1 /*
2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "code/codeCache.hpp"
28 #include "gc_implementation/g1/concurrentMark.inline.hpp"
29 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
30 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
31 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
32 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
33 #include "gc_implementation/g1/g1Log.hpp"
34 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
35 #include "gc_implementation/g1/g1RemSet.hpp"
36 #include "gc_implementation/g1/heapRegion.inline.hpp"
37 #include "gc_implementation/g1/heapRegionManager.inline.hpp"
38 #include "gc_implementation/g1/heapRegionRemSet.hpp"
39 #include "gc_implementation/g1/heapRegionSet.inline.hpp"
40 #include "gc_implementation/shared/vmGCOperations.hpp"
41 #include "gc_implementation/shared/gcTimer.hpp"
42 #include "gc_implementation/shared/gcTrace.hpp"
43 #include "gc_implementation/shared/gcTraceTime.hpp"
44 #include "memory/allocation.hpp"
45 #include "memory/genOopClosures.inline.hpp"
46 #include "memory/referencePolicy.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "oops/oop.inline.hpp"
49 #include "runtime/handles.inline.hpp"
50 #include "runtime/java.hpp"
51 #include "runtime/atomic.inline.hpp"
52 #include "runtime/prefetch.inline.hpp"
53 #include "services/memTracker.hpp"
54
55 // Concurrent marking bit map wrapper
56
57 CMBitMapRO::CMBitMapRO(int shifter) :
58 _bm(),
59 _shifter(shifter) {
60 _bmStartWord = 0;
61 _bmWordSize = 0;
62 }
63
64 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
65 const HeapWord* limit) const {
66 // First we must round addr *up* to a possible object boundary.
67 addr = (HeapWord*)align_size_up((intptr_t)addr,
68 HeapWordSize << _shifter);
69 size_t addrOffset = heapWordToOffset(addr);
70 if (limit == NULL) {
71 limit = _bmStartWord + _bmWordSize;
72 }
73 size_t limitOffset = heapWordToOffset(limit);
74 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
75 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
76 assert(nextAddr >= addr, "get_next_one postcondition");
77 assert(nextAddr == limit || isMarked(nextAddr),
78 "get_next_one postcondition");
79 return nextAddr;
80 }
81
82 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
83 const HeapWord* limit) const {
84 size_t addrOffset = heapWordToOffset(addr);
85 if (limit == NULL) {
86 limit = _bmStartWord + _bmWordSize;
87 }
88 size_t limitOffset = heapWordToOffset(limit);
89 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
90 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
91 assert(nextAddr >= addr, "get_next_one postcondition");
92 assert(nextAddr == limit || !isMarked(nextAddr),
93 "get_next_one postcondition");
94 return nextAddr;
95 }
96
97 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
98 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
99 return (int) (diff >> _shifter);
100 }
101
102 #ifndef PRODUCT
103 bool CMBitMapRO::covers(MemRegion heap_rs) const {
104 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
105 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
106 "size inconsistency");
107 return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
108 _bmWordSize == heap_rs.word_size();
109 }
110 #endif
111
112 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
113 _bm.print_on_error(st, prefix);
114 }
115
116 size_t CMBitMap::compute_size(size_t heap_size) {
117 return heap_size / mark_distance();
118 }
119
120 size_t CMBitMap::mark_distance() {
121 return MinObjAlignmentInBytes * BitsPerByte;
122 }
123
124 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
125 _bmStartWord = heap.start();
126 _bmWordSize = heap.word_size();
127
128 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
129 _bm.set_size(_bmWordSize >> _shifter);
130
131 storage->set_mapping_changed_listener(&_listener);
132 }
133
134 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions) {
135 // We need to clear the bitmap on commit, removing any existing information.
136 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
137 _bm->clearRange(mr);
138 }
139
140 // Closure used for clearing the given mark bitmap.
141 class ClearBitmapHRClosure : public HeapRegionClosure {
142 private:
143 ConcurrentMark* _cm;
144 CMBitMap* _bitmap;
145 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration.
146 public:
147 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
148 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
149 }
150
151 virtual bool doHeapRegion(HeapRegion* r) {
152 size_t const chunk_size_in_words = M / HeapWordSize;
153
154 HeapWord* cur = r->bottom();
155 HeapWord* const end = r->end();
156
157 while (cur < end) {
158 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
159 _bitmap->clearRange(mr);
160
161 cur += chunk_size_in_words;
162
163 // Abort iteration if after yielding the marking has been aborted.
164 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
165 return true;
166 }
167 // Repeat the asserts from before the start of the closure. We will do them
168 // as asserts here to minimize their overhead on the product. However, we
169 // will have them as guarantees at the beginning / end of the bitmap
170 // clearing to get some checking in the product.
171 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
172 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
173 }
174
175 return false;
176 }
177 };
178
179 void CMBitMap::clearAll() {
180 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
181 G1CollectedHeap::heap()->heap_region_iterate(&cl);
182 guarantee(cl.complete(), "Must have completed iteration.");
183 return;
184 }
185
186 void CMBitMap::markRange(MemRegion mr) {
187 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
188 assert(!mr.is_empty(), "unexpected empty region");
189 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
190 ((HeapWord *) mr.end())),
191 "markRange memory region end is not card aligned");
192 // convert address range into offset range
193 _bm.at_put_range(heapWordToOffset(mr.start()),
194 heapWordToOffset(mr.end()), true);
195 }
196
197 void CMBitMap::clearRange(MemRegion mr) {
198 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
199 assert(!mr.is_empty(), "unexpected empty region");
200 // convert address range into offset range
201 _bm.at_put_range(heapWordToOffset(mr.start()),
202 heapWordToOffset(mr.end()), false);
203 }
204
205 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
206 HeapWord* end_addr) {
207 HeapWord* start = getNextMarkedWordAddress(addr);
208 start = MIN2(start, end_addr);
209 HeapWord* end = getNextUnmarkedWordAddress(start);
210 end = MIN2(end, end_addr);
211 assert(start <= end, "Consistency check");
212 MemRegion mr(start, end);
213 if (!mr.is_empty()) {
214 clearRange(mr);
215 }
216 return mr;
217 }
218
219 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
220 _base(NULL), _cm(cm)
221 #ifdef ASSERT
222 , _drain_in_progress(false)
223 , _drain_in_progress_yields(false)
224 #endif
225 {}
226
227 bool CMMarkStack::allocate(size_t capacity) {
228 // allocate a stack of the requisite depth
229 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
230 if (!rs.is_reserved()) {
231 warning("ConcurrentMark MarkStack allocation failure");
232 return false;
233 }
234 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
235 if (!_virtual_space.initialize(rs, rs.size())) {
236 warning("ConcurrentMark MarkStack backing store failure");
237 // Release the virtual memory reserved for the marking stack
238 rs.release();
239 return false;
240 }
241 assert(_virtual_space.committed_size() == rs.size(),
242 "Didn't reserve backing store for all of ConcurrentMark stack?");
243 _base = (oop*) _virtual_space.low();
244 setEmpty();
245 _capacity = (jint) capacity;
246 _saved_index = -1;
247 _should_expand = false;
248 NOT_PRODUCT(_max_depth = 0);
249 return true;
250 }
251
252 void CMMarkStack::expand() {
253 // Called, during remark, if we've overflown the marking stack during marking.
254 assert(isEmpty(), "stack should been emptied while handling overflow");
255 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
256 // Clear expansion flag
257 _should_expand = false;
258 if (_capacity == (jint) MarkStackSizeMax) {
259 if (PrintGCDetails && Verbose) {
260 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
261 }
262 return;
263 }
264 // Double capacity if possible
265 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
266 // Do not give up existing stack until we have managed to
267 // get the double capacity that we desired.
268 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
269 sizeof(oop)));
270 if (rs.is_reserved()) {
271 // Release the backing store associated with old stack
272 _virtual_space.release();
273 // Reinitialize virtual space for new stack
274 if (!_virtual_space.initialize(rs, rs.size())) {
275 fatal("Not enough swap for expanded marking stack capacity");
276 }
277 _base = (oop*)(_virtual_space.low());
278 _index = 0;
279 _capacity = new_capacity;
280 } else {
281 if (PrintGCDetails && Verbose) {
282 // Failed to double capacity, continue;
283 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
284 SIZE_FORMAT"K to " SIZE_FORMAT"K",
285 _capacity / K, new_capacity / K);
286 }
287 }
288 }
289
290 void CMMarkStack::set_should_expand() {
291 // If we're resetting the marking state because of an
292 // marking stack overflow, record that we should, if
293 // possible, expand the stack.
294 _should_expand = _cm->has_overflown();
295 }
296
297 CMMarkStack::~CMMarkStack() {
298 if (_base != NULL) {
299 _base = NULL;
300 _virtual_space.release();
301 }
302 }
303
304 void CMMarkStack::par_push(oop ptr) {
305 while (true) {
306 if (isFull()) {
307 _overflow = true;
308 return;
309 }
310 // Otherwise...
311 jint index = _index;
312 jint next_index = index+1;
313 jint res = Atomic::cmpxchg(next_index, &_index, index);
314 if (res == index) {
315 _base[index] = ptr;
316 // Note that we don't maintain this atomically. We could, but it
317 // doesn't seem necessary.
318 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
319 return;
320 }
321 // Otherwise, we need to try again.
322 }
323 }
324
325 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
326 while (true) {
327 if (isFull()) {
328 _overflow = true;
329 return;
330 }
331 // Otherwise...
332 jint index = _index;
333 jint next_index = index + n;
334 if (next_index > _capacity) {
335 _overflow = true;
336 return;
337 }
338 jint res = Atomic::cmpxchg(next_index, &_index, index);
339 if (res == index) {
340 for (int i = 0; i < n; i++) {
341 int ind = index + i;
342 assert(ind < _capacity, "By overflow test above.");
343 _base[ind] = ptr_arr[i];
344 }
345 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
346 return;
347 }
348 // Otherwise, we need to try again.
349 }
350 }
351
352 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
353 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
354 jint start = _index;
355 jint next_index = start + n;
356 if (next_index > _capacity) {
357 _overflow = true;
358 return;
359 }
360 // Otherwise.
361 _index = next_index;
362 for (int i = 0; i < n; i++) {
363 int ind = start + i;
364 assert(ind < _capacity, "By overflow test above.");
365 _base[ind] = ptr_arr[i];
366 }
367 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
368 }
369
370 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
371 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
372 jint index = _index;
373 if (index == 0) {
374 *n = 0;
375 return false;
376 } else {
377 int k = MIN2(max, index);
378 jint new_ind = index - k;
379 for (int j = 0; j < k; j++) {
380 ptr_arr[j] = _base[new_ind + j];
381 }
382 _index = new_ind;
383 *n = k;
384 return true;
385 }
386 }
387
388 template<class OopClosureClass>
389 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
390 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
391 || SafepointSynchronize::is_at_safepoint(),
392 "Drain recursion must be yield-safe.");
393 bool res = true;
394 debug_only(_drain_in_progress = true);
395 debug_only(_drain_in_progress_yields = yield_after);
396 while (!isEmpty()) {
397 oop newOop = pop();
398 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
399 assert(newOop->is_oop(), "Expected an oop");
400 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
401 "only grey objects on this stack");
402 newOop->oop_iterate(cl);
403 if (yield_after && _cm->do_yield_check()) {
404 res = false;
405 break;
406 }
407 }
408 debug_only(_drain_in_progress = false);
409 return res;
410 }
411
412 void CMMarkStack::note_start_of_gc() {
413 assert(_saved_index == -1,
414 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
415 _saved_index = _index;
416 }
417
418 void CMMarkStack::note_end_of_gc() {
419 // This is intentionally a guarantee, instead of an assert. If we
420 // accidentally add something to the mark stack during GC, it
421 // will be a correctness issue so it's better if we crash. we'll
422 // only check this once per GC anyway, so it won't be a performance
423 // issue in any way.
424 guarantee(_saved_index == _index,
425 err_msg("saved index: %d index: %d", _saved_index, _index));
426 _saved_index = -1;
427 }
428
429 void CMMarkStack::oops_do(OopClosure* f) {
430 assert(_saved_index == _index,
431 err_msg("saved index: %d index: %d", _saved_index, _index));
432 for (int i = 0; i < _index; i += 1) {
433 f->do_oop(&_base[i]);
434 }
435 }
436
437 CMRootRegions::CMRootRegions() :
438 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
439 _should_abort(false), _next_survivor(NULL) { }
440
441 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
442 _young_list = g1h->young_list();
443 _cm = cm;
444 }
445
446 void CMRootRegions::prepare_for_scan() {
447 assert(!scan_in_progress(), "pre-condition");
448
449 // Currently, only survivors can be root regions.
450 assert(_next_survivor == NULL, "pre-condition");
451 _next_survivor = _young_list->first_survivor_region();
452 _scan_in_progress = (_next_survivor != NULL);
453 _should_abort = false;
454 }
455
456 HeapRegion* CMRootRegions::claim_next() {
457 if (_should_abort) {
458 // If someone has set the should_abort flag, we return NULL to
459 // force the caller to bail out of their loop.
460 return NULL;
461 }
462
463 // Currently, only survivors can be root regions.
464 HeapRegion* res = _next_survivor;
465 if (res != NULL) {
466 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
467 // Read it again in case it changed while we were waiting for the lock.
468 res = _next_survivor;
469 if (res != NULL) {
470 if (res == _young_list->last_survivor_region()) {
471 // We just claimed the last survivor so store NULL to indicate
472 // that we're done.
473 _next_survivor = NULL;
474 } else {
475 _next_survivor = res->get_next_young_region();
476 }
477 } else {
478 // Someone else claimed the last survivor while we were trying
479 // to take the lock so nothing else to do.
480 }
481 }
482 assert(res == NULL || res->is_survivor(), "post-condition");
483
484 return res;
485 }
486
487 void CMRootRegions::scan_finished() {
488 assert(scan_in_progress(), "pre-condition");
489
490 // Currently, only survivors can be root regions.
491 if (!_should_abort) {
492 assert(_next_survivor == NULL, "we should have claimed all survivors");
493 }
494 _next_survivor = NULL;
495
496 {
497 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
498 _scan_in_progress = false;
499 RootRegionScan_lock->notify_all();
500 }
501 }
502
503 bool CMRootRegions::wait_until_scan_finished() {
504 if (!scan_in_progress()) return false;
505
506 {
507 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
508 while (scan_in_progress()) {
509 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
510 }
511 }
512 return true;
513 }
514
515 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
516 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
517 #endif // _MSC_VER
518
519 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
520 return MAX2((n_par_threads + 2) / 4, 1U);
521 }
522
523 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
524 _g1h(g1h),
525 _markBitMap1(),
526 _markBitMap2(),
527 _parallel_marking_threads(0),
528 _max_parallel_marking_threads(0),
529 _sleep_factor(0.0),
530 _marking_task_overhead(1.0),
531 _cleanup_sleep_factor(0.0),
532 _cleanup_task_overhead(1.0),
533 _cleanup_list("Cleanup List"),
534 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
535 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
536 CardTableModRefBS::card_shift,
537 false /* in_resource_area*/),
538
539 _prevMarkBitMap(&_markBitMap1),
540 _nextMarkBitMap(&_markBitMap2),
541
542 _markStack(this),
543 // _finger set in set_non_marking_state
544
545 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
546 // _active_tasks set in set_non_marking_state
547 // _tasks set inside the constructor
548 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
549 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
550
551 _has_overflown(false),
552 _concurrent(false),
553 _has_aborted(false),
554 _aborted_gc_id(GCId::undefined()),
555 _restart_for_overflow(false),
556 _concurrent_marking_in_progress(false),
557
558 // _verbose_level set below
559
560 _init_times(),
561 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
562 _cleanup_times(),
563 _total_counting_time(0.0),
564 _total_rs_scrub_time(0.0),
565
566 _parallel_workers(NULL),
567
568 _count_card_bitmaps(NULL),
569 _count_marked_bytes(NULL),
570 _completed_initialization(false) {
571 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
572 if (verbose_level < no_verbose) {
573 verbose_level = no_verbose;
574 }
575 if (verbose_level > high_verbose) {
576 verbose_level = high_verbose;
577 }
578 _verbose_level = verbose_level;
579
580 if (verbose_low()) {
581 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
582 "heap end = " PTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
583 }
584
585 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
586 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
587
588 // Create & start a ConcurrentMark thread.
589 _cmThread = new ConcurrentMarkThread(this);
590 assert(cmThread() != NULL, "CM Thread should have been created");
591 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
592 if (_cmThread->osthread() == NULL) {
593 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
594 }
595
596 assert(CGC_lock != NULL, "Where's the CGC_lock?");
597 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
598 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
599
600 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
601 satb_qs.set_buffer_size(G1SATBBufferSize);
602
603 _root_regions.init(_g1h, this);
604
605 if (ConcGCThreads > ParallelGCThreads) {
606 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
607 "than ParallelGCThreads (" UINTX_FORMAT ").",
608 ConcGCThreads, ParallelGCThreads);
609 return;
610 }
611 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
612 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
613 // if both are set
614 _sleep_factor = 0.0;
615 _marking_task_overhead = 1.0;
616 } else if (G1MarkingOverheadPercent > 0) {
617 // We will calculate the number of parallel marking threads based
618 // on a target overhead with respect to the soft real-time goal
619 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
620 double overall_cm_overhead =
621 (double) MaxGCPauseMillis * marking_overhead /
622 (double) GCPauseIntervalMillis;
623 double cpu_ratio = 1.0 / (double) os::processor_count();
624 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
625 double marking_task_overhead =
626 overall_cm_overhead / marking_thread_num *
627 (double) os::processor_count();
628 double sleep_factor =
629 (1.0 - marking_task_overhead) / marking_task_overhead;
630
631 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
632 _sleep_factor = sleep_factor;
633 _marking_task_overhead = marking_task_overhead;
634 } else {
635 // Calculate the number of parallel marking threads by scaling
636 // the number of parallel GC threads.
637 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
638 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
639 _sleep_factor = 0.0;
640 _marking_task_overhead = 1.0;
641 }
642
643 assert(ConcGCThreads > 0, "Should have been set");
644 _parallel_marking_threads = (uint) ConcGCThreads;
645 _max_parallel_marking_threads = _parallel_marking_threads;
646
647 if (parallel_marking_threads() > 1) {
648 _cleanup_task_overhead = 1.0;
649 } else {
650 _cleanup_task_overhead = marking_task_overhead();
651 }
652 _cleanup_sleep_factor =
653 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
654
655 #if 0
656 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
657 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
658 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
659 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
660 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
661 #endif
662
663 guarantee(parallel_marking_threads() > 0, "peace of mind");
664 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
665 _max_parallel_marking_threads, false, true);
666 if (_parallel_workers == NULL) {
667 vm_exit_during_initialization("Failed necessary allocation.");
668 } else {
669 _parallel_workers->initialize_workers();
670 }
671
672 if (FLAG_IS_DEFAULT(MarkStackSize)) {
673 uintx mark_stack_size =
674 MIN2(MarkStackSizeMax,
675 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
676 // Verify that the calculated value for MarkStackSize is in range.
677 // It would be nice to use the private utility routine from Arguments.
678 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
679 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
680 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
681 mark_stack_size, (uintx) 1, MarkStackSizeMax);
682 return;
683 }
684 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
685 } else {
686 // Verify MarkStackSize is in range.
687 if (FLAG_IS_CMDLINE(MarkStackSize)) {
688 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
689 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
690 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
691 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
692 MarkStackSize, (uintx) 1, MarkStackSizeMax);
693 return;
694 }
695 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
696 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
697 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
698 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
699 MarkStackSize, MarkStackSizeMax);
700 return;
701 }
702 }
703 }
704 }
705
706 if (!_markStack.allocate(MarkStackSize)) {
707 warning("Failed to allocate CM marking stack");
708 return;
709 }
710
711 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
712 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
713
714 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
715 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
716
717 BitMap::idx_t card_bm_size = _card_bm.size();
718
719 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
720 _active_tasks = _max_worker_id;
721
722 size_t max_regions = (size_t) _g1h->max_regions();
723 for (uint i = 0; i < _max_worker_id; ++i) {
724 CMTaskQueue* task_queue = new CMTaskQueue();
725 task_queue->initialize();
726 _task_queues->register_queue(i, task_queue);
727
728 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
729 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
730
731 _tasks[i] = new CMTask(i, this,
732 _count_marked_bytes[i],
733 &_count_card_bitmaps[i],
734 task_queue, _task_queues);
735
736 _accum_task_vtime[i] = 0.0;
737 }
738
739 // Calculate the card number for the bottom of the heap. Used
740 // in biasing indexes into the accounting card bitmaps.
741 _heap_bottom_card_num =
742 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
743 CardTableModRefBS::card_shift);
744
745 // Clear all the liveness counting data
746 clear_all_count_data();
747
748 // so that the call below can read a sensible value
749 _heap_start = g1h->reserved_region().start();
750 set_non_marking_state();
751 _completed_initialization = true;
752 }
753
754 void ConcurrentMark::reset() {
755 // Starting values for these two. This should be called in a STW
756 // phase.
757 MemRegion reserved = _g1h->g1_reserved();
758 _heap_start = reserved.start();
759 _heap_end = reserved.end();
760
761 // Separated the asserts so that we know which one fires.
762 assert(_heap_start != NULL, "heap bounds should look ok");
763 assert(_heap_end != NULL, "heap bounds should look ok");
764 assert(_heap_start < _heap_end, "heap bounds should look ok");
765
766 // Reset all the marking data structures and any necessary flags
767 reset_marking_state();
768
769 if (verbose_low()) {
770 gclog_or_tty->print_cr("[global] resetting");
771 }
772
773 // We do reset all of them, since different phases will use
774 // different number of active threads. So, it's easiest to have all
775 // of them ready.
776 for (uint i = 0; i < _max_worker_id; ++i) {
777 _tasks[i]->reset(_nextMarkBitMap);
778 }
779
780 // we need this to make sure that the flag is on during the evac
781 // pause with initial mark piggy-backed
782 set_concurrent_marking_in_progress();
783 }
784
785
786 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
787 _markStack.set_should_expand();
788 _markStack.setEmpty(); // Also clears the _markStack overflow flag
789 if (clear_overflow) {
790 clear_has_overflown();
791 } else {
792 assert(has_overflown(), "pre-condition");
793 }
794 _finger = _heap_start;
795
796 for (uint i = 0; i < _max_worker_id; ++i) {
797 CMTaskQueue* queue = _task_queues->queue(i);
798 queue->set_empty();
799 }
800 }
801
802 void ConcurrentMark::set_concurrency(uint active_tasks) {
803 assert(active_tasks <= _max_worker_id, "we should not have more");
804
805 _active_tasks = active_tasks;
806 // Need to update the three data structures below according to the
807 // number of active threads for this phase.
808 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
809 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
810 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
811 }
812
813 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
814 set_concurrency(active_tasks);
815
816 _concurrent = concurrent;
817 // We propagate this to all tasks, not just the active ones.
818 for (uint i = 0; i < _max_worker_id; ++i)
819 _tasks[i]->set_concurrent(concurrent);
820
821 if (concurrent) {
822 set_concurrent_marking_in_progress();
823 } else {
824 // We currently assume that the concurrent flag has been set to
825 // false before we start remark. At this point we should also be
826 // in a STW phase.
827 assert(!concurrent_marking_in_progress(), "invariant");
828 assert(out_of_regions(),
829 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
830 p2i(_finger), p2i(_heap_end)));
831 }
832 }
833
834 void ConcurrentMark::set_non_marking_state() {
835 // We set the global marking state to some default values when we're
836 // not doing marking.
837 reset_marking_state();
838 _active_tasks = 0;
839 clear_concurrent_marking_in_progress();
840 }
841
842 ConcurrentMark::~ConcurrentMark() {
843 // The ConcurrentMark instance is never freed.
844 ShouldNotReachHere();
845 }
846
847 void ConcurrentMark::clearNextBitmap() {
848 G1CollectedHeap* g1h = G1CollectedHeap::heap();
849
850 // Make sure that the concurrent mark thread looks to still be in
851 // the current cycle.
852 guarantee(cmThread()->during_cycle(), "invariant");
853
854 // We are finishing up the current cycle by clearing the next
855 // marking bitmap and getting it ready for the next cycle. During
856 // this time no other cycle can start. So, let's make sure that this
857 // is the case.
858 guarantee(!g1h->mark_in_progress(), "invariant");
859
860 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
861 g1h->heap_region_iterate(&cl);
862
863 // Clear the liveness counting data. If the marking has been aborted, the abort()
864 // call already did that.
865 if (cl.complete()) {
866 clear_all_count_data();
867 }
868
869 // Repeat the asserts from above.
870 guarantee(cmThread()->during_cycle(), "invariant");
871 guarantee(!g1h->mark_in_progress(), "invariant");
872 }
873
874 class CheckBitmapClearHRClosure : public HeapRegionClosure {
875 CMBitMap* _bitmap;
876 bool _error;
877 public:
878 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
879 }
880
881 virtual bool doHeapRegion(HeapRegion* r) {
882 // This closure can be called concurrently to the mutator, so we must make sure
883 // that the result of the getNextMarkedWordAddress() call is compared to the
884 // value passed to it as limit to detect any found bits.
885 // We can use the region's orig_end() for the limit and the comparison value
886 // as it always contains the "real" end of the region that never changes and
887 // has no side effects.
888 // Due to the latter, there can also be no problem with the compiler generating
889 // reloads of the orig_end() call.
890 HeapWord* end = r->orig_end();
891 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
892 }
893 };
894
895 bool ConcurrentMark::nextMarkBitmapIsClear() {
896 CheckBitmapClearHRClosure cl(_nextMarkBitMap);
897 _g1h->heap_region_iterate(&cl);
898 return cl.complete();
899 }
900
901 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
902 public:
903 bool doHeapRegion(HeapRegion* r) {
904 if (!r->is_continues_humongous()) {
905 r->note_start_of_marking();
906 }
907 return false;
908 }
909 };
910
911 void ConcurrentMark::checkpointRootsInitialPre() {
912 G1CollectedHeap* g1h = G1CollectedHeap::heap();
913 G1CollectorPolicy* g1p = g1h->g1_policy();
914
915 _has_aborted = false;
916
917 #ifndef PRODUCT
918 if (G1PrintReachableAtInitialMark) {
919 print_reachable("at-cycle-start",
920 VerifyOption_G1UsePrevMarking, true /* all */);
921 }
922 #endif
923
924 // Initialize marking structures. This has to be done in a STW phase.
925 reset();
926
927 // For each region note start of marking.
928 NoteStartOfMarkHRClosure startcl;
929 g1h->heap_region_iterate(&startcl);
930 }
931
932
933 void ConcurrentMark::checkpointRootsInitialPost() {
934 G1CollectedHeap* g1h = G1CollectedHeap::heap();
935
936 // If we force an overflow during remark, the remark operation will
937 // actually abort and we'll restart concurrent marking. If we always
938 // force an overflow during remark we'll never actually complete the
939 // marking phase. So, we initialize this here, at the start of the
940 // cycle, so that at the remaining overflow number will decrease at
941 // every remark and we'll eventually not need to cause one.
942 force_overflow_stw()->init();
943
944 // Start Concurrent Marking weak-reference discovery.
945 ReferenceProcessor* rp = g1h->ref_processor_cm();
946 // enable ("weak") refs discovery
947 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
948 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
949
950 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
951 // This is the start of the marking cycle, we're expected all
952 // threads to have SATB queues with active set to false.
953 satb_mq_set.set_active_all_threads(true, /* new active value */
954 false /* expected_active */);
955
956 _root_regions.prepare_for_scan();
957
958 // update_g1_committed() will be called at the end of an evac pause
959 // when marking is on. So, it's also called at the end of the
960 // initial-mark pause to update the heap end, if the heap expands
961 // during it. No need to call it here.
962 }
963
964 /*
965 * Notice that in the next two methods, we actually leave the STS
966 * during the barrier sync and join it immediately afterwards. If we
967 * do not do this, the following deadlock can occur: one thread could
968 * be in the barrier sync code, waiting for the other thread to also
969 * sync up, whereas another one could be trying to yield, while also
970 * waiting for the other threads to sync up too.
971 *
972 * Note, however, that this code is also used during remark and in
973 * this case we should not attempt to leave / enter the STS, otherwise
974 * we'll either hit an assert (debug / fastdebug) or deadlock
975 * (product). So we should only leave / enter the STS if we are
976 * operating concurrently.
977 *
978 * Because the thread that does the sync barrier has left the STS, it
979 * is possible to be suspended for a Full GC or an evacuation pause
980 * could occur. This is actually safe, since the entering the sync
981 * barrier is one of the last things do_marking_step() does, and it
982 * doesn't manipulate any data structures afterwards.
983 */
984
985 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
986 if (verbose_low()) {
987 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
988 }
989
990 if (concurrent()) {
991 SuspendibleThreadSet::leave();
992 }
993
994 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
995
996 if (concurrent()) {
997 SuspendibleThreadSet::join();
998 }
999 // at this point everyone should have synced up and not be doing any
1000 // more work
1001
1002 if (verbose_low()) {
1003 if (barrier_aborted) {
1004 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1005 } else {
1006 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1007 }
1008 }
1009
1010 if (barrier_aborted) {
1011 // If the barrier aborted we ignore the overflow condition and
1012 // just abort the whole marking phase as quickly as possible.
1013 return;
1014 }
1015
1016 // If we're executing the concurrent phase of marking, reset the marking
1017 // state; otherwise the marking state is reset after reference processing,
1018 // during the remark pause.
1019 // If we reset here as a result of an overflow during the remark we will
1020 // see assertion failures from any subsequent set_concurrency_and_phase()
1021 // calls.
1022 if (concurrent()) {
1023 // let the task associated with with worker 0 do this
1024 if (worker_id == 0) {
1025 // task 0 is responsible for clearing the global data structures
1026 // We should be here because of an overflow. During STW we should
1027 // not clear the overflow flag since we rely on it being true when
1028 // we exit this method to abort the pause and restart concurrent
1029 // marking.
1030 reset_marking_state(true /* clear_overflow */);
1031 force_overflow()->update();
1032
1033 if (G1Log::fine()) {
1034 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1035 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1036 }
1037 }
1038 }
1039
1040 // after this, each task should reset its own data structures then
1041 // then go into the second barrier
1042 }
1043
1044 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1045 if (verbose_low()) {
1046 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1047 }
1048
1049 if (concurrent()) {
1050 SuspendibleThreadSet::leave();
1051 }
1052
1053 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1054
1055 if (concurrent()) {
1056 SuspendibleThreadSet::join();
1057 }
1058 // at this point everything should be re-initialized and ready to go
1059
1060 if (verbose_low()) {
1061 if (barrier_aborted) {
1062 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1063 } else {
1064 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1065 }
1066 }
1067 }
1068
1069 #ifndef PRODUCT
1070 void ForceOverflowSettings::init() {
1071 _num_remaining = G1ConcMarkForceOverflow;
1072 _force = false;
1073 update();
1074 }
1075
1076 void ForceOverflowSettings::update() {
1077 if (_num_remaining > 0) {
1078 _num_remaining -= 1;
1079 _force = true;
1080 } else {
1081 _force = false;
1082 }
1083 }
1084
1085 bool ForceOverflowSettings::should_force() {
1086 if (_force) {
1087 _force = false;
1088 return true;
1089 } else {
1090 return false;
1091 }
1092 }
1093 #endif // !PRODUCT
1094
1095 class CMConcurrentMarkingTask: public AbstractGangTask {
1096 private:
1097 ConcurrentMark* _cm;
1098 ConcurrentMarkThread* _cmt;
1099
1100 public:
1101 void work(uint worker_id) {
1102 assert(Thread::current()->is_ConcurrentGC_thread(),
1103 "this should only be done by a conc GC thread");
1104 ResourceMark rm;
1105
1106 double start_vtime = os::elapsedVTime();
1107
1108 SuspendibleThreadSet::join();
1109
1110 assert(worker_id < _cm->active_tasks(), "invariant");
1111 CMTask* the_task = _cm->task(worker_id);
1112 the_task->record_start_time();
1113 if (!_cm->has_aborted()) {
1114 do {
1115 double start_vtime_sec = os::elapsedVTime();
1116 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1117
1118 the_task->do_marking_step(mark_step_duration_ms,
1119 true /* do_termination */,
1120 false /* is_serial*/);
1121
1122 double end_vtime_sec = os::elapsedVTime();
1123 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1124 _cm->clear_has_overflown();
1125
1126 _cm->do_yield_check(worker_id);
1127
1128 jlong sleep_time_ms;
1129 if (!_cm->has_aborted() && the_task->has_aborted()) {
1130 sleep_time_ms =
1131 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1132 SuspendibleThreadSet::leave();
1133 os::sleep(Thread::current(), sleep_time_ms, false);
1134 SuspendibleThreadSet::join();
1135 }
1136 } while (!_cm->has_aborted() && the_task->has_aborted());
1137 }
1138 the_task->record_end_time();
1139 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1140
1141 SuspendibleThreadSet::leave();
1142
1143 double end_vtime = os::elapsedVTime();
1144 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1145 }
1146
1147 CMConcurrentMarkingTask(ConcurrentMark* cm,
1148 ConcurrentMarkThread* cmt) :
1149 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1150
1151 ~CMConcurrentMarkingTask() { }
1152 };
1153
1154 // Calculates the number of active workers for a concurrent
1155 // phase.
1156 uint ConcurrentMark::calc_parallel_marking_threads() {
1157 uint n_conc_workers = 0;
1158 if (!UseDynamicNumberOfGCThreads ||
1159 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1160 !ForceDynamicNumberOfGCThreads)) {
1161 n_conc_workers = max_parallel_marking_threads();
1162 } else {
1163 n_conc_workers =
1164 AdaptiveSizePolicy::calc_default_active_workers(
1165 max_parallel_marking_threads(),
1166 1, /* Minimum workers */
1167 parallel_marking_threads(),
1168 Threads::number_of_non_daemon_threads());
1169 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1170 // that scaling has already gone into "_max_parallel_marking_threads".
1171 }
1172 assert(n_conc_workers > 0, "Always need at least 1");
1173 return n_conc_workers;
1174 }
1175
1176 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1177 // Currently, only survivors can be root regions.
1178 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1179 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1180
1181 const uintx interval = PrefetchScanIntervalInBytes;
1182 HeapWord* curr = hr->bottom();
1183 const HeapWord* end = hr->top();
1184 while (curr < end) {
1185 Prefetch::read(curr, interval);
1186 oop obj = oop(curr);
1187 int size = obj->oop_iterate(&cl);
1188 assert(size == obj->size(), "sanity");
1189 curr += size;
1190 }
1191 }
1192
1193 class CMRootRegionScanTask : public AbstractGangTask {
1194 private:
1195 ConcurrentMark* _cm;
1196
1197 public:
1198 CMRootRegionScanTask(ConcurrentMark* cm) :
1199 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1200
1201 void work(uint worker_id) {
1202 assert(Thread::current()->is_ConcurrentGC_thread(),
1203 "this should only be done by a conc GC thread");
1204
1205 CMRootRegions* root_regions = _cm->root_regions();
1206 HeapRegion* hr = root_regions->claim_next();
1207 while (hr != NULL) {
1208 _cm->scanRootRegion(hr, worker_id);
1209 hr = root_regions->claim_next();
1210 }
1211 }
1212 };
1213
1214 void ConcurrentMark::scanRootRegions() {
1215 // Start of concurrent marking.
1216 ClassLoaderDataGraph::clear_claimed_marks();
1217
1218 // scan_in_progress() will have been set to true only if there was
1219 // at least one root region to scan. So, if it's false, we
1220 // should not attempt to do any further work.
1221 if (root_regions()->scan_in_progress()) {
1222 _parallel_marking_threads = calc_parallel_marking_threads();
1223 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1224 "Maximum number of marking threads exceeded");
1225 uint active_workers = MAX2(1U, parallel_marking_threads());
1226
1227 CMRootRegionScanTask task(this);
1228 if (use_parallel_marking_threads()) {
1229 _parallel_workers->set_active_workers((int) active_workers);
1230 _parallel_workers->run_task(&task);
1231 } else {
1232 task.work(0);
1233 }
1234
1235 // It's possible that has_aborted() is true here without actually
1236 // aborting the survivor scan earlier. This is OK as it's
1237 // mainly used for sanity checking.
1238 root_regions()->scan_finished();
1239 }
1240 }
1241
1242 void ConcurrentMark::markFromRoots() {
1243 // we might be tempted to assert that:
1244 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1245 // "inconsistent argument?");
1246 // However that wouldn't be right, because it's possible that
1247 // a safepoint is indeed in progress as a younger generation
1248 // stop-the-world GC happens even as we mark in this generation.
1249
1250 _restart_for_overflow = false;
1251 force_overflow_conc()->init();
1252
1253 // _g1h has _n_par_threads
1254 _parallel_marking_threads = calc_parallel_marking_threads();
1255 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1256 "Maximum number of marking threads exceeded");
1257
1258 uint active_workers = MAX2(1U, parallel_marking_threads());
1259
1260 // Parallel task terminator is set in "set_concurrency_and_phase()"
1261 set_concurrency_and_phase(active_workers, true /* concurrent */);
1262
1263 CMConcurrentMarkingTask markingTask(this, cmThread());
1264 if (use_parallel_marking_threads()) {
1265 _parallel_workers->set_active_workers((int)active_workers);
1266 // Don't set _n_par_threads because it affects MT in process_roots()
1267 // and the decisions on that MT processing is made elsewhere.
1268 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1269 _parallel_workers->run_task(&markingTask);
1270 } else {
1271 markingTask.work(0);
1272 }
1273 print_stats();
1274 }
1275
1276 // Helper class to get rid of some boilerplate code.
1277 class G1CMTraceTime : public GCTraceTime {
1278 static bool doit_and_prepend(bool doit) {
1279 if (doit) {
1280 gclog_or_tty->put(' ');
1281 }
1282 return doit;
1283 }
1284
1285 public:
1286 G1CMTraceTime(const char* title, bool doit)
1287 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
1288 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
1289 }
1290 };
1291
1292 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1293 // world is stopped at this checkpoint
1294 assert(SafepointSynchronize::is_at_safepoint(),
1295 "world should be stopped");
1296
1297 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1298
1299 // If a full collection has happened, we shouldn't do this.
1300 if (has_aborted()) {
1301 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1302 return;
1303 }
1304
1305 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1306
1307 if (VerifyDuringGC) {
1308 HandleMark hm; // handle scope
1309 Universe::heap()->prepare_for_verify();
1310 Universe::verify(VerifyOption_G1UsePrevMarking,
1311 " VerifyDuringGC:(before)");
1312 }
1313 g1h->check_bitmaps("Remark Start");
1314
1315 G1CollectorPolicy* g1p = g1h->g1_policy();
1316 g1p->record_concurrent_mark_remark_start();
1317
1318 double start = os::elapsedTime();
1319
1320 checkpointRootsFinalWork();
1321
1322 double mark_work_end = os::elapsedTime();
1323
1324 weakRefsWork(clear_all_soft_refs);
1325
1326 if (has_overflown()) {
1327 // Oops. We overflowed. Restart concurrent marking.
1328 _restart_for_overflow = true;
1329 if (G1TraceMarkStackOverflow) {
1330 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1331 }
1332
1333 // Verify the heap w.r.t. the previous marking bitmap.
1334 if (VerifyDuringGC) {
1335 HandleMark hm; // handle scope
1336 Universe::heap()->prepare_for_verify();
1337 Universe::verify(VerifyOption_G1UsePrevMarking,
1338 " VerifyDuringGC:(overflow)");
1339 }
1340
1341 // Clear the marking state because we will be restarting
1342 // marking due to overflowing the global mark stack.
1343 reset_marking_state();
1344 } else {
1345 {
1346 G1CMTraceTime trace("GC aggregate-data", G1Log::finer());
1347
1348 // Aggregate the per-task counting data that we have accumulated
1349 // while marking.
1350 aggregate_count_data();
1351 }
1352
1353 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1354 // We're done with marking.
1355 // This is the end of the marking cycle, we're expected all
1356 // threads to have SATB queues with active set to true.
1357 satb_mq_set.set_active_all_threads(false, /* new active value */
1358 true /* expected_active */);
1359
1360 if (VerifyDuringGC) {
1361 HandleMark hm; // handle scope
1362 Universe::heap()->prepare_for_verify();
1363 Universe::verify(VerifyOption_G1UseNextMarking,
1364 " VerifyDuringGC:(after)");
1365 }
1366 g1h->check_bitmaps("Remark End");
1367 assert(!restart_for_overflow(), "sanity");
1368 // Completely reset the marking state since marking completed
1369 set_non_marking_state();
1370 }
1371
1372 // Expand the marking stack, if we have to and if we can.
1373 if (_markStack.should_expand()) {
1374 _markStack.expand();
1375 }
1376
1377 // Statistics
1378 double now = os::elapsedTime();
1379 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1380 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1381 _remark_times.add((now - start) * 1000.0);
1382
1383 g1p->record_concurrent_mark_remark_end();
1384
1385 G1CMIsAliveClosure is_alive(g1h);
1386 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1387 }
1388
1389 // Base class of the closures that finalize and verify the
1390 // liveness counting data.
1391 class CMCountDataClosureBase: public HeapRegionClosure {
1392 protected:
1393 G1CollectedHeap* _g1h;
1394 ConcurrentMark* _cm;
1395 CardTableModRefBS* _ct_bs;
1396
1397 BitMap* _region_bm;
1398 BitMap* _card_bm;
1399
1400 // Takes a region that's not empty (i.e., it has at least one
1401 // live object in it and sets its corresponding bit on the region
1402 // bitmap to 1. If the region is "starts humongous" it will also set
1403 // to 1 the bits on the region bitmap that correspond to its
1404 // associated "continues humongous" regions.
1405 void set_bit_for_region(HeapRegion* hr) {
1406 assert(!hr->is_continues_humongous(), "should have filtered those out");
1407
1408 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1409 if (!hr->is_starts_humongous()) {
1410 // Normal (non-humongous) case: just set the bit.
1411 _region_bm->par_at_put(index, true);
1412 } else {
1413 // Starts humongous case: calculate how many regions are part of
1414 // this humongous region and then set the bit range.
1415 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1416 _region_bm->par_at_put_range(index, end_index, true);
1417 }
1418 }
1419
1420 public:
1421 CMCountDataClosureBase(G1CollectedHeap* g1h,
1422 BitMap* region_bm, BitMap* card_bm):
1423 _g1h(g1h), _cm(g1h->concurrent_mark()),
1424 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1425 _region_bm(region_bm), _card_bm(card_bm) { }
1426 };
1427
1428 // Closure that calculates the # live objects per region. Used
1429 // for verification purposes during the cleanup pause.
1430 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1431 CMBitMapRO* _bm;
1432 size_t _region_marked_bytes;
1433
1434 public:
1435 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1436 BitMap* region_bm, BitMap* card_bm) :
1437 CMCountDataClosureBase(g1h, region_bm, card_bm),
1438 _bm(bm), _region_marked_bytes(0) { }
1439
1440 bool doHeapRegion(HeapRegion* hr) {
1441
1442 if (hr->is_continues_humongous()) {
1443 // We will ignore these here and process them when their
1444 // associated "starts humongous" region is processed (see
1445 // set_bit_for_heap_region()). Note that we cannot rely on their
1446 // associated "starts humongous" region to have their bit set to
1447 // 1 since, due to the region chunking in the parallel region
1448 // iteration, a "continues humongous" region might be visited
1449 // before its associated "starts humongous".
1450 return false;
1451 }
1452
1453 HeapWord* ntams = hr->next_top_at_mark_start();
1454 HeapWord* start = hr->bottom();
1455
1456 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1457 err_msg("Preconditions not met - "
1458 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1459 p2i(start), p2i(ntams), p2i(hr->end())));
1460
1461 // Find the first marked object at or after "start".
1462 start = _bm->getNextMarkedWordAddress(start, ntams);
1463
1464 size_t marked_bytes = 0;
1465
1466 while (start < ntams) {
1467 oop obj = oop(start);
1468 int obj_sz = obj->size();
1469 HeapWord* obj_end = start + obj_sz;
1470
1471 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1472 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1473
1474 // Note: if we're looking at the last region in heap - obj_end
1475 // could be actually just beyond the end of the heap; end_idx
1476 // will then correspond to a (non-existent) card that is also
1477 // just beyond the heap.
1478 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1479 // end of object is not card aligned - increment to cover
1480 // all the cards spanned by the object
1481 end_idx += 1;
1482 }
1483
1484 // Set the bits in the card BM for the cards spanned by this object.
1485 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1486
1487 // Add the size of this object to the number of marked bytes.
1488 marked_bytes += (size_t)obj_sz * HeapWordSize;
1489
1490 // Find the next marked object after this one.
1491 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1492 }
1493
1494 // Mark the allocated-since-marking portion...
1495 HeapWord* top = hr->top();
1496 if (ntams < top) {
1497 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1498 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1499
1500 // Note: if we're looking at the last region in heap - top
1501 // could be actually just beyond the end of the heap; end_idx
1502 // will then correspond to a (non-existent) card that is also
1503 // just beyond the heap.
1504 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1505 // end of object is not card aligned - increment to cover
1506 // all the cards spanned by the object
1507 end_idx += 1;
1508 }
1509 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1510
1511 // This definitely means the region has live objects.
1512 set_bit_for_region(hr);
1513 }
1514
1515 // Update the live region bitmap.
1516 if (marked_bytes > 0) {
1517 set_bit_for_region(hr);
1518 }
1519
1520 // Set the marked bytes for the current region so that
1521 // it can be queried by a calling verification routine
1522 _region_marked_bytes = marked_bytes;
1523
1524 return false;
1525 }
1526
1527 size_t region_marked_bytes() const { return _region_marked_bytes; }
1528 };
1529
1530 // Heap region closure used for verifying the counting data
1531 // that was accumulated concurrently and aggregated during
1532 // the remark pause. This closure is applied to the heap
1533 // regions during the STW cleanup pause.
1534
1535 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1536 G1CollectedHeap* _g1h;
1537 ConcurrentMark* _cm;
1538 CalcLiveObjectsClosure _calc_cl;
1539 BitMap* _region_bm; // Region BM to be verified
1540 BitMap* _card_bm; // Card BM to be verified
1541 bool _verbose; // verbose output?
1542
1543 BitMap* _exp_region_bm; // Expected Region BM values
1544 BitMap* _exp_card_bm; // Expected card BM values
1545
1546 int _failures;
1547
1548 public:
1549 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1550 BitMap* region_bm,
1551 BitMap* card_bm,
1552 BitMap* exp_region_bm,
1553 BitMap* exp_card_bm,
1554 bool verbose) :
1555 _g1h(g1h), _cm(g1h->concurrent_mark()),
1556 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1557 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1558 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1559 _failures(0) { }
1560
1561 int failures() const { return _failures; }
1562
1563 bool doHeapRegion(HeapRegion* hr) {
1564 if (hr->is_continues_humongous()) {
1565 // We will ignore these here and process them when their
1566 // associated "starts humongous" region is processed (see
1567 // set_bit_for_heap_region()). Note that we cannot rely on their
1568 // associated "starts humongous" region to have their bit set to
1569 // 1 since, due to the region chunking in the parallel region
1570 // iteration, a "continues humongous" region might be visited
1571 // before its associated "starts humongous".
1572 return false;
1573 }
1574
1575 int failures = 0;
1576
1577 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1578 // this region and set the corresponding bits in the expected region
1579 // and card bitmaps.
1580 bool res = _calc_cl.doHeapRegion(hr);
1581 assert(res == false, "should be continuing");
1582
1583 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1584 Mutex::_no_safepoint_check_flag);
1585
1586 // Verify the marked bytes for this region.
1587 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1588 size_t act_marked_bytes = hr->next_marked_bytes();
1589
1590 // We're not OK if expected marked bytes > actual marked bytes. It means
1591 // we have missed accounting some objects during the actual marking.
1592 if (exp_marked_bytes > act_marked_bytes) {
1593 if (_verbose) {
1594 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1595 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1596 hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1597 }
1598 failures += 1;
1599 }
1600
1601 // Verify the bit, for this region, in the actual and expected
1602 // (which was just calculated) region bit maps.
1603 // We're not OK if the bit in the calculated expected region
1604 // bitmap is set and the bit in the actual region bitmap is not.
1605 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1606
1607 bool expected = _exp_region_bm->at(index);
1608 bool actual = _region_bm->at(index);
1609 if (expected && !actual) {
1610 if (_verbose) {
1611 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1612 "expected: %s, actual: %s",
1613 hr->hrm_index(),
1614 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1615 }
1616 failures += 1;
1617 }
1618
1619 // Verify that the card bit maps for the cards spanned by the current
1620 // region match. We have an error if we have a set bit in the expected
1621 // bit map and the corresponding bit in the actual bitmap is not set.
1622
1623 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1624 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1625
1626 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1627 expected = _exp_card_bm->at(i);
1628 actual = _card_bm->at(i);
1629
1630 if (expected && !actual) {
1631 if (_verbose) {
1632 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1633 "expected: %s, actual: %s",
1634 hr->hrm_index(), i,
1635 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1636 }
1637 failures += 1;
1638 }
1639 }
1640
1641 if (failures > 0 && _verbose) {
1642 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1643 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1644 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1645 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1646 }
1647
1648 _failures += failures;
1649
1650 // We could stop iteration over the heap when we
1651 // find the first violating region by returning true.
1652 return false;
1653 }
1654 };
1655
1656 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1657 protected:
1658 G1CollectedHeap* _g1h;
1659 ConcurrentMark* _cm;
1660 BitMap* _actual_region_bm;
1661 BitMap* _actual_card_bm;
1662
1663 uint _n_workers;
1664
1665 BitMap* _expected_region_bm;
1666 BitMap* _expected_card_bm;
1667
1668 int _failures;
1669 bool _verbose;
1670
1671 HeapRegionClaimer _hrclaimer;
1672
1673 public:
1674 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1675 BitMap* region_bm, BitMap* card_bm,
1676 BitMap* expected_region_bm, BitMap* expected_card_bm)
1677 : AbstractGangTask("G1 verify final counting"),
1678 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1679 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1680 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1681 _failures(0), _verbose(false),
1682 _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1683 assert(VerifyDuringGC, "don't call this otherwise");
1684 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1685 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1686
1687 _verbose = _cm->verbose_medium();
1688 }
1689
1690 void work(uint worker_id) {
1691 assert(worker_id < _n_workers, "invariant");
1692
1693 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1694 _actual_region_bm, _actual_card_bm,
1695 _expected_region_bm,
1696 _expected_card_bm,
1697 _verbose);
1698
1699 _g1h->heap_region_par_iterate(&verify_cl, worker_id, &_hrclaimer);
1700
1701 Atomic::add(verify_cl.failures(), &_failures);
1702 }
1703
1704 int failures() const { return _failures; }
1705 };
1706
1707 // Closure that finalizes the liveness counting data.
1708 // Used during the cleanup pause.
1709 // Sets the bits corresponding to the interval [NTAMS, top]
1710 // (which contains the implicitly live objects) in the
1711 // card liveness bitmap. Also sets the bit for each region,
1712 // containing live data, in the region liveness bitmap.
1713
1714 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1715 public:
1716 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1717 BitMap* region_bm,
1718 BitMap* card_bm) :
1719 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1720
1721 bool doHeapRegion(HeapRegion* hr) {
1722
1723 if (hr->is_continues_humongous()) {
1724 // We will ignore these here and process them when their
1725 // associated "starts humongous" region is processed (see
1726 // set_bit_for_heap_region()). Note that we cannot rely on their
1727 // associated "starts humongous" region to have their bit set to
1728 // 1 since, due to the region chunking in the parallel region
1729 // iteration, a "continues humongous" region might be visited
1730 // before its associated "starts humongous".
1731 return false;
1732 }
1733
1734 HeapWord* ntams = hr->next_top_at_mark_start();
1735 HeapWord* top = hr->top();
1736
1737 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1738
1739 // Mark the allocated-since-marking portion...
1740 if (ntams < top) {
1741 // This definitely means the region has live objects.
1742 set_bit_for_region(hr);
1743
1744 // Now set the bits in the card bitmap for [ntams, top)
1745 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1746 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1747
1748 // Note: if we're looking at the last region in heap - top
1749 // could be actually just beyond the end of the heap; end_idx
1750 // will then correspond to a (non-existent) card that is also
1751 // just beyond the heap.
1752 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1753 // end of object is not card aligned - increment to cover
1754 // all the cards spanned by the object
1755 end_idx += 1;
1756 }
1757
1758 assert(end_idx <= _card_bm->size(),
1759 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1760 end_idx, _card_bm->size()));
1761 assert(start_idx < _card_bm->size(),
1762 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1763 start_idx, _card_bm->size()));
1764
1765 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1766 }
1767
1768 // Set the bit for the region if it contains live data
1769 if (hr->next_marked_bytes() > 0) {
1770 set_bit_for_region(hr);
1771 }
1772
1773 return false;
1774 }
1775 };
1776
1777 class G1ParFinalCountTask: public AbstractGangTask {
1778 protected:
1779 G1CollectedHeap* _g1h;
1780 ConcurrentMark* _cm;
1781 BitMap* _actual_region_bm;
1782 BitMap* _actual_card_bm;
1783
1784 uint _n_workers;
1785 HeapRegionClaimer _hrclaimer;
1786
1787 public:
1788 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1789 : AbstractGangTask("G1 final counting"),
1790 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1791 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1792 _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1793 }
1794
1795 void work(uint worker_id) {
1796 assert(worker_id < _n_workers, "invariant");
1797
1798 FinalCountDataUpdateClosure final_update_cl(_g1h,
1799 _actual_region_bm,
1800 _actual_card_bm);
1801
1802 _g1h->heap_region_par_iterate(&final_update_cl, worker_id, &_hrclaimer);
1803 }
1804 };
1805
1806 class G1ParNoteEndTask;
1807
1808 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1809 G1CollectedHeap* _g1;
1810 size_t _max_live_bytes;
1811 uint _regions_claimed;
1812 size_t _freed_bytes;
1813 FreeRegionList* _local_cleanup_list;
1814 HeapRegionSetCount _old_regions_removed;
1815 HeapRegionSetCount _humongous_regions_removed;
1816 HRRSCleanupTask* _hrrs_cleanup_task;
1817 double _claimed_region_time;
1818 double _max_region_time;
1819
1820 public:
1821 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1822 FreeRegionList* local_cleanup_list,
1823 HRRSCleanupTask* hrrs_cleanup_task) :
1824 _g1(g1),
1825 _max_live_bytes(0), _regions_claimed(0),
1826 _freed_bytes(0),
1827 _claimed_region_time(0.0), _max_region_time(0.0),
1828 _local_cleanup_list(local_cleanup_list),
1829 _old_regions_removed(),
1830 _humongous_regions_removed(),
1831 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1832
1833 size_t freed_bytes() { return _freed_bytes; }
1834 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1835 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1836
1837 bool doHeapRegion(HeapRegion *hr) {
1838 if (hr->is_continues_humongous()) {
1839 return false;
1840 }
1841 // We use a claim value of zero here because all regions
1842 // were claimed with value 1 in the FinalCount task.
1843 _g1->reset_gc_time_stamps(hr);
1844 double start = os::elapsedTime();
1845 _regions_claimed++;
1846 hr->note_end_of_marking();
1847 _max_live_bytes += hr->max_live_bytes();
1848
1849 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1850 _freed_bytes += hr->used();
1851 hr->set_containing_set(NULL);
1852 if (hr->is_humongous()) {
1853 assert(hr->is_starts_humongous(), "we should only see starts humongous");
1854 _humongous_regions_removed.increment(1u, hr->capacity());
1855 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1856 } else {
1857 _old_regions_removed.increment(1u, hr->capacity());
1858 _g1->free_region(hr, _local_cleanup_list, true);
1859 }
1860 } else {
1861 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1862 }
1863
1864 double region_time = (os::elapsedTime() - start);
1865 _claimed_region_time += region_time;
1866 if (region_time > _max_region_time) {
1867 _max_region_time = region_time;
1868 }
1869 return false;
1870 }
1871
1872 size_t max_live_bytes() { return _max_live_bytes; }
1873 uint regions_claimed() { return _regions_claimed; }
1874 double claimed_region_time_sec() { return _claimed_region_time; }
1875 double max_region_time_sec() { return _max_region_time; }
1876 };
1877
1878 class G1ParNoteEndTask: public AbstractGangTask {
1879 friend class G1NoteEndOfConcMarkClosure;
1880
1881 protected:
1882 G1CollectedHeap* _g1h;
1883 size_t _max_live_bytes;
1884 size_t _freed_bytes;
1885 FreeRegionList* _cleanup_list;
1886 HeapRegionClaimer _hrclaimer;
1887
1888 public:
1889 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1890 AbstractGangTask("G1 note end"), _g1h(g1h), _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1891 }
1892
1893 void work(uint worker_id) {
1894 double start = os::elapsedTime();
1895 FreeRegionList local_cleanup_list("Local Cleanup List");
1896 HRRSCleanupTask hrrs_cleanup_task;
1897 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1898 &hrrs_cleanup_task);
1899 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1900 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1901
1902 // Now update the lists
1903 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1904 {
1905 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1906 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1907 _max_live_bytes += g1_note_end.max_live_bytes();
1908 _freed_bytes += g1_note_end.freed_bytes();
1909
1910 // If we iterate over the global cleanup list at the end of
1911 // cleanup to do this printing we will not guarantee to only
1912 // generate output for the newly-reclaimed regions (the list
1913 // might not be empty at the beginning of cleanup; we might
1914 // still be working on its previous contents). So we do the
1915 // printing here, before we append the new regions to the global
1916 // cleanup list.
1917
1918 G1HRPrinter* hr_printer = _g1h->hr_printer();
1919 if (hr_printer->is_active()) {
1920 FreeRegionListIterator iter(&local_cleanup_list);
1921 while (iter.more_available()) {
1922 HeapRegion* hr = iter.get_next();
1923 hr_printer->cleanup(hr);
1924 }
1925 }
1926
1927 _cleanup_list->add_ordered(&local_cleanup_list);
1928 assert(local_cleanup_list.is_empty(), "post-condition");
1929
1930 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1931 }
1932 }
1933 size_t max_live_bytes() { return _max_live_bytes; }
1934 size_t freed_bytes() { return _freed_bytes; }
1935 };
1936
1937 class G1ParScrubRemSetTask: public AbstractGangTask {
1938 protected:
1939 G1RemSet* _g1rs;
1940 BitMap* _region_bm;
1941 BitMap* _card_bm;
1942 HeapRegionClaimer _hrclaimer;
1943
1944 public:
1945 G1ParScrubRemSetTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm, uint n_workers) :
1946 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), _region_bm(region_bm), _card_bm(card_bm), _hrclaimer(n_workers) {
1947 }
1948
1949 void work(uint worker_id) {
1950 _g1rs->scrub(_region_bm, _card_bm, worker_id, &_hrclaimer);
1951 }
1952
1953 };
1954
1955 void ConcurrentMark::cleanup() {
1956 // world is stopped at this checkpoint
1957 assert(SafepointSynchronize::is_at_safepoint(),
1958 "world should be stopped");
1959 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1960
1961 // If a full collection has happened, we shouldn't do this.
1962 if (has_aborted()) {
1963 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1964 return;
1965 }
1966
1967 g1h->verify_region_sets_optional();
1968
1969 if (VerifyDuringGC) {
1970 HandleMark hm; // handle scope
1971 Universe::heap()->prepare_for_verify();
1972 Universe::verify(VerifyOption_G1UsePrevMarking,
1973 " VerifyDuringGC:(before)");
1974 }
1975 g1h->check_bitmaps("Cleanup Start");
1976
1977 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
1978 g1p->record_concurrent_mark_cleanup_start();
1979
1980 double start = os::elapsedTime();
1981
1982 HeapRegionRemSet::reset_for_cleanup_tasks();
1983
1984 uint n_workers;
1985
1986 // Do counting once more with the world stopped for good measure.
1987 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1988
1989 g1h->set_par_threads();
1990 n_workers = g1h->n_par_threads();
1991 assert(g1h->n_par_threads() == n_workers,
1992 "Should not have been reset");
1993 g1h->workers()->run_task(&g1_par_count_task);
1994 // Done with the parallel phase so reset to 0.
1995 g1h->set_par_threads(0);
1996
1997 if (VerifyDuringGC) {
1998 // Verify that the counting data accumulated during marking matches
1999 // that calculated by walking the marking bitmap.
2000
2001 // Bitmaps to hold expected values
2002 BitMap expected_region_bm(_region_bm.size(), true);
2003 BitMap expected_card_bm(_card_bm.size(), true);
2004
2005 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2006 &_region_bm,
2007 &_card_bm,
2008 &expected_region_bm,
2009 &expected_card_bm);
2010
2011 g1h->set_par_threads((int)n_workers);
2012 g1h->workers()->run_task(&g1_par_verify_task);
2013 // Done with the parallel phase so reset to 0.
2014 g1h->set_par_threads(0);
2015
2016 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2017 }
2018
2019 size_t start_used_bytes = g1h->used();
2020 g1h->set_marking_complete();
2021
2022 double count_end = os::elapsedTime();
2023 double this_final_counting_time = (count_end - start);
2024 _total_counting_time += this_final_counting_time;
2025
2026 if (G1PrintRegionLivenessInfo) {
2027 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2028 _g1h->heap_region_iterate(&cl);
2029 }
2030
2031 // Install newly created mark bitMap as "prev".
2032 swapMarkBitMaps();
2033
2034 g1h->reset_gc_time_stamp();
2035
2036 // Note end of marking in all heap regions.
2037 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
2038 g1h->set_par_threads((int)n_workers);
2039 g1h->workers()->run_task(&g1_par_note_end_task);
2040 g1h->set_par_threads(0);
2041 g1h->check_gc_time_stamps();
2042
2043 if (!cleanup_list_is_empty()) {
2044 // The cleanup list is not empty, so we'll have to process it
2045 // concurrently. Notify anyone else that might be wanting free
2046 // regions that there will be more free regions coming soon.
2047 g1h->set_free_regions_coming();
2048 }
2049
2050 // call below, since it affects the metric by which we sort the heap
2051 // regions.
2052 if (G1ScrubRemSets) {
2053 double rs_scrub_start = os::elapsedTime();
2054 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm, n_workers);
2055 g1h->set_par_threads((int)n_workers);
2056 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2057 g1h->set_par_threads(0);
2058
2059 double rs_scrub_end = os::elapsedTime();
2060 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2061 _total_rs_scrub_time += this_rs_scrub_time;
2062 }
2063
2064 // this will also free any regions totally full of garbage objects,
2065 // and sort the regions.
2066 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2067
2068 // Statistics.
2069 double end = os::elapsedTime();
2070 _cleanup_times.add((end - start) * 1000.0);
2071
2072 if (G1Log::fine()) {
2073 g1h->print_size_transition(gclog_or_tty,
2074 start_used_bytes,
2075 g1h->used(),
2076 g1h->capacity());
2077 }
2078
2079 // Clean up will have freed any regions completely full of garbage.
2080 // Update the soft reference policy with the new heap occupancy.
2081 Universe::update_heap_info_at_gc();
2082
2083 if (VerifyDuringGC) {
2084 HandleMark hm; // handle scope
2085 Universe::heap()->prepare_for_verify();
2086 Universe::verify(VerifyOption_G1UsePrevMarking,
2087 " VerifyDuringGC:(after)");
2088 }
2089
2090 g1h->check_bitmaps("Cleanup End");
2091
2092 g1h->verify_region_sets_optional();
2093
2094 // We need to make this be a "collection" so any collection pause that
2095 // races with it goes around and waits for completeCleanup to finish.
2096 g1h->increment_total_collections();
2097
2098 // Clean out dead classes and update Metaspace sizes.
2099 if (ClassUnloadingWithConcurrentMark) {
2100 ClassLoaderDataGraph::purge();
2101 }
2102 MetaspaceGC::compute_new_size();
2103
2104 // We reclaimed old regions so we should calculate the sizes to make
2105 // sure we update the old gen/space data.
2106 g1h->g1mm()->update_sizes();
2107
2108 g1h->trace_heap_after_concurrent_cycle();
2109 }
2110
2111 void ConcurrentMark::completeCleanup() {
2112 if (has_aborted()) return;
2113
2114 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2115
2116 _cleanup_list.verify_optional();
2117 FreeRegionList tmp_free_list("Tmp Free List");
2118
2119 if (G1ConcRegionFreeingVerbose) {
2120 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2121 "cleanup list has %u entries",
2122 _cleanup_list.length());
2123 }
2124
2125 // No one else should be accessing the _cleanup_list at this point,
2126 // so it is not necessary to take any locks
2127 while (!_cleanup_list.is_empty()) {
2128 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2129 assert(hr != NULL, "Got NULL from a non-empty list");
2130 hr->par_clear();
2131 tmp_free_list.add_ordered(hr);
2132
2133 // Instead of adding one region at a time to the secondary_free_list,
2134 // we accumulate them in the local list and move them a few at a
2135 // time. This also cuts down on the number of notify_all() calls
2136 // we do during this process. We'll also append the local list when
2137 // _cleanup_list is empty (which means we just removed the last
2138 // region from the _cleanup_list).
2139 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2140 _cleanup_list.is_empty()) {
2141 if (G1ConcRegionFreeingVerbose) {
2142 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2143 "appending %u entries to the secondary_free_list, "
2144 "cleanup list still has %u entries",
2145 tmp_free_list.length(),
2146 _cleanup_list.length());
2147 }
2148
2149 {
2150 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2151 g1h->secondary_free_list_add(&tmp_free_list);
2152 SecondaryFreeList_lock->notify_all();
2153 }
2154
2155 if (G1StressConcRegionFreeing) {
2156 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2157 os::sleep(Thread::current(), (jlong) 1, false);
2158 }
2159 }
2160 }
2161 }
2162 assert(tmp_free_list.is_empty(), "post-condition");
2163 }
2164
2165 // Supporting Object and Oop closures for reference discovery
2166 // and processing in during marking
2167
2168 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2169 HeapWord* addr = (HeapWord*)obj;
2170 return addr != NULL &&
2171 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2172 }
2173
2174 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2175 // Uses the CMTask associated with a worker thread (for serial reference
2176 // processing the CMTask for worker 0 is used) to preserve (mark) and
2177 // trace referent objects.
2178 //
2179 // Using the CMTask and embedded local queues avoids having the worker
2180 // threads operating on the global mark stack. This reduces the risk
2181 // of overflowing the stack - which we would rather avoid at this late
2182 // state. Also using the tasks' local queues removes the potential
2183 // of the workers interfering with each other that could occur if
2184 // operating on the global stack.
2185
2186 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2187 ConcurrentMark* _cm;
2188 CMTask* _task;
2189 int _ref_counter_limit;
2190 int _ref_counter;
2191 bool _is_serial;
2192 public:
2193 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2194 _cm(cm), _task(task), _is_serial(is_serial),
2195 _ref_counter_limit(G1RefProcDrainInterval) {
2196 assert(_ref_counter_limit > 0, "sanity");
2197 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2198 _ref_counter = _ref_counter_limit;
2199 }
2200
2201 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2202 virtual void do_oop( oop* p) { do_oop_work(p); }
2203
2204 template <class T> void do_oop_work(T* p) {
2205 if (!_cm->has_overflown()) {
2206 oop obj = oopDesc::load_decode_heap_oop(p);
2207 if (_cm->verbose_high()) {
2208 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2209 "*"PTR_FORMAT" = "PTR_FORMAT,
2210 _task->worker_id(), p2i(p), p2i((void*) obj));
2211 }
2212
2213 _task->deal_with_reference(obj);
2214 _ref_counter--;
2215
2216 if (_ref_counter == 0) {
2217 // We have dealt with _ref_counter_limit references, pushing them
2218 // and objects reachable from them on to the local stack (and
2219 // possibly the global stack). Call CMTask::do_marking_step() to
2220 // process these entries.
2221 //
2222 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2223 // there's nothing more to do (i.e. we're done with the entries that
2224 // were pushed as a result of the CMTask::deal_with_reference() calls
2225 // above) or we overflow.
2226 //
2227 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2228 // flag while there may still be some work to do. (See the comment at
2229 // the beginning of CMTask::do_marking_step() for those conditions -
2230 // one of which is reaching the specified time target.) It is only
2231 // when CMTask::do_marking_step() returns without setting the
2232 // has_aborted() flag that the marking step has completed.
2233 do {
2234 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2235 _task->do_marking_step(mark_step_duration_ms,
2236 false /* do_termination */,
2237 _is_serial);
2238 } while (_task->has_aborted() && !_cm->has_overflown());
2239 _ref_counter = _ref_counter_limit;
2240 }
2241 } else {
2242 if (_cm->verbose_high()) {
2243 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2244 }
2245 }
2246 }
2247 };
2248
2249 // 'Drain' oop closure used by both serial and parallel reference processing.
2250 // Uses the CMTask associated with a given worker thread (for serial
2251 // reference processing the CMtask for worker 0 is used). Calls the
2252 // do_marking_step routine, with an unbelievably large timeout value,
2253 // to drain the marking data structures of the remaining entries
2254 // added by the 'keep alive' oop closure above.
2255
2256 class G1CMDrainMarkingStackClosure: public VoidClosure {
2257 ConcurrentMark* _cm;
2258 CMTask* _task;
2259 bool _is_serial;
2260 public:
2261 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2262 _cm(cm), _task(task), _is_serial(is_serial) {
2263 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2264 }
2265
2266 void do_void() {
2267 do {
2268 if (_cm->verbose_high()) {
2269 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2270 _task->worker_id(), BOOL_TO_STR(_is_serial));
2271 }
2272
2273 // We call CMTask::do_marking_step() to completely drain the local
2274 // and global marking stacks of entries pushed by the 'keep alive'
2275 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2276 //
2277 // CMTask::do_marking_step() is called in a loop, which we'll exit
2278 // if there's nothing more to do (i.e. we've completely drained the
2279 // entries that were pushed as a a result of applying the 'keep alive'
2280 // closure to the entries on the discovered ref lists) or we overflow
2281 // the global marking stack.
2282 //
2283 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2284 // flag while there may still be some work to do. (See the comment at
2285 // the beginning of CMTask::do_marking_step() for those conditions -
2286 // one of which is reaching the specified time target.) It is only
2287 // when CMTask::do_marking_step() returns without setting the
2288 // has_aborted() flag that the marking step has completed.
2289
2290 _task->do_marking_step(1000000000.0 /* something very large */,
2291 true /* do_termination */,
2292 _is_serial);
2293 } while (_task->has_aborted() && !_cm->has_overflown());
2294 }
2295 };
2296
2297 // Implementation of AbstractRefProcTaskExecutor for parallel
2298 // reference processing at the end of G1 concurrent marking
2299
2300 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2301 private:
2302 G1CollectedHeap* _g1h;
2303 ConcurrentMark* _cm;
2304 WorkGang* _workers;
2305 int _active_workers;
2306
2307 public:
2308 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2309 ConcurrentMark* cm,
2310 WorkGang* workers,
2311 int n_workers) :
2312 _g1h(g1h), _cm(cm),
2313 _workers(workers), _active_workers(n_workers) { }
2314
2315 // Executes the given task using concurrent marking worker threads.
2316 virtual void execute(ProcessTask& task);
2317 virtual void execute(EnqueueTask& task);
2318 };
2319
2320 class G1CMRefProcTaskProxy: public AbstractGangTask {
2321 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2322 ProcessTask& _proc_task;
2323 G1CollectedHeap* _g1h;
2324 ConcurrentMark* _cm;
2325
2326 public:
2327 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2328 G1CollectedHeap* g1h,
2329 ConcurrentMark* cm) :
2330 AbstractGangTask("Process reference objects in parallel"),
2331 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2332 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2333 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2334 }
2335
2336 virtual void work(uint worker_id) {
2337 ResourceMark rm;
2338 HandleMark hm;
2339 CMTask* task = _cm->task(worker_id);
2340 G1CMIsAliveClosure g1_is_alive(_g1h);
2341 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2342 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2343
2344 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2345 }
2346 };
2347
2348 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2349 assert(_workers != NULL, "Need parallel worker threads.");
2350 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2351
2352 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2353
2354 // We need to reset the concurrency level before each
2355 // proxy task execution, so that the termination protocol
2356 // and overflow handling in CMTask::do_marking_step() knows
2357 // how many workers to wait for.
2358 _cm->set_concurrency(_active_workers);
2359 _g1h->set_par_threads(_active_workers);
2360 _workers->run_task(&proc_task_proxy);
2361 _g1h->set_par_threads(0);
2362 }
2363
2364 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2365 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2366 EnqueueTask& _enq_task;
2367
2368 public:
2369 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2370 AbstractGangTask("Enqueue reference objects in parallel"),
2371 _enq_task(enq_task) { }
2372
2373 virtual void work(uint worker_id) {
2374 _enq_task.work(worker_id);
2375 }
2376 };
2377
2378 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2379 assert(_workers != NULL, "Need parallel worker threads.");
2380 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2381
2382 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2383
2384 // Not strictly necessary but...
2385 //
2386 // We need to reset the concurrency level before each
2387 // proxy task execution, so that the termination protocol
2388 // and overflow handling in CMTask::do_marking_step() knows
2389 // how many workers to wait for.
2390 _cm->set_concurrency(_active_workers);
2391 _g1h->set_par_threads(_active_workers);
2392 _workers->run_task(&enq_task_proxy);
2393 _g1h->set_par_threads(0);
2394 }
2395
2396 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2397 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2398 }
2399
2400 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2401 if (has_overflown()) {
2402 // Skip processing the discovered references if we have
2403 // overflown the global marking stack. Reference objects
2404 // only get discovered once so it is OK to not
2405 // de-populate the discovered reference lists. We could have,
2406 // but the only benefit would be that, when marking restarts,
2407 // less reference objects are discovered.
2408 return;
2409 }
2410
2411 ResourceMark rm;
2412 HandleMark hm;
2413
2414 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2415
2416 // Is alive closure.
2417 G1CMIsAliveClosure g1_is_alive(g1h);
2418
2419 // Inner scope to exclude the cleaning of the string and symbol
2420 // tables from the displayed time.
2421 {
2422 G1CMTraceTime t("GC ref-proc", G1Log::finer());
2423
2424 ReferenceProcessor* rp = g1h->ref_processor_cm();
2425
2426 // See the comment in G1CollectedHeap::ref_processing_init()
2427 // about how reference processing currently works in G1.
2428
2429 // Set the soft reference policy
2430 rp->setup_policy(clear_all_soft_refs);
2431 assert(_markStack.isEmpty(), "mark stack should be empty");
2432
2433 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2434 // in serial reference processing. Note these closures are also
2435 // used for serially processing (by the the current thread) the
2436 // JNI references during parallel reference processing.
2437 //
2438 // These closures do not need to synchronize with the worker
2439 // threads involved in parallel reference processing as these
2440 // instances are executed serially by the current thread (e.g.
2441 // reference processing is not multi-threaded and is thus
2442 // performed by the current thread instead of a gang worker).
2443 //
2444 // The gang tasks involved in parallel reference processing create
2445 // their own instances of these closures, which do their own
2446 // synchronization among themselves.
2447 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2448 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2449
2450 // We need at least one active thread. If reference processing
2451 // is not multi-threaded we use the current (VMThread) thread,
2452 // otherwise we use the work gang from the G1CollectedHeap and
2453 // we utilize all the worker threads we can.
2454 bool processing_is_mt = rp->processing_is_mt();
2455 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2456 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2457
2458 // Parallel processing task executor.
2459 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2460 g1h->workers(), active_workers);
2461 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2462
2463 // Set the concurrency level. The phase was already set prior to
2464 // executing the remark task.
2465 set_concurrency(active_workers);
2466
2467 // Set the degree of MT processing here. If the discovery was done MT,
2468 // the number of threads involved during discovery could differ from
2469 // the number of active workers. This is OK as long as the discovered
2470 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2471 rp->set_active_mt_degree(active_workers);
2472
2473 // Process the weak references.
2474 const ReferenceProcessorStats& stats =
2475 rp->process_discovered_references(&g1_is_alive,
2476 &g1_keep_alive,
2477 &g1_drain_mark_stack,
2478 executor,
2479 g1h->gc_timer_cm(),
2480 concurrent_gc_id());
2481 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2482
2483 // The do_oop work routines of the keep_alive and drain_marking_stack
2484 // oop closures will set the has_overflown flag if we overflow the
2485 // global marking stack.
2486
2487 assert(_markStack.overflow() || _markStack.isEmpty(),
2488 "mark stack should be empty (unless it overflowed)");
2489
2490 if (_markStack.overflow()) {
2491 // This should have been done already when we tried to push an
2492 // entry on to the global mark stack. But let's do it again.
2493 set_has_overflown();
2494 }
2495
2496 assert(rp->num_q() == active_workers, "why not");
2497
2498 rp->enqueue_discovered_references(executor);
2499
2500 rp->verify_no_references_recorded();
2501 assert(!rp->discovery_enabled(), "Post condition");
2502 }
2503
2504 if (has_overflown()) {
2505 // We can not trust g1_is_alive if the marking stack overflowed
2506 return;
2507 }
2508
2509 assert(_markStack.isEmpty(), "Marking should have completed");
2510
2511 // Unload Klasses, String, Symbols, Code Cache, etc.
2512 {
2513 G1CMTraceTime trace("Unloading", G1Log::finer());
2514
2515 if (ClassUnloadingWithConcurrentMark) {
2516 bool purged_classes;
2517
2518 {
2519 G1CMTraceTime trace("System Dictionary Unloading", G1Log::finest());
2520 purged_classes = SystemDictionary::do_unloading(&g1_is_alive);
2521 }
2522
2523 {
2524 G1CMTraceTime trace("Parallel Unloading", G1Log::finest());
2525 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2526 }
2527 }
2528
2529 if (G1StringDedup::is_enabled()) {
2530 G1CMTraceTime trace("String Deduplication Unlink", G1Log::finest());
2531 G1StringDedup::unlink(&g1_is_alive);
2532 }
2533 }
2534 }
2535
2536 void ConcurrentMark::swapMarkBitMaps() {
2537 CMBitMapRO* temp = _prevMarkBitMap;
2538 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2539 _nextMarkBitMap = (CMBitMap*) temp;
2540 }
2541
2542 class CMObjectClosure;
2543
2544 // Closure for iterating over objects, currently only used for
2545 // processing SATB buffers.
2546 class CMObjectClosure : public ObjectClosure {
2547 private:
2548 CMTask* _task;
2549
2550 public:
2551 void do_object(oop obj) {
2552 _task->deal_with_reference(obj);
2553 }
2554
2555 CMObjectClosure(CMTask* task) : _task(task) { }
2556 };
2557
2558 class G1RemarkThreadsClosure : public ThreadClosure {
2559 CMObjectClosure _cm_obj;
2560 G1CMOopClosure _cm_cl;
2561 MarkingCodeBlobClosure _code_cl;
2562 int _thread_parity;
2563
2564 public:
2565 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task) :
2566 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2567 _thread_parity(SharedHeap::heap()->strong_roots_parity()) {}
2568
2569 void do_thread(Thread* thread) {
2570 if (thread->is_Java_thread()) {
2571 if (thread->claim_oops_do(true, _thread_parity)) {
2572 JavaThread* jt = (JavaThread*)thread;
2573
2574 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2575 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2576 // * Alive if on the stack of an executing method
2577 // * Weakly reachable otherwise
2578 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2579 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2580 jt->nmethods_do(&_code_cl);
2581
2582 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2583 }
2584 } else if (thread->is_VM_thread()) {
2585 if (thread->claim_oops_do(true, _thread_parity)) {
2586 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2587 }
2588 }
2589 }
2590 };
2591
2592 class CMRemarkTask: public AbstractGangTask {
2593 private:
2594 ConcurrentMark* _cm;
2595 public:
2596 void work(uint worker_id) {
2597 // Since all available tasks are actually started, we should
2598 // only proceed if we're supposed to be active.
2599 if (worker_id < _cm->active_tasks()) {
2600 CMTask* task = _cm->task(worker_id);
2601 task->record_start_time();
2602 {
2603 ResourceMark rm;
2604 HandleMark hm;
2605
2606 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
2607 Threads::threads_do(&threads_f);
2608 }
2609
2610 do {
2611 task->do_marking_step(1000000000.0 /* something very large */,
2612 true /* do_termination */,
2613 false /* is_serial */);
2614 } while (task->has_aborted() && !_cm->has_overflown());
2615 // If we overflow, then we do not want to restart. We instead
2616 // want to abort remark and do concurrent marking again.
2617 task->record_end_time();
2618 }
2619 }
2620
2621 CMRemarkTask(ConcurrentMark* cm, int active_workers) :
2622 AbstractGangTask("Par Remark"), _cm(cm) {
2623 _cm->terminator()->reset_for_reuse(active_workers);
2624 }
2625 };
2626
2627 void ConcurrentMark::checkpointRootsFinalWork() {
2628 ResourceMark rm;
2629 HandleMark hm;
2630 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2631
2632 G1CMTraceTime trace("Finalize Marking", G1Log::finer());
2633
2634 g1h->ensure_parsability(false);
2635
2636 G1CollectedHeap::StrongRootsScope srs(g1h);
2637 // this is remark, so we'll use up all active threads
2638 uint active_workers = g1h->workers()->active_workers();
2639 if (active_workers == 0) {
2640 assert(active_workers > 0, "Should have been set earlier");
2641 active_workers = (uint) ParallelGCThreads;
2642 g1h->workers()->set_active_workers(active_workers);
2643 }
2644 set_concurrency_and_phase(active_workers, false /* concurrent */);
2645 // Leave _parallel_marking_threads at it's
2646 // value originally calculated in the ConcurrentMark
2647 // constructor and pass values of the active workers
2648 // through the gang in the task.
2649
2650 CMRemarkTask remarkTask(this, active_workers);
2651 // We will start all available threads, even if we decide that the
2652 // active_workers will be fewer. The extra ones will just bail out
2653 // immediately.
2654 g1h->set_par_threads(active_workers);
2655 g1h->workers()->run_task(&remarkTask);
2656 g1h->set_par_threads(0);
2657
2658 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2659 guarantee(has_overflown() ||
2660 satb_mq_set.completed_buffers_num() == 0,
2661 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2662 BOOL_TO_STR(has_overflown()),
2663 satb_mq_set.completed_buffers_num()));
2664
2665 print_stats();
2666 }
2667
2668 #ifndef PRODUCT
2669
2670 class PrintReachableOopClosure: public OopClosure {
2671 private:
2672 G1CollectedHeap* _g1h;
2673 outputStream* _out;
2674 VerifyOption _vo;
2675 bool _all;
2676
2677 public:
2678 PrintReachableOopClosure(outputStream* out,
2679 VerifyOption vo,
2680 bool all) :
2681 _g1h(G1CollectedHeap::heap()),
2682 _out(out), _vo(vo), _all(all) { }
2683
2684 void do_oop(narrowOop* p) { do_oop_work(p); }
2685 void do_oop( oop* p) { do_oop_work(p); }
2686
2687 template <class T> void do_oop_work(T* p) {
2688 oop obj = oopDesc::load_decode_heap_oop(p);
2689 const char* str = NULL;
2690 const char* str2 = "";
2691
2692 if (obj == NULL) {
2693 str = "";
2694 } else if (!_g1h->is_in_g1_reserved(obj)) {
2695 str = " O";
2696 } else {
2697 HeapRegion* hr = _g1h->heap_region_containing(obj);
2698 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2699 bool marked = _g1h->is_marked(obj, _vo);
2700
2701 if (over_tams) {
2702 str = " >";
2703 if (marked) {
2704 str2 = " AND MARKED";
2705 }
2706 } else if (marked) {
2707 str = " M";
2708 } else {
2709 str = " NOT";
2710 }
2711 }
2712
2713 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2714 p2i(p), p2i((void*) obj), str, str2);
2715 }
2716 };
2717
2718 class PrintReachableObjectClosure : public ObjectClosure {
2719 private:
2720 G1CollectedHeap* _g1h;
2721 outputStream* _out;
2722 VerifyOption _vo;
2723 bool _all;
2724 HeapRegion* _hr;
2725
2726 public:
2727 PrintReachableObjectClosure(outputStream* out,
2728 VerifyOption vo,
2729 bool all,
2730 HeapRegion* hr) :
2731 _g1h(G1CollectedHeap::heap()),
2732 _out(out), _vo(vo), _all(all), _hr(hr) { }
2733
2734 void do_object(oop o) {
2735 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2736 bool marked = _g1h->is_marked(o, _vo);
2737 bool print_it = _all || over_tams || marked;
2738
2739 if (print_it) {
2740 _out->print_cr(" "PTR_FORMAT"%s",
2741 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2742 PrintReachableOopClosure oopCl(_out, _vo, _all);
2743 o->oop_iterate_no_header(&oopCl);
2744 }
2745 }
2746 };
2747
2748 class PrintReachableRegionClosure : public HeapRegionClosure {
2749 private:
2750 G1CollectedHeap* _g1h;
2751 outputStream* _out;
2752 VerifyOption _vo;
2753 bool _all;
2754
2755 public:
2756 bool doHeapRegion(HeapRegion* hr) {
2757 HeapWord* b = hr->bottom();
2758 HeapWord* e = hr->end();
2759 HeapWord* t = hr->top();
2760 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2761 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2762 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2763 _out->cr();
2764
2765 HeapWord* from = b;
2766 HeapWord* to = t;
2767
2768 if (to > from) {
2769 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2770 _out->cr();
2771 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2772 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2773 _out->cr();
2774 }
2775
2776 return false;
2777 }
2778
2779 PrintReachableRegionClosure(outputStream* out,
2780 VerifyOption vo,
2781 bool all) :
2782 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2783 };
2784
2785 void ConcurrentMark::print_reachable(const char* str,
2786 VerifyOption vo,
2787 bool all) {
2788 gclog_or_tty->cr();
2789 gclog_or_tty->print_cr("== Doing heap dump... ");
2790
2791 if (G1PrintReachableBaseFile == NULL) {
2792 gclog_or_tty->print_cr(" #### error: no base file defined");
2793 return;
2794 }
2795
2796 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2797 (JVM_MAXPATHLEN - 1)) {
2798 gclog_or_tty->print_cr(" #### error: file name too long");
2799 return;
2800 }
2801
2802 char file_name[JVM_MAXPATHLEN];
2803 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2804 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2805
2806 fileStream fout(file_name);
2807 if (!fout.is_open()) {
2808 gclog_or_tty->print_cr(" #### error: could not open file");
2809 return;
2810 }
2811
2812 outputStream* out = &fout;
2813 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2814 out->cr();
2815
2816 out->print_cr("--- ITERATING OVER REGIONS");
2817 out->cr();
2818 PrintReachableRegionClosure rcl(out, vo, all);
2819 _g1h->heap_region_iterate(&rcl);
2820 out->cr();
2821
2822 gclog_or_tty->print_cr(" done");
2823 gclog_or_tty->flush();
2824 }
2825
2826 #endif // PRODUCT
2827
2828 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2829 // Note we are overriding the read-only view of the prev map here, via
2830 // the cast.
2831 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2832 }
2833
2834 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2835 _nextMarkBitMap->clearRange(mr);
2836 }
2837
2838 HeapRegion*
2839 ConcurrentMark::claim_region(uint worker_id) {
2840 // "checkpoint" the finger
2841 HeapWord* finger = _finger;
2842
2843 // _heap_end will not change underneath our feet; it only changes at
2844 // yield points.
2845 while (finger < _heap_end) {
2846 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2847
2848 // Note on how this code handles humongous regions. In the
2849 // normal case the finger will reach the start of a "starts
2850 // humongous" (SH) region. Its end will either be the end of the
2851 // last "continues humongous" (CH) region in the sequence, or the
2852 // standard end of the SH region (if the SH is the only region in
2853 // the sequence). That way claim_region() will skip over the CH
2854 // regions. However, there is a subtle race between a CM thread
2855 // executing this method and a mutator thread doing a humongous
2856 // object allocation. The two are not mutually exclusive as the CM
2857 // thread does not need to hold the Heap_lock when it gets
2858 // here. So there is a chance that claim_region() will come across
2859 // a free region that's in the progress of becoming a SH or a CH
2860 // region. In the former case, it will either
2861 // a) Miss the update to the region's end, in which case it will
2862 // visit every subsequent CH region, will find their bitmaps
2863 // empty, and do nothing, or
2864 // b) Will observe the update of the region's end (in which case
2865 // it will skip the subsequent CH regions).
2866 // If it comes across a region that suddenly becomes CH, the
2867 // scenario will be similar to b). So, the race between
2868 // claim_region() and a humongous object allocation might force us
2869 // to do a bit of unnecessary work (due to some unnecessary bitmap
2870 // iterations) but it should not introduce and correctness issues.
2871 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2872
2873 // Above heap_region_containing_raw may return NULL as we always scan claim
2874 // until the end of the heap. In this case, just jump to the next region.
2875 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2876
2877 // Is the gap between reading the finger and doing the CAS too long?
2878 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2879 if (res == finger && curr_region != NULL) {
2880 // we succeeded
2881 HeapWord* bottom = curr_region->bottom();
2882 HeapWord* limit = curr_region->next_top_at_mark_start();
2883
2884 if (verbose_low()) {
2885 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2886 "["PTR_FORMAT", "PTR_FORMAT"), "
2887 "limit = "PTR_FORMAT,
2888 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2889 }
2890
2891 // notice that _finger == end cannot be guaranteed here since,
2892 // someone else might have moved the finger even further
2893 assert(_finger >= end, "the finger should have moved forward");
2894
2895 if (verbose_low()) {
2896 gclog_or_tty->print_cr("[%u] we were successful with region = "
2897 PTR_FORMAT, worker_id, p2i(curr_region));
2898 }
2899
2900 if (limit > bottom) {
2901 if (verbose_low()) {
2902 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2903 "returning it ", worker_id, p2i(curr_region));
2904 }
2905 return curr_region;
2906 } else {
2907 assert(limit == bottom,
2908 "the region limit should be at bottom");
2909 if (verbose_low()) {
2910 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
2911 "returning NULL", worker_id, p2i(curr_region));
2912 }
2913 // we return NULL and the caller should try calling
2914 // claim_region() again.
2915 return NULL;
2916 }
2917 } else {
2918 assert(_finger > finger, "the finger should have moved forward");
2919 if (verbose_low()) {
2920 if (curr_region == NULL) {
2921 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
2922 "global finger = "PTR_FORMAT", "
2923 "our finger = "PTR_FORMAT,
2924 worker_id, p2i(_finger), p2i(finger));
2925 } else {
2926 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2927 "global finger = "PTR_FORMAT", "
2928 "our finger = "PTR_FORMAT,
2929 worker_id, p2i(_finger), p2i(finger));
2930 }
2931 }
2932
2933 // read it again
2934 finger = _finger;
2935 }
2936 }
2937
2938 return NULL;
2939 }
2940
2941 #ifndef PRODUCT
2942 enum VerifyNoCSetOopsPhase {
2943 VerifyNoCSetOopsStack,
2944 VerifyNoCSetOopsQueues,
2945 VerifyNoCSetOopsSATBCompleted,
2946 VerifyNoCSetOopsSATBThread
2947 };
2948
2949 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
2950 private:
2951 G1CollectedHeap* _g1h;
2952 VerifyNoCSetOopsPhase _phase;
2953 int _info;
2954
2955 const char* phase_str() {
2956 switch (_phase) {
2957 case VerifyNoCSetOopsStack: return "Stack";
2958 case VerifyNoCSetOopsQueues: return "Queue";
2959 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2960 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
2961 default: ShouldNotReachHere();
2962 }
2963 return NULL;
2964 }
2965
2966 void do_object_work(oop obj) {
2967 guarantee(!_g1h->obj_in_cs(obj),
2968 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2969 p2i((void*) obj), phase_str(), _info));
2970 }
2971
2972 public:
2973 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2974
2975 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2976 _phase = phase;
2977 _info = info;
2978 }
2979
2980 virtual void do_oop(oop* p) {
2981 oop obj = oopDesc::load_decode_heap_oop(p);
2982 do_object_work(obj);
2983 }
2984
2985 virtual void do_oop(narrowOop* p) {
2986 // We should not come across narrow oops while scanning marking
2987 // stacks and SATB buffers.
2988 ShouldNotReachHere();
2989 }
2990
2991 virtual void do_object(oop obj) {
2992 do_object_work(obj);
2993 }
2994 };
2995
2996 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
2997 bool verify_enqueued_buffers,
2998 bool verify_thread_buffers,
2999 bool verify_fingers) {
3000 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3001 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3002 return;
3003 }
3004
3005 VerifyNoCSetOopsClosure cl;
3006
3007 if (verify_stacks) {
3008 // Verify entries on the global mark stack
3009 cl.set_phase(VerifyNoCSetOopsStack);
3010 _markStack.oops_do(&cl);
3011
3012 // Verify entries on the task queues
3013 for (uint i = 0; i < _max_worker_id; i += 1) {
3014 cl.set_phase(VerifyNoCSetOopsQueues, i);
3015 CMTaskQueue* queue = _task_queues->queue(i);
3016 queue->oops_do(&cl);
3017 }
3018 }
3019
3020 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3021
3022 // Verify entries on the enqueued SATB buffers
3023 if (verify_enqueued_buffers) {
3024 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3025 satb_qs.iterate_completed_buffers_read_only(&cl);
3026 }
3027
3028 // Verify entries on the per-thread SATB buffers
3029 if (verify_thread_buffers) {
3030 cl.set_phase(VerifyNoCSetOopsSATBThread);
3031 satb_qs.iterate_thread_buffers_read_only(&cl);
3032 }
3033
3034 if (verify_fingers) {
3035 // Verify the global finger
3036 HeapWord* global_finger = finger();
3037 if (global_finger != NULL && global_finger < _heap_end) {
3038 // The global finger always points to a heap region boundary. We
3039 // use heap_region_containing_raw() to get the containing region
3040 // given that the global finger could be pointing to a free region
3041 // which subsequently becomes continues humongous. If that
3042 // happens, heap_region_containing() will return the bottom of the
3043 // corresponding starts humongous region and the check below will
3044 // not hold any more.
3045 // Since we always iterate over all regions, we might get a NULL HeapRegion
3046 // here.
3047 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3048 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3049 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3050 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3051 }
3052
3053 // Verify the task fingers
3054 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3055 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3056 CMTask* task = _tasks[i];
3057 HeapWord* task_finger = task->finger();
3058 if (task_finger != NULL && task_finger < _heap_end) {
3059 // See above note on the global finger verification.
3060 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3061 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3062 !task_hr->in_collection_set(),
3063 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3064 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3065 }
3066 }
3067 }
3068 }
3069 #endif // PRODUCT
3070
3071 // Aggregate the counting data that was constructed concurrently
3072 // with marking.
3073 class AggregateCountDataHRClosure: public HeapRegionClosure {
3074 G1CollectedHeap* _g1h;
3075 ConcurrentMark* _cm;
3076 CardTableModRefBS* _ct_bs;
3077 BitMap* _cm_card_bm;
3078 uint _max_worker_id;
3079
3080 public:
3081 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3082 BitMap* cm_card_bm,
3083 uint max_worker_id) :
3084 _g1h(g1h), _cm(g1h->concurrent_mark()),
3085 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3086 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3087
3088 bool doHeapRegion(HeapRegion* hr) {
3089 if (hr->is_continues_humongous()) {
3090 // We will ignore these here and process them when their
3091 // associated "starts humongous" region is processed.
3092 // Note that we cannot rely on their associated
3093 // "starts humongous" region to have their bit set to 1
3094 // since, due to the region chunking in the parallel region
3095 // iteration, a "continues humongous" region might be visited
3096 // before its associated "starts humongous".
3097 return false;
3098 }
3099
3100 HeapWord* start = hr->bottom();
3101 HeapWord* limit = hr->next_top_at_mark_start();
3102 HeapWord* end = hr->end();
3103
3104 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3105 err_msg("Preconditions not met - "
3106 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3107 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3108 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3109
3110 assert(hr->next_marked_bytes() == 0, "Precondition");
3111
3112 if (start == limit) {
3113 // NTAMS of this region has not been set so nothing to do.
3114 return false;
3115 }
3116
3117 // 'start' should be in the heap.
3118 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3119 // 'end' *may* be just beyond the end of the heap (if hr is the last region)
3120 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3121
3122 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3123 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3124 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3125
3126 // If ntams is not card aligned then we bump card bitmap index
3127 // for limit so that we get the all the cards spanned by
3128 // the object ending at ntams.
3129 // Note: if this is the last region in the heap then ntams
3130 // could be actually just beyond the end of the the heap;
3131 // limit_idx will then correspond to a (non-existent) card
3132 // that is also outside the heap.
3133 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3134 limit_idx += 1;
3135 }
3136
3137 assert(limit_idx <= end_idx, "or else use atomics");
3138
3139 // Aggregate the "stripe" in the count data associated with hr.
3140 uint hrm_index = hr->hrm_index();
3141 size_t marked_bytes = 0;
3142
3143 for (uint i = 0; i < _max_worker_id; i += 1) {
3144 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3145 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3146
3147 // Fetch the marked_bytes in this region for task i and
3148 // add it to the running total for this region.
3149 marked_bytes += marked_bytes_array[hrm_index];
3150
3151 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3152 // into the global card bitmap.
3153 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3154
3155 while (scan_idx < limit_idx) {
3156 assert(task_card_bm->at(scan_idx) == true, "should be");
3157 _cm_card_bm->set_bit(scan_idx);
3158 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3159
3160 // BitMap::get_next_one_offset() can handle the case when
3161 // its left_offset parameter is greater than its right_offset
3162 // parameter. It does, however, have an early exit if
3163 // left_offset == right_offset. So let's limit the value
3164 // passed in for left offset here.
3165 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3166 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3167 }
3168 }
3169
3170 // Update the marked bytes for this region.
3171 hr->add_to_marked_bytes(marked_bytes);
3172
3173 // Next heap region
3174 return false;
3175 }
3176 };
3177
3178 class G1AggregateCountDataTask: public AbstractGangTask {
3179 protected:
3180 G1CollectedHeap* _g1h;
3181 ConcurrentMark* _cm;
3182 BitMap* _cm_card_bm;
3183 uint _max_worker_id;
3184 int _active_workers;
3185 HeapRegionClaimer _hrclaimer;
3186
3187 public:
3188 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3189 ConcurrentMark* cm,
3190 BitMap* cm_card_bm,
3191 uint max_worker_id,
3192 int n_workers) :
3193 AbstractGangTask("Count Aggregation"),
3194 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3195 _max_worker_id(max_worker_id),
3196 _active_workers(n_workers),
3197 _hrclaimer(_active_workers) {
3198 }
3199
3200 void work(uint worker_id) {
3201 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3202
3203 _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer);
3204 }
3205 };
3206
3207
3208 void ConcurrentMark::aggregate_count_data() {
3209 int n_workers = _g1h->workers()->active_workers();
3210
3211 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3212 _max_worker_id, n_workers);
3213
3214 _g1h->set_par_threads(n_workers);
3215 _g1h->workers()->run_task(&g1_par_agg_task);
3216 _g1h->set_par_threads(0);
3217 _g1h->allocation_context_stats().update_at_remark();
3218 }
3219
3220 // Clear the per-worker arrays used to store the per-region counting data
3221 void ConcurrentMark::clear_all_count_data() {
3222 // Clear the global card bitmap - it will be filled during
3223 // liveness count aggregation (during remark) and the
3224 // final counting task.
3225 _card_bm.clear();
3226
3227 // Clear the global region bitmap - it will be filled as part
3228 // of the final counting task.
3229 _region_bm.clear();
3230
3231 uint max_regions = _g1h->max_regions();
3232 assert(_max_worker_id > 0, "uninitialized");
3233
3234 for (uint i = 0; i < _max_worker_id; i += 1) {
3235 BitMap* task_card_bm = count_card_bitmap_for(i);
3236 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3237
3238 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3239 assert(marked_bytes_array != NULL, "uninitialized");
3240
3241 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3242 task_card_bm->clear();
3243 }
3244 }
3245
3246 void ConcurrentMark::print_stats() {
3247 if (verbose_stats()) {
3248 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3249 for (size_t i = 0; i < _active_tasks; ++i) {
3250 _tasks[i]->print_stats();
3251 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3252 }
3253 }
3254 }
3255
3256 // abandon current marking iteration due to a Full GC
3257 void ConcurrentMark::abort() {
3258 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3259 // concurrent bitmap clearing.
3260 _nextMarkBitMap->clearAll();
3261
3262 // Note we cannot clear the previous marking bitmap here
3263 // since VerifyDuringGC verifies the objects marked during
3264 // a full GC against the previous bitmap.
3265
3266 // Clear the liveness counting data
3267 clear_all_count_data();
3268 // Empty mark stack
3269 reset_marking_state();
3270 for (uint i = 0; i < _max_worker_id; ++i) {
3271 _tasks[i]->clear_region_fields();
3272 }
3273 _first_overflow_barrier_sync.abort();
3274 _second_overflow_barrier_sync.abort();
3275 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3276 if (!gc_id.is_undefined()) {
3277 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3278 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3279 _aborted_gc_id = gc_id;
3280 }
3281 _has_aborted = true;
3282
3283 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3284 satb_mq_set.abandon_partial_marking();
3285 // This can be called either during or outside marking, we'll read
3286 // the expected_active value from the SATB queue set.
3287 satb_mq_set.set_active_all_threads(
3288 false, /* new active value */
3289 satb_mq_set.is_active() /* expected_active */);
3290
3291 _g1h->trace_heap_after_concurrent_cycle();
3292 _g1h->register_concurrent_cycle_end();
3293 }
3294
3295 const GCId& ConcurrentMark::concurrent_gc_id() {
3296 if (has_aborted()) {
3297 return _aborted_gc_id;
3298 }
3299 return _g1h->gc_tracer_cm()->gc_id();
3300 }
3301
3302 static void print_ms_time_info(const char* prefix, const char* name,
3303 NumberSeq& ns) {
3304 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3305 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3306 if (ns.num() > 0) {
3307 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3308 prefix, ns.sd(), ns.maximum());
3309 }
3310 }
3311
3312 void ConcurrentMark::print_summary_info() {
3313 gclog_or_tty->print_cr(" Concurrent marking:");
3314 print_ms_time_info(" ", "init marks", _init_times);
3315 print_ms_time_info(" ", "remarks", _remark_times);
3316 {
3317 print_ms_time_info(" ", "final marks", _remark_mark_times);
3318 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3319
3320 }
3321 print_ms_time_info(" ", "cleanups", _cleanup_times);
3322 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3323 _total_counting_time,
3324 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3325 (double)_cleanup_times.num()
3326 : 0.0));
3327 if (G1ScrubRemSets) {
3328 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3329 _total_rs_scrub_time,
3330 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3331 (double)_cleanup_times.num()
3332 : 0.0));
3333 }
3334 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3335 (_init_times.sum() + _remark_times.sum() +
3336 _cleanup_times.sum())/1000.0);
3337 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3338 "(%8.2f s marking).",
3339 cmThread()->vtime_accum(),
3340 cmThread()->vtime_mark_accum());
3341 }
3342
3343 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3344 if (use_parallel_marking_threads()) {
3345 _parallel_workers->print_worker_threads_on(st);
3346 }
3347 }
3348
3349 void ConcurrentMark::print_on_error(outputStream* st) const {
3350 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3351 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3352 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3353 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3354 }
3355
3356 // We take a break if someone is trying to stop the world.
3357 bool ConcurrentMark::do_yield_check(uint worker_id) {
3358 if (SuspendibleThreadSet::should_yield()) {
3359 if (worker_id == 0) {
3360 _g1h->g1_policy()->record_concurrent_pause();
3361 }
3362 SuspendibleThreadSet::yield();
3363 return true;
3364 } else {
3365 return false;
3366 }
3367 }
3368
3369 #ifndef PRODUCT
3370 // for debugging purposes
3371 void ConcurrentMark::print_finger() {
3372 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3373 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3374 for (uint i = 0; i < _max_worker_id; ++i) {
3375 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3376 }
3377 gclog_or_tty->cr();
3378 }
3379 #endif
3380
3381 void CMTask::scan_object(oop obj) {
3382 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3383
3384 if (_cm->verbose_high()) {
3385 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3386 _worker_id, p2i((void*) obj));
3387 }
3388
3389 size_t obj_size = obj->size();
3390 _words_scanned += obj_size;
3391
3392 obj->oop_iterate(_cm_oop_closure);
3393 statsOnly( ++_objs_scanned );
3394 check_limits();
3395 }
3396
3397 // Closure for iteration over bitmaps
3398 class CMBitMapClosure : public BitMapClosure {
3399 private:
3400 // the bitmap that is being iterated over
3401 CMBitMap* _nextMarkBitMap;
3402 ConcurrentMark* _cm;
3403 CMTask* _task;
3404
3405 public:
3406 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3407 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3408
3409 bool do_bit(size_t offset) {
3410 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3411 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3412 assert( addr < _cm->finger(), "invariant");
3413
3414 statsOnly( _task->increase_objs_found_on_bitmap() );
3415 assert(addr >= _task->finger(), "invariant");
3416
3417 // We move that task's local finger along.
3418 _task->move_finger_to(addr);
3419
3420 _task->scan_object(oop(addr));
3421 // we only partially drain the local queue and global stack
3422 _task->drain_local_queue(true);
3423 _task->drain_global_stack(true);
3424
3425 // if the has_aborted flag has been raised, we need to bail out of
3426 // the iteration
3427 return !_task->has_aborted();
3428 }
3429 };
3430
3431 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3432 ConcurrentMark* cm,
3433 CMTask* task)
3434 : _g1h(g1h), _cm(cm), _task(task) {
3435 assert(_ref_processor == NULL, "should be initialized to NULL");
3436
3437 if (G1UseConcMarkReferenceProcessing) {
3438 _ref_processor = g1h->ref_processor_cm();
3439 assert(_ref_processor != NULL, "should not be NULL");
3440 }
3441 }
3442
3443 void CMTask::setup_for_region(HeapRegion* hr) {
3444 assert(hr != NULL,
3445 "claim_region() should have filtered out NULL regions");
3446 assert(!hr->is_continues_humongous(),
3447 "claim_region() should have filtered out continues humongous regions");
3448
3449 if (_cm->verbose_low()) {
3450 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3451 _worker_id, p2i(hr));
3452 }
3453
3454 _curr_region = hr;
3455 _finger = hr->bottom();
3456 update_region_limit();
3457 }
3458
3459 void CMTask::update_region_limit() {
3460 HeapRegion* hr = _curr_region;
3461 HeapWord* bottom = hr->bottom();
3462 HeapWord* limit = hr->next_top_at_mark_start();
3463
3464 if (limit == bottom) {
3465 if (_cm->verbose_low()) {
3466 gclog_or_tty->print_cr("[%u] found an empty region "
3467 "["PTR_FORMAT", "PTR_FORMAT")",
3468 _worker_id, p2i(bottom), p2i(limit));
3469 }
3470 // The region was collected underneath our feet.
3471 // We set the finger to bottom to ensure that the bitmap
3472 // iteration that will follow this will not do anything.
3473 // (this is not a condition that holds when we set the region up,
3474 // as the region is not supposed to be empty in the first place)
3475 _finger = bottom;
3476 } else if (limit >= _region_limit) {
3477 assert(limit >= _finger, "peace of mind");
3478 } else {
3479 assert(limit < _region_limit, "only way to get here");
3480 // This can happen under some pretty unusual circumstances. An
3481 // evacuation pause empties the region underneath our feet (NTAMS
3482 // at bottom). We then do some allocation in the region (NTAMS
3483 // stays at bottom), followed by the region being used as a GC
3484 // alloc region (NTAMS will move to top() and the objects
3485 // originally below it will be grayed). All objects now marked in
3486 // the region are explicitly grayed, if below the global finger,
3487 // and we do not need in fact to scan anything else. So, we simply
3488 // set _finger to be limit to ensure that the bitmap iteration
3489 // doesn't do anything.
3490 _finger = limit;
3491 }
3492
3493 _region_limit = limit;
3494 }
3495
3496 void CMTask::giveup_current_region() {
3497 assert(_curr_region != NULL, "invariant");
3498 if (_cm->verbose_low()) {
3499 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3500 _worker_id, p2i(_curr_region));
3501 }
3502 clear_region_fields();
3503 }
3504
3505 void CMTask::clear_region_fields() {
3506 // Values for these three fields that indicate that we're not
3507 // holding on to a region.
3508 _curr_region = NULL;
3509 _finger = NULL;
3510 _region_limit = NULL;
3511 }
3512
3513 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3514 if (cm_oop_closure == NULL) {
3515 assert(_cm_oop_closure != NULL, "invariant");
3516 } else {
3517 assert(_cm_oop_closure == NULL, "invariant");
3518 }
3519 _cm_oop_closure = cm_oop_closure;
3520 }
3521
3522 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3523 guarantee(nextMarkBitMap != NULL, "invariant");
3524
3525 if (_cm->verbose_low()) {
3526 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3527 }
3528
3529 _nextMarkBitMap = nextMarkBitMap;
3530 clear_region_fields();
3531
3532 _calls = 0;
3533 _elapsed_time_ms = 0.0;
3534 _termination_time_ms = 0.0;
3535 _termination_start_time_ms = 0.0;
3536
3537 #if _MARKING_STATS_
3538 _local_pushes = 0;
3539 _local_pops = 0;
3540 _local_max_size = 0;
3541 _objs_scanned = 0;
3542 _global_pushes = 0;
3543 _global_pops = 0;
3544 _global_max_size = 0;
3545 _global_transfers_to = 0;
3546 _global_transfers_from = 0;
3547 _regions_claimed = 0;
3548 _objs_found_on_bitmap = 0;
3549 _satb_buffers_processed = 0;
3550 _steal_attempts = 0;
3551 _steals = 0;
3552 _aborted = 0;
3553 _aborted_overflow = 0;
3554 _aborted_cm_aborted = 0;
3555 _aborted_yield = 0;
3556 _aborted_timed_out = 0;
3557 _aborted_satb = 0;
3558 _aborted_termination = 0;
3559 #endif // _MARKING_STATS_
3560 }
3561
3562 bool CMTask::should_exit_termination() {
3563 regular_clock_call();
3564 // This is called when we are in the termination protocol. We should
3565 // quit if, for some reason, this task wants to abort or the global
3566 // stack is not empty (this means that we can get work from it).
3567 return !_cm->mark_stack_empty() || has_aborted();
3568 }
3569
3570 void CMTask::reached_limit() {
3571 assert(_words_scanned >= _words_scanned_limit ||
3572 _refs_reached >= _refs_reached_limit ,
3573 "shouldn't have been called otherwise");
3574 regular_clock_call();
3575 }
3576
3577 void CMTask::regular_clock_call() {
3578 if (has_aborted()) return;
3579
3580 // First, we need to recalculate the words scanned and refs reached
3581 // limits for the next clock call.
3582 recalculate_limits();
3583
3584 // During the regular clock call we do the following
3585
3586 // (1) If an overflow has been flagged, then we abort.
3587 if (_cm->has_overflown()) {
3588 set_has_aborted();
3589 return;
3590 }
3591
3592 // If we are not concurrent (i.e. we're doing remark) we don't need
3593 // to check anything else. The other steps are only needed during
3594 // the concurrent marking phase.
3595 if (!concurrent()) return;
3596
3597 // (2) If marking has been aborted for Full GC, then we also abort.
3598 if (_cm->has_aborted()) {
3599 set_has_aborted();
3600 statsOnly( ++_aborted_cm_aborted );
3601 return;
3602 }
3603
3604 double curr_time_ms = os::elapsedVTime() * 1000.0;
3605
3606 // (3) If marking stats are enabled, then we update the step history.
3607 #if _MARKING_STATS_
3608 if (_words_scanned >= _words_scanned_limit) {
3609 ++_clock_due_to_scanning;
3610 }
3611 if (_refs_reached >= _refs_reached_limit) {
3612 ++_clock_due_to_marking;
3613 }
3614
3615 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3616 _interval_start_time_ms = curr_time_ms;
3617 _all_clock_intervals_ms.add(last_interval_ms);
3618
3619 if (_cm->verbose_medium()) {
3620 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3621 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3622 _worker_id, last_interval_ms,
3623 _words_scanned,
3624 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3625 _refs_reached,
3626 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3627 }
3628 #endif // _MARKING_STATS_
3629
3630 // (4) We check whether we should yield. If we have to, then we abort.
3631 if (SuspendibleThreadSet::should_yield()) {
3632 // We should yield. To do this we abort the task. The caller is
3633 // responsible for yielding.
3634 set_has_aborted();
3635 statsOnly( ++_aborted_yield );
3636 return;
3637 }
3638
3639 // (5) We check whether we've reached our time quota. If we have,
3640 // then we abort.
3641 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3642 if (elapsed_time_ms > _time_target_ms) {
3643 set_has_aborted();
3644 _has_timed_out = true;
3645 statsOnly( ++_aborted_timed_out );
3646 return;
3647 }
3648
3649 // (6) Finally, we check whether there are enough completed STAB
3650 // buffers available for processing. If there are, we abort.
3651 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3652 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3653 if (_cm->verbose_low()) {
3654 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3655 _worker_id);
3656 }
3657 // we do need to process SATB buffers, we'll abort and restart
3658 // the marking task to do so
3659 set_has_aborted();
3660 statsOnly( ++_aborted_satb );
3661 return;
3662 }
3663 }
3664
3665 void CMTask::recalculate_limits() {
3666 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3667 _words_scanned_limit = _real_words_scanned_limit;
3668
3669 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3670 _refs_reached_limit = _real_refs_reached_limit;
3671 }
3672
3673 void CMTask::decrease_limits() {
3674 // This is called when we believe that we're going to do an infrequent
3675 // operation which will increase the per byte scanned cost (i.e. move
3676 // entries to/from the global stack). It basically tries to decrease the
3677 // scanning limit so that the clock is called earlier.
3678
3679 if (_cm->verbose_medium()) {
3680 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3681 }
3682
3683 _words_scanned_limit = _real_words_scanned_limit -
3684 3 * words_scanned_period / 4;
3685 _refs_reached_limit = _real_refs_reached_limit -
3686 3 * refs_reached_period / 4;
3687 }
3688
3689 void CMTask::move_entries_to_global_stack() {
3690 // local array where we'll store the entries that will be popped
3691 // from the local queue
3692 oop buffer[global_stack_transfer_size];
3693
3694 int n = 0;
3695 oop obj;
3696 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3697 buffer[n] = obj;
3698 ++n;
3699 }
3700
3701 if (n > 0) {
3702 // we popped at least one entry from the local queue
3703
3704 statsOnly( ++_global_transfers_to; _local_pops += n );
3705
3706 if (!_cm->mark_stack_push(buffer, n)) {
3707 if (_cm->verbose_low()) {
3708 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3709 _worker_id);
3710 }
3711 set_has_aborted();
3712 } else {
3713 // the transfer was successful
3714
3715 if (_cm->verbose_medium()) {
3716 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3717 _worker_id, n);
3718 }
3719 statsOnly( int tmp_size = _cm->mark_stack_size();
3720 if (tmp_size > _global_max_size) {
3721 _global_max_size = tmp_size;
3722 }
3723 _global_pushes += n );
3724 }
3725 }
3726
3727 // this operation was quite expensive, so decrease the limits
3728 decrease_limits();
3729 }
3730
3731 void CMTask::get_entries_from_global_stack() {
3732 // local array where we'll store the entries that will be popped
3733 // from the global stack.
3734 oop buffer[global_stack_transfer_size];
3735 int n;
3736 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3737 assert(n <= global_stack_transfer_size,
3738 "we should not pop more than the given limit");
3739 if (n > 0) {
3740 // yes, we did actually pop at least one entry
3741
3742 statsOnly( ++_global_transfers_from; _global_pops += n );
3743 if (_cm->verbose_medium()) {
3744 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3745 _worker_id, n);
3746 }
3747 for (int i = 0; i < n; ++i) {
3748 bool success = _task_queue->push(buffer[i]);
3749 // We only call this when the local queue is empty or under a
3750 // given target limit. So, we do not expect this push to fail.
3751 assert(success, "invariant");
3752 }
3753
3754 statsOnly( int tmp_size = _task_queue->size();
3755 if (tmp_size > _local_max_size) {
3756 _local_max_size = tmp_size;
3757 }
3758 _local_pushes += n );
3759 }
3760
3761 // this operation was quite expensive, so decrease the limits
3762 decrease_limits();
3763 }
3764
3765 void CMTask::drain_local_queue(bool partially) {
3766 if (has_aborted()) return;
3767
3768 // Decide what the target size is, depending whether we're going to
3769 // drain it partially (so that other tasks can steal if they run out
3770 // of things to do) or totally (at the very end).
3771 size_t target_size;
3772 if (partially) {
3773 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3774 } else {
3775 target_size = 0;
3776 }
3777
3778 if (_task_queue->size() > target_size) {
3779 if (_cm->verbose_high()) {
3780 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3781 _worker_id, target_size);
3782 }
3783
3784 oop obj;
3785 bool ret = _task_queue->pop_local(obj);
3786 while (ret) {
3787 statsOnly( ++_local_pops );
3788
3789 if (_cm->verbose_high()) {
3790 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3791 p2i((void*) obj));
3792 }
3793
3794 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3795 assert(!_g1h->is_on_master_free_list(
3796 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3797
3798 scan_object(obj);
3799
3800 if (_task_queue->size() <= target_size || has_aborted()) {
3801 ret = false;
3802 } else {
3803 ret = _task_queue->pop_local(obj);
3804 }
3805 }
3806
3807 if (_cm->verbose_high()) {
3808 gclog_or_tty->print_cr("[%u] drained local queue, size = %u",
3809 _worker_id, _task_queue->size());
3810 }
3811 }
3812 }
3813
3814 void CMTask::drain_global_stack(bool partially) {
3815 if (has_aborted()) return;
3816
3817 // We have a policy to drain the local queue before we attempt to
3818 // drain the global stack.
3819 assert(partially || _task_queue->size() == 0, "invariant");
3820
3821 // Decide what the target size is, depending whether we're going to
3822 // drain it partially (so that other tasks can steal if they run out
3823 // of things to do) or totally (at the very end). Notice that,
3824 // because we move entries from the global stack in chunks or
3825 // because another task might be doing the same, we might in fact
3826 // drop below the target. But, this is not a problem.
3827 size_t target_size;
3828 if (partially) {
3829 target_size = _cm->partial_mark_stack_size_target();
3830 } else {
3831 target_size = 0;
3832 }
3833
3834 if (_cm->mark_stack_size() > target_size) {
3835 if (_cm->verbose_low()) {
3836 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3837 _worker_id, target_size);
3838 }
3839
3840 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3841 get_entries_from_global_stack();
3842 drain_local_queue(partially);
3843 }
3844
3845 if (_cm->verbose_low()) {
3846 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3847 _worker_id, _cm->mark_stack_size());
3848 }
3849 }
3850 }
3851
3852 // SATB Queue has several assumptions on whether to call the par or
3853 // non-par versions of the methods. this is why some of the code is
3854 // replicated. We should really get rid of the single-threaded version
3855 // of the code to simplify things.
3856 void CMTask::drain_satb_buffers() {
3857 if (has_aborted()) return;
3858
3859 // We set this so that the regular clock knows that we're in the
3860 // middle of draining buffers and doesn't set the abort flag when it
3861 // notices that SATB buffers are available for draining. It'd be
3862 // very counter productive if it did that. :-)
3863 _draining_satb_buffers = true;
3864
3865 CMObjectClosure oc(this);
3866 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3867 satb_mq_set.set_closure(_worker_id, &oc);
3868
3869 // This keeps claiming and applying the closure to completed buffers
3870 // until we run out of buffers or we need to abort.
3871 while (!has_aborted() &&
3872 satb_mq_set.apply_closure_to_completed_buffer(_worker_id)) {
3873 if (_cm->verbose_medium()) {
3874 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3875 }
3876 statsOnly( ++_satb_buffers_processed );
3877 regular_clock_call();
3878 }
3879
3880 _draining_satb_buffers = false;
3881
3882 assert(has_aborted() ||
3883 concurrent() ||
3884 satb_mq_set.completed_buffers_num() == 0, "invariant");
3885
3886 satb_mq_set.set_closure(_worker_id, NULL);
3887
3888 // again, this was a potentially expensive operation, decrease the
3889 // limits to get the regular clock call early
3890 decrease_limits();
3891 }
3892
3893 void CMTask::print_stats() {
3894 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3895 _worker_id, _calls);
3896 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3897 _elapsed_time_ms, _termination_time_ms);
3898 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3899 _step_times_ms.num(), _step_times_ms.avg(),
3900 _step_times_ms.sd());
3901 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3902 _step_times_ms.maximum(), _step_times_ms.sum());
3903
3904 #if _MARKING_STATS_
3905 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3906 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3907 _all_clock_intervals_ms.sd());
3908 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3909 _all_clock_intervals_ms.maximum(),
3910 _all_clock_intervals_ms.sum());
3911 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
3912 _clock_due_to_scanning, _clock_due_to_marking);
3913 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
3914 _objs_scanned, _objs_found_on_bitmap);
3915 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
3916 _local_pushes, _local_pops, _local_max_size);
3917 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
3918 _global_pushes, _global_pops, _global_max_size);
3919 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
3920 _global_transfers_to,_global_transfers_from);
3921 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
3922 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
3923 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
3924 _steal_attempts, _steals);
3925 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
3926 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
3927 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3928 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
3929 _aborted_timed_out, _aborted_satb, _aborted_termination);
3930 #endif // _MARKING_STATS_
3931 }
3932
3933 /*****************************************************************************
3934
3935 The do_marking_step(time_target_ms, ...) method is the building
3936 block of the parallel marking framework. It can be called in parallel
3937 with other invocations of do_marking_step() on different tasks
3938 (but only one per task, obviously) and concurrently with the
3939 mutator threads, or during remark, hence it eliminates the need
3940 for two versions of the code. When called during remark, it will
3941 pick up from where the task left off during the concurrent marking
3942 phase. Interestingly, tasks are also claimable during evacuation
3943 pauses too, since do_marking_step() ensures that it aborts before
3944 it needs to yield.
3945
3946 The data structures that it uses to do marking work are the
3947 following:
3948
3949 (1) Marking Bitmap. If there are gray objects that appear only
3950 on the bitmap (this happens either when dealing with an overflow
3951 or when the initial marking phase has simply marked the roots
3952 and didn't push them on the stack), then tasks claim heap
3953 regions whose bitmap they then scan to find gray objects. A
3954 global finger indicates where the end of the last claimed region
3955 is. A local finger indicates how far into the region a task has
3956 scanned. The two fingers are used to determine how to gray an
3957 object (i.e. whether simply marking it is OK, as it will be
3958 visited by a task in the future, or whether it needs to be also
3959 pushed on a stack).
3960
3961 (2) Local Queue. The local queue of the task which is accessed
3962 reasonably efficiently by the task. Other tasks can steal from
3963 it when they run out of work. Throughout the marking phase, a
3964 task attempts to keep its local queue short but not totally
3965 empty, so that entries are available for stealing by other
3966 tasks. Only when there is no more work, a task will totally
3967 drain its local queue.
3968
3969 (3) Global Mark Stack. This handles local queue overflow. During
3970 marking only sets of entries are moved between it and the local
3971 queues, as access to it requires a mutex and more fine-grain
3972 interaction with it which might cause contention. If it
3973 overflows, then the marking phase should restart and iterate
3974 over the bitmap to identify gray objects. Throughout the marking
3975 phase, tasks attempt to keep the global mark stack at a small
3976 length but not totally empty, so that entries are available for
3977 popping by other tasks. Only when there is no more work, tasks
3978 will totally drain the global mark stack.
3979
3980 (4) SATB Buffer Queue. This is where completed SATB buffers are
3981 made available. Buffers are regularly removed from this queue
3982 and scanned for roots, so that the queue doesn't get too
3983 long. During remark, all completed buffers are processed, as
3984 well as the filled in parts of any uncompleted buffers.
3985
3986 The do_marking_step() method tries to abort when the time target
3987 has been reached. There are a few other cases when the
3988 do_marking_step() method also aborts:
3989
3990 (1) When the marking phase has been aborted (after a Full GC).
3991
3992 (2) When a global overflow (on the global stack) has been
3993 triggered. Before the task aborts, it will actually sync up with
3994 the other tasks to ensure that all the marking data structures
3995 (local queues, stacks, fingers etc.) are re-initialized so that
3996 when do_marking_step() completes, the marking phase can
3997 immediately restart.
3998
3999 (3) When enough completed SATB buffers are available. The
4000 do_marking_step() method only tries to drain SATB buffers right
4001 at the beginning. So, if enough buffers are available, the
4002 marking step aborts and the SATB buffers are processed at
4003 the beginning of the next invocation.
4004
4005 (4) To yield. when we have to yield then we abort and yield
4006 right at the end of do_marking_step(). This saves us from a lot
4007 of hassle as, by yielding we might allow a Full GC. If this
4008 happens then objects will be compacted underneath our feet, the
4009 heap might shrink, etc. We save checking for this by just
4010 aborting and doing the yield right at the end.
4011
4012 From the above it follows that the do_marking_step() method should
4013 be called in a loop (or, otherwise, regularly) until it completes.
4014
4015 If a marking step completes without its has_aborted() flag being
4016 true, it means it has completed the current marking phase (and
4017 also all other marking tasks have done so and have all synced up).
4018
4019 A method called regular_clock_call() is invoked "regularly" (in
4020 sub ms intervals) throughout marking. It is this clock method that
4021 checks all the abort conditions which were mentioned above and
4022 decides when the task should abort. A work-based scheme is used to
4023 trigger this clock method: when the number of object words the
4024 marking phase has scanned or the number of references the marking
4025 phase has visited reach a given limit. Additional invocations to
4026 the method clock have been planted in a few other strategic places
4027 too. The initial reason for the clock method was to avoid calling
4028 vtime too regularly, as it is quite expensive. So, once it was in
4029 place, it was natural to piggy-back all the other conditions on it
4030 too and not constantly check them throughout the code.
4031
4032 If do_termination is true then do_marking_step will enter its
4033 termination protocol.
4034
4035 The value of is_serial must be true when do_marking_step is being
4036 called serially (i.e. by the VMThread) and do_marking_step should
4037 skip any synchronization in the termination and overflow code.
4038 Examples include the serial remark code and the serial reference
4039 processing closures.
4040
4041 The value of is_serial must be false when do_marking_step is
4042 being called by any of the worker threads in a work gang.
4043 Examples include the concurrent marking code (CMMarkingTask),
4044 the MT remark code, and the MT reference processing closures.
4045
4046 *****************************************************************************/
4047
4048 void CMTask::do_marking_step(double time_target_ms,
4049 bool do_termination,
4050 bool is_serial) {
4051 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4052 assert(concurrent() == _cm->concurrent(), "they should be the same");
4053
4054 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4055 assert(_task_queues != NULL, "invariant");
4056 assert(_task_queue != NULL, "invariant");
4057 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4058
4059 assert(!_claimed,
4060 "only one thread should claim this task at any one time");
4061
4062 // OK, this doesn't safeguard again all possible scenarios, as it is
4063 // possible for two threads to set the _claimed flag at the same
4064 // time. But it is only for debugging purposes anyway and it will
4065 // catch most problems.
4066 _claimed = true;
4067
4068 _start_time_ms = os::elapsedVTime() * 1000.0;
4069 statsOnly( _interval_start_time_ms = _start_time_ms );
4070
4071 // If do_stealing is true then do_marking_step will attempt to
4072 // steal work from the other CMTasks. It only makes sense to
4073 // enable stealing when the termination protocol is enabled
4074 // and do_marking_step() is not being called serially.
4075 bool do_stealing = do_termination && !is_serial;
4076
4077 double diff_prediction_ms =
4078 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4079 _time_target_ms = time_target_ms - diff_prediction_ms;
4080
4081 // set up the variables that are used in the work-based scheme to
4082 // call the regular clock method
4083 _words_scanned = 0;
4084 _refs_reached = 0;
4085 recalculate_limits();
4086
4087 // clear all flags
4088 clear_has_aborted();
4089 _has_timed_out = false;
4090 _draining_satb_buffers = false;
4091
4092 ++_calls;
4093
4094 if (_cm->verbose_low()) {
4095 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4096 "target = %1.2lfms >>>>>>>>>>",
4097 _worker_id, _calls, _time_target_ms);
4098 }
4099
4100 // Set up the bitmap and oop closures. Anything that uses them is
4101 // eventually called from this method, so it is OK to allocate these
4102 // statically.
4103 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4104 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4105 set_cm_oop_closure(&cm_oop_closure);
4106
4107 if (_cm->has_overflown()) {
4108 // This can happen if the mark stack overflows during a GC pause
4109 // and this task, after a yield point, restarts. We have to abort
4110 // as we need to get into the overflow protocol which happens
4111 // right at the end of this task.
4112 set_has_aborted();
4113 }
4114
4115 // First drain any available SATB buffers. After this, we will not
4116 // look at SATB buffers before the next invocation of this method.
4117 // If enough completed SATB buffers are queued up, the regular clock
4118 // will abort this task so that it restarts.
4119 drain_satb_buffers();
4120 // ...then partially drain the local queue and the global stack
4121 drain_local_queue(true);
4122 drain_global_stack(true);
4123
4124 do {
4125 if (!has_aborted() && _curr_region != NULL) {
4126 // This means that we're already holding on to a region.
4127 assert(_finger != NULL, "if region is not NULL, then the finger "
4128 "should not be NULL either");
4129
4130 // We might have restarted this task after an evacuation pause
4131 // which might have evacuated the region we're holding on to
4132 // underneath our feet. Let's read its limit again to make sure
4133 // that we do not iterate over a region of the heap that
4134 // contains garbage (update_region_limit() will also move
4135 // _finger to the start of the region if it is found empty).
4136 update_region_limit();
4137 // We will start from _finger not from the start of the region,
4138 // as we might be restarting this task after aborting half-way
4139 // through scanning this region. In this case, _finger points to
4140 // the address where we last found a marked object. If this is a
4141 // fresh region, _finger points to start().
4142 MemRegion mr = MemRegion(_finger, _region_limit);
4143
4144 if (_cm->verbose_low()) {
4145 gclog_or_tty->print_cr("[%u] we're scanning part "
4146 "["PTR_FORMAT", "PTR_FORMAT") "
4147 "of region "HR_FORMAT,
4148 _worker_id, p2i(_finger), p2i(_region_limit),
4149 HR_FORMAT_PARAMS(_curr_region));
4150 }
4151
4152 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
4153 "humongous regions should go around loop once only");
4154
4155 // Some special cases:
4156 // If the memory region is empty, we can just give up the region.
4157 // If the current region is humongous then we only need to check
4158 // the bitmap for the bit associated with the start of the object,
4159 // scan the object if it's live, and give up the region.
4160 // Otherwise, let's iterate over the bitmap of the part of the region
4161 // that is left.
4162 // If the iteration is successful, give up the region.
4163 if (mr.is_empty()) {
4164 giveup_current_region();
4165 regular_clock_call();
4166 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
4167 if (_nextMarkBitMap->isMarked(mr.start())) {
4168 // The object is marked - apply the closure
4169 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4170 bitmap_closure.do_bit(offset);
4171 }
4172 // Even if this task aborted while scanning the humongous object
4173 // we can (and should) give up the current region.
4174 giveup_current_region();
4175 regular_clock_call();
4176 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4177 giveup_current_region();
4178 regular_clock_call();
4179 } else {
4180 assert(has_aborted(), "currently the only way to do so");
4181 // The only way to abort the bitmap iteration is to return
4182 // false from the do_bit() method. However, inside the
4183 // do_bit() method we move the _finger to point to the
4184 // object currently being looked at. So, if we bail out, we
4185 // have definitely set _finger to something non-null.
4186 assert(_finger != NULL, "invariant");
4187
4188 // Region iteration was actually aborted. So now _finger
4189 // points to the address of the object we last scanned. If we
4190 // leave it there, when we restart this task, we will rescan
4191 // the object. It is easy to avoid this. We move the finger by
4192 // enough to point to the next possible object header (the
4193 // bitmap knows by how much we need to move it as it knows its
4194 // granularity).
4195 assert(_finger < _region_limit, "invariant");
4196 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4197 // Check if bitmap iteration was aborted while scanning the last object
4198 if (new_finger >= _region_limit) {
4199 giveup_current_region();
4200 } else {
4201 move_finger_to(new_finger);
4202 }
4203 }
4204 }
4205 // At this point we have either completed iterating over the
4206 // region we were holding on to, or we have aborted.
4207
4208 // We then partially drain the local queue and the global stack.
4209 // (Do we really need this?)
4210 drain_local_queue(true);
4211 drain_global_stack(true);
4212
4213 // Read the note on the claim_region() method on why it might
4214 // return NULL with potentially more regions available for
4215 // claiming and why we have to check out_of_regions() to determine
4216 // whether we're done or not.
4217 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4218 // We are going to try to claim a new region. We should have
4219 // given up on the previous one.
4220 // Separated the asserts so that we know which one fires.
4221 assert(_curr_region == NULL, "invariant");
4222 assert(_finger == NULL, "invariant");
4223 assert(_region_limit == NULL, "invariant");
4224 if (_cm->verbose_low()) {
4225 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4226 }
4227 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4228 if (claimed_region != NULL) {
4229 // Yes, we managed to claim one
4230 statsOnly( ++_regions_claimed );
4231
4232 if (_cm->verbose_low()) {
4233 gclog_or_tty->print_cr("[%u] we successfully claimed "
4234 "region "PTR_FORMAT,
4235 _worker_id, p2i(claimed_region));
4236 }
4237
4238 setup_for_region(claimed_region);
4239 assert(_curr_region == claimed_region, "invariant");
4240 }
4241 // It is important to call the regular clock here. It might take
4242 // a while to claim a region if, for example, we hit a large
4243 // block of empty regions. So we need to call the regular clock
4244 // method once round the loop to make sure it's called
4245 // frequently enough.
4246 regular_clock_call();
4247 }
4248
4249 if (!has_aborted() && _curr_region == NULL) {
4250 assert(_cm->out_of_regions(),
4251 "at this point we should be out of regions");
4252 }
4253 } while ( _curr_region != NULL && !has_aborted());
4254
4255 if (!has_aborted()) {
4256 // We cannot check whether the global stack is empty, since other
4257 // tasks might be pushing objects to it concurrently.
4258 assert(_cm->out_of_regions(),
4259 "at this point we should be out of regions");
4260
4261 if (_cm->verbose_low()) {
4262 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4263 }
4264
4265 // Try to reduce the number of available SATB buffers so that
4266 // remark has less work to do.
4267 drain_satb_buffers();
4268 }
4269
4270 // Since we've done everything else, we can now totally drain the
4271 // local queue and global stack.
4272 drain_local_queue(false);
4273 drain_global_stack(false);
4274
4275 // Attempt at work stealing from other task's queues.
4276 if (do_stealing && !has_aborted()) {
4277 // We have not aborted. This means that we have finished all that
4278 // we could. Let's try to do some stealing...
4279
4280 // We cannot check whether the global stack is empty, since other
4281 // tasks might be pushing objects to it concurrently.
4282 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4283 "only way to reach here");
4284
4285 if (_cm->verbose_low()) {
4286 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4287 }
4288
4289 while (!has_aborted()) {
4290 oop obj;
4291 statsOnly( ++_steal_attempts );
4292
4293 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4294 if (_cm->verbose_medium()) {
4295 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4296 _worker_id, p2i((void*) obj));
4297 }
4298
4299 statsOnly( ++_steals );
4300
4301 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4302 "any stolen object should be marked");
4303 scan_object(obj);
4304
4305 // And since we're towards the end, let's totally drain the
4306 // local queue and global stack.
4307 drain_local_queue(false);
4308 drain_global_stack(false);
4309 } else {
4310 break;
4311 }
4312 }
4313 }
4314
4315 // If we are about to wrap up and go into termination, check if we
4316 // should raise the overflow flag.
4317 if (do_termination && !has_aborted()) {
4318 if (_cm->force_overflow()->should_force()) {
4319 _cm->set_has_overflown();
4320 regular_clock_call();
4321 }
4322 }
4323
4324 // We still haven't aborted. Now, let's try to get into the
4325 // termination protocol.
4326 if (do_termination && !has_aborted()) {
4327 // We cannot check whether the global stack is empty, since other
4328 // tasks might be concurrently pushing objects on it.
4329 // Separated the asserts so that we know which one fires.
4330 assert(_cm->out_of_regions(), "only way to reach here");
4331 assert(_task_queue->size() == 0, "only way to reach here");
4332
4333 if (_cm->verbose_low()) {
4334 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4335 }
4336
4337 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4338
4339 // The CMTask class also extends the TerminatorTerminator class,
4340 // hence its should_exit_termination() method will also decide
4341 // whether to exit the termination protocol or not.
4342 bool finished = (is_serial ||
4343 _cm->terminator()->offer_termination(this));
4344 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4345 _termination_time_ms +=
4346 termination_end_time_ms - _termination_start_time_ms;
4347
4348 if (finished) {
4349 // We're all done.
4350
4351 if (_worker_id == 0) {
4352 // let's allow task 0 to do this
4353 if (concurrent()) {
4354 assert(_cm->concurrent_marking_in_progress(), "invariant");
4355 // we need to set this to false before the next
4356 // safepoint. This way we ensure that the marking phase
4357 // doesn't observe any more heap expansions.
4358 _cm->clear_concurrent_marking_in_progress();
4359 }
4360 }
4361
4362 // We can now guarantee that the global stack is empty, since
4363 // all other tasks have finished. We separated the guarantees so
4364 // that, if a condition is false, we can immediately find out
4365 // which one.
4366 guarantee(_cm->out_of_regions(), "only way to reach here");
4367 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4368 guarantee(_task_queue->size() == 0, "only way to reach here");
4369 guarantee(!_cm->has_overflown(), "only way to reach here");
4370 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4371
4372 if (_cm->verbose_low()) {
4373 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4374 }
4375 } else {
4376 // Apparently there's more work to do. Let's abort this task. It
4377 // will restart it and we can hopefully find more things to do.
4378
4379 if (_cm->verbose_low()) {
4380 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4381 _worker_id);
4382 }
4383
4384 set_has_aborted();
4385 statsOnly( ++_aborted_termination );
4386 }
4387 }
4388
4389 // Mainly for debugging purposes to make sure that a pointer to the
4390 // closure which was statically allocated in this frame doesn't
4391 // escape it by accident.
4392 set_cm_oop_closure(NULL);
4393 double end_time_ms = os::elapsedVTime() * 1000.0;
4394 double elapsed_time_ms = end_time_ms - _start_time_ms;
4395 // Update the step history.
4396 _step_times_ms.add(elapsed_time_ms);
4397
4398 if (has_aborted()) {
4399 // The task was aborted for some reason.
4400
4401 statsOnly( ++_aborted );
4402
4403 if (_has_timed_out) {
4404 double diff_ms = elapsed_time_ms - _time_target_ms;
4405 // Keep statistics of how well we did with respect to hitting
4406 // our target only if we actually timed out (if we aborted for
4407 // other reasons, then the results might get skewed).
4408 _marking_step_diffs_ms.add(diff_ms);
4409 }
4410
4411 if (_cm->has_overflown()) {
4412 // This is the interesting one. We aborted because a global
4413 // overflow was raised. This means we have to restart the
4414 // marking phase and start iterating over regions. However, in
4415 // order to do this we have to make sure that all tasks stop
4416 // what they are doing and re-initialize in a safe manner. We
4417 // will achieve this with the use of two barrier sync points.
4418
4419 if (_cm->verbose_low()) {
4420 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4421 }
4422
4423 if (!is_serial) {
4424 // We only need to enter the sync barrier if being called
4425 // from a parallel context
4426 _cm->enter_first_sync_barrier(_worker_id);
4427
4428 // When we exit this sync barrier we know that all tasks have
4429 // stopped doing marking work. So, it's now safe to
4430 // re-initialize our data structures. At the end of this method,
4431 // task 0 will clear the global data structures.
4432 }
4433
4434 statsOnly( ++_aborted_overflow );
4435
4436 // We clear the local state of this task...
4437 clear_region_fields();
4438
4439 if (!is_serial) {
4440 // ...and enter the second barrier.
4441 _cm->enter_second_sync_barrier(_worker_id);
4442 }
4443 // At this point, if we're during the concurrent phase of
4444 // marking, everything has been re-initialized and we're
4445 // ready to restart.
4446 }
4447
4448 if (_cm->verbose_low()) {
4449 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4450 "elapsed = %1.2lfms <<<<<<<<<<",
4451 _worker_id, _time_target_ms, elapsed_time_ms);
4452 if (_cm->has_aborted()) {
4453 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4454 _worker_id);
4455 }
4456 }
4457 } else {
4458 if (_cm->verbose_low()) {
4459 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4460 "elapsed = %1.2lfms <<<<<<<<<<",
4461 _worker_id, _time_target_ms, elapsed_time_ms);
4462 }
4463 }
4464
4465 _claimed = false;
4466 }
4467
4468 CMTask::CMTask(uint worker_id,
4469 ConcurrentMark* cm,
4470 size_t* marked_bytes,
4471 BitMap* card_bm,
4472 CMTaskQueue* task_queue,
4473 CMTaskQueueSet* task_queues)
4474 : _g1h(G1CollectedHeap::heap()),
4475 _worker_id(worker_id), _cm(cm),
4476 _claimed(false),
4477 _nextMarkBitMap(NULL), _hash_seed(17),
4478 _task_queue(task_queue),
4479 _task_queues(task_queues),
4480 _cm_oop_closure(NULL),
4481 _marked_bytes_array(marked_bytes),
4482 _card_bm(card_bm) {
4483 guarantee(task_queue != NULL, "invariant");
4484 guarantee(task_queues != NULL, "invariant");
4485
4486 statsOnly( _clock_due_to_scanning = 0;
4487 _clock_due_to_marking = 0 );
4488
4489 _marking_step_diffs_ms.add(0.5);
4490 }
4491
4492 // These are formatting macros that are used below to ensure
4493 // consistent formatting. The *_H_* versions are used to format the
4494 // header for a particular value and they should be kept consistent
4495 // with the corresponding macro. Also note that most of the macros add
4496 // the necessary white space (as a prefix) which makes them a bit
4497 // easier to compose.
4498
4499 // All the output lines are prefixed with this string to be able to
4500 // identify them easily in a large log file.
4501 #define G1PPRL_LINE_PREFIX "###"
4502
4503 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4504 #ifdef _LP64
4505 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4506 #else // _LP64
4507 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4508 #endif // _LP64
4509
4510 // For per-region info
4511 #define G1PPRL_TYPE_FORMAT " %-4s"
4512 #define G1PPRL_TYPE_H_FORMAT " %4s"
4513 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4514 #define G1PPRL_BYTE_H_FORMAT " %9s"
4515 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4516 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4517
4518 // For summary info
4519 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4520 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4521 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4522 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4523
4524 G1PrintRegionLivenessInfoClosure::
4525 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4526 : _out(out),
4527 _total_used_bytes(0), _total_capacity_bytes(0),
4528 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4529 _hum_used_bytes(0), _hum_capacity_bytes(0),
4530 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4531 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4532 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4533 MemRegion g1_reserved = g1h->g1_reserved();
4534 double now = os::elapsedTime();
4535
4536 // Print the header of the output.
4537 _out->cr();
4538 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4539 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4540 G1PPRL_SUM_ADDR_FORMAT("reserved")
4541 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4542 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4543 HeapRegion::GrainBytes);
4544 _out->print_cr(G1PPRL_LINE_PREFIX);
4545 _out->print_cr(G1PPRL_LINE_PREFIX
4546 G1PPRL_TYPE_H_FORMAT
4547 G1PPRL_ADDR_BASE_H_FORMAT
4548 G1PPRL_BYTE_H_FORMAT
4549 G1PPRL_BYTE_H_FORMAT
4550 G1PPRL_BYTE_H_FORMAT
4551 G1PPRL_DOUBLE_H_FORMAT
4552 G1PPRL_BYTE_H_FORMAT
4553 G1PPRL_BYTE_H_FORMAT,
4554 "type", "address-range",
4555 "used", "prev-live", "next-live", "gc-eff",
4556 "remset", "code-roots");
4557 _out->print_cr(G1PPRL_LINE_PREFIX
4558 G1PPRL_TYPE_H_FORMAT
4559 G1PPRL_ADDR_BASE_H_FORMAT
4560 G1PPRL_BYTE_H_FORMAT
4561 G1PPRL_BYTE_H_FORMAT
4562 G1PPRL_BYTE_H_FORMAT
4563 G1PPRL_DOUBLE_H_FORMAT
4564 G1PPRL_BYTE_H_FORMAT
4565 G1PPRL_BYTE_H_FORMAT,
4566 "", "",
4567 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4568 "(bytes)", "(bytes)");
4569 }
4570
4571 // It takes as a parameter a reference to one of the _hum_* fields, it
4572 // deduces the corresponding value for a region in a humongous region
4573 // series (either the region size, or what's left if the _hum_* field
4574 // is < the region size), and updates the _hum_* field accordingly.
4575 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4576 size_t bytes = 0;
4577 // The > 0 check is to deal with the prev and next live bytes which
4578 // could be 0.
4579 if (*hum_bytes > 0) {
4580 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4581 *hum_bytes -= bytes;
4582 }
4583 return bytes;
4584 }
4585
4586 // It deduces the values for a region in a humongous region series
4587 // from the _hum_* fields and updates those accordingly. It assumes
4588 // that that _hum_* fields have already been set up from the "starts
4589 // humongous" region and we visit the regions in address order.
4590 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4591 size_t* capacity_bytes,
4592 size_t* prev_live_bytes,
4593 size_t* next_live_bytes) {
4594 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4595 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4596 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4597 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4598 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4599 }
4600
4601 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4602 const char* type = r->get_type_str();
4603 HeapWord* bottom = r->bottom();
4604 HeapWord* end = r->end();
4605 size_t capacity_bytes = r->capacity();
4606 size_t used_bytes = r->used();
4607 size_t prev_live_bytes = r->live_bytes();
4608 size_t next_live_bytes = r->next_live_bytes();
4609 double gc_eff = r->gc_efficiency();
4610 size_t remset_bytes = r->rem_set()->mem_size();
4611 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4612
4613 if (r->is_starts_humongous()) {
4614 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4615 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4616 "they should have been zeroed after the last time we used them");
4617 // Set up the _hum_* fields.
4618 _hum_capacity_bytes = capacity_bytes;
4619 _hum_used_bytes = used_bytes;
4620 _hum_prev_live_bytes = prev_live_bytes;
4621 _hum_next_live_bytes = next_live_bytes;
4622 get_hum_bytes(&used_bytes, &capacity_bytes,
4623 &prev_live_bytes, &next_live_bytes);
4624 end = bottom + HeapRegion::GrainWords;
4625 } else if (r->is_continues_humongous()) {
4626 get_hum_bytes(&used_bytes, &capacity_bytes,
4627 &prev_live_bytes, &next_live_bytes);
4628 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4629 }
4630
4631 _total_used_bytes += used_bytes;
4632 _total_capacity_bytes += capacity_bytes;
4633 _total_prev_live_bytes += prev_live_bytes;
4634 _total_next_live_bytes += next_live_bytes;
4635 _total_remset_bytes += remset_bytes;
4636 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4637
4638 // Print a line for this particular region.
4639 _out->print_cr(G1PPRL_LINE_PREFIX
4640 G1PPRL_TYPE_FORMAT
4641 G1PPRL_ADDR_BASE_FORMAT
4642 G1PPRL_BYTE_FORMAT
4643 G1PPRL_BYTE_FORMAT
4644 G1PPRL_BYTE_FORMAT
4645 G1PPRL_DOUBLE_FORMAT
4646 G1PPRL_BYTE_FORMAT
4647 G1PPRL_BYTE_FORMAT,
4648 type, p2i(bottom), p2i(end),
4649 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4650 remset_bytes, strong_code_roots_bytes);
4651
4652 return false;
4653 }
4654
4655 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4656 // add static memory usages to remembered set sizes
4657 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4658 // Print the footer of the output.
4659 _out->print_cr(G1PPRL_LINE_PREFIX);
4660 _out->print_cr(G1PPRL_LINE_PREFIX
4661 " SUMMARY"
4662 G1PPRL_SUM_MB_FORMAT("capacity")
4663 G1PPRL_SUM_MB_PERC_FORMAT("used")
4664 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4665 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4666 G1PPRL_SUM_MB_FORMAT("remset")
4667 G1PPRL_SUM_MB_FORMAT("code-roots"),
4668 bytes_to_mb(_total_capacity_bytes),
4669 bytes_to_mb(_total_used_bytes),
4670 perc(_total_used_bytes, _total_capacity_bytes),
4671 bytes_to_mb(_total_prev_live_bytes),
4672 perc(_total_prev_live_bytes, _total_capacity_bytes),
4673 bytes_to_mb(_total_next_live_bytes),
4674 perc(_total_next_live_bytes, _total_capacity_bytes),
4675 bytes_to_mb(_total_remset_bytes),
4676 bytes_to_mb(_total_strong_code_roots_bytes));
4677 _out->cr();
4678 }