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/metadataOnStackMark.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/g1/concurrentMark.inline.hpp"
30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
32 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
33 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
34 #include "gc_implementation/g1/g1Log.hpp"
35 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
36 #include "gc_implementation/g1/g1RemSet.hpp"
37 #include "gc_implementation/g1/heapRegion.inline.hpp"
38 #include "gc_implementation/g1/heapRegionManager.inline.hpp"
39 #include "gc_implementation/g1/heapRegionRemSet.hpp"
40 #include "gc_implementation/g1/heapRegionSet.inline.hpp"
41 #include "gc_implementation/shared/vmGCOperations.hpp"
42 #include "gc_implementation/shared/gcTimer.hpp"
43 #include "gc_implementation/shared/gcTrace.hpp"
44 #include "gc_implementation/shared/gcTraceTime.hpp"
45 #include "memory/allocation.hpp"
46 #include "memory/genOopClosures.inline.hpp"
47 #include "memory/referencePolicy.hpp"
48 #include "memory/resourceArea.hpp"
49 #include "oops/oop.inline.hpp"
50 #include "runtime/handles.inline.hpp"
51 #include "runtime/java.hpp"
52 #include "runtime/atomic.inline.hpp"
53 #include "runtime/prefetch.inline.hpp"
54 #include "services/memTracker.hpp"
55
56 // Concurrent marking bit map wrapper
57
58 CMBitMapRO::CMBitMapRO(int shifter) :
59 _bm(),
60 _shifter(shifter) {
61 _bmStartWord = 0;
62 _bmWordSize = 0;
63 }
64
65 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
66 const HeapWord* limit) const {
67 // First we must round addr *up* to a possible object boundary.
68 addr = (HeapWord*)align_size_up((intptr_t)addr,
69 HeapWordSize << _shifter);
70 size_t addrOffset = heapWordToOffset(addr);
71 if (limit == NULL) {
72 limit = _bmStartWord + _bmWordSize;
73 }
74 size_t limitOffset = heapWordToOffset(limit);
75 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
76 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
77 assert(nextAddr >= addr, "get_next_one postcondition");
78 assert(nextAddr == limit || isMarked(nextAddr),
79 "get_next_one postcondition");
80 return nextAddr;
81 }
82
83 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
84 const HeapWord* limit) const {
85 size_t addrOffset = heapWordToOffset(addr);
86 if (limit == NULL) {
87 limit = _bmStartWord + _bmWordSize;
88 }
89 size_t limitOffset = heapWordToOffset(limit);
90 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
91 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
92 assert(nextAddr >= addr, "get_next_one postcondition");
93 assert(nextAddr == limit || !isMarked(nextAddr),
94 "get_next_one postcondition");
95 return nextAddr;
96 }
97
98 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
99 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
100 return (int) (diff >> _shifter);
101 }
102
103 #ifndef PRODUCT
104 bool CMBitMapRO::covers(MemRegion heap_rs) const {
105 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
106 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
107 "size inconsistency");
108 return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
109 _bmWordSize == heap_rs.word_size();
110 }
111 #endif
112
113 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
114 _bm.print_on_error(st, prefix);
115 }
116
117 size_t CMBitMap::compute_size(size_t heap_size) {
118 return heap_size / mark_distance();
119 }
120
121 size_t CMBitMap::mark_distance() {
122 return MinObjAlignmentInBytes * BitsPerByte;
123 }
124
125 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
126 _bmStartWord = heap.start();
127 _bmWordSize = heap.word_size();
128
129 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
130 _bm.set_size(_bmWordSize >> _shifter);
131
132 storage->set_mapping_changed_listener(&_listener);
133 }
134
135 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
136 if (zero_filled) {
137 return;
138 }
139 // We need to clear the bitmap on commit, removing any existing information.
140 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
141 _bm->clearRange(mr);
142 }
143
144 // Closure used for clearing the given mark bitmap.
145 class ClearBitmapHRClosure : public HeapRegionClosure {
146 private:
147 ConcurrentMark* _cm;
148 CMBitMap* _bitmap;
149 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration.
150 public:
151 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
152 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
153 }
154
155 virtual bool doHeapRegion(HeapRegion* r) {
156 size_t const chunk_size_in_words = M / HeapWordSize;
157
158 HeapWord* cur = r->bottom();
159 HeapWord* const end = r->end();
160
161 while (cur < end) {
162 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
163 _bitmap->clearRange(mr);
164
165 cur += chunk_size_in_words;
166
167 // Abort iteration if after yielding the marking has been aborted.
168 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
169 return true;
170 }
171 // Repeat the asserts from before the start of the closure. We will do them
172 // as asserts here to minimize their overhead on the product. However, we
173 // will have them as guarantees at the beginning / end of the bitmap
174 // clearing to get some checking in the product.
175 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
176 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
177 }
178
179 return false;
180 }
181 };
182
183 class ParClearNextMarkBitmapTask : public AbstractGangTask {
184 ClearBitmapHRClosure* _cl;
185 HeapRegionClaimer _hrclaimer;
186 bool _suspendible; // If the task is suspendible, workers must join the STS.
187
188 public:
189 ParClearNextMarkBitmapTask(ClearBitmapHRClosure *cl, uint n_workers, bool suspendible) :
190 _cl(cl), _suspendible(suspendible), AbstractGangTask("Parallel Clear Bitmap Task"), _hrclaimer(n_workers) {}
191
192 void work(uint worker_id) {
193 if (_suspendible) {
194 SuspendibleThreadSet::join();
195 }
196 G1CollectedHeap::heap()->heap_region_par_iterate(_cl, worker_id, &_hrclaimer, true);
197 if (_suspendible) {
198 SuspendibleThreadSet::leave();
199 }
200 }
201 };
202
203 void CMBitMap::clearAll() {
204 G1CollectedHeap* g1h = G1CollectedHeap::heap();
205 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
206 uint n_workers = g1h->workers()->active_workers();
207 ParClearNextMarkBitmapTask task(&cl, n_workers, false);
208 g1h->workers()->run_task(&task);
209 guarantee(cl.complete(), "Must have completed iteration.");
210 return;
211 }
212
213 void CMBitMap::markRange(MemRegion mr) {
214 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
215 assert(!mr.is_empty(), "unexpected empty region");
216 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
217 ((HeapWord *) mr.end())),
218 "markRange memory region end is not card aligned");
219 // convert address range into offset range
220 _bm.at_put_range(heapWordToOffset(mr.start()),
221 heapWordToOffset(mr.end()), true);
222 }
223
224 void CMBitMap::clearRange(MemRegion mr) {
225 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
226 assert(!mr.is_empty(), "unexpected empty region");
227 // convert address range into offset range
228 _bm.at_put_range(heapWordToOffset(mr.start()),
229 heapWordToOffset(mr.end()), false);
230 }
231
232 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
233 HeapWord* end_addr) {
234 HeapWord* start = getNextMarkedWordAddress(addr);
235 start = MIN2(start, end_addr);
236 HeapWord* end = getNextUnmarkedWordAddress(start);
237 end = MIN2(end, end_addr);
238 assert(start <= end, "Consistency check");
239 MemRegion mr(start, end);
240 if (!mr.is_empty()) {
241 clearRange(mr);
242 }
243 return mr;
244 }
245
246 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
247 _base(NULL), _cm(cm)
248 #ifdef ASSERT
249 , _drain_in_progress(false)
250 , _drain_in_progress_yields(false)
251 #endif
252 {}
253
254 bool CMMarkStack::allocate(size_t capacity) {
255 // allocate a stack of the requisite depth
256 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
257 if (!rs.is_reserved()) {
258 warning("ConcurrentMark MarkStack allocation failure");
259 return false;
260 }
261 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
262 if (!_virtual_space.initialize(rs, rs.size())) {
263 warning("ConcurrentMark MarkStack backing store failure");
264 // Release the virtual memory reserved for the marking stack
265 rs.release();
266 return false;
267 }
268 assert(_virtual_space.committed_size() == rs.size(),
269 "Didn't reserve backing store for all of ConcurrentMark stack?");
270 _base = (oop*) _virtual_space.low();
271 setEmpty();
272 _capacity = (jint) capacity;
273 _saved_index = -1;
274 _should_expand = false;
275 NOT_PRODUCT(_max_depth = 0);
276 return true;
277 }
278
279 void CMMarkStack::expand() {
280 // Called, during remark, if we've overflown the marking stack during marking.
281 assert(isEmpty(), "stack should been emptied while handling overflow");
282 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
283 // Clear expansion flag
284 _should_expand = false;
285 if (_capacity == (jint) MarkStackSizeMax) {
286 if (PrintGCDetails && Verbose) {
287 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
288 }
289 return;
290 }
291 // Double capacity if possible
292 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
293 // Do not give up existing stack until we have managed to
294 // get the double capacity that we desired.
295 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
296 sizeof(oop)));
297 if (rs.is_reserved()) {
298 // Release the backing store associated with old stack
299 _virtual_space.release();
300 // Reinitialize virtual space for new stack
301 if (!_virtual_space.initialize(rs, rs.size())) {
302 fatal("Not enough swap for expanded marking stack capacity");
303 }
304 _base = (oop*)(_virtual_space.low());
305 _index = 0;
306 _capacity = new_capacity;
307 } else {
308 if (PrintGCDetails && Verbose) {
309 // Failed to double capacity, continue;
310 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
311 SIZE_FORMAT"K to " SIZE_FORMAT"K",
312 _capacity / K, new_capacity / K);
313 }
314 }
315 }
316
317 void CMMarkStack::set_should_expand() {
318 // If we're resetting the marking state because of an
319 // marking stack overflow, record that we should, if
320 // possible, expand the stack.
321 _should_expand = _cm->has_overflown();
322 }
323
324 CMMarkStack::~CMMarkStack() {
325 if (_base != NULL) {
326 _base = NULL;
327 _virtual_space.release();
328 }
329 }
330
331 void CMMarkStack::par_push(oop ptr) {
332 while (true) {
333 if (isFull()) {
334 _overflow = true;
335 return;
336 }
337 // Otherwise...
338 jint index = _index;
339 jint next_index = index+1;
340 jint res = Atomic::cmpxchg(next_index, &_index, index);
341 if (res == index) {
342 _base[index] = ptr;
343 // Note that we don't maintain this atomically. We could, but it
344 // doesn't seem necessary.
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_adjoin_arr(oop* ptr_arr, int n) {
353 while (true) {
354 if (isFull()) {
355 _overflow = true;
356 return;
357 }
358 // Otherwise...
359 jint index = _index;
360 jint next_index = index + n;
361 if (next_index > _capacity) {
362 _overflow = true;
363 return;
364 }
365 jint res = Atomic::cmpxchg(next_index, &_index, index);
366 if (res == index) {
367 for (int i = 0; i < n; i++) {
368 int ind = index + i;
369 assert(ind < _capacity, "By overflow test above.");
370 _base[ind] = ptr_arr[i];
371 }
372 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
373 return;
374 }
375 // Otherwise, we need to try again.
376 }
377 }
378
379 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
380 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
381 jint start = _index;
382 jint next_index = start + n;
383 if (next_index > _capacity) {
384 _overflow = true;
385 return;
386 }
387 // Otherwise.
388 _index = next_index;
389 for (int i = 0; i < n; i++) {
390 int ind = start + i;
391 assert(ind < _capacity, "By overflow test above.");
392 _base[ind] = ptr_arr[i];
393 }
394 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
395 }
396
397 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
398 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
399 jint index = _index;
400 if (index == 0) {
401 *n = 0;
402 return false;
403 } else {
404 int k = MIN2(max, index);
405 jint new_ind = index - k;
406 for (int j = 0; j < k; j++) {
407 ptr_arr[j] = _base[new_ind + j];
408 }
409 _index = new_ind;
410 *n = k;
411 return true;
412 }
413 }
414
415 template<class OopClosureClass>
416 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
417 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
418 || SafepointSynchronize::is_at_safepoint(),
419 "Drain recursion must be yield-safe.");
420 bool res = true;
421 debug_only(_drain_in_progress = true);
422 debug_only(_drain_in_progress_yields = yield_after);
423 while (!isEmpty()) {
424 oop newOop = pop();
425 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
426 assert(newOop->is_oop(), "Expected an oop");
427 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
428 "only grey objects on this stack");
429 newOop->oop_iterate(cl);
430 if (yield_after && _cm->do_yield_check()) {
431 res = false;
432 break;
433 }
434 }
435 debug_only(_drain_in_progress = false);
436 return res;
437 }
438
439 void CMMarkStack::note_start_of_gc() {
440 assert(_saved_index == -1,
441 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
442 _saved_index = _index;
443 }
444
445 void CMMarkStack::note_end_of_gc() {
446 // This is intentionally a guarantee, instead of an assert. If we
447 // accidentally add something to the mark stack during GC, it
448 // will be a correctness issue so it's better if we crash. we'll
449 // only check this once per GC anyway, so it won't be a performance
450 // issue in any way.
451 guarantee(_saved_index == _index,
452 err_msg("saved index: %d index: %d", _saved_index, _index));
453 _saved_index = -1;
454 }
455
456 void CMMarkStack::oops_do(OopClosure* f) {
457 assert(_saved_index == _index,
458 err_msg("saved index: %d index: %d", _saved_index, _index));
459 for (int i = 0; i < _index; i += 1) {
460 f->do_oop(&_base[i]);
461 }
462 }
463
464 CMRootRegions::CMRootRegions() :
465 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
466 _should_abort(false), _next_survivor(NULL) { }
467
468 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
469 _young_list = g1h->young_list();
470 _cm = cm;
471 }
472
473 void CMRootRegions::prepare_for_scan() {
474 assert(!scan_in_progress(), "pre-condition");
475
476 // Currently, only survivors can be root regions.
477 assert(_next_survivor == NULL, "pre-condition");
478 _next_survivor = _young_list->first_survivor_region();
479 _scan_in_progress = (_next_survivor != NULL);
480 _should_abort = false;
481 }
482
483 HeapRegion* CMRootRegions::claim_next() {
484 if (_should_abort) {
485 // If someone has set the should_abort flag, we return NULL to
486 // force the caller to bail out of their loop.
487 return NULL;
488 }
489
490 // Currently, only survivors can be root regions.
491 HeapRegion* res = _next_survivor;
492 if (res != NULL) {
493 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
494 // Read it again in case it changed while we were waiting for the lock.
495 res = _next_survivor;
496 if (res != NULL) {
497 if (res == _young_list->last_survivor_region()) {
498 // We just claimed the last survivor so store NULL to indicate
499 // that we're done.
500 _next_survivor = NULL;
501 } else {
502 _next_survivor = res->get_next_young_region();
503 }
504 } else {
505 // Someone else claimed the last survivor while we were trying
506 // to take the lock so nothing else to do.
507 }
508 }
509 assert(res == NULL || res->is_survivor(), "post-condition");
510
511 return res;
512 }
513
514 void CMRootRegions::scan_finished() {
515 assert(scan_in_progress(), "pre-condition");
516
517 // Currently, only survivors can be root regions.
518 if (!_should_abort) {
519 assert(_next_survivor == NULL, "we should have claimed all survivors");
520 }
521 _next_survivor = NULL;
522
523 {
524 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
525 _scan_in_progress = false;
526 RootRegionScan_lock->notify_all();
527 }
528 }
529
530 bool CMRootRegions::wait_until_scan_finished() {
531 if (!scan_in_progress()) return false;
532
533 {
534 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
535 while (scan_in_progress()) {
536 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
537 }
538 }
539 return true;
540 }
541
542 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
543 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
544 #endif // _MSC_VER
545
546 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
547 return MAX2((n_par_threads + 2) / 4, 1U);
548 }
549
550 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
551 _g1h(g1h),
552 _markBitMap1(),
553 _markBitMap2(),
554 _parallel_marking_threads(0),
555 _max_parallel_marking_threads(0),
556 _sleep_factor(0.0),
557 _marking_task_overhead(1.0),
558 _cleanup_sleep_factor(0.0),
559 _cleanup_task_overhead(1.0),
560 _cleanup_list("Cleanup List"),
561 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
562 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
563 CardTableModRefBS::card_shift,
564 false /* in_resource_area*/),
565
566 _prevMarkBitMap(&_markBitMap1),
567 _nextMarkBitMap(&_markBitMap2),
568
569 _markStack(this),
570 // _finger set in set_non_marking_state
571
572 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
573 // _active_tasks set in set_non_marking_state
574 // _tasks set inside the constructor
575 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
576 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
577
578 _has_overflown(false),
579 _concurrent(false),
580 _has_aborted(false),
581 _aborted_gc_id(GCId::undefined()),
582 _restart_for_overflow(false),
583 _concurrent_marking_in_progress(false),
584
585 // _verbose_level set below
586
587 _init_times(),
588 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
589 _cleanup_times(),
590 _total_counting_time(0.0),
591 _total_rs_scrub_time(0.0),
592
593 _parallel_workers(NULL),
594
595 _count_card_bitmaps(NULL),
596 _count_marked_bytes(NULL),
597 _completed_initialization(false) {
598 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
599 if (verbose_level < no_verbose) {
600 verbose_level = no_verbose;
601 }
602 if (verbose_level > high_verbose) {
603 verbose_level = high_verbose;
604 }
605 _verbose_level = verbose_level;
606
607 if (verbose_low()) {
608 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
609 "heap end = " PTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
610 }
611
612 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
613 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
614
615 // Create & start a ConcurrentMark thread.
616 _cmThread = new ConcurrentMarkThread(this);
617 assert(cmThread() != NULL, "CM Thread should have been created");
618 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
619 if (_cmThread->osthread() == NULL) {
620 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
621 }
622
623 assert(CGC_lock != NULL, "Where's the CGC_lock?");
624 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
625 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
626
627 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
628 satb_qs.set_buffer_size(G1SATBBufferSize);
629
630 _root_regions.init(_g1h, this);
631
632 if (ConcGCThreads > ParallelGCThreads) {
633 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
634 "than ParallelGCThreads (" UINTX_FORMAT ").",
635 ConcGCThreads, ParallelGCThreads);
636 return;
637 }
638 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
639 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
640 // if both are set
641 _sleep_factor = 0.0;
642 _marking_task_overhead = 1.0;
643 } else if (G1MarkingOverheadPercent > 0) {
644 // We will calculate the number of parallel marking threads based
645 // on a target overhead with respect to the soft real-time goal
646 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
647 double overall_cm_overhead =
648 (double) MaxGCPauseMillis * marking_overhead /
649 (double) GCPauseIntervalMillis;
650 double cpu_ratio = 1.0 / (double) os::processor_count();
651 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
652 double marking_task_overhead =
653 overall_cm_overhead / marking_thread_num *
654 (double) os::processor_count();
655 double sleep_factor =
656 (1.0 - marking_task_overhead) / marking_task_overhead;
657
658 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
659 _sleep_factor = sleep_factor;
660 _marking_task_overhead = marking_task_overhead;
661 } else {
662 // Calculate the number of parallel marking threads by scaling
663 // the number of parallel GC threads.
664 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
665 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
666 _sleep_factor = 0.0;
667 _marking_task_overhead = 1.0;
668 }
669
670 assert(ConcGCThreads > 0, "Should have been set");
671 _parallel_marking_threads = (uint) ConcGCThreads;
672 _max_parallel_marking_threads = _parallel_marking_threads;
673
674 if (parallel_marking_threads() > 1) {
675 _cleanup_task_overhead = 1.0;
676 } else {
677 _cleanup_task_overhead = marking_task_overhead();
678 }
679 _cleanup_sleep_factor =
680 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
681
682 #if 0
683 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
684 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
685 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
686 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
687 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
688 #endif
689
690 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
691 _max_parallel_marking_threads, false, true);
692 if (_parallel_workers == NULL) {
693 vm_exit_during_initialization("Failed necessary allocation.");
694 } else {
695 _parallel_workers->initialize_workers();
696 }
697
698 if (FLAG_IS_DEFAULT(MarkStackSize)) {
699 uintx mark_stack_size =
700 MIN2(MarkStackSizeMax,
701 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
702 // Verify that the calculated value for MarkStackSize is in range.
703 // It would be nice to use the private utility routine from Arguments.
704 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
705 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
706 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
707 mark_stack_size, (uintx) 1, MarkStackSizeMax);
708 return;
709 }
710 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
711 } else {
712 // Verify MarkStackSize is in range.
713 if (FLAG_IS_CMDLINE(MarkStackSize)) {
714 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
715 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
716 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
717 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
718 MarkStackSize, (uintx) 1, MarkStackSizeMax);
719 return;
720 }
721 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
722 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
723 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
724 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
725 MarkStackSize, MarkStackSizeMax);
726 return;
727 }
728 }
729 }
730 }
731
732 if (!_markStack.allocate(MarkStackSize)) {
733 warning("Failed to allocate CM marking stack");
734 return;
735 }
736
737 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
738 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
739
740 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
741 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
742
743 BitMap::idx_t card_bm_size = _card_bm.size();
744
745 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
746 _active_tasks = _max_worker_id;
747
748 size_t max_regions = (size_t) _g1h->max_regions();
749 for (uint i = 0; i < _max_worker_id; ++i) {
750 CMTaskQueue* task_queue = new CMTaskQueue();
751 task_queue->initialize();
752 _task_queues->register_queue(i, task_queue);
753
754 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
755 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
756
757 _tasks[i] = new CMTask(i, this,
758 _count_marked_bytes[i],
759 &_count_card_bitmaps[i],
760 task_queue, _task_queues);
761
762 _accum_task_vtime[i] = 0.0;
763 }
764
765 // Calculate the card number for the bottom of the heap. Used
766 // in biasing indexes into the accounting card bitmaps.
767 _heap_bottom_card_num =
768 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
769 CardTableModRefBS::card_shift);
770
771 // Clear all the liveness counting data
772 clear_all_count_data();
773
774 // so that the call below can read a sensible value
775 _heap_start = g1h->reserved_region().start();
776 set_non_marking_state();
777 _completed_initialization = true;
778 }
779
780 void ConcurrentMark::reset() {
781 // Starting values for these two. This should be called in a STW
782 // phase.
783 MemRegion reserved = _g1h->g1_reserved();
784 _heap_start = reserved.start();
785 _heap_end = reserved.end();
786
787 // Separated the asserts so that we know which one fires.
788 assert(_heap_start != NULL, "heap bounds should look ok");
789 assert(_heap_end != NULL, "heap bounds should look ok");
790 assert(_heap_start < _heap_end, "heap bounds should look ok");
791
792 // Reset all the marking data structures and any necessary flags
793 reset_marking_state();
794
795 if (verbose_low()) {
796 gclog_or_tty->print_cr("[global] resetting");
797 }
798
799 // We do reset all of them, since different phases will use
800 // different number of active threads. So, it's easiest to have all
801 // of them ready.
802 for (uint i = 0; i < _max_worker_id; ++i) {
803 _tasks[i]->reset(_nextMarkBitMap);
804 }
805
806 // we need this to make sure that the flag is on during the evac
807 // pause with initial mark piggy-backed
808 set_concurrent_marking_in_progress();
809 }
810
811
812 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
813 _markStack.set_should_expand();
814 _markStack.setEmpty(); // Also clears the _markStack overflow flag
815 if (clear_overflow) {
816 clear_has_overflown();
817 } else {
818 assert(has_overflown(), "pre-condition");
819 }
820 _finger = _heap_start;
821
822 for (uint i = 0; i < _max_worker_id; ++i) {
823 CMTaskQueue* queue = _task_queues->queue(i);
824 queue->set_empty();
825 }
826 }
827
828 void ConcurrentMark::set_concurrency(uint active_tasks) {
829 assert(active_tasks <= _max_worker_id, "we should not have more");
830
831 _active_tasks = active_tasks;
832 // Need to update the three data structures below according to the
833 // number of active threads for this phase.
834 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
835 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
836 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
837 }
838
839 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
840 set_concurrency(active_tasks);
841
842 _concurrent = concurrent;
843 // We propagate this to all tasks, not just the active ones.
844 for (uint i = 0; i < _max_worker_id; ++i)
845 _tasks[i]->set_concurrent(concurrent);
846
847 if (concurrent) {
848 set_concurrent_marking_in_progress();
849 } else {
850 // We currently assume that the concurrent flag has been set to
851 // false before we start remark. At this point we should also be
852 // in a STW phase.
853 assert(!concurrent_marking_in_progress(), "invariant");
854 assert(out_of_regions(),
855 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
856 p2i(_finger), p2i(_heap_end)));
857 }
858 }
859
860 void ConcurrentMark::set_non_marking_state() {
861 // We set the global marking state to some default values when we're
862 // not doing marking.
863 reset_marking_state();
864 _active_tasks = 0;
865 clear_concurrent_marking_in_progress();
866 }
867
868 ConcurrentMark::~ConcurrentMark() {
869 // The ConcurrentMark instance is never freed.
870 ShouldNotReachHere();
871 }
872
873 void ConcurrentMark::clearNextBitmap() {
874 G1CollectedHeap* g1h = G1CollectedHeap::heap();
875
876 // Make sure that the concurrent mark thread looks to still be in
877 // the current cycle.
878 guarantee(cmThread()->during_cycle(), "invariant");
879
880 // We are finishing up the current cycle by clearing the next
881 // marking bitmap and getting it ready for the next cycle. During
882 // this time no other cycle can start. So, let's make sure that this
883 // is the case.
884 guarantee(!g1h->mark_in_progress(), "invariant");
885
886 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
887 ParClearNextMarkBitmapTask task(&cl, parallel_marking_threads(), true);
888 _parallel_workers->run_task(&task);
889
890 // Clear the liveness counting data. If the marking has been aborted, the abort()
891 // call already did that.
892 if (cl.complete()) {
893 clear_all_count_data();
894 }
895
896 // Repeat the asserts from above.
897 guarantee(cmThread()->during_cycle(), "invariant");
898 guarantee(!g1h->mark_in_progress(), "invariant");
899 }
900
901 class CheckBitmapClearHRClosure : public HeapRegionClosure {
902 CMBitMap* _bitmap;
903 bool _error;
904 public:
905 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
906 }
907
908 virtual bool doHeapRegion(HeapRegion* r) {
909 // This closure can be called concurrently to the mutator, so we must make sure
910 // that the result of the getNextMarkedWordAddress() call is compared to the
911 // value passed to it as limit to detect any found bits.
912 // We can use the region's orig_end() for the limit and the comparison value
913 // as it always contains the "real" end of the region that never changes and
914 // has no side effects.
915 // Due to the latter, there can also be no problem with the compiler generating
916 // reloads of the orig_end() call.
917 HeapWord* end = r->orig_end();
918 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
919 }
920 };
921
922 bool ConcurrentMark::nextMarkBitmapIsClear() {
923 CheckBitmapClearHRClosure cl(_nextMarkBitMap);
924 _g1h->heap_region_iterate(&cl);
925 return cl.complete();
926 }
927
928 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
929 public:
930 bool doHeapRegion(HeapRegion* r) {
931 if (!r->is_continues_humongous()) {
932 r->note_start_of_marking();
933 }
934 return false;
935 }
936 };
937
938 void ConcurrentMark::checkpointRootsInitialPre() {
939 G1CollectedHeap* g1h = G1CollectedHeap::heap();
940 G1CollectorPolicy* g1p = g1h->g1_policy();
941
942 _has_aborted = false;
943
944 #ifndef PRODUCT
945 if (G1PrintReachableAtInitialMark) {
946 print_reachable("at-cycle-start",
947 VerifyOption_G1UsePrevMarking, true /* all */);
948 }
949 #endif
950
951 // Initialize marking structures. This has to be done in a STW phase.
952 reset();
953
954 // For each region note start of marking.
955 NoteStartOfMarkHRClosure startcl;
956 g1h->heap_region_iterate(&startcl);
957 }
958
959
960 void ConcurrentMark::checkpointRootsInitialPost() {
961 G1CollectedHeap* g1h = G1CollectedHeap::heap();
962
963 // If we force an overflow during remark, the remark operation will
964 // actually abort and we'll restart concurrent marking. If we always
965 // force an overflow during remark we'll never actually complete the
966 // marking phase. So, we initialize this here, at the start of the
967 // cycle, so that at the remaining overflow number will decrease at
968 // every remark and we'll eventually not need to cause one.
969 force_overflow_stw()->init();
970
971 // Start Concurrent Marking weak-reference discovery.
972 ReferenceProcessor* rp = g1h->ref_processor_cm();
973 // enable ("weak") refs discovery
974 rp->enable_discovery();
975 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
976
977 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
978 // This is the start of the marking cycle, we're expected all
979 // threads to have SATB queues with active set to false.
980 satb_mq_set.set_active_all_threads(true, /* new active value */
981 false /* expected_active */);
982
983 _root_regions.prepare_for_scan();
984
985 // update_g1_committed() will be called at the end of an evac pause
986 // when marking is on. So, it's also called at the end of the
987 // initial-mark pause to update the heap end, if the heap expands
988 // during it. No need to call it here.
989 }
990
991 /*
992 * Notice that in the next two methods, we actually leave the STS
993 * during the barrier sync and join it immediately afterwards. If we
994 * do not do this, the following deadlock can occur: one thread could
995 * be in the barrier sync code, waiting for the other thread to also
996 * sync up, whereas another one could be trying to yield, while also
997 * waiting for the other threads to sync up too.
998 *
999 * Note, however, that this code is also used during remark and in
1000 * this case we should not attempt to leave / enter the STS, otherwise
1001 * we'll either hit an assert (debug / fastdebug) or deadlock
1002 * (product). So we should only leave / enter the STS if we are
1003 * operating concurrently.
1004 *
1005 * Because the thread that does the sync barrier has left the STS, it
1006 * is possible to be suspended for a Full GC or an evacuation pause
1007 * could occur. This is actually safe, since the entering the sync
1008 * barrier is one of the last things do_marking_step() does, and it
1009 * doesn't manipulate any data structures afterwards.
1010 */
1011
1012 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
1013 if (verbose_low()) {
1014 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
1015 }
1016
1017 if (concurrent()) {
1018 SuspendibleThreadSet::leave();
1019 }
1020
1021 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
1022
1023 if (concurrent()) {
1024 SuspendibleThreadSet::join();
1025 }
1026 // at this point everyone should have synced up and not be doing any
1027 // more work
1028
1029 if (verbose_low()) {
1030 if (barrier_aborted) {
1031 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1032 } else {
1033 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1034 }
1035 }
1036
1037 if (barrier_aborted) {
1038 // If the barrier aborted we ignore the overflow condition and
1039 // just abort the whole marking phase as quickly as possible.
1040 return;
1041 }
1042
1043 // If we're executing the concurrent phase of marking, reset the marking
1044 // state; otherwise the marking state is reset after reference processing,
1045 // during the remark pause.
1046 // If we reset here as a result of an overflow during the remark we will
1047 // see assertion failures from any subsequent set_concurrency_and_phase()
1048 // calls.
1049 if (concurrent()) {
1050 // let the task associated with with worker 0 do this
1051 if (worker_id == 0) {
1052 // task 0 is responsible for clearing the global data structures
1053 // We should be here because of an overflow. During STW we should
1054 // not clear the overflow flag since we rely on it being true when
1055 // we exit this method to abort the pause and restart concurrent
1056 // marking.
1057 reset_marking_state(true /* clear_overflow */);
1058 force_overflow()->update();
1059
1060 if (G1Log::fine()) {
1061 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1062 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1063 }
1064 }
1065 }
1066
1067 // after this, each task should reset its own data structures then
1068 // then go into the second barrier
1069 }
1070
1071 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1072 if (verbose_low()) {
1073 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1074 }
1075
1076 if (concurrent()) {
1077 SuspendibleThreadSet::leave();
1078 }
1079
1080 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1081
1082 if (concurrent()) {
1083 SuspendibleThreadSet::join();
1084 }
1085 // at this point everything should be re-initialized and ready to go
1086
1087 if (verbose_low()) {
1088 if (barrier_aborted) {
1089 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1090 } else {
1091 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1092 }
1093 }
1094 }
1095
1096 #ifndef PRODUCT
1097 void ForceOverflowSettings::init() {
1098 _num_remaining = G1ConcMarkForceOverflow;
1099 _force = false;
1100 update();
1101 }
1102
1103 void ForceOverflowSettings::update() {
1104 if (_num_remaining > 0) {
1105 _num_remaining -= 1;
1106 _force = true;
1107 } else {
1108 _force = false;
1109 }
1110 }
1111
1112 bool ForceOverflowSettings::should_force() {
1113 if (_force) {
1114 _force = false;
1115 return true;
1116 } else {
1117 return false;
1118 }
1119 }
1120 #endif // !PRODUCT
1121
1122 class CMConcurrentMarkingTask: public AbstractGangTask {
1123 private:
1124 ConcurrentMark* _cm;
1125 ConcurrentMarkThread* _cmt;
1126
1127 public:
1128 void work(uint worker_id) {
1129 assert(Thread::current()->is_ConcurrentGC_thread(),
1130 "this should only be done by a conc GC thread");
1131 ResourceMark rm;
1132
1133 double start_vtime = os::elapsedVTime();
1134
1135 SuspendibleThreadSet::join();
1136
1137 assert(worker_id < _cm->active_tasks(), "invariant");
1138 CMTask* the_task = _cm->task(worker_id);
1139 the_task->record_start_time();
1140 if (!_cm->has_aborted()) {
1141 do {
1142 double start_vtime_sec = os::elapsedVTime();
1143 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1144
1145 the_task->do_marking_step(mark_step_duration_ms,
1146 true /* do_termination */,
1147 false /* is_serial*/);
1148
1149 double end_vtime_sec = os::elapsedVTime();
1150 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1151 _cm->clear_has_overflown();
1152
1153 _cm->do_yield_check(worker_id);
1154
1155 jlong sleep_time_ms;
1156 if (!_cm->has_aborted() && the_task->has_aborted()) {
1157 sleep_time_ms =
1158 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1159 SuspendibleThreadSet::leave();
1160 os::sleep(Thread::current(), sleep_time_ms, false);
1161 SuspendibleThreadSet::join();
1162 }
1163 } while (!_cm->has_aborted() && the_task->has_aborted());
1164 }
1165 the_task->record_end_time();
1166 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1167
1168 SuspendibleThreadSet::leave();
1169
1170 double end_vtime = os::elapsedVTime();
1171 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1172 }
1173
1174 CMConcurrentMarkingTask(ConcurrentMark* cm,
1175 ConcurrentMarkThread* cmt) :
1176 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1177
1178 ~CMConcurrentMarkingTask() { }
1179 };
1180
1181 // Calculates the number of active workers for a concurrent
1182 // phase.
1183 uint ConcurrentMark::calc_parallel_marking_threads() {
1184 uint n_conc_workers = 0;
1185 if (!UseDynamicNumberOfGCThreads ||
1186 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1187 !ForceDynamicNumberOfGCThreads)) {
1188 n_conc_workers = max_parallel_marking_threads();
1189 } else {
1190 n_conc_workers =
1191 AdaptiveSizePolicy::calc_default_active_workers(
1192 max_parallel_marking_threads(),
1193 1, /* Minimum workers */
1194 parallel_marking_threads(),
1195 Threads::number_of_non_daemon_threads());
1196 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1197 // that scaling has already gone into "_max_parallel_marking_threads".
1198 }
1199 assert(n_conc_workers > 0, "Always need at least 1");
1200 return n_conc_workers;
1201 }
1202
1203 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1204 // Currently, only survivors can be root regions.
1205 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1206 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1207
1208 const uintx interval = PrefetchScanIntervalInBytes;
1209 HeapWord* curr = hr->bottom();
1210 const HeapWord* end = hr->top();
1211 while (curr < end) {
1212 Prefetch::read(curr, interval);
1213 oop obj = oop(curr);
1214 int size = obj->oop_iterate(&cl);
1215 assert(size == obj->size(), "sanity");
1216 curr += size;
1217 }
1218 }
1219
1220 class CMRootRegionScanTask : public AbstractGangTask {
1221 private:
1222 ConcurrentMark* _cm;
1223
1224 public:
1225 CMRootRegionScanTask(ConcurrentMark* cm) :
1226 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1227
1228 void work(uint worker_id) {
1229 assert(Thread::current()->is_ConcurrentGC_thread(),
1230 "this should only be done by a conc GC thread");
1231
1232 CMRootRegions* root_regions = _cm->root_regions();
1233 HeapRegion* hr = root_regions->claim_next();
1234 while (hr != NULL) {
1235 _cm->scanRootRegion(hr, worker_id);
1236 hr = root_regions->claim_next();
1237 }
1238 }
1239 };
1240
1241 void ConcurrentMark::scanRootRegions() {
1242 // Start of concurrent marking.
1243 ClassLoaderDataGraph::clear_claimed_marks();
1244
1245 // scan_in_progress() will have been set to true only if there was
1246 // at least one root region to scan. So, if it's false, we
1247 // should not attempt to do any further work.
1248 if (root_regions()->scan_in_progress()) {
1249 _parallel_marking_threads = calc_parallel_marking_threads();
1250 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1251 "Maximum number of marking threads exceeded");
1252 uint active_workers = MAX2(1U, parallel_marking_threads());
1253
1254 CMRootRegionScanTask task(this);
1255 _parallel_workers->set_active_workers(active_workers);
1256 _parallel_workers->run_task(&task);
1257
1258 // It's possible that has_aborted() is true here without actually
1259 // aborting the survivor scan earlier. This is OK as it's
1260 // mainly used for sanity checking.
1261 root_regions()->scan_finished();
1262 }
1263 }
1264
1265 void ConcurrentMark::markFromRoots() {
1266 // we might be tempted to assert that:
1267 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1268 // "inconsistent argument?");
1269 // However that wouldn't be right, because it's possible that
1270 // a safepoint is indeed in progress as a younger generation
1271 // stop-the-world GC happens even as we mark in this generation.
1272
1273 _restart_for_overflow = false;
1274 force_overflow_conc()->init();
1275
1276 // _g1h has _n_par_threads
1277 _parallel_marking_threads = calc_parallel_marking_threads();
1278 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1279 "Maximum number of marking threads exceeded");
1280
1281 uint active_workers = MAX2(1U, parallel_marking_threads());
1282
1283 // Parallel task terminator is set in "set_concurrency_and_phase()"
1284 set_concurrency_and_phase(active_workers, true /* concurrent */);
1285
1286 CMConcurrentMarkingTask markingTask(this, cmThread());
1287 _parallel_workers->set_active_workers(active_workers);
1288 // Don't set _n_par_threads because it affects MT in process_roots()
1289 // and the decisions on that MT processing is made elsewhere.
1290 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1291 _parallel_workers->run_task(&markingTask);
1292 print_stats();
1293 }
1294
1295 // Helper class to get rid of some boilerplate code.
1296 class G1CMTraceTime : public GCTraceTime {
1297 static bool doit_and_prepend(bool doit) {
1298 if (doit) {
1299 gclog_or_tty->put(' ');
1300 }
1301 return doit;
1302 }
1303
1304 public:
1305 G1CMTraceTime(const char* title, bool doit)
1306 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
1307 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
1308 }
1309 };
1310
1311 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1312 // world is stopped at this checkpoint
1313 assert(SafepointSynchronize::is_at_safepoint(),
1314 "world should be stopped");
1315
1316 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1317
1318 // If a full collection has happened, we shouldn't do this.
1319 if (has_aborted()) {
1320 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1321 return;
1322 }
1323
1324 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1325
1326 if (VerifyDuringGC) {
1327 HandleMark hm; // handle scope
1328 Universe::heap()->prepare_for_verify();
1329 Universe::verify(VerifyOption_G1UsePrevMarking,
1330 " VerifyDuringGC:(before)");
1331 }
1332 g1h->check_bitmaps("Remark Start");
1333
1334 G1CollectorPolicy* g1p = g1h->g1_policy();
1335 g1p->record_concurrent_mark_remark_start();
1336
1337 double start = os::elapsedTime();
1338
1339 checkpointRootsFinalWork();
1340
1341 double mark_work_end = os::elapsedTime();
1342
1343 weakRefsWork(clear_all_soft_refs);
1344
1345 if (has_overflown()) {
1346 // Oops. We overflowed. Restart concurrent marking.
1347 _restart_for_overflow = true;
1348 if (G1TraceMarkStackOverflow) {
1349 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1350 }
1351
1352 // Verify the heap w.r.t. the previous marking bitmap.
1353 if (VerifyDuringGC) {
1354 HandleMark hm; // handle scope
1355 Universe::heap()->prepare_for_verify();
1356 Universe::verify(VerifyOption_G1UsePrevMarking,
1357 " VerifyDuringGC:(overflow)");
1358 }
1359
1360 // Clear the marking state because we will be restarting
1361 // marking due to overflowing the global mark stack.
1362 reset_marking_state();
1363 } else {
1364 {
1365 G1CMTraceTime trace("GC aggregate-data", G1Log::finer());
1366
1367 // Aggregate the per-task counting data that we have accumulated
1368 // while marking.
1369 aggregate_count_data();
1370 }
1371
1372 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1373 // We're done with marking.
1374 // This is the end of the marking cycle, we're expected all
1375 // threads to have SATB queues with active set to true.
1376 satb_mq_set.set_active_all_threads(false, /* new active value */
1377 true /* expected_active */);
1378
1379 if (VerifyDuringGC) {
1380 HandleMark hm; // handle scope
1381 Universe::heap()->prepare_for_verify();
1382 Universe::verify(VerifyOption_G1UseNextMarking,
1383 " VerifyDuringGC:(after)");
1384 }
1385 g1h->check_bitmaps("Remark End");
1386 assert(!restart_for_overflow(), "sanity");
1387 // Completely reset the marking state since marking completed
1388 set_non_marking_state();
1389 }
1390
1391 // Expand the marking stack, if we have to and if we can.
1392 if (_markStack.should_expand()) {
1393 _markStack.expand();
1394 }
1395
1396 // Statistics
1397 double now = os::elapsedTime();
1398 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1399 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1400 _remark_times.add((now - start) * 1000.0);
1401
1402 g1p->record_concurrent_mark_remark_end();
1403
1404 G1CMIsAliveClosure is_alive(g1h);
1405 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1406 }
1407
1408 // Base class of the closures that finalize and verify the
1409 // liveness counting data.
1410 class CMCountDataClosureBase: public HeapRegionClosure {
1411 protected:
1412 G1CollectedHeap* _g1h;
1413 ConcurrentMark* _cm;
1414 CardTableModRefBS* _ct_bs;
1415
1416 BitMap* _region_bm;
1417 BitMap* _card_bm;
1418
1419 // Takes a region that's not empty (i.e., it has at least one
1420 // live object in it and sets its corresponding bit on the region
1421 // bitmap to 1. If the region is "starts humongous" it will also set
1422 // to 1 the bits on the region bitmap that correspond to its
1423 // associated "continues humongous" regions.
1424 void set_bit_for_region(HeapRegion* hr) {
1425 assert(!hr->is_continues_humongous(), "should have filtered those out");
1426
1427 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1428 if (!hr->is_starts_humongous()) {
1429 // Normal (non-humongous) case: just set the bit.
1430 _region_bm->par_at_put(index, true);
1431 } else {
1432 // Starts humongous case: calculate how many regions are part of
1433 // this humongous region and then set the bit range.
1434 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1435 _region_bm->par_at_put_range(index, end_index, true);
1436 }
1437 }
1438
1439 public:
1440 CMCountDataClosureBase(G1CollectedHeap* g1h,
1441 BitMap* region_bm, BitMap* card_bm):
1442 _g1h(g1h), _cm(g1h->concurrent_mark()),
1443 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1444 _region_bm(region_bm), _card_bm(card_bm) { }
1445 };
1446
1447 // Closure that calculates the # live objects per region. Used
1448 // for verification purposes during the cleanup pause.
1449 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1450 CMBitMapRO* _bm;
1451 size_t _region_marked_bytes;
1452
1453 public:
1454 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1455 BitMap* region_bm, BitMap* card_bm) :
1456 CMCountDataClosureBase(g1h, region_bm, card_bm),
1457 _bm(bm), _region_marked_bytes(0) { }
1458
1459 bool doHeapRegion(HeapRegion* hr) {
1460
1461 if (hr->is_continues_humongous()) {
1462 // We will ignore these here and process them when their
1463 // associated "starts humongous" region is processed (see
1464 // set_bit_for_heap_region()). Note that we cannot rely on their
1465 // associated "starts humongous" region to have their bit set to
1466 // 1 since, due to the region chunking in the parallel region
1467 // iteration, a "continues humongous" region might be visited
1468 // before its associated "starts humongous".
1469 return false;
1470 }
1471
1472 HeapWord* ntams = hr->next_top_at_mark_start();
1473 HeapWord* start = hr->bottom();
1474
1475 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1476 err_msg("Preconditions not met - "
1477 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1478 p2i(start), p2i(ntams), p2i(hr->end())));
1479
1480 // Find the first marked object at or after "start".
1481 start = _bm->getNextMarkedWordAddress(start, ntams);
1482
1483 size_t marked_bytes = 0;
1484
1485 while (start < ntams) {
1486 oop obj = oop(start);
1487 int obj_sz = obj->size();
1488 HeapWord* obj_end = start + obj_sz;
1489
1490 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1491 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1492
1493 // Note: if we're looking at the last region in heap - obj_end
1494 // could be actually just beyond the end of the heap; end_idx
1495 // will then correspond to a (non-existent) card that is also
1496 // just beyond the heap.
1497 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1498 // end of object is not card aligned - increment to cover
1499 // all the cards spanned by the object
1500 end_idx += 1;
1501 }
1502
1503 // Set the bits in the card BM for the cards spanned by this object.
1504 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1505
1506 // Add the size of this object to the number of marked bytes.
1507 marked_bytes += (size_t)obj_sz * HeapWordSize;
1508
1509 // Find the next marked object after this one.
1510 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1511 }
1512
1513 // Mark the allocated-since-marking portion...
1514 HeapWord* top = hr->top();
1515 if (ntams < top) {
1516 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1517 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1518
1519 // Note: if we're looking at the last region in heap - top
1520 // could be actually just beyond the end of the heap; end_idx
1521 // will then correspond to a (non-existent) card that is also
1522 // just beyond the heap.
1523 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1524 // end of object is not card aligned - increment to cover
1525 // all the cards spanned by the object
1526 end_idx += 1;
1527 }
1528 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1529
1530 // This definitely means the region has live objects.
1531 set_bit_for_region(hr);
1532 }
1533
1534 // Update the live region bitmap.
1535 if (marked_bytes > 0) {
1536 set_bit_for_region(hr);
1537 }
1538
1539 // Set the marked bytes for the current region so that
1540 // it can be queried by a calling verification routine
1541 _region_marked_bytes = marked_bytes;
1542
1543 return false;
1544 }
1545
1546 size_t region_marked_bytes() const { return _region_marked_bytes; }
1547 };
1548
1549 // Heap region closure used for verifying the counting data
1550 // that was accumulated concurrently and aggregated during
1551 // the remark pause. This closure is applied to the heap
1552 // regions during the STW cleanup pause.
1553
1554 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1555 G1CollectedHeap* _g1h;
1556 ConcurrentMark* _cm;
1557 CalcLiveObjectsClosure _calc_cl;
1558 BitMap* _region_bm; // Region BM to be verified
1559 BitMap* _card_bm; // Card BM to be verified
1560 bool _verbose; // verbose output?
1561
1562 BitMap* _exp_region_bm; // Expected Region BM values
1563 BitMap* _exp_card_bm; // Expected card BM values
1564
1565 int _failures;
1566
1567 public:
1568 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1569 BitMap* region_bm,
1570 BitMap* card_bm,
1571 BitMap* exp_region_bm,
1572 BitMap* exp_card_bm,
1573 bool verbose) :
1574 _g1h(g1h), _cm(g1h->concurrent_mark()),
1575 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1576 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1577 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1578 _failures(0) { }
1579
1580 int failures() const { return _failures; }
1581
1582 bool doHeapRegion(HeapRegion* hr) {
1583 if (hr->is_continues_humongous()) {
1584 // We will ignore these here and process them when their
1585 // associated "starts humongous" region is processed (see
1586 // set_bit_for_heap_region()). Note that we cannot rely on their
1587 // associated "starts humongous" region to have their bit set to
1588 // 1 since, due to the region chunking in the parallel region
1589 // iteration, a "continues humongous" region might be visited
1590 // before its associated "starts humongous".
1591 return false;
1592 }
1593
1594 int failures = 0;
1595
1596 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1597 // this region and set the corresponding bits in the expected region
1598 // and card bitmaps.
1599 bool res = _calc_cl.doHeapRegion(hr);
1600 assert(res == false, "should be continuing");
1601
1602 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1603 Mutex::_no_safepoint_check_flag);
1604
1605 // Verify the marked bytes for this region.
1606 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1607 size_t act_marked_bytes = hr->next_marked_bytes();
1608
1609 // We're not OK if expected marked bytes > actual marked bytes. It means
1610 // we have missed accounting some objects during the actual marking.
1611 if (exp_marked_bytes > act_marked_bytes) {
1612 if (_verbose) {
1613 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1614 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1615 hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1616 }
1617 failures += 1;
1618 }
1619
1620 // Verify the bit, for this region, in the actual and expected
1621 // (which was just calculated) region bit maps.
1622 // We're not OK if the bit in the calculated expected region
1623 // bitmap is set and the bit in the actual region bitmap is not.
1624 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1625
1626 bool expected = _exp_region_bm->at(index);
1627 bool actual = _region_bm->at(index);
1628 if (expected && !actual) {
1629 if (_verbose) {
1630 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1631 "expected: %s, actual: %s",
1632 hr->hrm_index(),
1633 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1634 }
1635 failures += 1;
1636 }
1637
1638 // Verify that the card bit maps for the cards spanned by the current
1639 // region match. We have an error if we have a set bit in the expected
1640 // bit map and the corresponding bit in the actual bitmap is not set.
1641
1642 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1643 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1644
1645 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1646 expected = _exp_card_bm->at(i);
1647 actual = _card_bm->at(i);
1648
1649 if (expected && !actual) {
1650 if (_verbose) {
1651 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1652 "expected: %s, actual: %s",
1653 hr->hrm_index(), i,
1654 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1655 }
1656 failures += 1;
1657 }
1658 }
1659
1660 if (failures > 0 && _verbose) {
1661 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1662 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1663 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1664 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1665 }
1666
1667 _failures += failures;
1668
1669 // We could stop iteration over the heap when we
1670 // find the first violating region by returning true.
1671 return false;
1672 }
1673 };
1674
1675 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1676 protected:
1677 G1CollectedHeap* _g1h;
1678 ConcurrentMark* _cm;
1679 BitMap* _actual_region_bm;
1680 BitMap* _actual_card_bm;
1681
1682 uint _n_workers;
1683
1684 BitMap* _expected_region_bm;
1685 BitMap* _expected_card_bm;
1686
1687 int _failures;
1688 bool _verbose;
1689
1690 HeapRegionClaimer _hrclaimer;
1691
1692 public:
1693 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1694 BitMap* region_bm, BitMap* card_bm,
1695 BitMap* expected_region_bm, BitMap* expected_card_bm)
1696 : AbstractGangTask("G1 verify final counting"),
1697 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1698 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1699 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1700 _failures(0), _verbose(false),
1701 _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1702 assert(VerifyDuringGC, "don't call this otherwise");
1703 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1704 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1705
1706 _verbose = _cm->verbose_medium();
1707 }
1708
1709 void work(uint worker_id) {
1710 assert(worker_id < _n_workers, "invariant");
1711
1712 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1713 _actual_region_bm, _actual_card_bm,
1714 _expected_region_bm,
1715 _expected_card_bm,
1716 _verbose);
1717
1718 _g1h->heap_region_par_iterate(&verify_cl, worker_id, &_hrclaimer);
1719
1720 Atomic::add(verify_cl.failures(), &_failures);
1721 }
1722
1723 int failures() const { return _failures; }
1724 };
1725
1726 // Closure that finalizes the liveness counting data.
1727 // Used during the cleanup pause.
1728 // Sets the bits corresponding to the interval [NTAMS, top]
1729 // (which contains the implicitly live objects) in the
1730 // card liveness bitmap. Also sets the bit for each region,
1731 // containing live data, in the region liveness bitmap.
1732
1733 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1734 public:
1735 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1736 BitMap* region_bm,
1737 BitMap* card_bm) :
1738 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1739
1740 bool doHeapRegion(HeapRegion* hr) {
1741
1742 if (hr->is_continues_humongous()) {
1743 // We will ignore these here and process them when their
1744 // associated "starts humongous" region is processed (see
1745 // set_bit_for_heap_region()). Note that we cannot rely on their
1746 // associated "starts humongous" region to have their bit set to
1747 // 1 since, due to the region chunking in the parallel region
1748 // iteration, a "continues humongous" region might be visited
1749 // before its associated "starts humongous".
1750 return false;
1751 }
1752
1753 HeapWord* ntams = hr->next_top_at_mark_start();
1754 HeapWord* top = hr->top();
1755
1756 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1757
1758 // Mark the allocated-since-marking portion...
1759 if (ntams < top) {
1760 // This definitely means the region has live objects.
1761 set_bit_for_region(hr);
1762
1763 // Now set the bits in the card bitmap for [ntams, top)
1764 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1765 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1766
1767 // Note: if we're looking at the last region in heap - top
1768 // could be actually just beyond the end of the heap; end_idx
1769 // will then correspond to a (non-existent) card that is also
1770 // just beyond the heap.
1771 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1772 // end of object is not card aligned - increment to cover
1773 // all the cards spanned by the object
1774 end_idx += 1;
1775 }
1776
1777 assert(end_idx <= _card_bm->size(),
1778 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1779 end_idx, _card_bm->size()));
1780 assert(start_idx < _card_bm->size(),
1781 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1782 start_idx, _card_bm->size()));
1783
1784 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1785 }
1786
1787 // Set the bit for the region if it contains live data
1788 if (hr->next_marked_bytes() > 0) {
1789 set_bit_for_region(hr);
1790 }
1791
1792 return false;
1793 }
1794 };
1795
1796 class G1ParFinalCountTask: public AbstractGangTask {
1797 protected:
1798 G1CollectedHeap* _g1h;
1799 ConcurrentMark* _cm;
1800 BitMap* _actual_region_bm;
1801 BitMap* _actual_card_bm;
1802
1803 uint _n_workers;
1804 HeapRegionClaimer _hrclaimer;
1805
1806 public:
1807 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1808 : AbstractGangTask("G1 final counting"),
1809 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1810 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1811 _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1812 }
1813
1814 void work(uint worker_id) {
1815 assert(worker_id < _n_workers, "invariant");
1816
1817 FinalCountDataUpdateClosure final_update_cl(_g1h,
1818 _actual_region_bm,
1819 _actual_card_bm);
1820
1821 _g1h->heap_region_par_iterate(&final_update_cl, worker_id, &_hrclaimer);
1822 }
1823 };
1824
1825 class G1ParNoteEndTask;
1826
1827 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1828 G1CollectedHeap* _g1;
1829 size_t _max_live_bytes;
1830 uint _regions_claimed;
1831 size_t _freed_bytes;
1832 FreeRegionList* _local_cleanup_list;
1833 HeapRegionSetCount _old_regions_removed;
1834 HeapRegionSetCount _humongous_regions_removed;
1835 HRRSCleanupTask* _hrrs_cleanup_task;
1836 double _claimed_region_time;
1837 double _max_region_time;
1838
1839 public:
1840 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1841 FreeRegionList* local_cleanup_list,
1842 HRRSCleanupTask* hrrs_cleanup_task) :
1843 _g1(g1),
1844 _max_live_bytes(0), _regions_claimed(0),
1845 _freed_bytes(0),
1846 _claimed_region_time(0.0), _max_region_time(0.0),
1847 _local_cleanup_list(local_cleanup_list),
1848 _old_regions_removed(),
1849 _humongous_regions_removed(),
1850 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1851
1852 size_t freed_bytes() { return _freed_bytes; }
1853 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1854 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1855
1856 bool doHeapRegion(HeapRegion *hr) {
1857 if (hr->is_continues_humongous()) {
1858 return false;
1859 }
1860 // We use a claim value of zero here because all regions
1861 // were claimed with value 1 in the FinalCount task.
1862 _g1->reset_gc_time_stamps(hr);
1863 double start = os::elapsedTime();
1864 _regions_claimed++;
1865 hr->note_end_of_marking();
1866 _max_live_bytes += hr->max_live_bytes();
1867
1868 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1869 _freed_bytes += hr->used();
1870 hr->set_containing_set(NULL);
1871 if (hr->is_humongous()) {
1872 assert(hr->is_starts_humongous(), "we should only see starts humongous");
1873 _humongous_regions_removed.increment(1u, hr->capacity());
1874 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1875 } else {
1876 _old_regions_removed.increment(1u, hr->capacity());
1877 _g1->free_region(hr, _local_cleanup_list, true);
1878 }
1879 } else {
1880 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1881 }
1882
1883 double region_time = (os::elapsedTime() - start);
1884 _claimed_region_time += region_time;
1885 if (region_time > _max_region_time) {
1886 _max_region_time = region_time;
1887 }
1888 return false;
1889 }
1890
1891 size_t max_live_bytes() { return _max_live_bytes; }
1892 uint regions_claimed() { return _regions_claimed; }
1893 double claimed_region_time_sec() { return _claimed_region_time; }
1894 double max_region_time_sec() { return _max_region_time; }
1895 };
1896
1897 class G1ParNoteEndTask: public AbstractGangTask {
1898 friend class G1NoteEndOfConcMarkClosure;
1899
1900 protected:
1901 G1CollectedHeap* _g1h;
1902 size_t _max_live_bytes;
1903 size_t _freed_bytes;
1904 FreeRegionList* _cleanup_list;
1905 HeapRegionClaimer _hrclaimer;
1906
1907 public:
1908 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1909 AbstractGangTask("G1 note end"), _g1h(g1h), _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1910 }
1911
1912 void work(uint worker_id) {
1913 double start = os::elapsedTime();
1914 FreeRegionList local_cleanup_list("Local Cleanup List");
1915 HRRSCleanupTask hrrs_cleanup_task;
1916 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1917 &hrrs_cleanup_task);
1918 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1919 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1920
1921 // Now update the lists
1922 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1923 {
1924 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1925 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1926 _max_live_bytes += g1_note_end.max_live_bytes();
1927 _freed_bytes += g1_note_end.freed_bytes();
1928
1929 // If we iterate over the global cleanup list at the end of
1930 // cleanup to do this printing we will not guarantee to only
1931 // generate output for the newly-reclaimed regions (the list
1932 // might not be empty at the beginning of cleanup; we might
1933 // still be working on its previous contents). So we do the
1934 // printing here, before we append the new regions to the global
1935 // cleanup list.
1936
1937 G1HRPrinter* hr_printer = _g1h->hr_printer();
1938 if (hr_printer->is_active()) {
1939 FreeRegionListIterator iter(&local_cleanup_list);
1940 while (iter.more_available()) {
1941 HeapRegion* hr = iter.get_next();
1942 hr_printer->cleanup(hr);
1943 }
1944 }
1945
1946 _cleanup_list->add_ordered(&local_cleanup_list);
1947 assert(local_cleanup_list.is_empty(), "post-condition");
1948
1949 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1950 }
1951 }
1952 size_t max_live_bytes() { return _max_live_bytes; }
1953 size_t freed_bytes() { return _freed_bytes; }
1954 };
1955
1956 class G1ParScrubRemSetTask: public AbstractGangTask {
1957 protected:
1958 G1RemSet* _g1rs;
1959 BitMap* _region_bm;
1960 BitMap* _card_bm;
1961 HeapRegionClaimer _hrclaimer;
1962
1963 public:
1964 G1ParScrubRemSetTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm, uint n_workers) :
1965 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), _region_bm(region_bm), _card_bm(card_bm), _hrclaimer(n_workers) {
1966 }
1967
1968 void work(uint worker_id) {
1969 _g1rs->scrub(_region_bm, _card_bm, worker_id, &_hrclaimer);
1970 }
1971
1972 };
1973
1974 void ConcurrentMark::cleanup() {
1975 // world is stopped at this checkpoint
1976 assert(SafepointSynchronize::is_at_safepoint(),
1977 "world should be stopped");
1978 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1979
1980 // If a full collection has happened, we shouldn't do this.
1981 if (has_aborted()) {
1982 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1983 return;
1984 }
1985
1986 g1h->verify_region_sets_optional();
1987
1988 if (VerifyDuringGC) {
1989 HandleMark hm; // handle scope
1990 Universe::heap()->prepare_for_verify();
1991 Universe::verify(VerifyOption_G1UsePrevMarking,
1992 " VerifyDuringGC:(before)");
1993 }
1994 g1h->check_bitmaps("Cleanup Start");
1995
1996 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
1997 g1p->record_concurrent_mark_cleanup_start();
1998
1999 double start = os::elapsedTime();
2000
2001 HeapRegionRemSet::reset_for_cleanup_tasks();
2002
2003 uint n_workers;
2004
2005 // Do counting once more with the world stopped for good measure.
2006 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2007
2008 g1h->set_par_threads();
2009 n_workers = g1h->n_par_threads();
2010 assert(g1h->n_par_threads() == n_workers,
2011 "Should not have been reset");
2012 g1h->workers()->run_task(&g1_par_count_task);
2013 // Done with the parallel phase so reset to 0.
2014 g1h->set_par_threads(0);
2015
2016 if (VerifyDuringGC) {
2017 // Verify that the counting data accumulated during marking matches
2018 // that calculated by walking the marking bitmap.
2019
2020 // Bitmaps to hold expected values
2021 BitMap expected_region_bm(_region_bm.size(), true);
2022 BitMap expected_card_bm(_card_bm.size(), true);
2023
2024 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2025 &_region_bm,
2026 &_card_bm,
2027 &expected_region_bm,
2028 &expected_card_bm);
2029
2030 g1h->set_par_threads((int)n_workers);
2031 g1h->workers()->run_task(&g1_par_verify_task);
2032 // Done with the parallel phase so reset to 0.
2033 g1h->set_par_threads(0);
2034
2035 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2036 }
2037
2038 size_t start_used_bytes = g1h->used();
2039 g1h->set_marking_complete();
2040
2041 double count_end = os::elapsedTime();
2042 double this_final_counting_time = (count_end - start);
2043 _total_counting_time += this_final_counting_time;
2044
2045 if (G1PrintRegionLivenessInfo) {
2046 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2047 _g1h->heap_region_iterate(&cl);
2048 }
2049
2050 // Install newly created mark bitMap as "prev".
2051 swapMarkBitMaps();
2052
2053 g1h->reset_gc_time_stamp();
2054
2055 // Note end of marking in all heap regions.
2056 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
2057 g1h->set_par_threads((int)n_workers);
2058 g1h->workers()->run_task(&g1_par_note_end_task);
2059 g1h->set_par_threads(0);
2060 g1h->check_gc_time_stamps();
2061
2062 if (!cleanup_list_is_empty()) {
2063 // The cleanup list is not empty, so we'll have to process it
2064 // concurrently. Notify anyone else that might be wanting free
2065 // regions that there will be more free regions coming soon.
2066 g1h->set_free_regions_coming();
2067 }
2068
2069 // call below, since it affects the metric by which we sort the heap
2070 // regions.
2071 if (G1ScrubRemSets) {
2072 double rs_scrub_start = os::elapsedTime();
2073 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm, n_workers);
2074 g1h->set_par_threads((int)n_workers);
2075 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2076 g1h->set_par_threads(0);
2077
2078 double rs_scrub_end = os::elapsedTime();
2079 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2080 _total_rs_scrub_time += this_rs_scrub_time;
2081 }
2082
2083 // this will also free any regions totally full of garbage objects,
2084 // and sort the regions.
2085 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2086
2087 // Statistics.
2088 double end = os::elapsedTime();
2089 _cleanup_times.add((end - start) * 1000.0);
2090
2091 if (G1Log::fine()) {
2092 g1h->print_size_transition(gclog_or_tty,
2093 start_used_bytes,
2094 g1h->used(),
2095 g1h->capacity());
2096 }
2097
2098 // Clean up will have freed any regions completely full of garbage.
2099 // Update the soft reference policy with the new heap occupancy.
2100 Universe::update_heap_info_at_gc();
2101
2102 if (VerifyDuringGC) {
2103 HandleMark hm; // handle scope
2104 Universe::heap()->prepare_for_verify();
2105 Universe::verify(VerifyOption_G1UsePrevMarking,
2106 " VerifyDuringGC:(after)");
2107 }
2108
2109 g1h->check_bitmaps("Cleanup End");
2110
2111 g1h->verify_region_sets_optional();
2112
2113 // We need to make this be a "collection" so any collection pause that
2114 // races with it goes around and waits for completeCleanup to finish.
2115 g1h->increment_total_collections();
2116
2117 // Clean out dead classes and update Metaspace sizes.
2118 if (ClassUnloadingWithConcurrentMark) {
2119 ClassLoaderDataGraph::purge();
2120 }
2121 MetaspaceGC::compute_new_size();
2122
2123 // We reclaimed old regions so we should calculate the sizes to make
2124 // sure we update the old gen/space data.
2125 g1h->g1mm()->update_sizes();
2126 g1h->allocation_context_stats().update_after_mark();
2127
2128 g1h->trace_heap_after_concurrent_cycle();
2129 }
2130
2131 void ConcurrentMark::completeCleanup() {
2132 if (has_aborted()) return;
2133
2134 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2135
2136 _cleanup_list.verify_optional();
2137 FreeRegionList tmp_free_list("Tmp Free List");
2138
2139 if (G1ConcRegionFreeingVerbose) {
2140 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2141 "cleanup list has %u entries",
2142 _cleanup_list.length());
2143 }
2144
2145 // No one else should be accessing the _cleanup_list at this point,
2146 // so it is not necessary to take any locks
2147 while (!_cleanup_list.is_empty()) {
2148 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2149 assert(hr != NULL, "Got NULL from a non-empty list");
2150 hr->par_clear();
2151 tmp_free_list.add_ordered(hr);
2152
2153 // Instead of adding one region at a time to the secondary_free_list,
2154 // we accumulate them in the local list and move them a few at a
2155 // time. This also cuts down on the number of notify_all() calls
2156 // we do during this process. We'll also append the local list when
2157 // _cleanup_list is empty (which means we just removed the last
2158 // region from the _cleanup_list).
2159 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2160 _cleanup_list.is_empty()) {
2161 if (G1ConcRegionFreeingVerbose) {
2162 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2163 "appending %u entries to the secondary_free_list, "
2164 "cleanup list still has %u entries",
2165 tmp_free_list.length(),
2166 _cleanup_list.length());
2167 }
2168
2169 {
2170 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2171 g1h->secondary_free_list_add(&tmp_free_list);
2172 SecondaryFreeList_lock->notify_all();
2173 }
2174
2175 if (G1StressConcRegionFreeing) {
2176 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2177 os::sleep(Thread::current(), (jlong) 1, false);
2178 }
2179 }
2180 }
2181 }
2182 assert(tmp_free_list.is_empty(), "post-condition");
2183 }
2184
2185 // Supporting Object and Oop closures for reference discovery
2186 // and processing in during marking
2187
2188 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2189 HeapWord* addr = (HeapWord*)obj;
2190 return addr != NULL &&
2191 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2192 }
2193
2194 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2195 // Uses the CMTask associated with a worker thread (for serial reference
2196 // processing the CMTask for worker 0 is used) to preserve (mark) and
2197 // trace referent objects.
2198 //
2199 // Using the CMTask and embedded local queues avoids having the worker
2200 // threads operating on the global mark stack. This reduces the risk
2201 // of overflowing the stack - which we would rather avoid at this late
2202 // state. Also using the tasks' local queues removes the potential
2203 // of the workers interfering with each other that could occur if
2204 // operating on the global stack.
2205
2206 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2207 ConcurrentMark* _cm;
2208 CMTask* _task;
2209 int _ref_counter_limit;
2210 int _ref_counter;
2211 bool _is_serial;
2212 public:
2213 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2214 _cm(cm), _task(task), _is_serial(is_serial),
2215 _ref_counter_limit(G1RefProcDrainInterval) {
2216 assert(_ref_counter_limit > 0, "sanity");
2217 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2218 _ref_counter = _ref_counter_limit;
2219 }
2220
2221 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2222 virtual void do_oop( oop* p) { do_oop_work(p); }
2223
2224 template <class T> void do_oop_work(T* p) {
2225 if (!_cm->has_overflown()) {
2226 oop obj = oopDesc::load_decode_heap_oop(p);
2227 if (_cm->verbose_high()) {
2228 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2229 "*"PTR_FORMAT" = "PTR_FORMAT,
2230 _task->worker_id(), p2i(p), p2i((void*) obj));
2231 }
2232
2233 _task->deal_with_reference(obj);
2234 _ref_counter--;
2235
2236 if (_ref_counter == 0) {
2237 // We have dealt with _ref_counter_limit references, pushing them
2238 // and objects reachable from them on to the local stack (and
2239 // possibly the global stack). Call CMTask::do_marking_step() to
2240 // process these entries.
2241 //
2242 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2243 // there's nothing more to do (i.e. we're done with the entries that
2244 // were pushed as a result of the CMTask::deal_with_reference() calls
2245 // above) or we overflow.
2246 //
2247 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2248 // flag while there may still be some work to do. (See the comment at
2249 // the beginning of CMTask::do_marking_step() for those conditions -
2250 // one of which is reaching the specified time target.) It is only
2251 // when CMTask::do_marking_step() returns without setting the
2252 // has_aborted() flag that the marking step has completed.
2253 do {
2254 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2255 _task->do_marking_step(mark_step_duration_ms,
2256 false /* do_termination */,
2257 _is_serial);
2258 } while (_task->has_aborted() && !_cm->has_overflown());
2259 _ref_counter = _ref_counter_limit;
2260 }
2261 } else {
2262 if (_cm->verbose_high()) {
2263 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2264 }
2265 }
2266 }
2267 };
2268
2269 // 'Drain' oop closure used by both serial and parallel reference processing.
2270 // Uses the CMTask associated with a given worker thread (for serial
2271 // reference processing the CMtask for worker 0 is used). Calls the
2272 // do_marking_step routine, with an unbelievably large timeout value,
2273 // to drain the marking data structures of the remaining entries
2274 // added by the 'keep alive' oop closure above.
2275
2276 class G1CMDrainMarkingStackClosure: public VoidClosure {
2277 ConcurrentMark* _cm;
2278 CMTask* _task;
2279 bool _is_serial;
2280 public:
2281 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2282 _cm(cm), _task(task), _is_serial(is_serial) {
2283 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2284 }
2285
2286 void do_void() {
2287 do {
2288 if (_cm->verbose_high()) {
2289 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2290 _task->worker_id(), BOOL_TO_STR(_is_serial));
2291 }
2292
2293 // We call CMTask::do_marking_step() to completely drain the local
2294 // and global marking stacks of entries pushed by the 'keep alive'
2295 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2296 //
2297 // CMTask::do_marking_step() is called in a loop, which we'll exit
2298 // if there's nothing more to do (i.e. we've completely drained the
2299 // entries that were pushed as a a result of applying the 'keep alive'
2300 // closure to the entries on the discovered ref lists) or we overflow
2301 // the global marking stack.
2302 //
2303 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2304 // flag while there may still be some work to do. (See the comment at
2305 // the beginning of CMTask::do_marking_step() for those conditions -
2306 // one of which is reaching the specified time target.) It is only
2307 // when CMTask::do_marking_step() returns without setting the
2308 // has_aborted() flag that the marking step has completed.
2309
2310 _task->do_marking_step(1000000000.0 /* something very large */,
2311 true /* do_termination */,
2312 _is_serial);
2313 } while (_task->has_aborted() && !_cm->has_overflown());
2314 }
2315 };
2316
2317 // Implementation of AbstractRefProcTaskExecutor for parallel
2318 // reference processing at the end of G1 concurrent marking
2319
2320 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2321 private:
2322 G1CollectedHeap* _g1h;
2323 ConcurrentMark* _cm;
2324 WorkGang* _workers;
2325 int _active_workers;
2326
2327 public:
2328 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2329 ConcurrentMark* cm,
2330 WorkGang* workers,
2331 int n_workers) :
2332 _g1h(g1h), _cm(cm),
2333 _workers(workers), _active_workers(n_workers) { }
2334
2335 // Executes the given task using concurrent marking worker threads.
2336 virtual void execute(ProcessTask& task);
2337 virtual void execute(EnqueueTask& task);
2338 };
2339
2340 class G1CMRefProcTaskProxy: public AbstractGangTask {
2341 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2342 ProcessTask& _proc_task;
2343 G1CollectedHeap* _g1h;
2344 ConcurrentMark* _cm;
2345
2346 public:
2347 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2348 G1CollectedHeap* g1h,
2349 ConcurrentMark* cm) :
2350 AbstractGangTask("Process reference objects in parallel"),
2351 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2352 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2353 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2354 }
2355
2356 virtual void work(uint worker_id) {
2357 ResourceMark rm;
2358 HandleMark hm;
2359 CMTask* task = _cm->task(worker_id);
2360 G1CMIsAliveClosure g1_is_alive(_g1h);
2361 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2362 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2363
2364 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2365 }
2366 };
2367
2368 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2369 assert(_workers != NULL, "Need parallel worker threads.");
2370 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2371
2372 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2373
2374 // We need to reset the concurrency level before each
2375 // proxy task execution, so that the termination protocol
2376 // and overflow handling in CMTask::do_marking_step() knows
2377 // how many workers to wait for.
2378 _cm->set_concurrency(_active_workers);
2379 _g1h->set_par_threads(_active_workers);
2380 _workers->run_task(&proc_task_proxy);
2381 _g1h->set_par_threads(0);
2382 }
2383
2384 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2385 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2386 EnqueueTask& _enq_task;
2387
2388 public:
2389 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2390 AbstractGangTask("Enqueue reference objects in parallel"),
2391 _enq_task(enq_task) { }
2392
2393 virtual void work(uint worker_id) {
2394 _enq_task.work(worker_id);
2395 }
2396 };
2397
2398 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2399 assert(_workers != NULL, "Need parallel worker threads.");
2400 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2401
2402 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2403
2404 // Not strictly necessary but...
2405 //
2406 // We need to reset the concurrency level before each
2407 // proxy task execution, so that the termination protocol
2408 // and overflow handling in CMTask::do_marking_step() knows
2409 // how many workers to wait for.
2410 _cm->set_concurrency(_active_workers);
2411 _g1h->set_par_threads(_active_workers);
2412 _workers->run_task(&enq_task_proxy);
2413 _g1h->set_par_threads(0);
2414 }
2415
2416 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2417 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2418 }
2419
2420 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2421 if (has_overflown()) {
2422 // Skip processing the discovered references if we have
2423 // overflown the global marking stack. Reference objects
2424 // only get discovered once so it is OK to not
2425 // de-populate the discovered reference lists. We could have,
2426 // but the only benefit would be that, when marking restarts,
2427 // less reference objects are discovered.
2428 return;
2429 }
2430
2431 ResourceMark rm;
2432 HandleMark hm;
2433
2434 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2435
2436 // Is alive closure.
2437 G1CMIsAliveClosure g1_is_alive(g1h);
2438
2439 // Inner scope to exclude the cleaning of the string and symbol
2440 // tables from the displayed time.
2441 {
2442 G1CMTraceTime t("GC ref-proc", G1Log::finer());
2443
2444 ReferenceProcessor* rp = g1h->ref_processor_cm();
2445
2446 // See the comment in G1CollectedHeap::ref_processing_init()
2447 // about how reference processing currently works in G1.
2448
2449 // Set the soft reference policy
2450 rp->setup_policy(clear_all_soft_refs);
2451 assert(_markStack.isEmpty(), "mark stack should be empty");
2452
2453 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2454 // in serial reference processing. Note these closures are also
2455 // used for serially processing (by the the current thread) the
2456 // JNI references during parallel reference processing.
2457 //
2458 // These closures do not need to synchronize with the worker
2459 // threads involved in parallel reference processing as these
2460 // instances are executed serially by the current thread (e.g.
2461 // reference processing is not multi-threaded and is thus
2462 // performed by the current thread instead of a gang worker).
2463 //
2464 // The gang tasks involved in parallel reference processing create
2465 // their own instances of these closures, which do their own
2466 // synchronization among themselves.
2467 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2468 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2469
2470 // We need at least one active thread. If reference processing
2471 // is not multi-threaded we use the current (VMThread) thread,
2472 // otherwise we use the work gang from the G1CollectedHeap and
2473 // we utilize all the worker threads we can.
2474 bool processing_is_mt = rp->processing_is_mt();
2475 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2476 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2477
2478 // Parallel processing task executor.
2479 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2480 g1h->workers(), active_workers);
2481 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2482
2483 // Set the concurrency level. The phase was already set prior to
2484 // executing the remark task.
2485 set_concurrency(active_workers);
2486
2487 // Set the degree of MT processing here. If the discovery was done MT,
2488 // the number of threads involved during discovery could differ from
2489 // the number of active workers. This is OK as long as the discovered
2490 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2491 rp->set_active_mt_degree(active_workers);
2492
2493 // Process the weak references.
2494 const ReferenceProcessorStats& stats =
2495 rp->process_discovered_references(&g1_is_alive,
2496 &g1_keep_alive,
2497 &g1_drain_mark_stack,
2498 executor,
2499 g1h->gc_timer_cm(),
2500 concurrent_gc_id());
2501 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2502
2503 // The do_oop work routines of the keep_alive and drain_marking_stack
2504 // oop closures will set the has_overflown flag if we overflow the
2505 // global marking stack.
2506
2507 assert(_markStack.overflow() || _markStack.isEmpty(),
2508 "mark stack should be empty (unless it overflowed)");
2509
2510 if (_markStack.overflow()) {
2511 // This should have been done already when we tried to push an
2512 // entry on to the global mark stack. But let's do it again.
2513 set_has_overflown();
2514 }
2515
2516 assert(rp->num_q() == active_workers, "why not");
2517
2518 rp->enqueue_discovered_references(executor);
2519
2520 rp->verify_no_references_recorded();
2521 assert(!rp->discovery_enabled(), "Post condition");
2522 }
2523
2524 if (has_overflown()) {
2525 // We can not trust g1_is_alive if the marking stack overflowed
2526 return;
2527 }
2528
2529 assert(_markStack.isEmpty(), "Marking should have completed");
2530
2531 // Unload Klasses, String, Symbols, Code Cache, etc.
2532 {
2533 G1CMTraceTime trace("Unloading", G1Log::finer());
2534
2535 if (ClassUnloadingWithConcurrentMark) {
2536 // Cleaning of klasses depends on correct information from MetadataMarkOnStack. The CodeCache::mark_on_stack
2537 // part is too slow to be done serially, so it is handled during the weakRefsWorkParallelPart phase.
2538 // Defer the cleaning until we have complete on_stack data.
2539 MetadataOnStackMark md_on_stack(false /* Don't visit the code cache at this point */);
2540
2541 bool purged_classes;
2542
2543 {
2544 G1CMTraceTime trace("System Dictionary Unloading", G1Log::finest());
2545 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2546 }
2547
2548 {
2549 G1CMTraceTime trace("Parallel Unloading", G1Log::finest());
2550 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2551 }
2552
2553 {
2554 G1CMTraceTime trace("Deallocate Metadata", G1Log::finest());
2555 ClassLoaderDataGraph::free_deallocate_lists();
2556 }
2557 }
2558
2559 if (G1StringDedup::is_enabled()) {
2560 G1CMTraceTime trace("String Deduplication Unlink", G1Log::finest());
2561 G1StringDedup::unlink(&g1_is_alive);
2562 }
2563 }
2564 }
2565
2566 void ConcurrentMark::swapMarkBitMaps() {
2567 CMBitMapRO* temp = _prevMarkBitMap;
2568 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2569 _nextMarkBitMap = (CMBitMap*) temp;
2570 }
2571
2572 class CMObjectClosure;
2573
2574 // Closure for iterating over objects, currently only used for
2575 // processing SATB buffers.
2576 class CMObjectClosure : public ObjectClosure {
2577 private:
2578 CMTask* _task;
2579
2580 public:
2581 void do_object(oop obj) {
2582 _task->deal_with_reference(obj);
2583 }
2584
2585 CMObjectClosure(CMTask* task) : _task(task) { }
2586 };
2587
2588 class G1RemarkThreadsClosure : public ThreadClosure {
2589 CMObjectClosure _cm_obj;
2590 G1CMOopClosure _cm_cl;
2591 MarkingCodeBlobClosure _code_cl;
2592 int _thread_parity;
2593
2594 public:
2595 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task) :
2596 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2597 _thread_parity(SharedHeap::heap()->strong_roots_parity()) {}
2598
2599 void do_thread(Thread* thread) {
2600 if (thread->is_Java_thread()) {
2601 if (thread->claim_oops_do(true, _thread_parity)) {
2602 JavaThread* jt = (JavaThread*)thread;
2603
2604 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2605 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2606 // * Alive if on the stack of an executing method
2607 // * Weakly reachable otherwise
2608 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2609 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2610 jt->nmethods_do(&_code_cl);
2611
2612 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2613 }
2614 } else if (thread->is_VM_thread()) {
2615 if (thread->claim_oops_do(true, _thread_parity)) {
2616 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2617 }
2618 }
2619 }
2620 };
2621
2622 class CMRemarkTask: public AbstractGangTask {
2623 private:
2624 ConcurrentMark* _cm;
2625 public:
2626 void work(uint worker_id) {
2627 // Since all available tasks are actually started, we should
2628 // only proceed if we're supposed to be active.
2629 if (worker_id < _cm->active_tasks()) {
2630 CMTask* task = _cm->task(worker_id);
2631 task->record_start_time();
2632 {
2633 ResourceMark rm;
2634 HandleMark hm;
2635
2636 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
2637 Threads::threads_do(&threads_f);
2638 }
2639
2640 do {
2641 task->do_marking_step(1000000000.0 /* something very large */,
2642 true /* do_termination */,
2643 false /* is_serial */);
2644 } while (task->has_aborted() && !_cm->has_overflown());
2645 // If we overflow, then we do not want to restart. We instead
2646 // want to abort remark and do concurrent marking again.
2647 task->record_end_time();
2648 }
2649 }
2650
2651 CMRemarkTask(ConcurrentMark* cm, int active_workers) :
2652 AbstractGangTask("Par Remark"), _cm(cm) {
2653 _cm->terminator()->reset_for_reuse(active_workers);
2654 }
2655 };
2656
2657 void ConcurrentMark::checkpointRootsFinalWork() {
2658 ResourceMark rm;
2659 HandleMark hm;
2660 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2661
2662 G1CMTraceTime trace("Finalize Marking", G1Log::finer());
2663
2664 g1h->ensure_parsability(false);
2665
2666 G1CollectedHeap::StrongRootsScope srs(g1h);
2667 // this is remark, so we'll use up all active threads
2668 uint active_workers = g1h->workers()->active_workers();
2669 if (active_workers == 0) {
2670 assert(active_workers > 0, "Should have been set earlier");
2671 active_workers = (uint) ParallelGCThreads;
2672 g1h->workers()->set_active_workers(active_workers);
2673 }
2674 set_concurrency_and_phase(active_workers, false /* concurrent */);
2675 // Leave _parallel_marking_threads at it's
2676 // value originally calculated in the ConcurrentMark
2677 // constructor and pass values of the active workers
2678 // through the gang in the task.
2679
2680 CMRemarkTask remarkTask(this, active_workers);
2681 // We will start all available threads, even if we decide that the
2682 // active_workers will be fewer. The extra ones will just bail out
2683 // immediately.
2684 g1h->set_par_threads(active_workers);
2685 g1h->workers()->run_task(&remarkTask);
2686 g1h->set_par_threads(0);
2687
2688 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2689 guarantee(has_overflown() ||
2690 satb_mq_set.completed_buffers_num() == 0,
2691 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2692 BOOL_TO_STR(has_overflown()),
2693 satb_mq_set.completed_buffers_num()));
2694
2695 print_stats();
2696 }
2697
2698 #ifndef PRODUCT
2699
2700 class PrintReachableOopClosure: public OopClosure {
2701 private:
2702 G1CollectedHeap* _g1h;
2703 outputStream* _out;
2704 VerifyOption _vo;
2705 bool _all;
2706
2707 public:
2708 PrintReachableOopClosure(outputStream* out,
2709 VerifyOption vo,
2710 bool all) :
2711 _g1h(G1CollectedHeap::heap()),
2712 _out(out), _vo(vo), _all(all) { }
2713
2714 void do_oop(narrowOop* p) { do_oop_work(p); }
2715 void do_oop( oop* p) { do_oop_work(p); }
2716
2717 template <class T> void do_oop_work(T* p) {
2718 oop obj = oopDesc::load_decode_heap_oop(p);
2719 const char* str = NULL;
2720 const char* str2 = "";
2721
2722 if (obj == NULL) {
2723 str = "";
2724 } else if (!_g1h->is_in_g1_reserved(obj)) {
2725 str = " O";
2726 } else {
2727 HeapRegion* hr = _g1h->heap_region_containing(obj);
2728 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2729 bool marked = _g1h->is_marked(obj, _vo);
2730
2731 if (over_tams) {
2732 str = " >";
2733 if (marked) {
2734 str2 = " AND MARKED";
2735 }
2736 } else if (marked) {
2737 str = " M";
2738 } else {
2739 str = " NOT";
2740 }
2741 }
2742
2743 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2744 p2i(p), p2i((void*) obj), str, str2);
2745 }
2746 };
2747
2748 class PrintReachableObjectClosure : public ObjectClosure {
2749 private:
2750 G1CollectedHeap* _g1h;
2751 outputStream* _out;
2752 VerifyOption _vo;
2753 bool _all;
2754 HeapRegion* _hr;
2755
2756 public:
2757 PrintReachableObjectClosure(outputStream* out,
2758 VerifyOption vo,
2759 bool all,
2760 HeapRegion* hr) :
2761 _g1h(G1CollectedHeap::heap()),
2762 _out(out), _vo(vo), _all(all), _hr(hr) { }
2763
2764 void do_object(oop o) {
2765 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2766 bool marked = _g1h->is_marked(o, _vo);
2767 bool print_it = _all || over_tams || marked;
2768
2769 if (print_it) {
2770 _out->print_cr(" "PTR_FORMAT"%s",
2771 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2772 PrintReachableOopClosure oopCl(_out, _vo, _all);
2773 o->oop_iterate_no_header(&oopCl);
2774 }
2775 }
2776 };
2777
2778 class PrintReachableRegionClosure : public HeapRegionClosure {
2779 private:
2780 G1CollectedHeap* _g1h;
2781 outputStream* _out;
2782 VerifyOption _vo;
2783 bool _all;
2784
2785 public:
2786 bool doHeapRegion(HeapRegion* hr) {
2787 HeapWord* b = hr->bottom();
2788 HeapWord* e = hr->end();
2789 HeapWord* t = hr->top();
2790 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2791 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2792 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2793 _out->cr();
2794
2795 HeapWord* from = b;
2796 HeapWord* to = t;
2797
2798 if (to > from) {
2799 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2800 _out->cr();
2801 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2802 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2803 _out->cr();
2804 }
2805
2806 return false;
2807 }
2808
2809 PrintReachableRegionClosure(outputStream* out,
2810 VerifyOption vo,
2811 bool all) :
2812 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2813 };
2814
2815 void ConcurrentMark::print_reachable(const char* str,
2816 VerifyOption vo,
2817 bool all) {
2818 gclog_or_tty->cr();
2819 gclog_or_tty->print_cr("== Doing heap dump... ");
2820
2821 if (G1PrintReachableBaseFile == NULL) {
2822 gclog_or_tty->print_cr(" #### error: no base file defined");
2823 return;
2824 }
2825
2826 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2827 (JVM_MAXPATHLEN - 1)) {
2828 gclog_or_tty->print_cr(" #### error: file name too long");
2829 return;
2830 }
2831
2832 char file_name[JVM_MAXPATHLEN];
2833 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2834 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2835
2836 fileStream fout(file_name);
2837 if (!fout.is_open()) {
2838 gclog_or_tty->print_cr(" #### error: could not open file");
2839 return;
2840 }
2841
2842 outputStream* out = &fout;
2843 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2844 out->cr();
2845
2846 out->print_cr("--- ITERATING OVER REGIONS");
2847 out->cr();
2848 PrintReachableRegionClosure rcl(out, vo, all);
2849 _g1h->heap_region_iterate(&rcl);
2850 out->cr();
2851
2852 gclog_or_tty->print_cr(" done");
2853 gclog_or_tty->flush();
2854 }
2855
2856 #endif // PRODUCT
2857
2858 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2859 // Note we are overriding the read-only view of the prev map here, via
2860 // the cast.
2861 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2862 }
2863
2864 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2865 _nextMarkBitMap->clearRange(mr);
2866 }
2867
2868 HeapRegion*
2869 ConcurrentMark::claim_region(uint worker_id) {
2870 // "checkpoint" the finger
2871 HeapWord* finger = _finger;
2872
2873 // _heap_end will not change underneath our feet; it only changes at
2874 // yield points.
2875 while (finger < _heap_end) {
2876 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2877
2878 // Note on how this code handles humongous regions. In the
2879 // normal case the finger will reach the start of a "starts
2880 // humongous" (SH) region. Its end will either be the end of the
2881 // last "continues humongous" (CH) region in the sequence, or the
2882 // standard end of the SH region (if the SH is the only region in
2883 // the sequence). That way claim_region() will skip over the CH
2884 // regions. However, there is a subtle race between a CM thread
2885 // executing this method and a mutator thread doing a humongous
2886 // object allocation. The two are not mutually exclusive as the CM
2887 // thread does not need to hold the Heap_lock when it gets
2888 // here. So there is a chance that claim_region() will come across
2889 // a free region that's in the progress of becoming a SH or a CH
2890 // region. In the former case, it will either
2891 // a) Miss the update to the region's end, in which case it will
2892 // visit every subsequent CH region, will find their bitmaps
2893 // empty, and do nothing, or
2894 // b) Will observe the update of the region's end (in which case
2895 // it will skip the subsequent CH regions).
2896 // If it comes across a region that suddenly becomes CH, the
2897 // scenario will be similar to b). So, the race between
2898 // claim_region() and a humongous object allocation might force us
2899 // to do a bit of unnecessary work (due to some unnecessary bitmap
2900 // iterations) but it should not introduce and correctness issues.
2901 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2902
2903 // Above heap_region_containing_raw may return NULL as we always scan claim
2904 // until the end of the heap. In this case, just jump to the next region.
2905 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2906
2907 // Is the gap between reading the finger and doing the CAS too long?
2908 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2909 if (res == finger && curr_region != NULL) {
2910 // we succeeded
2911 HeapWord* bottom = curr_region->bottom();
2912 HeapWord* limit = curr_region->next_top_at_mark_start();
2913
2914 if (verbose_low()) {
2915 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2916 "["PTR_FORMAT", "PTR_FORMAT"), "
2917 "limit = "PTR_FORMAT,
2918 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2919 }
2920
2921 // notice that _finger == end cannot be guaranteed here since,
2922 // someone else might have moved the finger even further
2923 assert(_finger >= end, "the finger should have moved forward");
2924
2925 if (verbose_low()) {
2926 gclog_or_tty->print_cr("[%u] we were successful with region = "
2927 PTR_FORMAT, worker_id, p2i(curr_region));
2928 }
2929
2930 if (limit > bottom) {
2931 if (verbose_low()) {
2932 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2933 "returning it ", worker_id, p2i(curr_region));
2934 }
2935 return curr_region;
2936 } else {
2937 assert(limit == bottom,
2938 "the region limit should be at bottom");
2939 if (verbose_low()) {
2940 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
2941 "returning NULL", worker_id, p2i(curr_region));
2942 }
2943 // we return NULL and the caller should try calling
2944 // claim_region() again.
2945 return NULL;
2946 }
2947 } else {
2948 assert(_finger > finger, "the finger should have moved forward");
2949 if (verbose_low()) {
2950 if (curr_region == NULL) {
2951 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
2952 "global finger = "PTR_FORMAT", "
2953 "our finger = "PTR_FORMAT,
2954 worker_id, p2i(_finger), p2i(finger));
2955 } else {
2956 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2957 "global finger = "PTR_FORMAT", "
2958 "our finger = "PTR_FORMAT,
2959 worker_id, p2i(_finger), p2i(finger));
2960 }
2961 }
2962
2963 // read it again
2964 finger = _finger;
2965 }
2966 }
2967
2968 return NULL;
2969 }
2970
2971 #ifndef PRODUCT
2972 enum VerifyNoCSetOopsPhase {
2973 VerifyNoCSetOopsStack,
2974 VerifyNoCSetOopsQueues,
2975 VerifyNoCSetOopsSATBCompleted,
2976 VerifyNoCSetOopsSATBThread
2977 };
2978
2979 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
2980 private:
2981 G1CollectedHeap* _g1h;
2982 VerifyNoCSetOopsPhase _phase;
2983 int _info;
2984
2985 const char* phase_str() {
2986 switch (_phase) {
2987 case VerifyNoCSetOopsStack: return "Stack";
2988 case VerifyNoCSetOopsQueues: return "Queue";
2989 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2990 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
2991 default: ShouldNotReachHere();
2992 }
2993 return NULL;
2994 }
2995
2996 void do_object_work(oop obj) {
2997 guarantee(!_g1h->obj_in_cs(obj),
2998 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2999 p2i((void*) obj), phase_str(), _info));
3000 }
3001
3002 public:
3003 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3004
3005 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3006 _phase = phase;
3007 _info = info;
3008 }
3009
3010 virtual void do_oop(oop* p) {
3011 oop obj = oopDesc::load_decode_heap_oop(p);
3012 do_object_work(obj);
3013 }
3014
3015 virtual void do_oop(narrowOop* p) {
3016 // We should not come across narrow oops while scanning marking
3017 // stacks and SATB buffers.
3018 ShouldNotReachHere();
3019 }
3020
3021 virtual void do_object(oop obj) {
3022 do_object_work(obj);
3023 }
3024 };
3025
3026 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3027 bool verify_enqueued_buffers,
3028 bool verify_thread_buffers,
3029 bool verify_fingers) {
3030 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3031 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3032 return;
3033 }
3034
3035 VerifyNoCSetOopsClosure cl;
3036
3037 if (verify_stacks) {
3038 // Verify entries on the global mark stack
3039 cl.set_phase(VerifyNoCSetOopsStack);
3040 _markStack.oops_do(&cl);
3041
3042 // Verify entries on the task queues
3043 for (uint i = 0; i < _max_worker_id; i += 1) {
3044 cl.set_phase(VerifyNoCSetOopsQueues, i);
3045 CMTaskQueue* queue = _task_queues->queue(i);
3046 queue->oops_do(&cl);
3047 }
3048 }
3049
3050 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3051
3052 // Verify entries on the enqueued SATB buffers
3053 if (verify_enqueued_buffers) {
3054 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3055 satb_qs.iterate_completed_buffers_read_only(&cl);
3056 }
3057
3058 // Verify entries on the per-thread SATB buffers
3059 if (verify_thread_buffers) {
3060 cl.set_phase(VerifyNoCSetOopsSATBThread);
3061 satb_qs.iterate_thread_buffers_read_only(&cl);
3062 }
3063
3064 if (verify_fingers) {
3065 // Verify the global finger
3066 HeapWord* global_finger = finger();
3067 if (global_finger != NULL && global_finger < _heap_end) {
3068 // The global finger always points to a heap region boundary. We
3069 // use heap_region_containing_raw() to get the containing region
3070 // given that the global finger could be pointing to a free region
3071 // which subsequently becomes continues humongous. If that
3072 // happens, heap_region_containing() will return the bottom of the
3073 // corresponding starts humongous region and the check below will
3074 // not hold any more.
3075 // Since we always iterate over all regions, we might get a NULL HeapRegion
3076 // here.
3077 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3078 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3079 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3080 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3081 }
3082
3083 // Verify the task fingers
3084 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3085 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3086 CMTask* task = _tasks[i];
3087 HeapWord* task_finger = task->finger();
3088 if (task_finger != NULL && task_finger < _heap_end) {
3089 // See above note on the global finger verification.
3090 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3091 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3092 !task_hr->in_collection_set(),
3093 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3094 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3095 }
3096 }
3097 }
3098 }
3099 #endif // PRODUCT
3100
3101 // Aggregate the counting data that was constructed concurrently
3102 // with marking.
3103 class AggregateCountDataHRClosure: public HeapRegionClosure {
3104 G1CollectedHeap* _g1h;
3105 ConcurrentMark* _cm;
3106 CardTableModRefBS* _ct_bs;
3107 BitMap* _cm_card_bm;
3108 uint _max_worker_id;
3109
3110 public:
3111 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3112 BitMap* cm_card_bm,
3113 uint max_worker_id) :
3114 _g1h(g1h), _cm(g1h->concurrent_mark()),
3115 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3116 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3117
3118 bool doHeapRegion(HeapRegion* hr) {
3119 if (hr->is_continues_humongous()) {
3120 // We will ignore these here and process them when their
3121 // associated "starts humongous" region is processed.
3122 // Note that we cannot rely on their associated
3123 // "starts humongous" region to have their bit set to 1
3124 // since, due to the region chunking in the parallel region
3125 // iteration, a "continues humongous" region might be visited
3126 // before its associated "starts humongous".
3127 return false;
3128 }
3129
3130 HeapWord* start = hr->bottom();
3131 HeapWord* limit = hr->next_top_at_mark_start();
3132 HeapWord* end = hr->end();
3133
3134 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3135 err_msg("Preconditions not met - "
3136 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3137 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3138 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3139
3140 assert(hr->next_marked_bytes() == 0, "Precondition");
3141
3142 if (start == limit) {
3143 // NTAMS of this region has not been set so nothing to do.
3144 return false;
3145 }
3146
3147 // 'start' should be in the heap.
3148 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3149 // 'end' *may* be just beyond the end of the heap (if hr is the last region)
3150 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3151
3152 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3153 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3154 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3155
3156 // If ntams is not card aligned then we bump card bitmap index
3157 // for limit so that we get the all the cards spanned by
3158 // the object ending at ntams.
3159 // Note: if this is the last region in the heap then ntams
3160 // could be actually just beyond the end of the the heap;
3161 // limit_idx will then correspond to a (non-existent) card
3162 // that is also outside the heap.
3163 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3164 limit_idx += 1;
3165 }
3166
3167 assert(limit_idx <= end_idx, "or else use atomics");
3168
3169 // Aggregate the "stripe" in the count data associated with hr.
3170 uint hrm_index = hr->hrm_index();
3171 size_t marked_bytes = 0;
3172
3173 for (uint i = 0; i < _max_worker_id; i += 1) {
3174 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3175 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3176
3177 // Fetch the marked_bytes in this region for task i and
3178 // add it to the running total for this region.
3179 marked_bytes += marked_bytes_array[hrm_index];
3180
3181 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3182 // into the global card bitmap.
3183 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3184
3185 while (scan_idx < limit_idx) {
3186 assert(task_card_bm->at(scan_idx) == true, "should be");
3187 _cm_card_bm->set_bit(scan_idx);
3188 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3189
3190 // BitMap::get_next_one_offset() can handle the case when
3191 // its left_offset parameter is greater than its right_offset
3192 // parameter. It does, however, have an early exit if
3193 // left_offset == right_offset. So let's limit the value
3194 // passed in for left offset here.
3195 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3196 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3197 }
3198 }
3199
3200 // Update the marked bytes for this region.
3201 hr->add_to_marked_bytes(marked_bytes);
3202
3203 // Next heap region
3204 return false;
3205 }
3206 };
3207
3208 class G1AggregateCountDataTask: public AbstractGangTask {
3209 protected:
3210 G1CollectedHeap* _g1h;
3211 ConcurrentMark* _cm;
3212 BitMap* _cm_card_bm;
3213 uint _max_worker_id;
3214 int _active_workers;
3215 HeapRegionClaimer _hrclaimer;
3216
3217 public:
3218 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3219 ConcurrentMark* cm,
3220 BitMap* cm_card_bm,
3221 uint max_worker_id,
3222 int n_workers) :
3223 AbstractGangTask("Count Aggregation"),
3224 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3225 _max_worker_id(max_worker_id),
3226 _active_workers(n_workers),
3227 _hrclaimer(_active_workers) {
3228 }
3229
3230 void work(uint worker_id) {
3231 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3232
3233 _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer);
3234 }
3235 };
3236
3237
3238 void ConcurrentMark::aggregate_count_data() {
3239 int n_workers = _g1h->workers()->active_workers();
3240
3241 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3242 _max_worker_id, n_workers);
3243
3244 _g1h->set_par_threads(n_workers);
3245 _g1h->workers()->run_task(&g1_par_agg_task);
3246 _g1h->set_par_threads(0);
3247 }
3248
3249 // Clear the per-worker arrays used to store the per-region counting data
3250 void ConcurrentMark::clear_all_count_data() {
3251 // Clear the global card bitmap - it will be filled during
3252 // liveness count aggregation (during remark) and the
3253 // final counting task.
3254 _card_bm.clear();
3255
3256 // Clear the global region bitmap - it will be filled as part
3257 // of the final counting task.
3258 _region_bm.clear();
3259
3260 uint max_regions = _g1h->max_regions();
3261 assert(_max_worker_id > 0, "uninitialized");
3262
3263 for (uint i = 0; i < _max_worker_id; i += 1) {
3264 BitMap* task_card_bm = count_card_bitmap_for(i);
3265 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3266
3267 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3268 assert(marked_bytes_array != NULL, "uninitialized");
3269
3270 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3271 task_card_bm->clear();
3272 }
3273 }
3274
3275 void ConcurrentMark::print_stats() {
3276 if (verbose_stats()) {
3277 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3278 for (size_t i = 0; i < _active_tasks; ++i) {
3279 _tasks[i]->print_stats();
3280 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3281 }
3282 }
3283 }
3284
3285 // abandon current marking iteration due to a Full GC
3286 void ConcurrentMark::abort() {
3287 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3288 // concurrent bitmap clearing.
3289 _nextMarkBitMap->clearAll();
3290
3291 // Note we cannot clear the previous marking bitmap here
3292 // since VerifyDuringGC verifies the objects marked during
3293 // a full GC against the previous bitmap.
3294
3295 // Clear the liveness counting data
3296 clear_all_count_data();
3297 // Empty mark stack
3298 reset_marking_state();
3299 for (uint i = 0; i < _max_worker_id; ++i) {
3300 _tasks[i]->clear_region_fields();
3301 }
3302 _first_overflow_barrier_sync.abort();
3303 _second_overflow_barrier_sync.abort();
3304 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3305 if (!gc_id.is_undefined()) {
3306 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3307 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3308 _aborted_gc_id = gc_id;
3309 }
3310 _has_aborted = true;
3311
3312 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3313 satb_mq_set.abandon_partial_marking();
3314 // This can be called either during or outside marking, we'll read
3315 // the expected_active value from the SATB queue set.
3316 satb_mq_set.set_active_all_threads(
3317 false, /* new active value */
3318 satb_mq_set.is_active() /* expected_active */);
3319
3320 _g1h->trace_heap_after_concurrent_cycle();
3321 _g1h->register_concurrent_cycle_end();
3322 }
3323
3324 const GCId& ConcurrentMark::concurrent_gc_id() {
3325 if (has_aborted()) {
3326 return _aborted_gc_id;
3327 }
3328 return _g1h->gc_tracer_cm()->gc_id();
3329 }
3330
3331 static void print_ms_time_info(const char* prefix, const char* name,
3332 NumberSeq& ns) {
3333 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3334 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3335 if (ns.num() > 0) {
3336 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3337 prefix, ns.sd(), ns.maximum());
3338 }
3339 }
3340
3341 void ConcurrentMark::print_summary_info() {
3342 gclog_or_tty->print_cr(" Concurrent marking:");
3343 print_ms_time_info(" ", "init marks", _init_times);
3344 print_ms_time_info(" ", "remarks", _remark_times);
3345 {
3346 print_ms_time_info(" ", "final marks", _remark_mark_times);
3347 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3348
3349 }
3350 print_ms_time_info(" ", "cleanups", _cleanup_times);
3351 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3352 _total_counting_time,
3353 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3354 (double)_cleanup_times.num()
3355 : 0.0));
3356 if (G1ScrubRemSets) {
3357 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3358 _total_rs_scrub_time,
3359 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3360 (double)_cleanup_times.num()
3361 : 0.0));
3362 }
3363 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3364 (_init_times.sum() + _remark_times.sum() +
3365 _cleanup_times.sum())/1000.0);
3366 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3367 "(%8.2f s marking).",
3368 cmThread()->vtime_accum(),
3369 cmThread()->vtime_mark_accum());
3370 }
3371
3372 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3373 _parallel_workers->print_worker_threads_on(st);
3374 }
3375
3376 void ConcurrentMark::print_on_error(outputStream* st) const {
3377 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3378 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3379 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3380 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3381 }
3382
3383 // We take a break if someone is trying to stop the world.
3384 bool ConcurrentMark::do_yield_check(uint worker_id) {
3385 if (SuspendibleThreadSet::should_yield()) {
3386 if (worker_id == 0) {
3387 _g1h->g1_policy()->record_concurrent_pause();
3388 }
3389 SuspendibleThreadSet::yield();
3390 return true;
3391 } else {
3392 return false;
3393 }
3394 }
3395
3396 #ifndef PRODUCT
3397 // for debugging purposes
3398 void ConcurrentMark::print_finger() {
3399 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3400 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3401 for (uint i = 0; i < _max_worker_id; ++i) {
3402 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3403 }
3404 gclog_or_tty->cr();
3405 }
3406 #endif
3407
3408 void CMTask::scan_object(oop obj) {
3409 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3410
3411 if (_cm->verbose_high()) {
3412 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3413 _worker_id, p2i((void*) obj));
3414 }
3415
3416 size_t obj_size = obj->size();
3417 _words_scanned += obj_size;
3418
3419 obj->oop_iterate(_cm_oop_closure);
3420 statsOnly( ++_objs_scanned );
3421 check_limits();
3422 }
3423
3424 // Closure for iteration over bitmaps
3425 class CMBitMapClosure : public BitMapClosure {
3426 private:
3427 // the bitmap that is being iterated over
3428 CMBitMap* _nextMarkBitMap;
3429 ConcurrentMark* _cm;
3430 CMTask* _task;
3431
3432 public:
3433 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3434 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3435
3436 bool do_bit(size_t offset) {
3437 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3438 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3439 assert( addr < _cm->finger(), "invariant");
3440
3441 statsOnly( _task->increase_objs_found_on_bitmap() );
3442 assert(addr >= _task->finger(), "invariant");
3443
3444 // We move that task's local finger along.
3445 _task->move_finger_to(addr);
3446
3447 _task->scan_object(oop(addr));
3448 // we only partially drain the local queue and global stack
3449 _task->drain_local_queue(true);
3450 _task->drain_global_stack(true);
3451
3452 // if the has_aborted flag has been raised, we need to bail out of
3453 // the iteration
3454 return !_task->has_aborted();
3455 }
3456 };
3457
3458 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3459 ConcurrentMark* cm,
3460 CMTask* task)
3461 : _g1h(g1h), _cm(cm), _task(task) {
3462 assert(_ref_processor == NULL, "should be initialized to NULL");
3463
3464 if (G1UseConcMarkReferenceProcessing) {
3465 _ref_processor = g1h->ref_processor_cm();
3466 assert(_ref_processor != NULL, "should not be NULL");
3467 }
3468 }
3469
3470 void CMTask::setup_for_region(HeapRegion* hr) {
3471 assert(hr != NULL,
3472 "claim_region() should have filtered out NULL regions");
3473 assert(!hr->is_continues_humongous(),
3474 "claim_region() should have filtered out continues humongous regions");
3475
3476 if (_cm->verbose_low()) {
3477 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3478 _worker_id, p2i(hr));
3479 }
3480
3481 _curr_region = hr;
3482 _finger = hr->bottom();
3483 update_region_limit();
3484 }
3485
3486 void CMTask::update_region_limit() {
3487 HeapRegion* hr = _curr_region;
3488 HeapWord* bottom = hr->bottom();
3489 HeapWord* limit = hr->next_top_at_mark_start();
3490
3491 if (limit == bottom) {
3492 if (_cm->verbose_low()) {
3493 gclog_or_tty->print_cr("[%u] found an empty region "
3494 "["PTR_FORMAT", "PTR_FORMAT")",
3495 _worker_id, p2i(bottom), p2i(limit));
3496 }
3497 // The region was collected underneath our feet.
3498 // We set the finger to bottom to ensure that the bitmap
3499 // iteration that will follow this will not do anything.
3500 // (this is not a condition that holds when we set the region up,
3501 // as the region is not supposed to be empty in the first place)
3502 _finger = bottom;
3503 } else if (limit >= _region_limit) {
3504 assert(limit >= _finger, "peace of mind");
3505 } else {
3506 assert(limit < _region_limit, "only way to get here");
3507 // This can happen under some pretty unusual circumstances. An
3508 // evacuation pause empties the region underneath our feet (NTAMS
3509 // at bottom). We then do some allocation in the region (NTAMS
3510 // stays at bottom), followed by the region being used as a GC
3511 // alloc region (NTAMS will move to top() and the objects
3512 // originally below it will be grayed). All objects now marked in
3513 // the region are explicitly grayed, if below the global finger,
3514 // and we do not need in fact to scan anything else. So, we simply
3515 // set _finger to be limit to ensure that the bitmap iteration
3516 // doesn't do anything.
3517 _finger = limit;
3518 }
3519
3520 _region_limit = limit;
3521 }
3522
3523 void CMTask::giveup_current_region() {
3524 assert(_curr_region != NULL, "invariant");
3525 if (_cm->verbose_low()) {
3526 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3527 _worker_id, p2i(_curr_region));
3528 }
3529 clear_region_fields();
3530 }
3531
3532 void CMTask::clear_region_fields() {
3533 // Values for these three fields that indicate that we're not
3534 // holding on to a region.
3535 _curr_region = NULL;
3536 _finger = NULL;
3537 _region_limit = NULL;
3538 }
3539
3540 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3541 if (cm_oop_closure == NULL) {
3542 assert(_cm_oop_closure != NULL, "invariant");
3543 } else {
3544 assert(_cm_oop_closure == NULL, "invariant");
3545 }
3546 _cm_oop_closure = cm_oop_closure;
3547 }
3548
3549 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3550 guarantee(nextMarkBitMap != NULL, "invariant");
3551
3552 if (_cm->verbose_low()) {
3553 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3554 }
3555
3556 _nextMarkBitMap = nextMarkBitMap;
3557 clear_region_fields();
3558
3559 _calls = 0;
3560 _elapsed_time_ms = 0.0;
3561 _termination_time_ms = 0.0;
3562 _termination_start_time_ms = 0.0;
3563
3564 #if _MARKING_STATS_
3565 _local_pushes = 0;
3566 _local_pops = 0;
3567 _local_max_size = 0;
3568 _objs_scanned = 0;
3569 _global_pushes = 0;
3570 _global_pops = 0;
3571 _global_max_size = 0;
3572 _global_transfers_to = 0;
3573 _global_transfers_from = 0;
3574 _regions_claimed = 0;
3575 _objs_found_on_bitmap = 0;
3576 _satb_buffers_processed = 0;
3577 _steal_attempts = 0;
3578 _steals = 0;
3579 _aborted = 0;
3580 _aborted_overflow = 0;
3581 _aborted_cm_aborted = 0;
3582 _aborted_yield = 0;
3583 _aborted_timed_out = 0;
3584 _aborted_satb = 0;
3585 _aborted_termination = 0;
3586 #endif // _MARKING_STATS_
3587 }
3588
3589 bool CMTask::should_exit_termination() {
3590 regular_clock_call();
3591 // This is called when we are in the termination protocol. We should
3592 // quit if, for some reason, this task wants to abort or the global
3593 // stack is not empty (this means that we can get work from it).
3594 return !_cm->mark_stack_empty() || has_aborted();
3595 }
3596
3597 void CMTask::reached_limit() {
3598 assert(_words_scanned >= _words_scanned_limit ||
3599 _refs_reached >= _refs_reached_limit ,
3600 "shouldn't have been called otherwise");
3601 regular_clock_call();
3602 }
3603
3604 void CMTask::regular_clock_call() {
3605 if (has_aborted()) return;
3606
3607 // First, we need to recalculate the words scanned and refs reached
3608 // limits for the next clock call.
3609 recalculate_limits();
3610
3611 // During the regular clock call we do the following
3612
3613 // (1) If an overflow has been flagged, then we abort.
3614 if (_cm->has_overflown()) {
3615 set_has_aborted();
3616 return;
3617 }
3618
3619 // If we are not concurrent (i.e. we're doing remark) we don't need
3620 // to check anything else. The other steps are only needed during
3621 // the concurrent marking phase.
3622 if (!concurrent()) return;
3623
3624 // (2) If marking has been aborted for Full GC, then we also abort.
3625 if (_cm->has_aborted()) {
3626 set_has_aborted();
3627 statsOnly( ++_aborted_cm_aborted );
3628 return;
3629 }
3630
3631 double curr_time_ms = os::elapsedVTime() * 1000.0;
3632
3633 // (3) If marking stats are enabled, then we update the step history.
3634 #if _MARKING_STATS_
3635 if (_words_scanned >= _words_scanned_limit) {
3636 ++_clock_due_to_scanning;
3637 }
3638 if (_refs_reached >= _refs_reached_limit) {
3639 ++_clock_due_to_marking;
3640 }
3641
3642 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3643 _interval_start_time_ms = curr_time_ms;
3644 _all_clock_intervals_ms.add(last_interval_ms);
3645
3646 if (_cm->verbose_medium()) {
3647 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3648 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3649 _worker_id, last_interval_ms,
3650 _words_scanned,
3651 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3652 _refs_reached,
3653 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3654 }
3655 #endif // _MARKING_STATS_
3656
3657 // (4) We check whether we should yield. If we have to, then we abort.
3658 if (SuspendibleThreadSet::should_yield()) {
3659 // We should yield. To do this we abort the task. The caller is
3660 // responsible for yielding.
3661 set_has_aborted();
3662 statsOnly( ++_aborted_yield );
3663 return;
3664 }
3665
3666 // (5) We check whether we've reached our time quota. If we have,
3667 // then we abort.
3668 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3669 if (elapsed_time_ms > _time_target_ms) {
3670 set_has_aborted();
3671 _has_timed_out = true;
3672 statsOnly( ++_aborted_timed_out );
3673 return;
3674 }
3675
3676 // (6) Finally, we check whether there are enough completed STAB
3677 // buffers available for processing. If there are, we abort.
3678 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3679 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3680 if (_cm->verbose_low()) {
3681 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3682 _worker_id);
3683 }
3684 // we do need to process SATB buffers, we'll abort and restart
3685 // the marking task to do so
3686 set_has_aborted();
3687 statsOnly( ++_aborted_satb );
3688 return;
3689 }
3690 }
3691
3692 void CMTask::recalculate_limits() {
3693 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3694 _words_scanned_limit = _real_words_scanned_limit;
3695
3696 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3697 _refs_reached_limit = _real_refs_reached_limit;
3698 }
3699
3700 void CMTask::decrease_limits() {
3701 // This is called when we believe that we're going to do an infrequent
3702 // operation which will increase the per byte scanned cost (i.e. move
3703 // entries to/from the global stack). It basically tries to decrease the
3704 // scanning limit so that the clock is called earlier.
3705
3706 if (_cm->verbose_medium()) {
3707 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3708 }
3709
3710 _words_scanned_limit = _real_words_scanned_limit -
3711 3 * words_scanned_period / 4;
3712 _refs_reached_limit = _real_refs_reached_limit -
3713 3 * refs_reached_period / 4;
3714 }
3715
3716 void CMTask::move_entries_to_global_stack() {
3717 // local array where we'll store the entries that will be popped
3718 // from the local queue
3719 oop buffer[global_stack_transfer_size];
3720
3721 int n = 0;
3722 oop obj;
3723 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3724 buffer[n] = obj;
3725 ++n;
3726 }
3727
3728 if (n > 0) {
3729 // we popped at least one entry from the local queue
3730
3731 statsOnly( ++_global_transfers_to; _local_pops += n );
3732
3733 if (!_cm->mark_stack_push(buffer, n)) {
3734 if (_cm->verbose_low()) {
3735 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3736 _worker_id);
3737 }
3738 set_has_aborted();
3739 } else {
3740 // the transfer was successful
3741
3742 if (_cm->verbose_medium()) {
3743 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3744 _worker_id, n);
3745 }
3746 statsOnly( int tmp_size = _cm->mark_stack_size();
3747 if (tmp_size > _global_max_size) {
3748 _global_max_size = tmp_size;
3749 }
3750 _global_pushes += n );
3751 }
3752 }
3753
3754 // this operation was quite expensive, so decrease the limits
3755 decrease_limits();
3756 }
3757
3758 void CMTask::get_entries_from_global_stack() {
3759 // local array where we'll store the entries that will be popped
3760 // from the global stack.
3761 oop buffer[global_stack_transfer_size];
3762 int n;
3763 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3764 assert(n <= global_stack_transfer_size,
3765 "we should not pop more than the given limit");
3766 if (n > 0) {
3767 // yes, we did actually pop at least one entry
3768
3769 statsOnly( ++_global_transfers_from; _global_pops += n );
3770 if (_cm->verbose_medium()) {
3771 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3772 _worker_id, n);
3773 }
3774 for (int i = 0; i < n; ++i) {
3775 bool success = _task_queue->push(buffer[i]);
3776 // We only call this when the local queue is empty or under a
3777 // given target limit. So, we do not expect this push to fail.
3778 assert(success, "invariant");
3779 }
3780
3781 statsOnly( int tmp_size = _task_queue->size();
3782 if (tmp_size > _local_max_size) {
3783 _local_max_size = tmp_size;
3784 }
3785 _local_pushes += n );
3786 }
3787
3788 // this operation was quite expensive, so decrease the limits
3789 decrease_limits();
3790 }
3791
3792 void CMTask::drain_local_queue(bool partially) {
3793 if (has_aborted()) return;
3794
3795 // Decide what the target size is, depending whether we're going to
3796 // drain it partially (so that other tasks can steal if they run out
3797 // of things to do) or totally (at the very end).
3798 size_t target_size;
3799 if (partially) {
3800 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3801 } else {
3802 target_size = 0;
3803 }
3804
3805 if (_task_queue->size() > target_size) {
3806 if (_cm->verbose_high()) {
3807 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3808 _worker_id, target_size);
3809 }
3810
3811 oop obj;
3812 bool ret = _task_queue->pop_local(obj);
3813 while (ret) {
3814 statsOnly( ++_local_pops );
3815
3816 if (_cm->verbose_high()) {
3817 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3818 p2i((void*) obj));
3819 }
3820
3821 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3822 assert(!_g1h->is_on_master_free_list(
3823 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3824
3825 scan_object(obj);
3826
3827 if (_task_queue->size() <= target_size || has_aborted()) {
3828 ret = false;
3829 } else {
3830 ret = _task_queue->pop_local(obj);
3831 }
3832 }
3833
3834 if (_cm->verbose_high()) {
3835 gclog_or_tty->print_cr("[%u] drained local queue, size = %u",
3836 _worker_id, _task_queue->size());
3837 }
3838 }
3839 }
3840
3841 void CMTask::drain_global_stack(bool partially) {
3842 if (has_aborted()) return;
3843
3844 // We have a policy to drain the local queue before we attempt to
3845 // drain the global stack.
3846 assert(partially || _task_queue->size() == 0, "invariant");
3847
3848 // Decide what the target size is, depending whether we're going to
3849 // drain it partially (so that other tasks can steal if they run out
3850 // of things to do) or totally (at the very end). Notice that,
3851 // because we move entries from the global stack in chunks or
3852 // because another task might be doing the same, we might in fact
3853 // drop below the target. But, this is not a problem.
3854 size_t target_size;
3855 if (partially) {
3856 target_size = _cm->partial_mark_stack_size_target();
3857 } else {
3858 target_size = 0;
3859 }
3860
3861 if (_cm->mark_stack_size() > target_size) {
3862 if (_cm->verbose_low()) {
3863 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3864 _worker_id, target_size);
3865 }
3866
3867 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3868 get_entries_from_global_stack();
3869 drain_local_queue(partially);
3870 }
3871
3872 if (_cm->verbose_low()) {
3873 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3874 _worker_id, _cm->mark_stack_size());
3875 }
3876 }
3877 }
3878
3879 // SATB Queue has several assumptions on whether to call the par or
3880 // non-par versions of the methods. this is why some of the code is
3881 // replicated. We should really get rid of the single-threaded version
3882 // of the code to simplify things.
3883 void CMTask::drain_satb_buffers() {
3884 if (has_aborted()) return;
3885
3886 // We set this so that the regular clock knows that we're in the
3887 // middle of draining buffers and doesn't set the abort flag when it
3888 // notices that SATB buffers are available for draining. It'd be
3889 // very counter productive if it did that. :-)
3890 _draining_satb_buffers = true;
3891
3892 CMObjectClosure oc(this);
3893 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3894 satb_mq_set.set_closure(_worker_id, &oc);
3895
3896 // This keeps claiming and applying the closure to completed buffers
3897 // until we run out of buffers or we need to abort.
3898 while (!has_aborted() &&
3899 satb_mq_set.apply_closure_to_completed_buffer(_worker_id)) {
3900 if (_cm->verbose_medium()) {
3901 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3902 }
3903 statsOnly( ++_satb_buffers_processed );
3904 regular_clock_call();
3905 }
3906
3907 _draining_satb_buffers = false;
3908
3909 assert(has_aborted() ||
3910 concurrent() ||
3911 satb_mq_set.completed_buffers_num() == 0, "invariant");
3912
3913 satb_mq_set.set_closure(_worker_id, NULL);
3914
3915 // again, this was a potentially expensive operation, decrease the
3916 // limits to get the regular clock call early
3917 decrease_limits();
3918 }
3919
3920 void CMTask::print_stats() {
3921 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3922 _worker_id, _calls);
3923 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3924 _elapsed_time_ms, _termination_time_ms);
3925 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3926 _step_times_ms.num(), _step_times_ms.avg(),
3927 _step_times_ms.sd());
3928 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3929 _step_times_ms.maximum(), _step_times_ms.sum());
3930
3931 #if _MARKING_STATS_
3932 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3933 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3934 _all_clock_intervals_ms.sd());
3935 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3936 _all_clock_intervals_ms.maximum(),
3937 _all_clock_intervals_ms.sum());
3938 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
3939 _clock_due_to_scanning, _clock_due_to_marking);
3940 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
3941 _objs_scanned, _objs_found_on_bitmap);
3942 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
3943 _local_pushes, _local_pops, _local_max_size);
3944 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
3945 _global_pushes, _global_pops, _global_max_size);
3946 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
3947 _global_transfers_to,_global_transfers_from);
3948 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
3949 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
3950 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
3951 _steal_attempts, _steals);
3952 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
3953 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
3954 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3955 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
3956 _aborted_timed_out, _aborted_satb, _aborted_termination);
3957 #endif // _MARKING_STATS_
3958 }
3959
3960 /*****************************************************************************
3961
3962 The do_marking_step(time_target_ms, ...) method is the building
3963 block of the parallel marking framework. It can be called in parallel
3964 with other invocations of do_marking_step() on different tasks
3965 (but only one per task, obviously) and concurrently with the
3966 mutator threads, or during remark, hence it eliminates the need
3967 for two versions of the code. When called during remark, it will
3968 pick up from where the task left off during the concurrent marking
3969 phase. Interestingly, tasks are also claimable during evacuation
3970 pauses too, since do_marking_step() ensures that it aborts before
3971 it needs to yield.
3972
3973 The data structures that it uses to do marking work are the
3974 following:
3975
3976 (1) Marking Bitmap. If there are gray objects that appear only
3977 on the bitmap (this happens either when dealing with an overflow
3978 or when the initial marking phase has simply marked the roots
3979 and didn't push them on the stack), then tasks claim heap
3980 regions whose bitmap they then scan to find gray objects. A
3981 global finger indicates where the end of the last claimed region
3982 is. A local finger indicates how far into the region a task has
3983 scanned. The two fingers are used to determine how to gray an
3984 object (i.e. whether simply marking it is OK, as it will be
3985 visited by a task in the future, or whether it needs to be also
3986 pushed on a stack).
3987
3988 (2) Local Queue. The local queue of the task which is accessed
3989 reasonably efficiently by the task. Other tasks can steal from
3990 it when they run out of work. Throughout the marking phase, a
3991 task attempts to keep its local queue short but not totally
3992 empty, so that entries are available for stealing by other
3993 tasks. Only when there is no more work, a task will totally
3994 drain its local queue.
3995
3996 (3) Global Mark Stack. This handles local queue overflow. During
3997 marking only sets of entries are moved between it and the local
3998 queues, as access to it requires a mutex and more fine-grain
3999 interaction with it which might cause contention. If it
4000 overflows, then the marking phase should restart and iterate
4001 over the bitmap to identify gray objects. Throughout the marking
4002 phase, tasks attempt to keep the global mark stack at a small
4003 length but not totally empty, so that entries are available for
4004 popping by other tasks. Only when there is no more work, tasks
4005 will totally drain the global mark stack.
4006
4007 (4) SATB Buffer Queue. This is where completed SATB buffers are
4008 made available. Buffers are regularly removed from this queue
4009 and scanned for roots, so that the queue doesn't get too
4010 long. During remark, all completed buffers are processed, as
4011 well as the filled in parts of any uncompleted buffers.
4012
4013 The do_marking_step() method tries to abort when the time target
4014 has been reached. There are a few other cases when the
4015 do_marking_step() method also aborts:
4016
4017 (1) When the marking phase has been aborted (after a Full GC).
4018
4019 (2) When a global overflow (on the global stack) has been
4020 triggered. Before the task aborts, it will actually sync up with
4021 the other tasks to ensure that all the marking data structures
4022 (local queues, stacks, fingers etc.) are re-initialized so that
4023 when do_marking_step() completes, the marking phase can
4024 immediately restart.
4025
4026 (3) When enough completed SATB buffers are available. The
4027 do_marking_step() method only tries to drain SATB buffers right
4028 at the beginning. So, if enough buffers are available, the
4029 marking step aborts and the SATB buffers are processed at
4030 the beginning of the next invocation.
4031
4032 (4) To yield. when we have to yield then we abort and yield
4033 right at the end of do_marking_step(). This saves us from a lot
4034 of hassle as, by yielding we might allow a Full GC. If this
4035 happens then objects will be compacted underneath our feet, the
4036 heap might shrink, etc. We save checking for this by just
4037 aborting and doing the yield right at the end.
4038
4039 From the above it follows that the do_marking_step() method should
4040 be called in a loop (or, otherwise, regularly) until it completes.
4041
4042 If a marking step completes without its has_aborted() flag being
4043 true, it means it has completed the current marking phase (and
4044 also all other marking tasks have done so and have all synced up).
4045
4046 A method called regular_clock_call() is invoked "regularly" (in
4047 sub ms intervals) throughout marking. It is this clock method that
4048 checks all the abort conditions which were mentioned above and
4049 decides when the task should abort. A work-based scheme is used to
4050 trigger this clock method: when the number of object words the
4051 marking phase has scanned or the number of references the marking
4052 phase has visited reach a given limit. Additional invocations to
4053 the method clock have been planted in a few other strategic places
4054 too. The initial reason for the clock method was to avoid calling
4055 vtime too regularly, as it is quite expensive. So, once it was in
4056 place, it was natural to piggy-back all the other conditions on it
4057 too and not constantly check them throughout the code.
4058
4059 If do_termination is true then do_marking_step will enter its
4060 termination protocol.
4061
4062 The value of is_serial must be true when do_marking_step is being
4063 called serially (i.e. by the VMThread) and do_marking_step should
4064 skip any synchronization in the termination and overflow code.
4065 Examples include the serial remark code and the serial reference
4066 processing closures.
4067
4068 The value of is_serial must be false when do_marking_step is
4069 being called by any of the worker threads in a work gang.
4070 Examples include the concurrent marking code (CMMarkingTask),
4071 the MT remark code, and the MT reference processing closures.
4072
4073 *****************************************************************************/
4074
4075 void CMTask::do_marking_step(double time_target_ms,
4076 bool do_termination,
4077 bool is_serial) {
4078 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4079 assert(concurrent() == _cm->concurrent(), "they should be the same");
4080
4081 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4082 assert(_task_queues != NULL, "invariant");
4083 assert(_task_queue != NULL, "invariant");
4084 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4085
4086 assert(!_claimed,
4087 "only one thread should claim this task at any one time");
4088
4089 // OK, this doesn't safeguard again all possible scenarios, as it is
4090 // possible for two threads to set the _claimed flag at the same
4091 // time. But it is only for debugging purposes anyway and it will
4092 // catch most problems.
4093 _claimed = true;
4094
4095 _start_time_ms = os::elapsedVTime() * 1000.0;
4096 statsOnly( _interval_start_time_ms = _start_time_ms );
4097
4098 // If do_stealing is true then do_marking_step will attempt to
4099 // steal work from the other CMTasks. It only makes sense to
4100 // enable stealing when the termination protocol is enabled
4101 // and do_marking_step() is not being called serially.
4102 bool do_stealing = do_termination && !is_serial;
4103
4104 double diff_prediction_ms =
4105 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4106 _time_target_ms = time_target_ms - diff_prediction_ms;
4107
4108 // set up the variables that are used in the work-based scheme to
4109 // call the regular clock method
4110 _words_scanned = 0;
4111 _refs_reached = 0;
4112 recalculate_limits();
4113
4114 // clear all flags
4115 clear_has_aborted();
4116 _has_timed_out = false;
4117 _draining_satb_buffers = false;
4118
4119 ++_calls;
4120
4121 if (_cm->verbose_low()) {
4122 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4123 "target = %1.2lfms >>>>>>>>>>",
4124 _worker_id, _calls, _time_target_ms);
4125 }
4126
4127 // Set up the bitmap and oop closures. Anything that uses them is
4128 // eventually called from this method, so it is OK to allocate these
4129 // statically.
4130 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4131 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4132 set_cm_oop_closure(&cm_oop_closure);
4133
4134 if (_cm->has_overflown()) {
4135 // This can happen if the mark stack overflows during a GC pause
4136 // and this task, after a yield point, restarts. We have to abort
4137 // as we need to get into the overflow protocol which happens
4138 // right at the end of this task.
4139 set_has_aborted();
4140 }
4141
4142 // First drain any available SATB buffers. After this, we will not
4143 // look at SATB buffers before the next invocation of this method.
4144 // If enough completed SATB buffers are queued up, the regular clock
4145 // will abort this task so that it restarts.
4146 drain_satb_buffers();
4147 // ...then partially drain the local queue and the global stack
4148 drain_local_queue(true);
4149 drain_global_stack(true);
4150
4151 do {
4152 if (!has_aborted() && _curr_region != NULL) {
4153 // This means that we're already holding on to a region.
4154 assert(_finger != NULL, "if region is not NULL, then the finger "
4155 "should not be NULL either");
4156
4157 // We might have restarted this task after an evacuation pause
4158 // which might have evacuated the region we're holding on to
4159 // underneath our feet. Let's read its limit again to make sure
4160 // that we do not iterate over a region of the heap that
4161 // contains garbage (update_region_limit() will also move
4162 // _finger to the start of the region if it is found empty).
4163 update_region_limit();
4164 // We will start from _finger not from the start of the region,
4165 // as we might be restarting this task after aborting half-way
4166 // through scanning this region. In this case, _finger points to
4167 // the address where we last found a marked object. If this is a
4168 // fresh region, _finger points to start().
4169 MemRegion mr = MemRegion(_finger, _region_limit);
4170
4171 if (_cm->verbose_low()) {
4172 gclog_or_tty->print_cr("[%u] we're scanning part "
4173 "["PTR_FORMAT", "PTR_FORMAT") "
4174 "of region "HR_FORMAT,
4175 _worker_id, p2i(_finger), p2i(_region_limit),
4176 HR_FORMAT_PARAMS(_curr_region));
4177 }
4178
4179 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
4180 "humongous regions should go around loop once only");
4181
4182 // Some special cases:
4183 // If the memory region is empty, we can just give up the region.
4184 // If the current region is humongous then we only need to check
4185 // the bitmap for the bit associated with the start of the object,
4186 // scan the object if it's live, and give up the region.
4187 // Otherwise, let's iterate over the bitmap of the part of the region
4188 // that is left.
4189 // If the iteration is successful, give up the region.
4190 if (mr.is_empty()) {
4191 giveup_current_region();
4192 regular_clock_call();
4193 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
4194 if (_nextMarkBitMap->isMarked(mr.start())) {
4195 // The object is marked - apply the closure
4196 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4197 bitmap_closure.do_bit(offset);
4198 }
4199 // Even if this task aborted while scanning the humongous object
4200 // we can (and should) give up the current region.
4201 giveup_current_region();
4202 regular_clock_call();
4203 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4204 giveup_current_region();
4205 regular_clock_call();
4206 } else {
4207 assert(has_aborted(), "currently the only way to do so");
4208 // The only way to abort the bitmap iteration is to return
4209 // false from the do_bit() method. However, inside the
4210 // do_bit() method we move the _finger to point to the
4211 // object currently being looked at. So, if we bail out, we
4212 // have definitely set _finger to something non-null.
4213 assert(_finger != NULL, "invariant");
4214
4215 // Region iteration was actually aborted. So now _finger
4216 // points to the address of the object we last scanned. If we
4217 // leave it there, when we restart this task, we will rescan
4218 // the object. It is easy to avoid this. We move the finger by
4219 // enough to point to the next possible object header (the
4220 // bitmap knows by how much we need to move it as it knows its
4221 // granularity).
4222 assert(_finger < _region_limit, "invariant");
4223 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4224 // Check if bitmap iteration was aborted while scanning the last object
4225 if (new_finger >= _region_limit) {
4226 giveup_current_region();
4227 } else {
4228 move_finger_to(new_finger);
4229 }
4230 }
4231 }
4232 // At this point we have either completed iterating over the
4233 // region we were holding on to, or we have aborted.
4234
4235 // We then partially drain the local queue and the global stack.
4236 // (Do we really need this?)
4237 drain_local_queue(true);
4238 drain_global_stack(true);
4239
4240 // Read the note on the claim_region() method on why it might
4241 // return NULL with potentially more regions available for
4242 // claiming and why we have to check out_of_regions() to determine
4243 // whether we're done or not.
4244 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4245 // We are going to try to claim a new region. We should have
4246 // given up on the previous one.
4247 // Separated the asserts so that we know which one fires.
4248 assert(_curr_region == NULL, "invariant");
4249 assert(_finger == NULL, "invariant");
4250 assert(_region_limit == NULL, "invariant");
4251 if (_cm->verbose_low()) {
4252 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4253 }
4254 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4255 if (claimed_region != NULL) {
4256 // Yes, we managed to claim one
4257 statsOnly( ++_regions_claimed );
4258
4259 if (_cm->verbose_low()) {
4260 gclog_or_tty->print_cr("[%u] we successfully claimed "
4261 "region "PTR_FORMAT,
4262 _worker_id, p2i(claimed_region));
4263 }
4264
4265 setup_for_region(claimed_region);
4266 assert(_curr_region == claimed_region, "invariant");
4267 }
4268 // It is important to call the regular clock here. It might take
4269 // a while to claim a region if, for example, we hit a large
4270 // block of empty regions. So we need to call the regular clock
4271 // method once round the loop to make sure it's called
4272 // frequently enough.
4273 regular_clock_call();
4274 }
4275
4276 if (!has_aborted() && _curr_region == NULL) {
4277 assert(_cm->out_of_regions(),
4278 "at this point we should be out of regions");
4279 }
4280 } while ( _curr_region != NULL && !has_aborted());
4281
4282 if (!has_aborted()) {
4283 // We cannot check whether the global stack is empty, since other
4284 // tasks might be pushing objects to it concurrently.
4285 assert(_cm->out_of_regions(),
4286 "at this point we should be out of regions");
4287
4288 if (_cm->verbose_low()) {
4289 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4290 }
4291
4292 // Try to reduce the number of available SATB buffers so that
4293 // remark has less work to do.
4294 drain_satb_buffers();
4295 }
4296
4297 // Since we've done everything else, we can now totally drain the
4298 // local queue and global stack.
4299 drain_local_queue(false);
4300 drain_global_stack(false);
4301
4302 // Attempt at work stealing from other task's queues.
4303 if (do_stealing && !has_aborted()) {
4304 // We have not aborted. This means that we have finished all that
4305 // we could. Let's try to do some stealing...
4306
4307 // We cannot check whether the global stack is empty, since other
4308 // tasks might be pushing objects to it concurrently.
4309 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4310 "only way to reach here");
4311
4312 if (_cm->verbose_low()) {
4313 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4314 }
4315
4316 while (!has_aborted()) {
4317 oop obj;
4318 statsOnly( ++_steal_attempts );
4319
4320 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4321 if (_cm->verbose_medium()) {
4322 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4323 _worker_id, p2i((void*) obj));
4324 }
4325
4326 statsOnly( ++_steals );
4327
4328 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4329 "any stolen object should be marked");
4330 scan_object(obj);
4331
4332 // And since we're towards the end, let's totally drain the
4333 // local queue and global stack.
4334 drain_local_queue(false);
4335 drain_global_stack(false);
4336 } else {
4337 break;
4338 }
4339 }
4340 }
4341
4342 // If we are about to wrap up and go into termination, check if we
4343 // should raise the overflow flag.
4344 if (do_termination && !has_aborted()) {
4345 if (_cm->force_overflow()->should_force()) {
4346 _cm->set_has_overflown();
4347 regular_clock_call();
4348 }
4349 }
4350
4351 // We still haven't aborted. Now, let's try to get into the
4352 // termination protocol.
4353 if (do_termination && !has_aborted()) {
4354 // We cannot check whether the global stack is empty, since other
4355 // tasks might be concurrently pushing objects on it.
4356 // Separated the asserts so that we know which one fires.
4357 assert(_cm->out_of_regions(), "only way to reach here");
4358 assert(_task_queue->size() == 0, "only way to reach here");
4359
4360 if (_cm->verbose_low()) {
4361 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4362 }
4363
4364 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4365
4366 // The CMTask class also extends the TerminatorTerminator class,
4367 // hence its should_exit_termination() method will also decide
4368 // whether to exit the termination protocol or not.
4369 bool finished = (is_serial ||
4370 _cm->terminator()->offer_termination(this));
4371 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4372 _termination_time_ms +=
4373 termination_end_time_ms - _termination_start_time_ms;
4374
4375 if (finished) {
4376 // We're all done.
4377
4378 if (_worker_id == 0) {
4379 // let's allow task 0 to do this
4380 if (concurrent()) {
4381 assert(_cm->concurrent_marking_in_progress(), "invariant");
4382 // we need to set this to false before the next
4383 // safepoint. This way we ensure that the marking phase
4384 // doesn't observe any more heap expansions.
4385 _cm->clear_concurrent_marking_in_progress();
4386 }
4387 }
4388
4389 // We can now guarantee that the global stack is empty, since
4390 // all other tasks have finished. We separated the guarantees so
4391 // that, if a condition is false, we can immediately find out
4392 // which one.
4393 guarantee(_cm->out_of_regions(), "only way to reach here");
4394 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4395 guarantee(_task_queue->size() == 0, "only way to reach here");
4396 guarantee(!_cm->has_overflown(), "only way to reach here");
4397 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4398
4399 if (_cm->verbose_low()) {
4400 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4401 }
4402 } else {
4403 // Apparently there's more work to do. Let's abort this task. It
4404 // will restart it and we can hopefully find more things to do.
4405
4406 if (_cm->verbose_low()) {
4407 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4408 _worker_id);
4409 }
4410
4411 set_has_aborted();
4412 statsOnly( ++_aborted_termination );
4413 }
4414 }
4415
4416 // Mainly for debugging purposes to make sure that a pointer to the
4417 // closure which was statically allocated in this frame doesn't
4418 // escape it by accident.
4419 set_cm_oop_closure(NULL);
4420 double end_time_ms = os::elapsedVTime() * 1000.0;
4421 double elapsed_time_ms = end_time_ms - _start_time_ms;
4422 // Update the step history.
4423 _step_times_ms.add(elapsed_time_ms);
4424
4425 if (has_aborted()) {
4426 // The task was aborted for some reason.
4427
4428 statsOnly( ++_aborted );
4429
4430 if (_has_timed_out) {
4431 double diff_ms = elapsed_time_ms - _time_target_ms;
4432 // Keep statistics of how well we did with respect to hitting
4433 // our target only if we actually timed out (if we aborted for
4434 // other reasons, then the results might get skewed).
4435 _marking_step_diffs_ms.add(diff_ms);
4436 }
4437
4438 if (_cm->has_overflown()) {
4439 // This is the interesting one. We aborted because a global
4440 // overflow was raised. This means we have to restart the
4441 // marking phase and start iterating over regions. However, in
4442 // order to do this we have to make sure that all tasks stop
4443 // what they are doing and re-initialize in a safe manner. We
4444 // will achieve this with the use of two barrier sync points.
4445
4446 if (_cm->verbose_low()) {
4447 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4448 }
4449
4450 if (!is_serial) {
4451 // We only need to enter the sync barrier if being called
4452 // from a parallel context
4453 _cm->enter_first_sync_barrier(_worker_id);
4454
4455 // When we exit this sync barrier we know that all tasks have
4456 // stopped doing marking work. So, it's now safe to
4457 // re-initialize our data structures. At the end of this method,
4458 // task 0 will clear the global data structures.
4459 }
4460
4461 statsOnly( ++_aborted_overflow );
4462
4463 // We clear the local state of this task...
4464 clear_region_fields();
4465
4466 if (!is_serial) {
4467 // ...and enter the second barrier.
4468 _cm->enter_second_sync_barrier(_worker_id);
4469 }
4470 // At this point, if we're during the concurrent phase of
4471 // marking, everything has been re-initialized and we're
4472 // ready to restart.
4473 }
4474
4475 if (_cm->verbose_low()) {
4476 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4477 "elapsed = %1.2lfms <<<<<<<<<<",
4478 _worker_id, _time_target_ms, elapsed_time_ms);
4479 if (_cm->has_aborted()) {
4480 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4481 _worker_id);
4482 }
4483 }
4484 } else {
4485 if (_cm->verbose_low()) {
4486 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4487 "elapsed = %1.2lfms <<<<<<<<<<",
4488 _worker_id, _time_target_ms, elapsed_time_ms);
4489 }
4490 }
4491
4492 _claimed = false;
4493 }
4494
4495 CMTask::CMTask(uint worker_id,
4496 ConcurrentMark* cm,
4497 size_t* marked_bytes,
4498 BitMap* card_bm,
4499 CMTaskQueue* task_queue,
4500 CMTaskQueueSet* task_queues)
4501 : _g1h(G1CollectedHeap::heap()),
4502 _worker_id(worker_id), _cm(cm),
4503 _claimed(false),
4504 _nextMarkBitMap(NULL), _hash_seed(17),
4505 _task_queue(task_queue),
4506 _task_queues(task_queues),
4507 _cm_oop_closure(NULL),
4508 _marked_bytes_array(marked_bytes),
4509 _card_bm(card_bm) {
4510 guarantee(task_queue != NULL, "invariant");
4511 guarantee(task_queues != NULL, "invariant");
4512
4513 statsOnly( _clock_due_to_scanning = 0;
4514 _clock_due_to_marking = 0 );
4515
4516 _marking_step_diffs_ms.add(0.5);
4517 }
4518
4519 // These are formatting macros that are used below to ensure
4520 // consistent formatting. The *_H_* versions are used to format the
4521 // header for a particular value and they should be kept consistent
4522 // with the corresponding macro. Also note that most of the macros add
4523 // the necessary white space (as a prefix) which makes them a bit
4524 // easier to compose.
4525
4526 // All the output lines are prefixed with this string to be able to
4527 // identify them easily in a large log file.
4528 #define G1PPRL_LINE_PREFIX "###"
4529
4530 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4531 #ifdef _LP64
4532 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4533 #else // _LP64
4534 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4535 #endif // _LP64
4536
4537 // For per-region info
4538 #define G1PPRL_TYPE_FORMAT " %-4s"
4539 #define G1PPRL_TYPE_H_FORMAT " %4s"
4540 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4541 #define G1PPRL_BYTE_H_FORMAT " %9s"
4542 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4543 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4544
4545 // For summary info
4546 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4547 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4548 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4549 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4550
4551 G1PrintRegionLivenessInfoClosure::
4552 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4553 : _out(out),
4554 _total_used_bytes(0), _total_capacity_bytes(0),
4555 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4556 _hum_used_bytes(0), _hum_capacity_bytes(0),
4557 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4558 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4559 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4560 MemRegion g1_reserved = g1h->g1_reserved();
4561 double now = os::elapsedTime();
4562
4563 // Print the header of the output.
4564 _out->cr();
4565 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4566 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4567 G1PPRL_SUM_ADDR_FORMAT("reserved")
4568 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4569 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4570 HeapRegion::GrainBytes);
4571 _out->print_cr(G1PPRL_LINE_PREFIX);
4572 _out->print_cr(G1PPRL_LINE_PREFIX
4573 G1PPRL_TYPE_H_FORMAT
4574 G1PPRL_ADDR_BASE_H_FORMAT
4575 G1PPRL_BYTE_H_FORMAT
4576 G1PPRL_BYTE_H_FORMAT
4577 G1PPRL_BYTE_H_FORMAT
4578 G1PPRL_DOUBLE_H_FORMAT
4579 G1PPRL_BYTE_H_FORMAT
4580 G1PPRL_BYTE_H_FORMAT,
4581 "type", "address-range",
4582 "used", "prev-live", "next-live", "gc-eff",
4583 "remset", "code-roots");
4584 _out->print_cr(G1PPRL_LINE_PREFIX
4585 G1PPRL_TYPE_H_FORMAT
4586 G1PPRL_ADDR_BASE_H_FORMAT
4587 G1PPRL_BYTE_H_FORMAT
4588 G1PPRL_BYTE_H_FORMAT
4589 G1PPRL_BYTE_H_FORMAT
4590 G1PPRL_DOUBLE_H_FORMAT
4591 G1PPRL_BYTE_H_FORMAT
4592 G1PPRL_BYTE_H_FORMAT,
4593 "", "",
4594 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4595 "(bytes)", "(bytes)");
4596 }
4597
4598 // It takes as a parameter a reference to one of the _hum_* fields, it
4599 // deduces the corresponding value for a region in a humongous region
4600 // series (either the region size, or what's left if the _hum_* field
4601 // is < the region size), and updates the _hum_* field accordingly.
4602 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4603 size_t bytes = 0;
4604 // The > 0 check is to deal with the prev and next live bytes which
4605 // could be 0.
4606 if (*hum_bytes > 0) {
4607 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4608 *hum_bytes -= bytes;
4609 }
4610 return bytes;
4611 }
4612
4613 // It deduces the values for a region in a humongous region series
4614 // from the _hum_* fields and updates those accordingly. It assumes
4615 // that that _hum_* fields have already been set up from the "starts
4616 // humongous" region and we visit the regions in address order.
4617 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4618 size_t* capacity_bytes,
4619 size_t* prev_live_bytes,
4620 size_t* next_live_bytes) {
4621 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4622 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4623 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4624 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4625 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4626 }
4627
4628 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4629 const char* type = r->get_type_str();
4630 HeapWord* bottom = r->bottom();
4631 HeapWord* end = r->end();
4632 size_t capacity_bytes = r->capacity();
4633 size_t used_bytes = r->used();
4634 size_t prev_live_bytes = r->live_bytes();
4635 size_t next_live_bytes = r->next_live_bytes();
4636 double gc_eff = r->gc_efficiency();
4637 size_t remset_bytes = r->rem_set()->mem_size();
4638 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4639
4640 if (r->is_starts_humongous()) {
4641 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4642 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4643 "they should have been zeroed after the last time we used them");
4644 // Set up the _hum_* fields.
4645 _hum_capacity_bytes = capacity_bytes;
4646 _hum_used_bytes = used_bytes;
4647 _hum_prev_live_bytes = prev_live_bytes;
4648 _hum_next_live_bytes = next_live_bytes;
4649 get_hum_bytes(&used_bytes, &capacity_bytes,
4650 &prev_live_bytes, &next_live_bytes);
4651 end = bottom + HeapRegion::GrainWords;
4652 } else if (r->is_continues_humongous()) {
4653 get_hum_bytes(&used_bytes, &capacity_bytes,
4654 &prev_live_bytes, &next_live_bytes);
4655 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4656 }
4657
4658 _total_used_bytes += used_bytes;
4659 _total_capacity_bytes += capacity_bytes;
4660 _total_prev_live_bytes += prev_live_bytes;
4661 _total_next_live_bytes += next_live_bytes;
4662 _total_remset_bytes += remset_bytes;
4663 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4664
4665 // Print a line for this particular region.
4666 _out->print_cr(G1PPRL_LINE_PREFIX
4667 G1PPRL_TYPE_FORMAT
4668 G1PPRL_ADDR_BASE_FORMAT
4669 G1PPRL_BYTE_FORMAT
4670 G1PPRL_BYTE_FORMAT
4671 G1PPRL_BYTE_FORMAT
4672 G1PPRL_DOUBLE_FORMAT
4673 G1PPRL_BYTE_FORMAT
4674 G1PPRL_BYTE_FORMAT,
4675 type, p2i(bottom), p2i(end),
4676 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4677 remset_bytes, strong_code_roots_bytes);
4678
4679 return false;
4680 }
4681
4682 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4683 // add static memory usages to remembered set sizes
4684 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4685 // Print the footer of the output.
4686 _out->print_cr(G1PPRL_LINE_PREFIX);
4687 _out->print_cr(G1PPRL_LINE_PREFIX
4688 " SUMMARY"
4689 G1PPRL_SUM_MB_FORMAT("capacity")
4690 G1PPRL_SUM_MB_PERC_FORMAT("used")
4691 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4692 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4693 G1PPRL_SUM_MB_FORMAT("remset")
4694 G1PPRL_SUM_MB_FORMAT("code-roots"),
4695 bytes_to_mb(_total_capacity_bytes),
4696 bytes_to_mb(_total_used_bytes),
4697 perc(_total_used_bytes, _total_capacity_bytes),
4698 bytes_to_mb(_total_prev_live_bytes),
4699 perc(_total_prev_live_bytes, _total_capacity_bytes),
4700 bytes_to_mb(_total_next_live_bytes),
4701 perc(_total_next_live_bytes, _total_capacity_bytes),
4702 bytes_to_mb(_total_remset_bytes),
4703 bytes_to_mb(_total_strong_code_roots_bytes));
4704 _out->cr();
4705 }