1 /*
2 * Copyright (c) 2001, 2013, 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/symbolTable.hpp"
27 #include "gc_implementation/g1/concurrentMark.inline.hpp"
28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
32 #include "gc_implementation/g1/g1Log.hpp"
33 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
34 #include "gc_implementation/g1/g1RemSet.hpp"
35 #include "gc_implementation/g1/heapRegion.inline.hpp"
36 #include "gc_implementation/g1/heapRegionRemSet.hpp"
37 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
38 #include "gc_implementation/shared/vmGCOperations.hpp"
39 #include "gc_implementation/shared/gcTimer.hpp"
40 #include "gc_implementation/shared/gcTrace.hpp"
41 #include "gc_implementation/shared/gcTraceTime.hpp"
42 #include "memory/genOopClosures.inline.hpp"
43 #include "memory/referencePolicy.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "oops/oop.inline.hpp"
46 #include "runtime/handles.inline.hpp"
47 #include "runtime/java.hpp"
48 #include "services/memTracker.hpp"
49
50 // Concurrent marking bit map wrapper
51
52 CMBitMapRO::CMBitMapRO(ReservedSpace rs, int shifter) :
53 _bm((uintptr_t*)NULL,0),
54 _shifter(shifter) {
55 _bmStartWord = (HeapWord*)(rs.base());
56 _bmWordSize = rs.size()/HeapWordSize; // rs.size() is in bytes
57 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
58 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
59
60 MemTracker::record_virtual_memory_type((address)brs.base(), mtGC);
61
62 guarantee(brs.is_reserved(), "couldn't allocate concurrent marking bit map");
63 // For now we'll just commit all of the bit map up fromt.
64 // Later on we'll try to be more parsimonious with swap.
65 guarantee(_virtual_space.initialize(brs, brs.size()),
66 "couldn't reseve backing store for concurrent marking bit map");
67 assert(_virtual_space.committed_size() == brs.size(),
68 "didn't reserve backing store for all of concurrent marking bit map?");
69 _bm.set_map((uintptr_t*)_virtual_space.low());
70 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
71 _bmWordSize, "inconsistency in bit map sizing");
72 _bm.set_size(_bmWordSize >> _shifter);
73 }
74
75 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr,
76 HeapWord* limit) const {
77 // First we must round addr *up* to a possible object boundary.
78 addr = (HeapWord*)align_size_up((intptr_t)addr,
79 HeapWordSize << _shifter);
80 size_t addrOffset = heapWordToOffset(addr);
81 if (limit == NULL) {
82 limit = _bmStartWord + _bmWordSize;
83 }
84 size_t limitOffset = heapWordToOffset(limit);
85 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
86 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
87 assert(nextAddr >= addr, "get_next_one postcondition");
88 assert(nextAddr == limit || isMarked(nextAddr),
89 "get_next_one postcondition");
90 return nextAddr;
91 }
92
93 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr,
94 HeapWord* limit) const {
95 size_t addrOffset = heapWordToOffset(addr);
96 if (limit == NULL) {
97 limit = _bmStartWord + _bmWordSize;
98 }
99 size_t limitOffset = heapWordToOffset(limit);
100 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
101 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
102 assert(nextAddr >= addr, "get_next_one postcondition");
103 assert(nextAddr == limit || !isMarked(nextAddr),
104 "get_next_one postcondition");
105 return nextAddr;
106 }
107
108 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
109 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
110 return (int) (diff >> _shifter);
111 }
112
113 #ifndef PRODUCT
114 bool CMBitMapRO::covers(ReservedSpace rs) const {
115 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
116 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
117 "size inconsistency");
118 return _bmStartWord == (HeapWord*)(rs.base()) &&
119 _bmWordSize == rs.size()>>LogHeapWordSize;
120 }
121 #endif
122
123 void CMBitMap::clearAll() {
124 _bm.clear();
125 return;
126 }
127
128 void CMBitMap::markRange(MemRegion mr) {
129 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
130 assert(!mr.is_empty(), "unexpected empty region");
131 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
132 ((HeapWord *) mr.end())),
133 "markRange memory region end is not card aligned");
134 // convert address range into offset range
135 _bm.at_put_range(heapWordToOffset(mr.start()),
136 heapWordToOffset(mr.end()), true);
137 }
138
139 void CMBitMap::clearRange(MemRegion mr) {
140 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
141 assert(!mr.is_empty(), "unexpected empty region");
142 // convert address range into offset range
143 _bm.at_put_range(heapWordToOffset(mr.start()),
144 heapWordToOffset(mr.end()), false);
145 }
146
147 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
148 HeapWord* end_addr) {
149 HeapWord* start = getNextMarkedWordAddress(addr);
150 start = MIN2(start, end_addr);
151 HeapWord* end = getNextUnmarkedWordAddress(start);
152 end = MIN2(end, end_addr);
153 assert(start <= end, "Consistency check");
154 MemRegion mr(start, end);
155 if (!mr.is_empty()) {
156 clearRange(mr);
157 }
158 return mr;
159 }
160
161 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
162 _base(NULL), _cm(cm)
163 #ifdef ASSERT
164 , _drain_in_progress(false)
165 , _drain_in_progress_yields(false)
166 #endif
167 {}
168
169 void CMMarkStack::allocate(size_t size) {
170 _base = NEW_C_HEAP_ARRAY(oop, size, mtGC);
171 if (_base == NULL) {
172 vm_exit_during_initialization("Failed to allocate CM region mark stack");
173 }
174 _index = 0;
175 _capacity = (jint) size;
176 _saved_index = -1;
177 NOT_PRODUCT(_max_depth = 0);
178 }
179
180 CMMarkStack::~CMMarkStack() {
181 if (_base != NULL) {
182 FREE_C_HEAP_ARRAY(oop, _base, mtGC);
183 }
184 }
185
186 void CMMarkStack::par_push(oop ptr) {
187 while (true) {
188 if (isFull()) {
189 _overflow = true;
190 return;
191 }
192 // Otherwise...
193 jint index = _index;
194 jint next_index = index+1;
195 jint res = Atomic::cmpxchg(next_index, &_index, index);
196 if (res == index) {
197 _base[index] = ptr;
198 // Note that we don't maintain this atomically. We could, but it
199 // doesn't seem necessary.
200 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
201 return;
202 }
203 // Otherwise, we need to try again.
204 }
205 }
206
207 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
208 while (true) {
209 if (isFull()) {
210 _overflow = true;
211 return;
212 }
213 // Otherwise...
214 jint index = _index;
215 jint next_index = index + n;
216 if (next_index > _capacity) {
217 _overflow = true;
218 return;
219 }
220 jint res = Atomic::cmpxchg(next_index, &_index, index);
221 if (res == index) {
222 for (int i = 0; i < n; i++) {
223 int ind = index + i;
224 assert(ind < _capacity, "By overflow test above.");
225 _base[ind] = ptr_arr[i];
226 }
227 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
228 return;
229 }
230 // Otherwise, we need to try again.
231 }
232 }
233
234
235 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
236 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
237 jint start = _index;
238 jint next_index = start + n;
239 if (next_index > _capacity) {
240 _overflow = true;
241 return;
242 }
243 // Otherwise.
244 _index = next_index;
245 for (int i = 0; i < n; i++) {
246 int ind = start + i;
247 assert(ind < _capacity, "By overflow test above.");
248 _base[ind] = ptr_arr[i];
249 }
250 }
251
252
253 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
254 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
255 jint index = _index;
256 if (index == 0) {
257 *n = 0;
258 return false;
259 } else {
260 int k = MIN2(max, index);
261 jint new_ind = index - k;
262 for (int j = 0; j < k; j++) {
263 ptr_arr[j] = _base[new_ind + j];
264 }
265 _index = new_ind;
266 *n = k;
267 return true;
268 }
269 }
270
271 template<class OopClosureClass>
272 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
273 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
274 || SafepointSynchronize::is_at_safepoint(),
275 "Drain recursion must be yield-safe.");
276 bool res = true;
277 debug_only(_drain_in_progress = true);
278 debug_only(_drain_in_progress_yields = yield_after);
279 while (!isEmpty()) {
280 oop newOop = pop();
281 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
282 assert(newOop->is_oop(), "Expected an oop");
283 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
284 "only grey objects on this stack");
285 newOop->oop_iterate(cl);
286 if (yield_after && _cm->do_yield_check()) {
287 res = false;
288 break;
289 }
290 }
291 debug_only(_drain_in_progress = false);
292 return res;
293 }
294
295 void CMMarkStack::note_start_of_gc() {
296 assert(_saved_index == -1,
297 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
298 _saved_index = _index;
299 }
300
301 void CMMarkStack::note_end_of_gc() {
302 // This is intentionally a guarantee, instead of an assert. If we
303 // accidentally add something to the mark stack during GC, it
304 // will be a correctness issue so it's better if we crash. we'll
305 // only check this once per GC anyway, so it won't be a performance
306 // issue in any way.
307 guarantee(_saved_index == _index,
308 err_msg("saved index: %d index: %d", _saved_index, _index));
309 _saved_index = -1;
310 }
311
312 void CMMarkStack::oops_do(OopClosure* f) {
313 assert(_saved_index == _index,
314 err_msg("saved index: %d index: %d", _saved_index, _index));
315 for (int i = 0; i < _index; i += 1) {
316 f->do_oop(&_base[i]);
317 }
318 }
319
320 bool ConcurrentMark::not_yet_marked(oop obj) const {
321 return (_g1h->is_obj_ill(obj)
322 || (_g1h->is_in_permanent(obj)
323 && !nextMarkBitMap()->isMarked((HeapWord*)obj)));
324 }
325
326 CMRootRegions::CMRootRegions() :
327 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
328 _should_abort(false), _next_survivor(NULL) { }
329
330 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
331 _young_list = g1h->young_list();
332 _cm = cm;
333 }
334
335 void CMRootRegions::prepare_for_scan() {
336 assert(!scan_in_progress(), "pre-condition");
337
338 // Currently, only survivors can be root regions.
339 assert(_next_survivor == NULL, "pre-condition");
340 _next_survivor = _young_list->first_survivor_region();
341 _scan_in_progress = (_next_survivor != NULL);
342 _should_abort = false;
343 }
344
345 HeapRegion* CMRootRegions::claim_next() {
346 if (_should_abort) {
347 // If someone has set the should_abort flag, we return NULL to
348 // force the caller to bail out of their loop.
349 return NULL;
350 }
351
352 // Currently, only survivors can be root regions.
353 HeapRegion* res = _next_survivor;
354 if (res != NULL) {
355 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
356 // Read it again in case it changed while we were waiting for the lock.
357 res = _next_survivor;
358 if (res != NULL) {
359 if (res == _young_list->last_survivor_region()) {
360 // We just claimed the last survivor so store NULL to indicate
361 // that we're done.
362 _next_survivor = NULL;
363 } else {
364 _next_survivor = res->get_next_young_region();
365 }
366 } else {
367 // Someone else claimed the last survivor while we were trying
368 // to take the lock so nothing else to do.
369 }
370 }
371 assert(res == NULL || res->is_survivor(), "post-condition");
372
373 return res;
374 }
375
376 void CMRootRegions::scan_finished() {
377 assert(scan_in_progress(), "pre-condition");
378
379 // Currently, only survivors can be root regions.
380 if (!_should_abort) {
381 assert(_next_survivor == NULL, "we should have claimed all survivors");
382 }
383 _next_survivor = NULL;
384
385 {
386 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
387 _scan_in_progress = false;
388 RootRegionScan_lock->notify_all();
389 }
390 }
391
392 bool CMRootRegions::wait_until_scan_finished() {
393 if (!scan_in_progress()) return false;
394
395 {
396 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
397 while (scan_in_progress()) {
398 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
399 }
400 }
401 return true;
402 }
403
404 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
405 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
406 #endif // _MSC_VER
407
408 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
409 return MAX2((n_par_threads + 2) / 4, 1U);
410 }
411
412 ConcurrentMark::ConcurrentMark(ReservedSpace rs, uint max_regions) :
413 _markBitMap1(rs, MinObjAlignment - 1),
414 _markBitMap2(rs, MinObjAlignment - 1),
415
416 _parallel_marking_threads(0),
417 _max_parallel_marking_threads(0),
418 _sleep_factor(0.0),
419 _marking_task_overhead(1.0),
420 _cleanup_sleep_factor(0.0),
421 _cleanup_task_overhead(1.0),
422 _cleanup_list("Cleanup List"),
423 _region_bm((BitMap::idx_t) max_regions, false /* in_resource_area*/),
424 _card_bm((rs.size() + CardTableModRefBS::card_size - 1) >>
425 CardTableModRefBS::card_shift,
426 false /* in_resource_area*/),
427
428 _prevMarkBitMap(&_markBitMap1),
429 _nextMarkBitMap(&_markBitMap2),
430
431 _markStack(this),
432 // _finger set in set_non_marking_state
433
434 _max_task_num(MAX2((uint)ParallelGCThreads, 1U)),
435 // _active_tasks set in set_non_marking_state
436 // _tasks set inside the constructor
437 _task_queues(new CMTaskQueueSet((int) _max_task_num)),
438 _terminator(ParallelTaskTerminator((int) _max_task_num, _task_queues)),
439
440 _has_overflown(false),
441 _concurrent(false),
442 _has_aborted(false),
443 _restart_for_overflow(false),
444 _concurrent_marking_in_progress(false),
445
446 // _verbose_level set below
447
448 _init_times(),
449 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
450 _cleanup_times(),
451 _total_counting_time(0.0),
452 _total_rs_scrub_time(0.0),
453
454 _parallel_workers(NULL),
455
456 _count_card_bitmaps(NULL),
457 _count_marked_bytes(NULL) {
458 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
459 if (verbose_level < no_verbose) {
460 verbose_level = no_verbose;
461 }
462 if (verbose_level > high_verbose) {
463 verbose_level = high_verbose;
464 }
465 _verbose_level = verbose_level;
466
467 if (verbose_low()) {
468 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
469 "heap end = "PTR_FORMAT, _heap_start, _heap_end);
470 }
471
472 _markStack.allocate(MarkStackSize);
473
474 // Create & start a ConcurrentMark thread.
475 _cmThread = new ConcurrentMarkThread(this);
476 assert(cmThread() != NULL, "CM Thread should have been created");
477 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
478
479 _g1h = G1CollectedHeap::heap();
480 assert(CGC_lock != NULL, "Where's the CGC_lock?");
481 assert(_markBitMap1.covers(rs), "_markBitMap1 inconsistency");
482 assert(_markBitMap2.covers(rs), "_markBitMap2 inconsistency");
483
484 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
485 satb_qs.set_buffer_size(G1SATBBufferSize);
486
487 _root_regions.init(_g1h, this);
488
489 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_task_num, mtGC);
490 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_task_num, mtGC);
491
492 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_task_num, mtGC);
493 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_task_num, mtGC);
494
495 BitMap::idx_t card_bm_size = _card_bm.size();
496
497 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
498 _active_tasks = _max_task_num;
499 for (int i = 0; i < (int) _max_task_num; ++i) {
500 CMTaskQueue* task_queue = new CMTaskQueue();
501 task_queue->initialize();
502 _task_queues->register_queue(i, task_queue);
503
504 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
505 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, (size_t) max_regions, mtGC);
506
507 _tasks[i] = new CMTask(i, this,
508 _count_marked_bytes[i],
509 &_count_card_bitmaps[i],
510 task_queue, _task_queues);
511
512 _accum_task_vtime[i] = 0.0;
513 }
514
515 // Calculate the card number for the bottom of the heap. Used
516 // in biasing indexes into the accounting card bitmaps.
517 _heap_bottom_card_num =
518 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
519 CardTableModRefBS::card_shift);
520
521 // Clear all the liveness counting data
522 clear_all_count_data();
523
524 if (ConcGCThreads > ParallelGCThreads) {
525 vm_exit_during_initialization("Can't have more ConcGCThreads "
526 "than ParallelGCThreads.");
527 }
528 if (ParallelGCThreads == 0) {
529 // if we are not running with any parallel GC threads we will not
530 // spawn any marking threads either
531 _parallel_marking_threads = 0;
532 _max_parallel_marking_threads = 0;
533 _sleep_factor = 0.0;
534 _marking_task_overhead = 1.0;
535 } else {
536 if (ConcGCThreads > 0) {
537 // notice that ConcGCThreads overwrites G1MarkingOverheadPercent
538 // if both are set
539
540 _parallel_marking_threads = (uint) ConcGCThreads;
541 _max_parallel_marking_threads = _parallel_marking_threads;
542 _sleep_factor = 0.0;
543 _marking_task_overhead = 1.0;
544 } else if (G1MarkingOverheadPercent > 0) {
545 // we will calculate the number of parallel marking threads
546 // based on a target overhead with respect to the soft real-time
547 // goal
548
549 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
550 double overall_cm_overhead =
551 (double) MaxGCPauseMillis * marking_overhead /
552 (double) GCPauseIntervalMillis;
553 double cpu_ratio = 1.0 / (double) os::processor_count();
554 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
555 double marking_task_overhead =
556 overall_cm_overhead / marking_thread_num *
557 (double) os::processor_count();
558 double sleep_factor =
559 (1.0 - marking_task_overhead) / marking_task_overhead;
560
561 _parallel_marking_threads = (uint) marking_thread_num;
562 _max_parallel_marking_threads = _parallel_marking_threads;
563 _sleep_factor = sleep_factor;
564 _marking_task_overhead = marking_task_overhead;
565 } else {
566 _parallel_marking_threads = scale_parallel_threads((uint)ParallelGCThreads);
567 _max_parallel_marking_threads = _parallel_marking_threads;
568 _sleep_factor = 0.0;
569 _marking_task_overhead = 1.0;
570 }
571
572 if (parallel_marking_threads() > 1) {
573 _cleanup_task_overhead = 1.0;
574 } else {
575 _cleanup_task_overhead = marking_task_overhead();
576 }
577 _cleanup_sleep_factor =
578 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
579
580 #if 0
581 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
582 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
583 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
584 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
585 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
586 #endif
587
588 guarantee(parallel_marking_threads() > 0, "peace of mind");
589 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
590 _max_parallel_marking_threads, false, true);
591 if (_parallel_workers == NULL) {
592 vm_exit_during_initialization("Failed necessary allocation.");
593 } else {
594 _parallel_workers->initialize_workers();
595 }
596 }
597
598 // so that the call below can read a sensible value
599 _heap_start = (HeapWord*) rs.base();
600 set_non_marking_state();
601 }
602
603 void ConcurrentMark::update_g1_committed(bool force) {
604 // If concurrent marking is not in progress, then we do not need to
605 // update _heap_end.
606 if (!concurrent_marking_in_progress() && !force) return;
607
608 MemRegion committed = _g1h->g1_committed();
609 assert(committed.start() == _heap_start, "start shouldn't change");
610 HeapWord* new_end = committed.end();
611 if (new_end > _heap_end) {
612 // The heap has been expanded.
613
614 _heap_end = new_end;
615 }
616 // Notice that the heap can also shrink. However, this only happens
617 // during a Full GC (at least currently) and the entire marking
618 // phase will bail out and the task will not be restarted. So, let's
619 // do nothing.
620 }
621
622 void ConcurrentMark::reset() {
623 // Starting values for these two. This should be called in a STW
624 // phase. CM will be notified of any future g1_committed expansions
625 // will be at the end of evacuation pauses, when tasks are
626 // inactive.
627 MemRegion committed = _g1h->g1_committed();
628 _heap_start = committed.start();
629 _heap_end = committed.end();
630
631 // Separated the asserts so that we know which one fires.
632 assert(_heap_start != NULL, "heap bounds should look ok");
633 assert(_heap_end != NULL, "heap bounds should look ok");
634 assert(_heap_start < _heap_end, "heap bounds should look ok");
635
636 // Reset all the marking data structures and any necessary flags
637 reset_marking_state();
638
639 if (verbose_low()) {
640 gclog_or_tty->print_cr("[global] resetting");
641 }
642
643 // We do reset all of them, since different phases will use
644 // different number of active threads. So, it's easiest to have all
645 // of them ready.
646 for (int i = 0; i < (int) _max_task_num; ++i) {
647 _tasks[i]->reset(_nextMarkBitMap);
648 }
649
650 // we need this to make sure that the flag is on during the evac
651 // pause with initial mark piggy-backed
652 set_concurrent_marking_in_progress();
653 }
654
655
656 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
657 _markStack.setEmpty();
658 _markStack.clear_overflow();
659 if (clear_overflow) {
660 clear_has_overflown();
661 } else {
662 assert(has_overflown(), "pre-condition");
663 }
664 _finger = _heap_start;
665
666 for (uint i = 0; i < _max_task_num; ++i) {
667 CMTaskQueue* queue = _task_queues->queue(i);
668 queue->set_empty();
669 }
670 }
671
672 void ConcurrentMark::set_phase(uint active_tasks, bool concurrent) {
673 assert(active_tasks <= _max_task_num, "we should not have more");
674
675 _active_tasks = active_tasks;
676 // Need to update the three data structures below according to the
677 // number of active threads for this phase.
678 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
679 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
680 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
681
682 _concurrent = concurrent;
683 // We propagate this to all tasks, not just the active ones.
684 for (int i = 0; i < (int) _max_task_num; ++i)
685 _tasks[i]->set_concurrent(concurrent);
686
687 if (concurrent) {
688 set_concurrent_marking_in_progress();
689 } else {
690 // We currently assume that the concurrent flag has been set to
691 // false before we start remark. At this point we should also be
692 // in a STW phase.
693 assert(!concurrent_marking_in_progress(), "invariant");
694 assert(_finger == _heap_end, "only way to get here");
695 update_g1_committed(true);
696 }
697 }
698
699 void ConcurrentMark::set_non_marking_state() {
700 // We set the global marking state to some default values when we're
701 // not doing marking.
702 reset_marking_state();
703 _active_tasks = 0;
704 clear_concurrent_marking_in_progress();
705 }
706
707 ConcurrentMark::~ConcurrentMark() {
708 // The ConcurrentMark instance is never freed.
709 ShouldNotReachHere();
710 }
711
712 void ConcurrentMark::clearNextBitmap() {
713 G1CollectedHeap* g1h = G1CollectedHeap::heap();
714 G1CollectorPolicy* g1p = g1h->g1_policy();
715
716 // Make sure that the concurrent mark thread looks to still be in
717 // the current cycle.
718 guarantee(cmThread()->during_cycle(), "invariant");
719
720 // We are finishing up the current cycle by clearing the next
721 // marking bitmap and getting it ready for the next cycle. During
722 // this time no other cycle can start. So, let's make sure that this
723 // is the case.
724 guarantee(!g1h->mark_in_progress(), "invariant");
725
726 // clear the mark bitmap (no grey objects to start with).
727 // We need to do this in chunks and offer to yield in between
728 // each chunk.
729 HeapWord* start = _nextMarkBitMap->startWord();
730 HeapWord* end = _nextMarkBitMap->endWord();
731 HeapWord* cur = start;
732 size_t chunkSize = M;
733 while (cur < end) {
734 HeapWord* next = cur + chunkSize;
735 if (next > end) {
736 next = end;
737 }
738 MemRegion mr(cur,next);
739 _nextMarkBitMap->clearRange(mr);
740 cur = next;
741 do_yield_check();
742
743 // Repeat the asserts from above. We'll do them as asserts here to
744 // minimize their overhead on the product. However, we'll have
745 // them as guarantees at the beginning / end of the bitmap
746 // clearing to get some checking in the product.
747 assert(cmThread()->during_cycle(), "invariant");
748 assert(!g1h->mark_in_progress(), "invariant");
749 }
750
751 // Clear the liveness counting data
752 clear_all_count_data();
753
754 // Repeat the asserts from above.
755 guarantee(cmThread()->during_cycle(), "invariant");
756 guarantee(!g1h->mark_in_progress(), "invariant");
757 }
758
759 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
760 public:
761 bool doHeapRegion(HeapRegion* r) {
762 if (!r->continuesHumongous()) {
763 r->note_start_of_marking();
764 }
765 return false;
766 }
767 };
768
769 void ConcurrentMark::checkpointRootsInitialPre() {
770 G1CollectedHeap* g1h = G1CollectedHeap::heap();
771 G1CollectorPolicy* g1p = g1h->g1_policy();
772
773 _has_aborted = false;
774
775 #ifndef PRODUCT
776 if (G1PrintReachableAtInitialMark) {
777 print_reachable("at-cycle-start",
778 VerifyOption_G1UsePrevMarking, true /* all */);
779 }
780 #endif
781
782 // Initialise marking structures. This has to be done in a STW phase.
783 reset();
784
785 // For each region note start of marking.
786 NoteStartOfMarkHRClosure startcl;
787 g1h->heap_region_iterate(&startcl);
788 }
789
790
791 void ConcurrentMark::checkpointRootsInitialPost() {
792 G1CollectedHeap* g1h = G1CollectedHeap::heap();
793
794 // If we force an overflow during remark, the remark operation will
795 // actually abort and we'll restart concurrent marking. If we always
796 // force an oveflow during remark we'll never actually complete the
797 // marking phase. So, we initilize this here, at the start of the
798 // cycle, so that at the remaining overflow number will decrease at
799 // every remark and we'll eventually not need to cause one.
800 force_overflow_stw()->init();
801
802 // Start Concurrent Marking weak-reference discovery.
803 ReferenceProcessor* rp = g1h->ref_processor_cm();
804 // enable ("weak") refs discovery
805 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
806 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
807
808 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
809 // This is the start of the marking cycle, we're expected all
810 // threads to have SATB queues with active set to false.
811 satb_mq_set.set_active_all_threads(true, /* new active value */
812 false /* expected_active */);
813
814 _root_regions.prepare_for_scan();
815
816 // update_g1_committed() will be called at the end of an evac pause
817 // when marking is on. So, it's also called at the end of the
818 // initial-mark pause to update the heap end, if the heap expands
819 // during it. No need to call it here.
820 }
821
822 /*
823 * Notice that in the next two methods, we actually leave the STS
824 * during the barrier sync and join it immediately afterwards. If we
825 * do not do this, the following deadlock can occur: one thread could
826 * be in the barrier sync code, waiting for the other thread to also
827 * sync up, whereas another one could be trying to yield, while also
828 * waiting for the other threads to sync up too.
829 *
830 * Note, however, that this code is also used during remark and in
831 * this case we should not attempt to leave / enter the STS, otherwise
832 * we'll either hit an asseert (debug / fastdebug) or deadlock
833 * (product). So we should only leave / enter the STS if we are
834 * operating concurrently.
835 *
836 * Because the thread that does the sync barrier has left the STS, it
837 * is possible to be suspended for a Full GC or an evacuation pause
838 * could occur. This is actually safe, since the entering the sync
839 * barrier is one of the last things do_marking_step() does, and it
840 * doesn't manipulate any data structures afterwards.
841 */
842
843 void ConcurrentMark::enter_first_sync_barrier(int task_num) {
844 if (verbose_low()) {
845 gclog_or_tty->print_cr("[%d] entering first barrier", task_num);
846 }
847
848 if (concurrent()) {
849 ConcurrentGCThread::stsLeave();
850 }
851 _first_overflow_barrier_sync.enter();
852 if (concurrent()) {
853 ConcurrentGCThread::stsJoin();
854 }
855 // at this point everyone should have synced up and not be doing any
856 // more work
857
858 if (verbose_low()) {
859 gclog_or_tty->print_cr("[%d] leaving first barrier", task_num);
860 }
861
862 // let task 0 do this
863 if (task_num == 0) {
864 // task 0 is responsible for clearing the global data structures
865 // We should be here because of an overflow. During STW we should
866 // not clear the overflow flag since we rely on it being true when
867 // we exit this method to abort the pause and restart concurent
868 // marking.
869 reset_marking_state(concurrent() /* clear_overflow */);
870 force_overflow()->update();
871
872 if (G1Log::fine()) {
873 gclog_or_tty->date_stamp(PrintGCDateStamps);
874 gclog_or_tty->stamp(PrintGCTimeStamps);
875 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
876 }
877 }
878
879 // after this, each task should reset its own data structures then
880 // then go into the second barrier
881 }
882
883 void ConcurrentMark::enter_second_sync_barrier(int task_num) {
884 if (verbose_low()) {
885 gclog_or_tty->print_cr("[%d] entering second barrier", task_num);
886 }
887
888 if (concurrent()) {
889 ConcurrentGCThread::stsLeave();
890 }
891 _second_overflow_barrier_sync.enter();
892 if (concurrent()) {
893 ConcurrentGCThread::stsJoin();
894 }
895 // at this point everything should be re-initialised and ready to go
896
897 if (verbose_low()) {
898 gclog_or_tty->print_cr("[%d] leaving second barrier", task_num);
899 }
900 }
901
902 #ifndef PRODUCT
903 void ForceOverflowSettings::init() {
904 _num_remaining = G1ConcMarkForceOverflow;
905 _force = false;
906 update();
907 }
908
909 void ForceOverflowSettings::update() {
910 if (_num_remaining > 0) {
911 _num_remaining -= 1;
912 _force = true;
913 } else {
914 _force = false;
915 }
916 }
917
918 bool ForceOverflowSettings::should_force() {
919 if (_force) {
920 _force = false;
921 return true;
922 } else {
923 return false;
924 }
925 }
926 #endif // !PRODUCT
927
928 class CMConcurrentMarkingTask: public AbstractGangTask {
929 private:
930 ConcurrentMark* _cm;
931 ConcurrentMarkThread* _cmt;
932
933 public:
934 void work(uint worker_id) {
935 assert(Thread::current()->is_ConcurrentGC_thread(),
936 "this should only be done by a conc GC thread");
937 ResourceMark rm;
938
939 double start_vtime = os::elapsedVTime();
940
941 ConcurrentGCThread::stsJoin();
942
943 assert(worker_id < _cm->active_tasks(), "invariant");
944 CMTask* the_task = _cm->task(worker_id);
945 the_task->record_start_time();
946 if (!_cm->has_aborted()) {
947 do {
948 double start_vtime_sec = os::elapsedVTime();
949 double start_time_sec = os::elapsedTime();
950 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
951
952 the_task->do_marking_step(mark_step_duration_ms,
953 true /* do_stealing */,
954 true /* do_termination */);
955
956 double end_time_sec = os::elapsedTime();
957 double end_vtime_sec = os::elapsedVTime();
958 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
959 double elapsed_time_sec = end_time_sec - start_time_sec;
960 _cm->clear_has_overflown();
961
962 bool ret = _cm->do_yield_check(worker_id);
963
964 jlong sleep_time_ms;
965 if (!_cm->has_aborted() && the_task->has_aborted()) {
966 sleep_time_ms =
967 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
968 ConcurrentGCThread::stsLeave();
969 os::sleep(Thread::current(), sleep_time_ms, false);
970 ConcurrentGCThread::stsJoin();
971 }
972 double end_time2_sec = os::elapsedTime();
973 double elapsed_time2_sec = end_time2_sec - start_time_sec;
974
975 #if 0
976 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, "
977 "overhead %1.4lf",
978 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms,
979 the_task->conc_overhead(os::elapsedTime()) * 8.0);
980 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms",
981 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0);
982 #endif
983 } while (!_cm->has_aborted() && the_task->has_aborted());
984 }
985 the_task->record_end_time();
986 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
987
988 ConcurrentGCThread::stsLeave();
989
990 double end_vtime = os::elapsedVTime();
991 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
992 }
993
994 CMConcurrentMarkingTask(ConcurrentMark* cm,
995 ConcurrentMarkThread* cmt) :
996 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
997
998 ~CMConcurrentMarkingTask() { }
999 };
1000
1001 // Calculates the number of active workers for a concurrent
1002 // phase.
1003 uint ConcurrentMark::calc_parallel_marking_threads() {
1004 if (G1CollectedHeap::use_parallel_gc_threads()) {
1005 uint n_conc_workers = 0;
1006 if (!UseDynamicNumberOfGCThreads ||
1007 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1008 !ForceDynamicNumberOfGCThreads)) {
1009 n_conc_workers = max_parallel_marking_threads();
1010 } else {
1011 n_conc_workers =
1012 AdaptiveSizePolicy::calc_default_active_workers(
1013 max_parallel_marking_threads(),
1014 1, /* Minimum workers */
1015 parallel_marking_threads(),
1016 Threads::number_of_non_daemon_threads());
1017 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1018 // that scaling has already gone into "_max_parallel_marking_threads".
1019 }
1020 assert(n_conc_workers > 0, "Always need at least 1");
1021 return n_conc_workers;
1022 }
1023 // If we are not running with any parallel GC threads we will not
1024 // have spawned any marking threads either. Hence the number of
1025 // concurrent workers should be 0.
1026 return 0;
1027 }
1028
1029 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1030 // Currently, only survivors can be root regions.
1031 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1032 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1033
1034 const uintx interval = PrefetchScanIntervalInBytes;
1035 HeapWord* curr = hr->bottom();
1036 const HeapWord* end = hr->top();
1037 while (curr < end) {
1038 Prefetch::read(curr, interval);
1039 oop obj = oop(curr);
1040 int size = obj->oop_iterate(&cl);
1041 assert(size == obj->size(), "sanity");
1042 curr += size;
1043 }
1044 }
1045
1046 class CMRootRegionScanTask : public AbstractGangTask {
1047 private:
1048 ConcurrentMark* _cm;
1049
1050 public:
1051 CMRootRegionScanTask(ConcurrentMark* cm) :
1052 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1053
1054 void work(uint worker_id) {
1055 assert(Thread::current()->is_ConcurrentGC_thread(),
1056 "this should only be done by a conc GC thread");
1057
1058 CMRootRegions* root_regions = _cm->root_regions();
1059 HeapRegion* hr = root_regions->claim_next();
1060 while (hr != NULL) {
1061 _cm->scanRootRegion(hr, worker_id);
1062 hr = root_regions->claim_next();
1063 }
1064 }
1065 };
1066
1067 void ConcurrentMark::scanRootRegions() {
1068 // scan_in_progress() will have been set to true only if there was
1069 // at least one root region to scan. So, if it's false, we
1070 // should not attempt to do any further work.
1071 if (root_regions()->scan_in_progress()) {
1072 _parallel_marking_threads = calc_parallel_marking_threads();
1073 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1074 "Maximum number of marking threads exceeded");
1075 uint active_workers = MAX2(1U, parallel_marking_threads());
1076
1077 CMRootRegionScanTask task(this);
1078 if (use_parallel_marking_threads()) {
1079 _parallel_workers->set_active_workers((int) active_workers);
1080 _parallel_workers->run_task(&task);
1081 } else {
1082 task.work(0);
1083 }
1084
1085 // It's possible that has_aborted() is true here without actually
1086 // aborting the survivor scan earlier. This is OK as it's
1087 // mainly used for sanity checking.
1088 root_regions()->scan_finished();
1089 }
1090 }
1091
1092 void ConcurrentMark::markFromRoots() {
1093 // we might be tempted to assert that:
1094 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1095 // "inconsistent argument?");
1096 // However that wouldn't be right, because it's possible that
1097 // a safepoint is indeed in progress as a younger generation
1098 // stop-the-world GC happens even as we mark in this generation.
1099
1100 _restart_for_overflow = false;
1101 force_overflow_conc()->init();
1102
1103 // _g1h has _n_par_threads
1104 _parallel_marking_threads = calc_parallel_marking_threads();
1105 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1106 "Maximum number of marking threads exceeded");
1107
1108 uint active_workers = MAX2(1U, parallel_marking_threads());
1109
1110 // Parallel task terminator is set in "set_phase()"
1111 set_phase(active_workers, true /* concurrent */);
1112
1113 CMConcurrentMarkingTask markingTask(this, cmThread());
1114 if (use_parallel_marking_threads()) {
1115 _parallel_workers->set_active_workers((int)active_workers);
1116 // Don't set _n_par_threads because it affects MT in proceess_strong_roots()
1117 // and the decisions on that MT processing is made elsewhere.
1118 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1119 _parallel_workers->run_task(&markingTask);
1120 } else {
1121 markingTask.work(0);
1122 }
1123 print_stats();
1124 }
1125
1126 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1127 // world is stopped at this checkpoint
1128 assert(SafepointSynchronize::is_at_safepoint(),
1129 "world should be stopped");
1130
1131 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1132
1133 // If a full collection has happened, we shouldn't do this.
1134 if (has_aborted()) {
1135 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1136 return;
1137 }
1138
1139 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1140
1141 if (VerifyDuringGC) {
1142 HandleMark hm; // handle scope
1143 gclog_or_tty->print(" VerifyDuringGC:(before)");
1144 Universe::heap()->prepare_for_verify();
1145 Universe::verify(/* silent */ false,
1146 /* option */ VerifyOption_G1UsePrevMarking);
1147 }
1148
1149 G1CollectorPolicy* g1p = g1h->g1_policy();
1150 g1p->record_concurrent_mark_remark_start();
1151
1152 double start = os::elapsedTime();
1153
1154 checkpointRootsFinalWork();
1155
1156 double mark_work_end = os::elapsedTime();
1157
1158 weakRefsWork(clear_all_soft_refs);
1159
1160 if (has_overflown()) {
1161 // Oops. We overflowed. Restart concurrent marking.
1162 _restart_for_overflow = true;
1163 // Clear the marking state because we will be restarting
1164 // marking due to overflowing the global mark stack.
1165 reset_marking_state();
1166 if (G1TraceMarkStackOverflow) {
1167 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1168 }
1169 } else {
1170 // Aggregate the per-task counting data that we have accumulated
1171 // while marking.
1172 aggregate_count_data();
1173
1174 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1175 // We're done with marking.
1176 // This is the end of the marking cycle, we're expected all
1177 // threads to have SATB queues with active set to true.
1178 satb_mq_set.set_active_all_threads(false, /* new active value */
1179 true /* expected_active */);
1180
1181 if (VerifyDuringGC) {
1182 HandleMark hm; // handle scope
1183 gclog_or_tty->print(" VerifyDuringGC:(after)");
1184 Universe::heap()->prepare_for_verify();
1185 Universe::verify(/* silent */ false,
1186 /* option */ VerifyOption_G1UseNextMarking);
1187 }
1188 assert(!restart_for_overflow(), "sanity");
1189 // Completely reset the marking state since marking completed
1190 set_non_marking_state();
1191 }
1192
1193 #if VERIFY_OBJS_PROCESSED
1194 _scan_obj_cl.objs_processed = 0;
1195 ThreadLocalObjQueue::objs_enqueued = 0;
1196 #endif
1197
1198 // Statistics
1199 double now = os::elapsedTime();
1200 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1201 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1202 _remark_times.add((now - start) * 1000.0);
1203
1204 g1p->record_concurrent_mark_remark_end();
1205
1206 G1CMIsAliveClosure is_alive(g1h);
1207 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1208 }
1209
1210 // Base class of the closures that finalize and verify the
1211 // liveness counting data.
1212 class CMCountDataClosureBase: public HeapRegionClosure {
1213 protected:
1214 G1CollectedHeap* _g1h;
1215 ConcurrentMark* _cm;
1216 CardTableModRefBS* _ct_bs;
1217
1218 BitMap* _region_bm;
1219 BitMap* _card_bm;
1220
1221 // Takes a region that's not empty (i.e., it has at least one
1222 // live object in it and sets its corresponding bit on the region
1223 // bitmap to 1. If the region is "starts humongous" it will also set
1224 // to 1 the bits on the region bitmap that correspond to its
1225 // associated "continues humongous" regions.
1226 void set_bit_for_region(HeapRegion* hr) {
1227 assert(!hr->continuesHumongous(), "should have filtered those out");
1228
1229 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1230 if (!hr->startsHumongous()) {
1231 // Normal (non-humongous) case: just set the bit.
1232 _region_bm->par_at_put(index, true);
1233 } else {
1234 // Starts humongous case: calculate how many regions are part of
1235 // this humongous region and then set the bit range.
1236 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1237 _region_bm->par_at_put_range(index, end_index, true);
1238 }
1239 }
1240
1241 public:
1242 CMCountDataClosureBase(G1CollectedHeap* g1h,
1243 BitMap* region_bm, BitMap* card_bm):
1244 _g1h(g1h), _cm(g1h->concurrent_mark()),
1245 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1246 _region_bm(region_bm), _card_bm(card_bm) { }
1247 };
1248
1249 // Closure that calculates the # live objects per region. Used
1250 // for verification purposes during the cleanup pause.
1251 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1252 CMBitMapRO* _bm;
1253 size_t _region_marked_bytes;
1254
1255 public:
1256 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1257 BitMap* region_bm, BitMap* card_bm) :
1258 CMCountDataClosureBase(g1h, region_bm, card_bm),
1259 _bm(bm), _region_marked_bytes(0) { }
1260
1261 bool doHeapRegion(HeapRegion* hr) {
1262
1263 if (hr->continuesHumongous()) {
1264 // We will ignore these here and process them when their
1265 // associated "starts humongous" region is processed (see
1266 // set_bit_for_heap_region()). Note that we cannot rely on their
1267 // associated "starts humongous" region to have their bit set to
1268 // 1 since, due to the region chunking in the parallel region
1269 // iteration, a "continues humongous" region might be visited
1270 // before its associated "starts humongous".
1271 return false;
1272 }
1273
1274 HeapWord* ntams = hr->next_top_at_mark_start();
1275 HeapWord* start = hr->bottom();
1276
1277 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1278 err_msg("Preconditions not met - "
1279 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1280 start, ntams, hr->end()));
1281
1282 // Find the first marked object at or after "start".
1283 start = _bm->getNextMarkedWordAddress(start, ntams);
1284
1285 size_t marked_bytes = 0;
1286
1287 while (start < ntams) {
1288 oop obj = oop(start);
1289 int obj_sz = obj->size();
1290 HeapWord* obj_end = start + obj_sz;
1291
1292 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1293 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1294
1295 // Note: if we're looking at the last region in heap - obj_end
1296 // could be actually just beyond the end of the heap; end_idx
1297 // will then correspond to a (non-existent) card that is also
1298 // just beyond the heap.
1299 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1300 // end of object is not card aligned - increment to cover
1301 // all the cards spanned by the object
1302 end_idx += 1;
1303 }
1304
1305 // Set the bits in the card BM for the cards spanned by this object.
1306 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1307
1308 // Add the size of this object to the number of marked bytes.
1309 marked_bytes += (size_t)obj_sz * HeapWordSize;
1310
1311 // Find the next marked object after this one.
1312 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1313 }
1314
1315 // Mark the allocated-since-marking portion...
1316 HeapWord* top = hr->top();
1317 if (ntams < top) {
1318 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1319 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1320
1321 // Note: if we're looking at the last region in heap - top
1322 // could be actually just beyond the end of the heap; end_idx
1323 // will then correspond to a (non-existent) card that is also
1324 // just beyond the heap.
1325 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1326 // end of object is not card aligned - increment to cover
1327 // all the cards spanned by the object
1328 end_idx += 1;
1329 }
1330 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1331
1332 // This definitely means the region has live objects.
1333 set_bit_for_region(hr);
1334 }
1335
1336 // Update the live region bitmap.
1337 if (marked_bytes > 0) {
1338 set_bit_for_region(hr);
1339 }
1340
1341 // Set the marked bytes for the current region so that
1342 // it can be queried by a calling verificiation routine
1343 _region_marked_bytes = marked_bytes;
1344
1345 return false;
1346 }
1347
1348 size_t region_marked_bytes() const { return _region_marked_bytes; }
1349 };
1350
1351 // Heap region closure used for verifying the counting data
1352 // that was accumulated concurrently and aggregated during
1353 // the remark pause. This closure is applied to the heap
1354 // regions during the STW cleanup pause.
1355
1356 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1357 G1CollectedHeap* _g1h;
1358 ConcurrentMark* _cm;
1359 CalcLiveObjectsClosure _calc_cl;
1360 BitMap* _region_bm; // Region BM to be verified
1361 BitMap* _card_bm; // Card BM to be verified
1362 bool _verbose; // verbose output?
1363
1364 BitMap* _exp_region_bm; // Expected Region BM values
1365 BitMap* _exp_card_bm; // Expected card BM values
1366
1367 int _failures;
1368
1369 public:
1370 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1371 BitMap* region_bm,
1372 BitMap* card_bm,
1373 BitMap* exp_region_bm,
1374 BitMap* exp_card_bm,
1375 bool verbose) :
1376 _g1h(g1h), _cm(g1h->concurrent_mark()),
1377 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1378 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1379 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1380 _failures(0) { }
1381
1382 int failures() const { return _failures; }
1383
1384 bool doHeapRegion(HeapRegion* hr) {
1385 if (hr->continuesHumongous()) {
1386 // We will ignore these here and process them when their
1387 // associated "starts humongous" region is processed (see
1388 // set_bit_for_heap_region()). Note that we cannot rely on their
1389 // associated "starts humongous" region to have their bit set to
1390 // 1 since, due to the region chunking in the parallel region
1391 // iteration, a "continues humongous" region might be visited
1392 // before its associated "starts humongous".
1393 return false;
1394 }
1395
1396 int failures = 0;
1397
1398 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1399 // this region and set the corresponding bits in the expected region
1400 // and card bitmaps.
1401 bool res = _calc_cl.doHeapRegion(hr);
1402 assert(res == false, "should be continuing");
1403
1404 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1405 Mutex::_no_safepoint_check_flag);
1406
1407 // Verify the marked bytes for this region.
1408 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1409 size_t act_marked_bytes = hr->next_marked_bytes();
1410
1411 // We're not OK if expected marked bytes > actual marked bytes. It means
1412 // we have missed accounting some objects during the actual marking.
1413 if (exp_marked_bytes > act_marked_bytes) {
1414 if (_verbose) {
1415 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1416 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1417 hr->hrs_index(), exp_marked_bytes, act_marked_bytes);
1418 }
1419 failures += 1;
1420 }
1421
1422 // Verify the bit, for this region, in the actual and expected
1423 // (which was just calculated) region bit maps.
1424 // We're not OK if the bit in the calculated expected region
1425 // bitmap is set and the bit in the actual region bitmap is not.
1426 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1427
1428 bool expected = _exp_region_bm->at(index);
1429 bool actual = _region_bm->at(index);
1430 if (expected && !actual) {
1431 if (_verbose) {
1432 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1433 "expected: %s, actual: %s",
1434 hr->hrs_index(),
1435 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1436 }
1437 failures += 1;
1438 }
1439
1440 // Verify that the card bit maps for the cards spanned by the current
1441 // region match. We have an error if we have a set bit in the expected
1442 // bit map and the corresponding bit in the actual bitmap is not set.
1443
1444 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1445 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1446
1447 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1448 expected = _exp_card_bm->at(i);
1449 actual = _card_bm->at(i);
1450
1451 if (expected && !actual) {
1452 if (_verbose) {
1453 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1454 "expected: %s, actual: %s",
1455 hr->hrs_index(), i,
1456 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1457 }
1458 failures += 1;
1459 }
1460 }
1461
1462 if (failures > 0 && _verbose) {
1463 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1464 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1465 HR_FORMAT_PARAMS(hr), hr->next_top_at_mark_start(),
1466 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1467 }
1468
1469 _failures += failures;
1470
1471 // We could stop iteration over the heap when we
1472 // find the first violating region by returning true.
1473 return false;
1474 }
1475 };
1476
1477
1478 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1479 protected:
1480 G1CollectedHeap* _g1h;
1481 ConcurrentMark* _cm;
1482 BitMap* _actual_region_bm;
1483 BitMap* _actual_card_bm;
1484
1485 uint _n_workers;
1486
1487 BitMap* _expected_region_bm;
1488 BitMap* _expected_card_bm;
1489
1490 int _failures;
1491 bool _verbose;
1492
1493 public:
1494 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1495 BitMap* region_bm, BitMap* card_bm,
1496 BitMap* expected_region_bm, BitMap* expected_card_bm)
1497 : AbstractGangTask("G1 verify final counting"),
1498 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1499 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1500 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1501 _failures(0), _verbose(false),
1502 _n_workers(0) {
1503 assert(VerifyDuringGC, "don't call this otherwise");
1504
1505 // Use the value already set as the number of active threads
1506 // in the call to run_task().
1507 if (G1CollectedHeap::use_parallel_gc_threads()) {
1508 assert( _g1h->workers()->active_workers() > 0,
1509 "Should have been previously set");
1510 _n_workers = _g1h->workers()->active_workers();
1511 } else {
1512 _n_workers = 1;
1513 }
1514
1515 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1516 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1517
1518 _verbose = _cm->verbose_medium();
1519 }
1520
1521 void work(uint worker_id) {
1522 assert(worker_id < _n_workers, "invariant");
1523
1524 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1525 _actual_region_bm, _actual_card_bm,
1526 _expected_region_bm,
1527 _expected_card_bm,
1528 _verbose);
1529
1530 if (G1CollectedHeap::use_parallel_gc_threads()) {
1531 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1532 worker_id,
1533 _n_workers,
1534 HeapRegion::VerifyCountClaimValue);
1535 } else {
1536 _g1h->heap_region_iterate(&verify_cl);
1537 }
1538
1539 Atomic::add(verify_cl.failures(), &_failures);
1540 }
1541
1542 int failures() const { return _failures; }
1543 };
1544
1545 // Closure that finalizes the liveness counting data.
1546 // Used during the cleanup pause.
1547 // Sets the bits corresponding to the interval [NTAMS, top]
1548 // (which contains the implicitly live objects) in the
1549 // card liveness bitmap. Also sets the bit for each region,
1550 // containing live data, in the region liveness bitmap.
1551
1552 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1553 public:
1554 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1555 BitMap* region_bm,
1556 BitMap* card_bm) :
1557 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1558
1559 bool doHeapRegion(HeapRegion* hr) {
1560
1561 if (hr->continuesHumongous()) {
1562 // We will ignore these here and process them when their
1563 // associated "starts humongous" region is processed (see
1564 // set_bit_for_heap_region()). Note that we cannot rely on their
1565 // associated "starts humongous" region to have their bit set to
1566 // 1 since, due to the region chunking in the parallel region
1567 // iteration, a "continues humongous" region might be visited
1568 // before its associated "starts humongous".
1569 return false;
1570 }
1571
1572 HeapWord* ntams = hr->next_top_at_mark_start();
1573 HeapWord* top = hr->top();
1574
1575 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1576
1577 // Mark the allocated-since-marking portion...
1578 if (ntams < top) {
1579 // This definitely means the region has live objects.
1580 set_bit_for_region(hr);
1581
1582 // Now set the bits in the card bitmap for [ntams, top)
1583 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1584 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1585
1586 // Note: if we're looking at the last region in heap - top
1587 // could be actually just beyond the end of the heap; end_idx
1588 // will then correspond to a (non-existent) card that is also
1589 // just beyond the heap.
1590 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1591 // end of object is not card aligned - increment to cover
1592 // all the cards spanned by the object
1593 end_idx += 1;
1594 }
1595
1596 assert(end_idx <= _card_bm->size(),
1597 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1598 end_idx, _card_bm->size()));
1599 assert(start_idx < _card_bm->size(),
1600 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1601 start_idx, _card_bm->size()));
1602
1603 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1604 }
1605
1606 // Set the bit for the region if it contains live data
1607 if (hr->next_marked_bytes() > 0) {
1608 set_bit_for_region(hr);
1609 }
1610
1611 return false;
1612 }
1613 };
1614
1615 class G1ParFinalCountTask: public AbstractGangTask {
1616 protected:
1617 G1CollectedHeap* _g1h;
1618 ConcurrentMark* _cm;
1619 BitMap* _actual_region_bm;
1620 BitMap* _actual_card_bm;
1621
1622 uint _n_workers;
1623
1624 public:
1625 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1626 : AbstractGangTask("G1 final counting"),
1627 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1628 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1629 _n_workers(0) {
1630 // Use the value already set as the number of active threads
1631 // in the call to run_task().
1632 if (G1CollectedHeap::use_parallel_gc_threads()) {
1633 assert( _g1h->workers()->active_workers() > 0,
1634 "Should have been previously set");
1635 _n_workers = _g1h->workers()->active_workers();
1636 } else {
1637 _n_workers = 1;
1638 }
1639 }
1640
1641 void work(uint worker_id) {
1642 assert(worker_id < _n_workers, "invariant");
1643
1644 FinalCountDataUpdateClosure final_update_cl(_g1h,
1645 _actual_region_bm,
1646 _actual_card_bm);
1647
1648 if (G1CollectedHeap::use_parallel_gc_threads()) {
1649 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1650 worker_id,
1651 _n_workers,
1652 HeapRegion::FinalCountClaimValue);
1653 } else {
1654 _g1h->heap_region_iterate(&final_update_cl);
1655 }
1656 }
1657 };
1658
1659 class G1ParNoteEndTask;
1660
1661 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1662 G1CollectedHeap* _g1;
1663 int _worker_num;
1664 size_t _max_live_bytes;
1665 uint _regions_claimed;
1666 size_t _freed_bytes;
1667 FreeRegionList* _local_cleanup_list;
1668 OldRegionSet* _old_proxy_set;
1669 HumongousRegionSet* _humongous_proxy_set;
1670 HRRSCleanupTask* _hrrs_cleanup_task;
1671 double _claimed_region_time;
1672 double _max_region_time;
1673
1674 public:
1675 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1676 int worker_num,
1677 FreeRegionList* local_cleanup_list,
1678 OldRegionSet* old_proxy_set,
1679 HumongousRegionSet* humongous_proxy_set,
1680 HRRSCleanupTask* hrrs_cleanup_task) :
1681 _g1(g1), _worker_num(worker_num),
1682 _max_live_bytes(0), _regions_claimed(0),
1683 _freed_bytes(0),
1684 _claimed_region_time(0.0), _max_region_time(0.0),
1685 _local_cleanup_list(local_cleanup_list),
1686 _old_proxy_set(old_proxy_set),
1687 _humongous_proxy_set(humongous_proxy_set),
1688 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1689
1690 size_t freed_bytes() { return _freed_bytes; }
1691
1692 bool doHeapRegion(HeapRegion *hr) {
1693 if (hr->continuesHumongous()) {
1694 return false;
1695 }
1696 // We use a claim value of zero here because all regions
1697 // were claimed with value 1 in the FinalCount task.
1698 _g1->reset_gc_time_stamps(hr);
1699 double start = os::elapsedTime();
1700 _regions_claimed++;
1701 hr->note_end_of_marking();
1702 _max_live_bytes += hr->max_live_bytes();
1703 _g1->free_region_if_empty(hr,
1704 &_freed_bytes,
1705 _local_cleanup_list,
1706 _old_proxy_set,
1707 _humongous_proxy_set,
1708 _hrrs_cleanup_task,
1709 true /* par */);
1710 double region_time = (os::elapsedTime() - start);
1711 _claimed_region_time += region_time;
1712 if (region_time > _max_region_time) {
1713 _max_region_time = region_time;
1714 }
1715 return false;
1716 }
1717
1718 size_t max_live_bytes() { return _max_live_bytes; }
1719 uint regions_claimed() { return _regions_claimed; }
1720 double claimed_region_time_sec() { return _claimed_region_time; }
1721 double max_region_time_sec() { return _max_region_time; }
1722 };
1723
1724 class G1ParNoteEndTask: public AbstractGangTask {
1725 friend class G1NoteEndOfConcMarkClosure;
1726
1727 protected:
1728 G1CollectedHeap* _g1h;
1729 size_t _max_live_bytes;
1730 size_t _freed_bytes;
1731 FreeRegionList* _cleanup_list;
1732
1733 public:
1734 G1ParNoteEndTask(G1CollectedHeap* g1h,
1735 FreeRegionList* cleanup_list) :
1736 AbstractGangTask("G1 note end"), _g1h(g1h),
1737 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1738
1739 void work(uint worker_id) {
1740 double start = os::elapsedTime();
1741 FreeRegionList local_cleanup_list("Local Cleanup List");
1742 OldRegionSet old_proxy_set("Local Cleanup Old Proxy Set");
1743 HumongousRegionSet humongous_proxy_set("Local Cleanup Humongous Proxy Set");
1744 HRRSCleanupTask hrrs_cleanup_task;
1745 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, worker_id, &local_cleanup_list,
1746 &old_proxy_set,
1747 &humongous_proxy_set,
1748 &hrrs_cleanup_task);
1749 if (G1CollectedHeap::use_parallel_gc_threads()) {
1750 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1751 _g1h->workers()->active_workers(),
1752 HeapRegion::NoteEndClaimValue);
1753 } else {
1754 _g1h->heap_region_iterate(&g1_note_end);
1755 }
1756 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1757
1758 // Now update the lists
1759 _g1h->update_sets_after_freeing_regions(g1_note_end.freed_bytes(),
1760 NULL /* free_list */,
1761 &old_proxy_set,
1762 &humongous_proxy_set,
1763 true /* par */);
1764 {
1765 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1766 _max_live_bytes += g1_note_end.max_live_bytes();
1767 _freed_bytes += g1_note_end.freed_bytes();
1768
1769 // If we iterate over the global cleanup list at the end of
1770 // cleanup to do this printing we will not guarantee to only
1771 // generate output for the newly-reclaimed regions (the list
1772 // might not be empty at the beginning of cleanup; we might
1773 // still be working on its previous contents). So we do the
1774 // printing here, before we append the new regions to the global
1775 // cleanup list.
1776
1777 G1HRPrinter* hr_printer = _g1h->hr_printer();
1778 if (hr_printer->is_active()) {
1779 HeapRegionLinkedListIterator iter(&local_cleanup_list);
1780 while (iter.more_available()) {
1781 HeapRegion* hr = iter.get_next();
1782 hr_printer->cleanup(hr);
1783 }
1784 }
1785
1786 _cleanup_list->add_as_tail(&local_cleanup_list);
1787 assert(local_cleanup_list.is_empty(), "post-condition");
1788
1789 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1790 }
1791 }
1792 size_t max_live_bytes() { return _max_live_bytes; }
1793 size_t freed_bytes() { return _freed_bytes; }
1794 };
1795
1796 class G1ParScrubRemSetTask: public AbstractGangTask {
1797 protected:
1798 G1RemSet* _g1rs;
1799 BitMap* _region_bm;
1800 BitMap* _card_bm;
1801 public:
1802 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1803 BitMap* region_bm, BitMap* card_bm) :
1804 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1805 _region_bm(region_bm), _card_bm(card_bm) { }
1806
1807 void work(uint worker_id) {
1808 if (G1CollectedHeap::use_parallel_gc_threads()) {
1809 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1810 HeapRegion::ScrubRemSetClaimValue);
1811 } else {
1812 _g1rs->scrub(_region_bm, _card_bm);
1813 }
1814 }
1815
1816 };
1817
1818 void ConcurrentMark::cleanup() {
1819 // world is stopped at this checkpoint
1820 assert(SafepointSynchronize::is_at_safepoint(),
1821 "world should be stopped");
1822 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1823
1824 // If a full collection has happened, we shouldn't do this.
1825 if (has_aborted()) {
1826 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1827 return;
1828 }
1829
1830 HRSPhaseSetter x(HRSPhaseCleanup);
1831 g1h->verify_region_sets_optional();
1832
1833 if (VerifyDuringGC) {
1834 HandleMark hm; // handle scope
1835 gclog_or_tty->print(" VerifyDuringGC:(before)");
1836 Universe::heap()->prepare_for_verify();
1837 Universe::verify(/* silent */ false,
1838 /* option */ VerifyOption_G1UsePrevMarking);
1839 }
1840
1841 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
1842 g1p->record_concurrent_mark_cleanup_start();
1843
1844 double start = os::elapsedTime();
1845
1846 HeapRegionRemSet::reset_for_cleanup_tasks();
1847
1848 uint n_workers;
1849
1850 // Do counting once more with the world stopped for good measure.
1851 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1852
1853 if (G1CollectedHeap::use_parallel_gc_threads()) {
1854 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
1855 "sanity check");
1856
1857 g1h->set_par_threads();
1858 n_workers = g1h->n_par_threads();
1859 assert(g1h->n_par_threads() == n_workers,
1860 "Should not have been reset");
1861 g1h->workers()->run_task(&g1_par_count_task);
1862 // Done with the parallel phase so reset to 0.
1863 g1h->set_par_threads(0);
1864
1865 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
1866 "sanity check");
1867 } else {
1868 n_workers = 1;
1869 g1_par_count_task.work(0);
1870 }
1871
1872 if (VerifyDuringGC) {
1873 // Verify that the counting data accumulated during marking matches
1874 // that calculated by walking the marking bitmap.
1875
1876 // Bitmaps to hold expected values
1877 BitMap expected_region_bm(_region_bm.size(), false);
1878 BitMap expected_card_bm(_card_bm.size(), false);
1879
1880 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
1881 &_region_bm,
1882 &_card_bm,
1883 &expected_region_bm,
1884 &expected_card_bm);
1885
1886 if (G1CollectedHeap::use_parallel_gc_threads()) {
1887 g1h->set_par_threads((int)n_workers);
1888 g1h->workers()->run_task(&g1_par_verify_task);
1889 // Done with the parallel phase so reset to 0.
1890 g1h->set_par_threads(0);
1891
1892 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
1893 "sanity check");
1894 } else {
1895 g1_par_verify_task.work(0);
1896 }
1897
1898 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
1899 }
1900
1901 size_t start_used_bytes = g1h->used();
1902 g1h->set_marking_complete();
1903
1904 double count_end = os::elapsedTime();
1905 double this_final_counting_time = (count_end - start);
1906 _total_counting_time += this_final_counting_time;
1907
1908 if (G1PrintRegionLivenessInfo) {
1909 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
1910 _g1h->heap_region_iterate(&cl);
1911 }
1912
1913 // Install newly created mark bitMap as "prev".
1914 swapMarkBitMaps();
1915
1916 g1h->reset_gc_time_stamp();
1917
1918 // Note end of marking in all heap regions.
1919 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
1920 if (G1CollectedHeap::use_parallel_gc_threads()) {
1921 g1h->set_par_threads((int)n_workers);
1922 g1h->workers()->run_task(&g1_par_note_end_task);
1923 g1h->set_par_threads(0);
1924
1925 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
1926 "sanity check");
1927 } else {
1928 g1_par_note_end_task.work(0);
1929 }
1930 g1h->check_gc_time_stamps();
1931
1932 if (!cleanup_list_is_empty()) {
1933 // The cleanup list is not empty, so we'll have to process it
1934 // concurrently. Notify anyone else that might be wanting free
1935 // regions that there will be more free regions coming soon.
1936 g1h->set_free_regions_coming();
1937 }
1938
1939 // call below, since it affects the metric by which we sort the heap
1940 // regions.
1941 if (G1ScrubRemSets) {
1942 double rs_scrub_start = os::elapsedTime();
1943 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
1944 if (G1CollectedHeap::use_parallel_gc_threads()) {
1945 g1h->set_par_threads((int)n_workers);
1946 g1h->workers()->run_task(&g1_par_scrub_rs_task);
1947 g1h->set_par_threads(0);
1948
1949 assert(g1h->check_heap_region_claim_values(
1950 HeapRegion::ScrubRemSetClaimValue),
1951 "sanity check");
1952 } else {
1953 g1_par_scrub_rs_task.work(0);
1954 }
1955
1956 double rs_scrub_end = os::elapsedTime();
1957 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
1958 _total_rs_scrub_time += this_rs_scrub_time;
1959 }
1960
1961 // this will also free any regions totally full of garbage objects,
1962 // and sort the regions.
1963 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
1964
1965 // Statistics.
1966 double end = os::elapsedTime();
1967 _cleanup_times.add((end - start) * 1000.0);
1968
1969 if (G1Log::fine()) {
1970 g1h->print_size_transition(gclog_or_tty,
1971 start_used_bytes,
1972 g1h->used(),
1973 g1h->capacity());
1974 }
1975
1976 // Clean up will have freed any regions completely full of garbage.
1977 // Update the soft reference policy with the new heap occupancy.
1978 Universe::update_heap_info_at_gc();
1979
1980 // We need to make this be a "collection" so any collection pause that
1981 // races with it goes around and waits for completeCleanup to finish.
1982 g1h->increment_total_collections();
1983
1984 // We reclaimed old regions so we should calculate the sizes to make
1985 // sure we update the old gen/space data.
1986 g1h->g1mm()->update_sizes();
1987
1988 if (VerifyDuringGC) {
1989 HandleMark hm; // handle scope
1990 gclog_or_tty->print(" VerifyDuringGC:(after)");
1991 Universe::heap()->prepare_for_verify();
1992 Universe::verify(/* silent */ false,
1993 /* option */ VerifyOption_G1UsePrevMarking);
1994 }
1995
1996 g1h->verify_region_sets_optional();
1997 g1h->trace_heap_after_concurrent_cycle();
1998 }
1999
2000 void ConcurrentMark::completeCleanup() {
2001 if (has_aborted()) return;
2002
2003 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2004
2005 _cleanup_list.verify_optional();
2006 FreeRegionList tmp_free_list("Tmp Free List");
2007
2008 if (G1ConcRegionFreeingVerbose) {
2009 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2010 "cleanup list has %u entries",
2011 _cleanup_list.length());
2012 }
2013
2014 // Noone else should be accessing the _cleanup_list at this point,
2015 // so it's not necessary to take any locks
2016 while (!_cleanup_list.is_empty()) {
2017 HeapRegion* hr = _cleanup_list.remove_head();
2018 assert(hr != NULL, "the list was not empty");
2019 hr->par_clear();
2020 tmp_free_list.add_as_tail(hr);
2021
2022 // Instead of adding one region at a time to the secondary_free_list,
2023 // we accumulate them in the local list and move them a few at a
2024 // time. This also cuts down on the number of notify_all() calls
2025 // we do during this process. We'll also append the local list when
2026 // _cleanup_list is empty (which means we just removed the last
2027 // region from the _cleanup_list).
2028 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2029 _cleanup_list.is_empty()) {
2030 if (G1ConcRegionFreeingVerbose) {
2031 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2032 "appending %u entries to the secondary_free_list, "
2033 "cleanup list still has %u entries",
2034 tmp_free_list.length(),
2035 _cleanup_list.length());
2036 }
2037
2038 {
2039 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2040 g1h->secondary_free_list_add_as_tail(&tmp_free_list);
2041 SecondaryFreeList_lock->notify_all();
2042 }
2043
2044 if (G1StressConcRegionFreeing) {
2045 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2046 os::sleep(Thread::current(), (jlong) 1, false);
2047 }
2048 }
2049 }
2050 }
2051 assert(tmp_free_list.is_empty(), "post-condition");
2052 }
2053
2054 // Support closures for reference procssing in G1
2055
2056 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2057 HeapWord* addr = (HeapWord*)obj;
2058 return addr != NULL &&
2059 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2060 }
2061
2062 class G1CMKeepAliveClosure: public OopClosure {
2063 G1CollectedHeap* _g1;
2064 ConcurrentMark* _cm;
2065 public:
2066 G1CMKeepAliveClosure(G1CollectedHeap* g1, ConcurrentMark* cm) :
2067 _g1(g1), _cm(cm) {
2068 assert(Thread::current()->is_VM_thread(), "otherwise fix worker id");
2069 }
2070
2071 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2072 virtual void do_oop( oop* p) { do_oop_work(p); }
2073
2074 template <class T> void do_oop_work(T* p) {
2075 oop obj = oopDesc::load_decode_heap_oop(p);
2076 HeapWord* addr = (HeapWord*)obj;
2077
2078 if (_cm->verbose_high()) {
2079 gclog_or_tty->print_cr("\t[0] we're looking at location "
2080 "*"PTR_FORMAT" = "PTR_FORMAT,
2081 p, (void*) obj);
2082 }
2083
2084 if (_g1->is_in_g1_reserved(addr) && _g1->is_obj_ill(obj)) {
2085 _cm->mark_and_count(obj);
2086 _cm->mark_stack_push(obj);
2087 }
2088 }
2089 };
2090
2091 class G1CMDrainMarkingStackClosure: public VoidClosure {
2092 ConcurrentMark* _cm;
2093 CMMarkStack* _markStack;
2094 G1CMKeepAliveClosure* _oopClosure;
2095 public:
2096 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMMarkStack* markStack,
2097 G1CMKeepAliveClosure* oopClosure) :
2098 _cm(cm),
2099 _markStack(markStack),
2100 _oopClosure(oopClosure) { }
2101
2102 void do_void() {
2103 _markStack->drain((OopClosure*)_oopClosure, _cm->nextMarkBitMap(), false);
2104 }
2105 };
2106
2107 // 'Keep Alive' closure used by parallel reference processing.
2108 // An instance of this closure is used in the parallel reference processing
2109 // code rather than an instance of G1CMKeepAliveClosure. We could have used
2110 // the G1CMKeepAliveClosure as it is MT-safe. Also reference objects are
2111 // placed on to discovered ref lists once so we can mark and push with no
2112 // need to check whether the object has already been marked. Using the
2113 // G1CMKeepAliveClosure would mean, however, having all the worker threads
2114 // operating on the global mark stack. This means that an individual
2115 // worker would be doing lock-free pushes while it processes its own
2116 // discovered ref list followed by drain call. If the discovered ref lists
2117 // are unbalanced then this could cause interference with the other
2118 // workers. Using a CMTask (and its embedded local data structures)
2119 // avoids that potential interference.
2120 class G1CMParKeepAliveAndDrainClosure: public OopClosure {
2121 ConcurrentMark* _cm;
2122 CMTask* _task;
2123 int _ref_counter_limit;
2124 int _ref_counter;
2125 public:
2126 G1CMParKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task) :
2127 _cm(cm), _task(task),
2128 _ref_counter_limit(G1RefProcDrainInterval) {
2129 assert(_ref_counter_limit > 0, "sanity");
2130 _ref_counter = _ref_counter_limit;
2131 }
2132
2133 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2134 virtual void do_oop( oop* p) { do_oop_work(p); }
2135
2136 template <class T> void do_oop_work(T* p) {
2137 if (!_cm->has_overflown()) {
2138 oop obj = oopDesc::load_decode_heap_oop(p);
2139 if (_cm->verbose_high()) {
2140 gclog_or_tty->print_cr("\t[%d] we're looking at location "
2141 "*"PTR_FORMAT" = "PTR_FORMAT,
2142 _task->task_id(), p, (void*) obj);
2143 }
2144
2145 _task->deal_with_reference(obj);
2146 _ref_counter--;
2147
2148 if (_ref_counter == 0) {
2149 // We have dealt with _ref_counter_limit references, pushing them and objects
2150 // reachable from them on to the local stack (and possibly the global stack).
2151 // Call do_marking_step() to process these entries. We call the routine in a
2152 // loop, which we'll exit if there's nothing more to do (i.e. we're done
2153 // with the entries that we've pushed as a result of the deal_with_reference
2154 // calls above) or we overflow.
2155 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() flag
2156 // while there may still be some work to do. (See the comment at the
2157 // beginning of CMTask::do_marking_step() for those conditions - one of which
2158 // is reaching the specified time target.) It is only when
2159 // CMTask::do_marking_step() returns without setting the has_aborted() flag
2160 // that the marking has completed.
2161 do {
2162 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2163 _task->do_marking_step(mark_step_duration_ms,
2164 false /* do_stealing */,
2165 false /* do_termination */);
2166 } while (_task->has_aborted() && !_cm->has_overflown());
2167 _ref_counter = _ref_counter_limit;
2168 }
2169 } else {
2170 if (_cm->verbose_high()) {
2171 gclog_or_tty->print_cr("\t[%d] CM Overflow", _task->task_id());
2172 }
2173 }
2174 }
2175 };
2176
2177 class G1CMParDrainMarkingStackClosure: public VoidClosure {
2178 ConcurrentMark* _cm;
2179 CMTask* _task;
2180 public:
2181 G1CMParDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task) :
2182 _cm(cm), _task(task) { }
2183
2184 void do_void() {
2185 do {
2186 if (_cm->verbose_high()) {
2187 gclog_or_tty->print_cr("\t[%d] Drain: Calling do marking_step",
2188 _task->task_id());
2189 }
2190
2191 // We call CMTask::do_marking_step() to completely drain the local and
2192 // global marking stacks. The routine is called in a loop, which we'll
2193 // exit if there's nothing more to do (i.e. we'completely drained the
2194 // entries that were pushed as a result of applying the
2195 // G1CMParKeepAliveAndDrainClosure to the entries on the discovered ref
2196 // lists above) or we overflow the global marking stack.
2197 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() flag
2198 // while there may still be some work to do. (See the comment at the
2199 // beginning of CMTask::do_marking_step() for those conditions - one of which
2200 // is reaching the specified time target.) It is only when
2201 // CMTask::do_marking_step() returns without setting the has_aborted() flag
2202 // that the marking has completed.
2203
2204 _task->do_marking_step(1000000000.0 /* something very large */,
2205 true /* do_stealing */,
2206 true /* do_termination */);
2207 } while (_task->has_aborted() && !_cm->has_overflown());
2208 }
2209 };
2210
2211 // Implementation of AbstractRefProcTaskExecutor for parallel
2212 // reference processing at the end of G1 concurrent marking
2213
2214 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2215 private:
2216 G1CollectedHeap* _g1h;
2217 ConcurrentMark* _cm;
2218 WorkGang* _workers;
2219 int _active_workers;
2220
2221 public:
2222 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2223 ConcurrentMark* cm,
2224 WorkGang* workers,
2225 int n_workers) :
2226 _g1h(g1h), _cm(cm),
2227 _workers(workers), _active_workers(n_workers) { }
2228
2229 // Executes the given task using concurrent marking worker threads.
2230 virtual void execute(ProcessTask& task);
2231 virtual void execute(EnqueueTask& task);
2232 };
2233
2234 class G1CMRefProcTaskProxy: public AbstractGangTask {
2235 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2236 ProcessTask& _proc_task;
2237 G1CollectedHeap* _g1h;
2238 ConcurrentMark* _cm;
2239
2240 public:
2241 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2242 G1CollectedHeap* g1h,
2243 ConcurrentMark* cm) :
2244 AbstractGangTask("Process reference objects in parallel"),
2245 _proc_task(proc_task), _g1h(g1h), _cm(cm) { }
2246
2247 virtual void work(uint worker_id) {
2248 CMTask* marking_task = _cm->task(worker_id);
2249 G1CMIsAliveClosure g1_is_alive(_g1h);
2250 G1CMParKeepAliveAndDrainClosure g1_par_keep_alive(_cm, marking_task);
2251 G1CMParDrainMarkingStackClosure g1_par_drain(_cm, marking_task);
2252
2253 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2254 }
2255 };
2256
2257 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2258 assert(_workers != NULL, "Need parallel worker threads.");
2259
2260 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2261
2262 // We need to reset the phase for each task execution so that
2263 // the termination protocol of CMTask::do_marking_step works.
2264 _cm->set_phase(_active_workers, false /* concurrent */);
2265 _g1h->set_par_threads(_active_workers);
2266 _workers->run_task(&proc_task_proxy);
2267 _g1h->set_par_threads(0);
2268 }
2269
2270 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2271 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2272 EnqueueTask& _enq_task;
2273
2274 public:
2275 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2276 AbstractGangTask("Enqueue reference objects in parallel"),
2277 _enq_task(enq_task) { }
2278
2279 virtual void work(uint worker_id) {
2280 _enq_task.work(worker_id);
2281 }
2282 };
2283
2284 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2285 assert(_workers != NULL, "Need parallel worker threads.");
2286
2287 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2288
2289 _g1h->set_par_threads(_active_workers);
2290 _workers->run_task(&enq_task_proxy);
2291 _g1h->set_par_threads(0);
2292 }
2293
2294 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2295 ResourceMark rm;
2296 HandleMark hm;
2297
2298 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2299
2300 // Is alive closure.
2301 G1CMIsAliveClosure g1_is_alive(g1h);
2302
2303 // Inner scope to exclude the cleaning of the string and symbol
2304 // tables from the displayed time.
2305 {
2306 if (G1Log::finer()) {
2307 gclog_or_tty->put(' ');
2308 }
2309 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm());
2310
2311 ReferenceProcessor* rp = g1h->ref_processor_cm();
2312
2313 // See the comment in G1CollectedHeap::ref_processing_init()
2314 // about how reference processing currently works in G1.
2315
2316 // Process weak references.
2317 rp->setup_policy(clear_all_soft_refs);
2318 assert(_markStack.isEmpty(), "mark stack should be empty");
2319
2320 G1CMKeepAliveClosure g1_keep_alive(g1h, this);
2321 G1CMDrainMarkingStackClosure
2322 g1_drain_mark_stack(this, &_markStack, &g1_keep_alive);
2323
2324 // We use the work gang from the G1CollectedHeap and we utilize all
2325 // the worker threads.
2326 uint active_workers = g1h->workers() ? g1h->workers()->active_workers() : 1U;
2327 active_workers = MAX2(MIN2(active_workers, _max_task_num), 1U);
2328
2329 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2330 g1h->workers(), active_workers);
2331
2332 ReferenceProcessorStats stats;
2333 if (rp->processing_is_mt()) {
2334 // Set the degree of MT here. If the discovery is done MT, there
2335 // may have been a different number of threads doing the discovery
2336 // and a different number of discovered lists may have Ref objects.
2337 // That is OK as long as the Reference lists are balanced (see
2338 // balance_all_queues() and balance_queues()).
2339 rp->set_active_mt_degree(active_workers);
2340
2341 stats = rp->process_discovered_references(&g1_is_alive,
2342 &g1_keep_alive,
2343 &g1_drain_mark_stack,
2344 &par_task_executor,
2345 g1h->gc_timer_cm());
2346
2347 // The work routines of the parallel keep_alive and drain_marking_stack
2348 // will set the has_overflown flag if we overflow the global marking
2349 // stack.
2350 } else {
2351 stats = rp->process_discovered_references(&g1_is_alive,
2352 &g1_keep_alive,
2353 &g1_drain_mark_stack,
2354 NULL,
2355 g1h->gc_timer_cm());
2356 }
2357
2358 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2359
2360 assert(_markStack.overflow() || _markStack.isEmpty(),
2361 "mark stack should be empty (unless it overflowed)");
2362 if (_markStack.overflow()) {
2363 // Should have been done already when we tried to push an
2364 // entry on to the global mark stack. But let's do it again.
2365 set_has_overflown();
2366 }
2367
2368 if (rp->processing_is_mt()) {
2369 assert(rp->num_q() == active_workers, "why not");
2370 rp->enqueue_discovered_references(&par_task_executor);
2371 } else {
2372 rp->enqueue_discovered_references();
2373 }
2374
2375 rp->verify_no_references_recorded();
2376 assert(!rp->discovery_enabled(), "Post condition");
2377 }
2378
2379 // Now clean up stale oops in StringTable
2380 StringTable::unlink(&g1_is_alive);
2381 // Clean up unreferenced symbols in symbol table.
2382 SymbolTable::unlink();
2383 }
2384
2385 void ConcurrentMark::swapMarkBitMaps() {
2386 CMBitMapRO* temp = _prevMarkBitMap;
2387 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2388 _nextMarkBitMap = (CMBitMap*) temp;
2389 }
2390
2391 class CMRemarkTask: public AbstractGangTask {
2392 private:
2393 ConcurrentMark *_cm;
2394
2395 public:
2396 void work(uint worker_id) {
2397 // Since all available tasks are actually started, we should
2398 // only proceed if we're supposed to be actived.
2399 if (worker_id < _cm->active_tasks()) {
2400 CMTask* task = _cm->task(worker_id);
2401 task->record_start_time();
2402 do {
2403 task->do_marking_step(1000000000.0 /* something very large */,
2404 true /* do_stealing */,
2405 true /* do_termination */);
2406 } while (task->has_aborted() && !_cm->has_overflown());
2407 // If we overflow, then we do not want to restart. We instead
2408 // want to abort remark and do concurrent marking again.
2409 task->record_end_time();
2410 }
2411 }
2412
2413 CMRemarkTask(ConcurrentMark* cm, int active_workers) :
2414 AbstractGangTask("Par Remark"), _cm(cm) {
2415 _cm->terminator()->reset_for_reuse(active_workers);
2416 }
2417 };
2418
2419 void ConcurrentMark::checkpointRootsFinalWork() {
2420 ResourceMark rm;
2421 HandleMark hm;
2422 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2423
2424 g1h->ensure_parsability(false);
2425
2426 if (G1CollectedHeap::use_parallel_gc_threads()) {
2427 G1CollectedHeap::StrongRootsScope srs(g1h);
2428 // this is remark, so we'll use up all active threads
2429 uint active_workers = g1h->workers()->active_workers();
2430 if (active_workers == 0) {
2431 assert(active_workers > 0, "Should have been set earlier");
2432 active_workers = (uint) ParallelGCThreads;
2433 g1h->workers()->set_active_workers(active_workers);
2434 }
2435 set_phase(active_workers, false /* concurrent */);
2436 // Leave _parallel_marking_threads at it's
2437 // value originally calculated in the ConcurrentMark
2438 // constructor and pass values of the active workers
2439 // through the gang in the task.
2440
2441 CMRemarkTask remarkTask(this, active_workers);
2442 g1h->set_par_threads(active_workers);
2443 g1h->workers()->run_task(&remarkTask);
2444 g1h->set_par_threads(0);
2445 } else {
2446 G1CollectedHeap::StrongRootsScope srs(g1h);
2447 // this is remark, so we'll use up all available threads
2448 uint active_workers = 1;
2449 set_phase(active_workers, false /* concurrent */);
2450
2451 CMRemarkTask remarkTask(this, active_workers);
2452 // We will start all available threads, even if we decide that the
2453 // active_workers will be fewer. The extra ones will just bail out
2454 // immediately.
2455 remarkTask.work(0);
2456 }
2457 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2458 guarantee(satb_mq_set.completed_buffers_num() == 0, "invariant");
2459
2460 print_stats();
2461
2462 #if VERIFY_OBJS_PROCESSED
2463 if (_scan_obj_cl.objs_processed != ThreadLocalObjQueue::objs_enqueued) {
2464 gclog_or_tty->print_cr("Processed = %d, enqueued = %d.",
2465 _scan_obj_cl.objs_processed,
2466 ThreadLocalObjQueue::objs_enqueued);
2467 guarantee(_scan_obj_cl.objs_processed ==
2468 ThreadLocalObjQueue::objs_enqueued,
2469 "Different number of objs processed and enqueued.");
2470 }
2471 #endif
2472 }
2473
2474 #ifndef PRODUCT
2475
2476 class PrintReachableOopClosure: public OopClosure {
2477 private:
2478 G1CollectedHeap* _g1h;
2479 outputStream* _out;
2480 VerifyOption _vo;
2481 bool _all;
2482
2483 public:
2484 PrintReachableOopClosure(outputStream* out,
2485 VerifyOption vo,
2486 bool all) :
2487 _g1h(G1CollectedHeap::heap()),
2488 _out(out), _vo(vo), _all(all) { }
2489
2490 void do_oop(narrowOop* p) { do_oop_work(p); }
2491 void do_oop( oop* p) { do_oop_work(p); }
2492
2493 template <class T> void do_oop_work(T* p) {
2494 oop obj = oopDesc::load_decode_heap_oop(p);
2495 const char* str = NULL;
2496 const char* str2 = "";
2497
2498 if (obj == NULL) {
2499 str = "";
2500 } else if (!_g1h->is_in_g1_reserved(obj)) {
2501 str = " O";
2502 } else {
2503 HeapRegion* hr = _g1h->heap_region_containing(obj);
2504 guarantee(hr != NULL, "invariant");
2505 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2506 bool marked = _g1h->is_marked(obj, _vo);
2507
2508 if (over_tams) {
2509 str = " >";
2510 if (marked) {
2511 str2 = " AND MARKED";
2512 }
2513 } else if (marked) {
2514 str = " M";
2515 } else {
2516 str = " NOT";
2517 }
2518 }
2519
2520 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2521 p, (void*) obj, str, str2);
2522 }
2523 };
2524
2525 class PrintReachableObjectClosure : public ObjectClosure {
2526 private:
2527 G1CollectedHeap* _g1h;
2528 outputStream* _out;
2529 VerifyOption _vo;
2530 bool _all;
2531 HeapRegion* _hr;
2532
2533 public:
2534 PrintReachableObjectClosure(outputStream* out,
2535 VerifyOption vo,
2536 bool all,
2537 HeapRegion* hr) :
2538 _g1h(G1CollectedHeap::heap()),
2539 _out(out), _vo(vo), _all(all), _hr(hr) { }
2540
2541 void do_object(oop o) {
2542 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2543 bool marked = _g1h->is_marked(o, _vo);
2544 bool print_it = _all || over_tams || marked;
2545
2546 if (print_it) {
2547 _out->print_cr(" "PTR_FORMAT"%s",
2548 o, (over_tams) ? " >" : (marked) ? " M" : "");
2549 PrintReachableOopClosure oopCl(_out, _vo, _all);
2550 o->oop_iterate(&oopCl);
2551 }
2552 }
2553 };
2554
2555 class PrintReachableRegionClosure : public HeapRegionClosure {
2556 private:
2557 G1CollectedHeap* _g1h;
2558 outputStream* _out;
2559 VerifyOption _vo;
2560 bool _all;
2561
2562 public:
2563 bool doHeapRegion(HeapRegion* hr) {
2564 HeapWord* b = hr->bottom();
2565 HeapWord* e = hr->end();
2566 HeapWord* t = hr->top();
2567 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2568 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2569 "TAMS: "PTR_FORMAT, b, e, t, p);
2570 _out->cr();
2571
2572 HeapWord* from = b;
2573 HeapWord* to = t;
2574
2575 if (to > from) {
2576 _out->print_cr("Objects in ["PTR_FORMAT", "PTR_FORMAT"]", from, to);
2577 _out->cr();
2578 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2579 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2580 _out->cr();
2581 }
2582
2583 return false;
2584 }
2585
2586 PrintReachableRegionClosure(outputStream* out,
2587 VerifyOption vo,
2588 bool all) :
2589 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2590 };
2591
2592 void ConcurrentMark::print_reachable(const char* str,
2593 VerifyOption vo,
2594 bool all) {
2595 gclog_or_tty->cr();
2596 gclog_or_tty->print_cr("== Doing heap dump... ");
2597
2598 if (G1PrintReachableBaseFile == NULL) {
2599 gclog_or_tty->print_cr(" #### error: no base file defined");
2600 return;
2601 }
2602
2603 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2604 (JVM_MAXPATHLEN - 1)) {
2605 gclog_or_tty->print_cr(" #### error: file name too long");
2606 return;
2607 }
2608
2609 char file_name[JVM_MAXPATHLEN];
2610 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2611 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2612
2613 fileStream fout(file_name);
2614 if (!fout.is_open()) {
2615 gclog_or_tty->print_cr(" #### error: could not open file");
2616 return;
2617 }
2618
2619 outputStream* out = &fout;
2620 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2621 out->cr();
2622
2623 out->print_cr("--- ITERATING OVER REGIONS");
2624 out->cr();
2625 PrintReachableRegionClosure rcl(out, vo, all);
2626 _g1h->heap_region_iterate(&rcl);
2627 out->cr();
2628
2629 gclog_or_tty->print_cr(" done");
2630 gclog_or_tty->flush();
2631 }
2632
2633 #endif // PRODUCT
2634
2635 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2636 // Note we are overriding the read-only view of the prev map here, via
2637 // the cast.
2638 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2639 }
2640
2641 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2642 _nextMarkBitMap->clearRange(mr);
2643 }
2644
2645 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) {
2646 clearRangePrevBitmap(mr);
2647 clearRangeNextBitmap(mr);
2648 }
2649
2650 HeapRegion*
2651 ConcurrentMark::claim_region(int task_num) {
2652 // "checkpoint" the finger
2653 HeapWord* finger = _finger;
2654
2655 // _heap_end will not change underneath our feet; it only changes at
2656 // yield points.
2657 while (finger < _heap_end) {
2658 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2659
2660 // Note on how this code handles humongous regions. In the
2661 // normal case the finger will reach the start of a "starts
2662 // humongous" (SH) region. Its end will either be the end of the
2663 // last "continues humongous" (CH) region in the sequence, or the
2664 // standard end of the SH region (if the SH is the only region in
2665 // the sequence). That way claim_region() will skip over the CH
2666 // regions. However, there is a subtle race between a CM thread
2667 // executing this method and a mutator thread doing a humongous
2668 // object allocation. The two are not mutually exclusive as the CM
2669 // thread does not need to hold the Heap_lock when it gets
2670 // here. So there is a chance that claim_region() will come across
2671 // a free region that's in the progress of becoming a SH or a CH
2672 // region. In the former case, it will either
2673 // a) Miss the update to the region's end, in which case it will
2674 // visit every subsequent CH region, will find their bitmaps
2675 // empty, and do nothing, or
2676 // b) Will observe the update of the region's end (in which case
2677 // it will skip the subsequent CH regions).
2678 // If it comes across a region that suddenly becomes CH, the
2679 // scenario will be similar to b). So, the race between
2680 // claim_region() and a humongous object allocation might force us
2681 // to do a bit of unnecessary work (due to some unnecessary bitmap
2682 // iterations) but it should not introduce and correctness issues.
2683 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2684 HeapWord* bottom = curr_region->bottom();
2685 HeapWord* end = curr_region->end();
2686 HeapWord* limit = curr_region->next_top_at_mark_start();
2687
2688 if (verbose_low()) {
2689 gclog_or_tty->print_cr("[%d] curr_region = "PTR_FORMAT" "
2690 "["PTR_FORMAT", "PTR_FORMAT"), "
2691 "limit = "PTR_FORMAT,
2692 task_num, curr_region, bottom, end, limit);
2693 }
2694
2695 // Is the gap between reading the finger and doing the CAS too long?
2696 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2697 if (res == finger) {
2698 // we succeeded
2699
2700 // notice that _finger == end cannot be guaranteed here since,
2701 // someone else might have moved the finger even further
2702 assert(_finger >= end, "the finger should have moved forward");
2703
2704 if (verbose_low()) {
2705 gclog_or_tty->print_cr("[%d] we were successful with region = "
2706 PTR_FORMAT, task_num, curr_region);
2707 }
2708
2709 if (limit > bottom) {
2710 if (verbose_low()) {
2711 gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is not empty, "
2712 "returning it ", task_num, curr_region);
2713 }
2714 return curr_region;
2715 } else {
2716 assert(limit == bottom,
2717 "the region limit should be at bottom");
2718 if (verbose_low()) {
2719 gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is empty, "
2720 "returning NULL", task_num, curr_region);
2721 }
2722 // we return NULL and the caller should try calling
2723 // claim_region() again.
2724 return NULL;
2725 }
2726 } else {
2727 assert(_finger > finger, "the finger should have moved forward");
2728 if (verbose_low()) {
2729 gclog_or_tty->print_cr("[%d] somebody else moved the finger, "
2730 "global finger = "PTR_FORMAT", "
2731 "our finger = "PTR_FORMAT,
2732 task_num, _finger, finger);
2733 }
2734
2735 // read it again
2736 finger = _finger;
2737 }
2738 }
2739
2740 return NULL;
2741 }
2742
2743 #ifndef PRODUCT
2744 enum VerifyNoCSetOopsPhase {
2745 VerifyNoCSetOopsStack,
2746 VerifyNoCSetOopsQueues,
2747 VerifyNoCSetOopsSATBCompleted,
2748 VerifyNoCSetOopsSATBThread
2749 };
2750
2751 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
2752 private:
2753 G1CollectedHeap* _g1h;
2754 VerifyNoCSetOopsPhase _phase;
2755 int _info;
2756
2757 const char* phase_str() {
2758 switch (_phase) {
2759 case VerifyNoCSetOopsStack: return "Stack";
2760 case VerifyNoCSetOopsQueues: return "Queue";
2761 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2762 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
2763 default: ShouldNotReachHere();
2764 }
2765 return NULL;
2766 }
2767
2768 void do_object_work(oop obj) {
2769 guarantee(!_g1h->obj_in_cs(obj),
2770 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2771 (void*) obj, phase_str(), _info));
2772 }
2773
2774 public:
2775 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2776
2777 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2778 _phase = phase;
2779 _info = info;
2780 }
2781
2782 virtual void do_oop(oop* p) {
2783 oop obj = oopDesc::load_decode_heap_oop(p);
2784 do_object_work(obj);
2785 }
2786
2787 virtual void do_oop(narrowOop* p) {
2788 // We should not come across narrow oops while scanning marking
2789 // stacks and SATB buffers.
2790 ShouldNotReachHere();
2791 }
2792
2793 virtual void do_object(oop obj) {
2794 do_object_work(obj);
2795 }
2796 };
2797
2798 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
2799 bool verify_enqueued_buffers,
2800 bool verify_thread_buffers,
2801 bool verify_fingers) {
2802 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2803 if (!G1CollectedHeap::heap()->mark_in_progress()) {
2804 return;
2805 }
2806
2807 VerifyNoCSetOopsClosure cl;
2808
2809 if (verify_stacks) {
2810 // Verify entries on the global mark stack
2811 cl.set_phase(VerifyNoCSetOopsStack);
2812 _markStack.oops_do(&cl);
2813
2814 // Verify entries on the task queues
2815 for (int i = 0; i < (int) _max_task_num; i += 1) {
2816 cl.set_phase(VerifyNoCSetOopsQueues, i);
2817 OopTaskQueue* queue = _task_queues->queue(i);
2818 queue->oops_do(&cl);
2819 }
2820 }
2821
2822 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
2823
2824 // Verify entries on the enqueued SATB buffers
2825 if (verify_enqueued_buffers) {
2826 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
2827 satb_qs.iterate_completed_buffers_read_only(&cl);
2828 }
2829
2830 // Verify entries on the per-thread SATB buffers
2831 if (verify_thread_buffers) {
2832 cl.set_phase(VerifyNoCSetOopsSATBThread);
2833 satb_qs.iterate_thread_buffers_read_only(&cl);
2834 }
2835
2836 if (verify_fingers) {
2837 // Verify the global finger
2838 HeapWord* global_finger = finger();
2839 if (global_finger != NULL && global_finger < _heap_end) {
2840 // The global finger always points to a heap region boundary. We
2841 // use heap_region_containing_raw() to get the containing region
2842 // given that the global finger could be pointing to a free region
2843 // which subsequently becomes continues humongous. If that
2844 // happens, heap_region_containing() will return the bottom of the
2845 // corresponding starts humongous region and the check below will
2846 // not hold any more.
2847 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
2848 guarantee(global_finger == global_hr->bottom(),
2849 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
2850 global_finger, HR_FORMAT_PARAMS(global_hr)));
2851 }
2852
2853 // Verify the task fingers
2854 assert(parallel_marking_threads() <= _max_task_num, "sanity");
2855 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
2856 CMTask* task = _tasks[i];
2857 HeapWord* task_finger = task->finger();
2858 if (task_finger != NULL && task_finger < _heap_end) {
2859 // See above note on the global finger verification.
2860 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
2861 guarantee(task_finger == task_hr->bottom() ||
2862 !task_hr->in_collection_set(),
2863 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
2864 task_finger, HR_FORMAT_PARAMS(task_hr)));
2865 }
2866 }
2867 }
2868 }
2869 #endif // PRODUCT
2870
2871 // Aggregate the counting data that was constructed concurrently
2872 // with marking.
2873 class AggregateCountDataHRClosure: public HeapRegionClosure {
2874 G1CollectedHeap* _g1h;
2875 ConcurrentMark* _cm;
2876 CardTableModRefBS* _ct_bs;
2877 BitMap* _cm_card_bm;
2878 size_t _max_task_num;
2879
2880 public:
2881 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
2882 BitMap* cm_card_bm,
2883 size_t max_task_num) :
2884 _g1h(g1h), _cm(g1h->concurrent_mark()),
2885 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
2886 _cm_card_bm(cm_card_bm), _max_task_num(max_task_num) { }
2887
2888 bool doHeapRegion(HeapRegion* hr) {
2889 if (hr->continuesHumongous()) {
2890 // We will ignore these here and process them when their
2891 // associated "starts humongous" region is processed.
2892 // Note that we cannot rely on their associated
2893 // "starts humongous" region to have their bit set to 1
2894 // since, due to the region chunking in the parallel region
2895 // iteration, a "continues humongous" region might be visited
2896 // before its associated "starts humongous".
2897 return false;
2898 }
2899
2900 HeapWord* start = hr->bottom();
2901 HeapWord* limit = hr->next_top_at_mark_start();
2902 HeapWord* end = hr->end();
2903
2904 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
2905 err_msg("Preconditions not met - "
2906 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
2907 "top: "PTR_FORMAT", end: "PTR_FORMAT,
2908 start, limit, hr->top(), hr->end()));
2909
2910 assert(hr->next_marked_bytes() == 0, "Precondition");
2911
2912 if (start == limit) {
2913 // NTAMS of this region has not been set so nothing to do.
2914 return false;
2915 }
2916
2917 // 'start' should be in the heap.
2918 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
2919 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
2920 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
2921
2922 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
2923 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
2924 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
2925
2926 // If ntams is not card aligned then we bump card bitmap index
2927 // for limit so that we get the all the cards spanned by
2928 // the object ending at ntams.
2929 // Note: if this is the last region in the heap then ntams
2930 // could be actually just beyond the end of the the heap;
2931 // limit_idx will then correspond to a (non-existent) card
2932 // that is also outside the heap.
2933 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
2934 limit_idx += 1;
2935 }
2936
2937 assert(limit_idx <= end_idx, "or else use atomics");
2938
2939 // Aggregate the "stripe" in the count data associated with hr.
2940 uint hrs_index = hr->hrs_index();
2941 size_t marked_bytes = 0;
2942
2943 for (int i = 0; (size_t)i < _max_task_num; i += 1) {
2944 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
2945 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
2946
2947 // Fetch the marked_bytes in this region for task i and
2948 // add it to the running total for this region.
2949 marked_bytes += marked_bytes_array[hrs_index];
2950
2951 // Now union the bitmaps[0,max_task_num)[start_idx..limit_idx)
2952 // into the global card bitmap.
2953 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
2954
2955 while (scan_idx < limit_idx) {
2956 assert(task_card_bm->at(scan_idx) == true, "should be");
2957 _cm_card_bm->set_bit(scan_idx);
2958 assert(_cm_card_bm->at(scan_idx) == true, "should be");
2959
2960 // BitMap::get_next_one_offset() can handle the case when
2961 // its left_offset parameter is greater than its right_offset
2962 // parameter. It does, however, have an early exit if
2963 // left_offset == right_offset. So let's limit the value
2964 // passed in for left offset here.
2965 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
2966 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
2967 }
2968 }
2969
2970 // Update the marked bytes for this region.
2971 hr->add_to_marked_bytes(marked_bytes);
2972
2973 // Next heap region
2974 return false;
2975 }
2976 };
2977
2978 class G1AggregateCountDataTask: public AbstractGangTask {
2979 protected:
2980 G1CollectedHeap* _g1h;
2981 ConcurrentMark* _cm;
2982 BitMap* _cm_card_bm;
2983 size_t _max_task_num;
2984 int _active_workers;
2985
2986 public:
2987 G1AggregateCountDataTask(G1CollectedHeap* g1h,
2988 ConcurrentMark* cm,
2989 BitMap* cm_card_bm,
2990 size_t max_task_num,
2991 int n_workers) :
2992 AbstractGangTask("Count Aggregation"),
2993 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
2994 _max_task_num(max_task_num),
2995 _active_workers(n_workers) { }
2996
2997 void work(uint worker_id) {
2998 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_task_num);
2999
3000 if (G1CollectedHeap::use_parallel_gc_threads()) {
3001 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3002 _active_workers,
3003 HeapRegion::AggregateCountClaimValue);
3004 } else {
3005 _g1h->heap_region_iterate(&cl);
3006 }
3007 }
3008 };
3009
3010
3011 void ConcurrentMark::aggregate_count_data() {
3012 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3013 _g1h->workers()->active_workers() :
3014 1);
3015
3016 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3017 _max_task_num, n_workers);
3018
3019 if (G1CollectedHeap::use_parallel_gc_threads()) {
3020 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3021 "sanity check");
3022 _g1h->set_par_threads(n_workers);
3023 _g1h->workers()->run_task(&g1_par_agg_task);
3024 _g1h->set_par_threads(0);
3025
3026 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3027 "sanity check");
3028 _g1h->reset_heap_region_claim_values();
3029 } else {
3030 g1_par_agg_task.work(0);
3031 }
3032 }
3033
3034 // Clear the per-worker arrays used to store the per-region counting data
3035 void ConcurrentMark::clear_all_count_data() {
3036 // Clear the global card bitmap - it will be filled during
3037 // liveness count aggregation (during remark) and the
3038 // final counting task.
3039 _card_bm.clear();
3040
3041 // Clear the global region bitmap - it will be filled as part
3042 // of the final counting task.
3043 _region_bm.clear();
3044
3045 uint max_regions = _g1h->max_regions();
3046 assert(_max_task_num != 0, "unitialized");
3047
3048 for (int i = 0; (size_t) i < _max_task_num; i += 1) {
3049 BitMap* task_card_bm = count_card_bitmap_for(i);
3050 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3051
3052 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3053 assert(marked_bytes_array != NULL, "uninitialized");
3054
3055 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3056 task_card_bm->clear();
3057 }
3058 }
3059
3060 void ConcurrentMark::print_stats() {
3061 if (verbose_stats()) {
3062 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3063 for (size_t i = 0; i < _active_tasks; ++i) {
3064 _tasks[i]->print_stats();
3065 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3066 }
3067 }
3068 }
3069
3070 // abandon current marking iteration due to a Full GC
3071 void ConcurrentMark::abort() {
3072 // Clear all marks to force marking thread to do nothing
3073 _nextMarkBitMap->clearAll();
3074 // Clear the liveness counting data
3075 clear_all_count_data();
3076 // Empty mark stack
3077 reset_marking_state();
3078 for (int i = 0; i < (int)_max_task_num; ++i) {
3079 _tasks[i]->clear_region_fields();
3080 }
3081 _has_aborted = true;
3082
3083 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3084 satb_mq_set.abandon_partial_marking();
3085 // This can be called either during or outside marking, we'll read
3086 // the expected_active value from the SATB queue set.
3087 satb_mq_set.set_active_all_threads(
3088 false, /* new active value */
3089 satb_mq_set.is_active() /* expected_active */);
3090
3091 _g1h->trace_heap_after_concurrent_cycle();
3092 _g1h->register_concurrent_cycle_end();
3093 }
3094
3095 static void print_ms_time_info(const char* prefix, const char* name,
3096 NumberSeq& ns) {
3097 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3098 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3099 if (ns.num() > 0) {
3100 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3101 prefix, ns.sd(), ns.maximum());
3102 }
3103 }
3104
3105 void ConcurrentMark::print_summary_info() {
3106 gclog_or_tty->print_cr(" Concurrent marking:");
3107 print_ms_time_info(" ", "init marks", _init_times);
3108 print_ms_time_info(" ", "remarks", _remark_times);
3109 {
3110 print_ms_time_info(" ", "final marks", _remark_mark_times);
3111 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3112
3113 }
3114 print_ms_time_info(" ", "cleanups", _cleanup_times);
3115 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3116 _total_counting_time,
3117 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3118 (double)_cleanup_times.num()
3119 : 0.0));
3120 if (G1ScrubRemSets) {
3121 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3122 _total_rs_scrub_time,
3123 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3124 (double)_cleanup_times.num()
3125 : 0.0));
3126 }
3127 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3128 (_init_times.sum() + _remark_times.sum() +
3129 _cleanup_times.sum())/1000.0);
3130 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3131 "(%8.2f s marking).",
3132 cmThread()->vtime_accum(),
3133 cmThread()->vtime_mark_accum());
3134 }
3135
3136 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3137 if (use_parallel_marking_threads()) {
3138 _parallel_workers->print_worker_threads_on(st);
3139 }
3140 }
3141
3142 // We take a break if someone is trying to stop the world.
3143 bool ConcurrentMark::do_yield_check(uint worker_id) {
3144 if (should_yield()) {
3145 if (worker_id == 0) {
3146 _g1h->g1_policy()->record_concurrent_pause();
3147 }
3148 cmThread()->yield();
3149 return true;
3150 } else {
3151 return false;
3152 }
3153 }
3154
3155 bool ConcurrentMark::should_yield() {
3156 return cmThread()->should_yield();
3157 }
3158
3159 bool ConcurrentMark::containing_card_is_marked(void* p) {
3160 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1);
3161 return _card_bm.at(offset >> CardTableModRefBS::card_shift);
3162 }
3163
3164 bool ConcurrentMark::containing_cards_are_marked(void* start,
3165 void* last) {
3166 return containing_card_is_marked(start) &&
3167 containing_card_is_marked(last);
3168 }
3169
3170 #ifndef PRODUCT
3171 // for debugging purposes
3172 void ConcurrentMark::print_finger() {
3173 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3174 _heap_start, _heap_end, _finger);
3175 for (int i = 0; i < (int) _max_task_num; ++i) {
3176 gclog_or_tty->print(" %d: "PTR_FORMAT, i, _tasks[i]->finger());
3177 }
3178 gclog_or_tty->print_cr("");
3179 }
3180 #endif
3181
3182 void CMTask::scan_object(oop obj) {
3183 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3184
3185 if (_cm->verbose_high()) {
3186 gclog_or_tty->print_cr("[%d] we're scanning object "PTR_FORMAT,
3187 _task_id, (void*) obj);
3188 }
3189
3190 size_t obj_size = obj->size();
3191 _words_scanned += obj_size;
3192
3193 obj->oop_iterate(_cm_oop_closure);
3194 statsOnly( ++_objs_scanned );
3195 check_limits();
3196 }
3197
3198 // Closure for iteration over bitmaps
3199 class CMBitMapClosure : public BitMapClosure {
3200 private:
3201 // the bitmap that is being iterated over
3202 CMBitMap* _nextMarkBitMap;
3203 ConcurrentMark* _cm;
3204 CMTask* _task;
3205
3206 public:
3207 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3208 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3209
3210 bool do_bit(size_t offset) {
3211 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3212 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3213 assert( addr < _cm->finger(), "invariant");
3214
3215 statsOnly( _task->increase_objs_found_on_bitmap() );
3216 assert(addr >= _task->finger(), "invariant");
3217
3218 // We move that task's local finger along.
3219 _task->move_finger_to(addr);
3220
3221 _task->scan_object(oop(addr));
3222 // we only partially drain the local queue and global stack
3223 _task->drain_local_queue(true);
3224 _task->drain_global_stack(true);
3225
3226 // if the has_aborted flag has been raised, we need to bail out of
3227 // the iteration
3228 return !_task->has_aborted();
3229 }
3230 };
3231
3232 // Closure for iterating over objects, currently only used for
3233 // processing SATB buffers.
3234 class CMObjectClosure : public ObjectClosure {
3235 private:
3236 CMTask* _task;
3237
3238 public:
3239 void do_object(oop obj) {
3240 _task->deal_with_reference(obj);
3241 }
3242
3243 CMObjectClosure(CMTask* task) : _task(task) { }
3244 };
3245
3246 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3247 ConcurrentMark* cm,
3248 CMTask* task)
3249 : _g1h(g1h), _cm(cm), _task(task) {
3250 assert(_ref_processor == NULL, "should be initialized to NULL");
3251
3252 if (G1UseConcMarkReferenceProcessing) {
3253 _ref_processor = g1h->ref_processor_cm();
3254 assert(_ref_processor != NULL, "should not be NULL");
3255 }
3256 }
3257
3258 void CMTask::setup_for_region(HeapRegion* hr) {
3259 // Separated the asserts so that we know which one fires.
3260 assert(hr != NULL,
3261 "claim_region() should have filtered out continues humongous regions");
3262 assert(!hr->continuesHumongous(),
3263 "claim_region() should have filtered out continues humongous regions");
3264
3265 if (_cm->verbose_low()) {
3266 gclog_or_tty->print_cr("[%d] setting up for region "PTR_FORMAT,
3267 _task_id, hr);
3268 }
3269
3270 _curr_region = hr;
3271 _finger = hr->bottom();
3272 update_region_limit();
3273 }
3274
3275 void CMTask::update_region_limit() {
3276 HeapRegion* hr = _curr_region;
3277 HeapWord* bottom = hr->bottom();
3278 HeapWord* limit = hr->next_top_at_mark_start();
3279
3280 if (limit == bottom) {
3281 if (_cm->verbose_low()) {
3282 gclog_or_tty->print_cr("[%d] found an empty region "
3283 "["PTR_FORMAT", "PTR_FORMAT")",
3284 _task_id, bottom, limit);
3285 }
3286 // The region was collected underneath our feet.
3287 // We set the finger to bottom to ensure that the bitmap
3288 // iteration that will follow this will not do anything.
3289 // (this is not a condition that holds when we set the region up,
3290 // as the region is not supposed to be empty in the first place)
3291 _finger = bottom;
3292 } else if (limit >= _region_limit) {
3293 assert(limit >= _finger, "peace of mind");
3294 } else {
3295 assert(limit < _region_limit, "only way to get here");
3296 // This can happen under some pretty unusual circumstances. An
3297 // evacuation pause empties the region underneath our feet (NTAMS
3298 // at bottom). We then do some allocation in the region (NTAMS
3299 // stays at bottom), followed by the region being used as a GC
3300 // alloc region (NTAMS will move to top() and the objects
3301 // originally below it will be grayed). All objects now marked in
3302 // the region are explicitly grayed, if below the global finger,
3303 // and we do not need in fact to scan anything else. So, we simply
3304 // set _finger to be limit to ensure that the bitmap iteration
3305 // doesn't do anything.
3306 _finger = limit;
3307 }
3308
3309 _region_limit = limit;
3310 }
3311
3312 void CMTask::giveup_current_region() {
3313 assert(_curr_region != NULL, "invariant");
3314 if (_cm->verbose_low()) {
3315 gclog_or_tty->print_cr("[%d] giving up region "PTR_FORMAT,
3316 _task_id, _curr_region);
3317 }
3318 clear_region_fields();
3319 }
3320
3321 void CMTask::clear_region_fields() {
3322 // Values for these three fields that indicate that we're not
3323 // holding on to a region.
3324 _curr_region = NULL;
3325 _finger = NULL;
3326 _region_limit = NULL;
3327 }
3328
3329 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3330 if (cm_oop_closure == NULL) {
3331 assert(_cm_oop_closure != NULL, "invariant");
3332 } else {
3333 assert(_cm_oop_closure == NULL, "invariant");
3334 }
3335 _cm_oop_closure = cm_oop_closure;
3336 }
3337
3338 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3339 guarantee(nextMarkBitMap != NULL, "invariant");
3340
3341 if (_cm->verbose_low()) {
3342 gclog_or_tty->print_cr("[%d] resetting", _task_id);
3343 }
3344
3345 _nextMarkBitMap = nextMarkBitMap;
3346 clear_region_fields();
3347
3348 _calls = 0;
3349 _elapsed_time_ms = 0.0;
3350 _termination_time_ms = 0.0;
3351 _termination_start_time_ms = 0.0;
3352
3353 #if _MARKING_STATS_
3354 _local_pushes = 0;
3355 _local_pops = 0;
3356 _local_max_size = 0;
3357 _objs_scanned = 0;
3358 _global_pushes = 0;
3359 _global_pops = 0;
3360 _global_max_size = 0;
3361 _global_transfers_to = 0;
3362 _global_transfers_from = 0;
3363 _regions_claimed = 0;
3364 _objs_found_on_bitmap = 0;
3365 _satb_buffers_processed = 0;
3366 _steal_attempts = 0;
3367 _steals = 0;
3368 _aborted = 0;
3369 _aborted_overflow = 0;
3370 _aborted_cm_aborted = 0;
3371 _aborted_yield = 0;
3372 _aborted_timed_out = 0;
3373 _aborted_satb = 0;
3374 _aborted_termination = 0;
3375 #endif // _MARKING_STATS_
3376 }
3377
3378 bool CMTask::should_exit_termination() {
3379 regular_clock_call();
3380 // This is called when we are in the termination protocol. We should
3381 // quit if, for some reason, this task wants to abort or the global
3382 // stack is not empty (this means that we can get work from it).
3383 return !_cm->mark_stack_empty() || has_aborted();
3384 }
3385
3386 void CMTask::reached_limit() {
3387 assert(_words_scanned >= _words_scanned_limit ||
3388 _refs_reached >= _refs_reached_limit ,
3389 "shouldn't have been called otherwise");
3390 regular_clock_call();
3391 }
3392
3393 void CMTask::regular_clock_call() {
3394 if (has_aborted()) return;
3395
3396 // First, we need to recalculate the words scanned and refs reached
3397 // limits for the next clock call.
3398 recalculate_limits();
3399
3400 // During the regular clock call we do the following
3401
3402 // (1) If an overflow has been flagged, then we abort.
3403 if (_cm->has_overflown()) {
3404 set_has_aborted();
3405 return;
3406 }
3407
3408 // If we are not concurrent (i.e. we're doing remark) we don't need
3409 // to check anything else. The other steps are only needed during
3410 // the concurrent marking phase.
3411 if (!concurrent()) return;
3412
3413 // (2) If marking has been aborted for Full GC, then we also abort.
3414 if (_cm->has_aborted()) {
3415 set_has_aborted();
3416 statsOnly( ++_aborted_cm_aborted );
3417 return;
3418 }
3419
3420 double curr_time_ms = os::elapsedVTime() * 1000.0;
3421
3422 // (3) If marking stats are enabled, then we update the step history.
3423 #if _MARKING_STATS_
3424 if (_words_scanned >= _words_scanned_limit) {
3425 ++_clock_due_to_scanning;
3426 }
3427 if (_refs_reached >= _refs_reached_limit) {
3428 ++_clock_due_to_marking;
3429 }
3430
3431 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3432 _interval_start_time_ms = curr_time_ms;
3433 _all_clock_intervals_ms.add(last_interval_ms);
3434
3435 if (_cm->verbose_medium()) {
3436 gclog_or_tty->print_cr("[%d] regular clock, interval = %1.2lfms, "
3437 "scanned = %d%s, refs reached = %d%s",
3438 _task_id, last_interval_ms,
3439 _words_scanned,
3440 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3441 _refs_reached,
3442 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3443 }
3444 #endif // _MARKING_STATS_
3445
3446 // (4) We check whether we should yield. If we have to, then we abort.
3447 if (_cm->should_yield()) {
3448 // We should yield. To do this we abort the task. The caller is
3449 // responsible for yielding.
3450 set_has_aborted();
3451 statsOnly( ++_aborted_yield );
3452 return;
3453 }
3454
3455 // (5) We check whether we've reached our time quota. If we have,
3456 // then we abort.
3457 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3458 if (elapsed_time_ms > _time_target_ms) {
3459 set_has_aborted();
3460 _has_timed_out = true;
3461 statsOnly( ++_aborted_timed_out );
3462 return;
3463 }
3464
3465 // (6) Finally, we check whether there are enough completed STAB
3466 // buffers available for processing. If there are, we abort.
3467 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3468 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3469 if (_cm->verbose_low()) {
3470 gclog_or_tty->print_cr("[%d] aborting to deal with pending SATB buffers",
3471 _task_id);
3472 }
3473 // we do need to process SATB buffers, we'll abort and restart
3474 // the marking task to do so
3475 set_has_aborted();
3476 statsOnly( ++_aborted_satb );
3477 return;
3478 }
3479 }
3480
3481 void CMTask::recalculate_limits() {
3482 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3483 _words_scanned_limit = _real_words_scanned_limit;
3484
3485 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3486 _refs_reached_limit = _real_refs_reached_limit;
3487 }
3488
3489 void CMTask::decrease_limits() {
3490 // This is called when we believe that we're going to do an infrequent
3491 // operation which will increase the per byte scanned cost (i.e. move
3492 // entries to/from the global stack). It basically tries to decrease the
3493 // scanning limit so that the clock is called earlier.
3494
3495 if (_cm->verbose_medium()) {
3496 gclog_or_tty->print_cr("[%d] decreasing limits", _task_id);
3497 }
3498
3499 _words_scanned_limit = _real_words_scanned_limit -
3500 3 * words_scanned_period / 4;
3501 _refs_reached_limit = _real_refs_reached_limit -
3502 3 * refs_reached_period / 4;
3503 }
3504
3505 void CMTask::move_entries_to_global_stack() {
3506 // local array where we'll store the entries that will be popped
3507 // from the local queue
3508 oop buffer[global_stack_transfer_size];
3509
3510 int n = 0;
3511 oop obj;
3512 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3513 buffer[n] = obj;
3514 ++n;
3515 }
3516
3517 if (n > 0) {
3518 // we popped at least one entry from the local queue
3519
3520 statsOnly( ++_global_transfers_to; _local_pops += n );
3521
3522 if (!_cm->mark_stack_push(buffer, n)) {
3523 if (_cm->verbose_low()) {
3524 gclog_or_tty->print_cr("[%d] aborting due to global stack overflow",
3525 _task_id);
3526 }
3527 set_has_aborted();
3528 } else {
3529 // the transfer was successful
3530
3531 if (_cm->verbose_medium()) {
3532 gclog_or_tty->print_cr("[%d] pushed %d entries to the global stack",
3533 _task_id, n);
3534 }
3535 statsOnly( int tmp_size = _cm->mark_stack_size();
3536 if (tmp_size > _global_max_size) {
3537 _global_max_size = tmp_size;
3538 }
3539 _global_pushes += n );
3540 }
3541 }
3542
3543 // this operation was quite expensive, so decrease the limits
3544 decrease_limits();
3545 }
3546
3547 void CMTask::get_entries_from_global_stack() {
3548 // local array where we'll store the entries that will be popped
3549 // from the global stack.
3550 oop buffer[global_stack_transfer_size];
3551 int n;
3552 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3553 assert(n <= global_stack_transfer_size,
3554 "we should not pop more than the given limit");
3555 if (n > 0) {
3556 // yes, we did actually pop at least one entry
3557
3558 statsOnly( ++_global_transfers_from; _global_pops += n );
3559 if (_cm->verbose_medium()) {
3560 gclog_or_tty->print_cr("[%d] popped %d entries from the global stack",
3561 _task_id, n);
3562 }
3563 for (int i = 0; i < n; ++i) {
3564 bool success = _task_queue->push(buffer[i]);
3565 // We only call this when the local queue is empty or under a
3566 // given target limit. So, we do not expect this push to fail.
3567 assert(success, "invariant");
3568 }
3569
3570 statsOnly( int tmp_size = _task_queue->size();
3571 if (tmp_size > _local_max_size) {
3572 _local_max_size = tmp_size;
3573 }
3574 _local_pushes += n );
3575 }
3576
3577 // this operation was quite expensive, so decrease the limits
3578 decrease_limits();
3579 }
3580
3581 void CMTask::drain_local_queue(bool partially) {
3582 if (has_aborted()) return;
3583
3584 // Decide what the target size is, depending whether we're going to
3585 // drain it partially (so that other tasks can steal if they run out
3586 // of things to do) or totally (at the very end).
3587 size_t target_size;
3588 if (partially) {
3589 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3590 } else {
3591 target_size = 0;
3592 }
3593
3594 if (_task_queue->size() > target_size) {
3595 if (_cm->verbose_high()) {
3596 gclog_or_tty->print_cr("[%d] draining local queue, target size = %d",
3597 _task_id, target_size);
3598 }
3599
3600 oop obj;
3601 bool ret = _task_queue->pop_local(obj);
3602 while (ret) {
3603 statsOnly( ++_local_pops );
3604
3605 if (_cm->verbose_high()) {
3606 gclog_or_tty->print_cr("[%d] popped "PTR_FORMAT, _task_id,
3607 (void*) obj);
3608 }
3609
3610 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3611 assert(!_g1h->is_on_master_free_list(
3612 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3613
3614 scan_object(obj);
3615
3616 if (_task_queue->size() <= target_size || has_aborted()) {
3617 ret = false;
3618 } else {
3619 ret = _task_queue->pop_local(obj);
3620 }
3621 }
3622
3623 if (_cm->verbose_high()) {
3624 gclog_or_tty->print_cr("[%d] drained local queue, size = %d",
3625 _task_id, _task_queue->size());
3626 }
3627 }
3628 }
3629
3630 void CMTask::drain_global_stack(bool partially) {
3631 if (has_aborted()) return;
3632
3633 // We have a policy to drain the local queue before we attempt to
3634 // drain the global stack.
3635 assert(partially || _task_queue->size() == 0, "invariant");
3636
3637 // Decide what the target size is, depending whether we're going to
3638 // drain it partially (so that other tasks can steal if they run out
3639 // of things to do) or totally (at the very end). Notice that,
3640 // because we move entries from the global stack in chunks or
3641 // because another task might be doing the same, we might in fact
3642 // drop below the target. But, this is not a problem.
3643 size_t target_size;
3644 if (partially) {
3645 target_size = _cm->partial_mark_stack_size_target();
3646 } else {
3647 target_size = 0;
3648 }
3649
3650 if (_cm->mark_stack_size() > target_size) {
3651 if (_cm->verbose_low()) {
3652 gclog_or_tty->print_cr("[%d] draining global_stack, target size %d",
3653 _task_id, target_size);
3654 }
3655
3656 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3657 get_entries_from_global_stack();
3658 drain_local_queue(partially);
3659 }
3660
3661 if (_cm->verbose_low()) {
3662 gclog_or_tty->print_cr("[%d] drained global stack, size = %d",
3663 _task_id, _cm->mark_stack_size());
3664 }
3665 }
3666 }
3667
3668 // SATB Queue has several assumptions on whether to call the par or
3669 // non-par versions of the methods. this is why some of the code is
3670 // replicated. We should really get rid of the single-threaded version
3671 // of the code to simplify things.
3672 void CMTask::drain_satb_buffers() {
3673 if (has_aborted()) return;
3674
3675 // We set this so that the regular clock knows that we're in the
3676 // middle of draining buffers and doesn't set the abort flag when it
3677 // notices that SATB buffers are available for draining. It'd be
3678 // very counter productive if it did that. :-)
3679 _draining_satb_buffers = true;
3680
3681 CMObjectClosure oc(this);
3682 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3683 if (G1CollectedHeap::use_parallel_gc_threads()) {
3684 satb_mq_set.set_par_closure(_task_id, &oc);
3685 } else {
3686 satb_mq_set.set_closure(&oc);
3687 }
3688
3689 // This keeps claiming and applying the closure to completed buffers
3690 // until we run out of buffers or we need to abort.
3691 if (G1CollectedHeap::use_parallel_gc_threads()) {
3692 while (!has_aborted() &&
3693 satb_mq_set.par_apply_closure_to_completed_buffer(_task_id)) {
3694 if (_cm->verbose_medium()) {
3695 gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id);
3696 }
3697 statsOnly( ++_satb_buffers_processed );
3698 regular_clock_call();
3699 }
3700 } else {
3701 while (!has_aborted() &&
3702 satb_mq_set.apply_closure_to_completed_buffer()) {
3703 if (_cm->verbose_medium()) {
3704 gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id);
3705 }
3706 statsOnly( ++_satb_buffers_processed );
3707 regular_clock_call();
3708 }
3709 }
3710
3711 if (!concurrent() && !has_aborted()) {
3712 // We should only do this during remark.
3713 if (G1CollectedHeap::use_parallel_gc_threads()) {
3714 satb_mq_set.par_iterate_closure_all_threads(_task_id);
3715 } else {
3716 satb_mq_set.iterate_closure_all_threads();
3717 }
3718 }
3719
3720 _draining_satb_buffers = false;
3721
3722 assert(has_aborted() ||
3723 concurrent() ||
3724 satb_mq_set.completed_buffers_num() == 0, "invariant");
3725
3726 if (G1CollectedHeap::use_parallel_gc_threads()) {
3727 satb_mq_set.set_par_closure(_task_id, NULL);
3728 } else {
3729 satb_mq_set.set_closure(NULL);
3730 }
3731
3732 // again, this was a potentially expensive operation, decrease the
3733 // limits to get the regular clock call early
3734 decrease_limits();
3735 }
3736
3737 void CMTask::print_stats() {
3738 gclog_or_tty->print_cr("Marking Stats, task = %d, calls = %d",
3739 _task_id, _calls);
3740 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3741 _elapsed_time_ms, _termination_time_ms);
3742 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3743 _step_times_ms.num(), _step_times_ms.avg(),
3744 _step_times_ms.sd());
3745 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3746 _step_times_ms.maximum(), _step_times_ms.sum());
3747
3748 #if _MARKING_STATS_
3749 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3750 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3751 _all_clock_intervals_ms.sd());
3752 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3753 _all_clock_intervals_ms.maximum(),
3754 _all_clock_intervals_ms.sum());
3755 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
3756 _clock_due_to_scanning, _clock_due_to_marking);
3757 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
3758 _objs_scanned, _objs_found_on_bitmap);
3759 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
3760 _local_pushes, _local_pops, _local_max_size);
3761 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
3762 _global_pushes, _global_pops, _global_max_size);
3763 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
3764 _global_transfers_to,_global_transfers_from);
3765 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
3766 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
3767 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
3768 _steal_attempts, _steals);
3769 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
3770 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
3771 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3772 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
3773 _aborted_timed_out, _aborted_satb, _aborted_termination);
3774 #endif // _MARKING_STATS_
3775 }
3776
3777 /*****************************************************************************
3778
3779 The do_marking_step(time_target_ms) method is the building block
3780 of the parallel marking framework. It can be called in parallel
3781 with other invocations of do_marking_step() on different tasks
3782 (but only one per task, obviously) and concurrently with the
3783 mutator threads, or during remark, hence it eliminates the need
3784 for two versions of the code. When called during remark, it will
3785 pick up from where the task left off during the concurrent marking
3786 phase. Interestingly, tasks are also claimable during evacuation
3787 pauses too, since do_marking_step() ensures that it aborts before
3788 it needs to yield.
3789
3790 The data structures that is uses to do marking work are the
3791 following:
3792
3793 (1) Marking Bitmap. If there are gray objects that appear only
3794 on the bitmap (this happens either when dealing with an overflow
3795 or when the initial marking phase has simply marked the roots
3796 and didn't push them on the stack), then tasks claim heap
3797 regions whose bitmap they then scan to find gray objects. A
3798 global finger indicates where the end of the last claimed region
3799 is. A local finger indicates how far into the region a task has
3800 scanned. The two fingers are used to determine how to gray an
3801 object (i.e. whether simply marking it is OK, as it will be
3802 visited by a task in the future, or whether it needs to be also
3803 pushed on a stack).
3804
3805 (2) Local Queue. The local queue of the task which is accessed
3806 reasonably efficiently by the task. Other tasks can steal from
3807 it when they run out of work. Throughout the marking phase, a
3808 task attempts to keep its local queue short but not totally
3809 empty, so that entries are available for stealing by other
3810 tasks. Only when there is no more work, a task will totally
3811 drain its local queue.
3812
3813 (3) Global Mark Stack. This handles local queue overflow. During
3814 marking only sets of entries are moved between it and the local
3815 queues, as access to it requires a mutex and more fine-grain
3816 interaction with it which might cause contention. If it
3817 overflows, then the marking phase should restart and iterate
3818 over the bitmap to identify gray objects. Throughout the marking
3819 phase, tasks attempt to keep the global mark stack at a small
3820 length but not totally empty, so that entries are available for
3821 popping by other tasks. Only when there is no more work, tasks
3822 will totally drain the global mark stack.
3823
3824 (4) SATB Buffer Queue. This is where completed SATB buffers are
3825 made available. Buffers are regularly removed from this queue
3826 and scanned for roots, so that the queue doesn't get too
3827 long. During remark, all completed buffers are processed, as
3828 well as the filled in parts of any uncompleted buffers.
3829
3830 The do_marking_step() method tries to abort when the time target
3831 has been reached. There are a few other cases when the
3832 do_marking_step() method also aborts:
3833
3834 (1) When the marking phase has been aborted (after a Full GC).
3835
3836 (2) When a global overflow (on the global stack) has been
3837 triggered. Before the task aborts, it will actually sync up with
3838 the other tasks to ensure that all the marking data structures
3839 (local queues, stacks, fingers etc.) are re-initialised so that
3840 when do_marking_step() completes, the marking phase can
3841 immediately restart.
3842
3843 (3) When enough completed SATB buffers are available. The
3844 do_marking_step() method only tries to drain SATB buffers right
3845 at the beginning. So, if enough buffers are available, the
3846 marking step aborts and the SATB buffers are processed at
3847 the beginning of the next invocation.
3848
3849 (4) To yield. when we have to yield then we abort and yield
3850 right at the end of do_marking_step(). This saves us from a lot
3851 of hassle as, by yielding we might allow a Full GC. If this
3852 happens then objects will be compacted underneath our feet, the
3853 heap might shrink, etc. We save checking for this by just
3854 aborting and doing the yield right at the end.
3855
3856 From the above it follows that the do_marking_step() method should
3857 be called in a loop (or, otherwise, regularly) until it completes.
3858
3859 If a marking step completes without its has_aborted() flag being
3860 true, it means it has completed the current marking phase (and
3861 also all other marking tasks have done so and have all synced up).
3862
3863 A method called regular_clock_call() is invoked "regularly" (in
3864 sub ms intervals) throughout marking. It is this clock method that
3865 checks all the abort conditions which were mentioned above and
3866 decides when the task should abort. A work-based scheme is used to
3867 trigger this clock method: when the number of object words the
3868 marking phase has scanned or the number of references the marking
3869 phase has visited reach a given limit. Additional invocations to
3870 the method clock have been planted in a few other strategic places
3871 too. The initial reason for the clock method was to avoid calling
3872 vtime too regularly, as it is quite expensive. So, once it was in
3873 place, it was natural to piggy-back all the other conditions on it
3874 too and not constantly check them throughout the code.
3875
3876 *****************************************************************************/
3877
3878 void CMTask::do_marking_step(double time_target_ms,
3879 bool do_stealing,
3880 bool do_termination) {
3881 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
3882 assert(concurrent() == _cm->concurrent(), "they should be the same");
3883
3884 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
3885 assert(_task_queues != NULL, "invariant");
3886 assert(_task_queue != NULL, "invariant");
3887 assert(_task_queues->queue(_task_id) == _task_queue, "invariant");
3888
3889 assert(!_claimed,
3890 "only one thread should claim this task at any one time");
3891
3892 // OK, this doesn't safeguard again all possible scenarios, as it is
3893 // possible for two threads to set the _claimed flag at the same
3894 // time. But it is only for debugging purposes anyway and it will
3895 // catch most problems.
3896 _claimed = true;
3897
3898 _start_time_ms = os::elapsedVTime() * 1000.0;
3899 statsOnly( _interval_start_time_ms = _start_time_ms );
3900
3901 double diff_prediction_ms =
3902 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
3903 _time_target_ms = time_target_ms - diff_prediction_ms;
3904
3905 // set up the variables that are used in the work-based scheme to
3906 // call the regular clock method
3907 _words_scanned = 0;
3908 _refs_reached = 0;
3909 recalculate_limits();
3910
3911 // clear all flags
3912 clear_has_aborted();
3913 _has_timed_out = false;
3914 _draining_satb_buffers = false;
3915
3916 ++_calls;
3917
3918 if (_cm->verbose_low()) {
3919 gclog_or_tty->print_cr("[%d] >>>>>>>>>> START, call = %d, "
3920 "target = %1.2lfms >>>>>>>>>>",
3921 _task_id, _calls, _time_target_ms);
3922 }
3923
3924 // Set up the bitmap and oop closures. Anything that uses them is
3925 // eventually called from this method, so it is OK to allocate these
3926 // statically.
3927 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
3928 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
3929 set_cm_oop_closure(&cm_oop_closure);
3930
3931 if (_cm->has_overflown()) {
3932 // This can happen if the mark stack overflows during a GC pause
3933 // and this task, after a yield point, restarts. We have to abort
3934 // as we need to get into the overflow protocol which happens
3935 // right at the end of this task.
3936 set_has_aborted();
3937 }
3938
3939 // First drain any available SATB buffers. After this, we will not
3940 // look at SATB buffers before the next invocation of this method.
3941 // If enough completed SATB buffers are queued up, the regular clock
3942 // will abort this task so that it restarts.
3943 drain_satb_buffers();
3944 // ...then partially drain the local queue and the global stack
3945 drain_local_queue(true);
3946 drain_global_stack(true);
3947
3948 do {
3949 if (!has_aborted() && _curr_region != NULL) {
3950 // This means that we're already holding on to a region.
3951 assert(_finger != NULL, "if region is not NULL, then the finger "
3952 "should not be NULL either");
3953
3954 // We might have restarted this task after an evacuation pause
3955 // which might have evacuated the region we're holding on to
3956 // underneath our feet. Let's read its limit again to make sure
3957 // that we do not iterate over a region of the heap that
3958 // contains garbage (update_region_limit() will also move
3959 // _finger to the start of the region if it is found empty).
3960 update_region_limit();
3961 // We will start from _finger not from the start of the region,
3962 // as we might be restarting this task after aborting half-way
3963 // through scanning this region. In this case, _finger points to
3964 // the address where we last found a marked object. If this is a
3965 // fresh region, _finger points to start().
3966 MemRegion mr = MemRegion(_finger, _region_limit);
3967
3968 if (_cm->verbose_low()) {
3969 gclog_or_tty->print_cr("[%d] we're scanning part "
3970 "["PTR_FORMAT", "PTR_FORMAT") "
3971 "of region "PTR_FORMAT,
3972 _task_id, _finger, _region_limit, _curr_region);
3973 }
3974
3975 // Let's iterate over the bitmap of the part of the
3976 // region that is left.
3977 if (mr.is_empty() || _nextMarkBitMap->iterate(&bitmap_closure, mr)) {
3978 // We successfully completed iterating over the region. Now,
3979 // let's give up the region.
3980 giveup_current_region();
3981 regular_clock_call();
3982 } else {
3983 assert(has_aborted(), "currently the only way to do so");
3984 // The only way to abort the bitmap iteration is to return
3985 // false from the do_bit() method. However, inside the
3986 // do_bit() method we move the _finger to point to the
3987 // object currently being looked at. So, if we bail out, we
3988 // have definitely set _finger to something non-null.
3989 assert(_finger != NULL, "invariant");
3990
3991 // Region iteration was actually aborted. So now _finger
3992 // points to the address of the object we last scanned. If we
3993 // leave it there, when we restart this task, we will rescan
3994 // the object. It is easy to avoid this. We move the finger by
3995 // enough to point to the next possible object header (the
3996 // bitmap knows by how much we need to move it as it knows its
3997 // granularity).
3998 assert(_finger < _region_limit, "invariant");
3999 HeapWord* new_finger = _nextMarkBitMap->nextWord(_finger);
4000 // Check if bitmap iteration was aborted while scanning the last object
4001 if (new_finger >= _region_limit) {
4002 giveup_current_region();
4003 } else {
4004 move_finger_to(new_finger);
4005 }
4006 }
4007 }
4008 // At this point we have either completed iterating over the
4009 // region we were holding on to, or we have aborted.
4010
4011 // We then partially drain the local queue and the global stack.
4012 // (Do we really need this?)
4013 drain_local_queue(true);
4014 drain_global_stack(true);
4015
4016 // Read the note on the claim_region() method on why it might
4017 // return NULL with potentially more regions available for
4018 // claiming and why we have to check out_of_regions() to determine
4019 // whether we're done or not.
4020 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4021 // We are going to try to claim a new region. We should have
4022 // given up on the previous one.
4023 // Separated the asserts so that we know which one fires.
4024 assert(_curr_region == NULL, "invariant");
4025 assert(_finger == NULL, "invariant");
4026 assert(_region_limit == NULL, "invariant");
4027 if (_cm->verbose_low()) {
4028 gclog_or_tty->print_cr("[%d] trying to claim a new region", _task_id);
4029 }
4030 HeapRegion* claimed_region = _cm->claim_region(_task_id);
4031 if (claimed_region != NULL) {
4032 // Yes, we managed to claim one
4033 statsOnly( ++_regions_claimed );
4034
4035 if (_cm->verbose_low()) {
4036 gclog_or_tty->print_cr("[%d] we successfully claimed "
4037 "region "PTR_FORMAT,
4038 _task_id, claimed_region);
4039 }
4040
4041 setup_for_region(claimed_region);
4042 assert(_curr_region == claimed_region, "invariant");
4043 }
4044 // It is important to call the regular clock here. It might take
4045 // a while to claim a region if, for example, we hit a large
4046 // block of empty regions. So we need to call the regular clock
4047 // method once round the loop to make sure it's called
4048 // frequently enough.
4049 regular_clock_call();
4050 }
4051
4052 if (!has_aborted() && _curr_region == NULL) {
4053 assert(_cm->out_of_regions(),
4054 "at this point we should be out of regions");
4055 }
4056 } while ( _curr_region != NULL && !has_aborted());
4057
4058 if (!has_aborted()) {
4059 // We cannot check whether the global stack is empty, since other
4060 // tasks might be pushing objects to it concurrently.
4061 assert(_cm->out_of_regions(),
4062 "at this point we should be out of regions");
4063
4064 if (_cm->verbose_low()) {
4065 gclog_or_tty->print_cr("[%d] all regions claimed", _task_id);
4066 }
4067
4068 // Try to reduce the number of available SATB buffers so that
4069 // remark has less work to do.
4070 drain_satb_buffers();
4071 }
4072
4073 // Since we've done everything else, we can now totally drain the
4074 // local queue and global stack.
4075 drain_local_queue(false);
4076 drain_global_stack(false);
4077
4078 // Attempt at work stealing from other task's queues.
4079 if (do_stealing && !has_aborted()) {
4080 // We have not aborted. This means that we have finished all that
4081 // we could. Let's try to do some stealing...
4082
4083 // We cannot check whether the global stack is empty, since other
4084 // tasks might be pushing objects to it concurrently.
4085 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4086 "only way to reach here");
4087
4088 if (_cm->verbose_low()) {
4089 gclog_or_tty->print_cr("[%d] starting to steal", _task_id);
4090 }
4091
4092 while (!has_aborted()) {
4093 oop obj;
4094 statsOnly( ++_steal_attempts );
4095
4096 if (_cm->try_stealing(_task_id, &_hash_seed, obj)) {
4097 if (_cm->verbose_medium()) {
4098 gclog_or_tty->print_cr("[%d] stolen "PTR_FORMAT" successfully",
4099 _task_id, (void*) obj);
4100 }
4101
4102 statsOnly( ++_steals );
4103
4104 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4105 "any stolen object should be marked");
4106 scan_object(obj);
4107
4108 // And since we're towards the end, let's totally drain the
4109 // local queue and global stack.
4110 drain_local_queue(false);
4111 drain_global_stack(false);
4112 } else {
4113 break;
4114 }
4115 }
4116 }
4117
4118 // If we are about to wrap up and go into termination, check if we
4119 // should raise the overflow flag.
4120 if (do_termination && !has_aborted()) {
4121 if (_cm->force_overflow()->should_force()) {
4122 _cm->set_has_overflown();
4123 regular_clock_call();
4124 }
4125 }
4126
4127 // We still haven't aborted. Now, let's try to get into the
4128 // termination protocol.
4129 if (do_termination && !has_aborted()) {
4130 // We cannot check whether the global stack is empty, since other
4131 // tasks might be concurrently pushing objects on it.
4132 // Separated the asserts so that we know which one fires.
4133 assert(_cm->out_of_regions(), "only way to reach here");
4134 assert(_task_queue->size() == 0, "only way to reach here");
4135
4136 if (_cm->verbose_low()) {
4137 gclog_or_tty->print_cr("[%d] starting termination protocol", _task_id);
4138 }
4139
4140 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4141 // The CMTask class also extends the TerminatorTerminator class,
4142 // hence its should_exit_termination() method will also decide
4143 // whether to exit the termination protocol or not.
4144 bool finished = _cm->terminator()->offer_termination(this);
4145 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4146 _termination_time_ms +=
4147 termination_end_time_ms - _termination_start_time_ms;
4148
4149 if (finished) {
4150 // We're all done.
4151
4152 if (_task_id == 0) {
4153 // let's allow task 0 to do this
4154 if (concurrent()) {
4155 assert(_cm->concurrent_marking_in_progress(), "invariant");
4156 // we need to set this to false before the next
4157 // safepoint. This way we ensure that the marking phase
4158 // doesn't observe any more heap expansions.
4159 _cm->clear_concurrent_marking_in_progress();
4160 }
4161 }
4162
4163 // We can now guarantee that the global stack is empty, since
4164 // all other tasks have finished. We separated the guarantees so
4165 // that, if a condition is false, we can immediately find out
4166 // which one.
4167 guarantee(_cm->out_of_regions(), "only way to reach here");
4168 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4169 guarantee(_task_queue->size() == 0, "only way to reach here");
4170 guarantee(!_cm->has_overflown(), "only way to reach here");
4171 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4172
4173 if (_cm->verbose_low()) {
4174 gclog_or_tty->print_cr("[%d] all tasks terminated", _task_id);
4175 }
4176 } else {
4177 // Apparently there's more work to do. Let's abort this task. It
4178 // will restart it and we can hopefully find more things to do.
4179
4180 if (_cm->verbose_low()) {
4181 gclog_or_tty->print_cr("[%d] apparently there is more work to do",
4182 _task_id);
4183 }
4184
4185 set_has_aborted();
4186 statsOnly( ++_aborted_termination );
4187 }
4188 }
4189
4190 // Mainly for debugging purposes to make sure that a pointer to the
4191 // closure which was statically allocated in this frame doesn't
4192 // escape it by accident.
4193 set_cm_oop_closure(NULL);
4194 double end_time_ms = os::elapsedVTime() * 1000.0;
4195 double elapsed_time_ms = end_time_ms - _start_time_ms;
4196 // Update the step history.
4197 _step_times_ms.add(elapsed_time_ms);
4198
4199 if (has_aborted()) {
4200 // The task was aborted for some reason.
4201
4202 statsOnly( ++_aborted );
4203
4204 if (_has_timed_out) {
4205 double diff_ms = elapsed_time_ms - _time_target_ms;
4206 // Keep statistics of how well we did with respect to hitting
4207 // our target only if we actually timed out (if we aborted for
4208 // other reasons, then the results might get skewed).
4209 _marking_step_diffs_ms.add(diff_ms);
4210 }
4211
4212 if (_cm->has_overflown()) {
4213 // This is the interesting one. We aborted because a global
4214 // overflow was raised. This means we have to restart the
4215 // marking phase and start iterating over regions. However, in
4216 // order to do this we have to make sure that all tasks stop
4217 // what they are doing and re-initialise in a safe manner. We
4218 // will achieve this with the use of two barrier sync points.
4219
4220 if (_cm->verbose_low()) {
4221 gclog_or_tty->print_cr("[%d] detected overflow", _task_id);
4222 }
4223
4224 _cm->enter_first_sync_barrier(_task_id);
4225 // When we exit this sync barrier we know that all tasks have
4226 // stopped doing marking work. So, it's now safe to
4227 // re-initialise our data structures. At the end of this method,
4228 // task 0 will clear the global data structures.
4229
4230 statsOnly( ++_aborted_overflow );
4231
4232 // We clear the local state of this task...
4233 clear_region_fields();
4234
4235 // ...and enter the second barrier.
4236 _cm->enter_second_sync_barrier(_task_id);
4237 // At this point everything has bee re-initialised and we're
4238 // ready to restart.
4239 }
4240
4241 if (_cm->verbose_low()) {
4242 gclog_or_tty->print_cr("[%d] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4243 "elapsed = %1.2lfms <<<<<<<<<<",
4244 _task_id, _time_target_ms, elapsed_time_ms);
4245 if (_cm->has_aborted()) {
4246 gclog_or_tty->print_cr("[%d] ========== MARKING ABORTED ==========",
4247 _task_id);
4248 }
4249 }
4250 } else {
4251 if (_cm->verbose_low()) {
4252 gclog_or_tty->print_cr("[%d] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4253 "elapsed = %1.2lfms <<<<<<<<<<",
4254 _task_id, _time_target_ms, elapsed_time_ms);
4255 }
4256 }
4257
4258 _claimed = false;
4259 }
4260
4261 CMTask::CMTask(int task_id,
4262 ConcurrentMark* cm,
4263 size_t* marked_bytes,
4264 BitMap* card_bm,
4265 CMTaskQueue* task_queue,
4266 CMTaskQueueSet* task_queues)
4267 : _g1h(G1CollectedHeap::heap()),
4268 _task_id(task_id), _cm(cm),
4269 _claimed(false),
4270 _nextMarkBitMap(NULL), _hash_seed(17),
4271 _task_queue(task_queue),
4272 _task_queues(task_queues),
4273 _cm_oop_closure(NULL),
4274 _marked_bytes_array(marked_bytes),
4275 _card_bm(card_bm) {
4276 guarantee(task_queue != NULL, "invariant");
4277 guarantee(task_queues != NULL, "invariant");
4278
4279 statsOnly( _clock_due_to_scanning = 0;
4280 _clock_due_to_marking = 0 );
4281
4282 _marking_step_diffs_ms.add(0.5);
4283 }
4284
4285 // These are formatting macros that are used below to ensure
4286 // consistent formatting. The *_H_* versions are used to format the
4287 // header for a particular value and they should be kept consistent
4288 // with the corresponding macro. Also note that most of the macros add
4289 // the necessary white space (as a prefix) which makes them a bit
4290 // easier to compose.
4291
4292 // All the output lines are prefixed with this string to be able to
4293 // identify them easily in a large log file.
4294 #define G1PPRL_LINE_PREFIX "###"
4295
4296 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4297 #ifdef _LP64
4298 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4299 #else // _LP64
4300 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4301 #endif // _LP64
4302
4303 // For per-region info
4304 #define G1PPRL_TYPE_FORMAT " %-4s"
4305 #define G1PPRL_TYPE_H_FORMAT " %4s"
4306 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4307 #define G1PPRL_BYTE_H_FORMAT " %9s"
4308 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4309 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4310
4311 // For summary info
4312 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4313 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4314 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4315 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4316
4317 G1PrintRegionLivenessInfoClosure::
4318 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4319 : _out(out),
4320 _total_used_bytes(0), _total_capacity_bytes(0),
4321 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4322 _hum_used_bytes(0), _hum_capacity_bytes(0),
4323 _hum_prev_live_bytes(0), _hum_next_live_bytes(0) {
4324 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4325 MemRegion g1_committed = g1h->g1_committed();
4326 MemRegion g1_reserved = g1h->g1_reserved();
4327 double now = os::elapsedTime();
4328
4329 // Print the header of the output.
4330 _out->cr();
4331 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4332 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4333 G1PPRL_SUM_ADDR_FORMAT("committed")
4334 G1PPRL_SUM_ADDR_FORMAT("reserved")
4335 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4336 g1_committed.start(), g1_committed.end(),
4337 g1_reserved.start(), g1_reserved.end(),
4338 HeapRegion::GrainBytes);
4339 _out->print_cr(G1PPRL_LINE_PREFIX);
4340 _out->print_cr(G1PPRL_LINE_PREFIX
4341 G1PPRL_TYPE_H_FORMAT
4342 G1PPRL_ADDR_BASE_H_FORMAT
4343 G1PPRL_BYTE_H_FORMAT
4344 G1PPRL_BYTE_H_FORMAT
4345 G1PPRL_BYTE_H_FORMAT
4346 G1PPRL_DOUBLE_H_FORMAT,
4347 "type", "address-range",
4348 "used", "prev-live", "next-live", "gc-eff");
4349 _out->print_cr(G1PPRL_LINE_PREFIX
4350 G1PPRL_TYPE_H_FORMAT
4351 G1PPRL_ADDR_BASE_H_FORMAT
4352 G1PPRL_BYTE_H_FORMAT
4353 G1PPRL_BYTE_H_FORMAT
4354 G1PPRL_BYTE_H_FORMAT
4355 G1PPRL_DOUBLE_H_FORMAT,
4356 "", "",
4357 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)");
4358 }
4359
4360 // It takes as a parameter a reference to one of the _hum_* fields, it
4361 // deduces the corresponding value for a region in a humongous region
4362 // series (either the region size, or what's left if the _hum_* field
4363 // is < the region size), and updates the _hum_* field accordingly.
4364 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4365 size_t bytes = 0;
4366 // The > 0 check is to deal with the prev and next live bytes which
4367 // could be 0.
4368 if (*hum_bytes > 0) {
4369 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4370 *hum_bytes -= bytes;
4371 }
4372 return bytes;
4373 }
4374
4375 // It deduces the values for a region in a humongous region series
4376 // from the _hum_* fields and updates those accordingly. It assumes
4377 // that that _hum_* fields have already been set up from the "starts
4378 // humongous" region and we visit the regions in address order.
4379 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4380 size_t* capacity_bytes,
4381 size_t* prev_live_bytes,
4382 size_t* next_live_bytes) {
4383 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4384 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4385 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4386 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4387 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4388 }
4389
4390 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4391 const char* type = "";
4392 HeapWord* bottom = r->bottom();
4393 HeapWord* end = r->end();
4394 size_t capacity_bytes = r->capacity();
4395 size_t used_bytes = r->used();
4396 size_t prev_live_bytes = r->live_bytes();
4397 size_t next_live_bytes = r->next_live_bytes();
4398 double gc_eff = r->gc_efficiency();
4399 if (r->used() == 0) {
4400 type = "FREE";
4401 } else if (r->is_survivor()) {
4402 type = "SURV";
4403 } else if (r->is_young()) {
4404 type = "EDEN";
4405 } else if (r->startsHumongous()) {
4406 type = "HUMS";
4407
4408 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4409 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4410 "they should have been zeroed after the last time we used them");
4411 // Set up the _hum_* fields.
4412 _hum_capacity_bytes = capacity_bytes;
4413 _hum_used_bytes = used_bytes;
4414 _hum_prev_live_bytes = prev_live_bytes;
4415 _hum_next_live_bytes = next_live_bytes;
4416 get_hum_bytes(&used_bytes, &capacity_bytes,
4417 &prev_live_bytes, &next_live_bytes);
4418 end = bottom + HeapRegion::GrainWords;
4419 } else if (r->continuesHumongous()) {
4420 type = "HUMC";
4421 get_hum_bytes(&used_bytes, &capacity_bytes,
4422 &prev_live_bytes, &next_live_bytes);
4423 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4424 } else {
4425 type = "OLD";
4426 }
4427
4428 _total_used_bytes += used_bytes;
4429 _total_capacity_bytes += capacity_bytes;
4430 _total_prev_live_bytes += prev_live_bytes;
4431 _total_next_live_bytes += next_live_bytes;
4432
4433 // Print a line for this particular region.
4434 _out->print_cr(G1PPRL_LINE_PREFIX
4435 G1PPRL_TYPE_FORMAT
4436 G1PPRL_ADDR_BASE_FORMAT
4437 G1PPRL_BYTE_FORMAT
4438 G1PPRL_BYTE_FORMAT
4439 G1PPRL_BYTE_FORMAT
4440 G1PPRL_DOUBLE_FORMAT,
4441 type, bottom, end,
4442 used_bytes, prev_live_bytes, next_live_bytes, gc_eff);
4443
4444 return false;
4445 }
4446
4447 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4448 // Print the footer of the output.
4449 _out->print_cr(G1PPRL_LINE_PREFIX);
4450 _out->print_cr(G1PPRL_LINE_PREFIX
4451 " SUMMARY"
4452 G1PPRL_SUM_MB_FORMAT("capacity")
4453 G1PPRL_SUM_MB_PERC_FORMAT("used")
4454 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4455 G1PPRL_SUM_MB_PERC_FORMAT("next-live"),
4456 bytes_to_mb(_total_capacity_bytes),
4457 bytes_to_mb(_total_used_bytes),
4458 perc(_total_used_bytes, _total_capacity_bytes),
4459 bytes_to_mb(_total_prev_live_bytes),
4460 perc(_total_prev_live_bytes, _total_capacity_bytes),
4461 bytes_to_mb(_total_next_live_bytes),
4462 perc(_total_next_live_bytes, _total_capacity_bytes));
4463 _out->cr();
4464 }