1 /*
2 * Copyright (c) 2001, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderDataGraph.hpp"
27 #include "code/codeCache.hpp"
28 #include "gc/g1/g1BarrierSet.hpp"
29 #include "gc/g1/g1CollectedHeap.inline.hpp"
30 #include "gc/g1/g1CollectorState.hpp"
31 #include "gc/g1/g1ConcurrentMark.inline.hpp"
32 #include "gc/g1/g1ConcurrentMarkThread.inline.hpp"
33 #include "gc/g1/g1HeapVerifier.hpp"
34 #include "gc/g1/g1OopClosures.inline.hpp"
35 #include "gc/g1/g1Policy.hpp"
36 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
37 #include "gc/g1/g1StringDedup.hpp"
38 #include "gc/g1/g1ThreadLocalData.hpp"
39 #include "gc/g1/heapRegion.inline.hpp"
40 #include "gc/g1/heapRegionRemSet.hpp"
41 #include "gc/g1/heapRegionSet.inline.hpp"
42 #include "gc/shared/gcId.hpp"
43 #include "gc/shared/gcTimer.hpp"
44 #include "gc/shared/gcTrace.hpp"
45 #include "gc/shared/gcTraceTime.inline.hpp"
46 #include "gc/shared/gcVMOperations.hpp"
47 #include "gc/shared/genOopClosures.inline.hpp"
48 #include "gc/shared/referencePolicy.hpp"
49 #include "gc/shared/strongRootsScope.hpp"
50 #include "gc/shared/suspendibleThreadSet.hpp"
51 #include "gc/shared/taskqueue.inline.hpp"
52 #include "gc/shared/weakProcessor.inline.hpp"
53 #include "gc/shared/workerPolicy.hpp"
54 #include "include/jvm.h"
55 #include "logging/log.hpp"
56 #include "memory/allocation.hpp"
57 #include "memory/resourceArea.hpp"
58 #include "oops/access.inline.hpp"
59 #include "oops/oop.inline.hpp"
60 #include "runtime/atomic.hpp"
61 #include "runtime/handles.inline.hpp"
62 #include "runtime/java.hpp"
63 #include "runtime/prefetch.inline.hpp"
64 #include "services/memTracker.hpp"
65 #include "utilities/align.hpp"
66 #include "utilities/growableArray.hpp"
67
68 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
69 assert(addr < _cm->finger(), "invariant");
70 assert(addr >= _task->finger(), "invariant");
71
72 // We move that task's local finger along.
73 _task->move_finger_to(addr);
74
75 _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
76 // we only partially drain the local queue and global stack
77 _task->drain_local_queue(true);
78 _task->drain_global_stack(true);
79
80 // if the has_aborted flag has been raised, we need to bail out of
81 // the iteration
82 return !_task->has_aborted();
83 }
84
85 G1CMMarkStack::G1CMMarkStack() :
86 _max_chunk_capacity(0),
87 _base(NULL),
88 _chunk_capacity(0) {
89 set_empty();
90 }
91
92 bool G1CMMarkStack::resize(size_t new_capacity) {
93 assert(is_empty(), "Only resize when stack is empty.");
94 assert(new_capacity <= _max_chunk_capacity,
95 "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
96
97 TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
98
99 if (new_base == NULL) {
100 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));
101 return false;
102 }
103 // Release old mapping.
104 if (_base != NULL) {
105 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
106 }
107
108 _base = new_base;
109 _chunk_capacity = new_capacity;
110 set_empty();
111
112 return true;
113 }
114
115 size_t G1CMMarkStack::capacity_alignment() {
116 return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
117 }
118
119 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
120 guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
121
122 size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
123
124 _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
125 size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
126
127 guarantee(initial_chunk_capacity <= _max_chunk_capacity,
128 "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
129 _max_chunk_capacity,
130 initial_chunk_capacity);
131
132 log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
133 initial_chunk_capacity, _max_chunk_capacity);
134
135 return resize(initial_chunk_capacity);
136 }
137
138 void G1CMMarkStack::expand() {
139 if (_chunk_capacity == _max_chunk_capacity) {
140 log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
141 return;
142 }
143 size_t old_capacity = _chunk_capacity;
144 // Double capacity if possible
145 size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
146
147 if (resize(new_capacity)) {
148 log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
149 old_capacity, new_capacity);
150 } else {
151 log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
152 old_capacity, new_capacity);
153 }
154 }
155
156 G1CMMarkStack::~G1CMMarkStack() {
157 if (_base != NULL) {
158 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
159 }
160 }
161
162 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
163 elem->next = *list;
164 *list = elem;
165 }
166
167 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
168 MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
169 add_chunk_to_list(&_chunk_list, elem);
170 _chunks_in_chunk_list++;
171 }
172
173 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
174 MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
175 add_chunk_to_list(&_free_list, elem);
176 }
177
178 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
179 TaskQueueEntryChunk* result = *list;
180 if (result != NULL) {
181 *list = (*list)->next;
182 }
183 return result;
184 }
185
186 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
187 MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
188 TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
189 if (result != NULL) {
190 _chunks_in_chunk_list--;
191 }
192 return result;
193 }
194
195 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
196 MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
197 return remove_chunk_from_list(&_free_list);
198 }
199
200 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
201 // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
202 // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
203 // wraparound of _hwm.
204 if (_hwm >= _chunk_capacity) {
205 return NULL;
206 }
207
208 size_t cur_idx = Atomic::add(1u, &_hwm) - 1;
209 if (cur_idx >= _chunk_capacity) {
210 return NULL;
211 }
212
213 TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
214 result->next = NULL;
215 return result;
216 }
217
218 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
219 // Get a new chunk.
220 TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
221
222 if (new_chunk == NULL) {
223 // Did not get a chunk from the free list. Allocate from backing memory.
224 new_chunk = allocate_new_chunk();
225
226 if (new_chunk == NULL) {
227 return false;
228 }
229 }
230
231 Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
232
233 add_chunk_to_chunk_list(new_chunk);
234
235 return true;
236 }
237
238 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
239 TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
240
241 if (cur == NULL) {
242 return false;
243 }
244
245 Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
246
247 add_chunk_to_free_list(cur);
248 return true;
249 }
250
251 void G1CMMarkStack::set_empty() {
252 _chunks_in_chunk_list = 0;
253 _hwm = 0;
254 _chunk_list = NULL;
255 _free_list = NULL;
256 }
257
258 G1CMRootRegions::G1CMRootRegions(uint const max_regions) :
259 _root_regions(NEW_C_HEAP_ARRAY(HeapRegion*, max_regions, mtGC)),
260 _max_regions(max_regions),
261 _num_root_regions(0),
262 _claimed_root_regions(0),
263 _scan_in_progress(false),
264 _should_abort(false) { }
265
266 G1CMRootRegions::~G1CMRootRegions() {
267 FREE_C_HEAP_ARRAY(HeapRegion*, _max_regions);
268 }
269
270 void G1CMRootRegions::reset() {
271 _num_root_regions = 0;
272 }
273
274 void G1CMRootRegions::add(HeapRegion* hr) {
275 assert_at_safepoint();
276 size_t idx = Atomic::add((size_t)1, &_num_root_regions) - 1;
277 assert(idx < _max_regions, "Trying to add more root regions than there is space " SIZE_FORMAT, _max_regions);
278 _root_regions[idx] = hr;
279 }
280
281 void G1CMRootRegions::prepare_for_scan() {
282 assert(!scan_in_progress(), "pre-condition");
283
284 _scan_in_progress = _num_root_regions > 0;
285
286 _claimed_root_regions = 0;
287 _should_abort = false;
288 }
289
290 HeapRegion* G1CMRootRegions::claim_next() {
291 if (_should_abort) {
292 // If someone has set the should_abort flag, we return NULL to
293 // force the caller to bail out of their loop.
294 return NULL;
295 }
296
297 if (_claimed_root_regions >= _num_root_regions) {
298 return NULL;
299 }
300
301 size_t claimed_index = Atomic::add((size_t)1, &_claimed_root_regions) - 1;
302 if (claimed_index < _num_root_regions) {
303 return _root_regions[claimed_index];
304 }
305 return NULL;
306 }
307
308 uint G1CMRootRegions::num_root_regions() const {
309 return (uint)_num_root_regions;
310 }
311
312 void G1CMRootRegions::notify_scan_done() {
313 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
314 _scan_in_progress = false;
315 RootRegionScan_lock->notify_all();
316 }
317
318 void G1CMRootRegions::cancel_scan() {
319 notify_scan_done();
320 }
321
322 void G1CMRootRegions::scan_finished() {
323 assert(scan_in_progress(), "pre-condition");
324
325 if (!_should_abort) {
326 assert(_claimed_root_regions >= num_root_regions(),
327 "we should have claimed all root regions, claimed " SIZE_FORMAT ", length = %u",
328 _claimed_root_regions, num_root_regions());
329 }
330
331 notify_scan_done();
332 }
333
334 bool G1CMRootRegions::wait_until_scan_finished() {
335 if (!scan_in_progress()) {
336 return false;
337 }
338
339 {
340 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
341 while (scan_in_progress()) {
342 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
343 }
344 }
345 return true;
346 }
347
348 // Returns the maximum number of workers to be used in a concurrent
349 // phase based on the number of GC workers being used in a STW
350 // phase.
351 static uint scale_concurrent_worker_threads(uint num_gc_workers) {
352 return MAX2((num_gc_workers + 2) / 4, 1U);
353 }
354
355 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
356 G1RegionToSpaceMapper* prev_bitmap_storage,
357 G1RegionToSpaceMapper* next_bitmap_storage) :
358 // _cm_thread set inside the constructor
359 _g1h(g1h),
360 _completed_initialization(false),
361
362 _mark_bitmap_1(),
363 _mark_bitmap_2(),
364 _prev_mark_bitmap(&_mark_bitmap_1),
365 _next_mark_bitmap(&_mark_bitmap_2),
366
367 _heap(_g1h->reserved_region()),
368
369 _root_regions(_g1h->max_regions()),
370
371 _global_mark_stack(),
372
373 // _finger set in set_non_marking_state
374
375 _worker_id_offset(DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads),
376 _max_num_tasks(ParallelGCThreads),
377 // _num_active_tasks set in set_non_marking_state()
378 // _tasks set inside the constructor
379
380 _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
381 _terminator((int) _max_num_tasks, _task_queues),
382
383 _first_overflow_barrier_sync(),
384 _second_overflow_barrier_sync(),
385
386 _has_overflown(false),
387 _concurrent(false),
388 _has_aborted(false),
389 _restart_for_overflow(false),
390 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
391 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
392
393 // _verbose_level set below
394
395 _init_times(),
396 _remark_times(),
397 _remark_mark_times(),
398 _remark_weak_ref_times(),
399 _cleanup_times(),
400 _total_cleanup_time(0.0),
401
402 _accum_task_vtime(NULL),
403
404 _concurrent_workers(NULL),
405 _num_concurrent_workers(0),
406 _max_concurrent_workers(0),
407
408 _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)),
409 _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC))
410 {
411 _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage);
412 _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage);
413
414 // Create & start ConcurrentMark thread.
415 _cm_thread = new G1ConcurrentMarkThread(this);
416 if (_cm_thread->osthread() == NULL) {
417 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
418 }
419
420 assert(CGC_lock != NULL, "CGC_lock must be initialized");
421
422 if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) {
423 // Calculate the number of concurrent worker threads by scaling
424 // the number of parallel GC threads.
425 uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads);
426 FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
427 }
428
429 assert(ConcGCThreads > 0, "ConcGCThreads have been set.");
430 if (ConcGCThreads > ParallelGCThreads) {
431 log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).",
432 ConcGCThreads, ParallelGCThreads);
433 return;
434 }
435
436 log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset);
437 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
438
439 _num_concurrent_workers = ConcGCThreads;
440 _max_concurrent_workers = _num_concurrent_workers;
441
442 _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true);
443 _concurrent_workers->initialize_workers();
444
445 if (FLAG_IS_DEFAULT(MarkStackSize)) {
446 size_t mark_stack_size =
447 MIN2(MarkStackSizeMax,
448 MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE)));
449 // Verify that the calculated value for MarkStackSize is in range.
450 // It would be nice to use the private utility routine from Arguments.
451 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
452 log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
453 "must be between 1 and " SIZE_FORMAT,
454 mark_stack_size, MarkStackSizeMax);
455 return;
456 }
457 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
458 } else {
459 // Verify MarkStackSize is in range.
460 if (FLAG_IS_CMDLINE(MarkStackSize)) {
461 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
462 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
463 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
464 "must be between 1 and " SIZE_FORMAT,
465 MarkStackSize, MarkStackSizeMax);
466 return;
467 }
468 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
469 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
470 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
471 " or for MarkStackSizeMax (" SIZE_FORMAT ")",
472 MarkStackSize, MarkStackSizeMax);
473 return;
474 }
475 }
476 }
477 }
478
479 if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
480 vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
481 }
482
483 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
484 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
485
486 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
487 _num_active_tasks = _max_num_tasks;
488
489 for (uint i = 0; i < _max_num_tasks; ++i) {
490 G1CMTaskQueue* task_queue = new G1CMTaskQueue();
491 task_queue->initialize();
492 _task_queues->register_queue(i, task_queue);
493
494 _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions());
495
496 _accum_task_vtime[i] = 0.0;
497 }
498
499 reset_at_marking_complete();
500 _completed_initialization = true;
501 }
502
503 void G1ConcurrentMark::reset() {
504 _has_aborted = false;
505
506 reset_marking_for_restart();
507
508 // Reset all tasks, since different phases will use different number of active
509 // threads. So, it's easiest to have all of them ready.
510 for (uint i = 0; i < _max_num_tasks; ++i) {
511 _tasks[i]->reset(_next_mark_bitmap);
512 }
513
514 uint max_regions = _g1h->max_regions();
515 for (uint i = 0; i < max_regions; i++) {
516 _top_at_rebuild_starts[i] = NULL;
517 _region_mark_stats[i].clear();
518 }
519 }
520
521 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) {
522 for (uint j = 0; j < _max_num_tasks; ++j) {
523 _tasks[j]->clear_mark_stats_cache(region_idx);
524 }
525 _top_at_rebuild_starts[region_idx] = NULL;
526 _region_mark_stats[region_idx].clear();
527 }
528
529 void G1ConcurrentMark::clear_statistics(HeapRegion* r) {
530 uint const region_idx = r->hrm_index();
531 if (r->is_humongous()) {
532 assert(r->is_starts_humongous(), "Got humongous continues region here");
533 uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size());
534 for (uint j = region_idx; j < (region_idx + size_in_regions); j++) {
535 clear_statistics_in_region(j);
536 }
537 } else {
538 clear_statistics_in_region(region_idx);
539 }
540 }
541
542 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) {
543 if (bitmap->is_marked(addr)) {
544 bitmap->clear(addr);
545 }
546 }
547
548 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) {
549 assert_at_safepoint_on_vm_thread();
550
551 // Need to clear all mark bits of the humongous object.
552 clear_mark_if_set(_prev_mark_bitmap, r->bottom());
553 clear_mark_if_set(_next_mark_bitmap, r->bottom());
554
555 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
556 return;
557 }
558
559 // Clear any statistics about the region gathered so far.
560 clear_statistics(r);
561 }
562
563 void G1ConcurrentMark::reset_marking_for_restart() {
564 _global_mark_stack.set_empty();
565
566 // Expand the marking stack, if we have to and if we can.
567 if (has_overflown()) {
568 _global_mark_stack.expand();
569
570 uint max_regions = _g1h->max_regions();
571 for (uint i = 0; i < max_regions; i++) {
572 _region_mark_stats[i].clear_during_overflow();
573 }
574 }
575
576 clear_has_overflown();
577 _finger = _heap.start();
578
579 for (uint i = 0; i < _max_num_tasks; ++i) {
580 G1CMTaskQueue* queue = _task_queues->queue(i);
581 queue->set_empty();
582 }
583 }
584
585 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
586 assert(active_tasks <= _max_num_tasks, "we should not have more");
587
588 _num_active_tasks = active_tasks;
589 // Need to update the three data structures below according to the
590 // number of active threads for this phase.
591 _terminator = TaskTerminator((int) active_tasks, _task_queues);
592 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
593 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
594 }
595
596 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
597 set_concurrency(active_tasks);
598
599 _concurrent = concurrent;
600
601 if (!concurrent) {
602 // At this point we should be in a STW phase, and completed marking.
603 assert_at_safepoint_on_vm_thread();
604 assert(out_of_regions(),
605 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
606 p2i(_finger), p2i(_heap.end()));
607 }
608 }
609
610 void G1ConcurrentMark::reset_at_marking_complete() {
611 // We set the global marking state to some default values when we're
612 // not doing marking.
613 reset_marking_for_restart();
614 _num_active_tasks = 0;
615 }
616
617 G1ConcurrentMark::~G1ConcurrentMark() {
618 FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts);
619 FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats);
620 // The G1ConcurrentMark instance is never freed.
621 ShouldNotReachHere();
622 }
623
624 class G1ClearBitMapTask : public AbstractGangTask {
625 public:
626 static size_t chunk_size() { return M; }
627
628 private:
629 // Heap region closure used for clearing the given mark bitmap.
630 class G1ClearBitmapHRClosure : public HeapRegionClosure {
631 private:
632 G1CMBitMap* _bitmap;
633 G1ConcurrentMark* _cm;
634 public:
635 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _bitmap(bitmap), _cm(cm) {
636 }
637
638 virtual bool do_heap_region(HeapRegion* r) {
639 size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
640
641 HeapWord* cur = r->bottom();
642 HeapWord* const end = r->end();
643
644 while (cur < end) {
645 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
646 _bitmap->clear_range(mr);
647
648 cur += chunk_size_in_words;
649
650 // Abort iteration if after yielding the marking has been aborted.
651 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
652 return true;
653 }
654 // Repeat the asserts from before the start of the closure. We will do them
655 // as asserts here to minimize their overhead on the product. However, we
656 // will have them as guarantees at the beginning / end of the bitmap
657 // clearing to get some checking in the product.
658 assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant");
659 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant");
660 }
661 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
662
663 return false;
664 }
665 };
666
667 G1ClearBitmapHRClosure _cl;
668 HeapRegionClaimer _hr_claimer;
669 bool _suspendible; // If the task is suspendible, workers must join the STS.
670
671 public:
672 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
673 AbstractGangTask("G1 Clear Bitmap"),
674 _cl(bitmap, suspendible ? cm : NULL),
675 _hr_claimer(n_workers),
676 _suspendible(suspendible)
677 { }
678
679 void work(uint worker_id) {
680 SuspendibleThreadSetJoiner sts_join(_suspendible);
681 G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
682 }
683
684 bool is_complete() {
685 return _cl.is_complete();
686 }
687 };
688
689 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
690 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
691
692 size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
693 size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
694
695 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
696
697 G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
698
699 log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
700 workers->run_task(&cl, num_workers);
701 guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
702 }
703
704 void G1ConcurrentMark::cleanup_for_next_mark() {
705 // Make sure that the concurrent mark thread looks to still be in
706 // the current cycle.
707 guarantee(cm_thread()->during_cycle(), "invariant");
708
709 // We are finishing up the current cycle by clearing the next
710 // marking bitmap and getting it ready for the next cycle. During
711 // this time no other cycle can start. So, let's make sure that this
712 // is the case.
713 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
714
715 clear_bitmap(_next_mark_bitmap, _concurrent_workers, true);
716
717 // Repeat the asserts from above.
718 guarantee(cm_thread()->during_cycle(), "invariant");
719 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
720 }
721
722 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
723 assert_at_safepoint_on_vm_thread();
724 clear_bitmap(_prev_mark_bitmap, workers, false);
725 }
726
727 class NoteStartOfMarkHRClosure : public HeapRegionClosure {
728 public:
729 bool do_heap_region(HeapRegion* r) {
730 r->note_start_of_marking();
731 return false;
732 }
733 };
734
735 void G1ConcurrentMark::pre_initial_mark() {
736 // Initialize marking structures. This has to be done in a STW phase.
737 reset();
738
739 // For each region note start of marking.
740 NoteStartOfMarkHRClosure startcl;
741 _g1h->heap_region_iterate(&startcl);
742
743 _root_regions.reset();
744 }
745
746
747 void G1ConcurrentMark::post_initial_mark() {
748 // Start Concurrent Marking weak-reference discovery.
749 ReferenceProcessor* rp = _g1h->ref_processor_cm();
750 // enable ("weak") refs discovery
751 rp->enable_discovery();
752 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
753
754 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
755 // This is the start of the marking cycle, we're expected all
756 // threads to have SATB queues with active set to false.
757 satb_mq_set.set_active_all_threads(true, /* new active value */
758 false /* expected_active */);
759
760 _root_regions.prepare_for_scan();
761
762 // update_g1_committed() will be called at the end of an evac pause
763 // when marking is on. So, it's also called at the end of the
764 // initial-mark pause to update the heap end, if the heap expands
765 // during it. No need to call it here.
766 }
767
768 /*
769 * Notice that in the next two methods, we actually leave the STS
770 * during the barrier sync and join it immediately afterwards. If we
771 * do not do this, the following deadlock can occur: one thread could
772 * be in the barrier sync code, waiting for the other thread to also
773 * sync up, whereas another one could be trying to yield, while also
774 * waiting for the other threads to sync up too.
775 *
776 * Note, however, that this code is also used during remark and in
777 * this case we should not attempt to leave / enter the STS, otherwise
778 * we'll either hit an assert (debug / fastdebug) or deadlock
779 * (product). So we should only leave / enter the STS if we are
780 * operating concurrently.
781 *
782 * Because the thread that does the sync barrier has left the STS, it
783 * is possible to be suspended for a Full GC or an evacuation pause
784 * could occur. This is actually safe, since the entering the sync
785 * barrier is one of the last things do_marking_step() does, and it
786 * doesn't manipulate any data structures afterwards.
787 */
788
789 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
790 bool barrier_aborted;
791 {
792 SuspendibleThreadSetLeaver sts_leave(concurrent());
793 barrier_aborted = !_first_overflow_barrier_sync.enter();
794 }
795
796 // at this point everyone should have synced up and not be doing any
797 // more work
798
799 if (barrier_aborted) {
800 // If the barrier aborted we ignore the overflow condition and
801 // just abort the whole marking phase as quickly as possible.
802 return;
803 }
804 }
805
806 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
807 SuspendibleThreadSetLeaver sts_leave(concurrent());
808 _second_overflow_barrier_sync.enter();
809
810 // at this point everything should be re-initialized and ready to go
811 }
812
813 class G1CMConcurrentMarkingTask : public AbstractGangTask {
814 G1ConcurrentMark* _cm;
815
816 public:
817 void work(uint worker_id) {
818 assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread");
819 ResourceMark rm;
820
821 double start_vtime = os::elapsedVTime();
822
823 {
824 SuspendibleThreadSetJoiner sts_join;
825
826 assert(worker_id < _cm->active_tasks(), "invariant");
827
828 G1CMTask* task = _cm->task(worker_id);
829 task->record_start_time();
830 if (!_cm->has_aborted()) {
831 do {
832 task->do_marking_step(G1ConcMarkStepDurationMillis,
833 true /* do_termination */,
834 false /* is_serial*/);
835
836 _cm->do_yield_check();
837 } while (!_cm->has_aborted() && task->has_aborted());
838 }
839 task->record_end_time();
840 guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
841 }
842
843 double end_vtime = os::elapsedVTime();
844 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
845 }
846
847 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) :
848 AbstractGangTask("Concurrent Mark"), _cm(cm) { }
849
850 ~G1CMConcurrentMarkingTask() { }
851 };
852
853 uint G1ConcurrentMark::calc_active_marking_workers() {
854 uint result = 0;
855 if (!UseDynamicNumberOfGCThreads ||
856 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
857 !ForceDynamicNumberOfGCThreads)) {
858 result = _max_concurrent_workers;
859 } else {
860 result =
861 WorkerPolicy::calc_default_active_workers(_max_concurrent_workers,
862 1, /* Minimum workers */
863 _num_concurrent_workers,
864 Threads::number_of_non_daemon_threads());
865 // Don't scale the result down by scale_concurrent_workers() because
866 // that scaling has already gone into "_max_concurrent_workers".
867 }
868 assert(result > 0 && result <= _max_concurrent_workers,
869 "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
870 _max_concurrent_workers, result);
871 return result;
872 }
873
874 void G1ConcurrentMark::scan_root_region(HeapRegion* hr, uint worker_id) {
875 assert(hr->is_old() || (hr->is_survivor() && hr->next_top_at_mark_start() == hr->bottom()),
876 "Root regions must be old or survivor but region %u is %s", hr->hrm_index(), hr->get_type_str());
877 G1RootRegionScanClosure cl(_g1h, this, worker_id);
878
879 const uintx interval = PrefetchScanIntervalInBytes;
880 HeapWord* curr = hr->next_top_at_mark_start();
881 const HeapWord* end = hr->top();
882 while (curr < end) {
883 Prefetch::read(curr, interval);
884 oop obj = oop(curr);
885 int size = obj->oop_iterate_size(&cl);
886 assert(size == obj->size(), "sanity");
887 curr += size;
888 }
889 }
890
891 class G1CMRootRegionScanTask : public AbstractGangTask {
892 G1ConcurrentMark* _cm;
893 public:
894 G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
895 AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
896
897 void work(uint worker_id) {
898 assert(Thread::current()->is_ConcurrentGC_thread(),
899 "this should only be done by a conc GC thread");
900
901 G1CMRootRegions* root_regions = _cm->root_regions();
902 HeapRegion* hr = root_regions->claim_next();
903 while (hr != NULL) {
904 _cm->scan_root_region(hr, worker_id);
905 hr = root_regions->claim_next();
906 }
907 }
908 };
909
910 void G1ConcurrentMark::scan_root_regions() {
911 // scan_in_progress() will have been set to true only if there was
912 // at least one root region to scan. So, if it's false, we
913 // should not attempt to do any further work.
914 if (root_regions()->scan_in_progress()) {
915 assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
916
917 _num_concurrent_workers = MIN2(calc_active_marking_workers(),
918 // We distribute work on a per-region basis, so starting
919 // more threads than that is useless.
920 root_regions()->num_root_regions());
921 assert(_num_concurrent_workers <= _max_concurrent_workers,
922 "Maximum number of marking threads exceeded");
923
924 G1CMRootRegionScanTask task(this);
925 log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
926 task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
927 _concurrent_workers->run_task(&task, _num_concurrent_workers);
928
929 // It's possible that has_aborted() is true here without actually
930 // aborting the survivor scan earlier. This is OK as it's
931 // mainly used for sanity checking.
932 root_regions()->scan_finished();
933 }
934 }
935
936 void G1ConcurrentMark::concurrent_cycle_start() {
937 _gc_timer_cm->register_gc_start();
938
939 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
940
941 _g1h->trace_heap_before_gc(_gc_tracer_cm);
942 }
943
944 void G1ConcurrentMark::concurrent_cycle_end() {
945 _g1h->collector_state()->set_clearing_next_bitmap(false);
946
947 _g1h->trace_heap_after_gc(_gc_tracer_cm);
948
949 if (has_aborted()) {
950 log_info(gc, marking)("Concurrent Mark Abort");
951 _gc_tracer_cm->report_concurrent_mode_failure();
952 }
953
954 _gc_timer_cm->register_gc_end();
955
956 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
957 }
958
959 void G1ConcurrentMark::mark_from_roots() {
960 _restart_for_overflow = false;
961
962 _num_concurrent_workers = calc_active_marking_workers();
963
964 uint active_workers = MAX2(1U, _num_concurrent_workers);
965
966 // Setting active workers is not guaranteed since fewer
967 // worker threads may currently exist and more may not be
968 // available.
969 active_workers = _concurrent_workers->update_active_workers(active_workers);
970 log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers());
971
972 // Parallel task terminator is set in "set_concurrency_and_phase()"
973 set_concurrency_and_phase(active_workers, true /* concurrent */);
974
975 G1CMConcurrentMarkingTask marking_task(this);
976 _concurrent_workers->run_task(&marking_task);
977 print_stats();
978 }
979
980 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) {
981 G1HeapVerifier* verifier = _g1h->verifier();
982
983 verifier->verify_region_sets_optional();
984
985 if (VerifyDuringGC) {
986 GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm);
987
988 size_t const BufLen = 512;
989 char buffer[BufLen];
990
991 jio_snprintf(buffer, BufLen, "During GC (%s)", caller);
992 verifier->verify(type, vo, buffer);
993 }
994
995 verifier->check_bitmaps(caller);
996 }
997
998 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask {
999 G1CollectedHeap* _g1h;
1000 G1ConcurrentMark* _cm;
1001 HeapRegionClaimer _hrclaimer;
1002 uint volatile _total_selected_for_rebuild;
1003
1004 G1PrintRegionLivenessInfoClosure _cl;
1005
1006 class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure {
1007 G1CollectedHeap* _g1h;
1008 G1ConcurrentMark* _cm;
1009
1010 G1PrintRegionLivenessInfoClosure* _cl;
1011
1012 uint _num_regions_selected_for_rebuild; // The number of regions actually selected for rebuild.
1013
1014 void update_remset_before_rebuild(HeapRegion* hr) {
1015 G1RemSetTrackingPolicy* tracking_policy = _g1h->g1_policy()->remset_tracker();
1016
1017 bool selected_for_rebuild;
1018 if (hr->is_humongous()) {
1019 bool const is_live = _cm->liveness(hr->humongous_start_region()->hrm_index()) > 0;
1020 selected_for_rebuild = tracking_policy->update_humongous_before_rebuild(hr, is_live);
1021 } else {
1022 size_t const live_bytes = _cm->liveness(hr->hrm_index());
1023 selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);
1024 }
1025 if (selected_for_rebuild) {
1026 _num_regions_selected_for_rebuild++;
1027 }
1028 _cm->update_top_at_rebuild_start(hr);
1029 }
1030
1031 // Distribute the given words across the humongous object starting with hr and
1032 // note end of marking.
1033 void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) {
1034 uint const region_idx = hr->hrm_index();
1035 size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size();
1036 uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words);
1037
1038 // "Distributing" zero words means that we only note end of marking for these
1039 // regions.
1040 assert(marked_words == 0 || obj_size_in_words == marked_words,
1041 "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT,
1042 obj_size_in_words, marked_words);
1043
1044 for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) {
1045 HeapRegion* const r = _g1h->region_at(i);
1046 size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words);
1047
1048 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)",
1049 words_to_add, i, r->get_type_str());
1050 add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize);
1051 marked_words -= words_to_add;
1052 }
1053 assert(marked_words == 0,
1054 SIZE_FORMAT " words left after distributing space across %u regions",
1055 marked_words, num_regions_in_humongous);
1056 }
1057
1058 void update_marked_bytes(HeapRegion* hr) {
1059 uint const region_idx = hr->hrm_index();
1060 size_t const marked_words = _cm->liveness(region_idx);
1061 // The marking attributes the object's size completely to the humongous starts
1062 // region. We need to distribute this value across the entire set of regions a
1063 // humongous object spans.
1064 if (hr->is_humongous()) {
1065 assert(hr->is_starts_humongous() || marked_words == 0,
1066 "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)",
1067 marked_words, region_idx, hr->get_type_str());
1068 if (hr->is_starts_humongous()) {
1069 distribute_marked_bytes(hr, marked_words);
1070 }
1071 } else {
1072 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str());
1073 add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize);
1074 }
1075 }
1076
1077 void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) {
1078 hr->add_to_marked_bytes(marked_bytes);
1079 _cl->do_heap_region(hr);
1080 hr->note_end_of_marking();
1081 }
1082
1083 public:
1084 G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) :
1085 _g1h(g1h), _cm(cm), _cl(cl), _num_regions_selected_for_rebuild(0) { }
1086
1087 virtual bool do_heap_region(HeapRegion* r) {
1088 update_remset_before_rebuild(r);
1089 update_marked_bytes(r);
1090
1091 return false;
1092 }
1093
1094 uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; }
1095 };
1096
1097 public:
1098 G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) :
1099 AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"),
1100 _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { }
1101
1102 virtual void work(uint worker_id) {
1103 G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl);
1104 _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id);
1105 Atomic::add(update_cl.num_selected_for_rebuild(), &_total_selected_for_rebuild);
1106 }
1107
1108 uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; }
1109
1110 // Number of regions for which roughly one thread should be spawned for this work.
1111 static const uint RegionsPerThread = 384;
1112 };
1113
1114 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure {
1115 G1CollectedHeap* _g1h;
1116 public:
1117 G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { }
1118
1119 virtual bool do_heap_region(HeapRegion* r) {
1120 _g1h->g1_policy()->remset_tracker()->update_after_rebuild(r);
1121 return false;
1122 }
1123 };
1124
1125 void G1ConcurrentMark::remark() {
1126 assert_at_safepoint_on_vm_thread();
1127
1128 // If a full collection has happened, we should not continue. However we might
1129 // have ended up here as the Remark VM operation has been scheduled already.
1130 if (has_aborted()) {
1131 return;
1132 }
1133
1134 G1Policy* g1p = _g1h->g1_policy();
1135 g1p->record_concurrent_mark_remark_start();
1136
1137 double start = os::elapsedTime();
1138
1139 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before");
1140
1141 {
1142 GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm);
1143 finalize_marking();
1144 }
1145
1146 double mark_work_end = os::elapsedTime();
1147
1148 bool const mark_finished = !has_overflown();
1149 if (mark_finished) {
1150 weak_refs_work(false /* clear_all_soft_refs */);
1151
1152 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1153 // We're done with marking.
1154 // This is the end of the marking cycle, we're expected all
1155 // threads to have SATB queues with active set to true.
1156 satb_mq_set.set_active_all_threads(false, /* new active value */
1157 true /* expected_active */);
1158
1159 {
1160 GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm);
1161 flush_all_task_caches();
1162 }
1163
1164 // Install newly created mark bitmap as "prev".
1165 swap_mark_bitmaps();
1166 {
1167 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm);
1168
1169 uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) /
1170 G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread;
1171 uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity);
1172
1173 G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers);
1174 log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions());
1175 _g1h->workers()->run_task(&cl, num_workers);
1176
1177 log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u",
1178 _g1h->num_regions(), cl.total_selected_for_rebuild());
1179 }
1180 {
1181 GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm);
1182 reclaim_empty_regions();
1183 }
1184
1185 // Clean out dead classes
1186 if (ClassUnloadingWithConcurrentMark) {
1187 GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm);
1188 ClassLoaderDataGraph::purge();
1189 }
1190
1191 _g1h->resize_heap_if_necessary();
1192
1193 compute_new_sizes();
1194
1195 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after");
1196
1197 assert(!restart_for_overflow(), "sanity");
1198 // Completely reset the marking state since marking completed
1199 reset_at_marking_complete();
1200 } else {
1201 // We overflowed. Restart concurrent marking.
1202 _restart_for_overflow = true;
1203
1204 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow");
1205
1206 // Clear the marking state because we will be restarting
1207 // marking due to overflowing the global mark stack.
1208 reset_marking_for_restart();
1209 }
1210
1211 {
1212 GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm);
1213 report_object_count(mark_finished);
1214 }
1215
1216 // Statistics
1217 double now = os::elapsedTime();
1218 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1219 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1220 _remark_times.add((now - start) * 1000.0);
1221
1222 g1p->record_concurrent_mark_remark_end();
1223 }
1224
1225 class G1ReclaimEmptyRegionsTask : public AbstractGangTask {
1226 // Per-region work during the Cleanup pause.
1227 class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure {
1228 G1CollectedHeap* _g1h;
1229 size_t _freed_bytes;
1230 FreeRegionList* _local_cleanup_list;
1231 uint _old_regions_removed;
1232 uint _humongous_regions_removed;
1233
1234 public:
1235 G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h,
1236 FreeRegionList* local_cleanup_list) :
1237 _g1h(g1h),
1238 _freed_bytes(0),
1239 _local_cleanup_list(local_cleanup_list),
1240 _old_regions_removed(0),
1241 _humongous_regions_removed(0) { }
1242
1243 size_t freed_bytes() { return _freed_bytes; }
1244 const uint old_regions_removed() { return _old_regions_removed; }
1245 const uint humongous_regions_removed() { return _humongous_regions_removed; }
1246
1247 bool do_heap_region(HeapRegion *hr) {
1248 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) {
1249 _freed_bytes += hr->used();
1250 hr->set_containing_set(NULL);
1251 if (hr->is_humongous()) {
1252 _humongous_regions_removed++;
1253 _g1h->free_humongous_region(hr, _local_cleanup_list);
1254 } else {
1255 _old_regions_removed++;
1256 _g1h->free_region(hr, _local_cleanup_list, false /* skip_remset */, false /* skip_hcc */, true /* locked */);
1257 }
1258 hr->clear_cardtable();
1259 _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index());
1260 log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom()));
1261 }
1262
1263 return false;
1264 }
1265 };
1266
1267 G1CollectedHeap* _g1h;
1268 FreeRegionList* _cleanup_list;
1269 HeapRegionClaimer _hrclaimer;
1270
1271 public:
1272 G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1273 AbstractGangTask("G1 Cleanup"),
1274 _g1h(g1h),
1275 _cleanup_list(cleanup_list),
1276 _hrclaimer(n_workers) {
1277 }
1278
1279 void work(uint worker_id) {
1280 FreeRegionList local_cleanup_list("Local Cleanup List");
1281 G1ReclaimEmptyRegionsClosure cl(_g1h, &local_cleanup_list);
1282 _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id);
1283 assert(cl.is_complete(), "Shouldn't have aborted!");
1284
1285 // Now update the old/humongous region sets
1286 _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed());
1287 {
1288 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1289 _g1h->decrement_summary_bytes(cl.freed_bytes());
1290
1291 _cleanup_list->add_ordered(&local_cleanup_list);
1292 assert(local_cleanup_list.is_empty(), "post-condition");
1293 }
1294 }
1295 };
1296
1297 void G1ConcurrentMark::reclaim_empty_regions() {
1298 WorkGang* workers = _g1h->workers();
1299 FreeRegionList empty_regions_list("Empty Regions After Mark List");
1300
1301 G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers());
1302 workers->run_task(&cl);
1303
1304 if (!empty_regions_list.is_empty()) {
1305 log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length());
1306 // Now print the empty regions list.
1307 G1HRPrinter* hrp = _g1h->hr_printer();
1308 if (hrp->is_active()) {
1309 FreeRegionListIterator iter(&empty_regions_list);
1310 while (iter.more_available()) {
1311 HeapRegion* hr = iter.get_next();
1312 hrp->cleanup(hr);
1313 }
1314 }
1315 // And actually make them available.
1316 _g1h->prepend_to_freelist(&empty_regions_list);
1317 }
1318 }
1319
1320 void G1ConcurrentMark::compute_new_sizes() {
1321 MetaspaceGC::compute_new_size();
1322
1323 // Cleanup will have freed any regions completely full of garbage.
1324 // Update the soft reference policy with the new heap occupancy.
1325 Universe::update_heap_info_at_gc();
1326
1327 // We reclaimed old regions so we should calculate the sizes to make
1328 // sure we update the old gen/space data.
1329 _g1h->g1mm()->update_sizes();
1330 }
1331
1332 void G1ConcurrentMark::cleanup() {
1333 assert_at_safepoint_on_vm_thread();
1334
1335 // If a full collection has happened, we shouldn't do this.
1336 if (has_aborted()) {
1337 return;
1338 }
1339
1340 G1Policy* g1p = _g1h->g1_policy();
1341 g1p->record_concurrent_mark_cleanup_start();
1342
1343 double start = os::elapsedTime();
1344
1345 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before");
1346
1347 {
1348 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm);
1349 G1UpdateRemSetTrackingAfterRebuild cl(_g1h);
1350 _g1h->heap_region_iterate(&cl);
1351 }
1352
1353 if (log_is_enabled(Trace, gc, liveness)) {
1354 G1PrintRegionLivenessInfoClosure cl("Post-Cleanup");
1355 _g1h->heap_region_iterate(&cl);
1356 }
1357
1358 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after");
1359
1360 // We need to make this be a "collection" so any collection pause that
1361 // races with it goes around and waits for Cleanup to finish.
1362 _g1h->increment_total_collections();
1363
1364 // Local statistics
1365 double recent_cleanup_time = (os::elapsedTime() - start);
1366 _total_cleanup_time += recent_cleanup_time;
1367 _cleanup_times.add(recent_cleanup_time);
1368
1369 {
1370 GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm);
1371 _g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1372 }
1373 }
1374
1375 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1376 // Uses the G1CMTask associated with a worker thread (for serial reference
1377 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1378 // trace referent objects.
1379 //
1380 // Using the G1CMTask and embedded local queues avoids having the worker
1381 // threads operating on the global mark stack. This reduces the risk
1382 // of overflowing the stack - which we would rather avoid at this late
1383 // state. Also using the tasks' local queues removes the potential
1384 // of the workers interfering with each other that could occur if
1385 // operating on the global stack.
1386
1387 class G1CMKeepAliveAndDrainClosure : public OopClosure {
1388 G1ConcurrentMark* _cm;
1389 G1CMTask* _task;
1390 uint _ref_counter_limit;
1391 uint _ref_counter;
1392 bool _is_serial;
1393 public:
1394 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1395 _cm(cm), _task(task), _ref_counter_limit(G1RefProcDrainInterval),
1396 _ref_counter(_ref_counter_limit), _is_serial(is_serial) {
1397 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1398 }
1399
1400 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1401 virtual void do_oop( oop* p) { do_oop_work(p); }
1402
1403 template <class T> void do_oop_work(T* p) {
1404 if (_cm->has_overflown()) {
1405 return;
1406 }
1407 if (!_task->deal_with_reference(p)) {
1408 // We did not add anything to the mark bitmap (or mark stack), so there is
1409 // no point trying to drain it.
1410 return;
1411 }
1412 _ref_counter--;
1413
1414 if (_ref_counter == 0) {
1415 // We have dealt with _ref_counter_limit references, pushing them
1416 // and objects reachable from them on to the local stack (and
1417 // possibly the global stack). Call G1CMTask::do_marking_step() to
1418 // process these entries.
1419 //
1420 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1421 // there's nothing more to do (i.e. we're done with the entries that
1422 // were pushed as a result of the G1CMTask::deal_with_reference() calls
1423 // above) or we overflow.
1424 //
1425 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1426 // flag while there may still be some work to do. (See the comment at
1427 // the beginning of G1CMTask::do_marking_step() for those conditions -
1428 // one of which is reaching the specified time target.) It is only
1429 // when G1CMTask::do_marking_step() returns without setting the
1430 // has_aborted() flag that the marking step has completed.
1431 do {
1432 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1433 _task->do_marking_step(mark_step_duration_ms,
1434 false /* do_termination */,
1435 _is_serial);
1436 } while (_task->has_aborted() && !_cm->has_overflown());
1437 _ref_counter = _ref_counter_limit;
1438 }
1439 }
1440 };
1441
1442 // 'Drain' oop closure used by both serial and parallel reference processing.
1443 // Uses the G1CMTask associated with a given worker thread (for serial
1444 // reference processing the G1CMtask for worker 0 is used). Calls the
1445 // do_marking_step routine, with an unbelievably large timeout value,
1446 // to drain the marking data structures of the remaining entries
1447 // added by the 'keep alive' oop closure above.
1448
1449 class G1CMDrainMarkingStackClosure : public VoidClosure {
1450 G1ConcurrentMark* _cm;
1451 G1CMTask* _task;
1452 bool _is_serial;
1453 public:
1454 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1455 _cm(cm), _task(task), _is_serial(is_serial) {
1456 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1457 }
1458
1459 void do_void() {
1460 do {
1461 // We call G1CMTask::do_marking_step() to completely drain the local
1462 // and global marking stacks of entries pushed by the 'keep alive'
1463 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1464 //
1465 // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1466 // if there's nothing more to do (i.e. we've completely drained the
1467 // entries that were pushed as a a result of applying the 'keep alive'
1468 // closure to the entries on the discovered ref lists) or we overflow
1469 // the global marking stack.
1470 //
1471 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1472 // flag while there may still be some work to do. (See the comment at
1473 // the beginning of G1CMTask::do_marking_step() for those conditions -
1474 // one of which is reaching the specified time target.) It is only
1475 // when G1CMTask::do_marking_step() returns without setting the
1476 // has_aborted() flag that the marking step has completed.
1477
1478 _task->do_marking_step(1000000000.0 /* something very large */,
1479 true /* do_termination */,
1480 _is_serial);
1481 } while (_task->has_aborted() && !_cm->has_overflown());
1482 }
1483 };
1484
1485 // Implementation of AbstractRefProcTaskExecutor for parallel
1486 // reference processing at the end of G1 concurrent marking
1487
1488 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor {
1489 private:
1490 G1CollectedHeap* _g1h;
1491 G1ConcurrentMark* _cm;
1492 WorkGang* _workers;
1493 uint _active_workers;
1494
1495 public:
1496 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1497 G1ConcurrentMark* cm,
1498 WorkGang* workers,
1499 uint n_workers) :
1500 _g1h(g1h), _cm(cm),
1501 _workers(workers), _active_workers(n_workers) { }
1502
1503 virtual void execute(ProcessTask& task, uint ergo_workers);
1504 };
1505
1506 class G1CMRefProcTaskProxy : public AbstractGangTask {
1507 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1508 ProcessTask& _proc_task;
1509 G1CollectedHeap* _g1h;
1510 G1ConcurrentMark* _cm;
1511
1512 public:
1513 G1CMRefProcTaskProxy(ProcessTask& proc_task,
1514 G1CollectedHeap* g1h,
1515 G1ConcurrentMark* cm) :
1516 AbstractGangTask("Process reference objects in parallel"),
1517 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1518 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1519 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1520 }
1521
1522 virtual void work(uint worker_id) {
1523 ResourceMark rm;
1524 HandleMark hm;
1525 G1CMTask* task = _cm->task(worker_id);
1526 G1CMIsAliveClosure g1_is_alive(_g1h);
1527 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1528 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1529
1530 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1531 }
1532 };
1533
1534 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task, uint ergo_workers) {
1535 assert(_workers != NULL, "Need parallel worker threads.");
1536 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1537 assert(_workers->active_workers() >= ergo_workers,
1538 "Ergonomically chosen workers(%u) should be less than or equal to active workers(%u)",
1539 ergo_workers, _workers->active_workers());
1540
1541 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1542
1543 // We need to reset the concurrency level before each
1544 // proxy task execution, so that the termination protocol
1545 // and overflow handling in G1CMTask::do_marking_step() knows
1546 // how many workers to wait for.
1547 _cm->set_concurrency(ergo_workers);
1548 _workers->run_task(&proc_task_proxy, ergo_workers);
1549 }
1550
1551 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) {
1552 ResourceMark rm;
1553 HandleMark hm;
1554
1555 // Is alive closure.
1556 G1CMIsAliveClosure g1_is_alive(_g1h);
1557
1558 // Inner scope to exclude the cleaning of the string table
1559 // from the displayed time.
1560 {
1561 GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm);
1562
1563 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1564
1565 // See the comment in G1CollectedHeap::ref_processing_init()
1566 // about how reference processing currently works in G1.
1567
1568 // Set the soft reference policy
1569 rp->setup_policy(clear_all_soft_refs);
1570 assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1571
1572 // Instances of the 'Keep Alive' and 'Complete GC' closures used
1573 // in serial reference processing. Note these closures are also
1574 // used for serially processing (by the the current thread) the
1575 // JNI references during parallel reference processing.
1576 //
1577 // These closures do not need to synchronize with the worker
1578 // threads involved in parallel reference processing as these
1579 // instances are executed serially by the current thread (e.g.
1580 // reference processing is not multi-threaded and is thus
1581 // performed by the current thread instead of a gang worker).
1582 //
1583 // The gang tasks involved in parallel reference processing create
1584 // their own instances of these closures, which do their own
1585 // synchronization among themselves.
1586 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1587 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1588
1589 // We need at least one active thread. If reference processing
1590 // is not multi-threaded we use the current (VMThread) thread,
1591 // otherwise we use the work gang from the G1CollectedHeap and
1592 // we utilize all the worker threads we can.
1593 bool processing_is_mt = rp->processing_is_mt();
1594 uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U);
1595 active_workers = MAX2(MIN2(active_workers, _max_num_tasks), 1U);
1596
1597 // Parallel processing task executor.
1598 G1CMRefProcTaskExecutor par_task_executor(_g1h, this,
1599 _g1h->workers(), active_workers);
1600 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1601
1602 // Set the concurrency level. The phase was already set prior to
1603 // executing the remark task.
1604 set_concurrency(active_workers);
1605
1606 // Set the degree of MT processing here. If the discovery was done MT,
1607 // the number of threads involved during discovery could differ from
1608 // the number of active workers. This is OK as long as the discovered
1609 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1610 rp->set_active_mt_degree(active_workers);
1611
1612 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues());
1613
1614 // Process the weak references.
1615 const ReferenceProcessorStats& stats =
1616 rp->process_discovered_references(&g1_is_alive,
1617 &g1_keep_alive,
1618 &g1_drain_mark_stack,
1619 executor,
1620 &pt);
1621 _gc_tracer_cm->report_gc_reference_stats(stats);
1622 pt.print_all_references();
1623
1624 // The do_oop work routines of the keep_alive and drain_marking_stack
1625 // oop closures will set the has_overflown flag if we overflow the
1626 // global marking stack.
1627
1628 assert(has_overflown() || _global_mark_stack.is_empty(),
1629 "Mark stack should be empty (unless it has overflown)");
1630
1631 assert(rp->num_queues() == active_workers, "why not");
1632
1633 rp->verify_no_references_recorded();
1634 assert(!rp->discovery_enabled(), "Post condition");
1635 }
1636
1637 if (has_overflown()) {
1638 // We can not trust g1_is_alive and the contents of the heap if the marking stack
1639 // overflowed while processing references. Exit the VM.
1640 fatal("Overflow during reference processing, can not continue. Please "
1641 "increase MarkStackSizeMax (current value: " SIZE_FORMAT ") and "
1642 "restart.", MarkStackSizeMax);
1643 return;
1644 }
1645
1646 assert(_global_mark_stack.is_empty(), "Marking should have completed");
1647
1648 {
1649 GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1650 WeakProcessor::weak_oops_do(_g1h->workers(), &g1_is_alive, &do_nothing_cl, 1);
1651 }
1652
1653 // Unload Klasses, String, Code Cache, etc.
1654 if (ClassUnloadingWithConcurrentMark) {
1655 GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1656 bool purged_classes = SystemDictionary::do_unloading(_gc_timer_cm);
1657 _g1h->complete_cleaning(&g1_is_alive, purged_classes);
1658 } else if (StringDedup::is_enabled()) {
1659 GCTraceTime(Debug, gc, phases) debug("String Deduplication", _gc_timer_cm);
1660 _g1h->string_dedup_cleaning(&g1_is_alive, NULL);
1661 }
1662 }
1663
1664 class G1PrecleanYieldClosure : public YieldClosure {
1665 G1ConcurrentMark* _cm;
1666
1667 public:
1668 G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { }
1669
1670 virtual bool should_return() {
1671 return _cm->has_aborted();
1672 }
1673
1674 virtual bool should_return_fine_grain() {
1675 _cm->do_yield_check();
1676 return _cm->has_aborted();
1677 }
1678 };
1679
1680 void G1ConcurrentMark::preclean() {
1681 assert(G1UseReferencePrecleaning, "Precleaning must be enabled.");
1682
1683 SuspendibleThreadSetJoiner joiner;
1684
1685 G1CMKeepAliveAndDrainClosure keep_alive(this, task(0), true /* is_serial */);
1686 G1CMDrainMarkingStackClosure drain_mark_stack(this, task(0), true /* is_serial */);
1687
1688 set_concurrency_and_phase(1, true);
1689
1690 G1PrecleanYieldClosure yield_cl(this);
1691
1692 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1693 // Precleaning is single threaded. Temporarily disable MT discovery.
1694 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
1695 rp->preclean_discovered_references(rp->is_alive_non_header(),
1696 &keep_alive,
1697 &drain_mark_stack,
1698 &yield_cl,
1699 _gc_timer_cm);
1700 }
1701
1702 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
1703 // the prev bitmap determining liveness.
1704 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {
1705 G1CollectedHeap* _g1h;
1706 public:
1707 G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
1708
1709 bool do_object_b(oop obj) {
1710 HeapWord* addr = (HeapWord*)obj;
1711 return addr != NULL &&
1712 (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_dead(obj));
1713 }
1714 };
1715
1716 void G1ConcurrentMark::report_object_count(bool mark_completed) {
1717 // Depending on the completion of the marking liveness needs to be determined
1718 // using either the next or prev bitmap.
1719 if (mark_completed) {
1720 G1ObjectCountIsAliveClosure is_alive(_g1h);
1721 _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1722 } else {
1723 G1CMIsAliveClosure is_alive(_g1h);
1724 _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1725 }
1726 }
1727
1728
1729 void G1ConcurrentMark::swap_mark_bitmaps() {
1730 G1CMBitMap* temp = _prev_mark_bitmap;
1731 _prev_mark_bitmap = _next_mark_bitmap;
1732 _next_mark_bitmap = temp;
1733 _g1h->collector_state()->set_clearing_next_bitmap(true);
1734 }
1735
1736 // Closure for marking entries in SATB buffers.
1737 class G1CMSATBBufferClosure : public SATBBufferClosure {
1738 private:
1739 G1CMTask* _task;
1740 G1CollectedHeap* _g1h;
1741
1742 // This is very similar to G1CMTask::deal_with_reference, but with
1743 // more relaxed requirements for the argument, so this must be more
1744 // circumspect about treating the argument as an object.
1745 void do_entry(void* entry) const {
1746 _task->increment_refs_reached();
1747 oop const obj = static_cast<oop>(entry);
1748 _task->make_reference_grey(obj);
1749 }
1750
1751 public:
1752 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1753 : _task(task), _g1h(g1h) { }
1754
1755 virtual void do_buffer(void** buffer, size_t size) {
1756 for (size_t i = 0; i < size; ++i) {
1757 do_entry(buffer[i]);
1758 }
1759 }
1760 };
1761
1762 class G1RemarkThreadsClosure : public ThreadClosure {
1763 G1CMSATBBufferClosure _cm_satb_cl;
1764 G1CMOopClosure _cm_cl;
1765 MarkingCodeBlobClosure _code_cl;
1766 int _thread_parity;
1767
1768 public:
1769 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1770 _cm_satb_cl(task, g1h),
1771 _cm_cl(g1h, task),
1772 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1773 _thread_parity(Threads::thread_claim_parity()) {}
1774
1775 void do_thread(Thread* thread) {
1776 if (thread->is_Java_thread()) {
1777 if (thread->claim_oops_do(true, _thread_parity)) {
1778 JavaThread* jt = (JavaThread*)thread;
1779
1780 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1781 // however the liveness of oops reachable from nmethods have very complex lifecycles:
1782 // * Alive if on the stack of an executing method
1783 // * Weakly reachable otherwise
1784 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1785 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1786 jt->nmethods_do(&_code_cl);
1787
1788 G1ThreadLocalData::satb_mark_queue(jt).apply_closure_and_empty(&_cm_satb_cl);
1789 }
1790 } else if (thread->is_VM_thread()) {
1791 if (thread->claim_oops_do(true, _thread_parity)) {
1792 G1BarrierSet::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1793 }
1794 }
1795 }
1796 };
1797
1798 class G1CMRemarkTask : public AbstractGangTask {
1799 G1ConcurrentMark* _cm;
1800 public:
1801 void work(uint worker_id) {
1802 G1CMTask* task = _cm->task(worker_id);
1803 task->record_start_time();
1804 {
1805 ResourceMark rm;
1806 HandleMark hm;
1807
1808 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1809 Threads::threads_do(&threads_f);
1810 }
1811
1812 do {
1813 task->do_marking_step(1000000000.0 /* something very large */,
1814 true /* do_termination */,
1815 false /* is_serial */);
1816 } while (task->has_aborted() && !_cm->has_overflown());
1817 // If we overflow, then we do not want to restart. We instead
1818 // want to abort remark and do concurrent marking again.
1819 task->record_end_time();
1820 }
1821
1822 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1823 AbstractGangTask("Par Remark"), _cm(cm) {
1824 _cm->terminator()->reset_for_reuse(active_workers);
1825 }
1826 };
1827
1828 void G1ConcurrentMark::finalize_marking() {
1829 ResourceMark rm;
1830 HandleMark hm;
1831
1832 _g1h->ensure_parsability(false);
1833
1834 // this is remark, so we'll use up all active threads
1835 uint active_workers = _g1h->workers()->active_workers();
1836 set_concurrency_and_phase(active_workers, false /* concurrent */);
1837 // Leave _parallel_marking_threads at it's
1838 // value originally calculated in the G1ConcurrentMark
1839 // constructor and pass values of the active workers
1840 // through the gang in the task.
1841
1842 {
1843 StrongRootsScope srs(active_workers);
1844
1845 G1CMRemarkTask remarkTask(this, active_workers);
1846 // We will start all available threads, even if we decide that the
1847 // active_workers will be fewer. The extra ones will just bail out
1848 // immediately.
1849 _g1h->workers()->run_task(&remarkTask);
1850 }
1851
1852 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1853 guarantee(has_overflown() ||
1854 satb_mq_set.completed_buffers_num() == 0,
1855 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1856 BOOL_TO_STR(has_overflown()),
1857 satb_mq_set.completed_buffers_num());
1858
1859 print_stats();
1860 }
1861
1862 void G1ConcurrentMark::flush_all_task_caches() {
1863 size_t hits = 0;
1864 size_t misses = 0;
1865 for (uint i = 0; i < _max_num_tasks; i++) {
1866 Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache();
1867 hits += stats.first;
1868 misses += stats.second;
1869 }
1870 size_t sum = hits + misses;
1871 log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf",
1872 hits, misses, percent_of(hits, sum));
1873 }
1874
1875 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1876 _prev_mark_bitmap->clear_range(mr);
1877 }
1878
1879 HeapRegion*
1880 G1ConcurrentMark::claim_region(uint worker_id) {
1881 // "checkpoint" the finger
1882 HeapWord* finger = _finger;
1883
1884 while (finger < _heap.end()) {
1885 assert(_g1h->is_in_g1_reserved(finger), "invariant");
1886
1887 HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1888 // Make sure that the reads below do not float before loading curr_region.
1889 OrderAccess::loadload();
1890 // Above heap_region_containing may return NULL as we always scan claim
1891 // until the end of the heap. In this case, just jump to the next region.
1892 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1893
1894 // Is the gap between reading the finger and doing the CAS too long?
1895 HeapWord* res = Atomic::cmpxchg(end, &_finger, finger);
1896 if (res == finger && curr_region != NULL) {
1897 // we succeeded
1898 HeapWord* bottom = curr_region->bottom();
1899 HeapWord* limit = curr_region->next_top_at_mark_start();
1900
1901 // notice that _finger == end cannot be guaranteed here since,
1902 // someone else might have moved the finger even further
1903 assert(_finger >= end, "the finger should have moved forward");
1904
1905 if (limit > bottom) {
1906 return curr_region;
1907 } else {
1908 assert(limit == bottom,
1909 "the region limit should be at bottom");
1910 // we return NULL and the caller should try calling
1911 // claim_region() again.
1912 return NULL;
1913 }
1914 } else {
1915 assert(_finger > finger, "the finger should have moved forward");
1916 // read it again
1917 finger = _finger;
1918 }
1919 }
1920
1921 return NULL;
1922 }
1923
1924 #ifndef PRODUCT
1925 class VerifyNoCSetOops {
1926 G1CollectedHeap* _g1h;
1927 const char* _phase;
1928 int _info;
1929
1930 public:
1931 VerifyNoCSetOops(const char* phase, int info = -1) :
1932 _g1h(G1CollectedHeap::heap()),
1933 _phase(phase),
1934 _info(info)
1935 { }
1936
1937 void operator()(G1TaskQueueEntry task_entry) const {
1938 if (task_entry.is_array_slice()) {
1939 guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1940 return;
1941 }
1942 guarantee(oopDesc::is_oop(task_entry.obj()),
1943 "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1944 p2i(task_entry.obj()), _phase, _info);
1945 guarantee(!_g1h->is_in_cset(task_entry.obj()),
1946 "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1947 p2i(task_entry.obj()), _phase, _info);
1948 }
1949 };
1950
1951 void G1ConcurrentMark::verify_no_cset_oops() {
1952 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1953 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
1954 return;
1955 }
1956
1957 // Verify entries on the global mark stack
1958 _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1959
1960 // Verify entries on the task queues
1961 for (uint i = 0; i < _max_num_tasks; ++i) {
1962 G1CMTaskQueue* queue = _task_queues->queue(i);
1963 queue->iterate(VerifyNoCSetOops("Queue", i));
1964 }
1965
1966 // Verify the global finger
1967 HeapWord* global_finger = finger();
1968 if (global_finger != NULL && global_finger < _heap.end()) {
1969 // Since we always iterate over all regions, we might get a NULL HeapRegion
1970 // here.
1971 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1972 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1973 "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1974 p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1975 }
1976
1977 // Verify the task fingers
1978 assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
1979 for (uint i = 0; i < _num_concurrent_workers; ++i) {
1980 G1CMTask* task = _tasks[i];
1981 HeapWord* task_finger = task->finger();
1982 if (task_finger != NULL && task_finger < _heap.end()) {
1983 // See above note on the global finger verification.
1984 HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
1985 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
1986 !task_hr->in_collection_set(),
1987 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
1988 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
1989 }
1990 }
1991 }
1992 #endif // PRODUCT
1993
1994 void G1ConcurrentMark::rebuild_rem_set_concurrently() {
1995 _g1h->g1_rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset);
1996 }
1997
1998 void G1ConcurrentMark::print_stats() {
1999 if (!log_is_enabled(Debug, gc, stats)) {
2000 return;
2001 }
2002 log_debug(gc, stats)("---------------------------------------------------------------------");
2003 for (size_t i = 0; i < _num_active_tasks; ++i) {
2004 _tasks[i]->print_stats();
2005 log_debug(gc, stats)("---------------------------------------------------------------------");
2006 }
2007 }
2008
2009 void G1ConcurrentMark::concurrent_cycle_abort() {
2010 if (!cm_thread()->during_cycle() || _has_aborted) {
2011 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2012 return;
2013 }
2014
2015 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2016 // concurrent bitmap clearing.
2017 {
2018 GCTraceTime(Debug, gc) debug("Clear Next Bitmap");
2019 clear_bitmap(_next_mark_bitmap, _g1h->workers(), false);
2020 }
2021 // Note we cannot clear the previous marking bitmap here
2022 // since VerifyDuringGC verifies the objects marked during
2023 // a full GC against the previous bitmap.
2024
2025 // Empty mark stack
2026 reset_marking_for_restart();
2027 for (uint i = 0; i < _max_num_tasks; ++i) {
2028 _tasks[i]->clear_region_fields();
2029 }
2030 _first_overflow_barrier_sync.abort();
2031 _second_overflow_barrier_sync.abort();
2032 _has_aborted = true;
2033
2034 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2035 satb_mq_set.abandon_partial_marking();
2036 // This can be called either during or outside marking, we'll read
2037 // the expected_active value from the SATB queue set.
2038 satb_mq_set.set_active_all_threads(
2039 false, /* new active value */
2040 satb_mq_set.is_active() /* expected_active */);
2041 }
2042
2043 static void print_ms_time_info(const char* prefix, const char* name,
2044 NumberSeq& ns) {
2045 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2046 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2047 if (ns.num() > 0) {
2048 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]",
2049 prefix, ns.sd(), ns.maximum());
2050 }
2051 }
2052
2053 void G1ConcurrentMark::print_summary_info() {
2054 Log(gc, marking) log;
2055 if (!log.is_trace()) {
2056 return;
2057 }
2058
2059 log.trace(" Concurrent marking:");
2060 print_ms_time_info(" ", "init marks", _init_times);
2061 print_ms_time_info(" ", "remarks", _remark_times);
2062 {
2063 print_ms_time_info(" ", "final marks", _remark_mark_times);
2064 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
2065
2066 }
2067 print_ms_time_info(" ", "cleanups", _cleanup_times);
2068 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2069 _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2070 log.trace(" Total stop_world time = %8.2f s.",
2071 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2072 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).",
2073 cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2074 }
2075
2076 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2077 _concurrent_workers->print_worker_threads_on(st);
2078 }
2079
2080 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2081 _concurrent_workers->threads_do(tc);
2082 }
2083
2084 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2085 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2086 p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2087 _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2088 _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2089 }
2090
2091 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2092 ReferenceProcessor* result = g1h->ref_processor_cm();
2093 assert(result != NULL, "CM reference processor should not be NULL");
2094 return result;
2095 }
2096
2097 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2098 G1CMTask* task)
2099 : MetadataVisitingOopIterateClosure(get_cm_oop_closure_ref_processor(g1h)),
2100 _g1h(g1h), _task(task)
2101 { }
2102
2103 void G1CMTask::setup_for_region(HeapRegion* hr) {
2104 assert(hr != NULL,
2105 "claim_region() should have filtered out NULL regions");
2106 _curr_region = hr;
2107 _finger = hr->bottom();
2108 update_region_limit();
2109 }
2110
2111 void G1CMTask::update_region_limit() {
2112 HeapRegion* hr = _curr_region;
2113 HeapWord* bottom = hr->bottom();
2114 HeapWord* limit = hr->next_top_at_mark_start();
2115
2116 if (limit == bottom) {
2117 // The region was collected underneath our feet.
2118 // We set the finger to bottom to ensure that the bitmap
2119 // iteration that will follow this will not do anything.
2120 // (this is not a condition that holds when we set the region up,
2121 // as the region is not supposed to be empty in the first place)
2122 _finger = bottom;
2123 } else if (limit >= _region_limit) {
2124 assert(limit >= _finger, "peace of mind");
2125 } else {
2126 assert(limit < _region_limit, "only way to get here");
2127 // This can happen under some pretty unusual circumstances. An
2128 // evacuation pause empties the region underneath our feet (NTAMS
2129 // at bottom). We then do some allocation in the region (NTAMS
2130 // stays at bottom), followed by the region being used as a GC
2131 // alloc region (NTAMS will move to top() and the objects
2132 // originally below it will be grayed). All objects now marked in
2133 // the region are explicitly grayed, if below the global finger,
2134 // and we do not need in fact to scan anything else. So, we simply
2135 // set _finger to be limit to ensure that the bitmap iteration
2136 // doesn't do anything.
2137 _finger = limit;
2138 }
2139
2140 _region_limit = limit;
2141 }
2142
2143 void G1CMTask::giveup_current_region() {
2144 assert(_curr_region != NULL, "invariant");
2145 clear_region_fields();
2146 }
2147
2148 void G1CMTask::clear_region_fields() {
2149 // Values for these three fields that indicate that we're not
2150 // holding on to a region.
2151 _curr_region = NULL;
2152 _finger = NULL;
2153 _region_limit = NULL;
2154 }
2155
2156 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2157 if (cm_oop_closure == NULL) {
2158 assert(_cm_oop_closure != NULL, "invariant");
2159 } else {
2160 assert(_cm_oop_closure == NULL, "invariant");
2161 }
2162 _cm_oop_closure = cm_oop_closure;
2163 }
2164
2165 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2166 guarantee(next_mark_bitmap != NULL, "invariant");
2167 _next_mark_bitmap = next_mark_bitmap;
2168 clear_region_fields();
2169
2170 _calls = 0;
2171 _elapsed_time_ms = 0.0;
2172 _termination_time_ms = 0.0;
2173 _termination_start_time_ms = 0.0;
2174
2175 _mark_stats_cache.reset();
2176 }
2177
2178 bool G1CMTask::should_exit_termination() {
2179 if (!regular_clock_call()) {
2180 return true;
2181 }
2182
2183 // This is called when we are in the termination protocol. We should
2184 // quit if, for some reason, this task wants to abort or the global
2185 // stack is not empty (this means that we can get work from it).
2186 return !_cm->mark_stack_empty() || has_aborted();
2187 }
2188
2189 void G1CMTask::reached_limit() {
2190 assert(_words_scanned >= _words_scanned_limit ||
2191 _refs_reached >= _refs_reached_limit ,
2192 "shouldn't have been called otherwise");
2193 abort_marking_if_regular_check_fail();
2194 }
2195
2196 bool G1CMTask::regular_clock_call() {
2197 if (has_aborted()) {
2198 return false;
2199 }
2200
2201 // First, we need to recalculate the words scanned and refs reached
2202 // limits for the next clock call.
2203 recalculate_limits();
2204
2205 // During the regular clock call we do the following
2206
2207 // (1) If an overflow has been flagged, then we abort.
2208 if (_cm->has_overflown()) {
2209 return false;
2210 }
2211
2212 // If we are not concurrent (i.e. we're doing remark) we don't need
2213 // to check anything else. The other steps are only needed during
2214 // the concurrent marking phase.
2215 if (!_cm->concurrent()) {
2216 return true;
2217 }
2218
2219 // (2) If marking has been aborted for Full GC, then we also abort.
2220 if (_cm->has_aborted()) {
2221 return false;
2222 }
2223
2224 double curr_time_ms = os::elapsedVTime() * 1000.0;
2225
2226 // (4) We check whether we should yield. If we have to, then we abort.
2227 if (SuspendibleThreadSet::should_yield()) {
2228 // We should yield. To do this we abort the task. The caller is
2229 // responsible for yielding.
2230 return false;
2231 }
2232
2233 // (5) We check whether we've reached our time quota. If we have,
2234 // then we abort.
2235 double elapsed_time_ms = curr_time_ms - _start_time_ms;
2236 if (elapsed_time_ms > _time_target_ms) {
2237 _has_timed_out = true;
2238 return false;
2239 }
2240
2241 // (6) Finally, we check whether there are enough completed STAB
2242 // buffers available for processing. If there are, we abort.
2243 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2244 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2245 // we do need to process SATB buffers, we'll abort and restart
2246 // the marking task to do so
2247 return false;
2248 }
2249 return true;
2250 }
2251
2252 void G1CMTask::recalculate_limits() {
2253 _real_words_scanned_limit = _words_scanned + words_scanned_period;
2254 _words_scanned_limit = _real_words_scanned_limit;
2255
2256 _real_refs_reached_limit = _refs_reached + refs_reached_period;
2257 _refs_reached_limit = _real_refs_reached_limit;
2258 }
2259
2260 void G1CMTask::decrease_limits() {
2261 // This is called when we believe that we're going to do an infrequent
2262 // operation which will increase the per byte scanned cost (i.e. move
2263 // entries to/from the global stack). It basically tries to decrease the
2264 // scanning limit so that the clock is called earlier.
2265
2266 _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2267 _refs_reached_limit = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2268 }
2269
2270 void G1CMTask::move_entries_to_global_stack() {
2271 // Local array where we'll store the entries that will be popped
2272 // from the local queue.
2273 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2274
2275 size_t n = 0;
2276 G1TaskQueueEntry task_entry;
2277 while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2278 buffer[n] = task_entry;
2279 ++n;
2280 }
2281 if (n < G1CMMarkStack::EntriesPerChunk) {
2282 buffer[n] = G1TaskQueueEntry();
2283 }
2284
2285 if (n > 0) {
2286 if (!_cm->mark_stack_push(buffer)) {
2287 set_has_aborted();
2288 }
2289 }
2290
2291 // This operation was quite expensive, so decrease the limits.
2292 decrease_limits();
2293 }
2294
2295 bool G1CMTask::get_entries_from_global_stack() {
2296 // Local array where we'll store the entries that will be popped
2297 // from the global stack.
2298 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2299
2300 if (!_cm->mark_stack_pop(buffer)) {
2301 return false;
2302 }
2303
2304 // We did actually pop at least one entry.
2305 for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2306 G1TaskQueueEntry task_entry = buffer[i];
2307 if (task_entry.is_null()) {
2308 break;
2309 }
2310 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()));
2311 bool success = _task_queue->push(task_entry);
2312 // We only call this when the local queue is empty or under a
2313 // given target limit. So, we do not expect this push to fail.
2314 assert(success, "invariant");
2315 }
2316
2317 // This operation was quite expensive, so decrease the limits
2318 decrease_limits();
2319 return true;
2320 }
2321
2322 void G1CMTask::drain_local_queue(bool partially) {
2323 if (has_aborted()) {
2324 return;
2325 }
2326
2327 // Decide what the target size is, depending whether we're going to
2328 // drain it partially (so that other tasks can steal if they run out
2329 // of things to do) or totally (at the very end).
2330 size_t target_size;
2331 if (partially) {
2332 target_size = MIN2((size_t)_task_queue->max_elems()/3, (size_t)GCDrainStackTargetSize);
2333 } else {
2334 target_size = 0;
2335 }
2336
2337 if (_task_queue->size() > target_size) {
2338 G1TaskQueueEntry entry;
2339 bool ret = _task_queue->pop_local(entry);
2340 while (ret) {
2341 scan_task_entry(entry);
2342 if (_task_queue->size() <= target_size || has_aborted()) {
2343 ret = false;
2344 } else {
2345 ret = _task_queue->pop_local(entry);
2346 }
2347 }
2348 }
2349 }
2350
2351 void G1CMTask::drain_global_stack(bool partially) {
2352 if (has_aborted()) {
2353 return;
2354 }
2355
2356 // We have a policy to drain the local queue before we attempt to
2357 // drain the global stack.
2358 assert(partially || _task_queue->size() == 0, "invariant");
2359
2360 // Decide what the target size is, depending whether we're going to
2361 // drain it partially (so that other tasks can steal if they run out
2362 // of things to do) or totally (at the very end).
2363 // Notice that when draining the global mark stack partially, due to the racyness
2364 // of the mark stack size update we might in fact drop below the target. But,
2365 // this is not a problem.
2366 // In case of total draining, we simply process until the global mark stack is
2367 // totally empty, disregarding the size counter.
2368 if (partially) {
2369 size_t const target_size = _cm->partial_mark_stack_size_target();
2370 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2371 if (get_entries_from_global_stack()) {
2372 drain_local_queue(partially);
2373 }
2374 }
2375 } else {
2376 while (!has_aborted() && get_entries_from_global_stack()) {
2377 drain_local_queue(partially);
2378 }
2379 }
2380 }
2381
2382 // SATB Queue has several assumptions on whether to call the par or
2383 // non-par versions of the methods. this is why some of the code is
2384 // replicated. We should really get rid of the single-threaded version
2385 // of the code to simplify things.
2386 void G1CMTask::drain_satb_buffers() {
2387 if (has_aborted()) {
2388 return;
2389 }
2390
2391 // We set this so that the regular clock knows that we're in the
2392 // middle of draining buffers and doesn't set the abort flag when it
2393 // notices that SATB buffers are available for draining. It'd be
2394 // very counter productive if it did that. :-)
2395 _draining_satb_buffers = true;
2396
2397 G1CMSATBBufferClosure satb_cl(this, _g1h);
2398 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2399
2400 // This keeps claiming and applying the closure to completed buffers
2401 // until we run out of buffers or we need to abort.
2402 while (!has_aborted() &&
2403 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2404 abort_marking_if_regular_check_fail();
2405 }
2406
2407 _draining_satb_buffers = false;
2408
2409 assert(has_aborted() ||
2410 _cm->concurrent() ||
2411 satb_mq_set.completed_buffers_num() == 0, "invariant");
2412
2413 // again, this was a potentially expensive operation, decrease the
2414 // limits to get the regular clock call early
2415 decrease_limits();
2416 }
2417
2418 void G1CMTask::clear_mark_stats_cache(uint region_idx) {
2419 _mark_stats_cache.reset(region_idx);
2420 }
2421
2422 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() {
2423 return _mark_stats_cache.evict_all();
2424 }
2425
2426 void G1CMTask::print_stats() {
2427 log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls);
2428 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2429 _elapsed_time_ms, _termination_time_ms);
2430 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms",
2431 _step_times_ms.num(),
2432 _step_times_ms.avg(),
2433 _step_times_ms.sd(),
2434 _step_times_ms.maximum(),
2435 _step_times_ms.sum());
2436 size_t const hits = _mark_stats_cache.hits();
2437 size_t const misses = _mark_stats_cache.misses();
2438 log_debug(gc, stats)(" Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f",
2439 hits, misses, percent_of(hits, hits + misses));
2440 }
2441
2442 bool G1ConcurrentMark::try_stealing(uint worker_id, G1TaskQueueEntry& task_entry) {
2443 return _task_queues->steal(worker_id, task_entry);
2444 }
2445
2446 /*****************************************************************************
2447
2448 The do_marking_step(time_target_ms, ...) method is the building
2449 block of the parallel marking framework. It can be called in parallel
2450 with other invocations of do_marking_step() on different tasks
2451 (but only one per task, obviously) and concurrently with the
2452 mutator threads, or during remark, hence it eliminates the need
2453 for two versions of the code. When called during remark, it will
2454 pick up from where the task left off during the concurrent marking
2455 phase. Interestingly, tasks are also claimable during evacuation
2456 pauses too, since do_marking_step() ensures that it aborts before
2457 it needs to yield.
2458
2459 The data structures that it uses to do marking work are the
2460 following:
2461
2462 (1) Marking Bitmap. If there are gray objects that appear only
2463 on the bitmap (this happens either when dealing with an overflow
2464 or when the initial marking phase has simply marked the roots
2465 and didn't push them on the stack), then tasks claim heap
2466 regions whose bitmap they then scan to find gray objects. A
2467 global finger indicates where the end of the last claimed region
2468 is. A local finger indicates how far into the region a task has
2469 scanned. The two fingers are used to determine how to gray an
2470 object (i.e. whether simply marking it is OK, as it will be
2471 visited by a task in the future, or whether it needs to be also
2472 pushed on a stack).
2473
2474 (2) Local Queue. The local queue of the task which is accessed
2475 reasonably efficiently by the task. Other tasks can steal from
2476 it when they run out of work. Throughout the marking phase, a
2477 task attempts to keep its local queue short but not totally
2478 empty, so that entries are available for stealing by other
2479 tasks. Only when there is no more work, a task will totally
2480 drain its local queue.
2481
2482 (3) Global Mark Stack. This handles local queue overflow. During
2483 marking only sets of entries are moved between it and the local
2484 queues, as access to it requires a mutex and more fine-grain
2485 interaction with it which might cause contention. If it
2486 overflows, then the marking phase should restart and iterate
2487 over the bitmap to identify gray objects. Throughout the marking
2488 phase, tasks attempt to keep the global mark stack at a small
2489 length but not totally empty, so that entries are available for
2490 popping by other tasks. Only when there is no more work, tasks
2491 will totally drain the global mark stack.
2492
2493 (4) SATB Buffer Queue. This is where completed SATB buffers are
2494 made available. Buffers are regularly removed from this queue
2495 and scanned for roots, so that the queue doesn't get too
2496 long. During remark, all completed buffers are processed, as
2497 well as the filled in parts of any uncompleted buffers.
2498
2499 The do_marking_step() method tries to abort when the time target
2500 has been reached. There are a few other cases when the
2501 do_marking_step() method also aborts:
2502
2503 (1) When the marking phase has been aborted (after a Full GC).
2504
2505 (2) When a global overflow (on the global stack) has been
2506 triggered. Before the task aborts, it will actually sync up with
2507 the other tasks to ensure that all the marking data structures
2508 (local queues, stacks, fingers etc.) are re-initialized so that
2509 when do_marking_step() completes, the marking phase can
2510 immediately restart.
2511
2512 (3) When enough completed SATB buffers are available. The
2513 do_marking_step() method only tries to drain SATB buffers right
2514 at the beginning. So, if enough buffers are available, the
2515 marking step aborts and the SATB buffers are processed at
2516 the beginning of the next invocation.
2517
2518 (4) To yield. when we have to yield then we abort and yield
2519 right at the end of do_marking_step(). This saves us from a lot
2520 of hassle as, by yielding we might allow a Full GC. If this
2521 happens then objects will be compacted underneath our feet, the
2522 heap might shrink, etc. We save checking for this by just
2523 aborting and doing the yield right at the end.
2524
2525 From the above it follows that the do_marking_step() method should
2526 be called in a loop (or, otherwise, regularly) until it completes.
2527
2528 If a marking step completes without its has_aborted() flag being
2529 true, it means it has completed the current marking phase (and
2530 also all other marking tasks have done so and have all synced up).
2531
2532 A method called regular_clock_call() is invoked "regularly" (in
2533 sub ms intervals) throughout marking. It is this clock method that
2534 checks all the abort conditions which were mentioned above and
2535 decides when the task should abort. A work-based scheme is used to
2536 trigger this clock method: when the number of object words the
2537 marking phase has scanned or the number of references the marking
2538 phase has visited reach a given limit. Additional invocations to
2539 the method clock have been planted in a few other strategic places
2540 too. The initial reason for the clock method was to avoid calling
2541 vtime too regularly, as it is quite expensive. So, once it was in
2542 place, it was natural to piggy-back all the other conditions on it
2543 too and not constantly check them throughout the code.
2544
2545 If do_termination is true then do_marking_step will enter its
2546 termination protocol.
2547
2548 The value of is_serial must be true when do_marking_step is being
2549 called serially (i.e. by the VMThread) and do_marking_step should
2550 skip any synchronization in the termination and overflow code.
2551 Examples include the serial remark code and the serial reference
2552 processing closures.
2553
2554 The value of is_serial must be false when do_marking_step is
2555 being called by any of the worker threads in a work gang.
2556 Examples include the concurrent marking code (CMMarkingTask),
2557 the MT remark code, and the MT reference processing closures.
2558
2559 *****************************************************************************/
2560
2561 void G1CMTask::do_marking_step(double time_target_ms,
2562 bool do_termination,
2563 bool is_serial) {
2564 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2565
2566 _start_time_ms = os::elapsedVTime() * 1000.0;
2567
2568 // If do_stealing is true then do_marking_step will attempt to
2569 // steal work from the other G1CMTasks. It only makes sense to
2570 // enable stealing when the termination protocol is enabled
2571 // and do_marking_step() is not being called serially.
2572 bool do_stealing = do_termination && !is_serial;
2573
2574 double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2575 _time_target_ms = time_target_ms - diff_prediction_ms;
2576
2577 // set up the variables that are used in the work-based scheme to
2578 // call the regular clock method
2579 _words_scanned = 0;
2580 _refs_reached = 0;
2581 recalculate_limits();
2582
2583 // clear all flags
2584 clear_has_aborted();
2585 _has_timed_out = false;
2586 _draining_satb_buffers = false;
2587
2588 ++_calls;
2589
2590 // Set up the bitmap and oop closures. Anything that uses them is
2591 // eventually called from this method, so it is OK to allocate these
2592 // statically.
2593 G1CMBitMapClosure bitmap_closure(this, _cm);
2594 G1CMOopClosure cm_oop_closure(_g1h, this);
2595 set_cm_oop_closure(&cm_oop_closure);
2596
2597 if (_cm->has_overflown()) {
2598 // This can happen if the mark stack overflows during a GC pause
2599 // and this task, after a yield point, restarts. We have to abort
2600 // as we need to get into the overflow protocol which happens
2601 // right at the end of this task.
2602 set_has_aborted();
2603 }
2604
2605 // First drain any available SATB buffers. After this, we will not
2606 // look at SATB buffers before the next invocation of this method.
2607 // If enough completed SATB buffers are queued up, the regular clock
2608 // will abort this task so that it restarts.
2609 drain_satb_buffers();
2610 // ...then partially drain the local queue and the global stack
2611 drain_local_queue(true);
2612 drain_global_stack(true);
2613
2614 do {
2615 if (!has_aborted() && _curr_region != NULL) {
2616 // This means that we're already holding on to a region.
2617 assert(_finger != NULL, "if region is not NULL, then the finger "
2618 "should not be NULL either");
2619
2620 // We might have restarted this task after an evacuation pause
2621 // which might have evacuated the region we're holding on to
2622 // underneath our feet. Let's read its limit again to make sure
2623 // that we do not iterate over a region of the heap that
2624 // contains garbage (update_region_limit() will also move
2625 // _finger to the start of the region if it is found empty).
2626 update_region_limit();
2627 // We will start from _finger not from the start of the region,
2628 // as we might be restarting this task after aborting half-way
2629 // through scanning this region. In this case, _finger points to
2630 // the address where we last found a marked object. If this is a
2631 // fresh region, _finger points to start().
2632 MemRegion mr = MemRegion(_finger, _region_limit);
2633
2634 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2635 "humongous regions should go around loop once only");
2636
2637 // Some special cases:
2638 // If the memory region is empty, we can just give up the region.
2639 // If the current region is humongous then we only need to check
2640 // the bitmap for the bit associated with the start of the object,
2641 // scan the object if it's live, and give up the region.
2642 // Otherwise, let's iterate over the bitmap of the part of the region
2643 // that is left.
2644 // If the iteration is successful, give up the region.
2645 if (mr.is_empty()) {
2646 giveup_current_region();
2647 abort_marking_if_regular_check_fail();
2648 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2649 if (_next_mark_bitmap->is_marked(mr.start())) {
2650 // The object is marked - apply the closure
2651 bitmap_closure.do_addr(mr.start());
2652 }
2653 // Even if this task aborted while scanning the humongous object
2654 // we can (and should) give up the current region.
2655 giveup_current_region();
2656 abort_marking_if_regular_check_fail();
2657 } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2658 giveup_current_region();
2659 abort_marking_if_regular_check_fail();
2660 } else {
2661 assert(has_aborted(), "currently the only way to do so");
2662 // The only way to abort the bitmap iteration is to return
2663 // false from the do_bit() method. However, inside the
2664 // do_bit() method we move the _finger to point to the
2665 // object currently being looked at. So, if we bail out, we
2666 // have definitely set _finger to something non-null.
2667 assert(_finger != NULL, "invariant");
2668
2669 // Region iteration was actually aborted. So now _finger
2670 // points to the address of the object we last scanned. If we
2671 // leave it there, when we restart this task, we will rescan
2672 // the object. It is easy to avoid this. We move the finger by
2673 // enough to point to the next possible object header.
2674 assert(_finger < _region_limit, "invariant");
2675 HeapWord* const new_finger = _finger + ((oop)_finger)->size();
2676 // Check if bitmap iteration was aborted while scanning the last object
2677 if (new_finger >= _region_limit) {
2678 giveup_current_region();
2679 } else {
2680 move_finger_to(new_finger);
2681 }
2682 }
2683 }
2684 // At this point we have either completed iterating over the
2685 // region we were holding on to, or we have aborted.
2686
2687 // We then partially drain the local queue and the global stack.
2688 // (Do we really need this?)
2689 drain_local_queue(true);
2690 drain_global_stack(true);
2691
2692 // Read the note on the claim_region() method on why it might
2693 // return NULL with potentially more regions available for
2694 // claiming and why we have to check out_of_regions() to determine
2695 // whether we're done or not.
2696 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2697 // We are going to try to claim a new region. We should have
2698 // given up on the previous one.
2699 // Separated the asserts so that we know which one fires.
2700 assert(_curr_region == NULL, "invariant");
2701 assert(_finger == NULL, "invariant");
2702 assert(_region_limit == NULL, "invariant");
2703 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2704 if (claimed_region != NULL) {
2705 // Yes, we managed to claim one
2706 setup_for_region(claimed_region);
2707 assert(_curr_region == claimed_region, "invariant");
2708 }
2709 // It is important to call the regular clock here. It might take
2710 // a while to claim a region if, for example, we hit a large
2711 // block of empty regions. So we need to call the regular clock
2712 // method once round the loop to make sure it's called
2713 // frequently enough.
2714 abort_marking_if_regular_check_fail();
2715 }
2716
2717 if (!has_aborted() && _curr_region == NULL) {
2718 assert(_cm->out_of_regions(),
2719 "at this point we should be out of regions");
2720 }
2721 } while ( _curr_region != NULL && !has_aborted());
2722
2723 if (!has_aborted()) {
2724 // We cannot check whether the global stack is empty, since other
2725 // tasks might be pushing objects to it concurrently.
2726 assert(_cm->out_of_regions(),
2727 "at this point we should be out of regions");
2728 // Try to reduce the number of available SATB buffers so that
2729 // remark has less work to do.
2730 drain_satb_buffers();
2731 }
2732
2733 // Since we've done everything else, we can now totally drain the
2734 // local queue and global stack.
2735 drain_local_queue(false);
2736 drain_global_stack(false);
2737
2738 // Attempt at work stealing from other task's queues.
2739 if (do_stealing && !has_aborted()) {
2740 // We have not aborted. This means that we have finished all that
2741 // we could. Let's try to do some stealing...
2742
2743 // We cannot check whether the global stack is empty, since other
2744 // tasks might be pushing objects to it concurrently.
2745 assert(_cm->out_of_regions() && _task_queue->size() == 0,
2746 "only way to reach here");
2747 while (!has_aborted()) {
2748 G1TaskQueueEntry entry;
2749 if (_cm->try_stealing(_worker_id, entry)) {
2750 scan_task_entry(entry);
2751
2752 // And since we're towards the end, let's totally drain the
2753 // local queue and global stack.
2754 drain_local_queue(false);
2755 drain_global_stack(false);
2756 } else {
2757 break;
2758 }
2759 }
2760 }
2761
2762 // We still haven't aborted. Now, let's try to get into the
2763 // termination protocol.
2764 if (do_termination && !has_aborted()) {
2765 // We cannot check whether the global stack is empty, since other
2766 // tasks might be concurrently pushing objects on it.
2767 // Separated the asserts so that we know which one fires.
2768 assert(_cm->out_of_regions(), "only way to reach here");
2769 assert(_task_queue->size() == 0, "only way to reach here");
2770 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2771
2772 // The G1CMTask class also extends the TerminatorTerminator class,
2773 // hence its should_exit_termination() method will also decide
2774 // whether to exit the termination protocol or not.
2775 bool finished = (is_serial ||
2776 _cm->terminator()->offer_termination(this));
2777 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2778 _termination_time_ms +=
2779 termination_end_time_ms - _termination_start_time_ms;
2780
2781 if (finished) {
2782 // We're all done.
2783
2784 // We can now guarantee that the global stack is empty, since
2785 // all other tasks have finished. We separated the guarantees so
2786 // that, if a condition is false, we can immediately find out
2787 // which one.
2788 guarantee(_cm->out_of_regions(), "only way to reach here");
2789 guarantee(_cm->mark_stack_empty(), "only way to reach here");
2790 guarantee(_task_queue->size() == 0, "only way to reach here");
2791 guarantee(!_cm->has_overflown(), "only way to reach here");
2792 guarantee(!has_aborted(), "should never happen if termination has completed");
2793 } else {
2794 // Apparently there's more work to do. Let's abort this task. It
2795 // will restart it and we can hopefully find more things to do.
2796 set_has_aborted();
2797 }
2798 }
2799
2800 // Mainly for debugging purposes to make sure that a pointer to the
2801 // closure which was statically allocated in this frame doesn't
2802 // escape it by accident.
2803 set_cm_oop_closure(NULL);
2804 double end_time_ms = os::elapsedVTime() * 1000.0;
2805 double elapsed_time_ms = end_time_ms - _start_time_ms;
2806 // Update the step history.
2807 _step_times_ms.add(elapsed_time_ms);
2808
2809 if (has_aborted()) {
2810 // The task was aborted for some reason.
2811 if (_has_timed_out) {
2812 double diff_ms = elapsed_time_ms - _time_target_ms;
2813 // Keep statistics of how well we did with respect to hitting
2814 // our target only if we actually timed out (if we aborted for
2815 // other reasons, then the results might get skewed).
2816 _marking_step_diffs_ms.add(diff_ms);
2817 }
2818
2819 if (_cm->has_overflown()) {
2820 // This is the interesting one. We aborted because a global
2821 // overflow was raised. This means we have to restart the
2822 // marking phase and start iterating over regions. However, in
2823 // order to do this we have to make sure that all tasks stop
2824 // what they are doing and re-initialize in a safe manner. We
2825 // will achieve this with the use of two barrier sync points.
2826
2827 if (!is_serial) {
2828 // We only need to enter the sync barrier if being called
2829 // from a parallel context
2830 _cm->enter_first_sync_barrier(_worker_id);
2831
2832 // When we exit this sync barrier we know that all tasks have
2833 // stopped doing marking work. So, it's now safe to
2834 // re-initialize our data structures.
2835 }
2836
2837 clear_region_fields();
2838 flush_mark_stats_cache();
2839
2840 if (!is_serial) {
2841 // If we're executing the concurrent phase of marking, reset the marking
2842 // state; otherwise the marking state is reset after reference processing,
2843 // during the remark pause.
2844 // If we reset here as a result of an overflow during the remark we will
2845 // see assertion failures from any subsequent set_concurrency_and_phase()
2846 // calls.
2847 if (_cm->concurrent() && _worker_id == 0) {
2848 // Worker 0 is responsible for clearing the global data structures because
2849 // of an overflow. During STW we should not clear the overflow flag (in
2850 // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit
2851 // method to abort the pause and restart concurrent marking.
2852 _cm->reset_marking_for_restart();
2853
2854 log_info(gc, marking)("Concurrent Mark reset for overflow");
2855 }
2856
2857 // ...and enter the second barrier.
2858 _cm->enter_second_sync_barrier(_worker_id);
2859 }
2860 // At this point, if we're during the concurrent phase of
2861 // marking, everything has been re-initialized and we're
2862 // ready to restart.
2863 }
2864 }
2865 }
2866
2867 G1CMTask::G1CMTask(uint worker_id,
2868 G1ConcurrentMark* cm,
2869 G1CMTaskQueue* task_queue,
2870 G1RegionMarkStats* mark_stats,
2871 uint max_regions) :
2872 _objArray_processor(this),
2873 _worker_id(worker_id),
2874 _g1h(G1CollectedHeap::heap()),
2875 _cm(cm),
2876 _next_mark_bitmap(NULL),
2877 _task_queue(task_queue),
2878 _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize),
2879 _calls(0),
2880 _time_target_ms(0.0),
2881 _start_time_ms(0.0),
2882 _cm_oop_closure(NULL),
2883 _curr_region(NULL),
2884 _finger(NULL),
2885 _region_limit(NULL),
2886 _words_scanned(0),
2887 _words_scanned_limit(0),
2888 _real_words_scanned_limit(0),
2889 _refs_reached(0),
2890 _refs_reached_limit(0),
2891 _real_refs_reached_limit(0),
2892 _has_aborted(false),
2893 _has_timed_out(false),
2894 _draining_satb_buffers(false),
2895 _step_times_ms(),
2896 _elapsed_time_ms(0.0),
2897 _termination_time_ms(0.0),
2898 _termination_start_time_ms(0.0),
2899 _marking_step_diffs_ms()
2900 {
2901 guarantee(task_queue != NULL, "invariant");
2902
2903 _marking_step_diffs_ms.add(0.5);
2904 }
2905
2906 // These are formatting macros that are used below to ensure
2907 // consistent formatting. The *_H_* versions are used to format the
2908 // header for a particular value and they should be kept consistent
2909 // with the corresponding macro. Also note that most of the macros add
2910 // the necessary white space (as a prefix) which makes them a bit
2911 // easier to compose.
2912
2913 // All the output lines are prefixed with this string to be able to
2914 // identify them easily in a large log file.
2915 #define G1PPRL_LINE_PREFIX "###"
2916
2917 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT
2918 #ifdef _LP64
2919 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
2920 #else // _LP64
2921 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
2922 #endif // _LP64
2923
2924 // For per-region info
2925 #define G1PPRL_TYPE_FORMAT " %-4s"
2926 #define G1PPRL_TYPE_H_FORMAT " %4s"
2927 #define G1PPRL_STATE_FORMAT " %-5s"
2928 #define G1PPRL_STATE_H_FORMAT " %5s"
2929 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9)
2930 #define G1PPRL_BYTE_H_FORMAT " %9s"
2931 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
2932 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
2933
2934 // For summary info
2935 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT
2936 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT
2937 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB"
2938 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2939
2940 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2941 _total_used_bytes(0), _total_capacity_bytes(0),
2942 _total_prev_live_bytes(0), _total_next_live_bytes(0),
2943 _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2944 {
2945 if (!log_is_enabled(Trace, gc, liveness)) {
2946 return;
2947 }
2948
2949 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2950 MemRegion g1_reserved = g1h->g1_reserved();
2951 double now = os::elapsedTime();
2952
2953 // Print the header of the output.
2954 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2955 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2956 G1PPRL_SUM_ADDR_FORMAT("reserved")
2957 G1PPRL_SUM_BYTE_FORMAT("region-size"),
2958 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2959 HeapRegion::GrainBytes);
2960 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2961 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2962 G1PPRL_TYPE_H_FORMAT
2963 G1PPRL_ADDR_BASE_H_FORMAT
2964 G1PPRL_BYTE_H_FORMAT
2965 G1PPRL_BYTE_H_FORMAT
2966 G1PPRL_BYTE_H_FORMAT
2967 G1PPRL_DOUBLE_H_FORMAT
2968 G1PPRL_BYTE_H_FORMAT
2969 G1PPRL_STATE_H_FORMAT
2970 G1PPRL_BYTE_H_FORMAT,
2971 "type", "address-range",
2972 "used", "prev-live", "next-live", "gc-eff",
2973 "remset", "state", "code-roots");
2974 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2975 G1PPRL_TYPE_H_FORMAT
2976 G1PPRL_ADDR_BASE_H_FORMAT
2977 G1PPRL_BYTE_H_FORMAT
2978 G1PPRL_BYTE_H_FORMAT
2979 G1PPRL_BYTE_H_FORMAT
2980 G1PPRL_DOUBLE_H_FORMAT
2981 G1PPRL_BYTE_H_FORMAT
2982 G1PPRL_STATE_H_FORMAT
2983 G1PPRL_BYTE_H_FORMAT,
2984 "", "",
2985 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
2986 "(bytes)", "", "(bytes)");
2987 }
2988
2989 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) {
2990 if (!log_is_enabled(Trace, gc, liveness)) {
2991 return false;
2992 }
2993
2994 const char* type = r->get_type_str();
2995 HeapWord* bottom = r->bottom();
2996 HeapWord* end = r->end();
2997 size_t capacity_bytes = r->capacity();
2998 size_t used_bytes = r->used();
2999 size_t prev_live_bytes = r->live_bytes();
3000 size_t next_live_bytes = r->next_live_bytes();
3001 double gc_eff = r->gc_efficiency();
3002 size_t remset_bytes = r->rem_set()->mem_size();
3003 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3004 const char* remset_type = r->rem_set()->get_short_state_str();
3005
3006 _total_used_bytes += used_bytes;
3007 _total_capacity_bytes += capacity_bytes;
3008 _total_prev_live_bytes += prev_live_bytes;
3009 _total_next_live_bytes += next_live_bytes;
3010 _total_remset_bytes += remset_bytes;
3011 _total_strong_code_roots_bytes += strong_code_roots_bytes;
3012
3013 // Print a line for this particular region.
3014 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3015 G1PPRL_TYPE_FORMAT
3016 G1PPRL_ADDR_BASE_FORMAT
3017 G1PPRL_BYTE_FORMAT
3018 G1PPRL_BYTE_FORMAT
3019 G1PPRL_BYTE_FORMAT
3020 G1PPRL_DOUBLE_FORMAT
3021 G1PPRL_BYTE_FORMAT
3022 G1PPRL_STATE_FORMAT
3023 G1PPRL_BYTE_FORMAT,
3024 type, p2i(bottom), p2i(end),
3025 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3026 remset_bytes, remset_type, strong_code_roots_bytes);
3027
3028 return false;
3029 }
3030
3031 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3032 if (!log_is_enabled(Trace, gc, liveness)) {
3033 return;
3034 }
3035
3036 // add static memory usages to remembered set sizes
3037 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3038 // Print the footer of the output.
3039 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3040 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3041 " SUMMARY"
3042 G1PPRL_SUM_MB_FORMAT("capacity")
3043 G1PPRL_SUM_MB_PERC_FORMAT("used")
3044 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3045 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3046 G1PPRL_SUM_MB_FORMAT("remset")
3047 G1PPRL_SUM_MB_FORMAT("code-roots"),
3048 bytes_to_mb(_total_capacity_bytes),
3049 bytes_to_mb(_total_used_bytes),
3050 percent_of(_total_used_bytes, _total_capacity_bytes),
3051 bytes_to_mb(_total_prev_live_bytes),
3052 percent_of(_total_prev_live_bytes, _total_capacity_bytes),
3053 bytes_to_mb(_total_next_live_bytes),
3054 percent_of(_total_next_live_bytes, _total_capacity_bytes),
3055 bytes_to_mb(_total_remset_bytes),
3056 bytes_to_mb(_total_strong_code_roots_bytes));
3057 }