
1682 // Set the soft reference policy
1683 rp->setup_policy(clear_all_soft_refs);
1684 assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1685
1686 // Instances of the 'Keep Alive' and 'Complete GC' closures used
1687 // in serial reference processing. Note these closures are also
1688 // used for serially processing (by the the current thread) the
1689 // JNI references during parallel reference processing.
1690 //
1691 // These closures do not need to synchronize with the worker
1692 // threads involved in parallel reference processing as these
1693 // instances are executed serially by the current thread (e.g.
1694 // reference processing is not multi-threaded and is thus
1695 // performed by the current thread instead of a gang worker).
1696 //
1697 // The gang tasks involved in parallel reference processing create
1698 // their own instances of these closures, which do their own
1699 // synchronization among themselves.
1700 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1701 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1702
1703 // We need at least one active thread. If reference processing
1704 // is not multi-threaded we use the current (VMThread) thread,
1705 // otherwise we use the work gang from the G1CollectedHeap and
1706 // we utilize all the worker threads we can.
1707 bool processing_is_mt = rp->processing_is_mt();
1708 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
1709 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
1710
1711 // Parallel processing task executor.
1712 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
1713 g1h->workers(), active_workers);
1714 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1715
1716 // Set the concurrency level. The phase was already set prior to
1717 // executing the remark task.
1718 set_concurrency(active_workers);
1719
1720 // Set the degree of MT processing here. If the discovery was done MT,
1721 // the number of threads involved during discovery could differ from
1722 // the number of active workers. This is OK as long as the discovered
1723 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1724 rp->set_active_mt_degree(active_workers);
1725
1726 // Process the weak references.
1727 const ReferenceProcessorStats& stats =
1728 rp->process_discovered_references(&g1_is_alive,
1729 &g1_keep_alive,
1730 &g1_drain_mark_stack,
1731 executor,
1732 _gc_timer_cm);
1733 _gc_tracer_cm->report_gc_reference_stats(stats);
1734
1735 // The do_oop work routines of the keep_alive and drain_marking_stack
1736 // oop closures will set the has_overflown flag if we overflow the
1737 // global marking stack.
1738
1739 assert(has_overflown() || _global_mark_stack.is_empty(),
1740 "Mark stack should be empty (unless it has overflown)");
1741
1742 assert(rp->num_q() == active_workers, "why not");
1743
1744 rp->enqueue_discovered_references(executor);
1745
1746 rp->verify_no_references_recorded();
1747 assert(!rp->discovery_enabled(), "Post condition");
1748 }
1749
1750 if (has_overflown()) {
1751 // We can not trust g1_is_alive if the marking stack overflowed
1752 return;
1753 }
1754
1755 assert(_global_mark_stack.is_empty(), "Marking should have completed");
1756
1757 // Unload Klasses, String, Symbols, Code Cache, etc.
1758 if (ClassUnloadingWithConcurrentMark) {
1759 GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1760 bool purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
1761 g1h->complete_cleaning(&g1_is_alive, purged_classes);
1762 } else {
1763 GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm);
1764 // No need to clean string table and symbol table as they are treated as strong roots when
1765 // class unloading is disabled.
1766 g1h->partial_cleaning(&g1_is_alive, false, false, G1StringDedup::is_enabled());
1767
1768 }
1769 }
1770
1771 void G1ConcurrentMark::swapMarkBitMaps() {
1772 G1CMBitMapRO* temp = _prevMarkBitMap;
1773 _prevMarkBitMap = (G1CMBitMapRO*)_nextMarkBitMap;
1774 _nextMarkBitMap = (G1CMBitMap*) temp;
1775 }
1776
1777 // Closure for marking entries in SATB buffers.
1778 class G1CMSATBBufferClosure : public SATBBufferClosure {
1779 private:
1780 G1CMTask* _task;
1781 G1CollectedHeap* _g1h;
1782
1783 // This is very similar to G1CMTask::deal_with_reference, but with
1784 // more relaxed requirements for the argument, so this must be more
1785 // circumspect about treating the argument as an object.
1786 void do_entry(void* entry) const {
1787 _task->increment_refs_reached();
1788 HeapRegion* hr = _g1h->heap_region_containing(entry);
1789 if (entry < hr->next_top_at_mark_start()) {
1790 // Until we get here, we don't know whether entry refers to a valid
1791 // object; it could instead have been a stale reference.
1792 oop obj = static_cast<oop>(entry);
1793 assert(obj->is_oop(true /* ignore mark word */),
1794 "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj));
1795 _task->make_reference_grey(obj);
1796 }
1797 }
1798
1799 public:
1800 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1801 : _task(task), _g1h(g1h) { }
1802
1803 virtual void do_buffer(void** buffer, size_t size) {
1804 for (size_t i = 0; i < size; ++i) {
1805 do_entry(buffer[i]);
1806 }
1807 }
1808 };
1809
1810 class G1RemarkThreadsClosure : public ThreadClosure {
1811 G1CMSATBBufferClosure _cm_satb_cl;
1812 G1CMOopClosure _cm_cl;
1813 MarkingCodeBlobClosure _code_cl;
1814 int _thread_parity;
1815
1816 public:
1817 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1818 _cm_satb_cl(task, g1h),
1819 _cm_cl(g1h, g1h->concurrent_mark(), task),
1820 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1821 _thread_parity(Threads::thread_claim_parity()) {}
1822
1823 void do_thread(Thread* thread) {
1824 if (thread->is_Java_thread()) {
1825 if (thread->claim_oops_do(true, _thread_parity)) {
1826 JavaThread* jt = (JavaThread*)thread;
1827
1828 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1829 // however the liveness of oops reachable from nmethods have very complex lifecycles:
1830 // * Alive if on the stack of an executing method
1831 // * Weakly reachable otherwise
1832 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1833 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1834 jt->nmethods_do(&_code_cl);
1835
1836 jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
1837 }
1838 } else if (thread->is_VM_thread()) {
1839 if (thread->claim_oops_do(true, _thread_parity)) {
1840 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1841 }
1842 }
1843 }
1844 };
1845
1846 class G1CMRemarkTask: public AbstractGangTask {
1847 private:
1848 G1ConcurrentMark* _cm;
1849 public:
1850 void work(uint worker_id) {
1851 // Since all available tasks are actually started, we should
1852 // only proceed if we're supposed to be active.
1853 if (worker_id < _cm->active_tasks()) {
1854 G1CMTask* task = _cm->task(worker_id);
1855 task->record_start_time();
1856 {
1857 ResourceMark rm;
1858 HandleMark hm;
1859
1860 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1861 Threads::threads_do(&threads_f);
1862 }
1863
1864 do {
1865 task->do_marking_step(1000000000.0 /* something very large */,
1866 true /* do_termination */,
1867 false /* is_serial */);
1868 } while (task->has_aborted() && !_cm->has_overflown());
1869 // If we overflow, then we do not want to restart. We instead
1870 // want to abort remark and do concurrent marking again.
1871 task->record_end_time();
1872 }
1873 }
1874
1875 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1876 AbstractGangTask("Par Remark"), _cm(cm) {
1877 _cm->terminator()->reset_for_reuse(active_workers);
1878 }
1879 };
1880
1881 void G1ConcurrentMark::checkpointRootsFinalWork() {
1882 ResourceMark rm;
1883 HandleMark hm;
1884 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1885
1886 GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm);
1887
1888 g1h->ensure_parsability(false);
1889
1890 // this is remark, so we'll use up all active threads
1891 uint active_workers = g1h->workers()->active_workers();
1892 set_concurrency_and_phase(active_workers, false /* concurrent */);
1893 // Leave _parallel_marking_threads at it's
1894 // value originally calculated in the G1ConcurrentMark
1895 // constructor and pass values of the active workers
1896 // through the gang in the task.
1897
1898 {
1899 StrongRootsScope srs(active_workers);
1900
1901 G1CMRemarkTask remarkTask(this, active_workers);
1902 // We will start all available threads, even if we decide that the
1903 // active_workers will be fewer. The extra ones will just bail out
1904 // immediately.
1905 g1h->workers()->run_task(&remarkTask);
1906 }
1907
1908 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1909 guarantee(has_overflown() ||
1910 satb_mq_set.completed_buffers_num() == 0,
1911 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1912 BOOL_TO_STR(has_overflown()),
1913 satb_mq_set.completed_buffers_num());
1914
1915 print_stats();
1916 }
1917
1918 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
1919 // Note we are overriding the read-only view of the prev map here, via
1920 // the cast.
1921 ((G1CMBitMap*)_prevMarkBitMap)->clear_range(mr);
1922 }
1923
1924 HeapRegion*
1925 G1ConcurrentMark::claim_region(uint worker_id) {
1926 // "checkpoint" the finger
1927 HeapWord* finger = _finger;
1928
1929 // _heap_end will not change underneath our feet; it only changes at
1930 // yield points.
1931 while (finger < _heap_end) {
1932 assert(_g1h->is_in_g1_reserved(finger), "invariant");
1933
1934 HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1935 // Make sure that the reads below do not float before loading curr_region.
1936 OrderAccess::loadload();
1937 // Above heap_region_containing may return NULL as we always scan claim
1938 // until the end of the heap. In this case, just jump to the next region.
1939 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1940
1941 // Is the gap between reading the finger and doing the CAS too long?
1942 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
1943 if (res == finger && curr_region != NULL) {
1944 // we succeeded
1945 HeapWord* bottom = curr_region->bottom();
1946 HeapWord* limit = curr_region->next_top_at_mark_start();
1947
1948 // notice that _finger == end cannot be guaranteed here since,
1949 // someone else might have moved the finger even further
1950 assert(_finger >= end, "the finger should have moved forward");
1951
1952 if (limit > bottom) {
1953 return curr_region;
1954 } else {
1955 assert(limit == bottom,
1956 "the region limit should be at bottom");
1957 // we return NULL and the caller should try calling
1958 // claim_region() again.
1959 return NULL;
1960 }
1961 } else {
1962 assert(_finger > finger, "the finger should have moved forward");
1963 // read it again
1964 finger = _finger;
1965 }
1966 }
1967
1968 return NULL;
1969 }
1970
1971 #ifndef PRODUCT
1972 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC {
1973 private:
1974 G1CollectedHeap* _g1h;
1975 const char* _phase;
1976 int _info;
1977
1978 public:
1979 VerifyNoCSetOops(const char* phase, int info = -1) :
1980 _g1h(G1CollectedHeap::heap()),
1981 _phase(phase),
1982 _info(info)
1983 { }
1984
1985 void operator()(G1TaskQueueEntry task_entry) const {
1986 if (task_entry.is_array_slice()) {
1987 guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1988 return;
1989 }
1990 guarantee(task_entry.obj()->is_oop(),
1991 "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1992 p2i(task_entry.obj()), _phase, _info);
1993 guarantee(!_g1h->is_in_cset(task_entry.obj()),
1994 "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1995 p2i(task_entry.obj()), _phase, _info);
1996 }
1997 };
1998
1999 void G1ConcurrentMark::verify_no_cset_oops() {
2000 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2001 if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) {
2002 return;
2003 }
2004
2005 // Verify entries on the global mark stack
2006 _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
2007
2008 // Verify entries on the task queues
2009 for (uint i = 0; i < _max_worker_id; ++i) {
2010 G1CMTaskQueue* queue = _task_queues->queue(i);
2011 queue->iterate(VerifyNoCSetOops("Queue", i));
2012 }
2013
2014 // Verify the global finger
2015 HeapWord* global_finger = finger();
2016 if (global_finger != NULL && global_finger < _heap_end) {
2017 // Since we always iterate over all regions, we might get a NULL HeapRegion
2018 // here.
2019 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
2020 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
2021 "global finger: " PTR_FORMAT " region: " HR_FORMAT,
2022 p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
2023 }
2024
2025 // Verify the task fingers
2026 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
2027 for (uint i = 0; i < parallel_marking_threads(); ++i) {
2028 G1CMTask* task = _tasks[i];
2029 HeapWord* task_finger = task->finger();
2030 if (task_finger != NULL && task_finger < _heap_end) {
2031 // See above note on the global finger verification.
2032 HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
2033 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
2034 !task_hr->in_collection_set(),
2035 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
2036 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
2037 }
2038 }
2039 }
2040 #endif // PRODUCT
2041 void G1ConcurrentMark::create_live_data() {
2042 _g1h->g1_rem_set()->create_card_live_data(_parallel_workers, _nextMarkBitMap);
2043 }
2044
2045 void G1ConcurrentMark::finalize_live_data() {
2046 _g1h->g1_rem_set()->finalize_card_live_data(_g1h->workers(), _nextMarkBitMap);
2047 }
2048
2049 void G1ConcurrentMark::verify_live_data() {
2050 _g1h->g1_rem_set()->verify_card_live_data(_g1h->workers(), _nextMarkBitMap);
2051 }
2052
2053 void G1ConcurrentMark::clear_live_data(WorkGang* workers) {
2054 _g1h->g1_rem_set()->clear_card_live_data(workers);
2055 }
2056
2057 #ifdef ASSERT
2058 void G1ConcurrentMark::verify_live_data_clear() {
2059 _g1h->g1_rem_set()->verify_card_live_data_is_clear();
2060 }
2061 #endif
2062
2063 void G1ConcurrentMark::print_stats() {
2064 if (!log_is_enabled(Debug, gc, stats)) {
2065 return;
2066 }
2067 log_debug(gc, stats)("---------------------------------------------------------------------");
2068 for (size_t i = 0; i < _active_tasks; ++i) {
2069 _tasks[i]->print_stats();
2070 log_debug(gc, stats)("---------------------------------------------------------------------");
2071 }
2072 }
2073
2074 void G1ConcurrentMark::abort() {
2075 if (!cmThread()->during_cycle() || _has_aborted) {
2076 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2077 return;
2078 }
2079
2080 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2081 // concurrent bitmap clearing.
2082 {
2083 GCTraceTime(Debug, gc)("Clear Next Bitmap");
2084 clear_bitmap(_nextMarkBitMap, _g1h->workers(), false);
2085 }
2086 // Note we cannot clear the previous marking bitmap here
2087 // since VerifyDuringGC verifies the objects marked during
2088 // a full GC against the previous bitmap.
2089
2090 {
2091 GCTraceTime(Debug, gc)("Clear Live Data");
2092 clear_live_data(_g1h->workers());
2093 }
2094 DEBUG_ONLY({
2095 GCTraceTime(Debug, gc)("Verify Live Data Clear");
2096 verify_live_data_clear();
2097 })
2098 // Empty mark stack
2099 reset_marking_state();
2100 for (uint i = 0; i < _max_worker_id; ++i) {
2101 _tasks[i]->clear_region_fields();
2102 }
2103 _first_overflow_barrier_sync.abort();
2104 _second_overflow_barrier_sync.abort();
2105 _has_aborted = true;
2106
2107 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2108 satb_mq_set.abandon_partial_marking();
2109 // This can be called either during or outside marking, we'll read
2110 // the expected_active value from the SATB queue set.
2111 satb_mq_set.set_active_all_threads(
2112 false, /* new active value */
2113 satb_mq_set.is_active() /* expected_active */);
2114 }
2115
2116 static void print_ms_time_info(const char* prefix, const char* name,
2117 NumberSeq& ns) {
2118 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2119 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2120 if (ns.num() > 0) {
2121 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]",
2122 prefix, ns.sd(), ns.maximum());
2123 }
2124 }
2125
2126 void G1ConcurrentMark::print_summary_info() {
2127 Log(gc, marking) log;
2128 if (!log.is_trace()) {
2129 return;
2130 }
2131
2132 log.trace(" Concurrent marking:");
2133 print_ms_time_info(" ", "init marks", _init_times);
2134 print_ms_time_info(" ", "remarks", _remark_times);
2135 {
2136 print_ms_time_info(" ", "final marks", _remark_mark_times);
2137 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
2138
2139 }
2140 print_ms_time_info(" ", "cleanups", _cleanup_times);
2141 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2142 _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2143 if (G1ScrubRemSets) {
2144 log.trace(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
2145 _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2146 }
2147 log.trace(" Total stop_world time = %8.2f s.",
2148 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2149 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).",
2150 cmThread()->vtime_accum(), cmThread()->vtime_mark_accum());
2151 }
2152
2153 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2154 _parallel_workers->print_worker_threads_on(st);
2155 }
2156
2157 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2158 _parallel_workers->threads_do(tc);
2159 }
2160
2161 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2162 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2163 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
2164 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
2165 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
2166 }
2167
2168 // Closure for iteration over bitmaps
2169 class G1CMBitMapClosure : public BitMapClosure {
2170 private:
2171 // the bitmap that is being iterated over
2172 G1CMBitMap* _nextMarkBitMap;
2173 G1ConcurrentMark* _cm;
2174 G1CMTask* _task;
2175
2176 public:
2177 G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm, G1CMBitMap* nextMarkBitMap) :
2178 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
2179
2180 bool do_bit(size_t offset) {
2181 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
2182 assert(_nextMarkBitMap->isMarked(addr), "invariant");
2183 assert( addr < _cm->finger(), "invariant");
2184 assert(addr >= _task->finger(), "invariant");
2185
2186 // We move that task's local finger along.
2187 _task->move_finger_to(addr);
2188
2189 _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
2190 // we only partially drain the local queue and global stack
2191 _task->drain_local_queue(true);
2192 _task->drain_global_stack(true);
2193
2194 // if the has_aborted flag has been raised, we need to bail out of
2195 // the iteration
2196 return !_task->has_aborted();
2197 }
2198 };
2199
2200 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2201 ReferenceProcessor* result = g1h->ref_processor_cm();
2202 assert(result != NULL, "CM reference processor should not be NULL");
2203 return result;
2204 }
2205
2206 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2207 G1ConcurrentMark* cm,
2208 G1CMTask* task)
2209 : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2210 _g1h(g1h), _cm(cm), _task(task)
2211 { }
2212
2213 void G1CMTask::setup_for_region(HeapRegion* hr) {
2214 assert(hr != NULL,
2215 "claim_region() should have filtered out NULL regions");
2216 _curr_region = hr;
2217 _finger = hr->bottom();
2218 update_region_limit();
2219 }
2220
2221 void G1CMTask::update_region_limit() {
2222 HeapRegion* hr = _curr_region;
2223 HeapWord* bottom = hr->bottom();
2224 HeapWord* limit = hr->next_top_at_mark_start();
2225
2226 if (limit == bottom) {
2227 // The region was collected underneath our feet.
2228 // We set the finger to bottom to ensure that the bitmap
2229 // iteration that will follow this will not do anything.
2230 // (this is not a condition that holds when we set the region up,
2231 // as the region is not supposed to be empty in the first place)
2232 _finger = bottom;
2233 } else if (limit >= _region_limit) {
2234 assert(limit >= _finger, "peace of mind");
2235 } else {
2236 assert(limit < _region_limit, "only way to get here");
2237 // This can happen under some pretty unusual circumstances. An
2238 // evacuation pause empties the region underneath our feet (NTAMS
2239 // at bottom). We then do some allocation in the region (NTAMS
2240 // stays at bottom), followed by the region being used as a GC
2241 // alloc region (NTAMS will move to top() and the objects
2242 // originally below it will be grayed). All objects now marked in
2243 // the region are explicitly grayed, if below the global finger,
2244 // and we do not need in fact to scan anything else. So, we simply
2245 // set _finger to be limit to ensure that the bitmap iteration
2246 // doesn't do anything.
2247 _finger = limit;
2248 }
2249
2250 _region_limit = limit;
2251 }
2252
2253 void G1CMTask::giveup_current_region() {
2254 assert(_curr_region != NULL, "invariant");
2255 clear_region_fields();
2256 }
2257
2258 void G1CMTask::clear_region_fields() {
2259 // Values for these three fields that indicate that we're not
2260 // holding on to a region.
2261 _curr_region = NULL;
2262 _finger = NULL;
2263 _region_limit = NULL;
2264 }
2265
2266 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2267 if (cm_oop_closure == NULL) {
2268 assert(_cm_oop_closure != NULL, "invariant");
2269 } else {
2270 assert(_cm_oop_closure == NULL, "invariant");
2271 }
2272 _cm_oop_closure = cm_oop_closure;
2273 }
2274
2275 void G1CMTask::reset(G1CMBitMap* nextMarkBitMap) {
2276 guarantee(nextMarkBitMap != NULL, "invariant");
2277 _nextMarkBitMap = nextMarkBitMap;
2278 clear_region_fields();
2279
2280 _calls = 0;
2281 _elapsed_time_ms = 0.0;
2282 _termination_time_ms = 0.0;
2283 _termination_start_time_ms = 0.0;
2284 }
2285
2286 bool G1CMTask::should_exit_termination() {
2287 regular_clock_call();
2288 // This is called when we are in the termination protocol. We should
2289 // quit if, for some reason, this task wants to abort or the global
2290 // stack is not empty (this means that we can get work from it).
2291 return !_cm->mark_stack_empty() || has_aborted();
2292 }
2293
2294 void G1CMTask::reached_limit() {
2295 assert(_words_scanned >= _words_scanned_limit ||
2296 _refs_reached >= _refs_reached_limit ,
2297 "shouldn't have been called otherwise");
2298 regular_clock_call();
2299 }
2300
2301 void G1CMTask::regular_clock_call() {
2302 if (has_aborted()) return;
2303
2304 // First, we need to recalculate the words scanned and refs reached
2305 // limits for the next clock call.
2306 recalculate_limits();
2307
2308 // During the regular clock call we do the following
2309
2310 // (1) If an overflow has been flagged, then we abort.
2311 if (_cm->has_overflown()) {
2312 set_has_aborted();
2313 return;
2314 }
2315
2316 // If we are not concurrent (i.e. we're doing remark) we don't need
2317 // to check anything else. The other steps are only needed during
2318 // the concurrent marking phase.
2319 if (!concurrent()) return;
2320
2321 // (2) If marking has been aborted for Full GC, then we also abort.
2322 if (_cm->has_aborted()) {
2323 set_has_aborted();
2324 return;
2325 }
2326
2327 double curr_time_ms = os::elapsedVTime() * 1000.0;
2328
2329 // (4) We check whether we should yield. If we have to, then we abort.
2330 if (SuspendibleThreadSet::should_yield()) {
2331 // We should yield. To do this we abort the task. The caller is
2332 // responsible for yielding.
2333 set_has_aborted();
2334 return;
2335 }
2336
2337 // (5) We check whether we've reached our time quota. If we have,
2338 // then we abort.
2339 double elapsed_time_ms = curr_time_ms - _start_time_ms;
2340 if (elapsed_time_ms > _time_target_ms) {
2341 set_has_aborted();
2342 _has_timed_out = true;
2343 return;
2344 }
2345
2346 // (6) Finally, we check whether there are enough completed STAB
2347 // buffers available for processing. If there are, we abort.
2348 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2349 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2350 // we do need to process SATB buffers, we'll abort and restart
2351 // the marking task to do so
2352 set_has_aborted();
2353 return;
2354 }
2355 }
2356
2357 void G1CMTask::recalculate_limits() {
2358 _real_words_scanned_limit = _words_scanned + words_scanned_period;
2359 _words_scanned_limit = _real_words_scanned_limit;
2360
2361 _real_refs_reached_limit = _refs_reached + refs_reached_period;
2362 _refs_reached_limit = _real_refs_reached_limit;
2363 }
2364
2365 void G1CMTask::decrease_limits() {
2366 // This is called when we believe that we're going to do an infrequent
2367 // operation which will increase the per byte scanned cost (i.e. move
2368 // entries to/from the global stack). It basically tries to decrease the
2369 // scanning limit so that the clock is called earlier.
2370
2371 _words_scanned_limit = _real_words_scanned_limit -
2372 3 * words_scanned_period / 4;
2373 _refs_reached_limit = _real_refs_reached_limit -
2374 3 * refs_reached_period / 4;
2375 }
2376
2377 void G1CMTask::move_entries_to_global_stack() {
2378 // Local array where we'll store the entries that will be popped
2379 // from the local queue.
2380 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2381
2382 size_t n = 0;
2383 G1TaskQueueEntry task_entry;
2384 while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2385 buffer[n] = task_entry;
2386 ++n;
2387 }
2388 if (n < G1CMMarkStack::EntriesPerChunk) {
2389 buffer[n] = G1TaskQueueEntry();
2390 }
2391
2392 if (n > 0) {
2393 if (!_cm->mark_stack_push(buffer)) {
2394 set_has_aborted();
2395 }
2396 }
2397
2398 // This operation was quite expensive, so decrease the limits.
2399 decrease_limits();
2400 }
2401
2402 bool G1CMTask::get_entries_from_global_stack() {
2403 // Local array where we'll store the entries that will be popped
2404 // from the global stack.
2405 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2406
2407 if (!_cm->mark_stack_pop(buffer)) {
2408 return false;
2409 }
2410
2411 // We did actually pop at least one entry.
2412 for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2413 G1TaskQueueEntry task_entry = buffer[i];
2414 if (task_entry.is_null()) {
2415 break;
2416 }
2417 assert(task_entry.is_array_slice() || task_entry.obj()->is_oop(), "Element " PTR_FORMAT " must be an array slice or oop", p2i(task_entry.obj()));
2418 bool success = _task_queue->push(task_entry);
2419 // We only call this when the local queue is empty or under a
2420 // given target limit. So, we do not expect this push to fail.
2421 assert(success, "invariant");
2422 }
2423
2424 // This operation was quite expensive, so decrease the limits
2425 decrease_limits();
2426 return true;
2427 }
2428
2429 void G1CMTask::drain_local_queue(bool partially) {
2430 if (has_aborted()) {
2431 return;
2432 }
2433
2434 // Decide what the target size is, depending whether we're going to
2435 // drain it partially (so that other tasks can steal if they run out
2436 // of things to do) or totally (at the very end).
2437 size_t target_size;
2438 if (partially) {
2439 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2440 } else {
2441 target_size = 0;
2442 }
2443
2444 if (_task_queue->size() > target_size) {
2445 G1TaskQueueEntry entry;
2446 bool ret = _task_queue->pop_local(entry);
2447 while (ret) {
2448 scan_task_entry(entry);
2449 if (_task_queue->size() <= target_size || has_aborted()) {
2450 ret = false;
2451 } else {
2452 ret = _task_queue->pop_local(entry);
2453 }
2454 }
2455 }
2456 }
2457
2458 void G1CMTask::drain_global_stack(bool partially) {
2459 if (has_aborted()) return;
2460
2461 // We have a policy to drain the local queue before we attempt to
2462 // drain the global stack.
2463 assert(partially || _task_queue->size() == 0, "invariant");
2464
2465 // Decide what the target size is, depending whether we're going to
2466 // drain it partially (so that other tasks can steal if they run out
2467 // of things to do) or totally (at the very end).
2468 // Notice that when draining the global mark stack partially, due to the racyness
2469 // of the mark stack size update we might in fact drop below the target. But,
2470 // this is not a problem.
2471 // In case of total draining, we simply process until the global mark stack is
2472 // totally empty, disregarding the size counter.
2473 if (partially) {
2474 size_t const target_size = _cm->partial_mark_stack_size_target();
2475 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2476 if (get_entries_from_global_stack()) {
2477 drain_local_queue(partially);
2478 }
2479 }
2480 } else {
2481 while (!has_aborted() && get_entries_from_global_stack()) {
2482 drain_local_queue(partially);
2483 }
2484 }
2485 }
2486
2487 // SATB Queue has several assumptions on whether to call the par or
2488 // non-par versions of the methods. this is why some of the code is
2489 // replicated. We should really get rid of the single-threaded version
2490 // of the code to simplify things.
2491 void G1CMTask::drain_satb_buffers() {
2492 if (has_aborted()) return;
2493
2494 // We set this so that the regular clock knows that we're in the
2495 // middle of draining buffers and doesn't set the abort flag when it
2496 // notices that SATB buffers are available for draining. It'd be
2497 // very counter productive if it did that. :-)
2498 _draining_satb_buffers = true;
2499
2500 G1CMSATBBufferClosure satb_cl(this, _g1h);
2501 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2502
2503 // This keeps claiming and applying the closure to completed buffers
2504 // until we run out of buffers or we need to abort.
2505 while (!has_aborted() &&
2506 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2507 regular_clock_call();
2508 }
2509
2510 _draining_satb_buffers = false;
2511
2512 assert(has_aborted() ||
2513 concurrent() ||
2514 satb_mq_set.completed_buffers_num() == 0, "invariant");
2515
2516 // again, this was a potentially expensive operation, decrease the
2517 // limits to get the regular clock call early
2518 decrease_limits();
2519 }
2520
2521 void G1CMTask::print_stats() {
2522 log_debug(gc, stats)("Marking Stats, task = %u, calls = %d",
2523 _worker_id, _calls);
2524 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2525 _elapsed_time_ms, _termination_time_ms);
2526 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
2527 _step_times_ms.num(), _step_times_ms.avg(),
2528 _step_times_ms.sd());
2529 log_debug(gc, stats)(" max = %1.2lfms, total = %1.2lfms",
2530 _step_times_ms.maximum(), _step_times_ms.sum());
2531 }
2532
2533 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry) {
2534 return _task_queues->steal(worker_id, hash_seed, task_entry);
2535 }
2536
2537 /*****************************************************************************
2538
2539 The do_marking_step(time_target_ms, ...) method is the building
2540 block of the parallel marking framework. It can be called in parallel
2541 with other invocations of do_marking_step() on different tasks
2542 (but only one per task, obviously) and concurrently with the
2543 mutator threads, or during remark, hence it eliminates the need
2544 for two versions of the code. When called during remark, it will
2545 pick up from where the task left off during the concurrent marking
2546 phase. Interestingly, tasks are also claimable during evacuation
2547 pauses too, since do_marking_step() ensures that it aborts before
2548 it needs to yield.
2549
2550 The data structures that it uses to do marking work are the
2551 following:
2552
2553 (1) Marking Bitmap. If there are gray objects that appear only
2554 on the bitmap (this happens either when dealing with an overflow
2555 or when the initial marking phase has simply marked the roots
2556 and didn't push them on the stack), then tasks claim heap
2557 regions whose bitmap they then scan to find gray objects. A
2558 global finger indicates where the end of the last claimed region
2559 is. A local finger indicates how far into the region a task has
2560 scanned. The two fingers are used to determine how to gray an
2561 object (i.e. whether simply marking it is OK, as it will be
2562 visited by a task in the future, or whether it needs to be also
2563 pushed on a stack).
2564
2565 (2) Local Queue. The local queue of the task which is accessed
2566 reasonably efficiently by the task. Other tasks can steal from
2567 it when they run out of work. Throughout the marking phase, a
2568 task attempts to keep its local queue short but not totally
2569 empty, so that entries are available for stealing by other
2570 tasks. Only when there is no more work, a task will totally
2571 drain its local queue.
2572
2573 (3) Global Mark Stack. This handles local queue overflow. During
2574 marking only sets of entries are moved between it and the local
2575 queues, as access to it requires a mutex and more fine-grain
2576 interaction with it which might cause contention. If it
2577 overflows, then the marking phase should restart and iterate
2578 over the bitmap to identify gray objects. Throughout the marking
2579 phase, tasks attempt to keep the global mark stack at a small
2580 length but not totally empty, so that entries are available for
2581 popping by other tasks. Only when there is no more work, tasks
2582 will totally drain the global mark stack.
2583
2584 (4) SATB Buffer Queue. This is where completed SATB buffers are
2585 made available. Buffers are regularly removed from this queue
2586 and scanned for roots, so that the queue doesn't get too
2587 long. During remark, all completed buffers are processed, as
2588 well as the filled in parts of any uncompleted buffers.
2589
2590 The do_marking_step() method tries to abort when the time target
2591 has been reached. There are a few other cases when the
2592 do_marking_step() method also aborts:
2593
2594 (1) When the marking phase has been aborted (after a Full GC).
2595
2596 (2) When a global overflow (on the global stack) has been
2597 triggered. Before the task aborts, it will actually sync up with
2598 the other tasks to ensure that all the marking data structures
2599 (local queues, stacks, fingers etc.) are re-initialized so that
2600 when do_marking_step() completes, the marking phase can
2601 immediately restart.
2602
2603 (3) When enough completed SATB buffers are available. The
2604 do_marking_step() method only tries to drain SATB buffers right
2605 at the beginning. So, if enough buffers are available, the
2606 marking step aborts and the SATB buffers are processed at
2607 the beginning of the next invocation.
2608
2609 (4) To yield. when we have to yield then we abort and yield
2610 right at the end of do_marking_step(). This saves us from a lot
2611 of hassle as, by yielding we might allow a Full GC. If this
2612 happens then objects will be compacted underneath our feet, the
2613 heap might shrink, etc. We save checking for this by just
2614 aborting and doing the yield right at the end.
2615
2616 From the above it follows that the do_marking_step() method should
2617 be called in a loop (or, otherwise, regularly) until it completes.
2618
2619 If a marking step completes without its has_aborted() flag being
2620 true, it means it has completed the current marking phase (and
2621 also all other marking tasks have done so and have all synced up).
2622
2623 A method called regular_clock_call() is invoked "regularly" (in
2624 sub ms intervals) throughout marking. It is this clock method that
2625 checks all the abort conditions which were mentioned above and
2626 decides when the task should abort. A work-based scheme is used to
2627 trigger this clock method: when the number of object words the
2628 marking phase has scanned or the number of references the marking
2629 phase has visited reach a given limit. Additional invocations to
2630 the method clock have been planted in a few other strategic places
2631 too. The initial reason for the clock method was to avoid calling
2632 vtime too regularly, as it is quite expensive. So, once it was in
2633 place, it was natural to piggy-back all the other conditions on it
2634 too and not constantly check them throughout the code.
2635
2636 If do_termination is true then do_marking_step will enter its
2637 termination protocol.
2638
2639 The value of is_serial must be true when do_marking_step is being
2640 called serially (i.e. by the VMThread) and do_marking_step should
2641 skip any synchronization in the termination and overflow code.
2642 Examples include the serial remark code and the serial reference
2643 processing closures.
2644
2645 The value of is_serial must be false when do_marking_step is
2646 being called by any of the worker threads in a work gang.
2647 Examples include the concurrent marking code (CMMarkingTask),
2648 the MT remark code, and the MT reference processing closures.
2649
2650 *****************************************************************************/
2651
2652 void G1CMTask::do_marking_step(double time_target_ms,
2653 bool do_termination,
2654 bool is_serial) {
2655 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2656 assert(concurrent() == _cm->concurrent(), "they should be the same");
2657
2658 G1Policy* g1_policy = _g1h->g1_policy();
2659 assert(_task_queues != NULL, "invariant");
2660 assert(_task_queue != NULL, "invariant");
2661 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
2662
2663 assert(!_claimed,
2664 "only one thread should claim this task at any one time");
2665
2666 // OK, this doesn't safeguard again all possible scenarios, as it is
2667 // possible for two threads to set the _claimed flag at the same
2668 // time. But it is only for debugging purposes anyway and it will
2669 // catch most problems.
2670 _claimed = true;
2671
2672 _start_time_ms = os::elapsedVTime() * 1000.0;
2673
2674 // If do_stealing is true then do_marking_step will attempt to
2675 // steal work from the other G1CMTasks. It only makes sense to
2676 // enable stealing when the termination protocol is enabled
2677 // and do_marking_step() is not being called serially.
2678 bool do_stealing = do_termination && !is_serial;
2679
2680 double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2681 _time_target_ms = time_target_ms - diff_prediction_ms;
2682
2683 // set up the variables that are used in the work-based scheme to
2684 // call the regular clock method
2685 _words_scanned = 0;
2686 _refs_reached = 0;
2687 recalculate_limits();
2688
2689 // clear all flags
2690 clear_has_aborted();
2691 _has_timed_out = false;
2692 _draining_satb_buffers = false;
2693
2694 ++_calls;
2695
2696 // Set up the bitmap and oop closures. Anything that uses them is
2697 // eventually called from this method, so it is OK to allocate these
2698 // statically.
2699 G1CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
2700 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
2701 set_cm_oop_closure(&cm_oop_closure);
2702
2703 if (_cm->has_overflown()) {
2704 // This can happen if the mark stack overflows during a GC pause
2705 // and this task, after a yield point, restarts. We have to abort
2706 // as we need to get into the overflow protocol which happens
2707 // right at the end of this task.
2708 set_has_aborted();
2709 }
2710
2711 // First drain any available SATB buffers. After this, we will not
2712 // look at SATB buffers before the next invocation of this method.
2713 // If enough completed SATB buffers are queued up, the regular clock
2714 // will abort this task so that it restarts.
2715 drain_satb_buffers();
2716 // ...then partially drain the local queue and the global stack
2717 drain_local_queue(true);
2718 drain_global_stack(true);
2719
2720 do {
2721 if (!has_aborted() && _curr_region != NULL) {
2722 // This means that we're already holding on to a region.
2723 assert(_finger != NULL, "if region is not NULL, then the finger "
2724 "should not be NULL either");
2725
2726 // We might have restarted this task after an evacuation pause
2727 // which might have evacuated the region we're holding on to
2728 // underneath our feet. Let's read its limit again to make sure
2729 // that we do not iterate over a region of the heap that
2730 // contains garbage (update_region_limit() will also move
2731 // _finger to the start of the region if it is found empty).
2732 update_region_limit();
2733 // We will start from _finger not from the start of the region,
2734 // as we might be restarting this task after aborting half-way
2735 // through scanning this region. In this case, _finger points to
2736 // the address where we last found a marked object. If this is a
2737 // fresh region, _finger points to start().
2738 MemRegion mr = MemRegion(_finger, _region_limit);
2739
2740 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2741 "humongous regions should go around loop once only");
2742
2743 // Some special cases:
2744 // If the memory region is empty, we can just give up the region.
2745 // If the current region is humongous then we only need to check
2746 // the bitmap for the bit associated with the start of the object,
2747 // scan the object if it's live, and give up the region.
2748 // Otherwise, let's iterate over the bitmap of the part of the region
2749 // that is left.
2750 // If the iteration is successful, give up the region.
2751 if (mr.is_empty()) {
2752 giveup_current_region();
2753 regular_clock_call();
2754 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2755 if (_nextMarkBitMap->isMarked(mr.start())) {
2756 // The object is marked - apply the closure
2757 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
2758 bitmap_closure.do_bit(offset);
2759 }
2760 // Even if this task aborted while scanning the humongous object
2761 // we can (and should) give up the current region.
2762 giveup_current_region();
2763 regular_clock_call();
2764 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
2765 giveup_current_region();
2766 regular_clock_call();
2767 } else {
2768 assert(has_aborted(), "currently the only way to do so");
2769 // The only way to abort the bitmap iteration is to return
2770 // false from the do_bit() method. However, inside the
2771 // do_bit() method we move the _finger to point to the
2772 // object currently being looked at. So, if we bail out, we
2773 // have definitely set _finger to something non-null.
2774 assert(_finger != NULL, "invariant");
2775
2776 // Region iteration was actually aborted. So now _finger
2777 // points to the address of the object we last scanned. If we
2778 // leave it there, when we restart this task, we will rescan
2779 // the object. It is easy to avoid this. We move the finger by
2780 // enough to point to the next possible object header (the
2781 // bitmap knows by how much we need to move it as it knows its
2782 // granularity).
2783 assert(_finger < _region_limit, "invariant");
2784 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
2785 // Check if bitmap iteration was aborted while scanning the last object
2786 if (new_finger >= _region_limit) {
2787 giveup_current_region();
2788 } else {
2789 move_finger_to(new_finger);
2790 }
2791 }
2792 }
2793 // At this point we have either completed iterating over the
2794 // region we were holding on to, or we have aborted.
2795
2796 // We then partially drain the local queue and the global stack.
2797 // (Do we really need this?)
2798 drain_local_queue(true);
2799 drain_global_stack(true);
2800
2801 // Read the note on the claim_region() method on why it might
2802 // return NULL with potentially more regions available for
2803 // claiming and why we have to check out_of_regions() to determine
2804 // whether we're done or not.
2805 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2806 // We are going to try to claim a new region. We should have
2807 // given up on the previous one.
2808 // Separated the asserts so that we know which one fires.
2809 assert(_curr_region == NULL, "invariant");
2810 assert(_finger == NULL, "invariant");
2811 assert(_region_limit == NULL, "invariant");
2812 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2813 if (claimed_region != NULL) {
2814 // Yes, we managed to claim one
2815 setup_for_region(claimed_region);
2816 assert(_curr_region == claimed_region, "invariant");
2817 }
2818 // It is important to call the regular clock here. It might take
2819 // a while to claim a region if, for example, we hit a large
2820 // block of empty regions. So we need to call the regular clock
2821 // method once round the loop to make sure it's called
2822 // frequently enough.
2823 regular_clock_call();
2824 }
2825
2826 if (!has_aborted() && _curr_region == NULL) {
2827 assert(_cm->out_of_regions(),
2828 "at this point we should be out of regions");
2829 }
2830 } while ( _curr_region != NULL && !has_aborted());
2831
2832 if (!has_aborted()) {
2833 // We cannot check whether the global stack is empty, since other
2834 // tasks might be pushing objects to it concurrently.
2835 assert(_cm->out_of_regions(),
2836 "at this point we should be out of regions");
2837 // Try to reduce the number of available SATB buffers so that
2838 // remark has less work to do.
2839 drain_satb_buffers();
2840 }
2841
2842 // Since we've done everything else, we can now totally drain the
2843 // local queue and global stack.
2844 drain_local_queue(false);
2845 drain_global_stack(false);
2846
2847 // Attempt at work stealing from other task's queues.
2848 if (do_stealing && !has_aborted()) {
2849 // We have not aborted. This means that we have finished all that
2850 // we could. Let's try to do some stealing...
2851
2852 // We cannot check whether the global stack is empty, since other
2853 // tasks might be pushing objects to it concurrently.
2854 assert(_cm->out_of_regions() && _task_queue->size() == 0,
2855 "only way to reach here");
2856 while (!has_aborted()) {
2857 G1TaskQueueEntry entry;
2858 if (_cm->try_stealing(_worker_id, &_hash_seed, entry)) {
2859 scan_task_entry(entry);
2860
2861 // And since we're towards the end, let's totally drain the
2862 // local queue and global stack.
2863 drain_local_queue(false);
2864 drain_global_stack(false);
2865 } else {
2866 break;
2867 }
2868 }
2869 }
2870
2871 // We still haven't aborted. Now, let's try to get into the
2872 // termination protocol.
2873 if (do_termination && !has_aborted()) {
2874 // We cannot check whether the global stack is empty, since other
2875 // tasks might be concurrently pushing objects on it.
2876 // Separated the asserts so that we know which one fires.
2877 assert(_cm->out_of_regions(), "only way to reach here");
2878 assert(_task_queue->size() == 0, "only way to reach here");
2879 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2880
2881 // The G1CMTask class also extends the TerminatorTerminator class,
2882 // hence its should_exit_termination() method will also decide
2883 // whether to exit the termination protocol or not.
2884 bool finished = (is_serial ||
2885 _cm->terminator()->offer_termination(this));
2886 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2887 _termination_time_ms +=
2888 termination_end_time_ms - _termination_start_time_ms;
2889
2890 if (finished) {
2891 // We're all done.
2892
2893 if (_worker_id == 0) {
2894 // let's allow task 0 to do this
2895 if (concurrent()) {
2896 assert(_cm->concurrent_marking_in_progress(), "invariant");
2897 // we need to set this to false before the next
2898 // safepoint. This way we ensure that the marking phase
2899 // doesn't observe any more heap expansions.
2900 _cm->clear_concurrent_marking_in_progress();
2901 }
2902 }
2903
2904 // We can now guarantee that the global stack is empty, since
2905 // all other tasks have finished. We separated the guarantees so
2906 // that, if a condition is false, we can immediately find out
2907 // which one.
2908 guarantee(_cm->out_of_regions(), "only way to reach here");
2909 guarantee(_cm->mark_stack_empty(), "only way to reach here");
2910 guarantee(_task_queue->size() == 0, "only way to reach here");
2911 guarantee(!_cm->has_overflown(), "only way to reach here");
2912 } else {
2913 // Apparently there's more work to do. Let's abort this task. It
2914 // will restart it and we can hopefully find more things to do.
2915 set_has_aborted();
2916 }
2917 }
2918
2919 // Mainly for debugging purposes to make sure that a pointer to the
2920 // closure which was statically allocated in this frame doesn't
2921 // escape it by accident.
2922 set_cm_oop_closure(NULL);
2923 double end_time_ms = os::elapsedVTime() * 1000.0;
2924 double elapsed_time_ms = end_time_ms - _start_time_ms;
2925 // Update the step history.
2926 _step_times_ms.add(elapsed_time_ms);
2927
2928 if (has_aborted()) {
2929 // The task was aborted for some reason.
2930 if (_has_timed_out) {
2931 double diff_ms = elapsed_time_ms - _time_target_ms;
2932 // Keep statistics of how well we did with respect to hitting
2933 // our target only if we actually timed out (if we aborted for
2934 // other reasons, then the results might get skewed).
2935 _marking_step_diffs_ms.add(diff_ms);
2936 }
2937
2938 if (_cm->has_overflown()) {
2939 // This is the interesting one. We aborted because a global
2940 // overflow was raised. This means we have to restart the
2941 // marking phase and start iterating over regions. However, in
2942 // order to do this we have to make sure that all tasks stop
2943 // what they are doing and re-initialize in a safe manner. We
2944 // will achieve this with the use of two barrier sync points.
2945
2946 if (!is_serial) {
2947 // We only need to enter the sync barrier if being called
2948 // from a parallel context
2949 _cm->enter_first_sync_barrier(_worker_id);
2950
2951 // When we exit this sync barrier we know that all tasks have
2952 // stopped doing marking work. So, it's now safe to
2953 // re-initialize our data structures. At the end of this method,
2954 // task 0 will clear the global data structures.
2955 }
2956
2957 // We clear the local state of this task...
2958 clear_region_fields();
2959
2960 if (!is_serial) {
2961 // ...and enter the second barrier.
2962 _cm->enter_second_sync_barrier(_worker_id);
2963 }
2964 // At this point, if we're during the concurrent phase of
2965 // marking, everything has been re-initialized and we're
2966 // ready to restart.
2967 }
2968 }
2969
2970 _claimed = false;
2971 }
2972
2973 G1CMTask::G1CMTask(uint worker_id,
2974 G1ConcurrentMark* cm,
2975 G1CMTaskQueue* task_queue,
2976 G1CMTaskQueueSet* task_queues)
2977 : _g1h(G1CollectedHeap::heap()),
2978 _worker_id(worker_id), _cm(cm),
2979 _objArray_processor(this),
2980 _claimed(false),
2981 _nextMarkBitMap(NULL), _hash_seed(17),
2982 _task_queue(task_queue),
2983 _task_queues(task_queues),
2984 _cm_oop_closure(NULL) {
2985 guarantee(task_queue != NULL, "invariant");
2986 guarantee(task_queues != NULL, "invariant");
2987
2988 _marking_step_diffs_ms.add(0.5);
2989 }
2990
2991 // These are formatting macros that are used below to ensure
2992 // consistent formatting. The *_H_* versions are used to format the
2993 // header for a particular value and they should be kept consistent
2994 // with the corresponding macro. Also note that most of the macros add
2995 // the necessary white space (as a prefix) which makes them a bit
2996 // easier to compose.
2997
2998 // All the output lines are prefixed with this string to be able to
2999 // identify them easily in a large log file.
3000 #define G1PPRL_LINE_PREFIX "###"
3001
3002 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT
3003 #ifdef _LP64
3004 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
3005 #else // _LP64
3006 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
3007 #endif // _LP64
3008
3009 // For per-region info
3010 #define G1PPRL_TYPE_FORMAT " %-4s"
3011 #define G1PPRL_TYPE_H_FORMAT " %4s"
3012 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9)
3013 #define G1PPRL_BYTE_H_FORMAT " %9s"
3014 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
3015 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
3016
3017 // For summary info
3018 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT
3019 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT
3020 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB"
3021 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
3022
3023 G1PrintRegionLivenessInfoClosure::
3024 G1PrintRegionLivenessInfoClosure(const char* phase_name)
3025 : _total_used_bytes(0), _total_capacity_bytes(0),
3026 _total_prev_live_bytes(0), _total_next_live_bytes(0),
3027 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
3028 G1CollectedHeap* g1h = G1CollectedHeap::heap();
3029 MemRegion g1_reserved = g1h->g1_reserved();
3030 double now = os::elapsedTime();
3031
3032 // Print the header of the output.
3033 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
3034 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
3035 G1PPRL_SUM_ADDR_FORMAT("reserved")
3036 G1PPRL_SUM_BYTE_FORMAT("region-size"),
3037 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
3038 HeapRegion::GrainBytes);
3039 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3040 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3041 G1PPRL_TYPE_H_FORMAT
3042 G1PPRL_ADDR_BASE_H_FORMAT
3043 G1PPRL_BYTE_H_FORMAT
3044 G1PPRL_BYTE_H_FORMAT
3045 G1PPRL_BYTE_H_FORMAT
3046 G1PPRL_DOUBLE_H_FORMAT
3047 G1PPRL_BYTE_H_FORMAT
3048 G1PPRL_BYTE_H_FORMAT,
3049 "type", "address-range",
3050 "used", "prev-live", "next-live", "gc-eff",
3051 "remset", "code-roots");
3052 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3053 G1PPRL_TYPE_H_FORMAT
3054 G1PPRL_ADDR_BASE_H_FORMAT
3055 G1PPRL_BYTE_H_FORMAT
3056 G1PPRL_BYTE_H_FORMAT
3057 G1PPRL_BYTE_H_FORMAT
3058 G1PPRL_DOUBLE_H_FORMAT
3059 G1PPRL_BYTE_H_FORMAT
3060 G1PPRL_BYTE_H_FORMAT,
3061 "", "",
3062 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3063 "(bytes)", "(bytes)");
3064 }
3065
3066 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
3067 const char* type = r->get_type_str();
3068 HeapWord* bottom = r->bottom();
3069 HeapWord* end = r->end();
3070 size_t capacity_bytes = r->capacity();
3071 size_t used_bytes = r->used();
3072 size_t prev_live_bytes = r->live_bytes();
3073 size_t next_live_bytes = r->next_live_bytes();
3074 double gc_eff = r->gc_efficiency();
3075 size_t remset_bytes = r->rem_set()->mem_size();
3076 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3077
3078 _total_used_bytes += used_bytes;
3079 _total_capacity_bytes += capacity_bytes;
3080 _total_prev_live_bytes += prev_live_bytes;
3081 _total_next_live_bytes += next_live_bytes;
3082 _total_remset_bytes += remset_bytes;
3083 _total_strong_code_roots_bytes += strong_code_roots_bytes;
3084
3085 // Print a line for this particular region.
3086 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3087 G1PPRL_TYPE_FORMAT
3088 G1PPRL_ADDR_BASE_FORMAT
3089 G1PPRL_BYTE_FORMAT
3090 G1PPRL_BYTE_FORMAT
3091 G1PPRL_BYTE_FORMAT
3092 G1PPRL_DOUBLE_FORMAT
3093 G1PPRL_BYTE_FORMAT
3094 G1PPRL_BYTE_FORMAT,
3095 type, p2i(bottom), p2i(end),
3096 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3097 remset_bytes, strong_code_roots_bytes);
3098
3099 return false;
3100 }
3101
3102 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3103 // add static memory usages to remembered set sizes
3104 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3105 // Print the footer of the output.
3106 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3107 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3108 " SUMMARY"
3109 G1PPRL_SUM_MB_FORMAT("capacity")
3110 G1PPRL_SUM_MB_PERC_FORMAT("used")
3111 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3112 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3113 G1PPRL_SUM_MB_FORMAT("remset")
3114 G1PPRL_SUM_MB_FORMAT("code-roots"),
3115 bytes_to_mb(_total_capacity_bytes),
3116 bytes_to_mb(_total_used_bytes),
3117 perc(_total_used_bytes, _total_capacity_bytes),
3118 bytes_to_mb(_total_prev_live_bytes),
3119 perc(_total_prev_live_bytes, _total_capacity_bytes),
3120 bytes_to_mb(_total_next_live_bytes),
3121 perc(_total_next_live_bytes, _total_capacity_bytes),
3122 bytes_to_mb(_total_remset_bytes),
3123 bytes_to_mb(_total_strong_code_roots_bytes));
3124 }