rev 8048 : 8058354: SPECjvm2008-Derby -2.7% performance regression on Solaris-X64 starting with 9-b29
Summary: Allow partial use of large pages for auxiliary data structures in G1.
Reviewed-by: jmasa
1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/metadataOnStackMark.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/g1/concurrentMark.inline.hpp"
30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
32 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
33 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
34 #include "gc_implementation/g1/g1Log.hpp"
35 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
36 #include "gc_implementation/g1/g1RemSet.hpp"
37 #include "gc_implementation/g1/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 Marker",
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 size_t mark_stack_size =
700 MIN2(MarkStackSizeMax,
701 MAX2(MarkStackSize, (size_t) (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 (" SIZE_FORMAT "): "
706 "must be between 1 and " SIZE_FORMAT,
707 mark_stack_size, MarkStackSizeMax);
708 return;
709 }
710 FLAG_SET_ERGO(size_t, 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 (" SIZE_FORMAT "): "
717 "must be between 1 and " SIZE_FORMAT,
718 MarkStackSize, MarkStackSizeMax);
719 return;
720 }
721 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
722 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
723 warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
724 " or for MarkStackSizeMax (" SIZE_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 uint max_regions = _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(barrier_set_cast<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 FreeRegionList local_cleanup_list("Local Cleanup List");
1914 HRRSCleanupTask hrrs_cleanup_task;
1915 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1916 &hrrs_cleanup_task);
1917 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1918 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1919
1920 // Now update the lists
1921 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1922 {
1923 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1924 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1925 _max_live_bytes += g1_note_end.max_live_bytes();
1926 _freed_bytes += g1_note_end.freed_bytes();
1927
1928 // If we iterate over the global cleanup list at the end of
1929 // cleanup to do this printing we will not guarantee to only
1930 // generate output for the newly-reclaimed regions (the list
1931 // might not be empty at the beginning of cleanup; we might
1932 // still be working on its previous contents). So we do the
1933 // printing here, before we append the new regions to the global
1934 // cleanup list.
1935
1936 G1HRPrinter* hr_printer = _g1h->hr_printer();
1937 if (hr_printer->is_active()) {
1938 FreeRegionListIterator iter(&local_cleanup_list);
1939 while (iter.more_available()) {
1940 HeapRegion* hr = iter.get_next();
1941 hr_printer->cleanup(hr);
1942 }
1943 }
1944
1945 _cleanup_list->add_ordered(&local_cleanup_list);
1946 assert(local_cleanup_list.is_empty(), "post-condition");
1947
1948 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1949 }
1950 }
1951 size_t max_live_bytes() { return _max_live_bytes; }
1952 size_t freed_bytes() { return _freed_bytes; }
1953 };
1954
1955 class G1ParScrubRemSetTask: public AbstractGangTask {
1956 protected:
1957 G1RemSet* _g1rs;
1958 BitMap* _region_bm;
1959 BitMap* _card_bm;
1960 HeapRegionClaimer _hrclaimer;
1961
1962 public:
1963 G1ParScrubRemSetTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm, uint n_workers) :
1964 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), _region_bm(region_bm), _card_bm(card_bm), _hrclaimer(n_workers) {
1965 }
1966
1967 void work(uint worker_id) {
1968 _g1rs->scrub(_region_bm, _card_bm, worker_id, &_hrclaimer);
1969 }
1970
1971 };
1972
1973 void ConcurrentMark::cleanup() {
1974 // world is stopped at this checkpoint
1975 assert(SafepointSynchronize::is_at_safepoint(),
1976 "world should be stopped");
1977 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1978
1979 // If a full collection has happened, we shouldn't do this.
1980 if (has_aborted()) {
1981 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1982 return;
1983 }
1984
1985 g1h->verify_region_sets_optional();
1986
1987 if (VerifyDuringGC) {
1988 HandleMark hm; // handle scope
1989 Universe::heap()->prepare_for_verify();
1990 Universe::verify(VerifyOption_G1UsePrevMarking,
1991 " VerifyDuringGC:(before)");
1992 }
1993 g1h->check_bitmaps("Cleanup Start");
1994
1995 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
1996 g1p->record_concurrent_mark_cleanup_start();
1997
1998 double start = os::elapsedTime();
1999
2000 HeapRegionRemSet::reset_for_cleanup_tasks();
2001
2002 uint n_workers;
2003
2004 // Do counting once more with the world stopped for good measure.
2005 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2006
2007 g1h->set_par_threads();
2008 n_workers = g1h->n_par_threads();
2009 assert(g1h->n_par_threads() == n_workers,
2010 "Should not have been reset");
2011 g1h->workers()->run_task(&g1_par_count_task);
2012 // Done with the parallel phase so reset to 0.
2013 g1h->set_par_threads(0);
2014
2015 if (VerifyDuringGC) {
2016 // Verify that the counting data accumulated during marking matches
2017 // that calculated by walking the marking bitmap.
2018
2019 // Bitmaps to hold expected values
2020 BitMap expected_region_bm(_region_bm.size(), true);
2021 BitMap expected_card_bm(_card_bm.size(), true);
2022
2023 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2024 &_region_bm,
2025 &_card_bm,
2026 &expected_region_bm,
2027 &expected_card_bm);
2028
2029 g1h->set_par_threads((int)n_workers);
2030 g1h->workers()->run_task(&g1_par_verify_task);
2031 // Done with the parallel phase so reset to 0.
2032 g1h->set_par_threads(0);
2033
2034 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2035 }
2036
2037 size_t start_used_bytes = g1h->used();
2038 g1h->set_marking_complete();
2039
2040 double count_end = os::elapsedTime();
2041 double this_final_counting_time = (count_end - start);
2042 _total_counting_time += this_final_counting_time;
2043
2044 if (G1PrintRegionLivenessInfo) {
2045 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2046 _g1h->heap_region_iterate(&cl);
2047 }
2048
2049 // Install newly created mark bitMap as "prev".
2050 swapMarkBitMaps();
2051
2052 g1h->reset_gc_time_stamp();
2053
2054 // Note end of marking in all heap regions.
2055 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
2056 g1h->set_par_threads((int)n_workers);
2057 g1h->workers()->run_task(&g1_par_note_end_task);
2058 g1h->set_par_threads(0);
2059 g1h->check_gc_time_stamps();
2060
2061 if (!cleanup_list_is_empty()) {
2062 // The cleanup list is not empty, so we'll have to process it
2063 // concurrently. Notify anyone else that might be wanting free
2064 // regions that there will be more free regions coming soon.
2065 g1h->set_free_regions_coming();
2066 }
2067
2068 // call below, since it affects the metric by which we sort the heap
2069 // regions.
2070 if (G1ScrubRemSets) {
2071 double rs_scrub_start = os::elapsedTime();
2072 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm, n_workers);
2073 g1h->set_par_threads((int)n_workers);
2074 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2075 g1h->set_par_threads(0);
2076
2077 double rs_scrub_end = os::elapsedTime();
2078 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2079 _total_rs_scrub_time += this_rs_scrub_time;
2080 }
2081
2082 // this will also free any regions totally full of garbage objects,
2083 // and sort the regions.
2084 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2085
2086 // Statistics.
2087 double end = os::elapsedTime();
2088 _cleanup_times.add((end - start) * 1000.0);
2089
2090 if (G1Log::fine()) {
2091 g1h->g1_policy()->print_heap_transition(start_used_bytes);
2092 }
2093
2094 // Clean up will have freed any regions completely full of garbage.
2095 // Update the soft reference policy with the new heap occupancy.
2096 Universe::update_heap_info_at_gc();
2097
2098 if (VerifyDuringGC) {
2099 HandleMark hm; // handle scope
2100 Universe::heap()->prepare_for_verify();
2101 Universe::verify(VerifyOption_G1UsePrevMarking,
2102 " VerifyDuringGC:(after)");
2103 }
2104
2105 g1h->check_bitmaps("Cleanup End");
2106
2107 g1h->verify_region_sets_optional();
2108
2109 // We need to make this be a "collection" so any collection pause that
2110 // races with it goes around and waits for completeCleanup to finish.
2111 g1h->increment_total_collections();
2112
2113 // Clean out dead classes and update Metaspace sizes.
2114 if (ClassUnloadingWithConcurrentMark) {
2115 ClassLoaderDataGraph::purge();
2116 }
2117 MetaspaceGC::compute_new_size();
2118
2119 // We reclaimed old regions so we should calculate the sizes to make
2120 // sure we update the old gen/space data.
2121 g1h->g1mm()->update_sizes();
2122 g1h->allocation_context_stats().update_after_mark();
2123
2124 g1h->trace_heap_after_concurrent_cycle();
2125 }
2126
2127 void ConcurrentMark::completeCleanup() {
2128 if (has_aborted()) return;
2129
2130 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2131
2132 _cleanup_list.verify_optional();
2133 FreeRegionList tmp_free_list("Tmp Free List");
2134
2135 if (G1ConcRegionFreeingVerbose) {
2136 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2137 "cleanup list has %u entries",
2138 _cleanup_list.length());
2139 }
2140
2141 // No one else should be accessing the _cleanup_list at this point,
2142 // so it is not necessary to take any locks
2143 while (!_cleanup_list.is_empty()) {
2144 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2145 assert(hr != NULL, "Got NULL from a non-empty list");
2146 hr->par_clear();
2147 tmp_free_list.add_ordered(hr);
2148
2149 // Instead of adding one region at a time to the secondary_free_list,
2150 // we accumulate them in the local list and move them a few at a
2151 // time. This also cuts down on the number of notify_all() calls
2152 // we do during this process. We'll also append the local list when
2153 // _cleanup_list is empty (which means we just removed the last
2154 // region from the _cleanup_list).
2155 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2156 _cleanup_list.is_empty()) {
2157 if (G1ConcRegionFreeingVerbose) {
2158 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2159 "appending %u entries to the secondary_free_list, "
2160 "cleanup list still has %u entries",
2161 tmp_free_list.length(),
2162 _cleanup_list.length());
2163 }
2164
2165 {
2166 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2167 g1h->secondary_free_list_add(&tmp_free_list);
2168 SecondaryFreeList_lock->notify_all();
2169 }
2170 #ifndef PRODUCT
2171 if (G1StressConcRegionFreeing) {
2172 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2173 os::sleep(Thread::current(), (jlong) 1, false);
2174 }
2175 }
2176 #endif
2177 }
2178 }
2179 assert(tmp_free_list.is_empty(), "post-condition");
2180 }
2181
2182 // Supporting Object and Oop closures for reference discovery
2183 // and processing in during marking
2184
2185 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2186 HeapWord* addr = (HeapWord*)obj;
2187 return addr != NULL &&
2188 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2189 }
2190
2191 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2192 // Uses the CMTask associated with a worker thread (for serial reference
2193 // processing the CMTask for worker 0 is used) to preserve (mark) and
2194 // trace referent objects.
2195 //
2196 // Using the CMTask and embedded local queues avoids having the worker
2197 // threads operating on the global mark stack. This reduces the risk
2198 // of overflowing the stack - which we would rather avoid at this late
2199 // state. Also using the tasks' local queues removes the potential
2200 // of the workers interfering with each other that could occur if
2201 // operating on the global stack.
2202
2203 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2204 ConcurrentMark* _cm;
2205 CMTask* _task;
2206 int _ref_counter_limit;
2207 int _ref_counter;
2208 bool _is_serial;
2209 public:
2210 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2211 _cm(cm), _task(task), _is_serial(is_serial),
2212 _ref_counter_limit(G1RefProcDrainInterval) {
2213 assert(_ref_counter_limit > 0, "sanity");
2214 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2215 _ref_counter = _ref_counter_limit;
2216 }
2217
2218 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2219 virtual void do_oop( oop* p) { do_oop_work(p); }
2220
2221 template <class T> void do_oop_work(T* p) {
2222 if (!_cm->has_overflown()) {
2223 oop obj = oopDesc::load_decode_heap_oop(p);
2224 if (_cm->verbose_high()) {
2225 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2226 "*"PTR_FORMAT" = "PTR_FORMAT,
2227 _task->worker_id(), p2i(p), p2i((void*) obj));
2228 }
2229
2230 _task->deal_with_reference(obj);
2231 _ref_counter--;
2232
2233 if (_ref_counter == 0) {
2234 // We have dealt with _ref_counter_limit references, pushing them
2235 // and objects reachable from them on to the local stack (and
2236 // possibly the global stack). Call CMTask::do_marking_step() to
2237 // process these entries.
2238 //
2239 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2240 // there's nothing more to do (i.e. we're done with the entries that
2241 // were pushed as a result of the CMTask::deal_with_reference() calls
2242 // above) or we overflow.
2243 //
2244 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2245 // flag while there may still be some work to do. (See the comment at
2246 // the beginning of CMTask::do_marking_step() for those conditions -
2247 // one of which is reaching the specified time target.) It is only
2248 // when CMTask::do_marking_step() returns without setting the
2249 // has_aborted() flag that the marking step has completed.
2250 do {
2251 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2252 _task->do_marking_step(mark_step_duration_ms,
2253 false /* do_termination */,
2254 _is_serial);
2255 } while (_task->has_aborted() && !_cm->has_overflown());
2256 _ref_counter = _ref_counter_limit;
2257 }
2258 } else {
2259 if (_cm->verbose_high()) {
2260 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2261 }
2262 }
2263 }
2264 };
2265
2266 // 'Drain' oop closure used by both serial and parallel reference processing.
2267 // Uses the CMTask associated with a given worker thread (for serial
2268 // reference processing the CMtask for worker 0 is used). Calls the
2269 // do_marking_step routine, with an unbelievably large timeout value,
2270 // to drain the marking data structures of the remaining entries
2271 // added by the 'keep alive' oop closure above.
2272
2273 class G1CMDrainMarkingStackClosure: public VoidClosure {
2274 ConcurrentMark* _cm;
2275 CMTask* _task;
2276 bool _is_serial;
2277 public:
2278 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2279 _cm(cm), _task(task), _is_serial(is_serial) {
2280 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2281 }
2282
2283 void do_void() {
2284 do {
2285 if (_cm->verbose_high()) {
2286 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2287 _task->worker_id(), BOOL_TO_STR(_is_serial));
2288 }
2289
2290 // We call CMTask::do_marking_step() to completely drain the local
2291 // and global marking stacks of entries pushed by the 'keep alive'
2292 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2293 //
2294 // CMTask::do_marking_step() is called in a loop, which we'll exit
2295 // if there's nothing more to do (i.e. we've completely drained the
2296 // entries that were pushed as a a result of applying the 'keep alive'
2297 // closure to the entries on the discovered ref lists) or we overflow
2298 // the global marking stack.
2299 //
2300 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2301 // flag while there may still be some work to do. (See the comment at
2302 // the beginning of CMTask::do_marking_step() for those conditions -
2303 // one of which is reaching the specified time target.) It is only
2304 // when CMTask::do_marking_step() returns without setting the
2305 // has_aborted() flag that the marking step has completed.
2306
2307 _task->do_marking_step(1000000000.0 /* something very large */,
2308 true /* do_termination */,
2309 _is_serial);
2310 } while (_task->has_aborted() && !_cm->has_overflown());
2311 }
2312 };
2313
2314 // Implementation of AbstractRefProcTaskExecutor for parallel
2315 // reference processing at the end of G1 concurrent marking
2316
2317 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2318 private:
2319 G1CollectedHeap* _g1h;
2320 ConcurrentMark* _cm;
2321 WorkGang* _workers;
2322 int _active_workers;
2323
2324 public:
2325 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2326 ConcurrentMark* cm,
2327 WorkGang* workers,
2328 int n_workers) :
2329 _g1h(g1h), _cm(cm),
2330 _workers(workers), _active_workers(n_workers) { }
2331
2332 // Executes the given task using concurrent marking worker threads.
2333 virtual void execute(ProcessTask& task);
2334 virtual void execute(EnqueueTask& task);
2335 };
2336
2337 class G1CMRefProcTaskProxy: public AbstractGangTask {
2338 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2339 ProcessTask& _proc_task;
2340 G1CollectedHeap* _g1h;
2341 ConcurrentMark* _cm;
2342
2343 public:
2344 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2345 G1CollectedHeap* g1h,
2346 ConcurrentMark* cm) :
2347 AbstractGangTask("Process reference objects in parallel"),
2348 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2349 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2350 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2351 }
2352
2353 virtual void work(uint worker_id) {
2354 ResourceMark rm;
2355 HandleMark hm;
2356 CMTask* task = _cm->task(worker_id);
2357 G1CMIsAliveClosure g1_is_alive(_g1h);
2358 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2359 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2360
2361 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2362 }
2363 };
2364
2365 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2366 assert(_workers != NULL, "Need parallel worker threads.");
2367 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2368
2369 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2370
2371 // We need to reset the concurrency level before each
2372 // proxy task execution, so that the termination protocol
2373 // and overflow handling in CMTask::do_marking_step() knows
2374 // how many workers to wait for.
2375 _cm->set_concurrency(_active_workers);
2376 _g1h->set_par_threads(_active_workers);
2377 _workers->run_task(&proc_task_proxy);
2378 _g1h->set_par_threads(0);
2379 }
2380
2381 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2382 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2383 EnqueueTask& _enq_task;
2384
2385 public:
2386 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2387 AbstractGangTask("Enqueue reference objects in parallel"),
2388 _enq_task(enq_task) { }
2389
2390 virtual void work(uint worker_id) {
2391 _enq_task.work(worker_id);
2392 }
2393 };
2394
2395 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2396 assert(_workers != NULL, "Need parallel worker threads.");
2397 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2398
2399 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2400
2401 // Not strictly necessary but...
2402 //
2403 // We need to reset the concurrency level before each
2404 // proxy task execution, so that the termination protocol
2405 // and overflow handling in CMTask::do_marking_step() knows
2406 // how many workers to wait for.
2407 _cm->set_concurrency(_active_workers);
2408 _g1h->set_par_threads(_active_workers);
2409 _workers->run_task(&enq_task_proxy);
2410 _g1h->set_par_threads(0);
2411 }
2412
2413 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2414 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2415 }
2416
2417 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2418 if (has_overflown()) {
2419 // Skip processing the discovered references if we have
2420 // overflown the global marking stack. Reference objects
2421 // only get discovered once so it is OK to not
2422 // de-populate the discovered reference lists. We could have,
2423 // but the only benefit would be that, when marking restarts,
2424 // less reference objects are discovered.
2425 return;
2426 }
2427
2428 ResourceMark rm;
2429 HandleMark hm;
2430
2431 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2432
2433 // Is alive closure.
2434 G1CMIsAliveClosure g1_is_alive(g1h);
2435
2436 // Inner scope to exclude the cleaning of the string and symbol
2437 // tables from the displayed time.
2438 {
2439 G1CMTraceTime t("GC ref-proc", G1Log::finer());
2440
2441 ReferenceProcessor* rp = g1h->ref_processor_cm();
2442
2443 // See the comment in G1CollectedHeap::ref_processing_init()
2444 // about how reference processing currently works in G1.
2445
2446 // Set the soft reference policy
2447 rp->setup_policy(clear_all_soft_refs);
2448 assert(_markStack.isEmpty(), "mark stack should be empty");
2449
2450 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2451 // in serial reference processing. Note these closures are also
2452 // used for serially processing (by the the current thread) the
2453 // JNI references during parallel reference processing.
2454 //
2455 // These closures do not need to synchronize with the worker
2456 // threads involved in parallel reference processing as these
2457 // instances are executed serially by the current thread (e.g.
2458 // reference processing is not multi-threaded and is thus
2459 // performed by the current thread instead of a gang worker).
2460 //
2461 // The gang tasks involved in parallel reference processing create
2462 // their own instances of these closures, which do their own
2463 // synchronization among themselves.
2464 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2465 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2466
2467 // We need at least one active thread. If reference processing
2468 // is not multi-threaded we use the current (VMThread) thread,
2469 // otherwise we use the work gang from the G1CollectedHeap and
2470 // we utilize all the worker threads we can.
2471 bool processing_is_mt = rp->processing_is_mt();
2472 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2473 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2474
2475 // Parallel processing task executor.
2476 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2477 g1h->workers(), active_workers);
2478 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2479
2480 // Set the concurrency level. The phase was already set prior to
2481 // executing the remark task.
2482 set_concurrency(active_workers);
2483
2484 // Set the degree of MT processing here. If the discovery was done MT,
2485 // the number of threads involved during discovery could differ from
2486 // the number of active workers. This is OK as long as the discovered
2487 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2488 rp->set_active_mt_degree(active_workers);
2489
2490 // Process the weak references.
2491 const ReferenceProcessorStats& stats =
2492 rp->process_discovered_references(&g1_is_alive,
2493 &g1_keep_alive,
2494 &g1_drain_mark_stack,
2495 executor,
2496 g1h->gc_timer_cm(),
2497 concurrent_gc_id());
2498 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2499
2500 // The do_oop work routines of the keep_alive and drain_marking_stack
2501 // oop closures will set the has_overflown flag if we overflow the
2502 // global marking stack.
2503
2504 assert(_markStack.overflow() || _markStack.isEmpty(),
2505 "mark stack should be empty (unless it overflowed)");
2506
2507 if (_markStack.overflow()) {
2508 // This should have been done already when we tried to push an
2509 // entry on to the global mark stack. But let's do it again.
2510 set_has_overflown();
2511 }
2512
2513 assert(rp->num_q() == active_workers, "why not");
2514
2515 rp->enqueue_discovered_references(executor);
2516
2517 rp->verify_no_references_recorded();
2518 assert(!rp->discovery_enabled(), "Post condition");
2519 }
2520
2521 if (has_overflown()) {
2522 // We can not trust g1_is_alive if the marking stack overflowed
2523 return;
2524 }
2525
2526 assert(_markStack.isEmpty(), "Marking should have completed");
2527
2528 // Unload Klasses, String, Symbols, Code Cache, etc.
2529 {
2530 G1CMTraceTime trace("Unloading", G1Log::finer());
2531
2532 if (ClassUnloadingWithConcurrentMark) {
2533 bool purged_classes;
2534
2535 {
2536 G1CMTraceTime trace("System Dictionary Unloading", G1Log::finest());
2537 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2538 }
2539
2540 {
2541 G1CMTraceTime trace("Parallel Unloading", G1Log::finest());
2542 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2543 }
2544 }
2545
2546 if (G1StringDedup::is_enabled()) {
2547 G1CMTraceTime trace("String Deduplication Unlink", G1Log::finest());
2548 G1StringDedup::unlink(&g1_is_alive);
2549 }
2550 }
2551 }
2552
2553 void ConcurrentMark::swapMarkBitMaps() {
2554 CMBitMapRO* temp = _prevMarkBitMap;
2555 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2556 _nextMarkBitMap = (CMBitMap*) temp;
2557 }
2558
2559 class CMObjectClosure;
2560
2561 // Closure for iterating over objects, currently only used for
2562 // processing SATB buffers.
2563 class CMObjectClosure : public ObjectClosure {
2564 private:
2565 CMTask* _task;
2566
2567 public:
2568 void do_object(oop obj) {
2569 _task->deal_with_reference(obj);
2570 }
2571
2572 CMObjectClosure(CMTask* task) : _task(task) { }
2573 };
2574
2575 class G1RemarkThreadsClosure : public ThreadClosure {
2576 CMObjectClosure _cm_obj;
2577 G1CMOopClosure _cm_cl;
2578 MarkingCodeBlobClosure _code_cl;
2579 int _thread_parity;
2580
2581 public:
2582 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task) :
2583 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2584 _thread_parity(Threads::thread_claim_parity()) {}
2585
2586 void do_thread(Thread* thread) {
2587 if (thread->is_Java_thread()) {
2588 if (thread->claim_oops_do(true, _thread_parity)) {
2589 JavaThread* jt = (JavaThread*)thread;
2590
2591 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2592 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2593 // * Alive if on the stack of an executing method
2594 // * Weakly reachable otherwise
2595 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2596 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2597 jt->nmethods_do(&_code_cl);
2598
2599 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2600 }
2601 } else if (thread->is_VM_thread()) {
2602 if (thread->claim_oops_do(true, _thread_parity)) {
2603 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2604 }
2605 }
2606 }
2607 };
2608
2609 class CMRemarkTask: public AbstractGangTask {
2610 private:
2611 ConcurrentMark* _cm;
2612 public:
2613 void work(uint worker_id) {
2614 // Since all available tasks are actually started, we should
2615 // only proceed if we're supposed to be active.
2616 if (worker_id < _cm->active_tasks()) {
2617 CMTask* task = _cm->task(worker_id);
2618 task->record_start_time();
2619 {
2620 ResourceMark rm;
2621 HandleMark hm;
2622
2623 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
2624 Threads::threads_do(&threads_f);
2625 }
2626
2627 do {
2628 task->do_marking_step(1000000000.0 /* something very large */,
2629 true /* do_termination */,
2630 false /* is_serial */);
2631 } while (task->has_aborted() && !_cm->has_overflown());
2632 // If we overflow, then we do not want to restart. We instead
2633 // want to abort remark and do concurrent marking again.
2634 task->record_end_time();
2635 }
2636 }
2637
2638 CMRemarkTask(ConcurrentMark* cm, int active_workers) :
2639 AbstractGangTask("Par Remark"), _cm(cm) {
2640 _cm->terminator()->reset_for_reuse(active_workers);
2641 }
2642 };
2643
2644 void ConcurrentMark::checkpointRootsFinalWork() {
2645 ResourceMark rm;
2646 HandleMark hm;
2647 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2648
2649 G1CMTraceTime trace("Finalize Marking", G1Log::finer());
2650
2651 g1h->ensure_parsability(false);
2652
2653 G1CollectedHeap::StrongRootsScope srs(g1h);
2654 // this is remark, so we'll use up all active threads
2655 uint active_workers = g1h->workers()->active_workers();
2656 if (active_workers == 0) {
2657 assert(active_workers > 0, "Should have been set earlier");
2658 active_workers = (uint) ParallelGCThreads;
2659 g1h->workers()->set_active_workers(active_workers);
2660 }
2661 set_concurrency_and_phase(active_workers, false /* concurrent */);
2662 // Leave _parallel_marking_threads at it's
2663 // value originally calculated in the ConcurrentMark
2664 // constructor and pass values of the active workers
2665 // through the gang in the task.
2666
2667 CMRemarkTask remarkTask(this, active_workers);
2668 // We will start all available threads, even if we decide that the
2669 // active_workers will be fewer. The extra ones will just bail out
2670 // immediately.
2671 g1h->set_par_threads(active_workers);
2672 g1h->workers()->run_task(&remarkTask);
2673 g1h->set_par_threads(0);
2674
2675 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2676 guarantee(has_overflown() ||
2677 satb_mq_set.completed_buffers_num() == 0,
2678 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2679 BOOL_TO_STR(has_overflown()),
2680 satb_mq_set.completed_buffers_num()));
2681
2682 print_stats();
2683 }
2684
2685 #ifndef PRODUCT
2686
2687 class PrintReachableOopClosure: public OopClosure {
2688 private:
2689 G1CollectedHeap* _g1h;
2690 outputStream* _out;
2691 VerifyOption _vo;
2692 bool _all;
2693
2694 public:
2695 PrintReachableOopClosure(outputStream* out,
2696 VerifyOption vo,
2697 bool all) :
2698 _g1h(G1CollectedHeap::heap()),
2699 _out(out), _vo(vo), _all(all) { }
2700
2701 void do_oop(narrowOop* p) { do_oop_work(p); }
2702 void do_oop( oop* p) { do_oop_work(p); }
2703
2704 template <class T> void do_oop_work(T* p) {
2705 oop obj = oopDesc::load_decode_heap_oop(p);
2706 const char* str = NULL;
2707 const char* str2 = "";
2708
2709 if (obj == NULL) {
2710 str = "";
2711 } else if (!_g1h->is_in_g1_reserved(obj)) {
2712 str = " O";
2713 } else {
2714 HeapRegion* hr = _g1h->heap_region_containing(obj);
2715 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2716 bool marked = _g1h->is_marked(obj, _vo);
2717
2718 if (over_tams) {
2719 str = " >";
2720 if (marked) {
2721 str2 = " AND MARKED";
2722 }
2723 } else if (marked) {
2724 str = " M";
2725 } else {
2726 str = " NOT";
2727 }
2728 }
2729
2730 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2731 p2i(p), p2i((void*) obj), str, str2);
2732 }
2733 };
2734
2735 class PrintReachableObjectClosure : public ObjectClosure {
2736 private:
2737 G1CollectedHeap* _g1h;
2738 outputStream* _out;
2739 VerifyOption _vo;
2740 bool _all;
2741 HeapRegion* _hr;
2742
2743 public:
2744 PrintReachableObjectClosure(outputStream* out,
2745 VerifyOption vo,
2746 bool all,
2747 HeapRegion* hr) :
2748 _g1h(G1CollectedHeap::heap()),
2749 _out(out), _vo(vo), _all(all), _hr(hr) { }
2750
2751 void do_object(oop o) {
2752 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2753 bool marked = _g1h->is_marked(o, _vo);
2754 bool print_it = _all || over_tams || marked;
2755
2756 if (print_it) {
2757 _out->print_cr(" "PTR_FORMAT"%s",
2758 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2759 PrintReachableOopClosure oopCl(_out, _vo, _all);
2760 o->oop_iterate_no_header(&oopCl);
2761 }
2762 }
2763 };
2764
2765 class PrintReachableRegionClosure : public HeapRegionClosure {
2766 private:
2767 G1CollectedHeap* _g1h;
2768 outputStream* _out;
2769 VerifyOption _vo;
2770 bool _all;
2771
2772 public:
2773 bool doHeapRegion(HeapRegion* hr) {
2774 HeapWord* b = hr->bottom();
2775 HeapWord* e = hr->end();
2776 HeapWord* t = hr->top();
2777 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2778 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2779 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2780 _out->cr();
2781
2782 HeapWord* from = b;
2783 HeapWord* to = t;
2784
2785 if (to > from) {
2786 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2787 _out->cr();
2788 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2789 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2790 _out->cr();
2791 }
2792
2793 return false;
2794 }
2795
2796 PrintReachableRegionClosure(outputStream* out,
2797 VerifyOption vo,
2798 bool all) :
2799 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2800 };
2801
2802 void ConcurrentMark::print_reachable(const char* str,
2803 VerifyOption vo,
2804 bool all) {
2805 gclog_or_tty->cr();
2806 gclog_or_tty->print_cr("== Doing heap dump... ");
2807
2808 if (G1PrintReachableBaseFile == NULL) {
2809 gclog_or_tty->print_cr(" #### error: no base file defined");
2810 return;
2811 }
2812
2813 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2814 (JVM_MAXPATHLEN - 1)) {
2815 gclog_or_tty->print_cr(" #### error: file name too long");
2816 return;
2817 }
2818
2819 char file_name[JVM_MAXPATHLEN];
2820 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2821 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2822
2823 fileStream fout(file_name);
2824 if (!fout.is_open()) {
2825 gclog_or_tty->print_cr(" #### error: could not open file");
2826 return;
2827 }
2828
2829 outputStream* out = &fout;
2830 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2831 out->cr();
2832
2833 out->print_cr("--- ITERATING OVER REGIONS");
2834 out->cr();
2835 PrintReachableRegionClosure rcl(out, vo, all);
2836 _g1h->heap_region_iterate(&rcl);
2837 out->cr();
2838
2839 gclog_or_tty->print_cr(" done");
2840 gclog_or_tty->flush();
2841 }
2842
2843 #endif // PRODUCT
2844
2845 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2846 // Note we are overriding the read-only view of the prev map here, via
2847 // the cast.
2848 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2849 }
2850
2851 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2852 _nextMarkBitMap->clearRange(mr);
2853 }
2854
2855 HeapRegion*
2856 ConcurrentMark::claim_region(uint worker_id) {
2857 // "checkpoint" the finger
2858 HeapWord* finger = _finger;
2859
2860 // _heap_end will not change underneath our feet; it only changes at
2861 // yield points.
2862 while (finger < _heap_end) {
2863 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2864
2865 // Note on how this code handles humongous regions. In the
2866 // normal case the finger will reach the start of a "starts
2867 // humongous" (SH) region. Its end will either be the end of the
2868 // last "continues humongous" (CH) region in the sequence, or the
2869 // standard end of the SH region (if the SH is the only region in
2870 // the sequence). That way claim_region() will skip over the CH
2871 // regions. However, there is a subtle race between a CM thread
2872 // executing this method and a mutator thread doing a humongous
2873 // object allocation. The two are not mutually exclusive as the CM
2874 // thread does not need to hold the Heap_lock when it gets
2875 // here. So there is a chance that claim_region() will come across
2876 // a free region that's in the progress of becoming a SH or a CH
2877 // region. In the former case, it will either
2878 // a) Miss the update to the region's end, in which case it will
2879 // visit every subsequent CH region, will find their bitmaps
2880 // empty, and do nothing, or
2881 // b) Will observe the update of the region's end (in which case
2882 // it will skip the subsequent CH regions).
2883 // If it comes across a region that suddenly becomes CH, the
2884 // scenario will be similar to b). So, the race between
2885 // claim_region() and a humongous object allocation might force us
2886 // to do a bit of unnecessary work (due to some unnecessary bitmap
2887 // iterations) but it should not introduce and correctness issues.
2888 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2889
2890 // Above heap_region_containing_raw may return NULL as we always scan claim
2891 // until the end of the heap. In this case, just jump to the next region.
2892 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2893
2894 // Is the gap between reading the finger and doing the CAS too long?
2895 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2896 if (res == finger && curr_region != NULL) {
2897 // we succeeded
2898 HeapWord* bottom = curr_region->bottom();
2899 HeapWord* limit = curr_region->next_top_at_mark_start();
2900
2901 if (verbose_low()) {
2902 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2903 "["PTR_FORMAT", "PTR_FORMAT"), "
2904 "limit = "PTR_FORMAT,
2905 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2906 }
2907
2908 // notice that _finger == end cannot be guaranteed here since,
2909 // someone else might have moved the finger even further
2910 assert(_finger >= end, "the finger should have moved forward");
2911
2912 if (verbose_low()) {
2913 gclog_or_tty->print_cr("[%u] we were successful with region = "
2914 PTR_FORMAT, worker_id, p2i(curr_region));
2915 }
2916
2917 if (limit > bottom) {
2918 if (verbose_low()) {
2919 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2920 "returning it ", worker_id, p2i(curr_region));
2921 }
2922 return curr_region;
2923 } else {
2924 assert(limit == bottom,
2925 "the region limit should be at bottom");
2926 if (verbose_low()) {
2927 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
2928 "returning NULL", worker_id, p2i(curr_region));
2929 }
2930 // we return NULL and the caller should try calling
2931 // claim_region() again.
2932 return NULL;
2933 }
2934 } else {
2935 assert(_finger > finger, "the finger should have moved forward");
2936 if (verbose_low()) {
2937 if (curr_region == NULL) {
2938 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
2939 "global finger = "PTR_FORMAT", "
2940 "our finger = "PTR_FORMAT,
2941 worker_id, p2i(_finger), p2i(finger));
2942 } else {
2943 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2944 "global finger = "PTR_FORMAT", "
2945 "our finger = "PTR_FORMAT,
2946 worker_id, p2i(_finger), p2i(finger));
2947 }
2948 }
2949
2950 // read it again
2951 finger = _finger;
2952 }
2953 }
2954
2955 return NULL;
2956 }
2957
2958 #ifndef PRODUCT
2959 enum VerifyNoCSetOopsPhase {
2960 VerifyNoCSetOopsStack,
2961 VerifyNoCSetOopsQueues,
2962 VerifyNoCSetOopsSATBCompleted,
2963 VerifyNoCSetOopsSATBThread
2964 };
2965
2966 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
2967 private:
2968 G1CollectedHeap* _g1h;
2969 VerifyNoCSetOopsPhase _phase;
2970 int _info;
2971
2972 const char* phase_str() {
2973 switch (_phase) {
2974 case VerifyNoCSetOopsStack: return "Stack";
2975 case VerifyNoCSetOopsQueues: return "Queue";
2976 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2977 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
2978 default: ShouldNotReachHere();
2979 }
2980 return NULL;
2981 }
2982
2983 void do_object_work(oop obj) {
2984 guarantee(!_g1h->obj_in_cs(obj),
2985 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2986 p2i((void*) obj), phase_str(), _info));
2987 }
2988
2989 public:
2990 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2991
2992 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2993 _phase = phase;
2994 _info = info;
2995 }
2996
2997 virtual void do_oop(oop* p) {
2998 oop obj = oopDesc::load_decode_heap_oop(p);
2999 do_object_work(obj);
3000 }
3001
3002 virtual void do_oop(narrowOop* p) {
3003 // We should not come across narrow oops while scanning marking
3004 // stacks and SATB buffers.
3005 ShouldNotReachHere();
3006 }
3007
3008 virtual void do_object(oop obj) {
3009 do_object_work(obj);
3010 }
3011 };
3012
3013 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3014 bool verify_enqueued_buffers,
3015 bool verify_thread_buffers,
3016 bool verify_fingers) {
3017 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3018 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3019 return;
3020 }
3021
3022 VerifyNoCSetOopsClosure cl;
3023
3024 if (verify_stacks) {
3025 // Verify entries on the global mark stack
3026 cl.set_phase(VerifyNoCSetOopsStack);
3027 _markStack.oops_do(&cl);
3028
3029 // Verify entries on the task queues
3030 for (uint i = 0; i < _max_worker_id; i += 1) {
3031 cl.set_phase(VerifyNoCSetOopsQueues, i);
3032 CMTaskQueue* queue = _task_queues->queue(i);
3033 queue->oops_do(&cl);
3034 }
3035 }
3036
3037 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3038
3039 // Verify entries on the enqueued SATB buffers
3040 if (verify_enqueued_buffers) {
3041 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3042 satb_qs.iterate_completed_buffers_read_only(&cl);
3043 }
3044
3045 // Verify entries on the per-thread SATB buffers
3046 if (verify_thread_buffers) {
3047 cl.set_phase(VerifyNoCSetOopsSATBThread);
3048 satb_qs.iterate_thread_buffers_read_only(&cl);
3049 }
3050
3051 if (verify_fingers) {
3052 // Verify the global finger
3053 HeapWord* global_finger = finger();
3054 if (global_finger != NULL && global_finger < _heap_end) {
3055 // The global finger always points to a heap region boundary. We
3056 // use heap_region_containing_raw() to get the containing region
3057 // given that the global finger could be pointing to a free region
3058 // which subsequently becomes continues humongous. If that
3059 // happens, heap_region_containing() will return the bottom of the
3060 // corresponding starts humongous region and the check below will
3061 // not hold any more.
3062 // Since we always iterate over all regions, we might get a NULL HeapRegion
3063 // here.
3064 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3065 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3066 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3067 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3068 }
3069
3070 // Verify the task fingers
3071 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3072 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3073 CMTask* task = _tasks[i];
3074 HeapWord* task_finger = task->finger();
3075 if (task_finger != NULL && task_finger < _heap_end) {
3076 // See above note on the global finger verification.
3077 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3078 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3079 !task_hr->in_collection_set(),
3080 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3081 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3082 }
3083 }
3084 }
3085 }
3086 #endif // PRODUCT
3087
3088 // Aggregate the counting data that was constructed concurrently
3089 // with marking.
3090 class AggregateCountDataHRClosure: public HeapRegionClosure {
3091 G1CollectedHeap* _g1h;
3092 ConcurrentMark* _cm;
3093 CardTableModRefBS* _ct_bs;
3094 BitMap* _cm_card_bm;
3095 uint _max_worker_id;
3096
3097 public:
3098 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3099 BitMap* cm_card_bm,
3100 uint max_worker_id) :
3101 _g1h(g1h), _cm(g1h->concurrent_mark()),
3102 _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
3103 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3104
3105 bool doHeapRegion(HeapRegion* hr) {
3106 if (hr->is_continues_humongous()) {
3107 // We will ignore these here and process them when their
3108 // associated "starts humongous" region is processed.
3109 // Note that we cannot rely on their associated
3110 // "starts humongous" region to have their bit set to 1
3111 // since, due to the region chunking in the parallel region
3112 // iteration, a "continues humongous" region might be visited
3113 // before its associated "starts humongous".
3114 return false;
3115 }
3116
3117 HeapWord* start = hr->bottom();
3118 HeapWord* limit = hr->next_top_at_mark_start();
3119 HeapWord* end = hr->end();
3120
3121 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3122 err_msg("Preconditions not met - "
3123 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3124 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3125 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3126
3127 assert(hr->next_marked_bytes() == 0, "Precondition");
3128
3129 if (start == limit) {
3130 // NTAMS of this region has not been set so nothing to do.
3131 return false;
3132 }
3133
3134 // 'start' should be in the heap.
3135 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3136 // 'end' *may* be just beyond the end of the heap (if hr is the last region)
3137 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3138
3139 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3140 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3141 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3142
3143 // If ntams is not card aligned then we bump card bitmap index
3144 // for limit so that we get the all the cards spanned by
3145 // the object ending at ntams.
3146 // Note: if this is the last region in the heap then ntams
3147 // could be actually just beyond the end of the the heap;
3148 // limit_idx will then correspond to a (non-existent) card
3149 // that is also outside the heap.
3150 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3151 limit_idx += 1;
3152 }
3153
3154 assert(limit_idx <= end_idx, "or else use atomics");
3155
3156 // Aggregate the "stripe" in the count data associated with hr.
3157 uint hrm_index = hr->hrm_index();
3158 size_t marked_bytes = 0;
3159
3160 for (uint i = 0; i < _max_worker_id; i += 1) {
3161 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3162 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3163
3164 // Fetch the marked_bytes in this region for task i and
3165 // add it to the running total for this region.
3166 marked_bytes += marked_bytes_array[hrm_index];
3167
3168 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3169 // into the global card bitmap.
3170 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3171
3172 while (scan_idx < limit_idx) {
3173 assert(task_card_bm->at(scan_idx) == true, "should be");
3174 _cm_card_bm->set_bit(scan_idx);
3175 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3176
3177 // BitMap::get_next_one_offset() can handle the case when
3178 // its left_offset parameter is greater than its right_offset
3179 // parameter. It does, however, have an early exit if
3180 // left_offset == right_offset. So let's limit the value
3181 // passed in for left offset here.
3182 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3183 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3184 }
3185 }
3186
3187 // Update the marked bytes for this region.
3188 hr->add_to_marked_bytes(marked_bytes);
3189
3190 // Next heap region
3191 return false;
3192 }
3193 };
3194
3195 class G1AggregateCountDataTask: public AbstractGangTask {
3196 protected:
3197 G1CollectedHeap* _g1h;
3198 ConcurrentMark* _cm;
3199 BitMap* _cm_card_bm;
3200 uint _max_worker_id;
3201 int _active_workers;
3202 HeapRegionClaimer _hrclaimer;
3203
3204 public:
3205 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3206 ConcurrentMark* cm,
3207 BitMap* cm_card_bm,
3208 uint max_worker_id,
3209 int n_workers) :
3210 AbstractGangTask("Count Aggregation"),
3211 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3212 _max_worker_id(max_worker_id),
3213 _active_workers(n_workers),
3214 _hrclaimer(_active_workers) {
3215 }
3216
3217 void work(uint worker_id) {
3218 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3219
3220 _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer);
3221 }
3222 };
3223
3224
3225 void ConcurrentMark::aggregate_count_data() {
3226 int n_workers = _g1h->workers()->active_workers();
3227
3228 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3229 _max_worker_id, n_workers);
3230
3231 _g1h->set_par_threads(n_workers);
3232 _g1h->workers()->run_task(&g1_par_agg_task);
3233 _g1h->set_par_threads(0);
3234 }
3235
3236 // Clear the per-worker arrays used to store the per-region counting data
3237 void ConcurrentMark::clear_all_count_data() {
3238 // Clear the global card bitmap - it will be filled during
3239 // liveness count aggregation (during remark) and the
3240 // final counting task.
3241 _card_bm.clear();
3242
3243 // Clear the global region bitmap - it will be filled as part
3244 // of the final counting task.
3245 _region_bm.clear();
3246
3247 uint max_regions = _g1h->max_regions();
3248 assert(_max_worker_id > 0, "uninitialized");
3249
3250 for (uint i = 0; i < _max_worker_id; i += 1) {
3251 BitMap* task_card_bm = count_card_bitmap_for(i);
3252 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3253
3254 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3255 assert(marked_bytes_array != NULL, "uninitialized");
3256
3257 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3258 task_card_bm->clear();
3259 }
3260 }
3261
3262 void ConcurrentMark::print_stats() {
3263 if (verbose_stats()) {
3264 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3265 for (size_t i = 0; i < _active_tasks; ++i) {
3266 _tasks[i]->print_stats();
3267 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3268 }
3269 }
3270 }
3271
3272 // abandon current marking iteration due to a Full GC
3273 void ConcurrentMark::abort() {
3274 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3275 // concurrent bitmap clearing.
3276 _nextMarkBitMap->clearAll();
3277
3278 // Note we cannot clear the previous marking bitmap here
3279 // since VerifyDuringGC verifies the objects marked during
3280 // a full GC against the previous bitmap.
3281
3282 // Clear the liveness counting data
3283 clear_all_count_data();
3284 // Empty mark stack
3285 reset_marking_state();
3286 for (uint i = 0; i < _max_worker_id; ++i) {
3287 _tasks[i]->clear_region_fields();
3288 }
3289 _first_overflow_barrier_sync.abort();
3290 _second_overflow_barrier_sync.abort();
3291 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3292 if (!gc_id.is_undefined()) {
3293 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3294 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3295 _aborted_gc_id = gc_id;
3296 }
3297 _has_aborted = true;
3298
3299 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3300 satb_mq_set.abandon_partial_marking();
3301 // This can be called either during or outside marking, we'll read
3302 // the expected_active value from the SATB queue set.
3303 satb_mq_set.set_active_all_threads(
3304 false, /* new active value */
3305 satb_mq_set.is_active() /* expected_active */);
3306
3307 _g1h->trace_heap_after_concurrent_cycle();
3308 _g1h->register_concurrent_cycle_end();
3309 }
3310
3311 const GCId& ConcurrentMark::concurrent_gc_id() {
3312 if (has_aborted()) {
3313 return _aborted_gc_id;
3314 }
3315 return _g1h->gc_tracer_cm()->gc_id();
3316 }
3317
3318 static void print_ms_time_info(const char* prefix, const char* name,
3319 NumberSeq& ns) {
3320 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3321 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3322 if (ns.num() > 0) {
3323 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3324 prefix, ns.sd(), ns.maximum());
3325 }
3326 }
3327
3328 void ConcurrentMark::print_summary_info() {
3329 gclog_or_tty->print_cr(" Concurrent marking:");
3330 print_ms_time_info(" ", "init marks", _init_times);
3331 print_ms_time_info(" ", "remarks", _remark_times);
3332 {
3333 print_ms_time_info(" ", "final marks", _remark_mark_times);
3334 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3335
3336 }
3337 print_ms_time_info(" ", "cleanups", _cleanup_times);
3338 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3339 _total_counting_time,
3340 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3341 (double)_cleanup_times.num()
3342 : 0.0));
3343 if (G1ScrubRemSets) {
3344 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3345 _total_rs_scrub_time,
3346 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3347 (double)_cleanup_times.num()
3348 : 0.0));
3349 }
3350 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3351 (_init_times.sum() + _remark_times.sum() +
3352 _cleanup_times.sum())/1000.0);
3353 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3354 "(%8.2f s marking).",
3355 cmThread()->vtime_accum(),
3356 cmThread()->vtime_mark_accum());
3357 }
3358
3359 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3360 _parallel_workers->print_worker_threads_on(st);
3361 }
3362
3363 void ConcurrentMark::print_on_error(outputStream* st) const {
3364 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3365 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3366 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3367 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3368 }
3369
3370 // We take a break if someone is trying to stop the world.
3371 bool ConcurrentMark::do_yield_check(uint worker_id) {
3372 if (SuspendibleThreadSet::should_yield()) {
3373 if (worker_id == 0) {
3374 _g1h->g1_policy()->record_concurrent_pause();
3375 }
3376 SuspendibleThreadSet::yield();
3377 return true;
3378 } else {
3379 return false;
3380 }
3381 }
3382
3383 #ifndef PRODUCT
3384 // for debugging purposes
3385 void ConcurrentMark::print_finger() {
3386 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3387 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3388 for (uint i = 0; i < _max_worker_id; ++i) {
3389 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3390 }
3391 gclog_or_tty->cr();
3392 }
3393 #endif
3394
3395 void CMTask::scan_object(oop obj) {
3396 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3397
3398 if (_cm->verbose_high()) {
3399 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3400 _worker_id, p2i((void*) obj));
3401 }
3402
3403 size_t obj_size = obj->size();
3404 _words_scanned += obj_size;
3405
3406 obj->oop_iterate(_cm_oop_closure);
3407 statsOnly( ++_objs_scanned );
3408 check_limits();
3409 }
3410
3411 // Closure for iteration over bitmaps
3412 class CMBitMapClosure : public BitMapClosure {
3413 private:
3414 // the bitmap that is being iterated over
3415 CMBitMap* _nextMarkBitMap;
3416 ConcurrentMark* _cm;
3417 CMTask* _task;
3418
3419 public:
3420 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3421 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3422
3423 bool do_bit(size_t offset) {
3424 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3425 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3426 assert( addr < _cm->finger(), "invariant");
3427
3428 statsOnly( _task->increase_objs_found_on_bitmap() );
3429 assert(addr >= _task->finger(), "invariant");
3430
3431 // We move that task's local finger along.
3432 _task->move_finger_to(addr);
3433
3434 _task->scan_object(oop(addr));
3435 // we only partially drain the local queue and global stack
3436 _task->drain_local_queue(true);
3437 _task->drain_global_stack(true);
3438
3439 // if the has_aborted flag has been raised, we need to bail out of
3440 // the iteration
3441 return !_task->has_aborted();
3442 }
3443 };
3444
3445 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3446 ConcurrentMark* cm,
3447 CMTask* task)
3448 : _g1h(g1h), _cm(cm), _task(task) {
3449 assert(_ref_processor == NULL, "should be initialized to NULL");
3450
3451 if (G1UseConcMarkReferenceProcessing) {
3452 _ref_processor = g1h->ref_processor_cm();
3453 assert(_ref_processor != NULL, "should not be NULL");
3454 }
3455 }
3456
3457 void CMTask::setup_for_region(HeapRegion* hr) {
3458 assert(hr != NULL,
3459 "claim_region() should have filtered out NULL regions");
3460 assert(!hr->is_continues_humongous(),
3461 "claim_region() should have filtered out continues humongous regions");
3462
3463 if (_cm->verbose_low()) {
3464 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3465 _worker_id, p2i(hr));
3466 }
3467
3468 _curr_region = hr;
3469 _finger = hr->bottom();
3470 update_region_limit();
3471 }
3472
3473 void CMTask::update_region_limit() {
3474 HeapRegion* hr = _curr_region;
3475 HeapWord* bottom = hr->bottom();
3476 HeapWord* limit = hr->next_top_at_mark_start();
3477
3478 if (limit == bottom) {
3479 if (_cm->verbose_low()) {
3480 gclog_or_tty->print_cr("[%u] found an empty region "
3481 "["PTR_FORMAT", "PTR_FORMAT")",
3482 _worker_id, p2i(bottom), p2i(limit));
3483 }
3484 // The region was collected underneath our feet.
3485 // We set the finger to bottom to ensure that the bitmap
3486 // iteration that will follow this will not do anything.
3487 // (this is not a condition that holds when we set the region up,
3488 // as the region is not supposed to be empty in the first place)
3489 _finger = bottom;
3490 } else if (limit >= _region_limit) {
3491 assert(limit >= _finger, "peace of mind");
3492 } else {
3493 assert(limit < _region_limit, "only way to get here");
3494 // This can happen under some pretty unusual circumstances. An
3495 // evacuation pause empties the region underneath our feet (NTAMS
3496 // at bottom). We then do some allocation in the region (NTAMS
3497 // stays at bottom), followed by the region being used as a GC
3498 // alloc region (NTAMS will move to top() and the objects
3499 // originally below it will be grayed). All objects now marked in
3500 // the region are explicitly grayed, if below the global finger,
3501 // and we do not need in fact to scan anything else. So, we simply
3502 // set _finger to be limit to ensure that the bitmap iteration
3503 // doesn't do anything.
3504 _finger = limit;
3505 }
3506
3507 _region_limit = limit;
3508 }
3509
3510 void CMTask::giveup_current_region() {
3511 assert(_curr_region != NULL, "invariant");
3512 if (_cm->verbose_low()) {
3513 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3514 _worker_id, p2i(_curr_region));
3515 }
3516 clear_region_fields();
3517 }
3518
3519 void CMTask::clear_region_fields() {
3520 // Values for these three fields that indicate that we're not
3521 // holding on to a region.
3522 _curr_region = NULL;
3523 _finger = NULL;
3524 _region_limit = NULL;
3525 }
3526
3527 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3528 if (cm_oop_closure == NULL) {
3529 assert(_cm_oop_closure != NULL, "invariant");
3530 } else {
3531 assert(_cm_oop_closure == NULL, "invariant");
3532 }
3533 _cm_oop_closure = cm_oop_closure;
3534 }
3535
3536 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3537 guarantee(nextMarkBitMap != NULL, "invariant");
3538
3539 if (_cm->verbose_low()) {
3540 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3541 }
3542
3543 _nextMarkBitMap = nextMarkBitMap;
3544 clear_region_fields();
3545
3546 _calls = 0;
3547 _elapsed_time_ms = 0.0;
3548 _termination_time_ms = 0.0;
3549 _termination_start_time_ms = 0.0;
3550
3551 #if _MARKING_STATS_
3552 _aborted = 0;
3553 _aborted_overflow = 0;
3554 _aborted_cm_aborted = 0;
3555 _aborted_yield = 0;
3556 _aborted_timed_out = 0;
3557 _aborted_satb = 0;
3558 _aborted_termination = 0;
3559 _steal_attempts = 0;
3560 _steals = 0;
3561 _local_pushes = 0;
3562 _local_pops = 0;
3563 _local_max_size = 0;
3564 _objs_scanned = 0;
3565 _global_pushes = 0;
3566 _global_pops = 0;
3567 _global_max_size = 0;
3568 _global_transfers_to = 0;
3569 _global_transfers_from = 0;
3570 _regions_claimed = 0;
3571 _objs_found_on_bitmap = 0;
3572 _satb_buffers_processed = 0;
3573 #endif // _MARKING_STATS_
3574 }
3575
3576 bool CMTask::should_exit_termination() {
3577 regular_clock_call();
3578 // This is called when we are in the termination protocol. We should
3579 // quit if, for some reason, this task wants to abort or the global
3580 // stack is not empty (this means that we can get work from it).
3581 return !_cm->mark_stack_empty() || has_aborted();
3582 }
3583
3584 void CMTask::reached_limit() {
3585 assert(_words_scanned >= _words_scanned_limit ||
3586 _refs_reached >= _refs_reached_limit ,
3587 "shouldn't have been called otherwise");
3588 regular_clock_call();
3589 }
3590
3591 void CMTask::regular_clock_call() {
3592 if (has_aborted()) return;
3593
3594 // First, we need to recalculate the words scanned and refs reached
3595 // limits for the next clock call.
3596 recalculate_limits();
3597
3598 // During the regular clock call we do the following
3599
3600 // (1) If an overflow has been flagged, then we abort.
3601 if (_cm->has_overflown()) {
3602 set_has_aborted();
3603 return;
3604 }
3605
3606 // If we are not concurrent (i.e. we're doing remark) we don't need
3607 // to check anything else. The other steps are only needed during
3608 // the concurrent marking phase.
3609 if (!concurrent()) return;
3610
3611 // (2) If marking has been aborted for Full GC, then we also abort.
3612 if (_cm->has_aborted()) {
3613 set_has_aborted();
3614 statsOnly( ++_aborted_cm_aborted );
3615 return;
3616 }
3617
3618 double curr_time_ms = os::elapsedVTime() * 1000.0;
3619
3620 // (3) If marking stats are enabled, then we update the step history.
3621 #if _MARKING_STATS_
3622 if (_words_scanned >= _words_scanned_limit) {
3623 ++_clock_due_to_scanning;
3624 }
3625 if (_refs_reached >= _refs_reached_limit) {
3626 ++_clock_due_to_marking;
3627 }
3628
3629 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3630 _interval_start_time_ms = curr_time_ms;
3631 _all_clock_intervals_ms.add(last_interval_ms);
3632
3633 if (_cm->verbose_medium()) {
3634 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3635 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3636 _worker_id, last_interval_ms,
3637 _words_scanned,
3638 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3639 _refs_reached,
3640 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3641 }
3642 #endif // _MARKING_STATS_
3643
3644 // (4) We check whether we should yield. If we have to, then we abort.
3645 if (SuspendibleThreadSet::should_yield()) {
3646 // We should yield. To do this we abort the task. The caller is
3647 // responsible for yielding.
3648 set_has_aborted();
3649 statsOnly( ++_aborted_yield );
3650 return;
3651 }
3652
3653 // (5) We check whether we've reached our time quota. If we have,
3654 // then we abort.
3655 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3656 if (elapsed_time_ms > _time_target_ms) {
3657 set_has_aborted();
3658 _has_timed_out = true;
3659 statsOnly( ++_aborted_timed_out );
3660 return;
3661 }
3662
3663 // (6) Finally, we check whether there are enough completed STAB
3664 // buffers available for processing. If there are, we abort.
3665 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3666 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3667 if (_cm->verbose_low()) {
3668 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3669 _worker_id);
3670 }
3671 // we do need to process SATB buffers, we'll abort and restart
3672 // the marking task to do so
3673 set_has_aborted();
3674 statsOnly( ++_aborted_satb );
3675 return;
3676 }
3677 }
3678
3679 void CMTask::recalculate_limits() {
3680 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3681 _words_scanned_limit = _real_words_scanned_limit;
3682
3683 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3684 _refs_reached_limit = _real_refs_reached_limit;
3685 }
3686
3687 void CMTask::decrease_limits() {
3688 // This is called when we believe that we're going to do an infrequent
3689 // operation which will increase the per byte scanned cost (i.e. move
3690 // entries to/from the global stack). It basically tries to decrease the
3691 // scanning limit so that the clock is called earlier.
3692
3693 if (_cm->verbose_medium()) {
3694 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3695 }
3696
3697 _words_scanned_limit = _real_words_scanned_limit -
3698 3 * words_scanned_period / 4;
3699 _refs_reached_limit = _real_refs_reached_limit -
3700 3 * refs_reached_period / 4;
3701 }
3702
3703 void CMTask::move_entries_to_global_stack() {
3704 // local array where we'll store the entries that will be popped
3705 // from the local queue
3706 oop buffer[global_stack_transfer_size];
3707
3708 int n = 0;
3709 oop obj;
3710 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3711 buffer[n] = obj;
3712 ++n;
3713 }
3714
3715 if (n > 0) {
3716 // we popped at least one entry from the local queue
3717
3718 statsOnly( ++_global_transfers_to; _local_pops += n );
3719
3720 if (!_cm->mark_stack_push(buffer, n)) {
3721 if (_cm->verbose_low()) {
3722 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3723 _worker_id);
3724 }
3725 set_has_aborted();
3726 } else {
3727 // the transfer was successful
3728
3729 if (_cm->verbose_medium()) {
3730 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3731 _worker_id, n);
3732 }
3733 statsOnly( size_t tmp_size = _cm->mark_stack_size();
3734 if (tmp_size > _global_max_size) {
3735 _global_max_size = tmp_size;
3736 }
3737 _global_pushes += n );
3738 }
3739 }
3740
3741 // this operation was quite expensive, so decrease the limits
3742 decrease_limits();
3743 }
3744
3745 void CMTask::get_entries_from_global_stack() {
3746 // local array where we'll store the entries that will be popped
3747 // from the global stack.
3748 oop buffer[global_stack_transfer_size];
3749 int n;
3750 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3751 assert(n <= global_stack_transfer_size,
3752 "we should not pop more than the given limit");
3753 if (n > 0) {
3754 // yes, we did actually pop at least one entry
3755
3756 statsOnly( ++_global_transfers_from; _global_pops += n );
3757 if (_cm->verbose_medium()) {
3758 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3759 _worker_id, n);
3760 }
3761 for (int i = 0; i < n; ++i) {
3762 bool success = _task_queue->push(buffer[i]);
3763 // We only call this when the local queue is empty or under a
3764 // given target limit. So, we do not expect this push to fail.
3765 assert(success, "invariant");
3766 }
3767
3768 statsOnly( size_t tmp_size = (size_t)_task_queue->size();
3769 if (tmp_size > _local_max_size) {
3770 _local_max_size = tmp_size;
3771 }
3772 _local_pushes += n );
3773 }
3774
3775 // this operation was quite expensive, so decrease the limits
3776 decrease_limits();
3777 }
3778
3779 void CMTask::drain_local_queue(bool partially) {
3780 if (has_aborted()) return;
3781
3782 // Decide what the target size is, depending whether we're going to
3783 // drain it partially (so that other tasks can steal if they run out
3784 // of things to do) or totally (at the very end).
3785 size_t target_size;
3786 if (partially) {
3787 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3788 } else {
3789 target_size = 0;
3790 }
3791
3792 if (_task_queue->size() > target_size) {
3793 if (_cm->verbose_high()) {
3794 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3795 _worker_id, target_size);
3796 }
3797
3798 oop obj;
3799 bool ret = _task_queue->pop_local(obj);
3800 while (ret) {
3801 statsOnly( ++_local_pops );
3802
3803 if (_cm->verbose_high()) {
3804 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3805 p2i((void*) obj));
3806 }
3807
3808 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3809 assert(!_g1h->is_on_master_free_list(
3810 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3811
3812 scan_object(obj);
3813
3814 if (_task_queue->size() <= target_size || has_aborted()) {
3815 ret = false;
3816 } else {
3817 ret = _task_queue->pop_local(obj);
3818 }
3819 }
3820
3821 if (_cm->verbose_high()) {
3822 gclog_or_tty->print_cr("[%u] drained local queue, size = %u",
3823 _worker_id, _task_queue->size());
3824 }
3825 }
3826 }
3827
3828 void CMTask::drain_global_stack(bool partially) {
3829 if (has_aborted()) return;
3830
3831 // We have a policy to drain the local queue before we attempt to
3832 // drain the global stack.
3833 assert(partially || _task_queue->size() == 0, "invariant");
3834
3835 // Decide what the target size is, depending whether we're going to
3836 // drain it partially (so that other tasks can steal if they run out
3837 // of things to do) or totally (at the very end). Notice that,
3838 // because we move entries from the global stack in chunks or
3839 // because another task might be doing the same, we might in fact
3840 // drop below the target. But, this is not a problem.
3841 size_t target_size;
3842 if (partially) {
3843 target_size = _cm->partial_mark_stack_size_target();
3844 } else {
3845 target_size = 0;
3846 }
3847
3848 if (_cm->mark_stack_size() > target_size) {
3849 if (_cm->verbose_low()) {
3850 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3851 _worker_id, target_size);
3852 }
3853
3854 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3855 get_entries_from_global_stack();
3856 drain_local_queue(partially);
3857 }
3858
3859 if (_cm->verbose_low()) {
3860 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3861 _worker_id, _cm->mark_stack_size());
3862 }
3863 }
3864 }
3865
3866 // SATB Queue has several assumptions on whether to call the par or
3867 // non-par versions of the methods. this is why some of the code is
3868 // replicated. We should really get rid of the single-threaded version
3869 // of the code to simplify things.
3870 void CMTask::drain_satb_buffers() {
3871 if (has_aborted()) return;
3872
3873 // We set this so that the regular clock knows that we're in the
3874 // middle of draining buffers and doesn't set the abort flag when it
3875 // notices that SATB buffers are available for draining. It'd be
3876 // very counter productive if it did that. :-)
3877 _draining_satb_buffers = true;
3878
3879 CMObjectClosure oc(this);
3880 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3881 satb_mq_set.set_closure(_worker_id, &oc);
3882
3883 // This keeps claiming and applying the closure to completed buffers
3884 // until we run out of buffers or we need to abort.
3885 while (!has_aborted() &&
3886 satb_mq_set.apply_closure_to_completed_buffer(_worker_id)) {
3887 if (_cm->verbose_medium()) {
3888 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3889 }
3890 statsOnly( ++_satb_buffers_processed );
3891 regular_clock_call();
3892 }
3893
3894 _draining_satb_buffers = false;
3895
3896 assert(has_aborted() ||
3897 concurrent() ||
3898 satb_mq_set.completed_buffers_num() == 0, "invariant");
3899
3900 satb_mq_set.set_closure(_worker_id, NULL);
3901
3902 // again, this was a potentially expensive operation, decrease the
3903 // limits to get the regular clock call early
3904 decrease_limits();
3905 }
3906
3907 void CMTask::print_stats() {
3908 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3909 _worker_id, _calls);
3910 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3911 _elapsed_time_ms, _termination_time_ms);
3912 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3913 _step_times_ms.num(), _step_times_ms.avg(),
3914 _step_times_ms.sd());
3915 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3916 _step_times_ms.maximum(), _step_times_ms.sum());
3917
3918 #if _MARKING_STATS_
3919 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3920 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3921 _all_clock_intervals_ms.sd());
3922 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3923 _all_clock_intervals_ms.maximum(),
3924 _all_clock_intervals_ms.sum());
3925 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = " SIZE_FORMAT ", marking = " SIZE_FORMAT,
3926 _clock_due_to_scanning, _clock_due_to_marking);
3927 gclog_or_tty->print_cr(" Objects: scanned = " SIZE_FORMAT ", found on the bitmap = " SIZE_FORMAT,
3928 _objs_scanned, _objs_found_on_bitmap);
3929 gclog_or_tty->print_cr(" Local Queue: pushes = " SIZE_FORMAT ", pops = " SIZE_FORMAT ", max size = " SIZE_FORMAT,
3930 _local_pushes, _local_pops, _local_max_size);
3931 gclog_or_tty->print_cr(" Global Stack: pushes = " SIZE_FORMAT ", pops = " SIZE_FORMAT ", max size = " SIZE_FORMAT,
3932 _global_pushes, _global_pops, _global_max_size);
3933 gclog_or_tty->print_cr(" transfers to = " SIZE_FORMAT ", transfers from = " SIZE_FORMAT,
3934 _global_transfers_to,_global_transfers_from);
3935 gclog_or_tty->print_cr(" Regions: claimed = " SIZE_FORMAT, _regions_claimed);
3936 gclog_or_tty->print_cr(" SATB buffers: processed = " SIZE_FORMAT, _satb_buffers_processed);
3937 gclog_or_tty->print_cr(" Steals: attempts = " SIZE_FORMAT ", successes = " SIZE_FORMAT,
3938 _steal_attempts, _steals);
3939 gclog_or_tty->print_cr(" Aborted: " SIZE_FORMAT ", due to", _aborted);
3940 gclog_or_tty->print_cr(" overflow: " SIZE_FORMAT ", global abort: " SIZE_FORMAT ", yield: " SIZE_FORMAT,
3941 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3942 gclog_or_tty->print_cr(" time out: " SIZE_FORMAT ", SATB: " SIZE_FORMAT ", termination: " SIZE_FORMAT,
3943 _aborted_timed_out, _aborted_satb, _aborted_termination);
3944 #endif // _MARKING_STATS_
3945 }
3946
3947 /*****************************************************************************
3948
3949 The do_marking_step(time_target_ms, ...) method is the building
3950 block of the parallel marking framework. It can be called in parallel
3951 with other invocations of do_marking_step() on different tasks
3952 (but only one per task, obviously) and concurrently with the
3953 mutator threads, or during remark, hence it eliminates the need
3954 for two versions of the code. When called during remark, it will
3955 pick up from where the task left off during the concurrent marking
3956 phase. Interestingly, tasks are also claimable during evacuation
3957 pauses too, since do_marking_step() ensures that it aborts before
3958 it needs to yield.
3959
3960 The data structures that it uses to do marking work are the
3961 following:
3962
3963 (1) Marking Bitmap. If there are gray objects that appear only
3964 on the bitmap (this happens either when dealing with an overflow
3965 or when the initial marking phase has simply marked the roots
3966 and didn't push them on the stack), then tasks claim heap
3967 regions whose bitmap they then scan to find gray objects. A
3968 global finger indicates where the end of the last claimed region
3969 is. A local finger indicates how far into the region a task has
3970 scanned. The two fingers are used to determine how to gray an
3971 object (i.e. whether simply marking it is OK, as it will be
3972 visited by a task in the future, or whether it needs to be also
3973 pushed on a stack).
3974
3975 (2) Local Queue. The local queue of the task which is accessed
3976 reasonably efficiently by the task. Other tasks can steal from
3977 it when they run out of work. Throughout the marking phase, a
3978 task attempts to keep its local queue short but not totally
3979 empty, so that entries are available for stealing by other
3980 tasks. Only when there is no more work, a task will totally
3981 drain its local queue.
3982
3983 (3) Global Mark Stack. This handles local queue overflow. During
3984 marking only sets of entries are moved between it and the local
3985 queues, as access to it requires a mutex and more fine-grain
3986 interaction with it which might cause contention. If it
3987 overflows, then the marking phase should restart and iterate
3988 over the bitmap to identify gray objects. Throughout the marking
3989 phase, tasks attempt to keep the global mark stack at a small
3990 length but not totally empty, so that entries are available for
3991 popping by other tasks. Only when there is no more work, tasks
3992 will totally drain the global mark stack.
3993
3994 (4) SATB Buffer Queue. This is where completed SATB buffers are
3995 made available. Buffers are regularly removed from this queue
3996 and scanned for roots, so that the queue doesn't get too
3997 long. During remark, all completed buffers are processed, as
3998 well as the filled in parts of any uncompleted buffers.
3999
4000 The do_marking_step() method tries to abort when the time target
4001 has been reached. There are a few other cases when the
4002 do_marking_step() method also aborts:
4003
4004 (1) When the marking phase has been aborted (after a Full GC).
4005
4006 (2) When a global overflow (on the global stack) has been
4007 triggered. Before the task aborts, it will actually sync up with
4008 the other tasks to ensure that all the marking data structures
4009 (local queues, stacks, fingers etc.) are re-initialized so that
4010 when do_marking_step() completes, the marking phase can
4011 immediately restart.
4012
4013 (3) When enough completed SATB buffers are available. The
4014 do_marking_step() method only tries to drain SATB buffers right
4015 at the beginning. So, if enough buffers are available, the
4016 marking step aborts and the SATB buffers are processed at
4017 the beginning of the next invocation.
4018
4019 (4) To yield. when we have to yield then we abort and yield
4020 right at the end of do_marking_step(). This saves us from a lot
4021 of hassle as, by yielding we might allow a Full GC. If this
4022 happens then objects will be compacted underneath our feet, the
4023 heap might shrink, etc. We save checking for this by just
4024 aborting and doing the yield right at the end.
4025
4026 From the above it follows that the do_marking_step() method should
4027 be called in a loop (or, otherwise, regularly) until it completes.
4028
4029 If a marking step completes without its has_aborted() flag being
4030 true, it means it has completed the current marking phase (and
4031 also all other marking tasks have done so and have all synced up).
4032
4033 A method called regular_clock_call() is invoked "regularly" (in
4034 sub ms intervals) throughout marking. It is this clock method that
4035 checks all the abort conditions which were mentioned above and
4036 decides when the task should abort. A work-based scheme is used to
4037 trigger this clock method: when the number of object words the
4038 marking phase has scanned or the number of references the marking
4039 phase has visited reach a given limit. Additional invocations to
4040 the method clock have been planted in a few other strategic places
4041 too. The initial reason for the clock method was to avoid calling
4042 vtime too regularly, as it is quite expensive. So, once it was in
4043 place, it was natural to piggy-back all the other conditions on it
4044 too and not constantly check them throughout the code.
4045
4046 If do_termination is true then do_marking_step will enter its
4047 termination protocol.
4048
4049 The value of is_serial must be true when do_marking_step is being
4050 called serially (i.e. by the VMThread) and do_marking_step should
4051 skip any synchronization in the termination and overflow code.
4052 Examples include the serial remark code and the serial reference
4053 processing closures.
4054
4055 The value of is_serial must be false when do_marking_step is
4056 being called by any of the worker threads in a work gang.
4057 Examples include the concurrent marking code (CMMarkingTask),
4058 the MT remark code, and the MT reference processing closures.
4059
4060 *****************************************************************************/
4061
4062 void CMTask::do_marking_step(double time_target_ms,
4063 bool do_termination,
4064 bool is_serial) {
4065 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4066 assert(concurrent() == _cm->concurrent(), "they should be the same");
4067
4068 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4069 assert(_task_queues != NULL, "invariant");
4070 assert(_task_queue != NULL, "invariant");
4071 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4072
4073 assert(!_claimed,
4074 "only one thread should claim this task at any one time");
4075
4076 // OK, this doesn't safeguard again all possible scenarios, as it is
4077 // possible for two threads to set the _claimed flag at the same
4078 // time. But it is only for debugging purposes anyway and it will
4079 // catch most problems.
4080 _claimed = true;
4081
4082 _start_time_ms = os::elapsedVTime() * 1000.0;
4083 statsOnly( _interval_start_time_ms = _start_time_ms );
4084
4085 // If do_stealing is true then do_marking_step will attempt to
4086 // steal work from the other CMTasks. It only makes sense to
4087 // enable stealing when the termination protocol is enabled
4088 // and do_marking_step() is not being called serially.
4089 bool do_stealing = do_termination && !is_serial;
4090
4091 double diff_prediction_ms =
4092 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4093 _time_target_ms = time_target_ms - diff_prediction_ms;
4094
4095 // set up the variables that are used in the work-based scheme to
4096 // call the regular clock method
4097 _words_scanned = 0;
4098 _refs_reached = 0;
4099 recalculate_limits();
4100
4101 // clear all flags
4102 clear_has_aborted();
4103 _has_timed_out = false;
4104 _draining_satb_buffers = false;
4105
4106 ++_calls;
4107
4108 if (_cm->verbose_low()) {
4109 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4110 "target = %1.2lfms >>>>>>>>>>",
4111 _worker_id, _calls, _time_target_ms);
4112 }
4113
4114 // Set up the bitmap and oop closures. Anything that uses them is
4115 // eventually called from this method, so it is OK to allocate these
4116 // statically.
4117 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4118 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4119 set_cm_oop_closure(&cm_oop_closure);
4120
4121 if (_cm->has_overflown()) {
4122 // This can happen if the mark stack overflows during a GC pause
4123 // and this task, after a yield point, restarts. We have to abort
4124 // as we need to get into the overflow protocol which happens
4125 // right at the end of this task.
4126 set_has_aborted();
4127 }
4128
4129 // First drain any available SATB buffers. After this, we will not
4130 // look at SATB buffers before the next invocation of this method.
4131 // If enough completed SATB buffers are queued up, the regular clock
4132 // will abort this task so that it restarts.
4133 drain_satb_buffers();
4134 // ...then partially drain the local queue and the global stack
4135 drain_local_queue(true);
4136 drain_global_stack(true);
4137
4138 do {
4139 if (!has_aborted() && _curr_region != NULL) {
4140 // This means that we're already holding on to a region.
4141 assert(_finger != NULL, "if region is not NULL, then the finger "
4142 "should not be NULL either");
4143
4144 // We might have restarted this task after an evacuation pause
4145 // which might have evacuated the region we're holding on to
4146 // underneath our feet. Let's read its limit again to make sure
4147 // that we do not iterate over a region of the heap that
4148 // contains garbage (update_region_limit() will also move
4149 // _finger to the start of the region if it is found empty).
4150 update_region_limit();
4151 // We will start from _finger not from the start of the region,
4152 // as we might be restarting this task after aborting half-way
4153 // through scanning this region. In this case, _finger points to
4154 // the address where we last found a marked object. If this is a
4155 // fresh region, _finger points to start().
4156 MemRegion mr = MemRegion(_finger, _region_limit);
4157
4158 if (_cm->verbose_low()) {
4159 gclog_or_tty->print_cr("[%u] we're scanning part "
4160 "["PTR_FORMAT", "PTR_FORMAT") "
4161 "of region "HR_FORMAT,
4162 _worker_id, p2i(_finger), p2i(_region_limit),
4163 HR_FORMAT_PARAMS(_curr_region));
4164 }
4165
4166 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
4167 "humongous regions should go around loop once only");
4168
4169 // Some special cases:
4170 // If the memory region is empty, we can just give up the region.
4171 // If the current region is humongous then we only need to check
4172 // the bitmap for the bit associated with the start of the object,
4173 // scan the object if it's live, and give up the region.
4174 // Otherwise, let's iterate over the bitmap of the part of the region
4175 // that is left.
4176 // If the iteration is successful, give up the region.
4177 if (mr.is_empty()) {
4178 giveup_current_region();
4179 regular_clock_call();
4180 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
4181 if (_nextMarkBitMap->isMarked(mr.start())) {
4182 // The object is marked - apply the closure
4183 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4184 bitmap_closure.do_bit(offset);
4185 }
4186 // Even if this task aborted while scanning the humongous object
4187 // we can (and should) give up the current region.
4188 giveup_current_region();
4189 regular_clock_call();
4190 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4191 giveup_current_region();
4192 regular_clock_call();
4193 } else {
4194 assert(has_aborted(), "currently the only way to do so");
4195 // The only way to abort the bitmap iteration is to return
4196 // false from the do_bit() method. However, inside the
4197 // do_bit() method we move the _finger to point to the
4198 // object currently being looked at. So, if we bail out, we
4199 // have definitely set _finger to something non-null.
4200 assert(_finger != NULL, "invariant");
4201
4202 // Region iteration was actually aborted. So now _finger
4203 // points to the address of the object we last scanned. If we
4204 // leave it there, when we restart this task, we will rescan
4205 // the object. It is easy to avoid this. We move the finger by
4206 // enough to point to the next possible object header (the
4207 // bitmap knows by how much we need to move it as it knows its
4208 // granularity).
4209 assert(_finger < _region_limit, "invariant");
4210 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4211 // Check if bitmap iteration was aborted while scanning the last object
4212 if (new_finger >= _region_limit) {
4213 giveup_current_region();
4214 } else {
4215 move_finger_to(new_finger);
4216 }
4217 }
4218 }
4219 // At this point we have either completed iterating over the
4220 // region we were holding on to, or we have aborted.
4221
4222 // We then partially drain the local queue and the global stack.
4223 // (Do we really need this?)
4224 drain_local_queue(true);
4225 drain_global_stack(true);
4226
4227 // Read the note on the claim_region() method on why it might
4228 // return NULL with potentially more regions available for
4229 // claiming and why we have to check out_of_regions() to determine
4230 // whether we're done or not.
4231 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4232 // We are going to try to claim a new region. We should have
4233 // given up on the previous one.
4234 // Separated the asserts so that we know which one fires.
4235 assert(_curr_region == NULL, "invariant");
4236 assert(_finger == NULL, "invariant");
4237 assert(_region_limit == NULL, "invariant");
4238 if (_cm->verbose_low()) {
4239 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4240 }
4241 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4242 if (claimed_region != NULL) {
4243 // Yes, we managed to claim one
4244 statsOnly( ++_regions_claimed );
4245
4246 if (_cm->verbose_low()) {
4247 gclog_or_tty->print_cr("[%u] we successfully claimed "
4248 "region "PTR_FORMAT,
4249 _worker_id, p2i(claimed_region));
4250 }
4251
4252 setup_for_region(claimed_region);
4253 assert(_curr_region == claimed_region, "invariant");
4254 }
4255 // It is important to call the regular clock here. It might take
4256 // a while to claim a region if, for example, we hit a large
4257 // block of empty regions. So we need to call the regular clock
4258 // method once round the loop to make sure it's called
4259 // frequently enough.
4260 regular_clock_call();
4261 }
4262
4263 if (!has_aborted() && _curr_region == NULL) {
4264 assert(_cm->out_of_regions(),
4265 "at this point we should be out of regions");
4266 }
4267 } while ( _curr_region != NULL && !has_aborted());
4268
4269 if (!has_aborted()) {
4270 // We cannot check whether the global stack is empty, since other
4271 // tasks might be pushing objects to it concurrently.
4272 assert(_cm->out_of_regions(),
4273 "at this point we should be out of regions");
4274
4275 if (_cm->verbose_low()) {
4276 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4277 }
4278
4279 // Try to reduce the number of available SATB buffers so that
4280 // remark has less work to do.
4281 drain_satb_buffers();
4282 }
4283
4284 // Since we've done everything else, we can now totally drain the
4285 // local queue and global stack.
4286 drain_local_queue(false);
4287 drain_global_stack(false);
4288
4289 // Attempt at work stealing from other task's queues.
4290 if (do_stealing && !has_aborted()) {
4291 // We have not aborted. This means that we have finished all that
4292 // we could. Let's try to do some stealing...
4293
4294 // We cannot check whether the global stack is empty, since other
4295 // tasks might be pushing objects to it concurrently.
4296 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4297 "only way to reach here");
4298
4299 if (_cm->verbose_low()) {
4300 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4301 }
4302
4303 while (!has_aborted()) {
4304 oop obj;
4305 statsOnly( ++_steal_attempts );
4306
4307 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4308 if (_cm->verbose_medium()) {
4309 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4310 _worker_id, p2i((void*) obj));
4311 }
4312
4313 statsOnly( ++_steals );
4314
4315 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4316 "any stolen object should be marked");
4317 scan_object(obj);
4318
4319 // And since we're towards the end, let's totally drain the
4320 // local queue and global stack.
4321 drain_local_queue(false);
4322 drain_global_stack(false);
4323 } else {
4324 break;
4325 }
4326 }
4327 }
4328
4329 // If we are about to wrap up and go into termination, check if we
4330 // should raise the overflow flag.
4331 if (do_termination && !has_aborted()) {
4332 if (_cm->force_overflow()->should_force()) {
4333 _cm->set_has_overflown();
4334 regular_clock_call();
4335 }
4336 }
4337
4338 // We still haven't aborted. Now, let's try to get into the
4339 // termination protocol.
4340 if (do_termination && !has_aborted()) {
4341 // We cannot check whether the global stack is empty, since other
4342 // tasks might be concurrently pushing objects on it.
4343 // Separated the asserts so that we know which one fires.
4344 assert(_cm->out_of_regions(), "only way to reach here");
4345 assert(_task_queue->size() == 0, "only way to reach here");
4346
4347 if (_cm->verbose_low()) {
4348 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4349 }
4350
4351 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4352
4353 // The CMTask class also extends the TerminatorTerminator class,
4354 // hence its should_exit_termination() method will also decide
4355 // whether to exit the termination protocol or not.
4356 bool finished = (is_serial ||
4357 _cm->terminator()->offer_termination(this));
4358 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4359 _termination_time_ms +=
4360 termination_end_time_ms - _termination_start_time_ms;
4361
4362 if (finished) {
4363 // We're all done.
4364
4365 if (_worker_id == 0) {
4366 // let's allow task 0 to do this
4367 if (concurrent()) {
4368 assert(_cm->concurrent_marking_in_progress(), "invariant");
4369 // we need to set this to false before the next
4370 // safepoint. This way we ensure that the marking phase
4371 // doesn't observe any more heap expansions.
4372 _cm->clear_concurrent_marking_in_progress();
4373 }
4374 }
4375
4376 // We can now guarantee that the global stack is empty, since
4377 // all other tasks have finished. We separated the guarantees so
4378 // that, if a condition is false, we can immediately find out
4379 // which one.
4380 guarantee(_cm->out_of_regions(), "only way to reach here");
4381 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4382 guarantee(_task_queue->size() == 0, "only way to reach here");
4383 guarantee(!_cm->has_overflown(), "only way to reach here");
4384 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4385
4386 if (_cm->verbose_low()) {
4387 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4388 }
4389 } else {
4390 // Apparently there's more work to do. Let's abort this task. It
4391 // will restart it and we can hopefully find more things to do.
4392
4393 if (_cm->verbose_low()) {
4394 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4395 _worker_id);
4396 }
4397
4398 set_has_aborted();
4399 statsOnly( ++_aborted_termination );
4400 }
4401 }
4402
4403 // Mainly for debugging purposes to make sure that a pointer to the
4404 // closure which was statically allocated in this frame doesn't
4405 // escape it by accident.
4406 set_cm_oop_closure(NULL);
4407 double end_time_ms = os::elapsedVTime() * 1000.0;
4408 double elapsed_time_ms = end_time_ms - _start_time_ms;
4409 // Update the step history.
4410 _step_times_ms.add(elapsed_time_ms);
4411
4412 if (has_aborted()) {
4413 // The task was aborted for some reason.
4414
4415 statsOnly( ++_aborted );
4416
4417 if (_has_timed_out) {
4418 double diff_ms = elapsed_time_ms - _time_target_ms;
4419 // Keep statistics of how well we did with respect to hitting
4420 // our target only if we actually timed out (if we aborted for
4421 // other reasons, then the results might get skewed).
4422 _marking_step_diffs_ms.add(diff_ms);
4423 }
4424
4425 if (_cm->has_overflown()) {
4426 // This is the interesting one. We aborted because a global
4427 // overflow was raised. This means we have to restart the
4428 // marking phase and start iterating over regions. However, in
4429 // order to do this we have to make sure that all tasks stop
4430 // what they are doing and re-initialize in a safe manner. We
4431 // will achieve this with the use of two barrier sync points.
4432
4433 if (_cm->verbose_low()) {
4434 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4435 }
4436
4437 if (!is_serial) {
4438 // We only need to enter the sync barrier if being called
4439 // from a parallel context
4440 _cm->enter_first_sync_barrier(_worker_id);
4441
4442 // When we exit this sync barrier we know that all tasks have
4443 // stopped doing marking work. So, it's now safe to
4444 // re-initialize our data structures. At the end of this method,
4445 // task 0 will clear the global data structures.
4446 }
4447
4448 statsOnly( ++_aborted_overflow );
4449
4450 // We clear the local state of this task...
4451 clear_region_fields();
4452
4453 if (!is_serial) {
4454 // ...and enter the second barrier.
4455 _cm->enter_second_sync_barrier(_worker_id);
4456 }
4457 // At this point, if we're during the concurrent phase of
4458 // marking, everything has been re-initialized and we're
4459 // ready to restart.
4460 }
4461
4462 if (_cm->verbose_low()) {
4463 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4464 "elapsed = %1.2lfms <<<<<<<<<<",
4465 _worker_id, _time_target_ms, elapsed_time_ms);
4466 if (_cm->has_aborted()) {
4467 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4468 _worker_id);
4469 }
4470 }
4471 } else {
4472 if (_cm->verbose_low()) {
4473 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4474 "elapsed = %1.2lfms <<<<<<<<<<",
4475 _worker_id, _time_target_ms, elapsed_time_ms);
4476 }
4477 }
4478
4479 _claimed = false;
4480 }
4481
4482 CMTask::CMTask(uint worker_id,
4483 ConcurrentMark* cm,
4484 size_t* marked_bytes,
4485 BitMap* card_bm,
4486 CMTaskQueue* task_queue,
4487 CMTaskQueueSet* task_queues)
4488 : _g1h(G1CollectedHeap::heap()),
4489 _worker_id(worker_id), _cm(cm),
4490 _claimed(false),
4491 _nextMarkBitMap(NULL), _hash_seed(17),
4492 _task_queue(task_queue),
4493 _task_queues(task_queues),
4494 _cm_oop_closure(NULL),
4495 _marked_bytes_array(marked_bytes),
4496 _card_bm(card_bm) {
4497 guarantee(task_queue != NULL, "invariant");
4498 guarantee(task_queues != NULL, "invariant");
4499
4500 statsOnly( _clock_due_to_scanning = 0;
4501 _clock_due_to_marking = 0 );
4502
4503 _marking_step_diffs_ms.add(0.5);
4504 }
4505
4506 // These are formatting macros that are used below to ensure
4507 // consistent formatting. The *_H_* versions are used to format the
4508 // header for a particular value and they should be kept consistent
4509 // with the corresponding macro. Also note that most of the macros add
4510 // the necessary white space (as a prefix) which makes them a bit
4511 // easier to compose.
4512
4513 // All the output lines are prefixed with this string to be able to
4514 // identify them easily in a large log file.
4515 #define G1PPRL_LINE_PREFIX "###"
4516
4517 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4518 #ifdef _LP64
4519 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4520 #else // _LP64
4521 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4522 #endif // _LP64
4523
4524 // For per-region info
4525 #define G1PPRL_TYPE_FORMAT " %-4s"
4526 #define G1PPRL_TYPE_H_FORMAT " %4s"
4527 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4528 #define G1PPRL_BYTE_H_FORMAT " %9s"
4529 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4530 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4531
4532 // For summary info
4533 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4534 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4535 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4536 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4537
4538 G1PrintRegionLivenessInfoClosure::
4539 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4540 : _out(out),
4541 _total_used_bytes(0), _total_capacity_bytes(0),
4542 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4543 _hum_used_bytes(0), _hum_capacity_bytes(0),
4544 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4545 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4546 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4547 MemRegion g1_reserved = g1h->g1_reserved();
4548 double now = os::elapsedTime();
4549
4550 // Print the header of the output.
4551 _out->cr();
4552 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4553 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4554 G1PPRL_SUM_ADDR_FORMAT("reserved")
4555 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4556 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4557 HeapRegion::GrainBytes);
4558 _out->print_cr(G1PPRL_LINE_PREFIX);
4559 _out->print_cr(G1PPRL_LINE_PREFIX
4560 G1PPRL_TYPE_H_FORMAT
4561 G1PPRL_ADDR_BASE_H_FORMAT
4562 G1PPRL_BYTE_H_FORMAT
4563 G1PPRL_BYTE_H_FORMAT
4564 G1PPRL_BYTE_H_FORMAT
4565 G1PPRL_DOUBLE_H_FORMAT
4566 G1PPRL_BYTE_H_FORMAT
4567 G1PPRL_BYTE_H_FORMAT,
4568 "type", "address-range",
4569 "used", "prev-live", "next-live", "gc-eff",
4570 "remset", "code-roots");
4571 _out->print_cr(G1PPRL_LINE_PREFIX
4572 G1PPRL_TYPE_H_FORMAT
4573 G1PPRL_ADDR_BASE_H_FORMAT
4574 G1PPRL_BYTE_H_FORMAT
4575 G1PPRL_BYTE_H_FORMAT
4576 G1PPRL_BYTE_H_FORMAT
4577 G1PPRL_DOUBLE_H_FORMAT
4578 G1PPRL_BYTE_H_FORMAT
4579 G1PPRL_BYTE_H_FORMAT,
4580 "", "",
4581 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4582 "(bytes)", "(bytes)");
4583 }
4584
4585 // It takes as a parameter a reference to one of the _hum_* fields, it
4586 // deduces the corresponding value for a region in a humongous region
4587 // series (either the region size, or what's left if the _hum_* field
4588 // is < the region size), and updates the _hum_* field accordingly.
4589 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4590 size_t bytes = 0;
4591 // The > 0 check is to deal with the prev and next live bytes which
4592 // could be 0.
4593 if (*hum_bytes > 0) {
4594 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4595 *hum_bytes -= bytes;
4596 }
4597 return bytes;
4598 }
4599
4600 // It deduces the values for a region in a humongous region series
4601 // from the _hum_* fields and updates those accordingly. It assumes
4602 // that that _hum_* fields have already been set up from the "starts
4603 // humongous" region and we visit the regions in address order.
4604 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4605 size_t* capacity_bytes,
4606 size_t* prev_live_bytes,
4607 size_t* next_live_bytes) {
4608 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4609 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4610 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4611 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4612 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4613 }
4614
4615 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4616 const char* type = r->get_type_str();
4617 HeapWord* bottom = r->bottom();
4618 HeapWord* end = r->end();
4619 size_t capacity_bytes = r->capacity();
4620 size_t used_bytes = r->used();
4621 size_t prev_live_bytes = r->live_bytes();
4622 size_t next_live_bytes = r->next_live_bytes();
4623 double gc_eff = r->gc_efficiency();
4624 size_t remset_bytes = r->rem_set()->mem_size();
4625 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4626
4627 if (r->is_starts_humongous()) {
4628 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4629 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4630 "they should have been zeroed after the last time we used them");
4631 // Set up the _hum_* fields.
4632 _hum_capacity_bytes = capacity_bytes;
4633 _hum_used_bytes = used_bytes;
4634 _hum_prev_live_bytes = prev_live_bytes;
4635 _hum_next_live_bytes = next_live_bytes;
4636 get_hum_bytes(&used_bytes, &capacity_bytes,
4637 &prev_live_bytes, &next_live_bytes);
4638 end = bottom + HeapRegion::GrainWords;
4639 } else if (r->is_continues_humongous()) {
4640 get_hum_bytes(&used_bytes, &capacity_bytes,
4641 &prev_live_bytes, &next_live_bytes);
4642 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4643 }
4644
4645 _total_used_bytes += used_bytes;
4646 _total_capacity_bytes += capacity_bytes;
4647 _total_prev_live_bytes += prev_live_bytes;
4648 _total_next_live_bytes += next_live_bytes;
4649 _total_remset_bytes += remset_bytes;
4650 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4651
4652 // Print a line for this particular region.
4653 _out->print_cr(G1PPRL_LINE_PREFIX
4654 G1PPRL_TYPE_FORMAT
4655 G1PPRL_ADDR_BASE_FORMAT
4656 G1PPRL_BYTE_FORMAT
4657 G1PPRL_BYTE_FORMAT
4658 G1PPRL_BYTE_FORMAT
4659 G1PPRL_DOUBLE_FORMAT
4660 G1PPRL_BYTE_FORMAT
4661 G1PPRL_BYTE_FORMAT,
4662 type, p2i(bottom), p2i(end),
4663 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4664 remset_bytes, strong_code_roots_bytes);
4665
4666 return false;
4667 }
4668
4669 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4670 // add static memory usages to remembered set sizes
4671 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4672 // Print the footer of the output.
4673 _out->print_cr(G1PPRL_LINE_PREFIX);
4674 _out->print_cr(G1PPRL_LINE_PREFIX
4675 " SUMMARY"
4676 G1PPRL_SUM_MB_FORMAT("capacity")
4677 G1PPRL_SUM_MB_PERC_FORMAT("used")
4678 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4679 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4680 G1PPRL_SUM_MB_FORMAT("remset")
4681 G1PPRL_SUM_MB_FORMAT("code-roots"),
4682 bytes_to_mb(_total_capacity_bytes),
4683 bytes_to_mb(_total_used_bytes),
4684 perc(_total_used_bytes, _total_capacity_bytes),
4685 bytes_to_mb(_total_prev_live_bytes),
4686 perc(_total_prev_live_bytes, _total_capacity_bytes),
4687 bytes_to_mb(_total_next_live_bytes),
4688 perc(_total_next_live_bytes, _total_capacity_bytes),
4689 bytes_to_mb(_total_remset_bytes),
4690 bytes_to_mb(_total_strong_code_roots_bytes));
4691 _out->cr();
4692 }
--- EOF ---