1 /*
2 * Copyright (c) 2001, 2020, 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/g1DirtyCardQueue.hpp"
34 #include "gc/g1/g1HeapVerifier.hpp"
35 #include "gc/g1/g1OopClosures.inline.hpp"
36 #include "gc/g1/g1Policy.hpp"
37 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
38 #include "gc/g1/g1StringDedup.hpp"
39 #include "gc/g1/g1ThreadLocalData.hpp"
40 #include "gc/g1/g1Trace.hpp"
41 #include "gc/g1/heapRegion.inline.hpp"
42 #include "gc/g1/heapRegionRemSet.hpp"
43 #include "gc/g1/heapRegionSet.inline.hpp"
44 #include "gc/shared/gcId.hpp"
45 #include "gc/shared/gcTimer.hpp"
46 #include "gc/shared/gcTraceTime.inline.hpp"
47 #include "gc/shared/gcVMOperations.hpp"
48 #include "gc/shared/genOopClosures.inline.hpp"
49 #include "gc/shared/referencePolicy.hpp"
50 #include "gc/shared/strongRootsScope.hpp"
51 #include "gc/shared/suspendibleThreadSet.hpp"
52 #include "gc/shared/taskTerminator.hpp"
53 #include "gc/shared/taskqueue.inline.hpp"
54 #include "gc/shared/weakProcessor.inline.hpp"
55 #include "gc/shared/workerPolicy.hpp"
56 #include "include/jvm.h"
57 #include "logging/log.hpp"
58 #include "memory/allocation.hpp"
59 #include "memory/iterator.hpp"
60 #include "memory/resourceArea.hpp"
61 #include "memory/universe.hpp"
62 #include "oops/access.inline.hpp"
63 #include "oops/oop.inline.hpp"
64 #include "runtime/atomic.hpp"
65 #include "runtime/handles.inline.hpp"
66 #include "runtime/java.hpp"
67 #include "runtime/orderAccess.hpp"
68 #include "runtime/prefetch.inline.hpp"
69 #include "services/memTracker.hpp"
70 #include "utilities/align.hpp"
71 #include "utilities/growableArray.hpp"
72
73 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
74 assert(addr < _cm->finger(), "invariant");
75 assert(addr >= _task->finger(), "invariant");
76
77 // We move that task's local finger along.
78 _task->move_finger_to(addr);
79
80 _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
81 // we only partially drain the local queue and global stack
82 _task->drain_local_queue(true);
83 _task->drain_global_stack(true);
84
85 // if the has_aborted flag has been raised, we need to bail out of
86 // the iteration
87 return !_task->has_aborted();
88 }
89
90 G1CMMarkStack::G1CMMarkStack() :
91 _max_chunk_capacity(0),
92 _base(NULL),
93 _chunk_capacity(0) {
94 set_empty();
95 }
96
97 bool G1CMMarkStack::resize(size_t new_capacity) {
98 assert(is_empty(), "Only resize when stack is empty.");
99 assert(new_capacity <= _max_chunk_capacity,
100 "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
101
102 TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
103
104 if (new_base == NULL) {
105 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));
106 return false;
107 }
108 // Release old mapping.
109 if (_base != NULL) {
110 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
111 }
112
113 _base = new_base;
114 _chunk_capacity = new_capacity;
115 set_empty();
116
117 return true;
118 }
119
120 size_t G1CMMarkStack::capacity_alignment() {
121 return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
122 }
123
124 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
125 guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
126
127 size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
128
129 _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
130 size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
131
132 guarantee(initial_chunk_capacity <= _max_chunk_capacity,
133 "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
134 _max_chunk_capacity,
135 initial_chunk_capacity);
136
137 log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
138 initial_chunk_capacity, _max_chunk_capacity);
139
140 return resize(initial_chunk_capacity);
141 }
142
143 void G1CMMarkStack::expand() {
144 if (_chunk_capacity == _max_chunk_capacity) {
145 log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
146 return;
147 }
148 size_t old_capacity = _chunk_capacity;
149 // Double capacity if possible
150 size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
151
152 if (resize(new_capacity)) {
153 log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
154 old_capacity, new_capacity);
155 } else {
156 log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
157 old_capacity, new_capacity);
158 }
159 }
160
161 G1CMMarkStack::~G1CMMarkStack() {
162 if (_base != NULL) {
163 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
164 }
165 }
166
167 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
168 elem->next = *list;
169 *list = elem;
170 }
171
172 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
173 MutexLocker x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
174 add_chunk_to_list(&_chunk_list, elem);
175 _chunks_in_chunk_list++;
176 }
177
178 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
179 MutexLocker x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
180 add_chunk_to_list(&_free_list, elem);
181 }
182
183 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
184 TaskQueueEntryChunk* result = *list;
185 if (result != NULL) {
186 *list = (*list)->next;
187 }
188 return result;
189 }
190
191 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
192 MutexLocker x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
193 TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
194 if (result != NULL) {
195 _chunks_in_chunk_list--;
196 }
197 return result;
198 }
199
200 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
201 MutexLocker x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
202 return remove_chunk_from_list(&_free_list);
203 }
204
205 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
206 // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
207 // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
208 // wraparound of _hwm.
209 if (_hwm >= _chunk_capacity) {
210 return NULL;
211 }
212
213 size_t cur_idx = Atomic::fetch_and_add(&_hwm, 1u);
214 if (cur_idx >= _chunk_capacity) {
215 return NULL;
216 }
217
218 TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
219 result->next = NULL;
220 return result;
221 }
222
223 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
224 // Get a new chunk.
225 TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
226
227 if (new_chunk == NULL) {
228 // Did not get a chunk from the free list. Allocate from backing memory.
229 new_chunk = allocate_new_chunk();
230
231 if (new_chunk == NULL) {
232 return false;
233 }
234 }
235
236 Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
237
238 add_chunk_to_chunk_list(new_chunk);
239
240 return true;
241 }
242
243 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
244 TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
245
246 if (cur == NULL) {
247 return false;
248 }
249
250 Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
251
252 add_chunk_to_free_list(cur);
253 return true;
254 }
255
256 void G1CMMarkStack::set_empty() {
257 _chunks_in_chunk_list = 0;
258 _hwm = 0;
259 _chunk_list = NULL;
260 _free_list = NULL;
261 }
262
263 G1CMRootMemRegions::G1CMRootMemRegions(uint const max_regions) :
264 _root_regions(MemRegion::create(max_regions, mtGC)),
265 _max_regions(max_regions),
266 _num_root_regions(0),
267 _claimed_root_regions(0),
268 _scan_in_progress(false),
269 _should_abort(false) { }
270
271 G1CMRootMemRegions::~G1CMRootMemRegions() {
272 FREE_C_HEAP_ARRAY(MemRegion, _root_regions);
273 }
274
275 void G1CMRootMemRegions::reset() {
276 _num_root_regions = 0;
277 }
278
279 void G1CMRootMemRegions::add(HeapWord* start, HeapWord* end) {
280 assert_at_safepoint();
281 size_t idx = Atomic::fetch_and_add(&_num_root_regions, 1u);
282 assert(idx < _max_regions, "Trying to add more root MemRegions than there is space " SIZE_FORMAT, _max_regions);
283 assert(start != NULL && end != NULL && start <= end, "Start (" PTR_FORMAT ") should be less or equal to "
284 "end (" PTR_FORMAT ")", p2i(start), p2i(end));
285 _root_regions[idx].set_start(start);
286 _root_regions[idx].set_end(end);
287 }
288
289 void G1CMRootMemRegions::prepare_for_scan() {
290 assert(!scan_in_progress(), "pre-condition");
291
292 _scan_in_progress = _num_root_regions > 0;
293
294 _claimed_root_regions = 0;
295 _should_abort = false;
296 }
297
298 const MemRegion* G1CMRootMemRegions::claim_next() {
299 if (_should_abort) {
300 // If someone has set the should_abort flag, we return NULL to
301 // force the caller to bail out of their loop.
302 return NULL;
303 }
304
305 if (_claimed_root_regions >= _num_root_regions) {
306 return NULL;
307 }
308
309 size_t claimed_index = Atomic::fetch_and_add(&_claimed_root_regions, 1u);
310 if (claimed_index < _num_root_regions) {
311 return &_root_regions[claimed_index];
312 }
313 return NULL;
314 }
315
316 uint G1CMRootMemRegions::num_root_regions() const {
317 return (uint)_num_root_regions;
318 }
319
320 void G1CMRootMemRegions::notify_scan_done() {
321 MutexLocker x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
322 _scan_in_progress = false;
323 RootRegionScan_lock->notify_all();
324 }
325
326 void G1CMRootMemRegions::cancel_scan() {
327 notify_scan_done();
328 }
329
330 void G1CMRootMemRegions::scan_finished() {
331 assert(scan_in_progress(), "pre-condition");
332
333 if (!_should_abort) {
334 assert(_claimed_root_regions >= num_root_regions(),
335 "we should have claimed all root regions, claimed " SIZE_FORMAT ", length = %u",
336 _claimed_root_regions, num_root_regions());
337 }
338
339 notify_scan_done();
340 }
341
342 bool G1CMRootMemRegions::wait_until_scan_finished() {
343 if (!scan_in_progress()) {
344 return false;
345 }
346
347 {
348 MonitorLocker ml(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
349 while (scan_in_progress()) {
350 ml.wait();
351 }
352 }
353 return true;
354 }
355
356 // Returns the maximum number of workers to be used in a concurrent
357 // phase based on the number of GC workers being used in a STW
358 // phase.
359 static uint scale_concurrent_worker_threads(uint num_gc_workers) {
360 return MAX2((num_gc_workers + 2) / 4, 1U);
361 }
362
363 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
364 G1RegionToSpaceMapper* prev_bitmap_storage,
365 G1RegionToSpaceMapper* next_bitmap_storage) :
366 // _cm_thread set inside the constructor
367 _g1h(g1h),
368 _completed_initialization(false),
369
370 _mark_bitmap_1(),
371 _mark_bitmap_2(),
372 _prev_mark_bitmap(&_mark_bitmap_1),
373 _next_mark_bitmap(&_mark_bitmap_2),
374
375 _heap(_g1h->reserved_region()),
376
377 _root_regions(_g1h->max_regions()),
378
379 _global_mark_stack(),
380
381 // _finger set in set_non_marking_state
382
383 _worker_id_offset(G1DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads),
384 _max_num_tasks(ParallelGCThreads),
385 // _num_active_tasks set in set_non_marking_state()
386 // _tasks set inside the constructor
387
388 _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
389 _terminator((int) _max_num_tasks, _task_queues),
390
391 _first_overflow_barrier_sync(),
392 _second_overflow_barrier_sync(),
393
394 _has_overflown(false),
395 _concurrent(false),
396 _has_aborted(false),
397 _restart_for_overflow(false),
398 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
399 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
400
401 // _verbose_level set below
402
403 _init_times(),
404 _remark_times(),
405 _remark_mark_times(),
406 _remark_weak_ref_times(),
407 _cleanup_times(),
408 _total_cleanup_time(0.0),
409
410 _accum_task_vtime(NULL),
411
412 _concurrent_workers(NULL),
413 _num_concurrent_workers(0),
414 _max_concurrent_workers(0),
415
416 _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)),
417 _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC))
418 {
419 _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage);
420 _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage);
421
422 // Create & start ConcurrentMark thread.
423 _cm_thread = new G1ConcurrentMarkThread(this);
424 if (_cm_thread->osthread() == NULL) {
425 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
426 }
427
428 assert(CGC_lock != NULL, "CGC_lock must be initialized");
429
430 if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) {
431 // Calculate the number of concurrent worker threads by scaling
432 // the number of parallel GC threads.
433 uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads);
434 FLAG_SET_ERGO(ConcGCThreads, marking_thread_num);
435 }
436
437 assert(ConcGCThreads > 0, "ConcGCThreads have been set.");
438 if (ConcGCThreads > ParallelGCThreads) {
439 log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).",
440 ConcGCThreads, ParallelGCThreads);
441 return;
442 }
443
444 log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset);
445 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
446
447 _num_concurrent_workers = ConcGCThreads;
448 _max_concurrent_workers = _num_concurrent_workers;
449
450 _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true);
451 _concurrent_workers->initialize_workers();
452
453 if (FLAG_IS_DEFAULT(MarkStackSize)) {
454 size_t mark_stack_size =
455 MIN2(MarkStackSizeMax,
456 MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE)));
457 // Verify that the calculated value for MarkStackSize is in range.
458 // It would be nice to use the private utility routine from Arguments.
459 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
460 log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
461 "must be between 1 and " SIZE_FORMAT,
462 mark_stack_size, MarkStackSizeMax);
463 return;
464 }
465 FLAG_SET_ERGO(MarkStackSize, mark_stack_size);
466 } else {
467 // Verify MarkStackSize is in range.
468 if (FLAG_IS_CMDLINE(MarkStackSize)) {
469 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
470 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
471 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
472 "must be between 1 and " SIZE_FORMAT,
473 MarkStackSize, MarkStackSizeMax);
474 return;
475 }
476 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
477 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
478 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
479 " or for MarkStackSizeMax (" SIZE_FORMAT ")",
480 MarkStackSize, MarkStackSizeMax);
481 return;
482 }
483 }
484 }
485 }
486
487 if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
488 vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
489 }
490
491 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
492 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
493
494 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
495 _num_active_tasks = _max_num_tasks;
496
497 for (uint i = 0; i < _max_num_tasks; ++i) {
498 G1CMTaskQueue* task_queue = new G1CMTaskQueue();
499 task_queue->initialize();
500 _task_queues->register_queue(i, task_queue);
501
502 _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions());
503
504 _accum_task_vtime[i] = 0.0;
505 }
506
507 reset_at_marking_complete();
508 _completed_initialization = true;
509 }
510
511 void G1ConcurrentMark::reset() {
512 _has_aborted = false;
513
514 reset_marking_for_restart();
515
516 // Reset all tasks, since different phases will use different number of active
517 // threads. So, it's easiest to have all of them ready.
518 for (uint i = 0; i < _max_num_tasks; ++i) {
519 _tasks[i]->reset(_next_mark_bitmap);
520 }
521
522 uint max_regions = _g1h->max_regions();
523 for (uint i = 0; i < max_regions; i++) {
524 _top_at_rebuild_starts[i] = NULL;
525 _region_mark_stats[i].clear();
526 }
527 }
528
529 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) {
530 for (uint j = 0; j < _max_num_tasks; ++j) {
531 _tasks[j]->clear_mark_stats_cache(region_idx);
532 }
533 _top_at_rebuild_starts[region_idx] = NULL;
534 _region_mark_stats[region_idx].clear();
535 }
536
537 void G1ConcurrentMark::clear_statistics(HeapRegion* r) {
538 uint const region_idx = r->hrm_index();
539 if (r->is_humongous()) {
540 assert(r->is_starts_humongous(), "Got humongous continues region here");
541 uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size());
542 for (uint j = region_idx; j < (region_idx + size_in_regions); j++) {
543 clear_statistics_in_region(j);
544 }
545 } else {
546 clear_statistics_in_region(region_idx);
547 }
548 }
549
550 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) {
551 if (bitmap->is_marked(addr)) {
552 bitmap->clear(addr);
553 }
554 }
555
556 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) {
557 assert_at_safepoint_on_vm_thread();
558
559 // Need to clear all mark bits of the humongous object.
560 clear_mark_if_set(_prev_mark_bitmap, r->bottom());
561 clear_mark_if_set(_next_mark_bitmap, r->bottom());
562
563 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
564 return;
565 }
566
567 // Clear any statistics about the region gathered so far.
568 clear_statistics(r);
569 }
570
571 void G1ConcurrentMark::reset_marking_for_restart() {
572 _global_mark_stack.set_empty();
573
574 // Expand the marking stack, if we have to and if we can.
575 if (has_overflown()) {
576 _global_mark_stack.expand();
577
578 uint max_regions = _g1h->max_regions();
579 for (uint i = 0; i < max_regions; i++) {
580 _region_mark_stats[i].clear_during_overflow();
581 }
582 }
583
584 clear_has_overflown();
585 _finger = _heap.start();
586
587 for (uint i = 0; i < _max_num_tasks; ++i) {
588 G1CMTaskQueue* queue = _task_queues->queue(i);
589 queue->set_empty();
590 }
591 }
592
593 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
594 assert(active_tasks <= _max_num_tasks, "we should not have more");
595
596 _num_active_tasks = active_tasks;
597 // Need to update the three data structures below according to the
598 // number of active threads for this phase.
599 _terminator.reset_for_reuse(active_tasks);
600 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
601 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
602 }
603
604 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
605 set_concurrency(active_tasks);
606
607 _concurrent = concurrent;
608
609 if (!concurrent) {
610 // At this point we should be in a STW phase, and completed marking.
611 assert_at_safepoint_on_vm_thread();
612 assert(out_of_regions(),
613 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
614 p2i(_finger), p2i(_heap.end()));
615 }
616 }
617
618 void G1ConcurrentMark::reset_at_marking_complete() {
619 // We set the global marking state to some default values when we're
620 // not doing marking.
621 reset_marking_for_restart();
622 _num_active_tasks = 0;
623 }
624
625 G1ConcurrentMark::~G1ConcurrentMark() {
626 FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts);
627 FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats);
628 // The G1ConcurrentMark instance is never freed.
629 ShouldNotReachHere();
630 }
631
632 class G1ClearBitMapTask : public AbstractGangTask {
633 public:
634 static size_t chunk_size() { return M; }
635
636 private:
637 // Heap region closure used for clearing the given mark bitmap.
638 class G1ClearBitmapHRClosure : public HeapRegionClosure {
639 private:
640 G1CMBitMap* _bitmap;
641 G1ConcurrentMark* _cm;
642 public:
643 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _bitmap(bitmap), _cm(cm) {
644 }
645
646 virtual bool do_heap_region(HeapRegion* r) {
647 size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
648
649 HeapWord* cur = r->bottom();
650 HeapWord* const end = r->end();
651
652 while (cur < end) {
653 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
654 _bitmap->clear_range(mr);
655
656 cur += chunk_size_in_words;
657
658 // Abort iteration if after yielding the marking has been aborted.
659 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
660 return true;
661 }
662 // Repeat the asserts from before the start of the closure. We will do them
663 // as asserts here to minimize their overhead on the product. However, we
664 // will have them as guarantees at the beginning / end of the bitmap
665 // clearing to get some checking in the product.
666 assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant");
667 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant");
668 }
669 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
670
671 return false;
672 }
673 };
674
675 G1ClearBitmapHRClosure _cl;
676 HeapRegionClaimer _hr_claimer;
677 bool _suspendible; // If the task is suspendible, workers must join the STS.
678
679 public:
680 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
681 AbstractGangTask("G1 Clear Bitmap"),
682 _cl(bitmap, suspendible ? cm : NULL),
683 _hr_claimer(n_workers),
684 _suspendible(suspendible)
685 { }
686
687 void work(uint worker_id) {
688 SuspendibleThreadSetJoiner sts_join(_suspendible);
689 G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
690 }
691
692 bool is_complete() {
693 return _cl.is_complete();
694 }
695 };
696
697 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
698 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
699
700 size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
701 size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
702
703 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
704
705 G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
706
707 log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
708 workers->run_task(&cl, num_workers);
709 guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
710 }
711
712 void G1ConcurrentMark::cleanup_for_next_mark() {
713 // Make sure that the concurrent mark thread looks to still be in
714 // the current cycle.
715 guarantee(cm_thread()->during_cycle(), "invariant");
716
717 // We are finishing up the current cycle by clearing the next
718 // marking bitmap and getting it ready for the next cycle. During
719 // this time no other cycle can start. So, let's make sure that this
720 // is the case.
721 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
722
723 clear_bitmap(_next_mark_bitmap, _concurrent_workers, true);
724
725 // Repeat the asserts from above.
726 guarantee(cm_thread()->during_cycle(), "invariant");
727 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
728 }
729
730 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
731 assert_at_safepoint_on_vm_thread();
732 clear_bitmap(_prev_mark_bitmap, workers, false);
733 }
734
735 class NoteStartOfMarkHRClosure : public HeapRegionClosure {
736 public:
737 bool do_heap_region(HeapRegion* r) {
738 r->note_start_of_marking();
739 return false;
740 }
741 };
742
743 void G1ConcurrentMark::pre_initial_mark() {
744 assert_at_safepoint_on_vm_thread();
745
746 // Reset marking state.
747 reset();
748
749 // For each region note start of marking.
750 NoteStartOfMarkHRClosure startcl;
751 _g1h->heap_region_iterate(&startcl);
752
753 _root_regions.reset();
754 }
755
756
757 void G1ConcurrentMark::post_initial_mark() {
758 // Start Concurrent Marking weak-reference discovery.
759 ReferenceProcessor* rp = _g1h->ref_processor_cm();
760 // enable ("weak") refs discovery
761 rp->enable_discovery();
762 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
763
764 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
765 // This is the start of the marking cycle, we're expected all
766 // threads to have SATB queues with active set to false.
767 satb_mq_set.set_active_all_threads(true, /* new active value */
768 false /* expected_active */);
769
770 _root_regions.prepare_for_scan();
771
772 // update_g1_committed() will be called at the end of an evac pause
773 // when marking is on. So, it's also called at the end of the
774 // initial-mark pause to update the heap end, if the heap expands
775 // during it. No need to call it here.
776 }
777
778 /*
779 * Notice that in the next two methods, we actually leave the STS
780 * during the barrier sync and join it immediately afterwards. If we
781 * do not do this, the following deadlock can occur: one thread could
782 * be in the barrier sync code, waiting for the other thread to also
783 * sync up, whereas another one could be trying to yield, while also
784 * waiting for the other threads to sync up too.
785 *
786 * Note, however, that this code is also used during remark and in
787 * this case we should not attempt to leave / enter the STS, otherwise
788 * we'll either hit an assert (debug / fastdebug) or deadlock
789 * (product). So we should only leave / enter the STS if we are
790 * operating concurrently.
791 *
792 * Because the thread that does the sync barrier has left the STS, it
793 * is possible to be suspended for a Full GC or an evacuation pause
794 * could occur. This is actually safe, since the entering the sync
795 * barrier is one of the last things do_marking_step() does, and it
796 * doesn't manipulate any data structures afterwards.
797 */
798
799 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
800 bool barrier_aborted;
801 {
802 SuspendibleThreadSetLeaver sts_leave(concurrent());
803 barrier_aborted = !_first_overflow_barrier_sync.enter();
804 }
805
806 // at this point everyone should have synced up and not be doing any
807 // more work
808
809 if (barrier_aborted) {
810 // If the barrier aborted we ignore the overflow condition and
811 // just abort the whole marking phase as quickly as possible.
812 return;
813 }
814 }
815
816 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
817 SuspendibleThreadSetLeaver sts_leave(concurrent());
818 _second_overflow_barrier_sync.enter();
819
820 // at this point everything should be re-initialized and ready to go
821 }
822
823 class G1CMConcurrentMarkingTask : public AbstractGangTask {
824 G1ConcurrentMark* _cm;
825
826 public:
827 void work(uint worker_id) {
828 assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread");
829 ResourceMark rm;
830
831 double start_vtime = os::elapsedVTime();
832
833 {
834 SuspendibleThreadSetJoiner sts_join;
835
836 assert(worker_id < _cm->active_tasks(), "invariant");
837
838 G1CMTask* task = _cm->task(worker_id);
839 task->record_start_time();
840 if (!_cm->has_aborted()) {
841 do {
842 task->do_marking_step(G1ConcMarkStepDurationMillis,
843 true /* do_termination */,
844 false /* is_serial*/);
845
846 _cm->do_yield_check();
847 } while (!_cm->has_aborted() && task->has_aborted());
848 }
849 task->record_end_time();
850 guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
851 }
852
853 double end_vtime = os::elapsedVTime();
854 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
855 }
856
857 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) :
858 AbstractGangTask("Concurrent Mark"), _cm(cm) { }
859
860 ~G1CMConcurrentMarkingTask() { }
861 };
862
863 uint G1ConcurrentMark::calc_active_marking_workers() {
864 uint result = 0;
865 if (!UseDynamicNumberOfGCThreads ||
866 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
867 !ForceDynamicNumberOfGCThreads)) {
868 result = _max_concurrent_workers;
869 } else {
870 result =
871 WorkerPolicy::calc_default_active_workers(_max_concurrent_workers,
872 1, /* Minimum workers */
873 _num_concurrent_workers,
874 Threads::number_of_non_daemon_threads());
875 // Don't scale the result down by scale_concurrent_workers() because
876 // that scaling has already gone into "_max_concurrent_workers".
877 }
878 assert(result > 0 && result <= _max_concurrent_workers,
879 "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
880 _max_concurrent_workers, result);
881 return result;
882 }
883
884 void G1ConcurrentMark::scan_root_region(const MemRegion* region, uint worker_id) {
885 #ifdef ASSERT
886 HeapWord* last = region->last();
887 HeapRegion* hr = _g1h->heap_region_containing(last);
888 assert(hr->is_old() || hr->next_top_at_mark_start() == hr->bottom(),
889 "Root regions must be old or survivor/eden but region %u is %s", hr->hrm_index(), hr->get_type_str());
890 assert(hr->next_top_at_mark_start() == region->start(),
891 "MemRegion start should be equal to nTAMS");
892 #endif
893
894 G1RootRegionScanClosure cl(_g1h, this, worker_id);
895
896 const uintx interval = PrefetchScanIntervalInBytes;
897 HeapWord* curr = region->start();
898 const HeapWord* end = region->end();
899 while (curr < end) {
900 Prefetch::read(curr, interval);
901 oop obj = oop(curr);
902 int size = obj->oop_iterate_size(&cl);
903 assert(size == obj->size(), "sanity");
904 curr += size;
905 }
906 }
907
908 class G1CMRootRegionScanTask : public AbstractGangTask {
909 G1ConcurrentMark* _cm;
910 public:
911 G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
912 AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
913
914 void work(uint worker_id) {
915 assert(Thread::current()->is_ConcurrentGC_thread(),
916 "this should only be done by a conc GC thread");
917
918 G1CMRootMemRegions* root_regions = _cm->root_regions();
919 const MemRegion* region = root_regions->claim_next();
920 while (region != NULL) {
921 _cm->scan_root_region(region, worker_id);
922 region = root_regions->claim_next();
923 }
924 }
925 };
926
927 void G1ConcurrentMark::scan_root_regions() {
928 // scan_in_progress() will have been set to true only if there was
929 // at least one root region to scan. So, if it's false, we
930 // should not attempt to do any further work.
931 if (root_regions()->scan_in_progress()) {
932 assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
933
934 _num_concurrent_workers = MIN2(calc_active_marking_workers(),
935 // We distribute work on a per-region basis, so starting
936 // more threads than that is useless.
937 root_regions()->num_root_regions());
938 assert(_num_concurrent_workers <= _max_concurrent_workers,
939 "Maximum number of marking threads exceeded");
940
941 G1CMRootRegionScanTask task(this);
942 log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
943 task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
944 _concurrent_workers->run_task(&task, _num_concurrent_workers);
945
946 // It's possible that has_aborted() is true here without actually
947 // aborting the survivor scan earlier. This is OK as it's
948 // mainly used for sanity checking.
949 root_regions()->scan_finished();
950 }
951 }
952
953 void G1ConcurrentMark::concurrent_cycle_start() {
954 _gc_timer_cm->register_gc_start();
955
956 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
957
958 _g1h->trace_heap_before_gc(_gc_tracer_cm);
959 }
960
961 void G1ConcurrentMark::concurrent_cycle_end() {
962 _g1h->collector_state()->set_clearing_next_bitmap(false);
963
964 _g1h->trace_heap_after_gc(_gc_tracer_cm);
965
966 if (has_aborted()) {
967 log_info(gc, marking)("Concurrent Mark Abort");
968 _gc_tracer_cm->report_concurrent_mode_failure();
969 }
970
971 _gc_timer_cm->register_gc_end();
972
973 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
974 }
975
976 void G1ConcurrentMark::mark_from_roots() {
977 _restart_for_overflow = false;
978
979 _num_concurrent_workers = calc_active_marking_workers();
980
981 uint active_workers = MAX2(1U, _num_concurrent_workers);
982
983 // Setting active workers is not guaranteed since fewer
984 // worker threads may currently exist and more may not be
985 // available.
986 active_workers = _concurrent_workers->update_active_workers(active_workers);
987 log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers());
988
989 // Parallel task terminator is set in "set_concurrency_and_phase()"
990 set_concurrency_and_phase(active_workers, true /* concurrent */);
991
992 G1CMConcurrentMarkingTask marking_task(this);
993 _concurrent_workers->run_task(&marking_task);
994 print_stats();
995 }
996
997 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) {
998 G1HeapVerifier* verifier = _g1h->verifier();
999
1000 verifier->verify_region_sets_optional();
1001
1002 if (VerifyDuringGC) {
1003 GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm);
1004
1005 size_t const BufLen = 512;
1006 char buffer[BufLen];
1007
1008 jio_snprintf(buffer, BufLen, "During GC (%s)", caller);
1009 verifier->verify(type, vo, buffer);
1010 }
1011
1012 verifier->check_bitmaps(caller);
1013 }
1014
1015 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask {
1016 G1CollectedHeap* _g1h;
1017 G1ConcurrentMark* _cm;
1018 HeapRegionClaimer _hrclaimer;
1019 uint volatile _total_selected_for_rebuild;
1020
1021 G1PrintRegionLivenessInfoClosure _cl;
1022
1023 class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure {
1024 G1CollectedHeap* _g1h;
1025 G1ConcurrentMark* _cm;
1026
1027 G1PrintRegionLivenessInfoClosure* _cl;
1028
1029 uint _num_regions_selected_for_rebuild; // The number of regions actually selected for rebuild.
1030
1031 void update_remset_before_rebuild(HeapRegion* hr) {
1032 G1RemSetTrackingPolicy* tracking_policy = _g1h->policy()->remset_tracker();
1033
1034 bool selected_for_rebuild;
1035 if (hr->is_humongous()) {
1036 bool const is_live = _cm->liveness(hr->humongous_start_region()->hrm_index()) > 0;
1037 selected_for_rebuild = tracking_policy->update_humongous_before_rebuild(hr, is_live);
1038 } else {
1039 size_t const live_bytes = _cm->liveness(hr->hrm_index());
1040 selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);
1041 }
1042 if (selected_for_rebuild) {
1043 _num_regions_selected_for_rebuild++;
1044 }
1045 _cm->update_top_at_rebuild_start(hr);
1046 }
1047
1048 // Distribute the given words across the humongous object starting with hr and
1049 // note end of marking.
1050 void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) {
1051 uint const region_idx = hr->hrm_index();
1052 size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size();
1053 uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words);
1054
1055 // "Distributing" zero words means that we only note end of marking for these
1056 // regions.
1057 assert(marked_words == 0 || obj_size_in_words == marked_words,
1058 "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT,
1059 obj_size_in_words, marked_words);
1060
1061 for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) {
1062 HeapRegion* const r = _g1h->region_at(i);
1063 size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words);
1064
1065 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)",
1066 words_to_add, i, r->get_type_str());
1067 add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize);
1068 marked_words -= words_to_add;
1069 }
1070 assert(marked_words == 0,
1071 SIZE_FORMAT " words left after distributing space across %u regions",
1072 marked_words, num_regions_in_humongous);
1073 }
1074
1075 void update_marked_bytes(HeapRegion* hr) {
1076 uint const region_idx = hr->hrm_index();
1077 size_t const marked_words = _cm->liveness(region_idx);
1078 // The marking attributes the object's size completely to the humongous starts
1079 // region. We need to distribute this value across the entire set of regions a
1080 // humongous object spans.
1081 if (hr->is_humongous()) {
1082 assert(hr->is_starts_humongous() || marked_words == 0,
1083 "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)",
1084 marked_words, region_idx, hr->get_type_str());
1085 if (hr->is_starts_humongous()) {
1086 distribute_marked_bytes(hr, marked_words);
1087 }
1088 } else {
1089 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str());
1090 add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize);
1091 }
1092 }
1093
1094 void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) {
1095 hr->add_to_marked_bytes(marked_bytes);
1096 _cl->do_heap_region(hr);
1097 hr->note_end_of_marking();
1098 }
1099
1100 public:
1101 G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) :
1102 _g1h(g1h), _cm(cm), _cl(cl), _num_regions_selected_for_rebuild(0) { }
1103
1104 virtual bool do_heap_region(HeapRegion* r) {
1105 update_remset_before_rebuild(r);
1106 update_marked_bytes(r);
1107
1108 return false;
1109 }
1110
1111 uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; }
1112 };
1113
1114 public:
1115 G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) :
1116 AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"),
1117 _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { }
1118
1119 virtual void work(uint worker_id) {
1120 G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl);
1121 _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id);
1122 Atomic::add(&_total_selected_for_rebuild, update_cl.num_selected_for_rebuild());
1123 }
1124
1125 uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; }
1126
1127 // Number of regions for which roughly one thread should be spawned for this work.
1128 static const uint RegionsPerThread = 384;
1129 };
1130
1131 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure {
1132 G1CollectedHeap* _g1h;
1133 public:
1134 G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { }
1135
1136 virtual bool do_heap_region(HeapRegion* r) {
1137 _g1h->policy()->remset_tracker()->update_after_rebuild(r);
1138 return false;
1139 }
1140 };
1141
1142 void G1ConcurrentMark::remark() {
1143 assert_at_safepoint_on_vm_thread();
1144
1145 // If a full collection has happened, we should not continue. However we might
1146 // have ended up here as the Remark VM operation has been scheduled already.
1147 if (has_aborted()) {
1148 return;
1149 }
1150
1151 G1Policy* policy = _g1h->policy();
1152 policy->record_concurrent_mark_remark_start();
1153
1154 double start = os::elapsedTime();
1155
1156 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before");
1157
1158 {
1159 GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm);
1160 finalize_marking();
1161 }
1162
1163 double mark_work_end = os::elapsedTime();
1164
1165 bool const mark_finished = !has_overflown();
1166 if (mark_finished) {
1167 weak_refs_work(false /* clear_all_soft_refs */);
1168
1169 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1170 // We're done with marking.
1171 // This is the end of the marking cycle, we're expected all
1172 // threads to have SATB queues with active set to true.
1173 satb_mq_set.set_active_all_threads(false, /* new active value */
1174 true /* expected_active */);
1175
1176 {
1177 GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm);
1178 flush_all_task_caches();
1179 }
1180
1181 // Install newly created mark bitmap as "prev".
1182 swap_mark_bitmaps();
1183 {
1184 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm);
1185
1186 uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) /
1187 G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread;
1188 uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity);
1189
1190 G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers);
1191 log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions());
1192 _g1h->workers()->run_task(&cl, num_workers);
1193
1194 log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u",
1195 _g1h->num_regions(), cl.total_selected_for_rebuild());
1196 }
1197 {
1198 GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm);
1199 reclaim_empty_regions();
1200 }
1201
1202 // Clean out dead classes
1203 if (ClassUnloadingWithConcurrentMark) {
1204 GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm);
1205 ClassLoaderDataGraph::purge();
1206 }
1207
1208 _g1h->resize_heap_if_necessary();
1209
1210 compute_new_sizes();
1211
1212 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after");
1213
1214 assert(!restart_for_overflow(), "sanity");
1215 // Completely reset the marking state since marking completed
1216 reset_at_marking_complete();
1217 } else {
1218 // We overflowed. Restart concurrent marking.
1219 _restart_for_overflow = true;
1220
1221 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow");
1222
1223 // Clear the marking state because we will be restarting
1224 // marking due to overflowing the global mark stack.
1225 reset_marking_for_restart();
1226 }
1227
1228 {
1229 GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm);
1230 report_object_count(mark_finished);
1231 }
1232
1233 // Statistics
1234 double now = os::elapsedTime();
1235 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1236 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1237 _remark_times.add((now - start) * 1000.0);
1238
1239 policy->record_concurrent_mark_remark_end();
1240 }
1241
1242 class G1ReclaimEmptyRegionsTask : public AbstractGangTask {
1243 // Per-region work during the Cleanup pause.
1244 class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure {
1245 G1CollectedHeap* _g1h;
1246 size_t _freed_bytes;
1247 FreeRegionList* _local_cleanup_list;
1248 uint _old_regions_removed;
1249 uint _humongous_regions_removed;
1250
1251 public:
1252 G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h,
1253 FreeRegionList* local_cleanup_list) :
1254 _g1h(g1h),
1255 _freed_bytes(0),
1256 _local_cleanup_list(local_cleanup_list),
1257 _old_regions_removed(0),
1258 _humongous_regions_removed(0) { }
1259
1260 size_t freed_bytes() { return _freed_bytes; }
1261 const uint old_regions_removed() { return _old_regions_removed; }
1262 const uint humongous_regions_removed() { return _humongous_regions_removed; }
1263
1264 bool do_heap_region(HeapRegion *hr) {
1265 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) {
1266 _freed_bytes += hr->used();
1267 hr->set_containing_set(NULL);
1268 if (hr->is_humongous()) {
1269 _humongous_regions_removed++;
1270 _g1h->free_humongous_region(hr, _local_cleanup_list);
1271 } else {
1272 _old_regions_removed++;
1273 _g1h->free_region(hr, _local_cleanup_list);
1274 }
1275 hr->clear_cardtable();
1276 _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index());
1277 log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom()));
1278 }
1279
1280 return false;
1281 }
1282 };
1283
1284 G1CollectedHeap* _g1h;
1285 FreeRegionList* _cleanup_list;
1286 HeapRegionClaimer _hrclaimer;
1287
1288 public:
1289 G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1290 AbstractGangTask("G1 Cleanup"),
1291 _g1h(g1h),
1292 _cleanup_list(cleanup_list),
1293 _hrclaimer(n_workers) {
1294 }
1295
1296 void work(uint worker_id) {
1297 FreeRegionList local_cleanup_list("Local Cleanup List");
1298 G1ReclaimEmptyRegionsClosure cl(_g1h, &local_cleanup_list);
1299 _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id);
1300 assert(cl.is_complete(), "Shouldn't have aborted!");
1301
1302 // Now update the old/humongous region sets
1303 _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed());
1304 {
1305 MutexLocker x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1306 _g1h->decrement_summary_bytes(cl.freed_bytes());
1307
1308 _cleanup_list->add_ordered(&local_cleanup_list);
1309 assert(local_cleanup_list.is_empty(), "post-condition");
1310 }
1311 }
1312 };
1313
1314 void G1ConcurrentMark::reclaim_empty_regions() {
1315 WorkGang* workers = _g1h->workers();
1316 FreeRegionList empty_regions_list("Empty Regions After Mark List");
1317
1318 G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers());
1319 workers->run_task(&cl);
1320
1321 if (!empty_regions_list.is_empty()) {
1322 log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length());
1323 // Now print the empty regions list.
1324 G1HRPrinter* hrp = _g1h->hr_printer();
1325 if (hrp->is_active()) {
1326 FreeRegionListIterator iter(&empty_regions_list);
1327 while (iter.more_available()) {
1328 HeapRegion* hr = iter.get_next();
1329 hrp->cleanup(hr);
1330 }
1331 }
1332 // And actually make them available.
1333 _g1h->prepend_to_freelist(&empty_regions_list);
1334 }
1335 }
1336
1337 void G1ConcurrentMark::compute_new_sizes() {
1338 MetaspaceGC::compute_new_size();
1339
1340 // Cleanup will have freed any regions completely full of garbage.
1341 // Update the soft reference policy with the new heap occupancy.
1342 Universe::update_heap_info_at_gc();
1343
1344 // We reclaimed old regions so we should calculate the sizes to make
1345 // sure we update the old gen/space data.
1346 _g1h->g1mm()->update_sizes();
1347 }
1348
1349 void G1ConcurrentMark::cleanup() {
1350 assert_at_safepoint_on_vm_thread();
1351
1352 // If a full collection has happened, we shouldn't do this.
1353 if (has_aborted()) {
1354 return;
1355 }
1356
1357 G1Policy* policy = _g1h->policy();
1358 policy->record_concurrent_mark_cleanup_start();
1359
1360 double start = os::elapsedTime();
1361
1362 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before");
1363
1364 {
1365 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm);
1366 G1UpdateRemSetTrackingAfterRebuild cl(_g1h);
1367 _g1h->heap_region_iterate(&cl);
1368 }
1369
1370 if (log_is_enabled(Trace, gc, liveness)) {
1371 G1PrintRegionLivenessInfoClosure cl("Post-Cleanup");
1372 _g1h->heap_region_iterate(&cl);
1373 }
1374
1375 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after");
1376
1377 // We need to make this be a "collection" so any collection pause that
1378 // races with it goes around and waits for Cleanup to finish.
1379 _g1h->increment_total_collections();
1380
1381 // Local statistics
1382 double recent_cleanup_time = (os::elapsedTime() - start);
1383 _total_cleanup_time += recent_cleanup_time;
1384 _cleanup_times.add(recent_cleanup_time);
1385
1386 {
1387 GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm);
1388 policy->record_concurrent_mark_cleanup_end();
1389 }
1390 }
1391
1392 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1393 // Uses the G1CMTask associated with a worker thread (for serial reference
1394 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1395 // trace referent objects.
1396 //
1397 // Using the G1CMTask and embedded local queues avoids having the worker
1398 // threads operating on the global mark stack. This reduces the risk
1399 // of overflowing the stack - which we would rather avoid at this late
1400 // state. Also using the tasks' local queues removes the potential
1401 // of the workers interfering with each other that could occur if
1402 // operating on the global stack.
1403
1404 class G1CMKeepAliveAndDrainClosure : public OopClosure {
1405 G1ConcurrentMark* _cm;
1406 G1CMTask* _task;
1407 uint _ref_counter_limit;
1408 uint _ref_counter;
1409 bool _is_serial;
1410 public:
1411 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1412 _cm(cm), _task(task), _ref_counter_limit(G1RefProcDrainInterval),
1413 _ref_counter(_ref_counter_limit), _is_serial(is_serial) {
1414 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1415 }
1416
1417 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1418 virtual void do_oop( oop* p) { do_oop_work(p); }
1419
1420 template <class T> void do_oop_work(T* p) {
1421 if (_cm->has_overflown()) {
1422 return;
1423 }
1424 if (!_task->deal_with_reference(p)) {
1425 // We did not add anything to the mark bitmap (or mark stack), so there is
1426 // no point trying to drain it.
1427 return;
1428 }
1429 _ref_counter--;
1430
1431 if (_ref_counter == 0) {
1432 // We have dealt with _ref_counter_limit references, pushing them
1433 // and objects reachable from them on to the local stack (and
1434 // possibly the global stack). Call G1CMTask::do_marking_step() to
1435 // process these entries.
1436 //
1437 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1438 // there's nothing more to do (i.e. we're done with the entries that
1439 // were pushed as a result of the G1CMTask::deal_with_reference() calls
1440 // above) or we overflow.
1441 //
1442 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1443 // flag while there may still be some work to do. (See the comment at
1444 // the beginning of G1CMTask::do_marking_step() for those conditions -
1445 // one of which is reaching the specified time target.) It is only
1446 // when G1CMTask::do_marking_step() returns without setting the
1447 // has_aborted() flag that the marking step has completed.
1448 do {
1449 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1450 _task->do_marking_step(mark_step_duration_ms,
1451 false /* do_termination */,
1452 _is_serial);
1453 } while (_task->has_aborted() && !_cm->has_overflown());
1454 _ref_counter = _ref_counter_limit;
1455 }
1456 }
1457 };
1458
1459 // 'Drain' oop closure used by both serial and parallel reference processing.
1460 // Uses the G1CMTask associated with a given worker thread (for serial
1461 // reference processing the G1CMtask for worker 0 is used). Calls the
1462 // do_marking_step routine, with an unbelievably large timeout value,
1463 // to drain the marking data structures of the remaining entries
1464 // added by the 'keep alive' oop closure above.
1465
1466 class G1CMDrainMarkingStackClosure : public VoidClosure {
1467 G1ConcurrentMark* _cm;
1468 G1CMTask* _task;
1469 bool _is_serial;
1470 public:
1471 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1472 _cm(cm), _task(task), _is_serial(is_serial) {
1473 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1474 }
1475
1476 void do_void() {
1477 do {
1478 // We call G1CMTask::do_marking_step() to completely drain the local
1479 // and global marking stacks of entries pushed by the 'keep alive'
1480 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1481 //
1482 // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1483 // if there's nothing more to do (i.e. we've completely drained the
1484 // entries that were pushed as a a result of applying the 'keep alive'
1485 // closure to the entries on the discovered ref lists) or we overflow
1486 // the global marking stack.
1487 //
1488 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1489 // flag while there may still be some work to do. (See the comment at
1490 // the beginning of G1CMTask::do_marking_step() for those conditions -
1491 // one of which is reaching the specified time target.) It is only
1492 // when G1CMTask::do_marking_step() returns without setting the
1493 // has_aborted() flag that the marking step has completed.
1494
1495 _task->do_marking_step(1000000000.0 /* something very large */,
1496 true /* do_termination */,
1497 _is_serial);
1498 } while (_task->has_aborted() && !_cm->has_overflown());
1499 }
1500 };
1501
1502 // Implementation of AbstractRefProcTaskExecutor for parallel
1503 // reference processing at the end of G1 concurrent marking
1504
1505 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor {
1506 private:
1507 G1CollectedHeap* _g1h;
1508 G1ConcurrentMark* _cm;
1509 WorkGang* _workers;
1510 uint _active_workers;
1511
1512 public:
1513 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1514 G1ConcurrentMark* cm,
1515 WorkGang* workers,
1516 uint n_workers) :
1517 _g1h(g1h), _cm(cm),
1518 _workers(workers), _active_workers(n_workers) { }
1519
1520 virtual void execute(ProcessTask& task, uint ergo_workers);
1521 };
1522
1523 class G1CMRefProcTaskProxy : public AbstractGangTask {
1524 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1525 ProcessTask& _proc_task;
1526 G1CollectedHeap* _g1h;
1527 G1ConcurrentMark* _cm;
1528
1529 public:
1530 G1CMRefProcTaskProxy(ProcessTask& proc_task,
1531 G1CollectedHeap* g1h,
1532 G1ConcurrentMark* cm) :
1533 AbstractGangTask("Process reference objects in parallel"),
1534 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1535 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1536 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1537 }
1538
1539 virtual void work(uint worker_id) {
1540 ResourceMark rm;
1541 HandleMark hm;
1542 G1CMTask* task = _cm->task(worker_id);
1543 G1CMIsAliveClosure g1_is_alive(_g1h);
1544 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1545 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1546
1547 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1548 }
1549 };
1550
1551 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task, uint ergo_workers) {
1552 assert(_workers != NULL, "Need parallel worker threads.");
1553 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1554 assert(_workers->active_workers() >= ergo_workers,
1555 "Ergonomically chosen workers(%u) should be less than or equal to active workers(%u)",
1556 ergo_workers, _workers->active_workers());
1557
1558 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1559
1560 // We need to reset the concurrency level before each
1561 // proxy task execution, so that the termination protocol
1562 // and overflow handling in G1CMTask::do_marking_step() knows
1563 // how many workers to wait for.
1564 _cm->set_concurrency(ergo_workers);
1565 _workers->run_task(&proc_task_proxy, ergo_workers);
1566 }
1567
1568 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) {
1569 ResourceMark rm;
1570 HandleMark hm;
1571
1572 // Is alive closure.
1573 G1CMIsAliveClosure g1_is_alive(_g1h);
1574
1575 // Inner scope to exclude the cleaning of the string table
1576 // from the displayed time.
1577 {
1578 GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm);
1579
1580 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1581
1582 // See the comment in G1CollectedHeap::ref_processing_init()
1583 // about how reference processing currently works in G1.
1584
1585 // Set the soft reference policy
1586 rp->setup_policy(clear_all_soft_refs);
1587 assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1588
1589 // Instances of the 'Keep Alive' and 'Complete GC' closures used
1590 // in serial reference processing. Note these closures are also
1591 // used for serially processing (by the the current thread) the
1592 // JNI references during parallel reference processing.
1593 //
1594 // These closures do not need to synchronize with the worker
1595 // threads involved in parallel reference processing as these
1596 // instances are executed serially by the current thread (e.g.
1597 // reference processing is not multi-threaded and is thus
1598 // performed by the current thread instead of a gang worker).
1599 //
1600 // The gang tasks involved in parallel reference processing create
1601 // their own instances of these closures, which do their own
1602 // synchronization among themselves.
1603 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1604 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1605
1606 // We need at least one active thread. If reference processing
1607 // is not multi-threaded we use the current (VMThread) thread,
1608 // otherwise we use the work gang from the G1CollectedHeap and
1609 // we utilize all the worker threads we can.
1610 bool processing_is_mt = rp->processing_is_mt();
1611 uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U);
1612 active_workers = clamp(active_workers, 1u, _max_num_tasks);
1613
1614 // Parallel processing task executor.
1615 G1CMRefProcTaskExecutor par_task_executor(_g1h, this,
1616 _g1h->workers(), active_workers);
1617 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1618
1619 // Set the concurrency level. The phase was already set prior to
1620 // executing the remark task.
1621 set_concurrency(active_workers);
1622
1623 // Set the degree of MT processing here. If the discovery was done MT,
1624 // the number of threads involved during discovery could differ from
1625 // the number of active workers. This is OK as long as the discovered
1626 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1627 rp->set_active_mt_degree(active_workers);
1628
1629 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues());
1630
1631 // Process the weak references.
1632 const ReferenceProcessorStats& stats =
1633 rp->process_discovered_references(&g1_is_alive,
1634 &g1_keep_alive,
1635 &g1_drain_mark_stack,
1636 executor,
1637 &pt);
1638 _gc_tracer_cm->report_gc_reference_stats(stats);
1639 pt.print_all_references();
1640
1641 // The do_oop work routines of the keep_alive and drain_marking_stack
1642 // oop closures will set the has_overflown flag if we overflow the
1643 // global marking stack.
1644
1645 assert(has_overflown() || _global_mark_stack.is_empty(),
1646 "Mark stack should be empty (unless it has overflown)");
1647
1648 assert(rp->num_queues() == active_workers, "why not");
1649
1650 rp->verify_no_references_recorded();
1651 assert(!rp->discovery_enabled(), "Post condition");
1652 }
1653
1654 if (has_overflown()) {
1655 // We can not trust g1_is_alive and the contents of the heap if the marking stack
1656 // overflowed while processing references. Exit the VM.
1657 fatal("Overflow during reference processing, can not continue. Please "
1658 "increase MarkStackSizeMax (current value: " SIZE_FORMAT ") and "
1659 "restart.", MarkStackSizeMax);
1660 return;
1661 }
1662
1663 assert(_global_mark_stack.is_empty(), "Marking should have completed");
1664
1665 {
1666 GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1667 WeakProcessor::weak_oops_do(_g1h->workers(), &g1_is_alive, &do_nothing_cl, 1);
1668 }
1669
1670 // Unload Klasses, String, Code Cache, etc.
1671 if (ClassUnloadingWithConcurrentMark) {
1672 GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1673 bool purged_classes = SystemDictionary::do_unloading(_gc_timer_cm);
1674 _g1h->complete_cleaning(&g1_is_alive, purged_classes);
1675 } else if (StringDedup::is_enabled()) {
1676 GCTraceTime(Debug, gc, phases) debug("String Deduplication", _gc_timer_cm);
1677 _g1h->string_dedup_cleaning(&g1_is_alive, NULL);
1678 }
1679 }
1680
1681 class G1PrecleanYieldClosure : public YieldClosure {
1682 G1ConcurrentMark* _cm;
1683
1684 public:
1685 G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { }
1686
1687 virtual bool should_return() {
1688 return _cm->has_aborted();
1689 }
1690
1691 virtual bool should_return_fine_grain() {
1692 _cm->do_yield_check();
1693 return _cm->has_aborted();
1694 }
1695 };
1696
1697 void G1ConcurrentMark::preclean() {
1698 assert(G1UseReferencePrecleaning, "Precleaning must be enabled.");
1699
1700 SuspendibleThreadSetJoiner joiner;
1701
1702 G1CMKeepAliveAndDrainClosure keep_alive(this, task(0), true /* is_serial */);
1703 G1CMDrainMarkingStackClosure drain_mark_stack(this, task(0), true /* is_serial */);
1704
1705 set_concurrency_and_phase(1, true);
1706
1707 G1PrecleanYieldClosure yield_cl(this);
1708
1709 ReferenceProcessor* rp = _g1h->ref_processor_cm();
1710 // Precleaning is single threaded. Temporarily disable MT discovery.
1711 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
1712 rp->preclean_discovered_references(rp->is_alive_non_header(),
1713 &keep_alive,
1714 &drain_mark_stack,
1715 &yield_cl,
1716 _gc_timer_cm);
1717 }
1718
1719 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
1720 // the prev bitmap determining liveness.
1721 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {
1722 G1CollectedHeap* _g1h;
1723 public:
1724 G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
1725
1726 bool do_object_b(oop obj) {
1727 return obj != NULL &&
1728 (!_g1h->is_in_g1_reserved(obj) || !_g1h->is_obj_dead(obj));
1729 }
1730 };
1731
1732 void G1ConcurrentMark::report_object_count(bool mark_completed) {
1733 // Depending on the completion of the marking liveness needs to be determined
1734 // using either the next or prev bitmap.
1735 if (mark_completed) {
1736 G1ObjectCountIsAliveClosure is_alive(_g1h);
1737 _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1738 } else {
1739 G1CMIsAliveClosure is_alive(_g1h);
1740 _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1741 }
1742 }
1743
1744
1745 void G1ConcurrentMark::swap_mark_bitmaps() {
1746 G1CMBitMap* temp = _prev_mark_bitmap;
1747 _prev_mark_bitmap = _next_mark_bitmap;
1748 _next_mark_bitmap = temp;
1749 _g1h->collector_state()->set_clearing_next_bitmap(true);
1750 }
1751
1752 // Closure for marking entries in SATB buffers.
1753 class G1CMSATBBufferClosure : public SATBBufferClosure {
1754 private:
1755 G1CMTask* _task;
1756 G1CollectedHeap* _g1h;
1757
1758 // This is very similar to G1CMTask::deal_with_reference, but with
1759 // more relaxed requirements for the argument, so this must be more
1760 // circumspect about treating the argument as an object.
1761 void do_entry(void* entry) const {
1762 _task->increment_refs_reached();
1763 oop const obj = static_cast<oop>(entry);
1764 _task->make_reference_grey(obj);
1765 }
1766
1767 public:
1768 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1769 : _task(task), _g1h(g1h) { }
1770
1771 virtual void do_buffer(void** buffer, size_t size) {
1772 for (size_t i = 0; i < size; ++i) {
1773 do_entry(buffer[i]);
1774 }
1775 }
1776 };
1777
1778 class G1RemarkThreadsClosure : public ThreadClosure {
1779 G1CMSATBBufferClosure _cm_satb_cl;
1780 G1CMOopClosure _cm_cl;
1781 MarkingCodeBlobClosure _code_cl;
1782 uintx _claim_token;
1783
1784 public:
1785 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1786 _cm_satb_cl(task, g1h),
1787 _cm_cl(g1h, task),
1788 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1789 _claim_token(Threads::thread_claim_token()) {}
1790
1791 void do_thread(Thread* thread) {
1792 if (thread->claim_threads_do(true, _claim_token)) {
1793 SATBMarkQueue& queue = G1ThreadLocalData::satb_mark_queue(thread);
1794 queue.apply_closure_and_empty(&_cm_satb_cl);
1795 if (thread->is_Java_thread()) {
1796 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1797 // however the liveness of oops reachable from nmethods have very complex lifecycles:
1798 // * Alive if on the stack of an executing method
1799 // * Weakly reachable otherwise
1800 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1801 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1802 JavaThread* jt = (JavaThread*)thread;
1803 jt->nmethods_do(&_code_cl);
1804 }
1805 }
1806 }
1807 };
1808
1809 class G1CMRemarkTask : public AbstractGangTask {
1810 G1ConcurrentMark* _cm;
1811 public:
1812 void work(uint worker_id) {
1813 G1CMTask* task = _cm->task(worker_id);
1814 task->record_start_time();
1815 {
1816 ResourceMark rm;
1817 HandleMark hm;
1818
1819 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1820 Threads::threads_do(&threads_f);
1821 }
1822
1823 do {
1824 task->do_marking_step(1000000000.0 /* something very large */,
1825 true /* do_termination */,
1826 false /* is_serial */);
1827 } while (task->has_aborted() && !_cm->has_overflown());
1828 // If we overflow, then we do not want to restart. We instead
1829 // want to abort remark and do concurrent marking again.
1830 task->record_end_time();
1831 }
1832
1833 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1834 AbstractGangTask("Par Remark"), _cm(cm) {
1835 _cm->terminator()->reset_for_reuse(active_workers);
1836 }
1837 };
1838
1839 void G1ConcurrentMark::finalize_marking() {
1840 ResourceMark rm;
1841 HandleMark hm;
1842
1843 _g1h->ensure_parsability(false);
1844
1845 // this is remark, so we'll use up all active threads
1846 uint active_workers = _g1h->workers()->active_workers();
1847 set_concurrency_and_phase(active_workers, false /* concurrent */);
1848 // Leave _parallel_marking_threads at it's
1849 // value originally calculated in the G1ConcurrentMark
1850 // constructor and pass values of the active workers
1851 // through the gang in the task.
1852
1853 {
1854 StrongRootsScope srs(active_workers);
1855
1856 G1CMRemarkTask remarkTask(this, active_workers);
1857 // We will start all available threads, even if we decide that the
1858 // active_workers will be fewer. The extra ones will just bail out
1859 // immediately.
1860 _g1h->workers()->run_task(&remarkTask);
1861 }
1862
1863 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1864 guarantee(has_overflown() ||
1865 satb_mq_set.completed_buffers_num() == 0,
1866 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1867 BOOL_TO_STR(has_overflown()),
1868 satb_mq_set.completed_buffers_num());
1869
1870 print_stats();
1871 }
1872
1873 void G1ConcurrentMark::flush_all_task_caches() {
1874 size_t hits = 0;
1875 size_t misses = 0;
1876 for (uint i = 0; i < _max_num_tasks; i++) {
1877 Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache();
1878 hits += stats.first;
1879 misses += stats.second;
1880 }
1881 size_t sum = hits + misses;
1882 log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf",
1883 hits, misses, percent_of(hits, sum));
1884 }
1885
1886 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1887 _prev_mark_bitmap->clear_range(mr);
1888 }
1889
1890 HeapRegion*
1891 G1ConcurrentMark::claim_region(uint worker_id) {
1892 // "checkpoint" the finger
1893 HeapWord* finger = _finger;
1894
1895 while (finger < _heap.end()) {
1896 assert(_g1h->is_in_g1_reserved(finger), "invariant");
1897
1898 HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1899 // Make sure that the reads below do not float before loading curr_region.
1900 OrderAccess::loadload();
1901 // Above heap_region_containing may return NULL as we always scan claim
1902 // until the end of the heap. In this case, just jump to the next region.
1903 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1904
1905 // Is the gap between reading the finger and doing the CAS too long?
1906 HeapWord* res = Atomic::cmpxchg(&_finger, finger, end);
1907 if (res == finger && curr_region != NULL) {
1908 // we succeeded
1909 HeapWord* bottom = curr_region->bottom();
1910 HeapWord* limit = curr_region->next_top_at_mark_start();
1911
1912 // notice that _finger == end cannot be guaranteed here since,
1913 // someone else might have moved the finger even further
1914 assert(_finger >= end, "the finger should have moved forward");
1915
1916 if (limit > bottom) {
1917 return curr_region;
1918 } else {
1919 assert(limit == bottom,
1920 "the region limit should be at bottom");
1921 // we return NULL and the caller should try calling
1922 // claim_region() again.
1923 return NULL;
1924 }
1925 } else {
1926 assert(_finger > finger, "the finger should have moved forward");
1927 // read it again
1928 finger = _finger;
1929 }
1930 }
1931
1932 return NULL;
1933 }
1934
1935 #ifndef PRODUCT
1936 class VerifyNoCSetOops {
1937 G1CollectedHeap* _g1h;
1938 const char* _phase;
1939 int _info;
1940
1941 public:
1942 VerifyNoCSetOops(const char* phase, int info = -1) :
1943 _g1h(G1CollectedHeap::heap()),
1944 _phase(phase),
1945 _info(info)
1946 { }
1947
1948 void operator()(G1TaskQueueEntry task_entry) const {
1949 if (task_entry.is_array_slice()) {
1950 guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1951 return;
1952 }
1953 guarantee(oopDesc::is_oop(task_entry.obj()),
1954 "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1955 p2i(task_entry.obj()), _phase, _info);
1956 HeapRegion* r = _g1h->heap_region_containing(task_entry.obj());
1957 guarantee(!(r->in_collection_set() || r->has_index_in_opt_cset()),
1958 "obj " PTR_FORMAT " from %s (%d) in region %u in (optional) collection set",
1959 p2i(task_entry.obj()), _phase, _info, r->hrm_index());
1960 }
1961 };
1962
1963 void G1ConcurrentMark::verify_no_collection_set_oops() {
1964 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1965 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
1966 return;
1967 }
1968
1969 // Verify entries on the global mark stack
1970 _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1971
1972 // Verify entries on the task queues
1973 for (uint i = 0; i < _max_num_tasks; ++i) {
1974 G1CMTaskQueue* queue = _task_queues->queue(i);
1975 queue->iterate(VerifyNoCSetOops("Queue", i));
1976 }
1977
1978 // Verify the global finger
1979 HeapWord* global_finger = finger();
1980 if (global_finger != NULL && global_finger < _heap.end()) {
1981 // Since we always iterate over all regions, we might get a NULL HeapRegion
1982 // here.
1983 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1984 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1985 "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1986 p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1987 }
1988
1989 // Verify the task fingers
1990 assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
1991 for (uint i = 0; i < _num_concurrent_workers; ++i) {
1992 G1CMTask* task = _tasks[i];
1993 HeapWord* task_finger = task->finger();
1994 if (task_finger != NULL && task_finger < _heap.end()) {
1995 // See above note on the global finger verification.
1996 HeapRegion* r = _g1h->heap_region_containing(task_finger);
1997 guarantee(r == NULL || task_finger == r->bottom() ||
1998 !r->in_collection_set() || !r->has_index_in_opt_cset(),
1999 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
2000 p2i(task_finger), HR_FORMAT_PARAMS(r));
2001 }
2002 }
2003 }
2004 #endif // PRODUCT
2005
2006 void G1ConcurrentMark::rebuild_rem_set_concurrently() {
2007 _g1h->rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset);
2008 }
2009
2010 void G1ConcurrentMark::print_stats() {
2011 if (!log_is_enabled(Debug, gc, stats)) {
2012 return;
2013 }
2014 log_debug(gc, stats)("---------------------------------------------------------------------");
2015 for (size_t i = 0; i < _num_active_tasks; ++i) {
2016 _tasks[i]->print_stats();
2017 log_debug(gc, stats)("---------------------------------------------------------------------");
2018 }
2019 }
2020
2021 void G1ConcurrentMark::concurrent_cycle_abort() {
2022 if (!cm_thread()->during_cycle() || _has_aborted) {
2023 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2024 return;
2025 }
2026
2027 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2028 // concurrent bitmap clearing.
2029 {
2030 GCTraceTime(Debug, gc) debug("Clear Next Bitmap");
2031 clear_bitmap(_next_mark_bitmap, _g1h->workers(), false);
2032 }
2033 // Note we cannot clear the previous marking bitmap here
2034 // since VerifyDuringGC verifies the objects marked during
2035 // a full GC against the previous bitmap.
2036
2037 // Empty mark stack
2038 reset_marking_for_restart();
2039 for (uint i = 0; i < _max_num_tasks; ++i) {
2040 _tasks[i]->clear_region_fields();
2041 }
2042 _first_overflow_barrier_sync.abort();
2043 _second_overflow_barrier_sync.abort();
2044 _has_aborted = true;
2045
2046 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2047 satb_mq_set.abandon_partial_marking();
2048 // This can be called either during or outside marking, we'll read
2049 // the expected_active value from the SATB queue set.
2050 satb_mq_set.set_active_all_threads(
2051 false, /* new active value */
2052 satb_mq_set.is_active() /* expected_active */);
2053 }
2054
2055 static void print_ms_time_info(const char* prefix, const char* name,
2056 NumberSeq& ns) {
2057 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2058 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2059 if (ns.num() > 0) {
2060 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]",
2061 prefix, ns.sd(), ns.maximum());
2062 }
2063 }
2064
2065 void G1ConcurrentMark::print_summary_info() {
2066 Log(gc, marking) log;
2067 if (!log.is_trace()) {
2068 return;
2069 }
2070
2071 log.trace(" Concurrent marking:");
2072 print_ms_time_info(" ", "init marks", _init_times);
2073 print_ms_time_info(" ", "remarks", _remark_times);
2074 {
2075 print_ms_time_info(" ", "final marks", _remark_mark_times);
2076 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
2077
2078 }
2079 print_ms_time_info(" ", "cleanups", _cleanup_times);
2080 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2081 _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2082 log.trace(" Total stop_world time = %8.2f s.",
2083 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2084 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).",
2085 cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2086 }
2087
2088 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2089 _concurrent_workers->print_worker_threads_on(st);
2090 }
2091
2092 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2093 _concurrent_workers->threads_do(tc);
2094 }
2095
2096 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2097 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2098 p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2099 _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2100 _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2101 }
2102
2103 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2104 ReferenceProcessor* result = g1h->ref_processor_cm();
2105 assert(result != NULL, "CM reference processor should not be NULL");
2106 return result;
2107 }
2108
2109 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2110 G1CMTask* task)
2111 : MetadataVisitingOopIterateClosure(get_cm_oop_closure_ref_processor(g1h)),
2112 _g1h(g1h), _task(task)
2113 { }
2114
2115 void G1CMTask::setup_for_region(HeapRegion* hr) {
2116 assert(hr != NULL,
2117 "claim_region() should have filtered out NULL regions");
2118 _curr_region = hr;
2119 _finger = hr->bottom();
2120 update_region_limit();
2121 }
2122
2123 void G1CMTask::update_region_limit() {
2124 HeapRegion* hr = _curr_region;
2125 HeapWord* bottom = hr->bottom();
2126 HeapWord* limit = hr->next_top_at_mark_start();
2127
2128 if (limit == bottom) {
2129 // The region was collected underneath our feet.
2130 // We set the finger to bottom to ensure that the bitmap
2131 // iteration that will follow this will not do anything.
2132 // (this is not a condition that holds when we set the region up,
2133 // as the region is not supposed to be empty in the first place)
2134 _finger = bottom;
2135 } else if (limit >= _region_limit) {
2136 assert(limit >= _finger, "peace of mind");
2137 } else {
2138 assert(limit < _region_limit, "only way to get here");
2139 // This can happen under some pretty unusual circumstances. An
2140 // evacuation pause empties the region underneath our feet (NTAMS
2141 // at bottom). We then do some allocation in the region (NTAMS
2142 // stays at bottom), followed by the region being used as a GC
2143 // alloc region (NTAMS will move to top() and the objects
2144 // originally below it will be grayed). All objects now marked in
2145 // the region are explicitly grayed, if below the global finger,
2146 // and we do not need in fact to scan anything else. So, we simply
2147 // set _finger to be limit to ensure that the bitmap iteration
2148 // doesn't do anything.
2149 _finger = limit;
2150 }
2151
2152 _region_limit = limit;
2153 }
2154
2155 void G1CMTask::giveup_current_region() {
2156 assert(_curr_region != NULL, "invariant");
2157 clear_region_fields();
2158 }
2159
2160 void G1CMTask::clear_region_fields() {
2161 // Values for these three fields that indicate that we're not
2162 // holding on to a region.
2163 _curr_region = NULL;
2164 _finger = NULL;
2165 _region_limit = NULL;
2166 }
2167
2168 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2169 if (cm_oop_closure == NULL) {
2170 assert(_cm_oop_closure != NULL, "invariant");
2171 } else {
2172 assert(_cm_oop_closure == NULL, "invariant");
2173 }
2174 _cm_oop_closure = cm_oop_closure;
2175 }
2176
2177 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2178 guarantee(next_mark_bitmap != NULL, "invariant");
2179 _next_mark_bitmap = next_mark_bitmap;
2180 clear_region_fields();
2181
2182 _calls = 0;
2183 _elapsed_time_ms = 0.0;
2184 _termination_time_ms = 0.0;
2185 _termination_start_time_ms = 0.0;
2186
2187 _mark_stats_cache.reset();
2188 }
2189
2190 bool G1CMTask::should_exit_termination() {
2191 if (!regular_clock_call()) {
2192 return true;
2193 }
2194
2195 // This is called when we are in the termination protocol. We should
2196 // quit if, for some reason, this task wants to abort or the global
2197 // stack is not empty (this means that we can get work from it).
2198 return !_cm->mark_stack_empty() || has_aborted();
2199 }
2200
2201 void G1CMTask::reached_limit() {
2202 assert(_words_scanned >= _words_scanned_limit ||
2203 _refs_reached >= _refs_reached_limit ,
2204 "shouldn't have been called otherwise");
2205 abort_marking_if_regular_check_fail();
2206 }
2207
2208 bool G1CMTask::regular_clock_call() {
2209 if (has_aborted()) {
2210 return false;
2211 }
2212
2213 // First, we need to recalculate the words scanned and refs reached
2214 // limits for the next clock call.
2215 recalculate_limits();
2216
2217 // During the regular clock call we do the following
2218
2219 // (1) If an overflow has been flagged, then we abort.
2220 if (_cm->has_overflown()) {
2221 return false;
2222 }
2223
2224 // If we are not concurrent (i.e. we're doing remark) we don't need
2225 // to check anything else. The other steps are only needed during
2226 // the concurrent marking phase.
2227 if (!_cm->concurrent()) {
2228 return true;
2229 }
2230
2231 // (2) If marking has been aborted for Full GC, then we also abort.
2232 if (_cm->has_aborted()) {
2233 return false;
2234 }
2235
2236 double curr_time_ms = os::elapsedVTime() * 1000.0;
2237
2238 // (4) We check whether we should yield. If we have to, then we abort.
2239 if (SuspendibleThreadSet::should_yield()) {
2240 // We should yield. To do this we abort the task. The caller is
2241 // responsible for yielding.
2242 return false;
2243 }
2244
2245 // (5) We check whether we've reached our time quota. If we have,
2246 // then we abort.
2247 double elapsed_time_ms = curr_time_ms - _start_time_ms;
2248 if (elapsed_time_ms > _time_target_ms) {
2249 _has_timed_out = true;
2250 return false;
2251 }
2252
2253 // (6) Finally, we check whether there are enough completed STAB
2254 // buffers available for processing. If there are, we abort.
2255 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2256 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2257 // we do need to process SATB buffers, we'll abort and restart
2258 // the marking task to do so
2259 return false;
2260 }
2261 return true;
2262 }
2263
2264 void G1CMTask::recalculate_limits() {
2265 _real_words_scanned_limit = _words_scanned + words_scanned_period;
2266 _words_scanned_limit = _real_words_scanned_limit;
2267
2268 _real_refs_reached_limit = _refs_reached + refs_reached_period;
2269 _refs_reached_limit = _real_refs_reached_limit;
2270 }
2271
2272 void G1CMTask::decrease_limits() {
2273 // This is called when we believe that we're going to do an infrequent
2274 // operation which will increase the per byte scanned cost (i.e. move
2275 // entries to/from the global stack). It basically tries to decrease the
2276 // scanning limit so that the clock is called earlier.
2277
2278 _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2279 _refs_reached_limit = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2280 }
2281
2282 void G1CMTask::move_entries_to_global_stack() {
2283 // Local array where we'll store the entries that will be popped
2284 // from the local queue.
2285 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2286
2287 size_t n = 0;
2288 G1TaskQueueEntry task_entry;
2289 while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2290 buffer[n] = task_entry;
2291 ++n;
2292 }
2293 if (n < G1CMMarkStack::EntriesPerChunk) {
2294 buffer[n] = G1TaskQueueEntry();
2295 }
2296
2297 if (n > 0) {
2298 if (!_cm->mark_stack_push(buffer)) {
2299 set_has_aborted();
2300 }
2301 }
2302
2303 // This operation was quite expensive, so decrease the limits.
2304 decrease_limits();
2305 }
2306
2307 bool G1CMTask::get_entries_from_global_stack() {
2308 // Local array where we'll store the entries that will be popped
2309 // from the global stack.
2310 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2311
2312 if (!_cm->mark_stack_pop(buffer)) {
2313 return false;
2314 }
2315
2316 // We did actually pop at least one entry.
2317 for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2318 G1TaskQueueEntry task_entry = buffer[i];
2319 if (task_entry.is_null()) {
2320 break;
2321 }
2322 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()));
2323 bool success = _task_queue->push(task_entry);
2324 // We only call this when the local queue is empty or under a
2325 // given target limit. So, we do not expect this push to fail.
2326 assert(success, "invariant");
2327 }
2328
2329 // This operation was quite expensive, so decrease the limits
2330 decrease_limits();
2331 return true;
2332 }
2333
2334 void G1CMTask::drain_local_queue(bool partially) {
2335 if (has_aborted()) {
2336 return;
2337 }
2338
2339 // Decide what the target size is, depending whether we're going to
2340 // drain it partially (so that other tasks can steal if they run out
2341 // of things to do) or totally (at the very end).
2342 size_t target_size;
2343 if (partially) {
2344 target_size = MIN2((size_t)_task_queue->max_elems()/3, (size_t)GCDrainStackTargetSize);
2345 } else {
2346 target_size = 0;
2347 }
2348
2349 if (_task_queue->size() > target_size) {
2350 G1TaskQueueEntry entry;
2351 bool ret = _task_queue->pop_local(entry);
2352 while (ret) {
2353 scan_task_entry(entry);
2354 if (_task_queue->size() <= target_size || has_aborted()) {
2355 ret = false;
2356 } else {
2357 ret = _task_queue->pop_local(entry);
2358 }
2359 }
2360 }
2361 }
2362
2363 void G1CMTask::drain_global_stack(bool partially) {
2364 if (has_aborted()) {
2365 return;
2366 }
2367
2368 // We have a policy to drain the local queue before we attempt to
2369 // drain the global stack.
2370 assert(partially || _task_queue->size() == 0, "invariant");
2371
2372 // Decide what the target size is, depending whether we're going to
2373 // drain it partially (so that other tasks can steal if they run out
2374 // of things to do) or totally (at the very end).
2375 // Notice that when draining the global mark stack partially, due to the racyness
2376 // of the mark stack size update we might in fact drop below the target. But,
2377 // this is not a problem.
2378 // In case of total draining, we simply process until the global mark stack is
2379 // totally empty, disregarding the size counter.
2380 if (partially) {
2381 size_t const target_size = _cm->partial_mark_stack_size_target();
2382 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2383 if (get_entries_from_global_stack()) {
2384 drain_local_queue(partially);
2385 }
2386 }
2387 } else {
2388 while (!has_aborted() && get_entries_from_global_stack()) {
2389 drain_local_queue(partially);
2390 }
2391 }
2392 }
2393
2394 // SATB Queue has several assumptions on whether to call the par or
2395 // non-par versions of the methods. this is why some of the code is
2396 // replicated. We should really get rid of the single-threaded version
2397 // of the code to simplify things.
2398 void G1CMTask::drain_satb_buffers() {
2399 if (has_aborted()) {
2400 return;
2401 }
2402
2403 // We set this so that the regular clock knows that we're in the
2404 // middle of draining buffers and doesn't set the abort flag when it
2405 // notices that SATB buffers are available for draining. It'd be
2406 // very counter productive if it did that. :-)
2407 _draining_satb_buffers = true;
2408
2409 G1CMSATBBufferClosure satb_cl(this, _g1h);
2410 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2411
2412 // This keeps claiming and applying the closure to completed buffers
2413 // until we run out of buffers or we need to abort.
2414 while (!has_aborted() &&
2415 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2416 abort_marking_if_regular_check_fail();
2417 }
2418
2419 // Can't assert qset is empty here, even if not aborted. If concurrent,
2420 // some other thread might be adding to the queue. If not concurrent,
2421 // some other thread might have won the race for the last buffer, but
2422 // has not yet decremented the count.
2423
2424 _draining_satb_buffers = false;
2425
2426 // again, this was a potentially expensive operation, decrease the
2427 // limits to get the regular clock call early
2428 decrease_limits();
2429 }
2430
2431 void G1CMTask::clear_mark_stats_cache(uint region_idx) {
2432 _mark_stats_cache.reset(region_idx);
2433 }
2434
2435 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() {
2436 return _mark_stats_cache.evict_all();
2437 }
2438
2439 void G1CMTask::print_stats() {
2440 log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls);
2441 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2442 _elapsed_time_ms, _termination_time_ms);
2443 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms",
2444 _step_times_ms.num(),
2445 _step_times_ms.avg(),
2446 _step_times_ms.sd(),
2447 _step_times_ms.maximum(),
2448 _step_times_ms.sum());
2449 size_t const hits = _mark_stats_cache.hits();
2450 size_t const misses = _mark_stats_cache.misses();
2451 log_debug(gc, stats)(" Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f",
2452 hits, misses, percent_of(hits, hits + misses));
2453 }
2454
2455 bool G1ConcurrentMark::try_stealing(uint worker_id, G1TaskQueueEntry& task_entry) {
2456 return _task_queues->steal(worker_id, task_entry);
2457 }
2458
2459 /*****************************************************************************
2460
2461 The do_marking_step(time_target_ms, ...) method is the building
2462 block of the parallel marking framework. It can be called in parallel
2463 with other invocations of do_marking_step() on different tasks
2464 (but only one per task, obviously) and concurrently with the
2465 mutator threads, or during remark, hence it eliminates the need
2466 for two versions of the code. When called during remark, it will
2467 pick up from where the task left off during the concurrent marking
2468 phase. Interestingly, tasks are also claimable during evacuation
2469 pauses too, since do_marking_step() ensures that it aborts before
2470 it needs to yield.
2471
2472 The data structures that it uses to do marking work are the
2473 following:
2474
2475 (1) Marking Bitmap. If there are gray objects that appear only
2476 on the bitmap (this happens either when dealing with an overflow
2477 or when the initial marking phase has simply marked the roots
2478 and didn't push them on the stack), then tasks claim heap
2479 regions whose bitmap they then scan to find gray objects. A
2480 global finger indicates where the end of the last claimed region
2481 is. A local finger indicates how far into the region a task has
2482 scanned. The two fingers are used to determine how to gray an
2483 object (i.e. whether simply marking it is OK, as it will be
2484 visited by a task in the future, or whether it needs to be also
2485 pushed on a stack).
2486
2487 (2) Local Queue. The local queue of the task which is accessed
2488 reasonably efficiently by the task. Other tasks can steal from
2489 it when they run out of work. Throughout the marking phase, a
2490 task attempts to keep its local queue short but not totally
2491 empty, so that entries are available for stealing by other
2492 tasks. Only when there is no more work, a task will totally
2493 drain its local queue.
2494
2495 (3) Global Mark Stack. This handles local queue overflow. During
2496 marking only sets of entries are moved between it and the local
2497 queues, as access to it requires a mutex and more fine-grain
2498 interaction with it which might cause contention. If it
2499 overflows, then the marking phase should restart and iterate
2500 over the bitmap to identify gray objects. Throughout the marking
2501 phase, tasks attempt to keep the global mark stack at a small
2502 length but not totally empty, so that entries are available for
2503 popping by other tasks. Only when there is no more work, tasks
2504 will totally drain the global mark stack.
2505
2506 (4) SATB Buffer Queue. This is where completed SATB buffers are
2507 made available. Buffers are regularly removed from this queue
2508 and scanned for roots, so that the queue doesn't get too
2509 long. During remark, all completed buffers are processed, as
2510 well as the filled in parts of any uncompleted buffers.
2511
2512 The do_marking_step() method tries to abort when the time target
2513 has been reached. There are a few other cases when the
2514 do_marking_step() method also aborts:
2515
2516 (1) When the marking phase has been aborted (after a Full GC).
2517
2518 (2) When a global overflow (on the global stack) has been
2519 triggered. Before the task aborts, it will actually sync up with
2520 the other tasks to ensure that all the marking data structures
2521 (local queues, stacks, fingers etc.) are re-initialized so that
2522 when do_marking_step() completes, the marking phase can
2523 immediately restart.
2524
2525 (3) When enough completed SATB buffers are available. The
2526 do_marking_step() method only tries to drain SATB buffers right
2527 at the beginning. So, if enough buffers are available, the
2528 marking step aborts and the SATB buffers are processed at
2529 the beginning of the next invocation.
2530
2531 (4) To yield. when we have to yield then we abort and yield
2532 right at the end of do_marking_step(). This saves us from a lot
2533 of hassle as, by yielding we might allow a Full GC. If this
2534 happens then objects will be compacted underneath our feet, the
2535 heap might shrink, etc. We save checking for this by just
2536 aborting and doing the yield right at the end.
2537
2538 From the above it follows that the do_marking_step() method should
2539 be called in a loop (or, otherwise, regularly) until it completes.
2540
2541 If a marking step completes without its has_aborted() flag being
2542 true, it means it has completed the current marking phase (and
2543 also all other marking tasks have done so and have all synced up).
2544
2545 A method called regular_clock_call() is invoked "regularly" (in
2546 sub ms intervals) throughout marking. It is this clock method that
2547 checks all the abort conditions which were mentioned above and
2548 decides when the task should abort. A work-based scheme is used to
2549 trigger this clock method: when the number of object words the
2550 marking phase has scanned or the number of references the marking
2551 phase has visited reach a given limit. Additional invocations to
2552 the method clock have been planted in a few other strategic places
2553 too. The initial reason for the clock method was to avoid calling
2554 vtime too regularly, as it is quite expensive. So, once it was in
2555 place, it was natural to piggy-back all the other conditions on it
2556 too and not constantly check them throughout the code.
2557
2558 If do_termination is true then do_marking_step will enter its
2559 termination protocol.
2560
2561 The value of is_serial must be true when do_marking_step is being
2562 called serially (i.e. by the VMThread) and do_marking_step should
2563 skip any synchronization in the termination and overflow code.
2564 Examples include the serial remark code and the serial reference
2565 processing closures.
2566
2567 The value of is_serial must be false when do_marking_step is
2568 being called by any of the worker threads in a work gang.
2569 Examples include the concurrent marking code (CMMarkingTask),
2570 the MT remark code, and the MT reference processing closures.
2571
2572 *****************************************************************************/
2573
2574 void G1CMTask::do_marking_step(double time_target_ms,
2575 bool do_termination,
2576 bool is_serial) {
2577 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2578
2579 _start_time_ms = os::elapsedVTime() * 1000.0;
2580
2581 // If do_stealing is true then do_marking_step will attempt to
2582 // steal work from the other G1CMTasks. It only makes sense to
2583 // enable stealing when the termination protocol is enabled
2584 // and do_marking_step() is not being called serially.
2585 bool do_stealing = do_termination && !is_serial;
2586
2587 G1Predictions const& predictor = _g1h->policy()->predictor();
2588 double diff_prediction_ms = predictor.predict_zero_bounded(&_marking_step_diff_ms);
2589 _time_target_ms = time_target_ms - diff_prediction_ms;
2590
2591 // set up the variables that are used in the work-based scheme to
2592 // call the regular clock method
2593 _words_scanned = 0;
2594 _refs_reached = 0;
2595 recalculate_limits();
2596
2597 // clear all flags
2598 clear_has_aborted();
2599 _has_timed_out = false;
2600 _draining_satb_buffers = false;
2601
2602 ++_calls;
2603
2604 // Set up the bitmap and oop closures. Anything that uses them is
2605 // eventually called from this method, so it is OK to allocate these
2606 // statically.
2607 G1CMBitMapClosure bitmap_closure(this, _cm);
2608 G1CMOopClosure cm_oop_closure(_g1h, this);
2609 set_cm_oop_closure(&cm_oop_closure);
2610
2611 if (_cm->has_overflown()) {
2612 // This can happen if the mark stack overflows during a GC pause
2613 // and this task, after a yield point, restarts. We have to abort
2614 // as we need to get into the overflow protocol which happens
2615 // right at the end of this task.
2616 set_has_aborted();
2617 }
2618
2619 // First drain any available SATB buffers. After this, we will not
2620 // look at SATB buffers before the next invocation of this method.
2621 // If enough completed SATB buffers are queued up, the regular clock
2622 // will abort this task so that it restarts.
2623 drain_satb_buffers();
2624 // ...then partially drain the local queue and the global stack
2625 drain_local_queue(true);
2626 drain_global_stack(true);
2627
2628 do {
2629 if (!has_aborted() && _curr_region != NULL) {
2630 // This means that we're already holding on to a region.
2631 assert(_finger != NULL, "if region is not NULL, then the finger "
2632 "should not be NULL either");
2633
2634 // We might have restarted this task after an evacuation pause
2635 // which might have evacuated the region we're holding on to
2636 // underneath our feet. Let's read its limit again to make sure
2637 // that we do not iterate over a region of the heap that
2638 // contains garbage (update_region_limit() will also move
2639 // _finger to the start of the region if it is found empty).
2640 update_region_limit();
2641 // We will start from _finger not from the start of the region,
2642 // as we might be restarting this task after aborting half-way
2643 // through scanning this region. In this case, _finger points to
2644 // the address where we last found a marked object. If this is a
2645 // fresh region, _finger points to start().
2646 MemRegion mr = MemRegion(_finger, _region_limit);
2647
2648 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2649 "humongous regions should go around loop once only");
2650
2651 // Some special cases:
2652 // If the memory region is empty, we can just give up the region.
2653 // If the current region is humongous then we only need to check
2654 // the bitmap for the bit associated with the start of the object,
2655 // scan the object if it's live, and give up the region.
2656 // Otherwise, let's iterate over the bitmap of the part of the region
2657 // that is left.
2658 // If the iteration is successful, give up the region.
2659 if (mr.is_empty()) {
2660 giveup_current_region();
2661 abort_marking_if_regular_check_fail();
2662 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2663 if (_next_mark_bitmap->is_marked(mr.start())) {
2664 // The object is marked - apply the closure
2665 bitmap_closure.do_addr(mr.start());
2666 }
2667 // Even if this task aborted while scanning the humongous object
2668 // we can (and should) give up the current region.
2669 giveup_current_region();
2670 abort_marking_if_regular_check_fail();
2671 } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2672 giveup_current_region();
2673 abort_marking_if_regular_check_fail();
2674 } else {
2675 assert(has_aborted(), "currently the only way to do so");
2676 // The only way to abort the bitmap iteration is to return
2677 // false from the do_bit() method. However, inside the
2678 // do_bit() method we move the _finger to point to the
2679 // object currently being looked at. So, if we bail out, we
2680 // have definitely set _finger to something non-null.
2681 assert(_finger != NULL, "invariant");
2682
2683 // Region iteration was actually aborted. So now _finger
2684 // points to the address of the object we last scanned. If we
2685 // leave it there, when we restart this task, we will rescan
2686 // the object. It is easy to avoid this. We move the finger by
2687 // enough to point to the next possible object header.
2688 assert(_finger < _region_limit, "invariant");
2689 HeapWord* const new_finger = _finger + ((oop)_finger)->size();
2690 // Check if bitmap iteration was aborted while scanning the last object
2691 if (new_finger >= _region_limit) {
2692 giveup_current_region();
2693 } else {
2694 move_finger_to(new_finger);
2695 }
2696 }
2697 }
2698 // At this point we have either completed iterating over the
2699 // region we were holding on to, or we have aborted.
2700
2701 // We then partially drain the local queue and the global stack.
2702 // (Do we really need this?)
2703 drain_local_queue(true);
2704 drain_global_stack(true);
2705
2706 // Read the note on the claim_region() method on why it might
2707 // return NULL with potentially more regions available for
2708 // claiming and why we have to check out_of_regions() to determine
2709 // whether we're done or not.
2710 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2711 // We are going to try to claim a new region. We should have
2712 // given up on the previous one.
2713 // Separated the asserts so that we know which one fires.
2714 assert(_curr_region == NULL, "invariant");
2715 assert(_finger == NULL, "invariant");
2716 assert(_region_limit == NULL, "invariant");
2717 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2718 if (claimed_region != NULL) {
2719 // Yes, we managed to claim one
2720 setup_for_region(claimed_region);
2721 assert(_curr_region == claimed_region, "invariant");
2722 }
2723 // It is important to call the regular clock here. It might take
2724 // a while to claim a region if, for example, we hit a large
2725 // block of empty regions. So we need to call the regular clock
2726 // method once round the loop to make sure it's called
2727 // frequently enough.
2728 abort_marking_if_regular_check_fail();
2729 }
2730
2731 if (!has_aborted() && _curr_region == NULL) {
2732 assert(_cm->out_of_regions(),
2733 "at this point we should be out of regions");
2734 }
2735 } while ( _curr_region != NULL && !has_aborted());
2736
2737 if (!has_aborted()) {
2738 // We cannot check whether the global stack is empty, since other
2739 // tasks might be pushing objects to it concurrently.
2740 assert(_cm->out_of_regions(),
2741 "at this point we should be out of regions");
2742 // Try to reduce the number of available SATB buffers so that
2743 // remark has less work to do.
2744 drain_satb_buffers();
2745 }
2746
2747 // Since we've done everything else, we can now totally drain the
2748 // local queue and global stack.
2749 drain_local_queue(false);
2750 drain_global_stack(false);
2751
2752 // Attempt at work stealing from other task's queues.
2753 if (do_stealing && !has_aborted()) {
2754 // We have not aborted. This means that we have finished all that
2755 // we could. Let's try to do some stealing...
2756
2757 // We cannot check whether the global stack is empty, since other
2758 // tasks might be pushing objects to it concurrently.
2759 assert(_cm->out_of_regions() && _task_queue->size() == 0,
2760 "only way to reach here");
2761 while (!has_aborted()) {
2762 G1TaskQueueEntry entry;
2763 if (_cm->try_stealing(_worker_id, entry)) {
2764 scan_task_entry(entry);
2765
2766 // And since we're towards the end, let's totally drain the
2767 // local queue and global stack.
2768 drain_local_queue(false);
2769 drain_global_stack(false);
2770 } else {
2771 break;
2772 }
2773 }
2774 }
2775
2776 // We still haven't aborted. Now, let's try to get into the
2777 // termination protocol.
2778 if (do_termination && !has_aborted()) {
2779 // We cannot check whether the global stack is empty, since other
2780 // tasks might be concurrently pushing objects on it.
2781 // Separated the asserts so that we know which one fires.
2782 assert(_cm->out_of_regions(), "only way to reach here");
2783 assert(_task_queue->size() == 0, "only way to reach here");
2784 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2785
2786 // The G1CMTask class also extends the TerminatorTerminator class,
2787 // hence its should_exit_termination() method will also decide
2788 // whether to exit the termination protocol or not.
2789 bool finished = (is_serial ||
2790 _cm->terminator()->offer_termination(this));
2791 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2792 _termination_time_ms +=
2793 termination_end_time_ms - _termination_start_time_ms;
2794
2795 if (finished) {
2796 // We're all done.
2797
2798 // We can now guarantee that the global stack is empty, since
2799 // all other tasks have finished. We separated the guarantees so
2800 // that, if a condition is false, we can immediately find out
2801 // which one.
2802 guarantee(_cm->out_of_regions(), "only way to reach here");
2803 guarantee(_cm->mark_stack_empty(), "only way to reach here");
2804 guarantee(_task_queue->size() == 0, "only way to reach here");
2805 guarantee(!_cm->has_overflown(), "only way to reach here");
2806 guarantee(!has_aborted(), "should never happen if termination has completed");
2807 } else {
2808 // Apparently there's more work to do. Let's abort this task. It
2809 // will restart it and we can hopefully find more things to do.
2810 set_has_aborted();
2811 }
2812 }
2813
2814 // Mainly for debugging purposes to make sure that a pointer to the
2815 // closure which was statically allocated in this frame doesn't
2816 // escape it by accident.
2817 set_cm_oop_closure(NULL);
2818 double end_time_ms = os::elapsedVTime() * 1000.0;
2819 double elapsed_time_ms = end_time_ms - _start_time_ms;
2820 // Update the step history.
2821 _step_times_ms.add(elapsed_time_ms);
2822
2823 if (has_aborted()) {
2824 // The task was aborted for some reason.
2825 if (_has_timed_out) {
2826 double diff_ms = elapsed_time_ms - _time_target_ms;
2827 // Keep statistics of how well we did with respect to hitting
2828 // our target only if we actually timed out (if we aborted for
2829 // other reasons, then the results might get skewed).
2830 _marking_step_diff_ms.add(diff_ms);
2831 }
2832
2833 if (_cm->has_overflown()) {
2834 // This is the interesting one. We aborted because a global
2835 // overflow was raised. This means we have to restart the
2836 // marking phase and start iterating over regions. However, in
2837 // order to do this we have to make sure that all tasks stop
2838 // what they are doing and re-initialize in a safe manner. We
2839 // will achieve this with the use of two barrier sync points.
2840
2841 if (!is_serial) {
2842 // We only need to enter the sync barrier if being called
2843 // from a parallel context
2844 _cm->enter_first_sync_barrier(_worker_id);
2845
2846 // When we exit this sync barrier we know that all tasks have
2847 // stopped doing marking work. So, it's now safe to
2848 // re-initialize our data structures.
2849 }
2850
2851 clear_region_fields();
2852 flush_mark_stats_cache();
2853
2854 if (!is_serial) {
2855 // If we're executing the concurrent phase of marking, reset the marking
2856 // state; otherwise the marking state is reset after reference processing,
2857 // during the remark pause.
2858 // If we reset here as a result of an overflow during the remark we will
2859 // see assertion failures from any subsequent set_concurrency_and_phase()
2860 // calls.
2861 if (_cm->concurrent() && _worker_id == 0) {
2862 // Worker 0 is responsible for clearing the global data structures because
2863 // of an overflow. During STW we should not clear the overflow flag (in
2864 // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit
2865 // method to abort the pause and restart concurrent marking.
2866 _cm->reset_marking_for_restart();
2867
2868 log_info(gc, marking)("Concurrent Mark reset for overflow");
2869 }
2870
2871 // ...and enter the second barrier.
2872 _cm->enter_second_sync_barrier(_worker_id);
2873 }
2874 // At this point, if we're during the concurrent phase of
2875 // marking, everything has been re-initialized and we're
2876 // ready to restart.
2877 }
2878 }
2879 }
2880
2881 G1CMTask::G1CMTask(uint worker_id,
2882 G1ConcurrentMark* cm,
2883 G1CMTaskQueue* task_queue,
2884 G1RegionMarkStats* mark_stats,
2885 uint max_regions) :
2886 _objArray_processor(this),
2887 _worker_id(worker_id),
2888 _g1h(G1CollectedHeap::heap()),
2889 _cm(cm),
2890 _next_mark_bitmap(NULL),
2891 _task_queue(task_queue),
2892 _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize),
2893 _calls(0),
2894 _time_target_ms(0.0),
2895 _start_time_ms(0.0),
2896 _cm_oop_closure(NULL),
2897 _curr_region(NULL),
2898 _finger(NULL),
2899 _region_limit(NULL),
2900 _words_scanned(0),
2901 _words_scanned_limit(0),
2902 _real_words_scanned_limit(0),
2903 _refs_reached(0),
2904 _refs_reached_limit(0),
2905 _real_refs_reached_limit(0),
2906 _has_aborted(false),
2907 _has_timed_out(false),
2908 _draining_satb_buffers(false),
2909 _step_times_ms(),
2910 _elapsed_time_ms(0.0),
2911 _termination_time_ms(0.0),
2912 _termination_start_time_ms(0.0),
2913 _marking_step_diff_ms()
2914 {
2915 guarantee(task_queue != NULL, "invariant");
2916
2917 _marking_step_diff_ms.add(0.5);
2918 }
2919
2920 // These are formatting macros that are used below to ensure
2921 // consistent formatting. The *_H_* versions are used to format the
2922 // header for a particular value and they should be kept consistent
2923 // with the corresponding macro. Also note that most of the macros add
2924 // the necessary white space (as a prefix) which makes them a bit
2925 // easier to compose.
2926
2927 // All the output lines are prefixed with this string to be able to
2928 // identify them easily in a large log file.
2929 #define G1PPRL_LINE_PREFIX "###"
2930
2931 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT
2932 #ifdef _LP64
2933 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
2934 #else // _LP64
2935 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
2936 #endif // _LP64
2937
2938 // For per-region info
2939 #define G1PPRL_TYPE_FORMAT " %-4s"
2940 #define G1PPRL_TYPE_H_FORMAT " %4s"
2941 #define G1PPRL_STATE_FORMAT " %-5s"
2942 #define G1PPRL_STATE_H_FORMAT " %5s"
2943 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9)
2944 #define G1PPRL_BYTE_H_FORMAT " %9s"
2945 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
2946 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
2947
2948 // For summary info
2949 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT
2950 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT
2951 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB"
2952 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2953
2954 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2955 _total_used_bytes(0), _total_capacity_bytes(0),
2956 _total_prev_live_bytes(0), _total_next_live_bytes(0),
2957 _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2958 {
2959 if (!log_is_enabled(Trace, gc, liveness)) {
2960 return;
2961 }
2962
2963 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2964 MemRegion g1_reserved = g1h->g1_reserved();
2965 double now = os::elapsedTime();
2966
2967 // Print the header of the output.
2968 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2969 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2970 G1PPRL_SUM_ADDR_FORMAT("reserved")
2971 G1PPRL_SUM_BYTE_FORMAT("region-size"),
2972 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2973 HeapRegion::GrainBytes);
2974 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2975 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2976 G1PPRL_TYPE_H_FORMAT
2977 G1PPRL_ADDR_BASE_H_FORMAT
2978 G1PPRL_BYTE_H_FORMAT
2979 G1PPRL_BYTE_H_FORMAT
2980 G1PPRL_BYTE_H_FORMAT
2981 G1PPRL_DOUBLE_H_FORMAT
2982 G1PPRL_BYTE_H_FORMAT
2983 G1PPRL_STATE_H_FORMAT
2984 G1PPRL_BYTE_H_FORMAT,
2985 "type", "address-range",
2986 "used", "prev-live", "next-live", "gc-eff",
2987 "remset", "state", "code-roots");
2988 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2989 G1PPRL_TYPE_H_FORMAT
2990 G1PPRL_ADDR_BASE_H_FORMAT
2991 G1PPRL_BYTE_H_FORMAT
2992 G1PPRL_BYTE_H_FORMAT
2993 G1PPRL_BYTE_H_FORMAT
2994 G1PPRL_DOUBLE_H_FORMAT
2995 G1PPRL_BYTE_H_FORMAT
2996 G1PPRL_STATE_H_FORMAT
2997 G1PPRL_BYTE_H_FORMAT,
2998 "", "",
2999 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3000 "(bytes)", "", "(bytes)");
3001 }
3002
3003 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) {
3004 if (!log_is_enabled(Trace, gc, liveness)) {
3005 return false;
3006 }
3007
3008 const char* type = r->get_type_str();
3009 HeapWord* bottom = r->bottom();
3010 HeapWord* end = r->end();
3011 size_t capacity_bytes = r->capacity();
3012 size_t used_bytes = r->used();
3013 size_t prev_live_bytes = r->live_bytes();
3014 size_t next_live_bytes = r->next_live_bytes();
3015 double gc_eff = r->gc_efficiency();
3016 size_t remset_bytes = r->rem_set()->mem_size();
3017 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3018 const char* remset_type = r->rem_set()->get_short_state_str();
3019
3020 _total_used_bytes += used_bytes;
3021 _total_capacity_bytes += capacity_bytes;
3022 _total_prev_live_bytes += prev_live_bytes;
3023 _total_next_live_bytes += next_live_bytes;
3024 _total_remset_bytes += remset_bytes;
3025 _total_strong_code_roots_bytes += strong_code_roots_bytes;
3026
3027 // Print a line for this particular region.
3028 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3029 G1PPRL_TYPE_FORMAT
3030 G1PPRL_ADDR_BASE_FORMAT
3031 G1PPRL_BYTE_FORMAT
3032 G1PPRL_BYTE_FORMAT
3033 G1PPRL_BYTE_FORMAT
3034 G1PPRL_DOUBLE_FORMAT
3035 G1PPRL_BYTE_FORMAT
3036 G1PPRL_STATE_FORMAT
3037 G1PPRL_BYTE_FORMAT,
3038 type, p2i(bottom), p2i(end),
3039 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3040 remset_bytes, remset_type, strong_code_roots_bytes);
3041
3042 return false;
3043 }
3044
3045 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3046 if (!log_is_enabled(Trace, gc, liveness)) {
3047 return;
3048 }
3049
3050 // add static memory usages to remembered set sizes
3051 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3052 // Print the footer of the output.
3053 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3054 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3055 " SUMMARY"
3056 G1PPRL_SUM_MB_FORMAT("capacity")
3057 G1PPRL_SUM_MB_PERC_FORMAT("used")
3058 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3059 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3060 G1PPRL_SUM_MB_FORMAT("remset")
3061 G1PPRL_SUM_MB_FORMAT("code-roots"),
3062 bytes_to_mb(_total_capacity_bytes),
3063 bytes_to_mb(_total_used_bytes),
3064 percent_of(_total_used_bytes, _total_capacity_bytes),
3065 bytes_to_mb(_total_prev_live_bytes),
3066 percent_of(_total_prev_live_bytes, _total_capacity_bytes),
3067 bytes_to_mb(_total_next_live_bytes),
3068 percent_of(_total_next_live_bytes, _total_capacity_bytes),
3069 bytes_to_mb(_total_remset_bytes),
3070 bytes_to_mb(_total_strong_code_roots_bytes));
3071 }