1 /*
2 * Copyright (c) 2001, 2018, 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/g1BarrierSet.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/g1ConcurrentMarkThread.inline.hpp"
34 #include "gc/g1/g1HeapVerifier.hpp"
35 #include "gc/g1/g1OopClosures.inline.hpp"
36 #include "gc/g1/g1Policy.hpp"
37 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
38 #include "gc/g1/g1StringDedup.hpp"
39 #include "gc/g1/g1ThreadLocalData.hpp"
40 #include "gc/g1/heapRegion.inline.hpp"
41 #include "gc/g1/heapRegionRemSet.hpp"
42 #include "gc/g1/heapRegionSet.inline.hpp"
43 #include "gc/shared/adaptiveSizePolicy.hpp"
44 #include "gc/shared/gcId.hpp"
45 #include "gc/shared/gcTimer.hpp"
46 #include "gc/shared/gcTrace.hpp"
47 #include "gc/shared/gcTraceTime.inline.hpp"
48 #include "gc/shared/genOopClosures.inline.hpp"
49 #include "gc/shared/referencePolicy.hpp"
50 #include "gc/shared/strongRootsScope.hpp"
51 #include "gc/shared/suspendibleThreadSet.hpp"
52 #include "gc/shared/taskqueue.inline.hpp"
53 #include "gc/shared/vmGCOperations.hpp"
54 #include "gc/shared/weakProcessor.hpp"
55 #include "include/jvm.h"
56 #include "logging/log.hpp"
57 #include "memory/allocation.hpp"
58 #include "memory/resourceArea.hpp"
59 #include "oops/access.inline.hpp"
60 #include "oops/oop.inline.hpp"
61 #include "runtime/atomic.hpp"
62 #include "runtime/handles.inline.hpp"
63 #include "runtime/java.hpp"
64 #include "runtime/prefetch.inline.hpp"
65 #include "services/memTracker.hpp"
66 #include "utilities/align.hpp"
67 #include "utilities/growableArray.hpp"
68
69 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
70 assert(addr < _cm->finger(), "invariant");
71 assert(addr >= _task->finger(), "invariant");
72
73 // We move that task's local finger along.
74 _task->move_finger_to(addr);
75
76 _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
77 // we only partially drain the local queue and global stack
78 _task->drain_local_queue(true);
79 _task->drain_global_stack(true);
80
81 // if the has_aborted flag has been raised, we need to bail out of
82 // the iteration
83 return !_task->has_aborted();
84 }
85
86 G1CMMarkStack::G1CMMarkStack() :
87 _max_chunk_capacity(0),
88 _base(NULL),
89 _chunk_capacity(0) {
90 set_empty();
91 }
92
93 bool G1CMMarkStack::resize(size_t new_capacity) {
94 assert(is_empty(), "Only resize when stack is empty.");
95 assert(new_capacity <= _max_chunk_capacity,
96 "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
97
98 TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
99
100 if (new_base == NULL) {
101 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));
102 return false;
103 }
104 // Release old mapping.
105 if (_base != NULL) {
106 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
107 }
108
109 _base = new_base;
110 _chunk_capacity = new_capacity;
111 set_empty();
112
113 return true;
114 }
115
116 size_t G1CMMarkStack::capacity_alignment() {
117 return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
118 }
119
120 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
121 guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
122
123 size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
124
125 _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
126 size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
127
128 guarantee(initial_chunk_capacity <= _max_chunk_capacity,
129 "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
130 _max_chunk_capacity,
131 initial_chunk_capacity);
132
133 log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
134 initial_chunk_capacity, _max_chunk_capacity);
135
136 return resize(initial_chunk_capacity);
137 }
138
139 void G1CMMarkStack::expand() {
140 if (_chunk_capacity == _max_chunk_capacity) {
141 log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
142 return;
143 }
144 size_t old_capacity = _chunk_capacity;
145 // Double capacity if possible
146 size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
147
148 if (resize(new_capacity)) {
149 log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
150 old_capacity, new_capacity);
151 } else {
152 log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
153 old_capacity, new_capacity);
154 }
155 }
156
157 G1CMMarkStack::~G1CMMarkStack() {
158 if (_base != NULL) {
159 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
160 }
161 }
162
163 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
164 elem->next = *list;
165 *list = elem;
166 }
167
168 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
169 MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
170 add_chunk_to_list(&_chunk_list, elem);
171 _chunks_in_chunk_list++;
172 }
173
174 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
175 MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
176 add_chunk_to_list(&_free_list, elem);
177 }
178
179 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
180 TaskQueueEntryChunk* result = *list;
181 if (result != NULL) {
182 *list = (*list)->next;
183 }
184 return result;
185 }
186
187 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
188 MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
189 TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
190 if (result != NULL) {
191 _chunks_in_chunk_list--;
192 }
193 return result;
194 }
195
196 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
197 MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
198 return remove_chunk_from_list(&_free_list);
199 }
200
201 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
202 // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
203 // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
204 // wraparound of _hwm.
205 if (_hwm >= _chunk_capacity) {
206 return NULL;
207 }
208
209 size_t cur_idx = Atomic::add(1u, &_hwm) - 1;
210 if (cur_idx >= _chunk_capacity) {
211 return NULL;
212 }
213
214 TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
215 result->next = NULL;
216 return result;
217 }
218
219 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
220 // Get a new chunk.
221 TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
222
223 if (new_chunk == NULL) {
224 // Did not get a chunk from the free list. Allocate from backing memory.
225 new_chunk = allocate_new_chunk();
226
227 if (new_chunk == NULL) {
228 return false;
229 }
230 }
231
232 Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
233
234 add_chunk_to_chunk_list(new_chunk);
235
236 return true;
237 }
238
239 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
240 TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
241
242 if (cur == NULL) {
243 return false;
244 }
245
246 Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
247
248 add_chunk_to_free_list(cur);
249 return true;
250 }
251
252 void G1CMMarkStack::set_empty() {
253 _chunks_in_chunk_list = 0;
254 _hwm = 0;
255 _chunk_list = NULL;
256 _free_list = NULL;
257 }
258
259 G1CMRootRegions::G1CMRootRegions() :
260 _survivors(NULL), _cm(NULL), _scan_in_progress(false),
261 _should_abort(false), _claimed_survivor_index(0) { }
262
263 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
264 _survivors = survivors;
265 _cm = cm;
266 }
267
268 void G1CMRootRegions::prepare_for_scan() {
269 assert(!scan_in_progress(), "pre-condition");
270
271 // Currently, only survivors can be root regions.
272 _claimed_survivor_index = 0;
273 _scan_in_progress = _survivors->regions()->is_nonempty();
274 _should_abort = false;
275 }
276
277 HeapRegion* G1CMRootRegions::claim_next() {
278 if (_should_abort) {
279 // If someone has set the should_abort flag, we return NULL to
280 // force the caller to bail out of their loop.
281 return NULL;
282 }
283
284 // Currently, only survivors can be root regions.
285 const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
286
287 int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
288 if (claimed_index < survivor_regions->length()) {
289 return survivor_regions->at(claimed_index);
290 }
291 return NULL;
292 }
293
294 uint G1CMRootRegions::num_root_regions() const {
295 return (uint)_survivors->regions()->length();
296 }
297
298 void G1CMRootRegions::notify_scan_done() {
299 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
300 _scan_in_progress = false;
301 RootRegionScan_lock->notify_all();
302 }
303
304 void G1CMRootRegions::cancel_scan() {
305 notify_scan_done();
306 }
307
308 void G1CMRootRegions::scan_finished() {
309 assert(scan_in_progress(), "pre-condition");
310
311 // Currently, only survivors can be root regions.
312 if (!_should_abort) {
313 assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
314 assert((uint)_claimed_survivor_index >= _survivors->length(),
315 "we should have claimed all survivors, claimed index = %u, length = %u",
316 (uint)_claimed_survivor_index, _survivors->length());
317 }
318
319 notify_scan_done();
320 }
321
322 bool G1CMRootRegions::wait_until_scan_finished() {
323 if (!scan_in_progress()) {
324 return false;
325 }
326
327 {
328 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
329 while (scan_in_progress()) {
330 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
331 }
332 }
333 return true;
334 }
335
336 // Returns the maximum number of workers to be used in a concurrent
337 // phase based on the number of GC workers being used in a STW
338 // phase.
339 static uint scale_concurrent_worker_threads(uint num_gc_workers) {
340 return MAX2((num_gc_workers + 2) / 4, 1U);
341 }
342
343 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
344 G1RegionToSpaceMapper* prev_bitmap_storage,
345 G1RegionToSpaceMapper* next_bitmap_storage) :
346 // _cm_thread set inside the constructor
347 _g1h(g1h),
348 _completed_initialization(false),
349
350 _mark_bitmap_1(),
351 _mark_bitmap_2(),
352 _prev_mark_bitmap(&_mark_bitmap_1),
353 _next_mark_bitmap(&_mark_bitmap_2),
354
355 _heap(_g1h->reserved_region()),
356
357 _root_regions(),
358
359 _global_mark_stack(),
360
361 // _finger set in set_non_marking_state
362
363 _worker_id_offset(DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads),
364 _max_num_tasks(ParallelGCThreads),
365 // _num_active_tasks set in set_non_marking_state()
366 // _tasks set inside the constructor
367
368 _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
369 _terminator(ParallelTaskTerminator((int) _max_num_tasks, _task_queues)),
370
371 _first_overflow_barrier_sync(),
372 _second_overflow_barrier_sync(),
373
374 _has_overflown(false),
375 _concurrent(false),
376 _has_aborted(false),
377 _restart_for_overflow(false),
378 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
379 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
380
381 // _verbose_level set below
382
383 _init_times(),
384 _remark_times(),
385 _remark_mark_times(),
386 _remark_weak_ref_times(),
387 _cleanup_times(),
388 _total_cleanup_time(0.0),
389
390 _accum_task_vtime(NULL),
391
392 _concurrent_workers(NULL),
393 _num_concurrent_workers(0),
394 _max_concurrent_workers(0),
395
396 _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)),
397 _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC))
398 {
399 _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage);
400 _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage);
401
402 // Create & start ConcurrentMark thread.
403 _cm_thread = new G1ConcurrentMarkThread(this);
404 if (_cm_thread->osthread() == NULL) {
405 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
406 }
407
408 assert(CGC_lock != NULL, "CGC_lock must be initialized");
409
410 SATBMarkQueueSet& satb_qs = G1BarrierSet::satb_mark_queue_set();
411 satb_qs.set_buffer_size(G1SATBBufferSize);
412
413 _root_regions.init(_g1h->survivor(), this);
414
415 if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) {
416 // Calculate the number of concurrent worker threads by scaling
417 // the number of parallel GC threads.
418 uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads);
419 FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
420 }
421
422 assert(ConcGCThreads > 0, "ConcGCThreads have been set.");
423 if (ConcGCThreads > ParallelGCThreads) {
424 log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).",
425 ConcGCThreads, ParallelGCThreads);
426 return;
427 }
428
429 log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset);
430 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
431
432 _num_concurrent_workers = ConcGCThreads;
433 _max_concurrent_workers = _num_concurrent_workers;
434
435 _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true);
436 _concurrent_workers->initialize_workers();
437
438 if (FLAG_IS_DEFAULT(MarkStackSize)) {
439 size_t mark_stack_size =
440 MIN2(MarkStackSizeMax,
441 MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE)));
442 // Verify that the calculated value for MarkStackSize is in range.
443 // It would be nice to use the private utility routine from Arguments.
444 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
445 log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
446 "must be between 1 and " SIZE_FORMAT,
447 mark_stack_size, MarkStackSizeMax);
448 return;
449 }
450 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
451 } else {
452 // Verify MarkStackSize is in range.
453 if (FLAG_IS_CMDLINE(MarkStackSize)) {
454 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
455 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
456 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
457 "must be between 1 and " SIZE_FORMAT,
458 MarkStackSize, MarkStackSizeMax);
459 return;
460 }
461 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
462 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
463 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
464 " or for MarkStackSizeMax (" SIZE_FORMAT ")",
465 MarkStackSize, MarkStackSizeMax);
466 return;
467 }
468 }
469 }
470 }
471
472 if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
473 vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
474 }
475
476 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
477 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
478
479 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
480 _num_active_tasks = _max_num_tasks;
481
482 for (uint i = 0; i < _max_num_tasks; ++i) {
483 G1CMTaskQueue* task_queue = new G1CMTaskQueue();
484 task_queue->initialize();
485 _task_queues->register_queue(i, task_queue);
486
487 _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions());
488
489 _accum_task_vtime[i] = 0.0;
490 }
491
492 reset_at_marking_complete();
493 _completed_initialization = true;
494 }
495
496 void G1ConcurrentMark::reset() {
497 _has_aborted = false;
498
499 reset_marking_for_restart();
500
501 // Reset all tasks, since different phases will use different number of active
502 // threads. So, it's easiest to have all of them ready.
503 for (uint i = 0; i < _max_num_tasks; ++i) {
504 _tasks[i]->reset(_next_mark_bitmap);
505 }
506
507 uint max_regions = _g1h->max_regions();
508 for (uint i = 0; i < max_regions; i++) {
509 _top_at_rebuild_starts[i] = NULL;
510 _region_mark_stats[i].clear();
511 }
512 }
513
514 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) {
515 for (uint j = 0; j < _max_num_tasks; ++j) {
516 _tasks[j]->clear_mark_stats_cache(region_idx);
517 }
518 _top_at_rebuild_starts[region_idx] = NULL;
519 _region_mark_stats[region_idx].clear();
520 }
521
522 void G1ConcurrentMark::clear_statistics(HeapRegion* r) {
523 uint const region_idx = r->hrm_index();
524 if (r->is_humongous()) {
525 assert(r->is_starts_humongous(), "Got humongous continues region here");
526 uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size());
527 for (uint j = region_idx; j < (region_idx + size_in_regions); j++) {
528 clear_statistics_in_region(j);
529 }
530 } else {
531 clear_statistics_in_region(region_idx);
532 }
533 }
534
535 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) {
536 if (bitmap->is_marked(addr)) {
537 bitmap->clear(addr);
538 }
539 }
540
541 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) {
542 assert_at_safepoint_on_vm_thread();
543
544 // Need to clear all mark bits of the humongous object.
545 clear_mark_if_set(_prev_mark_bitmap, r->bottom());
546 clear_mark_if_set(_next_mark_bitmap, r->bottom());
547
548 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
549 return;
550 }
551
552 // Clear any statistics about the region gathered so far.
553 clear_statistics(r);
554 }
555
556 void G1ConcurrentMark::reset_marking_for_restart() {
557 _global_mark_stack.set_empty();
558
559 // Expand the marking stack, if we have to and if we can.
560 if (has_overflown()) {
561 _global_mark_stack.expand();
562
563 uint max_regions = _g1h->max_regions();
564 for (uint i = 0; i < max_regions; i++) {
565 _region_mark_stats[i].clear_during_overflow();
566 }
567 }
568
569 clear_has_overflown();
570 _finger = _heap.start();
571
572 for (uint i = 0; i < _max_num_tasks; ++i) {
573 G1CMTaskQueue* queue = _task_queues->queue(i);
574 queue->set_empty();
575 }
576 }
577
578 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
579 assert(active_tasks <= _max_num_tasks, "we should not have more");
580
581 _num_active_tasks = active_tasks;
582 // Need to update the three data structures below according to the
583 // number of active threads for this phase.
584 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
585 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
586 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
587 }
588
589 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
590 set_concurrency(active_tasks);
591
592 _concurrent = concurrent;
593
594 if (!concurrent) {
595 // At this point we should be in a STW phase, and completed marking.
596 assert_at_safepoint_on_vm_thread();
597 assert(out_of_regions(),
598 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
599 p2i(_finger), p2i(_heap.end()));
600 }
601 }
602
603 void G1ConcurrentMark::reset_at_marking_complete() {
604 // We set the global marking state to some default values when we're
605 // not doing marking.
606 reset_marking_for_restart();
607 _num_active_tasks = 0;
608 }
609
610 G1ConcurrentMark::~G1ConcurrentMark() {
611 FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts);
612 FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats);
613 // The G1ConcurrentMark instance is never freed.
614 ShouldNotReachHere();
615 }
616
617 class G1ClearBitMapTask : public AbstractGangTask {
618 public:
619 static size_t chunk_size() { return M; }
620
621 private:
622 // Heap region closure used for clearing the given mark bitmap.
623 class G1ClearBitmapHRClosure : public HeapRegionClosure {
624 private:
625 G1CMBitMap* _bitmap;
626 G1ConcurrentMark* _cm;
627 public:
628 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
629 }
630
631 virtual bool do_heap_region(HeapRegion* r) {
632 size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
633
634 HeapWord* cur = r->bottom();
635 HeapWord* const end = r->end();
636
637 while (cur < end) {
638 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
639 _bitmap->clear_range(mr);
640
641 cur += chunk_size_in_words;
642
643 // Abort iteration if after yielding the marking has been aborted.
644 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
645 return true;
646 }
647 // Repeat the asserts from before the start of the closure. We will do them
648 // as asserts here to minimize their overhead on the product. However, we
649 // will have them as guarantees at the beginning / end of the bitmap
650 // clearing to get some checking in the product.
651 assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant");
652 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant");
653 }
654 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
655
656 return false;
657 }
658 };
659
660 G1ClearBitmapHRClosure _cl;
661 HeapRegionClaimer _hr_claimer;
662 bool _suspendible; // If the task is suspendible, workers must join the STS.
663
664 public:
665 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
666 AbstractGangTask("G1 Clear Bitmap"),
667 _cl(bitmap, suspendible ? cm : NULL),
668 _hr_claimer(n_workers),
669 _suspendible(suspendible)
670 { }
671
672 void work(uint worker_id) {
673 SuspendibleThreadSetJoiner sts_join(_suspendible);
674 G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
675 }
676
677 bool is_complete() {
678 return _cl.is_complete();
679 }
680 };
681
682 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
683 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
684
685 size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
686 size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
687
688 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
689
690 G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
691
692 log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
693 workers->run_task(&cl, num_workers);
694 guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
695 }
696
697 void G1ConcurrentMark::cleanup_for_next_mark() {
698 // Make sure that the concurrent mark thread looks to still be in
699 // the current cycle.
700 guarantee(cm_thread()->during_cycle(), "invariant");
701
702 // We are finishing up the current cycle by clearing the next
703 // marking bitmap and getting it ready for the next cycle. During
704 // this time no other cycle can start. So, let's make sure that this
705 // is the case.
706 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
707
708 clear_bitmap(_next_mark_bitmap, _concurrent_workers, true);
709
710 // Repeat the asserts from above.
711 guarantee(cm_thread()->during_cycle(), "invariant");
712 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
713 }
714
715 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
716 assert_at_safepoint_on_vm_thread();
717 clear_bitmap(_prev_mark_bitmap, workers, false);
718 }
719
720 class CheckBitmapClearHRClosure : public HeapRegionClosure {
721 G1CMBitMap* _bitmap;
722 public:
723 CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
724 }
725
726 virtual bool do_heap_region(HeapRegion* r) {
727 // This closure can be called concurrently to the mutator, so we must make sure
728 // that the result of the getNextMarkedWordAddress() call is compared to the
729 // value passed to it as limit to detect any found bits.
730 // end never changes in G1.
731 HeapWord* end = r->end();
732 return _bitmap->get_next_marked_addr(r->bottom(), end) != end;
733 }
734 };
735
736 bool G1ConcurrentMark::next_mark_bitmap_is_clear() {
737 CheckBitmapClearHRClosure cl(_next_mark_bitmap);
738 _g1h->heap_region_iterate(&cl);
739 return cl.is_complete();
740 }
741
742 class NoteStartOfMarkHRClosure : public HeapRegionClosure {
743 public:
744 bool do_heap_region(HeapRegion* r) {
745 r->note_start_of_marking();
746 return false;
747 }
748 };
749
750 void G1ConcurrentMark::pre_initial_mark() {
751 // Initialize marking structures. This has to be done in a STW phase.
752 reset();
753
754 // For each region note start of marking.
755 NoteStartOfMarkHRClosure startcl;
756 _g1h->heap_region_iterate(&startcl);
757 }
758
759
760 void G1ConcurrentMark::post_initial_mark() {
761 // Start Concurrent Marking weak-reference discovery.
762 ReferenceProcessor* rp = _g1h->ref_processor_cm();
763 // enable ("weak") refs discovery
764 rp->enable_discovery();
765 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
766
767 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
768 // This is the start of the marking cycle, we're expected all
769 // threads to have SATB queues with active set to false.
770 satb_mq_set.set_active_all_threads(true, /* new active value */
771 false /* expected_active */);
772
773 _root_regions.prepare_for_scan();
774
775 // update_g1_committed() will be called at the end of an evac pause
776 // when marking is on. So, it's also called at the end of the
777 // initial-mark pause to update the heap end, if the heap expands
778 // during it. No need to call it here.
779 }
780
781 /*
782 * Notice that in the next two methods, we actually leave the STS
783 * during the barrier sync and join it immediately afterwards. If we
784 * do not do this, the following deadlock can occur: one thread could
785 * be in the barrier sync code, waiting for the other thread to also
786 * sync up, whereas another one could be trying to yield, while also
787 * waiting for the other threads to sync up too.
788 *
789 * Note, however, that this code is also used during remark and in
790 * this case we should not attempt to leave / enter the STS, otherwise
791 * we'll either hit an assert (debug / fastdebug) or deadlock
792 * (product). So we should only leave / enter the STS if we are
793 * operating concurrently.
794 *
795 * Because the thread that does the sync barrier has left the STS, it
796 * is possible to be suspended for a Full GC or an evacuation pause
797 * could occur. This is actually safe, since the entering the sync
798 * barrier is one of the last things do_marking_step() does, and it
799 * doesn't manipulate any data structures afterwards.
800 */
801
802 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
803 bool barrier_aborted;
804 {
805 SuspendibleThreadSetLeaver sts_leave(concurrent());
806 barrier_aborted = !_first_overflow_barrier_sync.enter();
807 }
808
809 // at this point everyone should have synced up and not be doing any
810 // more work
811
812 if (barrier_aborted) {
813 // If the barrier aborted we ignore the overflow condition and
814 // just abort the whole marking phase as quickly as possible.
815 return;
816 }
817 }
818
819 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
820 SuspendibleThreadSetLeaver sts_leave(concurrent());
821 _second_overflow_barrier_sync.enter();
822
823 // at this point everything should be re-initialized and ready to go
824 }
825
826 class G1CMConcurrentMarkingTask : public AbstractGangTask {
827 G1ConcurrentMark* _cm;
828
829 public:
830 void work(uint worker_id) {
831 assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread");
832 ResourceMark rm;
833
834 double start_vtime = os::elapsedVTime();
835
836 {
837 SuspendibleThreadSetJoiner sts_join;
838
839 assert(worker_id < _cm->active_tasks(), "invariant");
840
841 G1CMTask* task = _cm->task(worker_id);
842 task->record_start_time();
843 if (!_cm->has_aborted()) {
844 do {
845 task->do_marking_step(G1ConcMarkStepDurationMillis,
846 true /* do_termination */,
847 false /* is_serial*/);
848
849 _cm->do_yield_check();
850 } while (!_cm->has_aborted() && task->has_aborted());
851 }
852 task->record_end_time();
853 guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
854 }
855
856 double end_vtime = os::elapsedVTime();
857 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
858 }
859
860 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) :
861 AbstractGangTask("Concurrent Mark"), _cm(cm) { }
862
863 ~G1CMConcurrentMarkingTask() { }
864 };
865
866 uint G1ConcurrentMark::calc_active_marking_workers() {
867 uint result = 0;
868 if (!UseDynamicNumberOfGCThreads ||
869 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
870 !ForceDynamicNumberOfGCThreads)) {
871 result = _max_concurrent_workers;
872 } else {
873 result =
874 AdaptiveSizePolicy::calc_default_active_workers(_max_concurrent_workers,
875 1, /* Minimum workers */
876 _num_concurrent_workers,
877 Threads::number_of_non_daemon_threads());
878 // Don't scale the result down by scale_concurrent_workers() because
879 // that scaling has already gone into "_max_concurrent_workers".
880 }
881 assert(result > 0 && result <= _max_concurrent_workers,
882 "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
883 _max_concurrent_workers, result);
884 return result;
885 }
886
887 void G1ConcurrentMark::scan_root_region(HeapRegion* hr, uint worker_id) {
888 // Currently, only survivors can be root regions.
889 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
890 G1RootRegionScanClosure cl(_g1h, this, worker_id);
891
892 const uintx interval = PrefetchScanIntervalInBytes;
893 HeapWord* curr = hr->bottom();
894 const HeapWord* end = hr->top();
895 while (curr < end) {
896 Prefetch::read(curr, interval);
897 oop obj = oop(curr);
898 int size = obj->oop_iterate_size(&cl);
899 assert(size == obj->size(), "sanity");
900 curr += size;
901 }
902 }
903
904 class G1CMRootRegionScanTask : public AbstractGangTask {
905 G1ConcurrentMark* _cm;
906 public:
907 G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
908 AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
909
910 void work(uint worker_id) {
911 assert(Thread::current()->is_ConcurrentGC_thread(),
912 "this should only be done by a conc GC thread");
913
914 G1CMRootRegions* root_regions = _cm->root_regions();
915 HeapRegion* hr = root_regions->claim_next();
916 while (hr != NULL) {
917 _cm->scan_root_region(hr, worker_id);
918 hr = root_regions->claim_next();
919 }
920 }
921 };
922
923 void G1ConcurrentMark::scan_root_regions() {
924 // scan_in_progress() will have been set to true only if there was
925 // at least one root region to scan. So, if it's false, we
926 // should not attempt to do any further work.
927 if (root_regions()->scan_in_progress()) {
928 assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
929
930 _num_concurrent_workers = MIN2(calc_active_marking_workers(),
931 // We distribute work on a per-region basis, so starting
932 // more threads than that is useless.
933 root_regions()->num_root_regions());
934 assert(_num_concurrent_workers <= _max_concurrent_workers,
935 "Maximum number of marking threads exceeded");
936
937 G1CMRootRegionScanTask task(this);
938 log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
939 task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
940 _concurrent_workers->run_task(&task, _num_concurrent_workers);
941
942 // It's possible that has_aborted() is true here without actually
943 // aborting the survivor scan earlier. This is OK as it's
944 // mainly used for sanity checking.
945 root_regions()->scan_finished();
946 }
947 }
948
949 void G1ConcurrentMark::concurrent_cycle_start() {
950 _gc_timer_cm->register_gc_start();
951
952 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
953
954 _g1h->trace_heap_before_gc(_gc_tracer_cm);
955 }
956
957 void G1ConcurrentMark::concurrent_cycle_end() {
958 _g1h->collector_state()->set_clearing_next_bitmap(false);
959
960 _g1h->trace_heap_after_gc(_gc_tracer_cm);
961
962 if (has_aborted()) {
963 log_info(gc, marking)("Concurrent Mark Abort");
964 _gc_tracer_cm->report_concurrent_mode_failure();
965 }
966
967 _gc_timer_cm->register_gc_end();
968
969 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
970 }
971
972 void G1ConcurrentMark::mark_from_roots() {
973 _restart_for_overflow = false;
974
975 _num_concurrent_workers = calc_active_marking_workers();
976
977 uint active_workers = MAX2(1U, _num_concurrent_workers);
978
979 // Setting active workers is not guaranteed since fewer
980 // worker threads may currently exist and more may not be
981 // available.
982 active_workers = _concurrent_workers->update_active_workers(active_workers);
983 log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers());
984
985 // Parallel task terminator is set in "set_concurrency_and_phase()"
986 set_concurrency_and_phase(active_workers, true /* concurrent */);
987
988 G1CMConcurrentMarkingTask marking_task(this);
989 _concurrent_workers->run_task(&marking_task);
990 print_stats();
991 }
992
993 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) {
994 G1HeapVerifier* verifier = _g1h->verifier();
995
996 verifier->verify_region_sets_optional();
997
998 if (VerifyDuringGC) {
999 GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm);
1000
1001 size_t const BufLen = 512;
1002 char buffer[BufLen];
1003
1004 jio_snprintf(buffer, BufLen, "During GC (%s)", caller);
1005 verifier->verify(type, vo, buffer);
1006 }
1007
1008 verifier->check_bitmaps(caller);
1009 }
1010
1011 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask {
1012 G1CollectedHeap* _g1h;
1013 G1ConcurrentMark* _cm;
1014 HeapRegionClaimer _hrclaimer;
1015 uint volatile _total_selected_for_rebuild;
1016
1017 G1PrintRegionLivenessInfoClosure _cl;
1018
1019 class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure {
1020 G1CollectedHeap* _g1h;
1021 G1ConcurrentMark* _cm;
1022
1023 G1PrintRegionLivenessInfoClosure* _cl;
1024
1025 uint _num_regions_selected_for_rebuild; // The number of regions actually selected for rebuild.
1026
1027 void update_remset_before_rebuild(HeapRegion * hr) {
1028 G1RemSetTrackingPolicy* tracking_policy = _g1h->g1_policy()->remset_tracker();
1029
1030 size_t const live_bytes = _cm->liveness(hr->hrm_index()) * HeapWordSize;
1031 bool selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);
1032 if (selected_for_rebuild) {
1033 _num_regions_selected_for_rebuild++;
1034 }
1035 _cm->update_top_at_rebuild_start(hr);
1036 }
1037
1038 // Distribute the given words across the humongous object starting with hr and
1039 // note end of marking.
1040 void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) {
1041 uint const region_idx = hr->hrm_index();
1042 size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size();
1043 uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words);
1044
1045 // "Distributing" zero words means that we only note end of marking for these
1046 // regions.
1047 assert(marked_words == 0 || obj_size_in_words == marked_words,
1048 "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT,
1049 obj_size_in_words, marked_words);
1050
1051 for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) {
1052 HeapRegion* const r = _g1h->region_at(i);
1053 size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words);
1054
1055 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)",
1056 words_to_add, i, r->get_type_str());
1057 add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize);
1058 marked_words -= words_to_add;
1059 }
1060 assert(marked_words == 0,
1061 SIZE_FORMAT " words left after distributing space across %u regions",
1062 marked_words, num_regions_in_humongous);
1063 }
1064
1065 void update_marked_bytes(HeapRegion* hr) {
1066 uint const region_idx = hr->hrm_index();
1067 size_t const marked_words = _cm->liveness(region_idx);
1068 // The marking attributes the object's size completely to the humongous starts
1069 // region. We need to distribute this value across the entire set of regions a
1070 // humongous object spans.
1071 if (hr->is_humongous()) {
1072 assert(hr->is_starts_humongous() || marked_words == 0,
1073 "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)",
1074 marked_words, region_idx, hr->get_type_str());
1075 if (hr->is_starts_humongous()) {
1076 distribute_marked_bytes(hr, marked_words);
1077 }
1078 } else {
1079 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str());
1080 add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize);
1081 }
1082 }
1083
1084 void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) {
1085 hr->add_to_marked_bytes(marked_bytes);
1086 _cl->do_heap_region(hr);
1087 hr->note_end_of_marking();
1088 }
1089
1090 public:
1091 G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) :
1092 _g1h(g1h), _cm(cm), _num_regions_selected_for_rebuild(0), _cl(cl) { }
1093
1094 virtual bool do_heap_region(HeapRegion* r) {
1095 update_remset_before_rebuild(r);
1096 update_marked_bytes(r);
1097
1098 return false;
1099 }
1100
1101 uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; }
1102 };
1103
1104 public:
1105 G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) :
1106 AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"),
1107 _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { }
1108
1109 virtual void work(uint worker_id) {
1110 G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl);
1111 _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id);
1112 Atomic::add(update_cl.num_selected_for_rebuild(), &_total_selected_for_rebuild);
1113 }
1114
1115 uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; }
1116
1117 // Number of regions for which roughly one thread should be spawned for this work.
1118 static const uint RegionsPerThread = 384;
1119 };
1120
1121 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure {
1122 G1CollectedHeap* _g1h;
1123 public:
1124 G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { }
1125
1126 virtual bool do_heap_region(HeapRegion* r) {
1127 _g1h->g1_policy()->remset_tracker()->update_after_rebuild(r);
1128 return false;
1129 }
1130 };
1131
1132 void G1ConcurrentMark::remark() {
1133 assert_at_safepoint_on_vm_thread();
1134
1135 // If a full collection has happened, we should not continue. However we might
1136 // have ended up here as the Remark VM operation has been scheduled already.
1137 if (has_aborted()) {
1138 return;
1139 }
1140
1141 G1Policy* g1p = _g1h->g1_policy();
1142 g1p->record_concurrent_mark_remark_start();
1143
1144 double start = os::elapsedTime();
1145
1146 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before");
1147
1148 {
1149 GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm);
1150 finalize_marking();
1151 }
1152
1153 double mark_work_end = os::elapsedTime();
1154
1155 bool const mark_finished = !has_overflown();
1156 if (mark_finished) {
1157 weak_refs_work(false /* clear_all_soft_refs */);
1158
1159 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1160 // We're done with marking.
1161 // This is the end of the marking cycle, we're expected all
1162 // threads to have SATB queues with active set to true.
1163 satb_mq_set.set_active_all_threads(false, /* new active value */
1164 true /* expected_active */);
1165
1166 {
1167 GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm);
1168 flush_all_task_caches();
1169 }
1170
1171 // Install newly created mark bitmap as "prev".
1172 swap_mark_bitmaps();
1173 {
1174 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm);
1175
1176 uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) /
1177 G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread;
1178 uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity);
1179
1180 G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers);
1181 log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions());
1182 _g1h->workers()->run_task(&cl, num_workers);
1183
1184 log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u",
1185 _g1h->num_regions(), cl.total_selected_for_rebuild());
1186 }
1187 {
1188 GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm);
1189 reclaim_empty_regions();
1190 }
1191
1192 // Clean out dead classes
1193 if (ClassUnloadingWithConcurrentMark) {
1194 GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm);
1195 ClassLoaderDataGraph::purge();
1196 }
1197
1198 compute_new_sizes();
1199
1200 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after");
1201
1202 assert(!restart_for_overflow(), "sanity");
1203 // Completely reset the marking state since marking completed
1204 reset_at_marking_complete();
1205 } else {
1206 // We overflowed. Restart concurrent marking.
1207 _restart_for_overflow = true;
1208
1209 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow");
1210
1211 // Clear the marking state because we will be restarting
1212 // marking due to overflowing the global mark stack.
1213 reset_marking_for_restart();
1214 }
1215
1216 {
1217 GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm);
1218 report_object_count(mark_finished);
1219 }
1220
1221 // Statistics
1222 double now = os::elapsedTime();
1223 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1224 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1225 _remark_times.add((now - start) * 1000.0);
1226
1227 g1p->record_concurrent_mark_remark_end();
1228 }
1229
1230 class G1ReclaimEmptyRegionsTask : public AbstractGangTask {
1231 // Per-region work during the Cleanup pause.
1232 class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure {
1233 G1CollectedHeap* _g1h;
1234 size_t _freed_bytes;
1235 FreeRegionList* _local_cleanup_list;
1236 uint _old_regions_removed;
1237 uint _humongous_regions_removed;
1238 HRRSCleanupTask* _hrrs_cleanup_task;
1239
1240 public:
1241 G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h,
1242 FreeRegionList* local_cleanup_list,
1243 HRRSCleanupTask* hrrs_cleanup_task) :
1244 _g1h(g1h),
1245 _freed_bytes(0),
1246 _local_cleanup_list(local_cleanup_list),
1247 _old_regions_removed(0),
1248 _humongous_regions_removed(0),
1249 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1250
1251 size_t freed_bytes() { return _freed_bytes; }
1252 const uint old_regions_removed() { return _old_regions_removed; }
1253 const uint humongous_regions_removed() { return _humongous_regions_removed; }
1254
1255 bool do_heap_region(HeapRegion *hr) {
1256 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) {
1257 _freed_bytes += hr->used();
1258 hr->set_containing_set(NULL);
1259 if (hr->is_humongous()) {
1260 _humongous_regions_removed++;
1261 _g1h->free_humongous_region(hr, _local_cleanup_list);
1262 } else {
1263 _old_regions_removed++;
1264 _g1h->free_region(hr, _local_cleanup_list, false /* skip_remset */, false /* skip_hcc */, true /* locked */);
1265 }
1266 hr->clear_cardtable();
1267 _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index());
1268 log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom()));
1269 } else {
1270 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1271 }
1272
1273 return false;
1274 }
1275 };
1276
1277 G1CollectedHeap* _g1h;
1278 FreeRegionList* _cleanup_list;
1279 HeapRegionClaimer _hrclaimer;
1280
1281 public:
1282 G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1283 AbstractGangTask("G1 Cleanup"),
1284 _g1h(g1h),
1285 _cleanup_list(cleanup_list),
1286 _hrclaimer(n_workers) {
1287
1288 HeapRegionRemSet::reset_for_cleanup_tasks();
1289 }
1290
1291 void work(uint worker_id) {
1292 FreeRegionList local_cleanup_list("Local Cleanup List");
1293 HRRSCleanupTask hrrs_cleanup_task;
1294 G1ReclaimEmptyRegionsClosure cl(_g1h,
1295 &local_cleanup_list,
1296 &hrrs_cleanup_task);
1297 _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id);
1298 assert(cl.is_complete(), "Shouldn't have aborted!");
1299
1300 // Now update the old/humongous region sets
1301 _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed());
1302 {
1303 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1304 _g1h->decrement_summary_bytes(cl.freed_bytes());
1305
1306 _cleanup_list->add_ordered(&local_cleanup_list);
1307 assert(local_cleanup_list.is_empty(), "post-condition");
1308
1309 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1310 }
1311 }
1312 };
1313
1314 void G1ConcurrentMark::reclaim_empty_regions() {
1315 WorkGang* workers = _g1h->workers();
1316 FreeRegionList empty_regions_list("Empty Regions After Mark List");
1317
1318 G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers());
1319 workers->run_task(&cl);
1320
1321 if (!empty_regions_list.is_empty()) {
1322 log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length());
1323 // Now print the empty regions list.
1324 G1HRPrinter* hrp = _g1h->hr_printer();
1325 if (hrp->is_active()) {
1326 FreeRegionListIterator iter(&empty_regions_list);
1327 while (iter.more_available()) {
1328 HeapRegion* hr = iter.get_next();
1329 hrp->cleanup(hr);
1330 }
1331 }
1332 // And actually make them available.
1333 _g1h->prepend_to_freelist(&empty_regions_list);
1334 }
1335 }
1336
1337 void G1ConcurrentMark::compute_new_sizes() {
1338 MetaspaceGC::compute_new_size();
1339
1340 // Cleanup will have freed any regions completely full of garbage.
1341 // Update the soft reference policy with the new heap occupancy.
1342 Universe::update_heap_info_at_gc();
1343
1344 // We reclaimed old regions so we should calculate the sizes to make
1345 // sure we update the old gen/space data.
1346 _g1h->g1mm()->update_sizes();
1347 }
1348
1349 void G1ConcurrentMark::cleanup() {
1350 assert_at_safepoint_on_vm_thread();
1351
1352 // If a full collection has happened, we shouldn't do this.
1353 if (has_aborted()) {
1354 return;
1355 }
1356
1357 G1Policy* g1p = _g1h->g1_policy();
1358 g1p->record_concurrent_mark_cleanup_start();
1359
1360 double start = os::elapsedTime();
1361
1362 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before");
1363
1364 {
1365 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm);
1366 G1UpdateRemSetTrackingAfterRebuild cl(_g1h);
1367 _g1h->heap_region_iterate(&cl);
1368 }
1369
1370 if (log_is_enabled(Trace, gc, liveness)) {
1371 G1PrintRegionLivenessInfoClosure cl("Post-Cleanup");
1372 _g1h->heap_region_iterate(&cl);
1373 }
1374
1375 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after");
1376
1377 // We need to make this be a "collection" so any collection pause that
1378 // races with it goes around and waits for Cleanup to finish.
1379 _g1h->increment_total_collections();
1380
1381 // Local statistics
1382 double recent_cleanup_time = (os::elapsedTime() - start);
1383 _total_cleanup_time += recent_cleanup_time;
1384 _cleanup_times.add(recent_cleanup_time);
1385
1386 {
1387 GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm);
1388 _g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1389 }
1390 }
1391
1392 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1393 // Uses the G1CMTask associated with a worker thread (for serial reference
1394 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1395 // trace referent objects.
1396 //
1397 // Using the G1CMTask and embedded local queues avoids having the worker
1398 // threads operating on the global mark stack. This reduces the risk
1399 // of overflowing the stack - which we would rather avoid at this late
1400 // state. Also using the tasks' local queues removes the potential
1401 // of the workers interfering with each other that could occur if
1402 // operating on the global stack.
1403
1404 class G1CMKeepAliveAndDrainClosure : public OopClosure {
1405 G1ConcurrentMark* _cm;
1406 G1CMTask* _task;
1407 uint _ref_counter_limit;
1408 uint _ref_counter;
1409 bool _is_serial;
1410 public:
1411 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1412 _cm(cm), _task(task), _is_serial(is_serial),
1413 _ref_counter_limit(G1RefProcDrainInterval) {
1414 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1415 _ref_counter = _ref_counter_limit;
1416 }
1417
1418 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1419 virtual void do_oop( oop* p) { do_oop_work(p); }
1420
1421 template <class T> void do_oop_work(T* p) {
1422 if (_cm->has_overflown()) {
1423 return;
1424 }
1425 if (!_task->deal_with_reference(p)) {
1426 // We did not add anything to the mark bitmap (or mark stack), so there is
1427 // no point trying to drain it.
1428 return;
1429 }
1430 _ref_counter--;
1431
1432 if (_ref_counter == 0) {
1433 // We have dealt with _ref_counter_limit references, pushing them
1434 // and objects reachable from them on to the local stack (and
1435 // possibly the global stack). Call G1CMTask::do_marking_step() to
1436 // process these entries.
1437 //
1438 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1439 // there's nothing more to do (i.e. we're done with the entries that
1440 // were pushed as a result of the G1CMTask::deal_with_reference() calls
1441 // above) or we overflow.
1442 //
1443 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1444 // flag while there may still be some work to do. (See the comment at
1445 // the beginning of G1CMTask::do_marking_step() for those conditions -
1446 // one of which is reaching the specified time target.) It is only
1447 // when G1CMTask::do_marking_step() returns without setting the
1448 // has_aborted() flag that the marking step has completed.
1449 do {
1450 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1451 _task->do_marking_step(mark_step_duration_ms,
1452 false /* do_termination */,
1453 _is_serial);
1454 } while (_task->has_aborted() && !_cm->has_overflown());
1455 _ref_counter = _ref_counter_limit;
1456 }
1457 }
1458 };
1459
1460 // 'Drain' oop closure used by both serial and parallel reference processing.
1461 // Uses the G1CMTask associated with a given worker thread (for serial
1462 // reference processing the G1CMtask for worker 0 is used). Calls the
1463 // do_marking_step routine, with an unbelievably large timeout value,
1464 // to drain the marking data structures of the remaining entries
1465 // added by the 'keep alive' oop closure above.
1466
1467 class G1CMDrainMarkingStackClosure : public VoidClosure {
1468 G1ConcurrentMark* _cm;
1469 G1CMTask* _task;
1470 bool _is_serial;
1471 public:
1472 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1473 _cm(cm), _task(task), _is_serial(is_serial) {
1474 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1475 }
1476
1477 void do_void() {
1478 do {
1479 // We call G1CMTask::do_marking_step() to completely drain the local
1480 // and global marking stacks of entries pushed by the 'keep alive'
1481 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1482 //
1483 // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1484 // if there's nothing more to do (i.e. we've completely drained the
1485 // entries that were pushed as a a result of applying the 'keep alive'
1486 // closure to the entries on the discovered ref lists) or we overflow
1487 // the global marking stack.
1488 //
1489 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1490 // flag while there may still be some work to do. (See the comment at
1491 // the beginning of G1CMTask::do_marking_step() for those conditions -
1492 // one of which is reaching the specified time target.) It is only
1493 // when G1CMTask::do_marking_step() returns without setting the
1494 // has_aborted() flag that the marking step has completed.
1495
1496 _task->do_marking_step(1000000000.0 /* something very large */,
1497 true /* do_termination */,
1498 _is_serial);
1499 } while (_task->has_aborted() && !_cm->has_overflown());
1500 }
1501 };
1502
1503 // Implementation of AbstractRefProcTaskExecutor for parallel
1504 // reference processing at the end of G1 concurrent marking
1505
1506 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor {
1507 private:
1508 G1CollectedHeap* _g1h;
1509 G1ConcurrentMark* _cm;
1510 WorkGang* _workers;
1511 uint _active_workers;
1512
1513 public:
1514 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1515 G1ConcurrentMark* cm,
1516 WorkGang* workers,
1517 uint n_workers) :
1518 _g1h(g1h), _cm(cm),
1519 _workers(workers), _active_workers(n_workers) { }
1520
1521 // Executes the given task using concurrent marking worker threads.
1522 virtual void execute(ProcessTask& task);
1523 virtual void execute(EnqueueTask& task);
1524 };
1525
1526 class G1CMRefProcTaskProxy : public AbstractGangTask {
1527 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1528 ProcessTask& _proc_task;
1529 G1CollectedHeap* _g1h;
1530 G1ConcurrentMark* _cm;
1531
1532 public:
1533 G1CMRefProcTaskProxy(ProcessTask& proc_task,
1534 G1CollectedHeap* g1h,
1535 G1ConcurrentMark* cm) :
1536 AbstractGangTask("Process reference objects in parallel"),
1537 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1538 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1539 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1540 }
1541
1542 virtual void work(uint worker_id) {
1543 ResourceMark rm;
1544 HandleMark hm;
1545 G1CMTask* task = _cm->task(worker_id);
1546 G1CMIsAliveClosure g1_is_alive(_g1h);
1547 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1548 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1549
1550 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1551 }
1552 };
1553
1554 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1555 assert(_workers != NULL, "Need parallel worker threads.");
1556 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1557
1558 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1559
1560 // We need to reset the concurrency level before each
1561 // proxy task execution, so that the termination protocol
1562 // and overflow handling in G1CMTask::do_marking_step() knows
1563 // how many workers to wait for.
1564 _cm->set_concurrency(_active_workers);
1565 _workers->run_task(&proc_task_proxy);
1566 }
1567
1568 class G1CMRefEnqueueTaskProxy : public AbstractGangTask {
1569 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1570 EnqueueTask& _enq_task;
1571
1572 public:
1573 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1574 AbstractGangTask("Enqueue reference objects in parallel"),
1575 _enq_task(enq_task) { }
1576
1577 virtual void work(uint worker_id) {
1578 _enq_task.work(worker_id);
1579 }
1580 };
1581
1582 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1583 assert(_workers != NULL, "Need parallel worker threads.");
1584 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1585
1586 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1587
1588 // Not strictly necessary but...
1589 //
1590 // We need to reset the concurrency level before each
1591 // proxy task execution, so that the termination protocol
1592 // and overflow handling in G1CMTask::do_marking_step() knows
1593 // how many workers to wait for.
1594 _cm->set_concurrency(_active_workers);
1595 _workers->run_task(&enq_task_proxy);
1596 }
1597
1598 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) {
1599 ResourceMark rm;
1600 HandleMark hm;
1601
1602 // Is alive closure.
1603 G1CMIsAliveClosure g1_is_alive(_g1h);
1604
1605 // Inner scope to exclude the cleaning of the string and symbol
1606 // tables from the displayed time.
1607 {
1608 GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm);
1609
1610 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1611
1612 // See the comment in G1CollectedHeap::ref_processing_init()
1613 // about how reference processing currently works in G1.
1614
1615 // Set the soft reference policy
1616 rp->setup_policy(clear_all_soft_refs);
1617 assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1618
1619 // Instances of the 'Keep Alive' and 'Complete GC' closures used
1620 // in serial reference processing. Note these closures are also
1621 // used for serially processing (by the the current thread) the
1622 // JNI references during parallel reference processing.
1623 //
1624 // These closures do not need to synchronize with the worker
1625 // threads involved in parallel reference processing as these
1626 // instances are executed serially by the current thread (e.g.
1627 // reference processing is not multi-threaded and is thus
1628 // performed by the current thread instead of a gang worker).
1629 //
1630 // The gang tasks involved in parallel reference processing create
1631 // their own instances of these closures, which do their own
1632 // synchronization among themselves.
1633 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1634 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1635
1636 // We need at least one active thread. If reference processing
1637 // is not multi-threaded we use the current (VMThread) thread,
1638 // otherwise we use the work gang from the G1CollectedHeap and
1639 // we utilize all the worker threads we can.
1640 bool processing_is_mt = rp->processing_is_mt();
1641 uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U);
1642 active_workers = MAX2(MIN2(active_workers, _max_num_tasks), 1U);
1643
1644 // Parallel processing task executor.
1645 G1CMRefProcTaskExecutor par_task_executor(_g1h, this,
1646 _g1h->workers(), active_workers);
1647 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1648
1649 // Set the concurrency level. The phase was already set prior to
1650 // executing the remark task.
1651 set_concurrency(active_workers);
1652
1653 // Set the degree of MT processing here. If the discovery was done MT,
1654 // the number of threads involved during discovery could differ from
1655 // the number of active workers. This is OK as long as the discovered
1656 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1657 rp->set_active_mt_degree(active_workers);
1658
1659 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_queues());
1660
1661 // Process the weak references.
1662 const ReferenceProcessorStats& stats =
1663 rp->process_discovered_references(&g1_is_alive,
1664 &g1_keep_alive,
1665 &g1_drain_mark_stack,
1666 executor,
1667 &pt);
1668 _gc_tracer_cm->report_gc_reference_stats(stats);
1669 pt.print_all_references();
1670
1671 // The do_oop work routines of the keep_alive and drain_marking_stack
1672 // oop closures will set the has_overflown flag if we overflow the
1673 // global marking stack.
1674
1675 assert(has_overflown() || _global_mark_stack.is_empty(),
1676 "Mark stack should be empty (unless it has overflown)");
1677
1678 assert(rp->num_queues() == active_workers, "why not");
1679
1680 rp->verify_no_references_recorded();
1681 assert(!rp->discovery_enabled(), "Post condition");
1682 }
1683
1684 if (has_overflown()) {
1685 // We can not trust g1_is_alive if the marking stack overflowed
1686 return;
1687 }
1688
1689 assert(_global_mark_stack.is_empty(), "Marking should have completed");
1690
1691 {
1692 GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1693 WeakProcessor::weak_oops_do(&g1_is_alive, &do_nothing_cl);
1694 }
1695
1696 // Unload Klasses, String, Symbols, Code Cache, etc.
1697 if (ClassUnloadingWithConcurrentMark) {
1698 GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1699 bool purged_classes = SystemDictionary::do_unloading(&g1_is_alive, _gc_timer_cm, false /* Defer cleaning */);
1700 _g1h->complete_cleaning(&g1_is_alive, purged_classes);
1701 } else {
1702 GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm);
1703 // No need to clean string table and symbol table as they are treated as strong roots when
1704 // class unloading is disabled.
1705 _g1h->partial_cleaning(&g1_is_alive, false, false, G1StringDedup::is_enabled());
1706 }
1707 }
1708
1709 class G1PrecleanYieldClosure : public YieldClosure {
1710 G1ConcurrentMark* _cm;
1711
1712 public:
1713 G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { }
1714
1715 virtual bool should_return() {
1716 return _cm->has_aborted();
1717 }
1718
1719 virtual bool should_return_fine_grain() {
1720 _cm->do_yield_check();
1721 return _cm->has_aborted();
1722 }
1723 };
1724
1725 void G1ConcurrentMark::preclean() {
1726 assert(G1UseReferencePrecleaning, "Precleaning must be enabled.");
1727
1728 SuspendibleThreadSetJoiner joiner;
1729
1730 G1CMKeepAliveAndDrainClosure keep_alive(this, task(0), true /* is_serial */);
1731 G1CMDrainMarkingStackClosure drain_mark_stack(this, task(0), true /* is_serial */);
1732
1733 set_concurrency_and_phase(1, true);
1734
1735 G1PrecleanYieldClosure yield_cl(this);
1736
1737 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1738 // Precleaning is single threaded. Temporarily disable MT discovery.
1739 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
1740 rp->preclean_discovered_references(rp->is_alive_non_header(),
1741 &keep_alive,
1742 &drain_mark_stack,
1743 &yield_cl,
1744 _gc_timer_cm);
1745 }
1746
1747 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
1748 // the prev bitmap determining liveness.
1749 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {
1750 G1CollectedHeap* _g1h;
1751 public:
1752 G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
1753
1754 bool do_object_b(oop obj) {
1755 HeapWord* addr = (HeapWord*)obj;
1756 return addr != NULL &&
1757 (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_dead(obj));
1758 }
1759 };
1760
1761 void G1ConcurrentMark::report_object_count(bool mark_completed) {
1762 // Depending on the completion of the marking liveness needs to be determined
1763 // using either the next or prev bitmap.
1764 if (mark_completed) {
1765 G1ObjectCountIsAliveClosure is_alive(_g1h);
1766 _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1767 } else {
1768 G1CMIsAliveClosure is_alive(_g1h);
1769 _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1770 }
1771 }
1772
1773
1774 void G1ConcurrentMark::swap_mark_bitmaps() {
1775 G1CMBitMap* temp = _prev_mark_bitmap;
1776 _prev_mark_bitmap = _next_mark_bitmap;
1777 _next_mark_bitmap = temp;
1778 _g1h->collector_state()->set_clearing_next_bitmap(true);
1779 }
1780
1781 // Closure for marking entries in SATB buffers.
1782 class G1CMSATBBufferClosure : public SATBBufferClosure {
1783 private:
1784 G1CMTask* _task;
1785 G1CollectedHeap* _g1h;
1786
1787 // This is very similar to G1CMTask::deal_with_reference, but with
1788 // more relaxed requirements for the argument, so this must be more
1789 // circumspect about treating the argument as an object.
1790 void do_entry(void* entry) const {
1791 _task->increment_refs_reached();
1792 oop const obj = static_cast<oop>(entry);
1793 _task->make_reference_grey(obj);
1794 }
1795
1796 public:
1797 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1798 : _task(task), _g1h(g1h) { }
1799
1800 virtual void do_buffer(void** buffer, size_t size) {
1801 for (size_t i = 0; i < size; ++i) {
1802 do_entry(buffer[i]);
1803 }
1804 }
1805 };
1806
1807 class G1RemarkThreadsClosure : public ThreadClosure {
1808 G1CMSATBBufferClosure _cm_satb_cl;
1809 G1CMOopClosure _cm_cl;
1810 MarkingCodeBlobClosure _code_cl;
1811 int _thread_parity;
1812
1813 public:
1814 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1815 _cm_satb_cl(task, g1h),
1816 _cm_cl(g1h, task),
1817 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1818 _thread_parity(Threads::thread_claim_parity()) {}
1819
1820 void do_thread(Thread* thread) {
1821 if (thread->is_Java_thread()) {
1822 if (thread->claim_oops_do(true, _thread_parity)) {
1823 JavaThread* jt = (JavaThread*)thread;
1824
1825 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1826 // however the liveness of oops reachable from nmethods have very complex lifecycles:
1827 // * Alive if on the stack of an executing method
1828 // * Weakly reachable otherwise
1829 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1830 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1831 jt->nmethods_do(&_code_cl);
1832
1833 G1ThreadLocalData::satb_mark_queue(jt).apply_closure_and_empty(&_cm_satb_cl);
1834 }
1835 } else if (thread->is_VM_thread()) {
1836 if (thread->claim_oops_do(true, _thread_parity)) {
1837 G1BarrierSet::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1838 }
1839 }
1840 }
1841 };
1842
1843 class G1CMRemarkTask : public AbstractGangTask {
1844 G1ConcurrentMark* _cm;
1845 public:
1846 void work(uint worker_id) {
1847 G1CMTask* task = _cm->task(worker_id);
1848 task->record_start_time();
1849 {
1850 ResourceMark rm;
1851 HandleMark hm;
1852
1853 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1854 Threads::threads_do(&threads_f);
1855 }
1856
1857 do {
1858 task->do_marking_step(1000000000.0 /* something very large */,
1859 true /* do_termination */,
1860 false /* is_serial */);
1861 } while (task->has_aborted() && !_cm->has_overflown());
1862 // If we overflow, then we do not want to restart. We instead
1863 // want to abort remark and do concurrent marking again.
1864 task->record_end_time();
1865 }
1866
1867 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1868 AbstractGangTask("Par Remark"), _cm(cm) {
1869 _cm->terminator()->reset_for_reuse(active_workers);
1870 }
1871 };
1872
1873 void G1ConcurrentMark::finalize_marking() {
1874 ResourceMark rm;
1875 HandleMark hm;
1876
1877 _g1h->ensure_parsability(false);
1878
1879 // this is remark, so we'll use up all active threads
1880 uint active_workers = _g1h->workers()->active_workers();
1881 set_concurrency_and_phase(active_workers, false /* concurrent */);
1882 // Leave _parallel_marking_threads at it's
1883 // value originally calculated in the G1ConcurrentMark
1884 // constructor and pass values of the active workers
1885 // through the gang in the task.
1886
1887 {
1888 StrongRootsScope srs(active_workers);
1889
1890 G1CMRemarkTask remarkTask(this, active_workers);
1891 // We will start all available threads, even if we decide that the
1892 // active_workers will be fewer. The extra ones will just bail out
1893 // immediately.
1894 _g1h->workers()->run_task(&remarkTask);
1895 }
1896
1897 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1898 guarantee(has_overflown() ||
1899 satb_mq_set.completed_buffers_num() == 0,
1900 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1901 BOOL_TO_STR(has_overflown()),
1902 satb_mq_set.completed_buffers_num());
1903
1904 print_stats();
1905 }
1906
1907 void G1ConcurrentMark::flush_all_task_caches() {
1908 size_t hits = 0;
1909 size_t misses = 0;
1910 for (uint i = 0; i < _max_num_tasks; i++) {
1911 Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache();
1912 hits += stats.first;
1913 misses += stats.second;
1914 }
1915 size_t sum = hits + misses;
1916 log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf",
1917 hits, misses, percent_of(hits, sum));
1918 }
1919
1920 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1921 _prev_mark_bitmap->clear_range(mr);
1922 }
1923
1924 HeapRegion*
1925 G1ConcurrentMark::claim_region(uint worker_id) {
1926 // "checkpoint" the finger
1927 HeapWord* finger = _finger;
1928
1929 while (finger < _heap.end()) {
1930 assert(_g1h->is_in_g1_reserved(finger), "invariant");
1931
1932 HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1933 // Make sure that the reads below do not float before loading curr_region.
1934 OrderAccess::loadload();
1935 // Above heap_region_containing may return NULL as we always scan claim
1936 // until the end of the heap. In this case, just jump to the next region.
1937 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1938
1939 // Is the gap between reading the finger and doing the CAS too long?
1940 HeapWord* res = Atomic::cmpxchg(end, &_finger, finger);
1941 if (res == finger && curr_region != NULL) {
1942 // we succeeded
1943 HeapWord* bottom = curr_region->bottom();
1944 HeapWord* limit = curr_region->next_top_at_mark_start();
1945
1946 // notice that _finger == end cannot be guaranteed here since,
1947 // someone else might have moved the finger even further
1948 assert(_finger >= end, "the finger should have moved forward");
1949
1950 if (limit > bottom) {
1951 return curr_region;
1952 } else {
1953 assert(limit == bottom,
1954 "the region limit should be at bottom");
1955 // we return NULL and the caller should try calling
1956 // claim_region() again.
1957 return NULL;
1958 }
1959 } else {
1960 assert(_finger > finger, "the finger should have moved forward");
1961 // read it again
1962 finger = _finger;
1963 }
1964 }
1965
1966 return NULL;
1967 }
1968
1969 #ifndef PRODUCT
1970 class VerifyNoCSetOops {
1971 G1CollectedHeap* _g1h;
1972 const char* _phase;
1973 int _info;
1974
1975 public:
1976 VerifyNoCSetOops(const char* phase, int info = -1) :
1977 _g1h(G1CollectedHeap::heap()),
1978 _phase(phase),
1979 _info(info)
1980 { }
1981
1982 void operator()(G1TaskQueueEntry task_entry) const {
1983 if (task_entry.is_array_slice()) {
1984 guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1985 return;
1986 }
1987 guarantee(oopDesc::is_oop(task_entry.obj()),
1988 "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1989 p2i(task_entry.obj()), _phase, _info);
1990 guarantee(!_g1h->is_in_cset(task_entry.obj()),
1991 "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1992 p2i(task_entry.obj()), _phase, _info);
1993 }
1994 };
1995
1996 void G1ConcurrentMark::verify_no_cset_oops() {
1997 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1998 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
1999 return;
2000 }
2001
2002 // Verify entries on the global mark stack
2003 _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
2004
2005 // Verify entries on the task queues
2006 for (uint i = 0; i < _max_num_tasks; ++i) {
2007 G1CMTaskQueue* queue = _task_queues->queue(i);
2008 queue->iterate(VerifyNoCSetOops("Queue", i));
2009 }
2010
2011 // Verify the global finger
2012 HeapWord* global_finger = finger();
2013 if (global_finger != NULL && global_finger < _heap.end()) {
2014 // Since we always iterate over all regions, we might get a NULL HeapRegion
2015 // here.
2016 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
2017 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
2018 "global finger: " PTR_FORMAT " region: " HR_FORMAT,
2019 p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
2020 }
2021
2022 // Verify the task fingers
2023 assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
2024 for (uint i = 0; i < _num_concurrent_workers; ++i) {
2025 G1CMTask* task = _tasks[i];
2026 HeapWord* task_finger = task->finger();
2027 if (task_finger != NULL && task_finger < _heap.end()) {
2028 // See above note on the global finger verification.
2029 HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
2030 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
2031 !task_hr->in_collection_set(),
2032 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
2033 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
2034 }
2035 }
2036 }
2037 #endif // PRODUCT
2038
2039 void G1ConcurrentMark::rebuild_rem_set_concurrently() {
2040 _g1h->g1_rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset);
2041 }
2042
2043 void G1ConcurrentMark::print_stats() {
2044 if (!log_is_enabled(Debug, gc, stats)) {
2045 return;
2046 }
2047 log_debug(gc, stats)("---------------------------------------------------------------------");
2048 for (size_t i = 0; i < _num_active_tasks; ++i) {
2049 _tasks[i]->print_stats();
2050 log_debug(gc, stats)("---------------------------------------------------------------------");
2051 }
2052 }
2053
2054 void G1ConcurrentMark::concurrent_cycle_abort() {
2055 if (!cm_thread()->during_cycle() || _has_aborted) {
2056 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2057 return;
2058 }
2059
2060 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2061 // concurrent bitmap clearing.
2062 {
2063 GCTraceTime(Debug, gc) debug("Clear Next Bitmap");
2064 clear_bitmap(_next_mark_bitmap, _g1h->workers(), false);
2065 }
2066 // Note we cannot clear the previous marking bitmap here
2067 // since VerifyDuringGC verifies the objects marked during
2068 // a full GC against the previous bitmap.
2069
2070 // Empty mark stack
2071 reset_marking_for_restart();
2072 for (uint i = 0; i < _max_num_tasks; ++i) {
2073 _tasks[i]->clear_region_fields();
2074 }
2075 _first_overflow_barrier_sync.abort();
2076 _second_overflow_barrier_sync.abort();
2077 _has_aborted = true;
2078
2079 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2080 satb_mq_set.abandon_partial_marking();
2081 // This can be called either during or outside marking, we'll read
2082 // the expected_active value from the SATB queue set.
2083 satb_mq_set.set_active_all_threads(
2084 false, /* new active value */
2085 satb_mq_set.is_active() /* expected_active */);
2086 }
2087
2088 static void print_ms_time_info(const char* prefix, const char* name,
2089 NumberSeq& ns) {
2090 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2091 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2092 if (ns.num() > 0) {
2093 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]",
2094 prefix, ns.sd(), ns.maximum());
2095 }
2096 }
2097
2098 void G1ConcurrentMark::print_summary_info() {
2099 Log(gc, marking) log;
2100 if (!log.is_trace()) {
2101 return;
2102 }
2103
2104 log.trace(" Concurrent marking:");
2105 print_ms_time_info(" ", "init marks", _init_times);
2106 print_ms_time_info(" ", "remarks", _remark_times);
2107 {
2108 print_ms_time_info(" ", "final marks", _remark_mark_times);
2109 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
2110
2111 }
2112 print_ms_time_info(" ", "cleanups", _cleanup_times);
2113 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2114 _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2115 log.trace(" Total stop_world time = %8.2f s.",
2116 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2117 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).",
2118 cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2119 }
2120
2121 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2122 _concurrent_workers->print_worker_threads_on(st);
2123 }
2124
2125 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2126 _concurrent_workers->threads_do(tc);
2127 }
2128
2129 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2130 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2131 p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2132 _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2133 _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2134 }
2135
2136 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2137 ReferenceProcessor* result = g1h->ref_processor_cm();
2138 assert(result != NULL, "CM reference processor should not be NULL");
2139 return result;
2140 }
2141
2142 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2143 G1CMTask* task)
2144 : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2145 _g1h(g1h), _task(task)
2146 { }
2147
2148 void G1CMTask::setup_for_region(HeapRegion* hr) {
2149 assert(hr != NULL,
2150 "claim_region() should have filtered out NULL regions");
2151 _curr_region = hr;
2152 _finger = hr->bottom();
2153 update_region_limit();
2154 }
2155
2156 void G1CMTask::update_region_limit() {
2157 HeapRegion* hr = _curr_region;
2158 HeapWord* bottom = hr->bottom();
2159 HeapWord* limit = hr->next_top_at_mark_start();
2160
2161 if (limit == bottom) {
2162 // The region was collected underneath our feet.
2163 // We set the finger to bottom to ensure that the bitmap
2164 // iteration that will follow this will not do anything.
2165 // (this is not a condition that holds when we set the region up,
2166 // as the region is not supposed to be empty in the first place)
2167 _finger = bottom;
2168 } else if (limit >= _region_limit) {
2169 assert(limit >= _finger, "peace of mind");
2170 } else {
2171 assert(limit < _region_limit, "only way to get here");
2172 // This can happen under some pretty unusual circumstances. An
2173 // evacuation pause empties the region underneath our feet (NTAMS
2174 // at bottom). We then do some allocation in the region (NTAMS
2175 // stays at bottom), followed by the region being used as a GC
2176 // alloc region (NTAMS will move to top() and the objects
2177 // originally below it will be grayed). All objects now marked in
2178 // the region are explicitly grayed, if below the global finger,
2179 // and we do not need in fact to scan anything else. So, we simply
2180 // set _finger to be limit to ensure that the bitmap iteration
2181 // doesn't do anything.
2182 _finger = limit;
2183 }
2184
2185 _region_limit = limit;
2186 }
2187
2188 void G1CMTask::giveup_current_region() {
2189 assert(_curr_region != NULL, "invariant");
2190 clear_region_fields();
2191 }
2192
2193 void G1CMTask::clear_region_fields() {
2194 // Values for these three fields that indicate that we're not
2195 // holding on to a region.
2196 _curr_region = NULL;
2197 _finger = NULL;
2198 _region_limit = NULL;
2199 }
2200
2201 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2202 if (cm_oop_closure == NULL) {
2203 assert(_cm_oop_closure != NULL, "invariant");
2204 } else {
2205 assert(_cm_oop_closure == NULL, "invariant");
2206 }
2207 _cm_oop_closure = cm_oop_closure;
2208 }
2209
2210 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2211 guarantee(next_mark_bitmap != NULL, "invariant");
2212 _next_mark_bitmap = next_mark_bitmap;
2213 clear_region_fields();
2214
2215 _calls = 0;
2216 _elapsed_time_ms = 0.0;
2217 _termination_time_ms = 0.0;
2218 _termination_start_time_ms = 0.0;
2219
2220 _mark_stats_cache.reset();
2221 }
2222
2223 bool G1CMTask::should_exit_termination() {
2224 regular_clock_call();
2225 // This is called when we are in the termination protocol. We should
2226 // quit if, for some reason, this task wants to abort or the global
2227 // stack is not empty (this means that we can get work from it).
2228 return !_cm->mark_stack_empty() || has_aborted();
2229 }
2230
2231 void G1CMTask::reached_limit() {
2232 assert(_words_scanned >= _words_scanned_limit ||
2233 _refs_reached >= _refs_reached_limit ,
2234 "shouldn't have been called otherwise");
2235 regular_clock_call();
2236 }
2237
2238 void G1CMTask::regular_clock_call() {
2239 if (has_aborted()) {
2240 return;
2241 }
2242
2243 // First, we need to recalculate the words scanned and refs reached
2244 // limits for the next clock call.
2245 recalculate_limits();
2246
2247 // During the regular clock call we do the following
2248
2249 // (1) If an overflow has been flagged, then we abort.
2250 if (_cm->has_overflown()) {
2251 set_has_aborted();
2252 return;
2253 }
2254
2255 // If we are not concurrent (i.e. we're doing remark) we don't need
2256 // to check anything else. The other steps are only needed during
2257 // the concurrent marking phase.
2258 if (!_cm->concurrent()) {
2259 return;
2260 }
2261
2262 // (2) If marking has been aborted for Full GC, then we also abort.
2263 if (_cm->has_aborted()) {
2264 set_has_aborted();
2265 return;
2266 }
2267
2268 double curr_time_ms = os::elapsedVTime() * 1000.0;
2269
2270 // (4) We check whether we should yield. If we have to, then we abort.
2271 if (SuspendibleThreadSet::should_yield()) {
2272 // We should yield. To do this we abort the task. The caller is
2273 // responsible for yielding.
2274 set_has_aborted();
2275 return;
2276 }
2277
2278 // (5) We check whether we've reached our time quota. If we have,
2279 // then we abort.
2280 double elapsed_time_ms = curr_time_ms - _start_time_ms;
2281 if (elapsed_time_ms > _time_target_ms) {
2282 set_has_aborted();
2283 _has_timed_out = true;
2284 return;
2285 }
2286
2287 // (6) Finally, we check whether there are enough completed STAB
2288 // buffers available for processing. If there are, we abort.
2289 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2290 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2291 // we do need to process SATB buffers, we'll abort and restart
2292 // the marking task to do so
2293 set_has_aborted();
2294 return;
2295 }
2296 }
2297
2298 void G1CMTask::recalculate_limits() {
2299 _real_words_scanned_limit = _words_scanned + words_scanned_period;
2300 _words_scanned_limit = _real_words_scanned_limit;
2301
2302 _real_refs_reached_limit = _refs_reached + refs_reached_period;
2303 _refs_reached_limit = _real_refs_reached_limit;
2304 }
2305
2306 void G1CMTask::decrease_limits() {
2307 // This is called when we believe that we're going to do an infrequent
2308 // operation which will increase the per byte scanned cost (i.e. move
2309 // entries to/from the global stack). It basically tries to decrease the
2310 // scanning limit so that the clock is called earlier.
2311
2312 _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2313 _refs_reached_limit = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2314 }
2315
2316 void G1CMTask::move_entries_to_global_stack() {
2317 // Local array where we'll store the entries that will be popped
2318 // from the local queue.
2319 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2320
2321 size_t n = 0;
2322 G1TaskQueueEntry task_entry;
2323 while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2324 buffer[n] = task_entry;
2325 ++n;
2326 }
2327 if (n < G1CMMarkStack::EntriesPerChunk) {
2328 buffer[n] = G1TaskQueueEntry();
2329 }
2330
2331 if (n > 0) {
2332 if (!_cm->mark_stack_push(buffer)) {
2333 set_has_aborted();
2334 }
2335 }
2336
2337 // This operation was quite expensive, so decrease the limits.
2338 decrease_limits();
2339 }
2340
2341 bool G1CMTask::get_entries_from_global_stack() {
2342 // Local array where we'll store the entries that will be popped
2343 // from the global stack.
2344 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2345
2346 if (!_cm->mark_stack_pop(buffer)) {
2347 return false;
2348 }
2349
2350 // We did actually pop at least one entry.
2351 for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2352 G1TaskQueueEntry task_entry = buffer[i];
2353 if (task_entry.is_null()) {
2354 break;
2355 }
2356 assert(task_entry.is_array_slice() || oopDesc::is_oop(task_entry.obj()), "Element " PTR_FORMAT " must be an array slice or oop", p2i(task_entry.obj()));
2357 bool success = _task_queue->push(task_entry);
2358 // We only call this when the local queue is empty or under a
2359 // given target limit. So, we do not expect this push to fail.
2360 assert(success, "invariant");
2361 }
2362
2363 // This operation was quite expensive, so decrease the limits
2364 decrease_limits();
2365 return true;
2366 }
2367
2368 void G1CMTask::drain_local_queue(bool partially) {
2369 if (has_aborted()) {
2370 return;
2371 }
2372
2373 // Decide what the target size is, depending whether we're going to
2374 // drain it partially (so that other tasks can steal if they run out
2375 // of things to do) or totally (at the very end).
2376 size_t target_size;
2377 if (partially) {
2378 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2379 } else {
2380 target_size = 0;
2381 }
2382
2383 if (_task_queue->size() > target_size) {
2384 G1TaskQueueEntry entry;
2385 bool ret = _task_queue->pop_local(entry);
2386 while (ret) {
2387 scan_task_entry(entry);
2388 if (_task_queue->size() <= target_size || has_aborted()) {
2389 ret = false;
2390 } else {
2391 ret = _task_queue->pop_local(entry);
2392 }
2393 }
2394 }
2395 }
2396
2397 void G1CMTask::drain_global_stack(bool partially) {
2398 if (has_aborted()) {
2399 return;
2400 }
2401
2402 // We have a policy to drain the local queue before we attempt to
2403 // drain the global stack.
2404 assert(partially || _task_queue->size() == 0, "invariant");
2405
2406 // Decide what the target size is, depending whether we're going to
2407 // drain it partially (so that other tasks can steal if they run out
2408 // of things to do) or totally (at the very end).
2409 // Notice that when draining the global mark stack partially, due to the racyness
2410 // of the mark stack size update we might in fact drop below the target. But,
2411 // this is not a problem.
2412 // In case of total draining, we simply process until the global mark stack is
2413 // totally empty, disregarding the size counter.
2414 if (partially) {
2415 size_t const target_size = _cm->partial_mark_stack_size_target();
2416 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2417 if (get_entries_from_global_stack()) {
2418 drain_local_queue(partially);
2419 }
2420 }
2421 } else {
2422 while (!has_aborted() && get_entries_from_global_stack()) {
2423 drain_local_queue(partially);
2424 }
2425 }
2426 }
2427
2428 // SATB Queue has several assumptions on whether to call the par or
2429 // non-par versions of the methods. this is why some of the code is
2430 // replicated. We should really get rid of the single-threaded version
2431 // of the code to simplify things.
2432 void G1CMTask::drain_satb_buffers() {
2433 if (has_aborted()) {
2434 return;
2435 }
2436
2437 // We set this so that the regular clock knows that we're in the
2438 // middle of draining buffers and doesn't set the abort flag when it
2439 // notices that SATB buffers are available for draining. It'd be
2440 // very counter productive if it did that. :-)
2441 _draining_satb_buffers = true;
2442
2443 G1CMSATBBufferClosure satb_cl(this, _g1h);
2444 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2445
2446 // This keeps claiming and applying the closure to completed buffers
2447 // until we run out of buffers or we need to abort.
2448 while (!has_aborted() &&
2449 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2450 regular_clock_call();
2451 }
2452
2453 _draining_satb_buffers = false;
2454
2455 assert(has_aborted() ||
2456 _cm->concurrent() ||
2457 satb_mq_set.completed_buffers_num() == 0, "invariant");
2458
2459 // again, this was a potentially expensive operation, decrease the
2460 // limits to get the regular clock call early
2461 decrease_limits();
2462 }
2463
2464 void G1CMTask::clear_mark_stats_cache(uint region_idx) {
2465 _mark_stats_cache.reset(region_idx);
2466 }
2467
2468 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() {
2469 return _mark_stats_cache.evict_all();
2470 }
2471
2472 void G1CMTask::print_stats() {
2473 log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls);
2474 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2475 _elapsed_time_ms, _termination_time_ms);
2476 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms",
2477 _step_times_ms.num(),
2478 _step_times_ms.avg(),
2479 _step_times_ms.sd(),
2480 _step_times_ms.maximum(),
2481 _step_times_ms.sum());
2482 size_t const hits = _mark_stats_cache.hits();
2483 size_t const misses = _mark_stats_cache.misses();
2484 log_debug(gc, stats)(" Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f",
2485 hits, misses, percent_of(hits, hits + misses));
2486 }
2487
2488 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry) {
2489 return _task_queues->steal(worker_id, hash_seed, task_entry);
2490 }
2491
2492 /*****************************************************************************
2493
2494 The do_marking_step(time_target_ms, ...) method is the building
2495 block of the parallel marking framework. It can be called in parallel
2496 with other invocations of do_marking_step() on different tasks
2497 (but only one per task, obviously) and concurrently with the
2498 mutator threads, or during remark, hence it eliminates the need
2499 for two versions of the code. When called during remark, it will
2500 pick up from where the task left off during the concurrent marking
2501 phase. Interestingly, tasks are also claimable during evacuation
2502 pauses too, since do_marking_step() ensures that it aborts before
2503 it needs to yield.
2504
2505 The data structures that it uses to do marking work are the
2506 following:
2507
2508 (1) Marking Bitmap. If there are gray objects that appear only
2509 on the bitmap (this happens either when dealing with an overflow
2510 or when the initial marking phase has simply marked the roots
2511 and didn't push them on the stack), then tasks claim heap
2512 regions whose bitmap they then scan to find gray objects. A
2513 global finger indicates where the end of the last claimed region
2514 is. A local finger indicates how far into the region a task has
2515 scanned. The two fingers are used to determine how to gray an
2516 object (i.e. whether simply marking it is OK, as it will be
2517 visited by a task in the future, or whether it needs to be also
2518 pushed on a stack).
2519
2520 (2) Local Queue. The local queue of the task which is accessed
2521 reasonably efficiently by the task. Other tasks can steal from
2522 it when they run out of work. Throughout the marking phase, a
2523 task attempts to keep its local queue short but not totally
2524 empty, so that entries are available for stealing by other
2525 tasks. Only when there is no more work, a task will totally
2526 drain its local queue.
2527
2528 (3) Global Mark Stack. This handles local queue overflow. During
2529 marking only sets of entries are moved between it and the local
2530 queues, as access to it requires a mutex and more fine-grain
2531 interaction with it which might cause contention. If it
2532 overflows, then the marking phase should restart and iterate
2533 over the bitmap to identify gray objects. Throughout the marking
2534 phase, tasks attempt to keep the global mark stack at a small
2535 length but not totally empty, so that entries are available for
2536 popping by other tasks. Only when there is no more work, tasks
2537 will totally drain the global mark stack.
2538
2539 (4) SATB Buffer Queue. This is where completed SATB buffers are
2540 made available. Buffers are regularly removed from this queue
2541 and scanned for roots, so that the queue doesn't get too
2542 long. During remark, all completed buffers are processed, as
2543 well as the filled in parts of any uncompleted buffers.
2544
2545 The do_marking_step() method tries to abort when the time target
2546 has been reached. There are a few other cases when the
2547 do_marking_step() method also aborts:
2548
2549 (1) When the marking phase has been aborted (after a Full GC).
2550
2551 (2) When a global overflow (on the global stack) has been
2552 triggered. Before the task aborts, it will actually sync up with
2553 the other tasks to ensure that all the marking data structures
2554 (local queues, stacks, fingers etc.) are re-initialized so that
2555 when do_marking_step() completes, the marking phase can
2556 immediately restart.
2557
2558 (3) When enough completed SATB buffers are available. The
2559 do_marking_step() method only tries to drain SATB buffers right
2560 at the beginning. So, if enough buffers are available, the
2561 marking step aborts and the SATB buffers are processed at
2562 the beginning of the next invocation.
2563
2564 (4) To yield. when we have to yield then we abort and yield
2565 right at the end of do_marking_step(). This saves us from a lot
2566 of hassle as, by yielding we might allow a Full GC. If this
2567 happens then objects will be compacted underneath our feet, the
2568 heap might shrink, etc. We save checking for this by just
2569 aborting and doing the yield right at the end.
2570
2571 From the above it follows that the do_marking_step() method should
2572 be called in a loop (or, otherwise, regularly) until it completes.
2573
2574 If a marking step completes without its has_aborted() flag being
2575 true, it means it has completed the current marking phase (and
2576 also all other marking tasks have done so and have all synced up).
2577
2578 A method called regular_clock_call() is invoked "regularly" (in
2579 sub ms intervals) throughout marking. It is this clock method that
2580 checks all the abort conditions which were mentioned above and
2581 decides when the task should abort. A work-based scheme is used to
2582 trigger this clock method: when the number of object words the
2583 marking phase has scanned or the number of references the marking
2584 phase has visited reach a given limit. Additional invocations to
2585 the method clock have been planted in a few other strategic places
2586 too. The initial reason for the clock method was to avoid calling
2587 vtime too regularly, as it is quite expensive. So, once it was in
2588 place, it was natural to piggy-back all the other conditions on it
2589 too and not constantly check them throughout the code.
2590
2591 If do_termination is true then do_marking_step will enter its
2592 termination protocol.
2593
2594 The value of is_serial must be true when do_marking_step is being
2595 called serially (i.e. by the VMThread) and do_marking_step should
2596 skip any synchronization in the termination and overflow code.
2597 Examples include the serial remark code and the serial reference
2598 processing closures.
2599
2600 The value of is_serial must be false when do_marking_step is
2601 being called by any of the worker threads in a work gang.
2602 Examples include the concurrent marking code (CMMarkingTask),
2603 the MT remark code, and the MT reference processing closures.
2604
2605 *****************************************************************************/
2606
2607 void G1CMTask::do_marking_step(double time_target_ms,
2608 bool do_termination,
2609 bool is_serial) {
2610 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2611
2612 _start_time_ms = os::elapsedVTime() * 1000.0;
2613
2614 // If do_stealing is true then do_marking_step will attempt to
2615 // steal work from the other G1CMTasks. It only makes sense to
2616 // enable stealing when the termination protocol is enabled
2617 // and do_marking_step() is not being called serially.
2618 bool do_stealing = do_termination && !is_serial;
2619
2620 double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2621 _time_target_ms = time_target_ms - diff_prediction_ms;
2622
2623 // set up the variables that are used in the work-based scheme to
2624 // call the regular clock method
2625 _words_scanned = 0;
2626 _refs_reached = 0;
2627 recalculate_limits();
2628
2629 // clear all flags
2630 clear_has_aborted();
2631 _has_timed_out = false;
2632 _draining_satb_buffers = false;
2633
2634 ++_calls;
2635
2636 // Set up the bitmap and oop closures. Anything that uses them is
2637 // eventually called from this method, so it is OK to allocate these
2638 // statically.
2639 G1CMBitMapClosure bitmap_closure(this, _cm);
2640 G1CMOopClosure cm_oop_closure(_g1h, this);
2641 set_cm_oop_closure(&cm_oop_closure);
2642
2643 if (_cm->has_overflown()) {
2644 // This can happen if the mark stack overflows during a GC pause
2645 // and this task, after a yield point, restarts. We have to abort
2646 // as we need to get into the overflow protocol which happens
2647 // right at the end of this task.
2648 set_has_aborted();
2649 }
2650
2651 // First drain any available SATB buffers. After this, we will not
2652 // look at SATB buffers before the next invocation of this method.
2653 // If enough completed SATB buffers are queued up, the regular clock
2654 // will abort this task so that it restarts.
2655 drain_satb_buffers();
2656 // ...then partially drain the local queue and the global stack
2657 drain_local_queue(true);
2658 drain_global_stack(true);
2659
2660 do {
2661 if (!has_aborted() && _curr_region != NULL) {
2662 // This means that we're already holding on to a region.
2663 assert(_finger != NULL, "if region is not NULL, then the finger "
2664 "should not be NULL either");
2665
2666 // We might have restarted this task after an evacuation pause
2667 // which might have evacuated the region we're holding on to
2668 // underneath our feet. Let's read its limit again to make sure
2669 // that we do not iterate over a region of the heap that
2670 // contains garbage (update_region_limit() will also move
2671 // _finger to the start of the region if it is found empty).
2672 update_region_limit();
2673 // We will start from _finger not from the start of the region,
2674 // as we might be restarting this task after aborting half-way
2675 // through scanning this region. In this case, _finger points to
2676 // the address where we last found a marked object. If this is a
2677 // fresh region, _finger points to start().
2678 MemRegion mr = MemRegion(_finger, _region_limit);
2679
2680 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2681 "humongous regions should go around loop once only");
2682
2683 // Some special cases:
2684 // If the memory region is empty, we can just give up the region.
2685 // If the current region is humongous then we only need to check
2686 // the bitmap for the bit associated with the start of the object,
2687 // scan the object if it's live, and give up the region.
2688 // Otherwise, let's iterate over the bitmap of the part of the region
2689 // that is left.
2690 // If the iteration is successful, give up the region.
2691 if (mr.is_empty()) {
2692 giveup_current_region();
2693 regular_clock_call();
2694 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2695 if (_next_mark_bitmap->is_marked(mr.start())) {
2696 // The object is marked - apply the closure
2697 bitmap_closure.do_addr(mr.start());
2698 }
2699 // Even if this task aborted while scanning the humongous object
2700 // we can (and should) give up the current region.
2701 giveup_current_region();
2702 regular_clock_call();
2703 } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2704 giveup_current_region();
2705 regular_clock_call();
2706 } else {
2707 assert(has_aborted(), "currently the only way to do so");
2708 // The only way to abort the bitmap iteration is to return
2709 // false from the do_bit() method. However, inside the
2710 // do_bit() method we move the _finger to point to the
2711 // object currently being looked at. So, if we bail out, we
2712 // have definitely set _finger to something non-null.
2713 assert(_finger != NULL, "invariant");
2714
2715 // Region iteration was actually aborted. So now _finger
2716 // points to the address of the object we last scanned. If we
2717 // leave it there, when we restart this task, we will rescan
2718 // the object. It is easy to avoid this. We move the finger by
2719 // enough to point to the next possible object header.
2720 assert(_finger < _region_limit, "invariant");
2721 HeapWord* const new_finger = _finger + ((oop)_finger)->size();
2722 // Check if bitmap iteration was aborted while scanning the last object
2723 if (new_finger >= _region_limit) {
2724 giveup_current_region();
2725 } else {
2726 move_finger_to(new_finger);
2727 }
2728 }
2729 }
2730 // At this point we have either completed iterating over the
2731 // region we were holding on to, or we have aborted.
2732
2733 // We then partially drain the local queue and the global stack.
2734 // (Do we really need this?)
2735 drain_local_queue(true);
2736 drain_global_stack(true);
2737
2738 // Read the note on the claim_region() method on why it might
2739 // return NULL with potentially more regions available for
2740 // claiming and why we have to check out_of_regions() to determine
2741 // whether we're done or not.
2742 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2743 // We are going to try to claim a new region. We should have
2744 // given up on the previous one.
2745 // Separated the asserts so that we know which one fires.
2746 assert(_curr_region == NULL, "invariant");
2747 assert(_finger == NULL, "invariant");
2748 assert(_region_limit == NULL, "invariant");
2749 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2750 if (claimed_region != NULL) {
2751 // Yes, we managed to claim one
2752 setup_for_region(claimed_region);
2753 assert(_curr_region == claimed_region, "invariant");
2754 }
2755 // It is important to call the regular clock here. It might take
2756 // a while to claim a region if, for example, we hit a large
2757 // block of empty regions. So we need to call the regular clock
2758 // method once round the loop to make sure it's called
2759 // frequently enough.
2760 regular_clock_call();
2761 }
2762
2763 if (!has_aborted() && _curr_region == NULL) {
2764 assert(_cm->out_of_regions(),
2765 "at this point we should be out of regions");
2766 }
2767 } while ( _curr_region != NULL && !has_aborted());
2768
2769 if (!has_aborted()) {
2770 // We cannot check whether the global stack is empty, since other
2771 // tasks might be pushing objects to it concurrently.
2772 assert(_cm->out_of_regions(),
2773 "at this point we should be out of regions");
2774 // Try to reduce the number of available SATB buffers so that
2775 // remark has less work to do.
2776 drain_satb_buffers();
2777 }
2778
2779 // Since we've done everything else, we can now totally drain the
2780 // local queue and global stack.
2781 drain_local_queue(false);
2782 drain_global_stack(false);
2783
2784 // Attempt at work stealing from other task's queues.
2785 if (do_stealing && !has_aborted()) {
2786 // We have not aborted. This means that we have finished all that
2787 // we could. Let's try to do some stealing...
2788
2789 // We cannot check whether the global stack is empty, since other
2790 // tasks might be pushing objects to it concurrently.
2791 assert(_cm->out_of_regions() && _task_queue->size() == 0,
2792 "only way to reach here");
2793 while (!has_aborted()) {
2794 G1TaskQueueEntry entry;
2795 if (_cm->try_stealing(_worker_id, &_hash_seed, entry)) {
2796 scan_task_entry(entry);
2797
2798 // And since we're towards the end, let's totally drain the
2799 // local queue and global stack.
2800 drain_local_queue(false);
2801 drain_global_stack(false);
2802 } else {
2803 break;
2804 }
2805 }
2806 }
2807
2808 // We still haven't aborted. Now, let's try to get into the
2809 // termination protocol.
2810 if (do_termination && !has_aborted()) {
2811 // We cannot check whether the global stack is empty, since other
2812 // tasks might be concurrently pushing objects on it.
2813 // Separated the asserts so that we know which one fires.
2814 assert(_cm->out_of_regions(), "only way to reach here");
2815 assert(_task_queue->size() == 0, "only way to reach here");
2816 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2817
2818 // The G1CMTask class also extends the TerminatorTerminator class,
2819 // hence its should_exit_termination() method will also decide
2820 // whether to exit the termination protocol or not.
2821 bool finished = (is_serial ||
2822 _cm->terminator()->offer_termination(this));
2823 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2824 _termination_time_ms +=
2825 termination_end_time_ms - _termination_start_time_ms;
2826
2827 if (finished) {
2828 // We're all done.
2829
2830 // We can now guarantee that the global stack is empty, since
2831 // all other tasks have finished. We separated the guarantees so
2832 // that, if a condition is false, we can immediately find out
2833 // which one.
2834 guarantee(_cm->out_of_regions(), "only way to reach here");
2835 guarantee(_cm->mark_stack_empty(), "only way to reach here");
2836 guarantee(_task_queue->size() == 0, "only way to reach here");
2837 guarantee(!_cm->has_overflown(), "only way to reach here");
2838 } else {
2839 // Apparently there's more work to do. Let's abort this task. It
2840 // will restart it and we can hopefully find more things to do.
2841 set_has_aborted();
2842 }
2843 }
2844
2845 // Mainly for debugging purposes to make sure that a pointer to the
2846 // closure which was statically allocated in this frame doesn't
2847 // escape it by accident.
2848 set_cm_oop_closure(NULL);
2849 double end_time_ms = os::elapsedVTime() * 1000.0;
2850 double elapsed_time_ms = end_time_ms - _start_time_ms;
2851 // Update the step history.
2852 _step_times_ms.add(elapsed_time_ms);
2853
2854 if (has_aborted()) {
2855 // The task was aborted for some reason.
2856 if (_has_timed_out) {
2857 double diff_ms = elapsed_time_ms - _time_target_ms;
2858 // Keep statistics of how well we did with respect to hitting
2859 // our target only if we actually timed out (if we aborted for
2860 // other reasons, then the results might get skewed).
2861 _marking_step_diffs_ms.add(diff_ms);
2862 }
2863
2864 if (_cm->has_overflown()) {
2865 // This is the interesting one. We aborted because a global
2866 // overflow was raised. This means we have to restart the
2867 // marking phase and start iterating over regions. However, in
2868 // order to do this we have to make sure that all tasks stop
2869 // what they are doing and re-initialize in a safe manner. We
2870 // will achieve this with the use of two barrier sync points.
2871
2872 if (!is_serial) {
2873 // We only need to enter the sync barrier if being called
2874 // from a parallel context
2875 _cm->enter_first_sync_barrier(_worker_id);
2876
2877 // When we exit this sync barrier we know that all tasks have
2878 // stopped doing marking work. So, it's now safe to
2879 // re-initialize our data structures.
2880 }
2881
2882 clear_region_fields();
2883 flush_mark_stats_cache();
2884
2885 if (!is_serial) {
2886 // If we're executing the concurrent phase of marking, reset the marking
2887 // state; otherwise the marking state is reset after reference processing,
2888 // during the remark pause.
2889 // If we reset here as a result of an overflow during the remark we will
2890 // see assertion failures from any subsequent set_concurrency_and_phase()
2891 // calls.
2892 if (_cm->concurrent() && _worker_id == 0) {
2893 // Worker 0 is responsible for clearing the global data structures because
2894 // of an overflow. During STW we should not clear the overflow flag (in
2895 // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit
2896 // method to abort the pause and restart concurrent marking.
2897 _cm->reset_marking_for_restart();
2898
2899 log_info(gc, marking)("Concurrent Mark reset for overflow");
2900 }
2901
2902 // ...and enter the second barrier.
2903 _cm->enter_second_sync_barrier(_worker_id);
2904 }
2905 // At this point, if we're during the concurrent phase of
2906 // marking, everything has been re-initialized and we're
2907 // ready to restart.
2908 }
2909 }
2910 }
2911
2912 G1CMTask::G1CMTask(uint worker_id,
2913 G1ConcurrentMark* cm,
2914 G1CMTaskQueue* task_queue,
2915 G1RegionMarkStats* mark_stats,
2916 uint max_regions) :
2917 _objArray_processor(this),
2918 _worker_id(worker_id),
2919 _g1h(G1CollectedHeap::heap()),
2920 _cm(cm),
2921 _next_mark_bitmap(NULL),
2922 _task_queue(task_queue),
2923 _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize),
2924 _calls(0),
2925 _time_target_ms(0.0),
2926 _start_time_ms(0.0),
2927 _cm_oop_closure(NULL),
2928 _curr_region(NULL),
2929 _finger(NULL),
2930 _region_limit(NULL),
2931 _words_scanned(0),
2932 _words_scanned_limit(0),
2933 _real_words_scanned_limit(0),
2934 _refs_reached(0),
2935 _refs_reached_limit(0),
2936 _real_refs_reached_limit(0),
2937 _hash_seed(17),
2938 _has_aborted(false),
2939 _has_timed_out(false),
2940 _draining_satb_buffers(false),
2941 _step_times_ms(),
2942 _elapsed_time_ms(0.0),
2943 _termination_time_ms(0.0),
2944 _termination_start_time_ms(0.0),
2945 _marking_step_diffs_ms()
2946 {
2947 guarantee(task_queue != NULL, "invariant");
2948
2949 _marking_step_diffs_ms.add(0.5);
2950 }
2951
2952 // These are formatting macros that are used below to ensure
2953 // consistent formatting. The *_H_* versions are used to format the
2954 // header for a particular value and they should be kept consistent
2955 // with the corresponding macro. Also note that most of the macros add
2956 // the necessary white space (as a prefix) which makes them a bit
2957 // easier to compose.
2958
2959 // All the output lines are prefixed with this string to be able to
2960 // identify them easily in a large log file.
2961 #define G1PPRL_LINE_PREFIX "###"
2962
2963 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT
2964 #ifdef _LP64
2965 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
2966 #else // _LP64
2967 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
2968 #endif // _LP64
2969
2970 // For per-region info
2971 #define G1PPRL_TYPE_FORMAT " %-4s"
2972 #define G1PPRL_TYPE_H_FORMAT " %4s"
2973 #define G1PPRL_STATE_FORMAT " %-5s"
2974 #define G1PPRL_STATE_H_FORMAT " %5s"
2975 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9)
2976 #define G1PPRL_BYTE_H_FORMAT " %9s"
2977 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
2978 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
2979
2980 // For summary info
2981 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT
2982 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT
2983 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB"
2984 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2985
2986 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2987 _total_used_bytes(0), _total_capacity_bytes(0),
2988 _total_prev_live_bytes(0), _total_next_live_bytes(0),
2989 _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2990 {
2991 if (!log_is_enabled(Trace, gc, liveness)) {
2992 return;
2993 }
2994
2995 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2996 MemRegion g1_reserved = g1h->g1_reserved();
2997 double now = os::elapsedTime();
2998
2999 // Print the header of the output.
3000 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
3001 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
3002 G1PPRL_SUM_ADDR_FORMAT("reserved")
3003 G1PPRL_SUM_BYTE_FORMAT("region-size"),
3004 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
3005 HeapRegion::GrainBytes);
3006 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3007 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3008 G1PPRL_TYPE_H_FORMAT
3009 G1PPRL_ADDR_BASE_H_FORMAT
3010 G1PPRL_BYTE_H_FORMAT
3011 G1PPRL_BYTE_H_FORMAT
3012 G1PPRL_BYTE_H_FORMAT
3013 G1PPRL_DOUBLE_H_FORMAT
3014 G1PPRL_BYTE_H_FORMAT
3015 G1PPRL_STATE_H_FORMAT
3016 G1PPRL_BYTE_H_FORMAT,
3017 "type", "address-range",
3018 "used", "prev-live", "next-live", "gc-eff",
3019 "remset", "state", "code-roots");
3020 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3021 G1PPRL_TYPE_H_FORMAT
3022 G1PPRL_ADDR_BASE_H_FORMAT
3023 G1PPRL_BYTE_H_FORMAT
3024 G1PPRL_BYTE_H_FORMAT
3025 G1PPRL_BYTE_H_FORMAT
3026 G1PPRL_DOUBLE_H_FORMAT
3027 G1PPRL_BYTE_H_FORMAT
3028 G1PPRL_STATE_H_FORMAT
3029 G1PPRL_BYTE_H_FORMAT,
3030 "", "",
3031 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3032 "(bytes)", "", "(bytes)");
3033 }
3034
3035 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) {
3036 if (!log_is_enabled(Trace, gc, liveness)) {
3037 return false;
3038 }
3039
3040 const char* type = r->get_type_str();
3041 HeapWord* bottom = r->bottom();
3042 HeapWord* end = r->end();
3043 size_t capacity_bytes = r->capacity();
3044 size_t used_bytes = r->used();
3045 size_t prev_live_bytes = r->live_bytes();
3046 size_t next_live_bytes = r->next_live_bytes();
3047 double gc_eff = r->gc_efficiency();
3048 size_t remset_bytes = r->rem_set()->mem_size();
3049 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3050 const char* remset_type = r->rem_set()->get_short_state_str();
3051
3052 _total_used_bytes += used_bytes;
3053 _total_capacity_bytes += capacity_bytes;
3054 _total_prev_live_bytes += prev_live_bytes;
3055 _total_next_live_bytes += next_live_bytes;
3056 _total_remset_bytes += remset_bytes;
3057 _total_strong_code_roots_bytes += strong_code_roots_bytes;
3058
3059 // Print a line for this particular region.
3060 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3061 G1PPRL_TYPE_FORMAT
3062 G1PPRL_ADDR_BASE_FORMAT
3063 G1PPRL_BYTE_FORMAT
3064 G1PPRL_BYTE_FORMAT
3065 G1PPRL_BYTE_FORMAT
3066 G1PPRL_DOUBLE_FORMAT
3067 G1PPRL_BYTE_FORMAT
3068 G1PPRL_STATE_FORMAT
3069 G1PPRL_BYTE_FORMAT,
3070 type, p2i(bottom), p2i(end),
3071 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3072 remset_bytes, remset_type, strong_code_roots_bytes);
3073
3074 return false;
3075 }
3076
3077 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3078 if (!log_is_enabled(Trace, gc, liveness)) {
3079 return;
3080 }
3081
3082 // add static memory usages to remembered set sizes
3083 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3084 // Print the footer of the output.
3085 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3086 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3087 " SUMMARY"
3088 G1PPRL_SUM_MB_FORMAT("capacity")
3089 G1PPRL_SUM_MB_PERC_FORMAT("used")
3090 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3091 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3092 G1PPRL_SUM_MB_FORMAT("remset")
3093 G1PPRL_SUM_MB_FORMAT("code-roots"),
3094 bytes_to_mb(_total_capacity_bytes),
3095 bytes_to_mb(_total_used_bytes),
3096 percent_of(_total_used_bytes, _total_capacity_bytes),
3097 bytes_to_mb(_total_prev_live_bytes),
3098 percent_of(_total_prev_live_bytes, _total_capacity_bytes),
3099 bytes_to_mb(_total_next_live_bytes),
3100 percent_of(_total_next_live_bytes, _total_capacity_bytes),
3101 bytes_to_mb(_total_remset_bytes),
3102 bytes_to_mb(_total_strong_code_roots_bytes));
3103 }