1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/metadataOnStackMark.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/g1/concurrentMark.inline.hpp"
30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
32 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
33 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
34 #include "gc_implementation/g1/g1Log.hpp"
35 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
36 #include "gc_implementation/g1/g1RemSet.hpp"
37 #include "gc_implementation/g1/g1StringDedup.hpp"
38 #include "gc_implementation/g1/heapRegion.inline.hpp"
39 #include "gc_implementation/g1/heapRegionManager.inline.hpp"
40 #include "gc_implementation/g1/heapRegionRemSet.hpp"
41 #include "gc_implementation/g1/heapRegionSet.inline.hpp"
42 #include "gc_implementation/shared/vmGCOperations.hpp"
43 #include "gc_implementation/shared/gcTimer.hpp"
44 #include "gc_implementation/shared/gcTrace.hpp"
45 #include "gc_implementation/shared/gcTraceTime.hpp"
46 #include "memory/allocation.hpp"
47 #include "memory/genOopClosures.inline.hpp"
48 #include "memory/referencePolicy.hpp"
49 #include "memory/resourceArea.hpp"
50 #include "memory/strongRootsScope.hpp"
51 #include "oops/oop.inline.hpp"
52 #include "runtime/handles.inline.hpp"
53 #include "runtime/java.hpp"
54 #include "runtime/atomic.inline.hpp"
55 #include "runtime/prefetch.inline.hpp"
56 #include "services/memTracker.hpp"
57 #include "utilities/taskqueue.inline.hpp"
58
59 // Concurrent marking bit map wrapper
60
61 CMBitMapRO::CMBitMapRO(int shifter) :
62 _bm(),
63 _shifter(shifter) {
64 _bmStartWord = 0;
65 _bmWordSize = 0;
66 }
67
68 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
69 const HeapWord* limit) const {
70 // First we must round addr *up* to a possible object boundary.
71 addr = (HeapWord*)align_size_up((intptr_t)addr,
72 HeapWordSize << _shifter);
73 size_t addrOffset = heapWordToOffset(addr);
74 if (limit == NULL) {
75 limit = _bmStartWord + _bmWordSize;
76 }
77 size_t limitOffset = heapWordToOffset(limit);
78 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
79 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
80 assert(nextAddr >= addr, "get_next_one postcondition");
81 assert(nextAddr == limit || isMarked(nextAddr),
82 "get_next_one postcondition");
83 return nextAddr;
84 }
85
86 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
87 const HeapWord* limit) const {
88 size_t addrOffset = heapWordToOffset(addr);
89 if (limit == NULL) {
90 limit = _bmStartWord + _bmWordSize;
91 }
92 size_t limitOffset = heapWordToOffset(limit);
93 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
94 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
95 assert(nextAddr >= addr, "get_next_one postcondition");
96 assert(nextAddr == limit || !isMarked(nextAddr),
97 "get_next_one postcondition");
98 return nextAddr;
99 }
100
101 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
102 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
103 return (int) (diff >> _shifter);
104 }
105
106 #ifndef PRODUCT
107 bool CMBitMapRO::covers(MemRegion heap_rs) const {
108 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
109 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
110 "size inconsistency");
111 return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
112 _bmWordSize == heap_rs.word_size();
113 }
114 #endif
115
116 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
117 _bm.print_on_error(st, prefix);
118 }
119
120 size_t CMBitMap::compute_size(size_t heap_size) {
121 return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
122 }
123
124 size_t CMBitMap::mark_distance() {
125 return MinObjAlignmentInBytes * BitsPerByte;
126 }
127
128 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
129 _bmStartWord = heap.start();
130 _bmWordSize = heap.word_size();
131
132 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
133 _bm.set_size(_bmWordSize >> _shifter);
134
135 storage->set_mapping_changed_listener(&_listener);
136 }
137
138 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
139 if (zero_filled) {
140 return;
141 }
142 // We need to clear the bitmap on commit, removing any existing information.
143 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
144 _bm->clearRange(mr);
145 }
146
147 // Closure used for clearing the given mark bitmap.
148 class ClearBitmapHRClosure : public HeapRegionClosure {
149 private:
150 ConcurrentMark* _cm;
151 CMBitMap* _bitmap;
152 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration.
153 public:
154 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
155 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
156 }
157
158 virtual bool doHeapRegion(HeapRegion* r) {
159 size_t const chunk_size_in_words = M / HeapWordSize;
160
161 HeapWord* cur = r->bottom();
162 HeapWord* const end = r->end();
163
164 while (cur < end) {
165 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
166 _bitmap->clearRange(mr);
167
168 cur += chunk_size_in_words;
169
170 // Abort iteration if after yielding the marking has been aborted.
171 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
172 return true;
173 }
174 // Repeat the asserts from before the start of the closure. We will do them
175 // as asserts here to minimize their overhead on the product. However, we
176 // will have them as guarantees at the beginning / end of the bitmap
177 // clearing to get some checking in the product.
178 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
179 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
180 }
181
182 return false;
183 }
184 };
185
186 class ParClearNextMarkBitmapTask : public AbstractGangTask {
187 ClearBitmapHRClosure* _cl;
188 HeapRegionClaimer _hrclaimer;
189 bool _suspendible; // If the task is suspendible, workers must join the STS.
190
191 public:
192 ParClearNextMarkBitmapTask(ClearBitmapHRClosure *cl, uint n_workers, bool suspendible) :
193 _cl(cl), _suspendible(suspendible), AbstractGangTask("Parallel Clear Bitmap Task"), _hrclaimer(n_workers) {}
194
195 void work(uint worker_id) {
196 if (_suspendible) {
197 SuspendibleThreadSet::join();
198 }
199 G1CollectedHeap::heap()->heap_region_par_iterate(_cl, worker_id, &_hrclaimer, true);
200 if (_suspendible) {
201 SuspendibleThreadSet::leave();
202 }
203 }
204 };
205
206 void CMBitMap::clearAll() {
207 G1CollectedHeap* g1h = G1CollectedHeap::heap();
208 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
209 uint n_workers = g1h->workers()->active_workers();
210 ParClearNextMarkBitmapTask task(&cl, n_workers, false);
211 g1h->workers()->run_task(&task);
212 guarantee(cl.complete(), "Must have completed iteration.");
213 return;
214 }
215
216 void CMBitMap::markRange(MemRegion mr) {
217 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
218 assert(!mr.is_empty(), "unexpected empty region");
219 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
220 ((HeapWord *) mr.end())),
221 "markRange memory region end is not card aligned");
222 // convert address range into offset range
223 _bm.at_put_range(heapWordToOffset(mr.start()),
224 heapWordToOffset(mr.end()), true);
225 }
226
227 void CMBitMap::clearRange(MemRegion mr) {
228 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
229 assert(!mr.is_empty(), "unexpected empty region");
230 // convert address range into offset range
231 _bm.at_put_range(heapWordToOffset(mr.start()),
232 heapWordToOffset(mr.end()), false);
233 }
234
235 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
236 HeapWord* end_addr) {
237 HeapWord* start = getNextMarkedWordAddress(addr);
238 start = MIN2(start, end_addr);
239 HeapWord* end = getNextUnmarkedWordAddress(start);
240 end = MIN2(end, end_addr);
241 assert(start <= end, "Consistency check");
242 MemRegion mr(start, end);
243 if (!mr.is_empty()) {
244 clearRange(mr);
245 }
246 return mr;
247 }
248
249 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
250 _base(NULL), _cm(cm)
251 #ifdef ASSERT
252 , _drain_in_progress(false)
253 , _drain_in_progress_yields(false)
254 #endif
255 {}
256
257 bool CMMarkStack::allocate(size_t capacity) {
258 // allocate a stack of the requisite depth
259 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
260 if (!rs.is_reserved()) {
261 warning("ConcurrentMark MarkStack allocation failure");
262 return false;
263 }
264 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
265 if (!_virtual_space.initialize(rs, rs.size())) {
266 warning("ConcurrentMark MarkStack backing store failure");
267 // Release the virtual memory reserved for the marking stack
268 rs.release();
269 return false;
270 }
271 assert(_virtual_space.committed_size() == rs.size(),
272 "Didn't reserve backing store for all of ConcurrentMark stack?");
273 _base = (oop*) _virtual_space.low();
274 setEmpty();
275 _capacity = (jint) capacity;
276 _saved_index = -1;
277 _should_expand = false;
278 NOT_PRODUCT(_max_depth = 0);
279 return true;
280 }
281
282 void CMMarkStack::expand() {
283 // Called, during remark, if we've overflown the marking stack during marking.
284 assert(isEmpty(), "stack should been emptied while handling overflow");
285 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
286 // Clear expansion flag
287 _should_expand = false;
288 if (_capacity == (jint) MarkStackSizeMax) {
289 if (PrintGCDetails && Verbose) {
290 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
291 }
292 return;
293 }
294 // Double capacity if possible
295 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
296 // Do not give up existing stack until we have managed to
297 // get the double capacity that we desired.
298 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
299 sizeof(oop)));
300 if (rs.is_reserved()) {
301 // Release the backing store associated with old stack
302 _virtual_space.release();
303 // Reinitialize virtual space for new stack
304 if (!_virtual_space.initialize(rs, rs.size())) {
305 fatal("Not enough swap for expanded marking stack capacity");
306 }
307 _base = (oop*)(_virtual_space.low());
308 _index = 0;
309 _capacity = new_capacity;
310 } else {
311 if (PrintGCDetails && Verbose) {
312 // Failed to double capacity, continue;
313 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
314 SIZE_FORMAT"K to " SIZE_FORMAT"K",
315 _capacity / K, new_capacity / K);
316 }
317 }
318 }
319
320 void CMMarkStack::set_should_expand() {
321 // If we're resetting the marking state because of an
322 // marking stack overflow, record that we should, if
323 // possible, expand the stack.
324 _should_expand = _cm->has_overflown();
325 }
326
327 CMMarkStack::~CMMarkStack() {
328 if (_base != NULL) {
329 _base = NULL;
330 _virtual_space.release();
331 }
332 }
333
334 void CMMarkStack::par_push(oop ptr) {
335 while (true) {
336 if (isFull()) {
337 _overflow = true;
338 return;
339 }
340 // Otherwise...
341 jint index = _index;
342 jint next_index = index+1;
343 jint res = Atomic::cmpxchg(next_index, &_index, index);
344 if (res == index) {
345 _base[index] = ptr;
346 // Note that we don't maintain this atomically. We could, but it
347 // doesn't seem necessary.
348 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
349 return;
350 }
351 // Otherwise, we need to try again.
352 }
353 }
354
355 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
356 while (true) {
357 if (isFull()) {
358 _overflow = true;
359 return;
360 }
361 // Otherwise...
362 jint index = _index;
363 jint next_index = index + n;
364 if (next_index > _capacity) {
365 _overflow = true;
366 return;
367 }
368 jint res = Atomic::cmpxchg(next_index, &_index, index);
369 if (res == index) {
370 for (int i = 0; i < n; i++) {
371 int ind = index + i;
372 assert(ind < _capacity, "By overflow test above.");
373 _base[ind] = ptr_arr[i];
374 }
375 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
376 return;
377 }
378 // Otherwise, we need to try again.
379 }
380 }
381
382 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
383 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
384 jint start = _index;
385 jint next_index = start + n;
386 if (next_index > _capacity) {
387 _overflow = true;
388 return;
389 }
390 // Otherwise.
391 _index = next_index;
392 for (int i = 0; i < n; i++) {
393 int ind = start + i;
394 assert(ind < _capacity, "By overflow test above.");
395 _base[ind] = ptr_arr[i];
396 }
397 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
398 }
399
400 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
401 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
402 jint index = _index;
403 if (index == 0) {
404 *n = 0;
405 return false;
406 } else {
407 int k = MIN2(max, index);
408 jint new_ind = index - k;
409 for (int j = 0; j < k; j++) {
410 ptr_arr[j] = _base[new_ind + j];
411 }
412 _index = new_ind;
413 *n = k;
414 return true;
415 }
416 }
417
418 template<class OopClosureClass>
419 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
420 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
421 || SafepointSynchronize::is_at_safepoint(),
422 "Drain recursion must be yield-safe.");
423 bool res = true;
424 debug_only(_drain_in_progress = true);
425 debug_only(_drain_in_progress_yields = yield_after);
426 while (!isEmpty()) {
427 oop newOop = pop();
428 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
429 assert(newOop->is_oop(), "Expected an oop");
430 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
431 "only grey objects on this stack");
432 newOop->oop_iterate(cl);
433 if (yield_after && _cm->do_yield_check()) {
434 res = false;
435 break;
436 }
437 }
438 debug_only(_drain_in_progress = false);
439 return res;
440 }
441
442 void CMMarkStack::note_start_of_gc() {
443 assert(_saved_index == -1,
444 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
445 _saved_index = _index;
446 }
447
448 void CMMarkStack::note_end_of_gc() {
449 // This is intentionally a guarantee, instead of an assert. If we
450 // accidentally add something to the mark stack during GC, it
451 // will be a correctness issue so it's better if we crash. we'll
452 // only check this once per GC anyway, so it won't be a performance
453 // issue in any way.
454 guarantee(_saved_index == _index,
455 err_msg("saved index: %d index: %d", _saved_index, _index));
456 _saved_index = -1;
457 }
458
459 void CMMarkStack::oops_do(OopClosure* f) {
460 assert(_saved_index == _index,
461 err_msg("saved index: %d index: %d", _saved_index, _index));
462 for (int i = 0; i < _index; i += 1) {
463 f->do_oop(&_base[i]);
464 }
465 }
466
467 CMRootRegions::CMRootRegions() :
468 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
469 _should_abort(false), _next_survivor(NULL) { }
470
471 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
472 _young_list = g1h->young_list();
473 _cm = cm;
474 }
475
476 void CMRootRegions::prepare_for_scan() {
477 assert(!scan_in_progress(), "pre-condition");
478
479 // Currently, only survivors can be root regions.
480 assert(_next_survivor == NULL, "pre-condition");
481 _next_survivor = _young_list->first_survivor_region();
482 _scan_in_progress = (_next_survivor != NULL);
483 _should_abort = false;
484 }
485
486 HeapRegion* CMRootRegions::claim_next() {
487 if (_should_abort) {
488 // If someone has set the should_abort flag, we return NULL to
489 // force the caller to bail out of their loop.
490 return NULL;
491 }
492
493 // Currently, only survivors can be root regions.
494 HeapRegion* res = _next_survivor;
495 if (res != NULL) {
496 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
497 // Read it again in case it changed while we were waiting for the lock.
498 res = _next_survivor;
499 if (res != NULL) {
500 if (res == _young_list->last_survivor_region()) {
501 // We just claimed the last survivor so store NULL to indicate
502 // that we're done.
503 _next_survivor = NULL;
504 } else {
505 _next_survivor = res->get_next_young_region();
506 }
507 } else {
508 // Someone else claimed the last survivor while we were trying
509 // to take the lock so nothing else to do.
510 }
511 }
512 assert(res == NULL || res->is_survivor(), "post-condition");
513
514 return res;
515 }
516
517 void CMRootRegions::scan_finished() {
518 assert(scan_in_progress(), "pre-condition");
519
520 // Currently, only survivors can be root regions.
521 if (!_should_abort) {
522 assert(_next_survivor == NULL, "we should have claimed all survivors");
523 }
524 _next_survivor = NULL;
525
526 {
527 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
528 _scan_in_progress = false;
529 RootRegionScan_lock->notify_all();
530 }
531 }
532
533 bool CMRootRegions::wait_until_scan_finished() {
534 if (!scan_in_progress()) return false;
535
536 {
537 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
538 while (scan_in_progress()) {
539 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
540 }
541 }
542 return true;
543 }
544
545 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
546 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
547 #endif // _MSC_VER
548
549 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
550 return MAX2((n_par_threads + 2) / 4, 1U);
551 }
552
553 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
554 _g1h(g1h),
555 _markBitMap1(),
556 _markBitMap2(),
557 _parallel_marking_threads(0),
558 _max_parallel_marking_threads(0),
559 _sleep_factor(0.0),
560 _marking_task_overhead(1.0),
561 _cleanup_sleep_factor(0.0),
562 _cleanup_task_overhead(1.0),
563 _cleanup_list("Cleanup List"),
564 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
565 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
566 CardTableModRefBS::card_shift,
567 false /* in_resource_area*/),
568
569 _prevMarkBitMap(&_markBitMap1),
570 _nextMarkBitMap(&_markBitMap2),
571
572 _markStack(this),
573 // _finger set in set_non_marking_state
574
575 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
576 // _active_tasks set in set_non_marking_state
577 // _tasks set inside the constructor
578 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
579 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
580
581 _has_overflown(false),
582 _concurrent(false),
583 _has_aborted(false),
584 _aborted_gc_id(GCId::undefined()),
585 _restart_for_overflow(false),
586 _concurrent_marking_in_progress(false),
587
588 // _verbose_level set below
589
590 _init_times(),
591 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
592 _cleanup_times(),
593 _total_counting_time(0.0),
594 _total_rs_scrub_time(0.0),
595
596 _parallel_workers(NULL),
597
598 _count_card_bitmaps(NULL),
599 _count_marked_bytes(NULL),
600 _completed_initialization(false) {
601 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
602 if (verbose_level < no_verbose) {
603 verbose_level = no_verbose;
604 }
605 if (verbose_level > high_verbose) {
606 verbose_level = high_verbose;
607 }
608 _verbose_level = verbose_level;
609
610 if (verbose_low()) {
611 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
612 "heap end = " PTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
613 }
614
615 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
616 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
617
618 // Create & start a ConcurrentMark thread.
619 _cmThread = new ConcurrentMarkThread(this);
620 assert(cmThread() != NULL, "CM Thread should have been created");
621 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
622 if (_cmThread->osthread() == NULL) {
623 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
624 }
625
626 assert(CGC_lock != NULL, "Where's the CGC_lock?");
627 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
628 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
629
630 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
631 satb_qs.set_buffer_size(G1SATBBufferSize);
632
633 _root_regions.init(_g1h, this);
634
635 if (ConcGCThreads > ParallelGCThreads) {
636 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
637 "than ParallelGCThreads (" UINTX_FORMAT ").",
638 ConcGCThreads, ParallelGCThreads);
639 return;
640 }
641 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
642 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
643 // if both are set
644 _sleep_factor = 0.0;
645 _marking_task_overhead = 1.0;
646 } else if (G1MarkingOverheadPercent > 0) {
647 // We will calculate the number of parallel marking threads based
648 // on a target overhead with respect to the soft real-time goal
649 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
650 double overall_cm_overhead =
651 (double) MaxGCPauseMillis * marking_overhead /
652 (double) GCPauseIntervalMillis;
653 double cpu_ratio = 1.0 / (double) os::processor_count();
654 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
655 double marking_task_overhead =
656 overall_cm_overhead / marking_thread_num *
657 (double) os::processor_count();
658 double sleep_factor =
659 (1.0 - marking_task_overhead) / marking_task_overhead;
660
661 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
662 _sleep_factor = sleep_factor;
663 _marking_task_overhead = marking_task_overhead;
664 } else {
665 // Calculate the number of parallel marking threads by scaling
666 // the number of parallel GC threads.
667 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
668 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
669 _sleep_factor = 0.0;
670 _marking_task_overhead = 1.0;
671 }
672
673 assert(ConcGCThreads > 0, "Should have been set");
674 _parallel_marking_threads = (uint) ConcGCThreads;
675 _max_parallel_marking_threads = _parallel_marking_threads;
676
677 if (parallel_marking_threads() > 1) {
678 _cleanup_task_overhead = 1.0;
679 } else {
680 _cleanup_task_overhead = marking_task_overhead();
681 }
682 _cleanup_sleep_factor =
683 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
684
685 #if 0
686 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
687 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
688 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
689 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
690 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
691 #endif
692
693 _parallel_workers = new FlexibleWorkGang("G1 Marker",
694 _max_parallel_marking_threads, false, true);
695 if (_parallel_workers == NULL) {
696 vm_exit_during_initialization("Failed necessary allocation.");
697 } else {
698 _parallel_workers->initialize_workers();
699 }
700
701 if (FLAG_IS_DEFAULT(MarkStackSize)) {
702 size_t mark_stack_size =
703 MIN2(MarkStackSizeMax,
704 MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE)));
705 // Verify that the calculated value for MarkStackSize is in range.
706 // It would be nice to use the private utility routine from Arguments.
707 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
708 warning("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
709 "must be between 1 and " SIZE_FORMAT,
710 mark_stack_size, MarkStackSizeMax);
711 return;
712 }
713 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
714 } else {
715 // Verify MarkStackSize is in range.
716 if (FLAG_IS_CMDLINE(MarkStackSize)) {
717 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
718 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
719 warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
720 "must be between 1 and " SIZE_FORMAT,
721 MarkStackSize, MarkStackSizeMax);
722 return;
723 }
724 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
725 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
726 warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
727 " or for MarkStackSizeMax (" SIZE_FORMAT ")",
728 MarkStackSize, MarkStackSizeMax);
729 return;
730 }
731 }
732 }
733 }
734
735 if (!_markStack.allocate(MarkStackSize)) {
736 warning("Failed to allocate CM marking stack");
737 return;
738 }
739
740 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
741 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
742
743 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
744 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
745
746 BitMap::idx_t card_bm_size = _card_bm.size();
747
748 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
749 _active_tasks = _max_worker_id;
750
751 uint max_regions = _g1h->max_regions();
752 for (uint i = 0; i < _max_worker_id; ++i) {
753 CMTaskQueue* task_queue = new CMTaskQueue();
754 task_queue->initialize();
755 _task_queues->register_queue(i, task_queue);
756
757 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
758 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
759
760 _tasks[i] = new CMTask(i, this,
761 _count_marked_bytes[i],
762 &_count_card_bitmaps[i],
763 task_queue, _task_queues);
764
765 _accum_task_vtime[i] = 0.0;
766 }
767
768 // Calculate the card number for the bottom of the heap. Used
769 // in biasing indexes into the accounting card bitmaps.
770 _heap_bottom_card_num =
771 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
772 CardTableModRefBS::card_shift);
773
774 // Clear all the liveness counting data
775 clear_all_count_data();
776
777 // so that the call below can read a sensible value
778 _heap_start = g1h->reserved_region().start();
779 set_non_marking_state();
780 _completed_initialization = true;
781 }
782
783 void ConcurrentMark::reset() {
784 // Starting values for these two. This should be called in a STW
785 // phase.
786 MemRegion reserved = _g1h->g1_reserved();
787 _heap_start = reserved.start();
788 _heap_end = reserved.end();
789
790 // Separated the asserts so that we know which one fires.
791 assert(_heap_start != NULL, "heap bounds should look ok");
792 assert(_heap_end != NULL, "heap bounds should look ok");
793 assert(_heap_start < _heap_end, "heap bounds should look ok");
794
795 // Reset all the marking data structures and any necessary flags
796 reset_marking_state();
797
798 if (verbose_low()) {
799 gclog_or_tty->print_cr("[global] resetting");
800 }
801
802 // We do reset all of them, since different phases will use
803 // different number of active threads. So, it's easiest to have all
804 // of them ready.
805 for (uint i = 0; i < _max_worker_id; ++i) {
806 _tasks[i]->reset(_nextMarkBitMap);
807 }
808
809 // we need this to make sure that the flag is on during the evac
810 // pause with initial mark piggy-backed
811 set_concurrent_marking_in_progress();
812 }
813
814
815 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
816 _markStack.set_should_expand();
817 _markStack.setEmpty(); // Also clears the _markStack overflow flag
818 if (clear_overflow) {
819 clear_has_overflown();
820 } else {
821 assert(has_overflown(), "pre-condition");
822 }
823 _finger = _heap_start;
824
825 for (uint i = 0; i < _max_worker_id; ++i) {
826 CMTaskQueue* queue = _task_queues->queue(i);
827 queue->set_empty();
828 }
829 }
830
831 void ConcurrentMark::set_concurrency(uint active_tasks) {
832 assert(active_tasks <= _max_worker_id, "we should not have more");
833
834 _active_tasks = active_tasks;
835 // Need to update the three data structures below according to the
836 // number of active threads for this phase.
837 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
838 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
839 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
840 }
841
842 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
843 set_concurrency(active_tasks);
844
845 _concurrent = concurrent;
846 // We propagate this to all tasks, not just the active ones.
847 for (uint i = 0; i < _max_worker_id; ++i)
848 _tasks[i]->set_concurrent(concurrent);
849
850 if (concurrent) {
851 set_concurrent_marking_in_progress();
852 } else {
853 // We currently assume that the concurrent flag has been set to
854 // false before we start remark. At this point we should also be
855 // in a STW phase.
856 assert(!concurrent_marking_in_progress(), "invariant");
857 assert(out_of_regions(),
858 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
859 p2i(_finger), p2i(_heap_end)));
860 }
861 }
862
863 void ConcurrentMark::set_non_marking_state() {
864 // We set the global marking state to some default values when we're
865 // not doing marking.
866 reset_marking_state();
867 _active_tasks = 0;
868 clear_concurrent_marking_in_progress();
869 }
870
871 ConcurrentMark::~ConcurrentMark() {
872 // The ConcurrentMark instance is never freed.
873 ShouldNotReachHere();
874 }
875
876 void ConcurrentMark::clearNextBitmap() {
877 G1CollectedHeap* g1h = G1CollectedHeap::heap();
878
879 // Make sure that the concurrent mark thread looks to still be in
880 // the current cycle.
881 guarantee(cmThread()->during_cycle(), "invariant");
882
883 // We are finishing up the current cycle by clearing the next
884 // marking bitmap and getting it ready for the next cycle. During
885 // this time no other cycle can start. So, let's make sure that this
886 // is the case.
887 guarantee(!g1h->mark_in_progress(), "invariant");
888
889 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
890 ParClearNextMarkBitmapTask task(&cl, parallel_marking_threads(), true);
891 _parallel_workers->run_task(&task);
892
893 // Clear the liveness counting data. If the marking has been aborted, the abort()
894 // call already did that.
895 if (cl.complete()) {
896 clear_all_count_data();
897 }
898
899 // Repeat the asserts from above.
900 guarantee(cmThread()->during_cycle(), "invariant");
901 guarantee(!g1h->mark_in_progress(), "invariant");
902 }
903
904 class CheckBitmapClearHRClosure : public HeapRegionClosure {
905 CMBitMap* _bitmap;
906 bool _error;
907 public:
908 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
909 }
910
911 virtual bool doHeapRegion(HeapRegion* r) {
912 // This closure can be called concurrently to the mutator, so we must make sure
913 // that the result of the getNextMarkedWordAddress() call is compared to the
914 // value passed to it as limit to detect any found bits.
915 // We can use the region's orig_end() for the limit and the comparison value
916 // as it always contains the "real" end of the region that never changes and
917 // has no side effects.
918 // Due to the latter, there can also be no problem with the compiler generating
919 // reloads of the orig_end() call.
920 HeapWord* end = r->orig_end();
921 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
922 }
923 };
924
925 bool ConcurrentMark::nextMarkBitmapIsClear() {
926 CheckBitmapClearHRClosure cl(_nextMarkBitMap);
927 _g1h->heap_region_iterate(&cl);
928 return cl.complete();
929 }
930
931 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
932 public:
933 bool doHeapRegion(HeapRegion* r) {
934 if (!r->is_continues_humongous()) {
935 r->note_start_of_marking();
936 }
937 return false;
938 }
939 };
940
941 void ConcurrentMark::checkpointRootsInitialPre() {
942 G1CollectedHeap* g1h = G1CollectedHeap::heap();
943 G1CollectorPolicy* g1p = g1h->g1_policy();
944
945 _has_aborted = false;
946
947 // Initialize marking structures. This has to be done in a STW phase.
948 reset();
949
950 // For each region note start of marking.
951 NoteStartOfMarkHRClosure startcl;
952 g1h->heap_region_iterate(&startcl);
953 }
954
955
956 void ConcurrentMark::checkpointRootsInitialPost() {
957 G1CollectedHeap* g1h = G1CollectedHeap::heap();
958
959 // If we force an overflow during remark, the remark operation will
960 // actually abort and we'll restart concurrent marking. If we always
961 // force an overflow during remark we'll never actually complete the
962 // marking phase. So, we initialize this here, at the start of the
963 // cycle, so that at the remaining overflow number will decrease at
964 // every remark and we'll eventually not need to cause one.
965 force_overflow_stw()->init();
966
967 // Start Concurrent Marking weak-reference discovery.
968 ReferenceProcessor* rp = g1h->ref_processor_cm();
969 // enable ("weak") refs discovery
970 rp->enable_discovery();
971 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
972
973 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
974 // This is the start of the marking cycle, we're expected all
975 // threads to have SATB queues with active set to false.
976 satb_mq_set.set_active_all_threads(true, /* new active value */
977 false /* expected_active */);
978
979 _root_regions.prepare_for_scan();
980
981 // update_g1_committed() will be called at the end of an evac pause
982 // when marking is on. So, it's also called at the end of the
983 // initial-mark pause to update the heap end, if the heap expands
984 // during it. No need to call it here.
985 }
986
987 /*
988 * Notice that in the next two methods, we actually leave the STS
989 * during the barrier sync and join it immediately afterwards. If we
990 * do not do this, the following deadlock can occur: one thread could
991 * be in the barrier sync code, waiting for the other thread to also
992 * sync up, whereas another one could be trying to yield, while also
993 * waiting for the other threads to sync up too.
994 *
995 * Note, however, that this code is also used during remark and in
996 * this case we should not attempt to leave / enter the STS, otherwise
997 * we'll either hit an assert (debug / fastdebug) or deadlock
998 * (product). So we should only leave / enter the STS if we are
999 * operating concurrently.
1000 *
1001 * Because the thread that does the sync barrier has left the STS, it
1002 * is possible to be suspended for a Full GC or an evacuation pause
1003 * could occur. This is actually safe, since the entering the sync
1004 * barrier is one of the last things do_marking_step() does, and it
1005 * doesn't manipulate any data structures afterwards.
1006 */
1007
1008 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
1009 if (verbose_low()) {
1010 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
1011 }
1012
1013 if (concurrent()) {
1014 SuspendibleThreadSet::leave();
1015 }
1016
1017 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
1018
1019 if (concurrent()) {
1020 SuspendibleThreadSet::join();
1021 }
1022 // at this point everyone should have synced up and not be doing any
1023 // more work
1024
1025 if (verbose_low()) {
1026 if (barrier_aborted) {
1027 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1028 } else {
1029 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1030 }
1031 }
1032
1033 if (barrier_aborted) {
1034 // If the barrier aborted we ignore the overflow condition and
1035 // just abort the whole marking phase as quickly as possible.
1036 return;
1037 }
1038
1039 // If we're executing the concurrent phase of marking, reset the marking
1040 // state; otherwise the marking state is reset after reference processing,
1041 // during the remark pause.
1042 // If we reset here as a result of an overflow during the remark we will
1043 // see assertion failures from any subsequent set_concurrency_and_phase()
1044 // calls.
1045 if (concurrent()) {
1046 // let the task associated with with worker 0 do this
1047 if (worker_id == 0) {
1048 // task 0 is responsible for clearing the global data structures
1049 // We should be here because of an overflow. During STW we should
1050 // not clear the overflow flag since we rely on it being true when
1051 // we exit this method to abort the pause and restart concurrent
1052 // marking.
1053 reset_marking_state(true /* clear_overflow */);
1054 force_overflow()->update();
1055
1056 if (G1Log::fine()) {
1057 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1058 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1059 }
1060 }
1061 }
1062
1063 // after this, each task should reset its own data structures then
1064 // then go into the second barrier
1065 }
1066
1067 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1068 if (verbose_low()) {
1069 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1070 }
1071
1072 if (concurrent()) {
1073 SuspendibleThreadSet::leave();
1074 }
1075
1076 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1077
1078 if (concurrent()) {
1079 SuspendibleThreadSet::join();
1080 }
1081 // at this point everything should be re-initialized and ready to go
1082
1083 if (verbose_low()) {
1084 if (barrier_aborted) {
1085 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1086 } else {
1087 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1088 }
1089 }
1090 }
1091
1092 #ifndef PRODUCT
1093 void ForceOverflowSettings::init() {
1094 _num_remaining = G1ConcMarkForceOverflow;
1095 _force = false;
1096 update();
1097 }
1098
1099 void ForceOverflowSettings::update() {
1100 if (_num_remaining > 0) {
1101 _num_remaining -= 1;
1102 _force = true;
1103 } else {
1104 _force = false;
1105 }
1106 }
1107
1108 bool ForceOverflowSettings::should_force() {
1109 if (_force) {
1110 _force = false;
1111 return true;
1112 } else {
1113 return false;
1114 }
1115 }
1116 #endif // !PRODUCT
1117
1118 class CMConcurrentMarkingTask: public AbstractGangTask {
1119 private:
1120 ConcurrentMark* _cm;
1121 ConcurrentMarkThread* _cmt;
1122
1123 public:
1124 void work(uint worker_id) {
1125 assert(Thread::current()->is_ConcurrentGC_thread(),
1126 "this should only be done by a conc GC thread");
1127 ResourceMark rm;
1128
1129 double start_vtime = os::elapsedVTime();
1130
1131 SuspendibleThreadSet::join();
1132
1133 assert(worker_id < _cm->active_tasks(), "invariant");
1134 CMTask* the_task = _cm->task(worker_id);
1135 the_task->record_start_time();
1136 if (!_cm->has_aborted()) {
1137 do {
1138 double start_vtime_sec = os::elapsedVTime();
1139 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1140
1141 the_task->do_marking_step(mark_step_duration_ms,
1142 true /* do_termination */,
1143 false /* is_serial*/);
1144
1145 double end_vtime_sec = os::elapsedVTime();
1146 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1147 _cm->clear_has_overflown();
1148
1149 _cm->do_yield_check(worker_id);
1150
1151 jlong sleep_time_ms;
1152 if (!_cm->has_aborted() && the_task->has_aborted()) {
1153 sleep_time_ms =
1154 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1155 SuspendibleThreadSet::leave();
1156 os::sleep(Thread::current(), sleep_time_ms, false);
1157 SuspendibleThreadSet::join();
1158 }
1159 } while (!_cm->has_aborted() && the_task->has_aborted());
1160 }
1161 the_task->record_end_time();
1162 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1163
1164 SuspendibleThreadSet::leave();
1165
1166 double end_vtime = os::elapsedVTime();
1167 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1168 }
1169
1170 CMConcurrentMarkingTask(ConcurrentMark* cm,
1171 ConcurrentMarkThread* cmt) :
1172 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1173
1174 ~CMConcurrentMarkingTask() { }
1175 };
1176
1177 // Calculates the number of active workers for a concurrent
1178 // phase.
1179 uint ConcurrentMark::calc_parallel_marking_threads() {
1180 uint n_conc_workers = 0;
1181 if (!UseDynamicNumberOfGCThreads ||
1182 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1183 !ForceDynamicNumberOfGCThreads)) {
1184 n_conc_workers = max_parallel_marking_threads();
1185 } else {
1186 n_conc_workers =
1187 AdaptiveSizePolicy::calc_default_active_workers(
1188 max_parallel_marking_threads(),
1189 1, /* Minimum workers */
1190 parallel_marking_threads(),
1191 Threads::number_of_non_daemon_threads());
1192 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1193 // that scaling has already gone into "_max_parallel_marking_threads".
1194 }
1195 assert(n_conc_workers > 0, "Always need at least 1");
1196 return n_conc_workers;
1197 }
1198
1199 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1200 // Currently, only survivors can be root regions.
1201 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1202 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1203
1204 const uintx interval = PrefetchScanIntervalInBytes;
1205 HeapWord* curr = hr->bottom();
1206 const HeapWord* end = hr->top();
1207 while (curr < end) {
1208 Prefetch::read(curr, interval);
1209 oop obj = oop(curr);
1210 int size = obj->oop_iterate(&cl);
1211 assert(size == obj->size(), "sanity");
1212 curr += size;
1213 }
1214 }
1215
1216 class CMRootRegionScanTask : public AbstractGangTask {
1217 private:
1218 ConcurrentMark* _cm;
1219
1220 public:
1221 CMRootRegionScanTask(ConcurrentMark* cm) :
1222 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1223
1224 void work(uint worker_id) {
1225 assert(Thread::current()->is_ConcurrentGC_thread(),
1226 "this should only be done by a conc GC thread");
1227
1228 CMRootRegions* root_regions = _cm->root_regions();
1229 HeapRegion* hr = root_regions->claim_next();
1230 while (hr != NULL) {
1231 _cm->scanRootRegion(hr, worker_id);
1232 hr = root_regions->claim_next();
1233 }
1234 }
1235 };
1236
1237 void ConcurrentMark::scanRootRegions() {
1238 // Start of concurrent marking.
1239 ClassLoaderDataGraph::clear_claimed_marks();
1240
1241 // scan_in_progress() will have been set to true only if there was
1242 // at least one root region to scan. So, if it's false, we
1243 // should not attempt to do any further work.
1244 if (root_regions()->scan_in_progress()) {
1245 _parallel_marking_threads = calc_parallel_marking_threads();
1246 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1247 "Maximum number of marking threads exceeded");
1248 uint active_workers = MAX2(1U, parallel_marking_threads());
1249
1250 CMRootRegionScanTask task(this);
1251 _parallel_workers->set_active_workers(active_workers);
1252 _parallel_workers->run_task(&task);
1253
1254 // It's possible that has_aborted() is true here without actually
1255 // aborting the survivor scan earlier. This is OK as it's
1256 // mainly used for sanity checking.
1257 root_regions()->scan_finished();
1258 }
1259 }
1260
1261 void ConcurrentMark::markFromRoots() {
1262 // we might be tempted to assert that:
1263 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1264 // "inconsistent argument?");
1265 // However that wouldn't be right, because it's possible that
1266 // a safepoint is indeed in progress as a younger generation
1267 // stop-the-world GC happens even as we mark in this generation.
1268
1269 _restart_for_overflow = false;
1270 force_overflow_conc()->init();
1271
1272 // _g1h has _n_par_threads
1273 _parallel_marking_threads = calc_parallel_marking_threads();
1274 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1275 "Maximum number of marking threads exceeded");
1276
1277 uint active_workers = MAX2(1U, parallel_marking_threads());
1278
1279 // Parallel task terminator is set in "set_concurrency_and_phase()"
1280 set_concurrency_and_phase(active_workers, true /* concurrent */);
1281
1282 CMConcurrentMarkingTask markingTask(this, cmThread());
1283 _parallel_workers->set_active_workers(active_workers);
1284 // Don't set _n_par_threads because it affects MT in process_roots()
1285 // and the decisions on that MT processing is made elsewhere.
1286 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1287 _parallel_workers->run_task(&markingTask);
1288 print_stats();
1289 }
1290
1291 // Helper class to get rid of some boilerplate code.
1292 class G1CMTraceTime : public GCTraceTime {
1293 static bool doit_and_prepend(bool doit) {
1294 if (doit) {
1295 gclog_or_tty->put(' ');
1296 }
1297 return doit;
1298 }
1299
1300 public:
1301 G1CMTraceTime(const char* title, bool doit)
1302 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
1303 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
1304 }
1305 };
1306
1307 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1308 // world is stopped at this checkpoint
1309 assert(SafepointSynchronize::is_at_safepoint(),
1310 "world should be stopped");
1311
1312 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1313
1314 // If a full collection has happened, we shouldn't do this.
1315 if (has_aborted()) {
1316 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1317 return;
1318 }
1319
1320 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1321
1322 if (VerifyDuringGC) {
1323 HandleMark hm; // handle scope
1324 g1h->prepare_for_verify();
1325 Universe::verify(VerifyOption_G1UsePrevMarking,
1326 " VerifyDuringGC:(before)");
1327 }
1328 g1h->check_bitmaps("Remark Start");
1329
1330 G1CollectorPolicy* g1p = g1h->g1_policy();
1331 g1p->record_concurrent_mark_remark_start();
1332
1333 double start = os::elapsedTime();
1334
1335 checkpointRootsFinalWork();
1336
1337 double mark_work_end = os::elapsedTime();
1338
1339 weakRefsWork(clear_all_soft_refs);
1340
1341 if (has_overflown()) {
1342 // Oops. We overflowed. Restart concurrent marking.
1343 _restart_for_overflow = true;
1344 if (G1TraceMarkStackOverflow) {
1345 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1346 }
1347
1348 // Verify the heap w.r.t. the previous marking bitmap.
1349 if (VerifyDuringGC) {
1350 HandleMark hm; // handle scope
1351 g1h->prepare_for_verify();
1352 Universe::verify(VerifyOption_G1UsePrevMarking,
1353 " VerifyDuringGC:(overflow)");
1354 }
1355
1356 // Clear the marking state because we will be restarting
1357 // marking due to overflowing the global mark stack.
1358 reset_marking_state();
1359 } else {
1360 {
1361 G1CMTraceTime trace("GC aggregate-data", G1Log::finer());
1362
1363 // Aggregate the per-task counting data that we have accumulated
1364 // while marking.
1365 aggregate_count_data();
1366 }
1367
1368 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1369 // We're done with marking.
1370 // This is the end of the marking cycle, we're expected all
1371 // threads to have SATB queues with active set to true.
1372 satb_mq_set.set_active_all_threads(false, /* new active value */
1373 true /* expected_active */);
1374
1375 if (VerifyDuringGC) {
1376 HandleMark hm; // handle scope
1377 g1h->prepare_for_verify();
1378 Universe::verify(VerifyOption_G1UseNextMarking,
1379 " VerifyDuringGC:(after)");
1380 }
1381 g1h->check_bitmaps("Remark End");
1382 assert(!restart_for_overflow(), "sanity");
1383 // Completely reset the marking state since marking completed
1384 set_non_marking_state();
1385 }
1386
1387 // Expand the marking stack, if we have to and if we can.
1388 if (_markStack.should_expand()) {
1389 _markStack.expand();
1390 }
1391
1392 // Statistics
1393 double now = os::elapsedTime();
1394 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1395 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1396 _remark_times.add((now - start) * 1000.0);
1397
1398 g1p->record_concurrent_mark_remark_end();
1399
1400 G1CMIsAliveClosure is_alive(g1h);
1401 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1402 }
1403
1404 // Base class of the closures that finalize and verify the
1405 // liveness counting data.
1406 class CMCountDataClosureBase: public HeapRegionClosure {
1407 protected:
1408 G1CollectedHeap* _g1h;
1409 ConcurrentMark* _cm;
1410 CardTableModRefBS* _ct_bs;
1411
1412 BitMap* _region_bm;
1413 BitMap* _card_bm;
1414
1415 // Takes a region that's not empty (i.e., it has at least one
1416 // live object in it and sets its corresponding bit on the region
1417 // bitmap to 1. If the region is "starts humongous" it will also set
1418 // to 1 the bits on the region bitmap that correspond to its
1419 // associated "continues humongous" regions.
1420 void set_bit_for_region(HeapRegion* hr) {
1421 assert(!hr->is_continues_humongous(), "should have filtered those out");
1422
1423 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1424 if (!hr->is_starts_humongous()) {
1425 // Normal (non-humongous) case: just set the bit.
1426 _region_bm->par_at_put(index, true);
1427 } else {
1428 // Starts humongous case: calculate how many regions are part of
1429 // this humongous region and then set the bit range.
1430 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1431 _region_bm->par_at_put_range(index, end_index, true);
1432 }
1433 }
1434
1435 public:
1436 CMCountDataClosureBase(G1CollectedHeap* g1h,
1437 BitMap* region_bm, BitMap* card_bm):
1438 _g1h(g1h), _cm(g1h->concurrent_mark()),
1439 _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
1440 _region_bm(region_bm), _card_bm(card_bm) { }
1441 };
1442
1443 // Closure that calculates the # live objects per region. Used
1444 // for verification purposes during the cleanup pause.
1445 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1446 CMBitMapRO* _bm;
1447 size_t _region_marked_bytes;
1448
1449 public:
1450 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1451 BitMap* region_bm, BitMap* card_bm) :
1452 CMCountDataClosureBase(g1h, region_bm, card_bm),
1453 _bm(bm), _region_marked_bytes(0) { }
1454
1455 bool doHeapRegion(HeapRegion* hr) {
1456
1457 if (hr->is_continues_humongous()) {
1458 // We will ignore these here and process them when their
1459 // associated "starts humongous" region is processed (see
1460 // set_bit_for_heap_region()). Note that we cannot rely on their
1461 // associated "starts humongous" region to have their bit set to
1462 // 1 since, due to the region chunking in the parallel region
1463 // iteration, a "continues humongous" region might be visited
1464 // before its associated "starts humongous".
1465 return false;
1466 }
1467
1468 HeapWord* ntams = hr->next_top_at_mark_start();
1469 HeapWord* start = hr->bottom();
1470
1471 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1472 err_msg("Preconditions not met - "
1473 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1474 p2i(start), p2i(ntams), p2i(hr->end())));
1475
1476 // Find the first marked object at or after "start".
1477 start = _bm->getNextMarkedWordAddress(start, ntams);
1478
1479 size_t marked_bytes = 0;
1480
1481 while (start < ntams) {
1482 oop obj = oop(start);
1483 int obj_sz = obj->size();
1484 HeapWord* obj_end = start + obj_sz;
1485
1486 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1487 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1488
1489 // Note: if we're looking at the last region in heap - obj_end
1490 // could be actually just beyond the end of the heap; end_idx
1491 // will then correspond to a (non-existent) card that is also
1492 // just beyond the heap.
1493 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1494 // end of object is not card aligned - increment to cover
1495 // all the cards spanned by the object
1496 end_idx += 1;
1497 }
1498
1499 // Set the bits in the card BM for the cards spanned by this object.
1500 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1501
1502 // Add the size of this object to the number of marked bytes.
1503 marked_bytes += (size_t)obj_sz * HeapWordSize;
1504
1505 // Find the next marked object after this one.
1506 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1507 }
1508
1509 // Mark the allocated-since-marking portion...
1510 HeapWord* top = hr->top();
1511 if (ntams < top) {
1512 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1513 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1514
1515 // Note: if we're looking at the last region in heap - top
1516 // could be actually just beyond the end of the heap; end_idx
1517 // will then correspond to a (non-existent) card that is also
1518 // just beyond the heap.
1519 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1520 // end of object is not card aligned - increment to cover
1521 // all the cards spanned by the object
1522 end_idx += 1;
1523 }
1524 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1525
1526 // This definitely means the region has live objects.
1527 set_bit_for_region(hr);
1528 }
1529
1530 // Update the live region bitmap.
1531 if (marked_bytes > 0) {
1532 set_bit_for_region(hr);
1533 }
1534
1535 // Set the marked bytes for the current region so that
1536 // it can be queried by a calling verification routine
1537 _region_marked_bytes = marked_bytes;
1538
1539 return false;
1540 }
1541
1542 size_t region_marked_bytes() const { return _region_marked_bytes; }
1543 };
1544
1545 // Heap region closure used for verifying the counting data
1546 // that was accumulated concurrently and aggregated during
1547 // the remark pause. This closure is applied to the heap
1548 // regions during the STW cleanup pause.
1549
1550 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1551 G1CollectedHeap* _g1h;
1552 ConcurrentMark* _cm;
1553 CalcLiveObjectsClosure _calc_cl;
1554 BitMap* _region_bm; // Region BM to be verified
1555 BitMap* _card_bm; // Card BM to be verified
1556 bool _verbose; // verbose output?
1557
1558 BitMap* _exp_region_bm; // Expected Region BM values
1559 BitMap* _exp_card_bm; // Expected card BM values
1560
1561 int _failures;
1562
1563 public:
1564 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1565 BitMap* region_bm,
1566 BitMap* card_bm,
1567 BitMap* exp_region_bm,
1568 BitMap* exp_card_bm,
1569 bool verbose) :
1570 _g1h(g1h), _cm(g1h->concurrent_mark()),
1571 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1572 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1573 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1574 _failures(0) { }
1575
1576 int failures() const { return _failures; }
1577
1578 bool doHeapRegion(HeapRegion* hr) {
1579 if (hr->is_continues_humongous()) {
1580 // We will ignore these here and process them when their
1581 // associated "starts humongous" region is processed (see
1582 // set_bit_for_heap_region()). Note that we cannot rely on their
1583 // associated "starts humongous" region to have their bit set to
1584 // 1 since, due to the region chunking in the parallel region
1585 // iteration, a "continues humongous" region might be visited
1586 // before its associated "starts humongous".
1587 return false;
1588 }
1589
1590 int failures = 0;
1591
1592 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1593 // this region and set the corresponding bits in the expected region
1594 // and card bitmaps.
1595 bool res = _calc_cl.doHeapRegion(hr);
1596 assert(res == false, "should be continuing");
1597
1598 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1599 Mutex::_no_safepoint_check_flag);
1600
1601 // Verify the marked bytes for this region.
1602 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1603 size_t act_marked_bytes = hr->next_marked_bytes();
1604
1605 // We're not OK if expected marked bytes > actual marked bytes. It means
1606 // we have missed accounting some objects during the actual marking.
1607 if (exp_marked_bytes > act_marked_bytes) {
1608 if (_verbose) {
1609 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1610 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1611 hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1612 }
1613 failures += 1;
1614 }
1615
1616 // Verify the bit, for this region, in the actual and expected
1617 // (which was just calculated) region bit maps.
1618 // We're not OK if the bit in the calculated expected region
1619 // bitmap is set and the bit in the actual region bitmap is not.
1620 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1621
1622 bool expected = _exp_region_bm->at(index);
1623 bool actual = _region_bm->at(index);
1624 if (expected && !actual) {
1625 if (_verbose) {
1626 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1627 "expected: %s, actual: %s",
1628 hr->hrm_index(),
1629 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1630 }
1631 failures += 1;
1632 }
1633
1634 // Verify that the card bit maps for the cards spanned by the current
1635 // region match. We have an error if we have a set bit in the expected
1636 // bit map and the corresponding bit in the actual bitmap is not set.
1637
1638 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1639 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1640
1641 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1642 expected = _exp_card_bm->at(i);
1643 actual = _card_bm->at(i);
1644
1645 if (expected && !actual) {
1646 if (_verbose) {
1647 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1648 "expected: %s, actual: %s",
1649 hr->hrm_index(), i,
1650 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1651 }
1652 failures += 1;
1653 }
1654 }
1655
1656 if (failures > 0 && _verbose) {
1657 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1658 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1659 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1660 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1661 }
1662
1663 _failures += failures;
1664
1665 // We could stop iteration over the heap when we
1666 // find the first violating region by returning true.
1667 return false;
1668 }
1669 };
1670
1671 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1672 protected:
1673 G1CollectedHeap* _g1h;
1674 ConcurrentMark* _cm;
1675 BitMap* _actual_region_bm;
1676 BitMap* _actual_card_bm;
1677
1678 uint _n_workers;
1679
1680 BitMap* _expected_region_bm;
1681 BitMap* _expected_card_bm;
1682
1683 int _failures;
1684 bool _verbose;
1685
1686 HeapRegionClaimer _hrclaimer;
1687
1688 public:
1689 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1690 BitMap* region_bm, BitMap* card_bm,
1691 BitMap* expected_region_bm, BitMap* expected_card_bm)
1692 : AbstractGangTask("G1 verify final counting"),
1693 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1694 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1695 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1696 _failures(0), _verbose(false),
1697 _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1698 assert(VerifyDuringGC, "don't call this otherwise");
1699 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1700 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1701
1702 _verbose = _cm->verbose_medium();
1703 }
1704
1705 void work(uint worker_id) {
1706 assert(worker_id < _n_workers, "invariant");
1707
1708 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1709 _actual_region_bm, _actual_card_bm,
1710 _expected_region_bm,
1711 _expected_card_bm,
1712 _verbose);
1713
1714 _g1h->heap_region_par_iterate(&verify_cl, worker_id, &_hrclaimer);
1715
1716 Atomic::add(verify_cl.failures(), &_failures);
1717 }
1718
1719 int failures() const { return _failures; }
1720 };
1721
1722 // Closure that finalizes the liveness counting data.
1723 // Used during the cleanup pause.
1724 // Sets the bits corresponding to the interval [NTAMS, top]
1725 // (which contains the implicitly live objects) in the
1726 // card liveness bitmap. Also sets the bit for each region,
1727 // containing live data, in the region liveness bitmap.
1728
1729 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1730 public:
1731 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1732 BitMap* region_bm,
1733 BitMap* card_bm) :
1734 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1735
1736 bool doHeapRegion(HeapRegion* hr) {
1737
1738 if (hr->is_continues_humongous()) {
1739 // We will ignore these here and process them when their
1740 // associated "starts humongous" region is processed (see
1741 // set_bit_for_heap_region()). Note that we cannot rely on their
1742 // associated "starts humongous" region to have their bit set to
1743 // 1 since, due to the region chunking in the parallel region
1744 // iteration, a "continues humongous" region might be visited
1745 // before its associated "starts humongous".
1746 return false;
1747 }
1748
1749 HeapWord* ntams = hr->next_top_at_mark_start();
1750 HeapWord* top = hr->top();
1751
1752 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1753
1754 // Mark the allocated-since-marking portion...
1755 if (ntams < top) {
1756 // This definitely means the region has live objects.
1757 set_bit_for_region(hr);
1758
1759 // Now set the bits in the card bitmap for [ntams, top)
1760 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1761 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1762
1763 // Note: if we're looking at the last region in heap - top
1764 // could be actually just beyond the end of the heap; end_idx
1765 // will then correspond to a (non-existent) card that is also
1766 // just beyond the heap.
1767 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1768 // end of object is not card aligned - increment to cover
1769 // all the cards spanned by the object
1770 end_idx += 1;
1771 }
1772
1773 assert(end_idx <= _card_bm->size(),
1774 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1775 end_idx, _card_bm->size()));
1776 assert(start_idx < _card_bm->size(),
1777 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1778 start_idx, _card_bm->size()));
1779
1780 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1781 }
1782
1783 // Set the bit for the region if it contains live data
1784 if (hr->next_marked_bytes() > 0) {
1785 set_bit_for_region(hr);
1786 }
1787
1788 return false;
1789 }
1790 };
1791
1792 class G1ParFinalCountTask: public AbstractGangTask {
1793 protected:
1794 G1CollectedHeap* _g1h;
1795 ConcurrentMark* _cm;
1796 BitMap* _actual_region_bm;
1797 BitMap* _actual_card_bm;
1798
1799 uint _n_workers;
1800 HeapRegionClaimer _hrclaimer;
1801
1802 public:
1803 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1804 : AbstractGangTask("G1 final counting"),
1805 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1806 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1807 _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1808 }
1809
1810 void work(uint worker_id) {
1811 assert(worker_id < _n_workers, "invariant");
1812
1813 FinalCountDataUpdateClosure final_update_cl(_g1h,
1814 _actual_region_bm,
1815 _actual_card_bm);
1816
1817 _g1h->heap_region_par_iterate(&final_update_cl, worker_id, &_hrclaimer);
1818 }
1819 };
1820
1821 class G1ParNoteEndTask;
1822
1823 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1824 G1CollectedHeap* _g1;
1825 size_t _max_live_bytes;
1826 uint _regions_claimed;
1827 size_t _freed_bytes;
1828 FreeRegionList* _local_cleanup_list;
1829 HeapRegionSetCount _old_regions_removed;
1830 HeapRegionSetCount _humongous_regions_removed;
1831 HRRSCleanupTask* _hrrs_cleanup_task;
1832 double _claimed_region_time;
1833 double _max_region_time;
1834
1835 public:
1836 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1837 FreeRegionList* local_cleanup_list,
1838 HRRSCleanupTask* hrrs_cleanup_task) :
1839 _g1(g1),
1840 _max_live_bytes(0), _regions_claimed(0),
1841 _freed_bytes(0),
1842 _claimed_region_time(0.0), _max_region_time(0.0),
1843 _local_cleanup_list(local_cleanup_list),
1844 _old_regions_removed(),
1845 _humongous_regions_removed(),
1846 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1847
1848 size_t freed_bytes() { return _freed_bytes; }
1849 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1850 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1851
1852 bool doHeapRegion(HeapRegion *hr) {
1853 if (hr->is_continues_humongous()) {
1854 return false;
1855 }
1856 // We use a claim value of zero here because all regions
1857 // were claimed with value 1 in the FinalCount task.
1858 _g1->reset_gc_time_stamps(hr);
1859 double start = os::elapsedTime();
1860 _regions_claimed++;
1861 hr->note_end_of_marking();
1862 _max_live_bytes += hr->max_live_bytes();
1863
1864 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1865 _freed_bytes += hr->used();
1866 hr->set_containing_set(NULL);
1867 if (hr->is_humongous()) {
1868 assert(hr->is_starts_humongous(), "we should only see starts humongous");
1869 _humongous_regions_removed.increment(1u, hr->capacity());
1870 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1871 } else {
1872 _old_regions_removed.increment(1u, hr->capacity());
1873 _g1->free_region(hr, _local_cleanup_list, true);
1874 }
1875 } else {
1876 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1877 }
1878
1879 double region_time = (os::elapsedTime() - start);
1880 _claimed_region_time += region_time;
1881 if (region_time > _max_region_time) {
1882 _max_region_time = region_time;
1883 }
1884 return false;
1885 }
1886
1887 size_t max_live_bytes() { return _max_live_bytes; }
1888 uint regions_claimed() { return _regions_claimed; }
1889 double claimed_region_time_sec() { return _claimed_region_time; }
1890 double max_region_time_sec() { return _max_region_time; }
1891 };
1892
1893 class G1ParNoteEndTask: public AbstractGangTask {
1894 friend class G1NoteEndOfConcMarkClosure;
1895
1896 protected:
1897 G1CollectedHeap* _g1h;
1898 size_t _max_live_bytes;
1899 size_t _freed_bytes;
1900 FreeRegionList* _cleanup_list;
1901 HeapRegionClaimer _hrclaimer;
1902
1903 public:
1904 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1905 AbstractGangTask("G1 note end"), _g1h(g1h), _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1906 }
1907
1908 void work(uint worker_id) {
1909 FreeRegionList local_cleanup_list("Local Cleanup List");
1910 HRRSCleanupTask hrrs_cleanup_task;
1911 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1912 &hrrs_cleanup_task);
1913 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1914 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1915
1916 // Now update the lists
1917 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1918 {
1919 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1920 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1921 _max_live_bytes += g1_note_end.max_live_bytes();
1922 _freed_bytes += g1_note_end.freed_bytes();
1923
1924 // If we iterate over the global cleanup list at the end of
1925 // cleanup to do this printing we will not guarantee to only
1926 // generate output for the newly-reclaimed regions (the list
1927 // might not be empty at the beginning of cleanup; we might
1928 // still be working on its previous contents). So we do the
1929 // printing here, before we append the new regions to the global
1930 // cleanup list.
1931
1932 G1HRPrinter* hr_printer = _g1h->hr_printer();
1933 if (hr_printer->is_active()) {
1934 FreeRegionListIterator iter(&local_cleanup_list);
1935 while (iter.more_available()) {
1936 HeapRegion* hr = iter.get_next();
1937 hr_printer->cleanup(hr);
1938 }
1939 }
1940
1941 _cleanup_list->add_ordered(&local_cleanup_list);
1942 assert(local_cleanup_list.is_empty(), "post-condition");
1943
1944 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1945 }
1946 }
1947 size_t max_live_bytes() { return _max_live_bytes; }
1948 size_t freed_bytes() { return _freed_bytes; }
1949 };
1950
1951 class G1ParScrubRemSetTask: public AbstractGangTask {
1952 protected:
1953 G1RemSet* _g1rs;
1954 BitMap* _region_bm;
1955 BitMap* _card_bm;
1956 HeapRegionClaimer _hrclaimer;
1957
1958 public:
1959 G1ParScrubRemSetTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm, uint n_workers) :
1960 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), _region_bm(region_bm), _card_bm(card_bm), _hrclaimer(n_workers) {
1961 }
1962
1963 void work(uint worker_id) {
1964 _g1rs->scrub(_region_bm, _card_bm, worker_id, &_hrclaimer);
1965 }
1966
1967 };
1968
1969 void ConcurrentMark::cleanup() {
1970 // world is stopped at this checkpoint
1971 assert(SafepointSynchronize::is_at_safepoint(),
1972 "world should be stopped");
1973 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1974
1975 // If a full collection has happened, we shouldn't do this.
1976 if (has_aborted()) {
1977 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1978 return;
1979 }
1980
1981 g1h->verify_region_sets_optional();
1982
1983 if (VerifyDuringGC) {
1984 HandleMark hm; // handle scope
1985 g1h->prepare_for_verify();
1986 Universe::verify(VerifyOption_G1UsePrevMarking,
1987 " VerifyDuringGC:(before)");
1988 }
1989 g1h->check_bitmaps("Cleanup Start");
1990
1991 G1CollectorPolicy* g1p = g1h->g1_policy();
1992 g1p->record_concurrent_mark_cleanup_start();
1993
1994 double start = os::elapsedTime();
1995
1996 HeapRegionRemSet::reset_for_cleanup_tasks();
1997
1998 // Do counting once more with the world stopped for good measure.
1999 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2000
2001 g1h->set_par_threads();
2002 uint n_workers = _g1h->workers()->active_workers();
2003 g1h->workers()->run_task(&g1_par_count_task);
2004 // Done with the parallel phase so reset to 0.
2005 g1h->set_par_threads(0);
2006
2007 if (VerifyDuringGC) {
2008 // Verify that the counting data accumulated during marking matches
2009 // that calculated by walking the marking bitmap.
2010
2011 // Bitmaps to hold expected values
2012 BitMap expected_region_bm(_region_bm.size(), true);
2013 BitMap expected_card_bm(_card_bm.size(), true);
2014
2015 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2016 &_region_bm,
2017 &_card_bm,
2018 &expected_region_bm,
2019 &expected_card_bm);
2020
2021 g1h->set_par_threads((int)n_workers);
2022 g1h->workers()->run_task(&g1_par_verify_task);
2023 // Done with the parallel phase so reset to 0.
2024 g1h->set_par_threads(0);
2025
2026 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2027 }
2028
2029 size_t start_used_bytes = g1h->used();
2030 g1h->set_marking_complete();
2031
2032 double count_end = os::elapsedTime();
2033 double this_final_counting_time = (count_end - start);
2034 _total_counting_time += this_final_counting_time;
2035
2036 if (G1PrintRegionLivenessInfo) {
2037 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2038 _g1h->heap_region_iterate(&cl);
2039 }
2040
2041 // Install newly created mark bitMap as "prev".
2042 swapMarkBitMaps();
2043
2044 g1h->reset_gc_time_stamp();
2045
2046 // Note end of marking in all heap regions.
2047 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
2048 g1h->set_par_threads((int)n_workers);
2049 g1h->workers()->run_task(&g1_par_note_end_task);
2050 g1h->set_par_threads(0);
2051 g1h->check_gc_time_stamps();
2052
2053 if (!cleanup_list_is_empty()) {
2054 // The cleanup list is not empty, so we'll have to process it
2055 // concurrently. Notify anyone else that might be wanting free
2056 // regions that there will be more free regions coming soon.
2057 g1h->set_free_regions_coming();
2058 }
2059
2060 // call below, since it affects the metric by which we sort the heap
2061 // regions.
2062 if (G1ScrubRemSets) {
2063 double rs_scrub_start = os::elapsedTime();
2064 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm, n_workers);
2065 g1h->set_par_threads((int)n_workers);
2066 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2067 g1h->set_par_threads(0);
2068
2069 double rs_scrub_end = os::elapsedTime();
2070 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2071 _total_rs_scrub_time += this_rs_scrub_time;
2072 }
2073
2074 // this will also free any regions totally full of garbage objects,
2075 // and sort the regions.
2076 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2077
2078 // Statistics.
2079 double end = os::elapsedTime();
2080 _cleanup_times.add((end - start) * 1000.0);
2081
2082 if (G1Log::fine()) {
2083 g1h->g1_policy()->print_heap_transition(start_used_bytes);
2084 }
2085
2086 // Clean up will have freed any regions completely full of garbage.
2087 // Update the soft reference policy with the new heap occupancy.
2088 Universe::update_heap_info_at_gc();
2089
2090 if (VerifyDuringGC) {
2091 HandleMark hm; // handle scope
2092 g1h->prepare_for_verify();
2093 Universe::verify(VerifyOption_G1UsePrevMarking,
2094 " VerifyDuringGC:(after)");
2095 }
2096
2097 g1h->check_bitmaps("Cleanup End");
2098
2099 g1h->verify_region_sets_optional();
2100
2101 // We need to make this be a "collection" so any collection pause that
2102 // races with it goes around and waits for completeCleanup to finish.
2103 g1h->increment_total_collections();
2104
2105 // Clean out dead classes and update Metaspace sizes.
2106 if (ClassUnloadingWithConcurrentMark) {
2107 ClassLoaderDataGraph::purge();
2108 }
2109 MetaspaceGC::compute_new_size();
2110
2111 // We reclaimed old regions so we should calculate the sizes to make
2112 // sure we update the old gen/space data.
2113 g1h->g1mm()->update_sizes();
2114 g1h->allocation_context_stats().update_after_mark();
2115
2116 g1h->trace_heap_after_concurrent_cycle();
2117 }
2118
2119 void ConcurrentMark::completeCleanup() {
2120 if (has_aborted()) return;
2121
2122 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2123
2124 _cleanup_list.verify_optional();
2125 FreeRegionList tmp_free_list("Tmp Free List");
2126
2127 if (G1ConcRegionFreeingVerbose) {
2128 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2129 "cleanup list has %u entries",
2130 _cleanup_list.length());
2131 }
2132
2133 // No one else should be accessing the _cleanup_list at this point,
2134 // so it is not necessary to take any locks
2135 while (!_cleanup_list.is_empty()) {
2136 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2137 assert(hr != NULL, "Got NULL from a non-empty list");
2138 hr->par_clear();
2139 tmp_free_list.add_ordered(hr);
2140
2141 // Instead of adding one region at a time to the secondary_free_list,
2142 // we accumulate them in the local list and move them a few at a
2143 // time. This also cuts down on the number of notify_all() calls
2144 // we do during this process. We'll also append the local list when
2145 // _cleanup_list is empty (which means we just removed the last
2146 // region from the _cleanup_list).
2147 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2148 _cleanup_list.is_empty()) {
2149 if (G1ConcRegionFreeingVerbose) {
2150 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2151 "appending %u entries to the secondary_free_list, "
2152 "cleanup list still has %u entries",
2153 tmp_free_list.length(),
2154 _cleanup_list.length());
2155 }
2156
2157 {
2158 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2159 g1h->secondary_free_list_add(&tmp_free_list);
2160 SecondaryFreeList_lock->notify_all();
2161 }
2162 #ifndef PRODUCT
2163 if (G1StressConcRegionFreeing) {
2164 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2165 os::sleep(Thread::current(), (jlong) 1, false);
2166 }
2167 }
2168 #endif
2169 }
2170 }
2171 assert(tmp_free_list.is_empty(), "post-condition");
2172 }
2173
2174 // Supporting Object and Oop closures for reference discovery
2175 // and processing in during marking
2176
2177 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2178 HeapWord* addr = (HeapWord*)obj;
2179 return addr != NULL &&
2180 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2181 }
2182
2183 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2184 // Uses the CMTask associated with a worker thread (for serial reference
2185 // processing the CMTask for worker 0 is used) to preserve (mark) and
2186 // trace referent objects.
2187 //
2188 // Using the CMTask and embedded local queues avoids having the worker
2189 // threads operating on the global mark stack. This reduces the risk
2190 // of overflowing the stack - which we would rather avoid at this late
2191 // state. Also using the tasks' local queues removes the potential
2192 // of the workers interfering with each other that could occur if
2193 // operating on the global stack.
2194
2195 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2196 ConcurrentMark* _cm;
2197 CMTask* _task;
2198 int _ref_counter_limit;
2199 int _ref_counter;
2200 bool _is_serial;
2201 public:
2202 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2203 _cm(cm), _task(task), _is_serial(is_serial),
2204 _ref_counter_limit(G1RefProcDrainInterval) {
2205 assert(_ref_counter_limit > 0, "sanity");
2206 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2207 _ref_counter = _ref_counter_limit;
2208 }
2209
2210 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2211 virtual void do_oop( oop* p) { do_oop_work(p); }
2212
2213 template <class T> void do_oop_work(T* p) {
2214 if (!_cm->has_overflown()) {
2215 oop obj = oopDesc::load_decode_heap_oop(p);
2216 if (_cm->verbose_high()) {
2217 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2218 "*"PTR_FORMAT" = "PTR_FORMAT,
2219 _task->worker_id(), p2i(p), p2i((void*) obj));
2220 }
2221
2222 _task->deal_with_reference(obj);
2223 _ref_counter--;
2224
2225 if (_ref_counter == 0) {
2226 // We have dealt with _ref_counter_limit references, pushing them
2227 // and objects reachable from them on to the local stack (and
2228 // possibly the global stack). Call CMTask::do_marking_step() to
2229 // process these entries.
2230 //
2231 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2232 // there's nothing more to do (i.e. we're done with the entries that
2233 // were pushed as a result of the CMTask::deal_with_reference() calls
2234 // above) or we overflow.
2235 //
2236 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2237 // flag while there may still be some work to do. (See the comment at
2238 // the beginning of CMTask::do_marking_step() for those conditions -
2239 // one of which is reaching the specified time target.) It is only
2240 // when CMTask::do_marking_step() returns without setting the
2241 // has_aborted() flag that the marking step has completed.
2242 do {
2243 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2244 _task->do_marking_step(mark_step_duration_ms,
2245 false /* do_termination */,
2246 _is_serial);
2247 } while (_task->has_aborted() && !_cm->has_overflown());
2248 _ref_counter = _ref_counter_limit;
2249 }
2250 } else {
2251 if (_cm->verbose_high()) {
2252 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2253 }
2254 }
2255 }
2256 };
2257
2258 // 'Drain' oop closure used by both serial and parallel reference processing.
2259 // Uses the CMTask associated with a given worker thread (for serial
2260 // reference processing the CMtask for worker 0 is used). Calls the
2261 // do_marking_step routine, with an unbelievably large timeout value,
2262 // to drain the marking data structures of the remaining entries
2263 // added by the 'keep alive' oop closure above.
2264
2265 class G1CMDrainMarkingStackClosure: public VoidClosure {
2266 ConcurrentMark* _cm;
2267 CMTask* _task;
2268 bool _is_serial;
2269 public:
2270 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2271 _cm(cm), _task(task), _is_serial(is_serial) {
2272 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2273 }
2274
2275 void do_void() {
2276 do {
2277 if (_cm->verbose_high()) {
2278 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2279 _task->worker_id(), BOOL_TO_STR(_is_serial));
2280 }
2281
2282 // We call CMTask::do_marking_step() to completely drain the local
2283 // and global marking stacks of entries pushed by the 'keep alive'
2284 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2285 //
2286 // CMTask::do_marking_step() is called in a loop, which we'll exit
2287 // if there's nothing more to do (i.e. we've completely drained the
2288 // entries that were pushed as a a result of applying the 'keep alive'
2289 // closure to the entries on the discovered ref lists) or we overflow
2290 // the global marking stack.
2291 //
2292 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2293 // flag while there may still be some work to do. (See the comment at
2294 // the beginning of CMTask::do_marking_step() for those conditions -
2295 // one of which is reaching the specified time target.) It is only
2296 // when CMTask::do_marking_step() returns without setting the
2297 // has_aborted() flag that the marking step has completed.
2298
2299 _task->do_marking_step(1000000000.0 /* something very large */,
2300 true /* do_termination */,
2301 _is_serial);
2302 } while (_task->has_aborted() && !_cm->has_overflown());
2303 }
2304 };
2305
2306 // Implementation of AbstractRefProcTaskExecutor for parallel
2307 // reference processing at the end of G1 concurrent marking
2308
2309 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2310 private:
2311 G1CollectedHeap* _g1h;
2312 ConcurrentMark* _cm;
2313 WorkGang* _workers;
2314 uint _active_workers;
2315
2316 public:
2317 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2318 ConcurrentMark* cm,
2319 WorkGang* workers,
2320 uint n_workers) :
2321 _g1h(g1h), _cm(cm),
2322 _workers(workers), _active_workers(n_workers) { }
2323
2324 // Executes the given task using concurrent marking worker threads.
2325 virtual void execute(ProcessTask& task);
2326 virtual void execute(EnqueueTask& task);
2327 };
2328
2329 class G1CMRefProcTaskProxy: public AbstractGangTask {
2330 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2331 ProcessTask& _proc_task;
2332 G1CollectedHeap* _g1h;
2333 ConcurrentMark* _cm;
2334
2335 public:
2336 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2337 G1CollectedHeap* g1h,
2338 ConcurrentMark* cm) :
2339 AbstractGangTask("Process reference objects in parallel"),
2340 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2341 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2342 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2343 }
2344
2345 virtual void work(uint worker_id) {
2346 ResourceMark rm;
2347 HandleMark hm;
2348 CMTask* task = _cm->task(worker_id);
2349 G1CMIsAliveClosure g1_is_alive(_g1h);
2350 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2351 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2352
2353 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2354 }
2355 };
2356
2357 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2358 assert(_workers != NULL, "Need parallel worker threads.");
2359 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2360
2361 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2362
2363 // We need to reset the concurrency level before each
2364 // proxy task execution, so that the termination protocol
2365 // and overflow handling in CMTask::do_marking_step() knows
2366 // how many workers to wait for.
2367 _cm->set_concurrency(_active_workers);
2368 _g1h->set_par_threads(_active_workers);
2369 _workers->run_task(&proc_task_proxy);
2370 _g1h->set_par_threads(0);
2371 }
2372
2373 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2374 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2375 EnqueueTask& _enq_task;
2376
2377 public:
2378 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2379 AbstractGangTask("Enqueue reference objects in parallel"),
2380 _enq_task(enq_task) { }
2381
2382 virtual void work(uint worker_id) {
2383 _enq_task.work(worker_id);
2384 }
2385 };
2386
2387 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2388 assert(_workers != NULL, "Need parallel worker threads.");
2389 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2390
2391 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2392
2393 // Not strictly necessary but...
2394 //
2395 // We need to reset the concurrency level before each
2396 // proxy task execution, so that the termination protocol
2397 // and overflow handling in CMTask::do_marking_step() knows
2398 // how many workers to wait for.
2399 _cm->set_concurrency(_active_workers);
2400 _g1h->set_par_threads(_active_workers);
2401 _workers->run_task(&enq_task_proxy);
2402 _g1h->set_par_threads(0);
2403 }
2404
2405 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2406 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2407 }
2408
2409 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2410 if (has_overflown()) {
2411 // Skip processing the discovered references if we have
2412 // overflown the global marking stack. Reference objects
2413 // only get discovered once so it is OK to not
2414 // de-populate the discovered reference lists. We could have,
2415 // but the only benefit would be that, when marking restarts,
2416 // less reference objects are discovered.
2417 return;
2418 }
2419
2420 ResourceMark rm;
2421 HandleMark hm;
2422
2423 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2424
2425 // Is alive closure.
2426 G1CMIsAliveClosure g1_is_alive(g1h);
2427
2428 // Inner scope to exclude the cleaning of the string and symbol
2429 // tables from the displayed time.
2430 {
2431 G1CMTraceTime t("GC ref-proc", G1Log::finer());
2432
2433 ReferenceProcessor* rp = g1h->ref_processor_cm();
2434
2435 // See the comment in G1CollectedHeap::ref_processing_init()
2436 // about how reference processing currently works in G1.
2437
2438 // Set the soft reference policy
2439 rp->setup_policy(clear_all_soft_refs);
2440 assert(_markStack.isEmpty(), "mark stack should be empty");
2441
2442 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2443 // in serial reference processing. Note these closures are also
2444 // used for serially processing (by the the current thread) the
2445 // JNI references during parallel reference processing.
2446 //
2447 // These closures do not need to synchronize with the worker
2448 // threads involved in parallel reference processing as these
2449 // instances are executed serially by the current thread (e.g.
2450 // reference processing is not multi-threaded and is thus
2451 // performed by the current thread instead of a gang worker).
2452 //
2453 // The gang tasks involved in parallel reference processing create
2454 // their own instances of these closures, which do their own
2455 // synchronization among themselves.
2456 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2457 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2458
2459 // We need at least one active thread. If reference processing
2460 // is not multi-threaded we use the current (VMThread) thread,
2461 // otherwise we use the work gang from the G1CollectedHeap and
2462 // we utilize all the worker threads we can.
2463 bool processing_is_mt = rp->processing_is_mt();
2464 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2465 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2466
2467 // Parallel processing task executor.
2468 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2469 g1h->workers(), active_workers);
2470 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2471
2472 // Set the concurrency level. The phase was already set prior to
2473 // executing the remark task.
2474 set_concurrency(active_workers);
2475
2476 // Set the degree of MT processing here. If the discovery was done MT,
2477 // the number of threads involved during discovery could differ from
2478 // the number of active workers. This is OK as long as the discovered
2479 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2480 rp->set_active_mt_degree(active_workers);
2481
2482 // Process the weak references.
2483 const ReferenceProcessorStats& stats =
2484 rp->process_discovered_references(&g1_is_alive,
2485 &g1_keep_alive,
2486 &g1_drain_mark_stack,
2487 executor,
2488 g1h->gc_timer_cm(),
2489 concurrent_gc_id());
2490 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2491
2492 // The do_oop work routines of the keep_alive and drain_marking_stack
2493 // oop closures will set the has_overflown flag if we overflow the
2494 // global marking stack.
2495
2496 assert(_markStack.overflow() || _markStack.isEmpty(),
2497 "mark stack should be empty (unless it overflowed)");
2498
2499 if (_markStack.overflow()) {
2500 // This should have been done already when we tried to push an
2501 // entry on to the global mark stack. But let's do it again.
2502 set_has_overflown();
2503 }
2504
2505 assert(rp->num_q() == active_workers, "why not");
2506
2507 rp->enqueue_discovered_references(executor);
2508
2509 rp->verify_no_references_recorded();
2510 assert(!rp->discovery_enabled(), "Post condition");
2511 }
2512
2513 if (has_overflown()) {
2514 // We can not trust g1_is_alive if the marking stack overflowed
2515 return;
2516 }
2517
2518 assert(_markStack.isEmpty(), "Marking should have completed");
2519
2520 // Unload Klasses, String, Symbols, Code Cache, etc.
2521 {
2522 G1CMTraceTime trace("Unloading", G1Log::finer());
2523
2524 if (ClassUnloadingWithConcurrentMark) {
2525 bool purged_classes;
2526
2527 {
2528 G1CMTraceTime trace("System Dictionary Unloading", G1Log::finest());
2529 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2530 }
2531
2532 {
2533 G1CMTraceTime trace("Parallel Unloading", G1Log::finest());
2534 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2535 }
2536 }
2537
2538 if (G1StringDedup::is_enabled()) {
2539 G1CMTraceTime trace("String Deduplication Unlink", G1Log::finest());
2540 G1StringDedup::unlink(&g1_is_alive);
2541 }
2542 }
2543 }
2544
2545 void ConcurrentMark::swapMarkBitMaps() {
2546 CMBitMapRO* temp = _prevMarkBitMap;
2547 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2548 _nextMarkBitMap = (CMBitMap*) temp;
2549 }
2550
2551 // Closure for marking entries in SATB buffers.
2552 class CMSATBBufferClosure : public SATBBufferClosure {
2553 private:
2554 CMTask* _task;
2555 G1CollectedHeap* _g1h;
2556
2557 // This is very similar to CMTask::deal_with_reference, but with
2558 // more relaxed requirements for the argument, so this must be more
2559 // circumspect about treating the argument as an object.
2560 void do_entry(void* entry) const {
2561 _task->increment_refs_reached();
2562 HeapRegion* hr = _g1h->heap_region_containing_raw(entry);
2563 if (entry < hr->next_top_at_mark_start()) {
2564 // Until we get here, we don't know whether entry refers to a valid
2565 // object; it could instead have been a stale reference.
2566 oop obj = static_cast<oop>(entry);
2567 assert(obj->is_oop(true /* ignore mark word */),
2568 err_msg("Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj)));
2569 _task->make_reference_grey(obj, hr);
2570 }
2571 }
2572
2573 public:
2574 CMSATBBufferClosure(CMTask* task, G1CollectedHeap* g1h)
2575 : _task(task), _g1h(g1h) { }
2576
2577 virtual void do_buffer(void** buffer, size_t size) {
2578 for (size_t i = 0; i < size; ++i) {
2579 do_entry(buffer[i]);
2580 }
2581 }
2582 };
2583
2584 class G1RemarkThreadsClosure : public ThreadClosure {
2585 CMSATBBufferClosure _cm_satb_cl;
2586 G1CMOopClosure _cm_cl;
2587 MarkingCodeBlobClosure _code_cl;
2588 int _thread_parity;
2589
2590 public:
2591 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task) :
2592 _cm_satb_cl(task, g1h),
2593 _cm_cl(g1h, g1h->concurrent_mark(), task),
2594 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2595 _thread_parity(Threads::thread_claim_parity()) {}
2596
2597 void do_thread(Thread* thread) {
2598 if (thread->is_Java_thread()) {
2599 if (thread->claim_oops_do(true, _thread_parity)) {
2600 JavaThread* jt = (JavaThread*)thread;
2601
2602 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2603 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2604 // * Alive if on the stack of an executing method
2605 // * Weakly reachable otherwise
2606 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2607 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2608 jt->nmethods_do(&_code_cl);
2609
2610 jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
2611 }
2612 } else if (thread->is_VM_thread()) {
2613 if (thread->claim_oops_do(true, _thread_parity)) {
2614 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
2615 }
2616 }
2617 }
2618 };
2619
2620 class CMRemarkTask: public AbstractGangTask {
2621 private:
2622 ConcurrentMark* _cm;
2623 public:
2624 void work(uint worker_id) {
2625 // Since all available tasks are actually started, we should
2626 // only proceed if we're supposed to be active.
2627 if (worker_id < _cm->active_tasks()) {
2628 CMTask* task = _cm->task(worker_id);
2629 task->record_start_time();
2630 {
2631 ResourceMark rm;
2632 HandleMark hm;
2633
2634 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
2635 Threads::threads_do(&threads_f);
2636 }
2637
2638 do {
2639 task->do_marking_step(1000000000.0 /* something very large */,
2640 true /* do_termination */,
2641 false /* is_serial */);
2642 } while (task->has_aborted() && !_cm->has_overflown());
2643 // If we overflow, then we do not want to restart. We instead
2644 // want to abort remark and do concurrent marking again.
2645 task->record_end_time();
2646 }
2647 }
2648
2649 CMRemarkTask(ConcurrentMark* cm, uint active_workers) :
2650 AbstractGangTask("Par Remark"), _cm(cm) {
2651 _cm->terminator()->reset_for_reuse(active_workers);
2652 }
2653 };
2654
2655 void ConcurrentMark::checkpointRootsFinalWork() {
2656 ResourceMark rm;
2657 HandleMark hm;
2658 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2659
2660 G1CMTraceTime trace("Finalize Marking", G1Log::finer());
2661
2662 g1h->ensure_parsability(false);
2663
2664 // this is remark, so we'll use up all active threads
2665 uint active_workers = g1h->workers()->active_workers();
2666 if (active_workers == 0) {
2667 assert(active_workers > 0, "Should have been set earlier");
2668 active_workers = (uint) ParallelGCThreads;
2669 g1h->workers()->set_active_workers(active_workers);
2670 }
2671 set_concurrency_and_phase(active_workers, false /* concurrent */);
2672 // Leave _parallel_marking_threads at it's
2673 // value originally calculated in the ConcurrentMark
2674 // constructor and pass values of the active workers
2675 // through the gang in the task.
2676
2677 {
2678 StrongRootsScope srs(active_workers);
2679
2680 CMRemarkTask remarkTask(this, active_workers);
2681 // We will start all available threads, even if we decide that the
2682 // active_workers will be fewer. The extra ones will just bail out
2683 // immediately.
2684 g1h->set_par_threads(active_workers);
2685 g1h->workers()->run_task(&remarkTask);
2686 g1h->set_par_threads(0);
2687 }
2688
2689 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2690 guarantee(has_overflown() ||
2691 satb_mq_set.completed_buffers_num() == 0,
2692 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2693 BOOL_TO_STR(has_overflown()),
2694 satb_mq_set.completed_buffers_num()));
2695
2696 print_stats();
2697 }
2698
2699 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2700 // Note we are overriding the read-only view of the prev map here, via
2701 // the cast.
2702 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2703 }
2704
2705 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2706 _nextMarkBitMap->clearRange(mr);
2707 }
2708
2709 HeapRegion*
2710 ConcurrentMark::claim_region(uint worker_id) {
2711 // "checkpoint" the finger
2712 HeapWord* finger = _finger;
2713
2714 // _heap_end will not change underneath our feet; it only changes at
2715 // yield points.
2716 while (finger < _heap_end) {
2717 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2718
2719 // Note on how this code handles humongous regions. In the
2720 // normal case the finger will reach the start of a "starts
2721 // humongous" (SH) region. Its end will either be the end of the
2722 // last "continues humongous" (CH) region in the sequence, or the
2723 // standard end of the SH region (if the SH is the only region in
2724 // the sequence). That way claim_region() will skip over the CH
2725 // regions. However, there is a subtle race between a CM thread
2726 // executing this method and a mutator thread doing a humongous
2727 // object allocation. The two are not mutually exclusive as the CM
2728 // thread does not need to hold the Heap_lock when it gets
2729 // here. So there is a chance that claim_region() will come across
2730 // a free region that's in the progress of becoming a SH or a CH
2731 // region. In the former case, it will either
2732 // a) Miss the update to the region's end, in which case it will
2733 // visit every subsequent CH region, will find their bitmaps
2734 // empty, and do nothing, or
2735 // b) Will observe the update of the region's end (in which case
2736 // it will skip the subsequent CH regions).
2737 // If it comes across a region that suddenly becomes CH, the
2738 // scenario will be similar to b). So, the race between
2739 // claim_region() and a humongous object allocation might force us
2740 // to do a bit of unnecessary work (due to some unnecessary bitmap
2741 // iterations) but it should not introduce and correctness issues.
2742 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2743
2744 // Above heap_region_containing_raw may return NULL as we always scan claim
2745 // until the end of the heap. In this case, just jump to the next region.
2746 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2747
2748 // Is the gap between reading the finger and doing the CAS too long?
2749 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2750 if (res == finger && curr_region != NULL) {
2751 // we succeeded
2752 HeapWord* bottom = curr_region->bottom();
2753 HeapWord* limit = curr_region->next_top_at_mark_start();
2754
2755 if (verbose_low()) {
2756 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2757 "["PTR_FORMAT", "PTR_FORMAT"), "
2758 "limit = "PTR_FORMAT,
2759 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2760 }
2761
2762 // notice that _finger == end cannot be guaranteed here since,
2763 // someone else might have moved the finger even further
2764 assert(_finger >= end, "the finger should have moved forward");
2765
2766 if (verbose_low()) {
2767 gclog_or_tty->print_cr("[%u] we were successful with region = "
2768 PTR_FORMAT, worker_id, p2i(curr_region));
2769 }
2770
2771 if (limit > bottom) {
2772 if (verbose_low()) {
2773 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2774 "returning it ", worker_id, p2i(curr_region));
2775 }
2776 return curr_region;
2777 } else {
2778 assert(limit == bottom,
2779 "the region limit should be at bottom");
2780 if (verbose_low()) {
2781 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
2782 "returning NULL", worker_id, p2i(curr_region));
2783 }
2784 // we return NULL and the caller should try calling
2785 // claim_region() again.
2786 return NULL;
2787 }
2788 } else {
2789 assert(_finger > finger, "the finger should have moved forward");
2790 if (verbose_low()) {
2791 if (curr_region == NULL) {
2792 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
2793 "global finger = "PTR_FORMAT", "
2794 "our finger = "PTR_FORMAT,
2795 worker_id, p2i(_finger), p2i(finger));
2796 } else {
2797 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2798 "global finger = "PTR_FORMAT", "
2799 "our finger = "PTR_FORMAT,
2800 worker_id, p2i(_finger), p2i(finger));
2801 }
2802 }
2803
2804 // read it again
2805 finger = _finger;
2806 }
2807 }
2808
2809 return NULL;
2810 }
2811
2812 #ifndef PRODUCT
2813 enum VerifyNoCSetOopsPhase {
2814 VerifyNoCSetOopsStack,
2815 VerifyNoCSetOopsQueues
2816 };
2817
2818 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
2819 private:
2820 G1CollectedHeap* _g1h;
2821 VerifyNoCSetOopsPhase _phase;
2822 int _info;
2823
2824 const char* phase_str() {
2825 switch (_phase) {
2826 case VerifyNoCSetOopsStack: return "Stack";
2827 case VerifyNoCSetOopsQueues: return "Queue";
2828 default: ShouldNotReachHere();
2829 }
2830 return NULL;
2831 }
2832
2833 void do_object_work(oop obj) {
2834 guarantee(!_g1h->obj_in_cs(obj),
2835 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2836 p2i((void*) obj), phase_str(), _info));
2837 }
2838
2839 public:
2840 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2841
2842 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2843 _phase = phase;
2844 _info = info;
2845 }
2846
2847 virtual void do_oop(oop* p) {
2848 oop obj = oopDesc::load_decode_heap_oop(p);
2849 do_object_work(obj);
2850 }
2851
2852 virtual void do_oop(narrowOop* p) {
2853 // We should not come across narrow oops while scanning marking
2854 // stacks
2855 ShouldNotReachHere();
2856 }
2857
2858 virtual void do_object(oop obj) {
2859 do_object_work(obj);
2860 }
2861 };
2862
2863 void ConcurrentMark::verify_no_cset_oops() {
2864 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2865 if (!G1CollectedHeap::heap()->mark_in_progress()) {
2866 return;
2867 }
2868
2869 VerifyNoCSetOopsClosure cl;
2870
2871 // Verify entries on the global mark stack
2872 cl.set_phase(VerifyNoCSetOopsStack);
2873 _markStack.oops_do(&cl);
2874
2875 // Verify entries on the task queues
2876 for (uint i = 0; i < _max_worker_id; i += 1) {
2877 cl.set_phase(VerifyNoCSetOopsQueues, i);
2878 CMTaskQueue* queue = _task_queues->queue(i);
2879 queue->oops_do(&cl);
2880 }
2881
2882 // Verify the global finger
2883 HeapWord* global_finger = finger();
2884 if (global_finger != NULL && global_finger < _heap_end) {
2885 // The global finger always points to a heap region boundary. We
2886 // use heap_region_containing_raw() to get the containing region
2887 // given that the global finger could be pointing to a free region
2888 // which subsequently becomes continues humongous. If that
2889 // happens, heap_region_containing() will return the bottom of the
2890 // corresponding starts humongous region and the check below will
2891 // not hold any more.
2892 // Since we always iterate over all regions, we might get a NULL HeapRegion
2893 // here.
2894 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
2895 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
2896 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
2897 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
2898 }
2899
2900 // Verify the task fingers
2901 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
2902 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
2903 CMTask* task = _tasks[i];
2904 HeapWord* task_finger = task->finger();
2905 if (task_finger != NULL && task_finger < _heap_end) {
2906 // See above note on the global finger verification.
2907 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
2908 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
2909 !task_hr->in_collection_set(),
2910 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
2911 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
2912 }
2913 }
2914 }
2915 #endif // PRODUCT
2916
2917 // Aggregate the counting data that was constructed concurrently
2918 // with marking.
2919 class AggregateCountDataHRClosure: public HeapRegionClosure {
2920 G1CollectedHeap* _g1h;
2921 ConcurrentMark* _cm;
2922 CardTableModRefBS* _ct_bs;
2923 BitMap* _cm_card_bm;
2924 uint _max_worker_id;
2925
2926 public:
2927 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
2928 BitMap* cm_card_bm,
2929 uint max_worker_id) :
2930 _g1h(g1h), _cm(g1h->concurrent_mark()),
2931 _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
2932 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
2933
2934 bool doHeapRegion(HeapRegion* hr) {
2935 if (hr->is_continues_humongous()) {
2936 // We will ignore these here and process them when their
2937 // associated "starts humongous" region is processed.
2938 // Note that we cannot rely on their associated
2939 // "starts humongous" region to have their bit set to 1
2940 // since, due to the region chunking in the parallel region
2941 // iteration, a "continues humongous" region might be visited
2942 // before its associated "starts humongous".
2943 return false;
2944 }
2945
2946 HeapWord* start = hr->bottom();
2947 HeapWord* limit = hr->next_top_at_mark_start();
2948 HeapWord* end = hr->end();
2949
2950 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
2951 err_msg("Preconditions not met - "
2952 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
2953 "top: "PTR_FORMAT", end: "PTR_FORMAT,
2954 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
2955
2956 assert(hr->next_marked_bytes() == 0, "Precondition");
2957
2958 if (start == limit) {
2959 // NTAMS of this region has not been set so nothing to do.
2960 return false;
2961 }
2962
2963 // 'start' should be in the heap.
2964 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
2965 // 'end' *may* be just beyond the end of the heap (if hr is the last region)
2966 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
2967
2968 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
2969 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
2970 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
2971
2972 // If ntams is not card aligned then we bump card bitmap index
2973 // for limit so that we get the all the cards spanned by
2974 // the object ending at ntams.
2975 // Note: if this is the last region in the heap then ntams
2976 // could be actually just beyond the end of the the heap;
2977 // limit_idx will then correspond to a (non-existent) card
2978 // that is also outside the heap.
2979 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
2980 limit_idx += 1;
2981 }
2982
2983 assert(limit_idx <= end_idx, "or else use atomics");
2984
2985 // Aggregate the "stripe" in the count data associated with hr.
2986 uint hrm_index = hr->hrm_index();
2987 size_t marked_bytes = 0;
2988
2989 for (uint i = 0; i < _max_worker_id; i += 1) {
2990 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
2991 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
2992
2993 // Fetch the marked_bytes in this region for task i and
2994 // add it to the running total for this region.
2995 marked_bytes += marked_bytes_array[hrm_index];
2996
2997 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
2998 // into the global card bitmap.
2999 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3000
3001 while (scan_idx < limit_idx) {
3002 assert(task_card_bm->at(scan_idx) == true, "should be");
3003 _cm_card_bm->set_bit(scan_idx);
3004 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3005
3006 // BitMap::get_next_one_offset() can handle the case when
3007 // its left_offset parameter is greater than its right_offset
3008 // parameter. It does, however, have an early exit if
3009 // left_offset == right_offset. So let's limit the value
3010 // passed in for left offset here.
3011 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3012 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3013 }
3014 }
3015
3016 // Update the marked bytes for this region.
3017 hr->add_to_marked_bytes(marked_bytes);
3018
3019 // Next heap region
3020 return false;
3021 }
3022 };
3023
3024 class G1AggregateCountDataTask: public AbstractGangTask {
3025 protected:
3026 G1CollectedHeap* _g1h;
3027 ConcurrentMark* _cm;
3028 BitMap* _cm_card_bm;
3029 uint _max_worker_id;
3030 uint _active_workers;
3031 HeapRegionClaimer _hrclaimer;
3032
3033 public:
3034 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3035 ConcurrentMark* cm,
3036 BitMap* cm_card_bm,
3037 uint max_worker_id,
3038 uint n_workers) :
3039 AbstractGangTask("Count Aggregation"),
3040 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3041 _max_worker_id(max_worker_id),
3042 _active_workers(n_workers),
3043 _hrclaimer(_active_workers) {
3044 }
3045
3046 void work(uint worker_id) {
3047 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3048
3049 _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer);
3050 }
3051 };
3052
3053
3054 void ConcurrentMark::aggregate_count_data() {
3055 uint n_workers = _g1h->workers()->active_workers();
3056
3057 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3058 _max_worker_id, n_workers);
3059
3060 _g1h->set_par_threads(n_workers);
3061 _g1h->workers()->run_task(&g1_par_agg_task);
3062 _g1h->set_par_threads(0);
3063 }
3064
3065 // Clear the per-worker arrays used to store the per-region counting data
3066 void ConcurrentMark::clear_all_count_data() {
3067 // Clear the global card bitmap - it will be filled during
3068 // liveness count aggregation (during remark) and the
3069 // final counting task.
3070 _card_bm.clear();
3071
3072 // Clear the global region bitmap - it will be filled as part
3073 // of the final counting task.
3074 _region_bm.clear();
3075
3076 uint max_regions = _g1h->max_regions();
3077 assert(_max_worker_id > 0, "uninitialized");
3078
3079 for (uint i = 0; i < _max_worker_id; i += 1) {
3080 BitMap* task_card_bm = count_card_bitmap_for(i);
3081 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3082
3083 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3084 assert(marked_bytes_array != NULL, "uninitialized");
3085
3086 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3087 task_card_bm->clear();
3088 }
3089 }
3090
3091 void ConcurrentMark::print_stats() {
3092 if (verbose_stats()) {
3093 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3094 for (size_t i = 0; i < _active_tasks; ++i) {
3095 _tasks[i]->print_stats();
3096 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3097 }
3098 }
3099 }
3100
3101 // abandon current marking iteration due to a Full GC
3102 void ConcurrentMark::abort() {
3103 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3104 // concurrent bitmap clearing.
3105 _nextMarkBitMap->clearAll();
3106
3107 // Note we cannot clear the previous marking bitmap here
3108 // since VerifyDuringGC verifies the objects marked during
3109 // a full GC against the previous bitmap.
3110
3111 // Clear the liveness counting data
3112 clear_all_count_data();
3113 // Empty mark stack
3114 reset_marking_state();
3115 for (uint i = 0; i < _max_worker_id; ++i) {
3116 _tasks[i]->clear_region_fields();
3117 }
3118 _first_overflow_barrier_sync.abort();
3119 _second_overflow_barrier_sync.abort();
3120 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3121 if (!gc_id.is_undefined()) {
3122 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3123 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3124 _aborted_gc_id = gc_id;
3125 }
3126 _has_aborted = true;
3127
3128 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3129 satb_mq_set.abandon_partial_marking();
3130 // This can be called either during or outside marking, we'll read
3131 // the expected_active value from the SATB queue set.
3132 satb_mq_set.set_active_all_threads(
3133 false, /* new active value */
3134 satb_mq_set.is_active() /* expected_active */);
3135
3136 _g1h->trace_heap_after_concurrent_cycle();
3137 _g1h->register_concurrent_cycle_end();
3138 }
3139
3140 const GCId& ConcurrentMark::concurrent_gc_id() {
3141 if (has_aborted()) {
3142 return _aborted_gc_id;
3143 }
3144 return _g1h->gc_tracer_cm()->gc_id();
3145 }
3146
3147 static void print_ms_time_info(const char* prefix, const char* name,
3148 NumberSeq& ns) {
3149 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3150 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3151 if (ns.num() > 0) {
3152 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3153 prefix, ns.sd(), ns.maximum());
3154 }
3155 }
3156
3157 void ConcurrentMark::print_summary_info() {
3158 gclog_or_tty->print_cr(" Concurrent marking:");
3159 print_ms_time_info(" ", "init marks", _init_times);
3160 print_ms_time_info(" ", "remarks", _remark_times);
3161 {
3162 print_ms_time_info(" ", "final marks", _remark_mark_times);
3163 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3164
3165 }
3166 print_ms_time_info(" ", "cleanups", _cleanup_times);
3167 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3168 _total_counting_time,
3169 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3170 (double)_cleanup_times.num()
3171 : 0.0));
3172 if (G1ScrubRemSets) {
3173 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3174 _total_rs_scrub_time,
3175 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3176 (double)_cleanup_times.num()
3177 : 0.0));
3178 }
3179 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3180 (_init_times.sum() + _remark_times.sum() +
3181 _cleanup_times.sum())/1000.0);
3182 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3183 "(%8.2f s marking).",
3184 cmThread()->vtime_accum(),
3185 cmThread()->vtime_mark_accum());
3186 }
3187
3188 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3189 _parallel_workers->print_worker_threads_on(st);
3190 }
3191
3192 void ConcurrentMark::print_on_error(outputStream* st) const {
3193 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3194 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3195 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3196 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3197 }
3198
3199 // We take a break if someone is trying to stop the world.
3200 bool ConcurrentMark::do_yield_check(uint worker_id) {
3201 if (SuspendibleThreadSet::should_yield()) {
3202 if (worker_id == 0) {
3203 _g1h->g1_policy()->record_concurrent_pause();
3204 }
3205 SuspendibleThreadSet::yield();
3206 return true;
3207 } else {
3208 return false;
3209 }
3210 }
3211
3212 #ifndef PRODUCT
3213 // for debugging purposes
3214 void ConcurrentMark::print_finger() {
3215 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3216 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3217 for (uint i = 0; i < _max_worker_id; ++i) {
3218 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3219 }
3220 gclog_or_tty->cr();
3221 }
3222 #endif
3223
3224 template<bool scan>
3225 inline void CMTask::process_grey_object(oop obj) {
3226 assert(scan || obj->is_typeArray(), "Skipping scan of grey non-typeArray");
3227 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3228
3229 if (_cm->verbose_high()) {
3230 gclog_or_tty->print_cr("[%u] processing grey object " PTR_FORMAT,
3231 _worker_id, p2i((void*) obj));
3232 }
3233
3234 size_t obj_size = obj->size();
3235 _words_scanned += obj_size;
3236
3237 if (scan) {
3238 obj->oop_iterate(_cm_oop_closure);
3239 }
3240 statsOnly( ++_objs_scanned );
3241 check_limits();
3242 }
3243
3244 template void CMTask::process_grey_object<true>(oop);
3245 template void CMTask::process_grey_object<false>(oop);
3246
3247 // Closure for iteration over bitmaps
3248 class CMBitMapClosure : public BitMapClosure {
3249 private:
3250 // the bitmap that is being iterated over
3251 CMBitMap* _nextMarkBitMap;
3252 ConcurrentMark* _cm;
3253 CMTask* _task;
3254
3255 public:
3256 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3257 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3258
3259 bool do_bit(size_t offset) {
3260 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3261 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3262 assert( addr < _cm->finger(), "invariant");
3263
3264 statsOnly( _task->increase_objs_found_on_bitmap() );
3265 assert(addr >= _task->finger(), "invariant");
3266
3267 // We move that task's local finger along.
3268 _task->move_finger_to(addr);
3269
3270 _task->scan_object(oop(addr));
3271 // we only partially drain the local queue and global stack
3272 _task->drain_local_queue(true);
3273 _task->drain_global_stack(true);
3274
3275 // if the has_aborted flag has been raised, we need to bail out of
3276 // the iteration
3277 return !_task->has_aborted();
3278 }
3279 };
3280
3281 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3282 ConcurrentMark* cm,
3283 CMTask* task)
3284 : _g1h(g1h), _cm(cm), _task(task) {
3285 assert(_ref_processor == NULL, "should be initialized to NULL");
3286
3287 if (G1UseConcMarkReferenceProcessing) {
3288 _ref_processor = g1h->ref_processor_cm();
3289 assert(_ref_processor != NULL, "should not be NULL");
3290 }
3291 }
3292
3293 void CMTask::setup_for_region(HeapRegion* hr) {
3294 assert(hr != NULL,
3295 "claim_region() should have filtered out NULL regions");
3296 assert(!hr->is_continues_humongous(),
3297 "claim_region() should have filtered out continues humongous regions");
3298
3299 if (_cm->verbose_low()) {
3300 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3301 _worker_id, p2i(hr));
3302 }
3303
3304 _curr_region = hr;
3305 _finger = hr->bottom();
3306 update_region_limit();
3307 }
3308
3309 void CMTask::update_region_limit() {
3310 HeapRegion* hr = _curr_region;
3311 HeapWord* bottom = hr->bottom();
3312 HeapWord* limit = hr->next_top_at_mark_start();
3313
3314 if (limit == bottom) {
3315 if (_cm->verbose_low()) {
3316 gclog_or_tty->print_cr("[%u] found an empty region "
3317 "["PTR_FORMAT", "PTR_FORMAT")",
3318 _worker_id, p2i(bottom), p2i(limit));
3319 }
3320 // The region was collected underneath our feet.
3321 // We set the finger to bottom to ensure that the bitmap
3322 // iteration that will follow this will not do anything.
3323 // (this is not a condition that holds when we set the region up,
3324 // as the region is not supposed to be empty in the first place)
3325 _finger = bottom;
3326 } else if (limit >= _region_limit) {
3327 assert(limit >= _finger, "peace of mind");
3328 } else {
3329 assert(limit < _region_limit, "only way to get here");
3330 // This can happen under some pretty unusual circumstances. An
3331 // evacuation pause empties the region underneath our feet (NTAMS
3332 // at bottom). We then do some allocation in the region (NTAMS
3333 // stays at bottom), followed by the region being used as a GC
3334 // alloc region (NTAMS will move to top() and the objects
3335 // originally below it will be grayed). All objects now marked in
3336 // the region are explicitly grayed, if below the global finger,
3337 // and we do not need in fact to scan anything else. So, we simply
3338 // set _finger to be limit to ensure that the bitmap iteration
3339 // doesn't do anything.
3340 _finger = limit;
3341 }
3342
3343 _region_limit = limit;
3344 }
3345
3346 void CMTask::giveup_current_region() {
3347 assert(_curr_region != NULL, "invariant");
3348 if (_cm->verbose_low()) {
3349 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3350 _worker_id, p2i(_curr_region));
3351 }
3352 clear_region_fields();
3353 }
3354
3355 void CMTask::clear_region_fields() {
3356 // Values for these three fields that indicate that we're not
3357 // holding on to a region.
3358 _curr_region = NULL;
3359 _finger = NULL;
3360 _region_limit = NULL;
3361 }
3362
3363 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3364 if (cm_oop_closure == NULL) {
3365 assert(_cm_oop_closure != NULL, "invariant");
3366 } else {
3367 assert(_cm_oop_closure == NULL, "invariant");
3368 }
3369 _cm_oop_closure = cm_oop_closure;
3370 }
3371
3372 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3373 guarantee(nextMarkBitMap != NULL, "invariant");
3374
3375 if (_cm->verbose_low()) {
3376 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3377 }
3378
3379 _nextMarkBitMap = nextMarkBitMap;
3380 clear_region_fields();
3381
3382 _calls = 0;
3383 _elapsed_time_ms = 0.0;
3384 _termination_time_ms = 0.0;
3385 _termination_start_time_ms = 0.0;
3386
3387 #if _MARKING_STATS_
3388 _aborted = 0;
3389 _aborted_overflow = 0;
3390 _aborted_cm_aborted = 0;
3391 _aborted_yield = 0;
3392 _aborted_timed_out = 0;
3393 _aborted_satb = 0;
3394 _aborted_termination = 0;
3395 _steal_attempts = 0;
3396 _steals = 0;
3397 _local_pushes = 0;
3398 _local_pops = 0;
3399 _local_max_size = 0;
3400 _objs_scanned = 0;
3401 _global_pushes = 0;
3402 _global_pops = 0;
3403 _global_max_size = 0;
3404 _global_transfers_to = 0;
3405 _global_transfers_from = 0;
3406 _regions_claimed = 0;
3407 _objs_found_on_bitmap = 0;
3408 _satb_buffers_processed = 0;
3409 #endif // _MARKING_STATS_
3410 }
3411
3412 bool CMTask::should_exit_termination() {
3413 regular_clock_call();
3414 // This is called when we are in the termination protocol. We should
3415 // quit if, for some reason, this task wants to abort or the global
3416 // stack is not empty (this means that we can get work from it).
3417 return !_cm->mark_stack_empty() || has_aborted();
3418 }
3419
3420 void CMTask::reached_limit() {
3421 assert(_words_scanned >= _words_scanned_limit ||
3422 _refs_reached >= _refs_reached_limit ,
3423 "shouldn't have been called otherwise");
3424 regular_clock_call();
3425 }
3426
3427 void CMTask::regular_clock_call() {
3428 if (has_aborted()) return;
3429
3430 // First, we need to recalculate the words scanned and refs reached
3431 // limits for the next clock call.
3432 recalculate_limits();
3433
3434 // During the regular clock call we do the following
3435
3436 // (1) If an overflow has been flagged, then we abort.
3437 if (_cm->has_overflown()) {
3438 set_has_aborted();
3439 return;
3440 }
3441
3442 // If we are not concurrent (i.e. we're doing remark) we don't need
3443 // to check anything else. The other steps are only needed during
3444 // the concurrent marking phase.
3445 if (!concurrent()) return;
3446
3447 // (2) If marking has been aborted for Full GC, then we also abort.
3448 if (_cm->has_aborted()) {
3449 set_has_aborted();
3450 statsOnly( ++_aborted_cm_aborted );
3451 return;
3452 }
3453
3454 double curr_time_ms = os::elapsedVTime() * 1000.0;
3455
3456 // (3) If marking stats are enabled, then we update the step history.
3457 #if _MARKING_STATS_
3458 if (_words_scanned >= _words_scanned_limit) {
3459 ++_clock_due_to_scanning;
3460 }
3461 if (_refs_reached >= _refs_reached_limit) {
3462 ++_clock_due_to_marking;
3463 }
3464
3465 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3466 _interval_start_time_ms = curr_time_ms;
3467 _all_clock_intervals_ms.add(last_interval_ms);
3468
3469 if (_cm->verbose_medium()) {
3470 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3471 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3472 _worker_id, last_interval_ms,
3473 _words_scanned,
3474 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3475 _refs_reached,
3476 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3477 }
3478 #endif // _MARKING_STATS_
3479
3480 // (4) We check whether we should yield. If we have to, then we abort.
3481 if (SuspendibleThreadSet::should_yield()) {
3482 // We should yield. To do this we abort the task. The caller is
3483 // responsible for yielding.
3484 set_has_aborted();
3485 statsOnly( ++_aborted_yield );
3486 return;
3487 }
3488
3489 // (5) We check whether we've reached our time quota. If we have,
3490 // then we abort.
3491 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3492 if (elapsed_time_ms > _time_target_ms) {
3493 set_has_aborted();
3494 _has_timed_out = true;
3495 statsOnly( ++_aborted_timed_out );
3496 return;
3497 }
3498
3499 // (6) Finally, we check whether there are enough completed STAB
3500 // buffers available for processing. If there are, we abort.
3501 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3502 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3503 if (_cm->verbose_low()) {
3504 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3505 _worker_id);
3506 }
3507 // we do need to process SATB buffers, we'll abort and restart
3508 // the marking task to do so
3509 set_has_aborted();
3510 statsOnly( ++_aborted_satb );
3511 return;
3512 }
3513 }
3514
3515 void CMTask::recalculate_limits() {
3516 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3517 _words_scanned_limit = _real_words_scanned_limit;
3518
3519 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3520 _refs_reached_limit = _real_refs_reached_limit;
3521 }
3522
3523 void CMTask::decrease_limits() {
3524 // This is called when we believe that we're going to do an infrequent
3525 // operation which will increase the per byte scanned cost (i.e. move
3526 // entries to/from the global stack). It basically tries to decrease the
3527 // scanning limit so that the clock is called earlier.
3528
3529 if (_cm->verbose_medium()) {
3530 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3531 }
3532
3533 _words_scanned_limit = _real_words_scanned_limit -
3534 3 * words_scanned_period / 4;
3535 _refs_reached_limit = _real_refs_reached_limit -
3536 3 * refs_reached_period / 4;
3537 }
3538
3539 void CMTask::move_entries_to_global_stack() {
3540 // local array where we'll store the entries that will be popped
3541 // from the local queue
3542 oop buffer[global_stack_transfer_size];
3543
3544 int n = 0;
3545 oop obj;
3546 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3547 buffer[n] = obj;
3548 ++n;
3549 }
3550
3551 if (n > 0) {
3552 // we popped at least one entry from the local queue
3553
3554 statsOnly( ++_global_transfers_to; _local_pops += n );
3555
3556 if (!_cm->mark_stack_push(buffer, n)) {
3557 if (_cm->verbose_low()) {
3558 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3559 _worker_id);
3560 }
3561 set_has_aborted();
3562 } else {
3563 // the transfer was successful
3564
3565 if (_cm->verbose_medium()) {
3566 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3567 _worker_id, n);
3568 }
3569 statsOnly( size_t tmp_size = _cm->mark_stack_size();
3570 if (tmp_size > _global_max_size) {
3571 _global_max_size = tmp_size;
3572 }
3573 _global_pushes += n );
3574 }
3575 }
3576
3577 // this operation was quite expensive, so decrease the limits
3578 decrease_limits();
3579 }
3580
3581 void CMTask::get_entries_from_global_stack() {
3582 // local array where we'll store the entries that will be popped
3583 // from the global stack.
3584 oop buffer[global_stack_transfer_size];
3585 int n;
3586 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3587 assert(n <= global_stack_transfer_size,
3588 "we should not pop more than the given limit");
3589 if (n > 0) {
3590 // yes, we did actually pop at least one entry
3591
3592 statsOnly( ++_global_transfers_from; _global_pops += n );
3593 if (_cm->verbose_medium()) {
3594 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3595 _worker_id, n);
3596 }
3597 for (int i = 0; i < n; ++i) {
3598 bool success = _task_queue->push(buffer[i]);
3599 // We only call this when the local queue is empty or under a
3600 // given target limit. So, we do not expect this push to fail.
3601 assert(success, "invariant");
3602 }
3603
3604 statsOnly( size_t tmp_size = (size_t)_task_queue->size();
3605 if (tmp_size > _local_max_size) {
3606 _local_max_size = tmp_size;
3607 }
3608 _local_pushes += n );
3609 }
3610
3611 // this operation was quite expensive, so decrease the limits
3612 decrease_limits();
3613 }
3614
3615 void CMTask::drain_local_queue(bool partially) {
3616 if (has_aborted()) return;
3617
3618 // Decide what the target size is, depending whether we're going to
3619 // drain it partially (so that other tasks can steal if they run out
3620 // of things to do) or totally (at the very end).
3621 size_t target_size;
3622 if (partially) {
3623 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3624 } else {
3625 target_size = 0;
3626 }
3627
3628 if (_task_queue->size() > target_size) {
3629 if (_cm->verbose_high()) {
3630 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3631 _worker_id, target_size);
3632 }
3633
3634 oop obj;
3635 bool ret = _task_queue->pop_local(obj);
3636 while (ret) {
3637 statsOnly( ++_local_pops );
3638
3639 if (_cm->verbose_high()) {
3640 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3641 p2i((void*) obj));
3642 }
3643
3644 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3645 assert(!_g1h->is_on_master_free_list(
3646 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3647
3648 scan_object(obj);
3649
3650 if (_task_queue->size() <= target_size || has_aborted()) {
3651 ret = false;
3652 } else {
3653 ret = _task_queue->pop_local(obj);
3654 }
3655 }
3656
3657 if (_cm->verbose_high()) {
3658 gclog_or_tty->print_cr("[%u] drained local queue, size = %u",
3659 _worker_id, _task_queue->size());
3660 }
3661 }
3662 }
3663
3664 void CMTask::drain_global_stack(bool partially) {
3665 if (has_aborted()) return;
3666
3667 // We have a policy to drain the local queue before we attempt to
3668 // drain the global stack.
3669 assert(partially || _task_queue->size() == 0, "invariant");
3670
3671 // Decide what the target size is, depending whether we're going to
3672 // drain it partially (so that other tasks can steal if they run out
3673 // of things to do) or totally (at the very end). Notice that,
3674 // because we move entries from the global stack in chunks or
3675 // because another task might be doing the same, we might in fact
3676 // drop below the target. But, this is not a problem.
3677 size_t target_size;
3678 if (partially) {
3679 target_size = _cm->partial_mark_stack_size_target();
3680 } else {
3681 target_size = 0;
3682 }
3683
3684 if (_cm->mark_stack_size() > target_size) {
3685 if (_cm->verbose_low()) {
3686 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3687 _worker_id, target_size);
3688 }
3689
3690 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3691 get_entries_from_global_stack();
3692 drain_local_queue(partially);
3693 }
3694
3695 if (_cm->verbose_low()) {
3696 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3697 _worker_id, _cm->mark_stack_size());
3698 }
3699 }
3700 }
3701
3702 // SATB Queue has several assumptions on whether to call the par or
3703 // non-par versions of the methods. this is why some of the code is
3704 // replicated. We should really get rid of the single-threaded version
3705 // of the code to simplify things.
3706 void CMTask::drain_satb_buffers() {
3707 if (has_aborted()) return;
3708
3709 // We set this so that the regular clock knows that we're in the
3710 // middle of draining buffers and doesn't set the abort flag when it
3711 // notices that SATB buffers are available for draining. It'd be
3712 // very counter productive if it did that. :-)
3713 _draining_satb_buffers = true;
3714
3715 CMSATBBufferClosure satb_cl(this, _g1h);
3716 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3717
3718 // This keeps claiming and applying the closure to completed buffers
3719 // until we run out of buffers or we need to abort.
3720 while (!has_aborted() &&
3721 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
3722 if (_cm->verbose_medium()) {
3723 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3724 }
3725 statsOnly( ++_satb_buffers_processed );
3726 regular_clock_call();
3727 }
3728
3729 _draining_satb_buffers = false;
3730
3731 assert(has_aborted() ||
3732 concurrent() ||
3733 satb_mq_set.completed_buffers_num() == 0, "invariant");
3734
3735 // again, this was a potentially expensive operation, decrease the
3736 // limits to get the regular clock call early
3737 decrease_limits();
3738 }
3739
3740 void CMTask::print_stats() {
3741 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3742 _worker_id, _calls);
3743 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3744 _elapsed_time_ms, _termination_time_ms);
3745 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3746 _step_times_ms.num(), _step_times_ms.avg(),
3747 _step_times_ms.sd());
3748 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3749 _step_times_ms.maximum(), _step_times_ms.sum());
3750
3751 #if _MARKING_STATS_
3752 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3753 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3754 _all_clock_intervals_ms.sd());
3755 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3756 _all_clock_intervals_ms.maximum(),
3757 _all_clock_intervals_ms.sum());
3758 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = " SIZE_FORMAT ", marking = " SIZE_FORMAT,
3759 _clock_due_to_scanning, _clock_due_to_marking);
3760 gclog_or_tty->print_cr(" Objects: scanned = " SIZE_FORMAT ", found on the bitmap = " SIZE_FORMAT,
3761 _objs_scanned, _objs_found_on_bitmap);
3762 gclog_or_tty->print_cr(" Local Queue: pushes = " SIZE_FORMAT ", pops = " SIZE_FORMAT ", max size = " SIZE_FORMAT,
3763 _local_pushes, _local_pops, _local_max_size);
3764 gclog_or_tty->print_cr(" Global Stack: pushes = " SIZE_FORMAT ", pops = " SIZE_FORMAT ", max size = " SIZE_FORMAT,
3765 _global_pushes, _global_pops, _global_max_size);
3766 gclog_or_tty->print_cr(" transfers to = " SIZE_FORMAT ", transfers from = " SIZE_FORMAT,
3767 _global_transfers_to,_global_transfers_from);
3768 gclog_or_tty->print_cr(" Regions: claimed = " SIZE_FORMAT, _regions_claimed);
3769 gclog_or_tty->print_cr(" SATB buffers: processed = " SIZE_FORMAT, _satb_buffers_processed);
3770 gclog_or_tty->print_cr(" Steals: attempts = " SIZE_FORMAT ", successes = " SIZE_FORMAT,
3771 _steal_attempts, _steals);
3772 gclog_or_tty->print_cr(" Aborted: " SIZE_FORMAT ", due to", _aborted);
3773 gclog_or_tty->print_cr(" overflow: " SIZE_FORMAT ", global abort: " SIZE_FORMAT ", yield: " SIZE_FORMAT,
3774 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3775 gclog_or_tty->print_cr(" time out: " SIZE_FORMAT ", SATB: " SIZE_FORMAT ", termination: " SIZE_FORMAT,
3776 _aborted_timed_out, _aborted_satb, _aborted_termination);
3777 #endif // _MARKING_STATS_
3778 }
3779
3780 bool ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) {
3781 return _task_queues->steal(worker_id, hash_seed, obj);
3782 }
3783
3784 /*****************************************************************************
3785
3786 The do_marking_step(time_target_ms, ...) method is the building
3787 block of the parallel marking framework. It can be called in parallel
3788 with other invocations of do_marking_step() on different tasks
3789 (but only one per task, obviously) and concurrently with the
3790 mutator threads, or during remark, hence it eliminates the need
3791 for two versions of the code. When called during remark, it will
3792 pick up from where the task left off during the concurrent marking
3793 phase. Interestingly, tasks are also claimable during evacuation
3794 pauses too, since do_marking_step() ensures that it aborts before
3795 it needs to yield.
3796
3797 The data structures that it uses to do marking work are the
3798 following:
3799
3800 (1) Marking Bitmap. If there are gray objects that appear only
3801 on the bitmap (this happens either when dealing with an overflow
3802 or when the initial marking phase has simply marked the roots
3803 and didn't push them on the stack), then tasks claim heap
3804 regions whose bitmap they then scan to find gray objects. A
3805 global finger indicates where the end of the last claimed region
3806 is. A local finger indicates how far into the region a task has
3807 scanned. The two fingers are used to determine how to gray an
3808 object (i.e. whether simply marking it is OK, as it will be
3809 visited by a task in the future, or whether it needs to be also
3810 pushed on a stack).
3811
3812 (2) Local Queue. The local queue of the task which is accessed
3813 reasonably efficiently by the task. Other tasks can steal from
3814 it when they run out of work. Throughout the marking phase, a
3815 task attempts to keep its local queue short but not totally
3816 empty, so that entries are available for stealing by other
3817 tasks. Only when there is no more work, a task will totally
3818 drain its local queue.
3819
3820 (3) Global Mark Stack. This handles local queue overflow. During
3821 marking only sets of entries are moved between it and the local
3822 queues, as access to it requires a mutex and more fine-grain
3823 interaction with it which might cause contention. If it
3824 overflows, then the marking phase should restart and iterate
3825 over the bitmap to identify gray objects. Throughout the marking
3826 phase, tasks attempt to keep the global mark stack at a small
3827 length but not totally empty, so that entries are available for
3828 popping by other tasks. Only when there is no more work, tasks
3829 will totally drain the global mark stack.
3830
3831 (4) SATB Buffer Queue. This is where completed SATB buffers are
3832 made available. Buffers are regularly removed from this queue
3833 and scanned for roots, so that the queue doesn't get too
3834 long. During remark, all completed buffers are processed, as
3835 well as the filled in parts of any uncompleted buffers.
3836
3837 The do_marking_step() method tries to abort when the time target
3838 has been reached. There are a few other cases when the
3839 do_marking_step() method also aborts:
3840
3841 (1) When the marking phase has been aborted (after a Full GC).
3842
3843 (2) When a global overflow (on the global stack) has been
3844 triggered. Before the task aborts, it will actually sync up with
3845 the other tasks to ensure that all the marking data structures
3846 (local queues, stacks, fingers etc.) are re-initialized so that
3847 when do_marking_step() completes, the marking phase can
3848 immediately restart.
3849
3850 (3) When enough completed SATB buffers are available. The
3851 do_marking_step() method only tries to drain SATB buffers right
3852 at the beginning. So, if enough buffers are available, the
3853 marking step aborts and the SATB buffers are processed at
3854 the beginning of the next invocation.
3855
3856 (4) To yield. when we have to yield then we abort and yield
3857 right at the end of do_marking_step(). This saves us from a lot
3858 of hassle as, by yielding we might allow a Full GC. If this
3859 happens then objects will be compacted underneath our feet, the
3860 heap might shrink, etc. We save checking for this by just
3861 aborting and doing the yield right at the end.
3862
3863 From the above it follows that the do_marking_step() method should
3864 be called in a loop (or, otherwise, regularly) until it completes.
3865
3866 If a marking step completes without its has_aborted() flag being
3867 true, it means it has completed the current marking phase (and
3868 also all other marking tasks have done so and have all synced up).
3869
3870 A method called regular_clock_call() is invoked "regularly" (in
3871 sub ms intervals) throughout marking. It is this clock method that
3872 checks all the abort conditions which were mentioned above and
3873 decides when the task should abort. A work-based scheme is used to
3874 trigger this clock method: when the number of object words the
3875 marking phase has scanned or the number of references the marking
3876 phase has visited reach a given limit. Additional invocations to
3877 the method clock have been planted in a few other strategic places
3878 too. The initial reason for the clock method was to avoid calling
3879 vtime too regularly, as it is quite expensive. So, once it was in
3880 place, it was natural to piggy-back all the other conditions on it
3881 too and not constantly check them throughout the code.
3882
3883 If do_termination is true then do_marking_step will enter its
3884 termination protocol.
3885
3886 The value of is_serial must be true when do_marking_step is being
3887 called serially (i.e. by the VMThread) and do_marking_step should
3888 skip any synchronization in the termination and overflow code.
3889 Examples include the serial remark code and the serial reference
3890 processing closures.
3891
3892 The value of is_serial must be false when do_marking_step is
3893 being called by any of the worker threads in a work gang.
3894 Examples include the concurrent marking code (CMMarkingTask),
3895 the MT remark code, and the MT reference processing closures.
3896
3897 *****************************************************************************/
3898
3899 void CMTask::do_marking_step(double time_target_ms,
3900 bool do_termination,
3901 bool is_serial) {
3902 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
3903 assert(concurrent() == _cm->concurrent(), "they should be the same");
3904
3905 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
3906 assert(_task_queues != NULL, "invariant");
3907 assert(_task_queue != NULL, "invariant");
3908 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
3909
3910 assert(!_claimed,
3911 "only one thread should claim this task at any one time");
3912
3913 // OK, this doesn't safeguard again all possible scenarios, as it is
3914 // possible for two threads to set the _claimed flag at the same
3915 // time. But it is only for debugging purposes anyway and it will
3916 // catch most problems.
3917 _claimed = true;
3918
3919 _start_time_ms = os::elapsedVTime() * 1000.0;
3920 statsOnly( _interval_start_time_ms = _start_time_ms );
3921
3922 // If do_stealing is true then do_marking_step will attempt to
3923 // steal work from the other CMTasks. It only makes sense to
3924 // enable stealing when the termination protocol is enabled
3925 // and do_marking_step() is not being called serially.
3926 bool do_stealing = do_termination && !is_serial;
3927
3928 double diff_prediction_ms =
3929 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
3930 _time_target_ms = time_target_ms - diff_prediction_ms;
3931
3932 // set up the variables that are used in the work-based scheme to
3933 // call the regular clock method
3934 _words_scanned = 0;
3935 _refs_reached = 0;
3936 recalculate_limits();
3937
3938 // clear all flags
3939 clear_has_aborted();
3940 _has_timed_out = false;
3941 _draining_satb_buffers = false;
3942
3943 ++_calls;
3944
3945 if (_cm->verbose_low()) {
3946 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
3947 "target = %1.2lfms >>>>>>>>>>",
3948 _worker_id, _calls, _time_target_ms);
3949 }
3950
3951 // Set up the bitmap and oop closures. Anything that uses them is
3952 // eventually called from this method, so it is OK to allocate these
3953 // statically.
3954 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
3955 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
3956 set_cm_oop_closure(&cm_oop_closure);
3957
3958 if (_cm->has_overflown()) {
3959 // This can happen if the mark stack overflows during a GC pause
3960 // and this task, after a yield point, restarts. We have to abort
3961 // as we need to get into the overflow protocol which happens
3962 // right at the end of this task.
3963 set_has_aborted();
3964 }
3965
3966 // First drain any available SATB buffers. After this, we will not
3967 // look at SATB buffers before the next invocation of this method.
3968 // If enough completed SATB buffers are queued up, the regular clock
3969 // will abort this task so that it restarts.
3970 drain_satb_buffers();
3971 // ...then partially drain the local queue and the global stack
3972 drain_local_queue(true);
3973 drain_global_stack(true);
3974
3975 do {
3976 if (!has_aborted() && _curr_region != NULL) {
3977 // This means that we're already holding on to a region.
3978 assert(_finger != NULL, "if region is not NULL, then the finger "
3979 "should not be NULL either");
3980
3981 // We might have restarted this task after an evacuation pause
3982 // which might have evacuated the region we're holding on to
3983 // underneath our feet. Let's read its limit again to make sure
3984 // that we do not iterate over a region of the heap that
3985 // contains garbage (update_region_limit() will also move
3986 // _finger to the start of the region if it is found empty).
3987 update_region_limit();
3988 // We will start from _finger not from the start of the region,
3989 // as we might be restarting this task after aborting half-way
3990 // through scanning this region. In this case, _finger points to
3991 // the address where we last found a marked object. If this is a
3992 // fresh region, _finger points to start().
3993 MemRegion mr = MemRegion(_finger, _region_limit);
3994
3995 if (_cm->verbose_low()) {
3996 gclog_or_tty->print_cr("[%u] we're scanning part "
3997 "["PTR_FORMAT", "PTR_FORMAT") "
3998 "of region "HR_FORMAT,
3999 _worker_id, p2i(_finger), p2i(_region_limit),
4000 HR_FORMAT_PARAMS(_curr_region));
4001 }
4002
4003 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
4004 "humongous regions should go around loop once only");
4005
4006 // Some special cases:
4007 // If the memory region is empty, we can just give up the region.
4008 // If the current region is humongous then we only need to check
4009 // the bitmap for the bit associated with the start of the object,
4010 // scan the object if it's live, and give up the region.
4011 // Otherwise, let's iterate over the bitmap of the part of the region
4012 // that is left.
4013 // If the iteration is successful, give up the region.
4014 if (mr.is_empty()) {
4015 giveup_current_region();
4016 regular_clock_call();
4017 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
4018 if (_nextMarkBitMap->isMarked(mr.start())) {
4019 // The object is marked - apply the closure
4020 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4021 bitmap_closure.do_bit(offset);
4022 }
4023 // Even if this task aborted while scanning the humongous object
4024 // we can (and should) give up the current region.
4025 giveup_current_region();
4026 regular_clock_call();
4027 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4028 giveup_current_region();
4029 regular_clock_call();
4030 } else {
4031 assert(has_aborted(), "currently the only way to do so");
4032 // The only way to abort the bitmap iteration is to return
4033 // false from the do_bit() method. However, inside the
4034 // do_bit() method we move the _finger to point to the
4035 // object currently being looked at. So, if we bail out, we
4036 // have definitely set _finger to something non-null.
4037 assert(_finger != NULL, "invariant");
4038
4039 // Region iteration was actually aborted. So now _finger
4040 // points to the address of the object we last scanned. If we
4041 // leave it there, when we restart this task, we will rescan
4042 // the object. It is easy to avoid this. We move the finger by
4043 // enough to point to the next possible object header (the
4044 // bitmap knows by how much we need to move it as it knows its
4045 // granularity).
4046 assert(_finger < _region_limit, "invariant");
4047 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4048 // Check if bitmap iteration was aborted while scanning the last object
4049 if (new_finger >= _region_limit) {
4050 giveup_current_region();
4051 } else {
4052 move_finger_to(new_finger);
4053 }
4054 }
4055 }
4056 // At this point we have either completed iterating over the
4057 // region we were holding on to, or we have aborted.
4058
4059 // We then partially drain the local queue and the global stack.
4060 // (Do we really need this?)
4061 drain_local_queue(true);
4062 drain_global_stack(true);
4063
4064 // Read the note on the claim_region() method on why it might
4065 // return NULL with potentially more regions available for
4066 // claiming and why we have to check out_of_regions() to determine
4067 // whether we're done or not.
4068 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4069 // We are going to try to claim a new region. We should have
4070 // given up on the previous one.
4071 // Separated the asserts so that we know which one fires.
4072 assert(_curr_region == NULL, "invariant");
4073 assert(_finger == NULL, "invariant");
4074 assert(_region_limit == NULL, "invariant");
4075 if (_cm->verbose_low()) {
4076 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4077 }
4078 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4079 if (claimed_region != NULL) {
4080 // Yes, we managed to claim one
4081 statsOnly( ++_regions_claimed );
4082
4083 if (_cm->verbose_low()) {
4084 gclog_or_tty->print_cr("[%u] we successfully claimed "
4085 "region "PTR_FORMAT,
4086 _worker_id, p2i(claimed_region));
4087 }
4088
4089 setup_for_region(claimed_region);
4090 assert(_curr_region == claimed_region, "invariant");
4091 }
4092 // It is important to call the regular clock here. It might take
4093 // a while to claim a region if, for example, we hit a large
4094 // block of empty regions. So we need to call the regular clock
4095 // method once round the loop to make sure it's called
4096 // frequently enough.
4097 regular_clock_call();
4098 }
4099
4100 if (!has_aborted() && _curr_region == NULL) {
4101 assert(_cm->out_of_regions(),
4102 "at this point we should be out of regions");
4103 }
4104 } while ( _curr_region != NULL && !has_aborted());
4105
4106 if (!has_aborted()) {
4107 // We cannot check whether the global stack is empty, since other
4108 // tasks might be pushing objects to it concurrently.
4109 assert(_cm->out_of_regions(),
4110 "at this point we should be out of regions");
4111
4112 if (_cm->verbose_low()) {
4113 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4114 }
4115
4116 // Try to reduce the number of available SATB buffers so that
4117 // remark has less work to do.
4118 drain_satb_buffers();
4119 }
4120
4121 // Since we've done everything else, we can now totally drain the
4122 // local queue and global stack.
4123 drain_local_queue(false);
4124 drain_global_stack(false);
4125
4126 // Attempt at work stealing from other task's queues.
4127 if (do_stealing && !has_aborted()) {
4128 // We have not aborted. This means that we have finished all that
4129 // we could. Let's try to do some stealing...
4130
4131 // We cannot check whether the global stack is empty, since other
4132 // tasks might be pushing objects to it concurrently.
4133 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4134 "only way to reach here");
4135
4136 if (_cm->verbose_low()) {
4137 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4138 }
4139
4140 while (!has_aborted()) {
4141 oop obj;
4142 statsOnly( ++_steal_attempts );
4143
4144 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4145 if (_cm->verbose_medium()) {
4146 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4147 _worker_id, p2i((void*) obj));
4148 }
4149
4150 statsOnly( ++_steals );
4151
4152 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4153 "any stolen object should be marked");
4154 scan_object(obj);
4155
4156 // And since we're towards the end, let's totally drain the
4157 // local queue and global stack.
4158 drain_local_queue(false);
4159 drain_global_stack(false);
4160 } else {
4161 break;
4162 }
4163 }
4164 }
4165
4166 // If we are about to wrap up and go into termination, check if we
4167 // should raise the overflow flag.
4168 if (do_termination && !has_aborted()) {
4169 if (_cm->force_overflow()->should_force()) {
4170 _cm->set_has_overflown();
4171 regular_clock_call();
4172 }
4173 }
4174
4175 // We still haven't aborted. Now, let's try to get into the
4176 // termination protocol.
4177 if (do_termination && !has_aborted()) {
4178 // We cannot check whether the global stack is empty, since other
4179 // tasks might be concurrently pushing objects on it.
4180 // Separated the asserts so that we know which one fires.
4181 assert(_cm->out_of_regions(), "only way to reach here");
4182 assert(_task_queue->size() == 0, "only way to reach here");
4183
4184 if (_cm->verbose_low()) {
4185 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4186 }
4187
4188 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4189
4190 // The CMTask class also extends the TerminatorTerminator class,
4191 // hence its should_exit_termination() method will also decide
4192 // whether to exit the termination protocol or not.
4193 bool finished = (is_serial ||
4194 _cm->terminator()->offer_termination(this));
4195 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4196 _termination_time_ms +=
4197 termination_end_time_ms - _termination_start_time_ms;
4198
4199 if (finished) {
4200 // We're all done.
4201
4202 if (_worker_id == 0) {
4203 // let's allow task 0 to do this
4204 if (concurrent()) {
4205 assert(_cm->concurrent_marking_in_progress(), "invariant");
4206 // we need to set this to false before the next
4207 // safepoint. This way we ensure that the marking phase
4208 // doesn't observe any more heap expansions.
4209 _cm->clear_concurrent_marking_in_progress();
4210 }
4211 }
4212
4213 // We can now guarantee that the global stack is empty, since
4214 // all other tasks have finished. We separated the guarantees so
4215 // that, if a condition is false, we can immediately find out
4216 // which one.
4217 guarantee(_cm->out_of_regions(), "only way to reach here");
4218 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4219 guarantee(_task_queue->size() == 0, "only way to reach here");
4220 guarantee(!_cm->has_overflown(), "only way to reach here");
4221 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4222
4223 if (_cm->verbose_low()) {
4224 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4225 }
4226 } else {
4227 // Apparently there's more work to do. Let's abort this task. It
4228 // will restart it and we can hopefully find more things to do.
4229
4230 if (_cm->verbose_low()) {
4231 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4232 _worker_id);
4233 }
4234
4235 set_has_aborted();
4236 statsOnly( ++_aborted_termination );
4237 }
4238 }
4239
4240 // Mainly for debugging purposes to make sure that a pointer to the
4241 // closure which was statically allocated in this frame doesn't
4242 // escape it by accident.
4243 set_cm_oop_closure(NULL);
4244 double end_time_ms = os::elapsedVTime() * 1000.0;
4245 double elapsed_time_ms = end_time_ms - _start_time_ms;
4246 // Update the step history.
4247 _step_times_ms.add(elapsed_time_ms);
4248
4249 if (has_aborted()) {
4250 // The task was aborted for some reason.
4251
4252 statsOnly( ++_aborted );
4253
4254 if (_has_timed_out) {
4255 double diff_ms = elapsed_time_ms - _time_target_ms;
4256 // Keep statistics of how well we did with respect to hitting
4257 // our target only if we actually timed out (if we aborted for
4258 // other reasons, then the results might get skewed).
4259 _marking_step_diffs_ms.add(diff_ms);
4260 }
4261
4262 if (_cm->has_overflown()) {
4263 // This is the interesting one. We aborted because a global
4264 // overflow was raised. This means we have to restart the
4265 // marking phase and start iterating over regions. However, in
4266 // order to do this we have to make sure that all tasks stop
4267 // what they are doing and re-initialize in a safe manner. We
4268 // will achieve this with the use of two barrier sync points.
4269
4270 if (_cm->verbose_low()) {
4271 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4272 }
4273
4274 if (!is_serial) {
4275 // We only need to enter the sync barrier if being called
4276 // from a parallel context
4277 _cm->enter_first_sync_barrier(_worker_id);
4278
4279 // When we exit this sync barrier we know that all tasks have
4280 // stopped doing marking work. So, it's now safe to
4281 // re-initialize our data structures. At the end of this method,
4282 // task 0 will clear the global data structures.
4283 }
4284
4285 statsOnly( ++_aborted_overflow );
4286
4287 // We clear the local state of this task...
4288 clear_region_fields();
4289
4290 if (!is_serial) {
4291 // ...and enter the second barrier.
4292 _cm->enter_second_sync_barrier(_worker_id);
4293 }
4294 // At this point, if we're during the concurrent phase of
4295 // marking, everything has been re-initialized and we're
4296 // ready to restart.
4297 }
4298
4299 if (_cm->verbose_low()) {
4300 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4301 "elapsed = %1.2lfms <<<<<<<<<<",
4302 _worker_id, _time_target_ms, elapsed_time_ms);
4303 if (_cm->has_aborted()) {
4304 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4305 _worker_id);
4306 }
4307 }
4308 } else {
4309 if (_cm->verbose_low()) {
4310 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4311 "elapsed = %1.2lfms <<<<<<<<<<",
4312 _worker_id, _time_target_ms, elapsed_time_ms);
4313 }
4314 }
4315
4316 _claimed = false;
4317 }
4318
4319 CMTask::CMTask(uint worker_id,
4320 ConcurrentMark* cm,
4321 size_t* marked_bytes,
4322 BitMap* card_bm,
4323 CMTaskQueue* task_queue,
4324 CMTaskQueueSet* task_queues)
4325 : _g1h(G1CollectedHeap::heap()),
4326 _worker_id(worker_id), _cm(cm),
4327 _claimed(false),
4328 _nextMarkBitMap(NULL), _hash_seed(17),
4329 _task_queue(task_queue),
4330 _task_queues(task_queues),
4331 _cm_oop_closure(NULL),
4332 _marked_bytes_array(marked_bytes),
4333 _card_bm(card_bm) {
4334 guarantee(task_queue != NULL, "invariant");
4335 guarantee(task_queues != NULL, "invariant");
4336
4337 statsOnly( _clock_due_to_scanning = 0;
4338 _clock_due_to_marking = 0 );
4339
4340 _marking_step_diffs_ms.add(0.5);
4341 }
4342
4343 // These are formatting macros that are used below to ensure
4344 // consistent formatting. The *_H_* versions are used to format the
4345 // header for a particular value and they should be kept consistent
4346 // with the corresponding macro. Also note that most of the macros add
4347 // the necessary white space (as a prefix) which makes them a bit
4348 // easier to compose.
4349
4350 // All the output lines are prefixed with this string to be able to
4351 // identify them easily in a large log file.
4352 #define G1PPRL_LINE_PREFIX "###"
4353
4354 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4355 #ifdef _LP64
4356 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4357 #else // _LP64
4358 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4359 #endif // _LP64
4360
4361 // For per-region info
4362 #define G1PPRL_TYPE_FORMAT " %-4s"
4363 #define G1PPRL_TYPE_H_FORMAT " %4s"
4364 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4365 #define G1PPRL_BYTE_H_FORMAT " %9s"
4366 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4367 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4368
4369 // For summary info
4370 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4371 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4372 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4373 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4374
4375 G1PrintRegionLivenessInfoClosure::
4376 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4377 : _out(out),
4378 _total_used_bytes(0), _total_capacity_bytes(0),
4379 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4380 _hum_used_bytes(0), _hum_capacity_bytes(0),
4381 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4382 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4383 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4384 MemRegion g1_reserved = g1h->g1_reserved();
4385 double now = os::elapsedTime();
4386
4387 // Print the header of the output.
4388 _out->cr();
4389 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4390 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4391 G1PPRL_SUM_ADDR_FORMAT("reserved")
4392 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4393 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4394 HeapRegion::GrainBytes);
4395 _out->print_cr(G1PPRL_LINE_PREFIX);
4396 _out->print_cr(G1PPRL_LINE_PREFIX
4397 G1PPRL_TYPE_H_FORMAT
4398 G1PPRL_ADDR_BASE_H_FORMAT
4399 G1PPRL_BYTE_H_FORMAT
4400 G1PPRL_BYTE_H_FORMAT
4401 G1PPRL_BYTE_H_FORMAT
4402 G1PPRL_DOUBLE_H_FORMAT
4403 G1PPRL_BYTE_H_FORMAT
4404 G1PPRL_BYTE_H_FORMAT,
4405 "type", "address-range",
4406 "used", "prev-live", "next-live", "gc-eff",
4407 "remset", "code-roots");
4408 _out->print_cr(G1PPRL_LINE_PREFIX
4409 G1PPRL_TYPE_H_FORMAT
4410 G1PPRL_ADDR_BASE_H_FORMAT
4411 G1PPRL_BYTE_H_FORMAT
4412 G1PPRL_BYTE_H_FORMAT
4413 G1PPRL_BYTE_H_FORMAT
4414 G1PPRL_DOUBLE_H_FORMAT
4415 G1PPRL_BYTE_H_FORMAT
4416 G1PPRL_BYTE_H_FORMAT,
4417 "", "",
4418 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4419 "(bytes)", "(bytes)");
4420 }
4421
4422 // It takes as a parameter a reference to one of the _hum_* fields, it
4423 // deduces the corresponding value for a region in a humongous region
4424 // series (either the region size, or what's left if the _hum_* field
4425 // is < the region size), and updates the _hum_* field accordingly.
4426 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4427 size_t bytes = 0;
4428 // The > 0 check is to deal with the prev and next live bytes which
4429 // could be 0.
4430 if (*hum_bytes > 0) {
4431 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4432 *hum_bytes -= bytes;
4433 }
4434 return bytes;
4435 }
4436
4437 // It deduces the values for a region in a humongous region series
4438 // from the _hum_* fields and updates those accordingly. It assumes
4439 // that that _hum_* fields have already been set up from the "starts
4440 // humongous" region and we visit the regions in address order.
4441 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4442 size_t* capacity_bytes,
4443 size_t* prev_live_bytes,
4444 size_t* next_live_bytes) {
4445 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4446 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4447 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4448 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4449 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4450 }
4451
4452 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4453 const char* type = r->get_type_str();
4454 HeapWord* bottom = r->bottom();
4455 HeapWord* end = r->end();
4456 size_t capacity_bytes = r->capacity();
4457 size_t used_bytes = r->used();
4458 size_t prev_live_bytes = r->live_bytes();
4459 size_t next_live_bytes = r->next_live_bytes();
4460 double gc_eff = r->gc_efficiency();
4461 size_t remset_bytes = r->rem_set()->mem_size();
4462 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4463
4464 if (r->is_starts_humongous()) {
4465 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4466 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4467 "they should have been zeroed after the last time we used them");
4468 // Set up the _hum_* fields.
4469 _hum_capacity_bytes = capacity_bytes;
4470 _hum_used_bytes = used_bytes;
4471 _hum_prev_live_bytes = prev_live_bytes;
4472 _hum_next_live_bytes = next_live_bytes;
4473 get_hum_bytes(&used_bytes, &capacity_bytes,
4474 &prev_live_bytes, &next_live_bytes);
4475 end = bottom + HeapRegion::GrainWords;
4476 } else if (r->is_continues_humongous()) {
4477 get_hum_bytes(&used_bytes, &capacity_bytes,
4478 &prev_live_bytes, &next_live_bytes);
4479 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4480 }
4481
4482 _total_used_bytes += used_bytes;
4483 _total_capacity_bytes += capacity_bytes;
4484 _total_prev_live_bytes += prev_live_bytes;
4485 _total_next_live_bytes += next_live_bytes;
4486 _total_remset_bytes += remset_bytes;
4487 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4488
4489 // Print a line for this particular region.
4490 _out->print_cr(G1PPRL_LINE_PREFIX
4491 G1PPRL_TYPE_FORMAT
4492 G1PPRL_ADDR_BASE_FORMAT
4493 G1PPRL_BYTE_FORMAT
4494 G1PPRL_BYTE_FORMAT
4495 G1PPRL_BYTE_FORMAT
4496 G1PPRL_DOUBLE_FORMAT
4497 G1PPRL_BYTE_FORMAT
4498 G1PPRL_BYTE_FORMAT,
4499 type, p2i(bottom), p2i(end),
4500 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4501 remset_bytes, strong_code_roots_bytes);
4502
4503 return false;
4504 }
4505
4506 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4507 // add static memory usages to remembered set sizes
4508 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4509 // Print the footer of the output.
4510 _out->print_cr(G1PPRL_LINE_PREFIX);
4511 _out->print_cr(G1PPRL_LINE_PREFIX
4512 " SUMMARY"
4513 G1PPRL_SUM_MB_FORMAT("capacity")
4514 G1PPRL_SUM_MB_PERC_FORMAT("used")
4515 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4516 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4517 G1PPRL_SUM_MB_FORMAT("remset")
4518 G1PPRL_SUM_MB_FORMAT("code-roots"),
4519 bytes_to_mb(_total_capacity_bytes),
4520 bytes_to_mb(_total_used_bytes),
4521 perc(_total_used_bytes, _total_capacity_bytes),
4522 bytes_to_mb(_total_prev_live_bytes),
4523 perc(_total_prev_live_bytes, _total_capacity_bytes),
4524 bytes_to_mb(_total_next_live_bytes),
4525 perc(_total_next_live_bytes, _total_capacity_bytes),
4526 bytes_to_mb(_total_remset_bytes),
4527 bytes_to_mb(_total_strong_code_roots_bytes));
4528 _out->cr();
4529 }