Mon, 08 Dec 2014 18:57:33 +0100
8067655: Clean up G1 remembered set oop iteration
Summary: Pass on the static type G1ParPushHeapRSClosure to allow oop_iterate devirtualization
Reviewed-by: jmasa, kbarrett
1 /*
2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
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/prefetch.inline.hpp"
53 #include "services/memTracker.hpp"
55 // Concurrent marking bit map wrapper
57 CMBitMapRO::CMBitMapRO(int shifter) :
58 _bm(),
59 _shifter(shifter) {
60 _bmStartWord = 0;
61 _bmWordSize = 0;
62 }
64 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
65 const HeapWord* limit) const {
66 // First we must round addr *up* to a possible object boundary.
67 addr = (HeapWord*)align_size_up((intptr_t)addr,
68 HeapWordSize << _shifter);
69 size_t addrOffset = heapWordToOffset(addr);
70 if (limit == NULL) {
71 limit = _bmStartWord + _bmWordSize;
72 }
73 size_t limitOffset = heapWordToOffset(limit);
74 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
75 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
76 assert(nextAddr >= addr, "get_next_one postcondition");
77 assert(nextAddr == limit || isMarked(nextAddr),
78 "get_next_one postcondition");
79 return nextAddr;
80 }
82 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
83 const HeapWord* limit) const {
84 size_t addrOffset = heapWordToOffset(addr);
85 if (limit == NULL) {
86 limit = _bmStartWord + _bmWordSize;
87 }
88 size_t limitOffset = heapWordToOffset(limit);
89 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
90 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
91 assert(nextAddr >= addr, "get_next_one postcondition");
92 assert(nextAddr == limit || !isMarked(nextAddr),
93 "get_next_one postcondition");
94 return nextAddr;
95 }
97 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
98 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
99 return (int) (diff >> _shifter);
100 }
102 #ifndef PRODUCT
103 bool CMBitMapRO::covers(MemRegion heap_rs) const {
104 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
105 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
106 "size inconsistency");
107 return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
108 _bmWordSize == heap_rs.word_size();
109 }
110 #endif
112 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
113 _bm.print_on_error(st, prefix);
114 }
116 size_t CMBitMap::compute_size(size_t heap_size) {
117 return heap_size / mark_distance();
118 }
120 size_t CMBitMap::mark_distance() {
121 return MinObjAlignmentInBytes * BitsPerByte;
122 }
124 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
125 _bmStartWord = heap.start();
126 _bmWordSize = heap.word_size();
128 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
129 _bm.set_size(_bmWordSize >> _shifter);
131 storage->set_mapping_changed_listener(&_listener);
132 }
134 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
135 if (zero_filled) {
136 return;
137 }
138 // We need to clear the bitmap on commit, removing any existing information.
139 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
140 _bm->clearRange(mr);
141 }
143 // Closure used for clearing the given mark bitmap.
144 class ClearBitmapHRClosure : public HeapRegionClosure {
145 private:
146 ConcurrentMark* _cm;
147 CMBitMap* _bitmap;
148 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration.
149 public:
150 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
151 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
152 }
154 virtual bool doHeapRegion(HeapRegion* r) {
155 size_t const chunk_size_in_words = M / HeapWordSize;
157 HeapWord* cur = r->bottom();
158 HeapWord* const end = r->end();
160 while (cur < end) {
161 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
162 _bitmap->clearRange(mr);
164 cur += chunk_size_in_words;
166 // Abort iteration if after yielding the marking has been aborted.
167 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
168 return true;
169 }
170 // Repeat the asserts from before the start of the closure. We will do them
171 // as asserts here to minimize their overhead on the product. However, we
172 // will have them as guarantees at the beginning / end of the bitmap
173 // clearing to get some checking in the product.
174 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
175 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
176 }
178 return false;
179 }
180 };
182 void CMBitMap::clearAll() {
183 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
184 G1CollectedHeap::heap()->heap_region_iterate(&cl);
185 guarantee(cl.complete(), "Must have completed iteration.");
186 return;
187 }
189 void CMBitMap::markRange(MemRegion mr) {
190 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
191 assert(!mr.is_empty(), "unexpected empty region");
192 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
193 ((HeapWord *) mr.end())),
194 "markRange memory region end is not card aligned");
195 // convert address range into offset range
196 _bm.at_put_range(heapWordToOffset(mr.start()),
197 heapWordToOffset(mr.end()), true);
198 }
200 void CMBitMap::clearRange(MemRegion mr) {
201 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
202 assert(!mr.is_empty(), "unexpected empty region");
203 // convert address range into offset range
204 _bm.at_put_range(heapWordToOffset(mr.start()),
205 heapWordToOffset(mr.end()), false);
206 }
208 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
209 HeapWord* end_addr) {
210 HeapWord* start = getNextMarkedWordAddress(addr);
211 start = MIN2(start, end_addr);
212 HeapWord* end = getNextUnmarkedWordAddress(start);
213 end = MIN2(end, end_addr);
214 assert(start <= end, "Consistency check");
215 MemRegion mr(start, end);
216 if (!mr.is_empty()) {
217 clearRange(mr);
218 }
219 return mr;
220 }
222 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
223 _base(NULL), _cm(cm)
224 #ifdef ASSERT
225 , _drain_in_progress(false)
226 , _drain_in_progress_yields(false)
227 #endif
228 {}
230 bool CMMarkStack::allocate(size_t capacity) {
231 // allocate a stack of the requisite depth
232 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
233 if (!rs.is_reserved()) {
234 warning("ConcurrentMark MarkStack allocation failure");
235 return false;
236 }
237 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
238 if (!_virtual_space.initialize(rs, rs.size())) {
239 warning("ConcurrentMark MarkStack backing store failure");
240 // Release the virtual memory reserved for the marking stack
241 rs.release();
242 return false;
243 }
244 assert(_virtual_space.committed_size() == rs.size(),
245 "Didn't reserve backing store for all of ConcurrentMark stack?");
246 _base = (oop*) _virtual_space.low();
247 setEmpty();
248 _capacity = (jint) capacity;
249 _saved_index = -1;
250 _should_expand = false;
251 NOT_PRODUCT(_max_depth = 0);
252 return true;
253 }
255 void CMMarkStack::expand() {
256 // Called, during remark, if we've overflown the marking stack during marking.
257 assert(isEmpty(), "stack should been emptied while handling overflow");
258 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
259 // Clear expansion flag
260 _should_expand = false;
261 if (_capacity == (jint) MarkStackSizeMax) {
262 if (PrintGCDetails && Verbose) {
263 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
264 }
265 return;
266 }
267 // Double capacity if possible
268 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
269 // Do not give up existing stack until we have managed to
270 // get the double capacity that we desired.
271 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
272 sizeof(oop)));
273 if (rs.is_reserved()) {
274 // Release the backing store associated with old stack
275 _virtual_space.release();
276 // Reinitialize virtual space for new stack
277 if (!_virtual_space.initialize(rs, rs.size())) {
278 fatal("Not enough swap for expanded marking stack capacity");
279 }
280 _base = (oop*)(_virtual_space.low());
281 _index = 0;
282 _capacity = new_capacity;
283 } else {
284 if (PrintGCDetails && Verbose) {
285 // Failed to double capacity, continue;
286 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
287 SIZE_FORMAT"K to " SIZE_FORMAT"K",
288 _capacity / K, new_capacity / K);
289 }
290 }
291 }
293 void CMMarkStack::set_should_expand() {
294 // If we're resetting the marking state because of an
295 // marking stack overflow, record that we should, if
296 // possible, expand the stack.
297 _should_expand = _cm->has_overflown();
298 }
300 CMMarkStack::~CMMarkStack() {
301 if (_base != NULL) {
302 _base = NULL;
303 _virtual_space.release();
304 }
305 }
307 void CMMarkStack::par_push(oop ptr) {
308 while (true) {
309 if (isFull()) {
310 _overflow = true;
311 return;
312 }
313 // Otherwise...
314 jint index = _index;
315 jint next_index = index+1;
316 jint res = Atomic::cmpxchg(next_index, &_index, index);
317 if (res == index) {
318 _base[index] = ptr;
319 // Note that we don't maintain this atomically. We could, but it
320 // doesn't seem necessary.
321 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
322 return;
323 }
324 // Otherwise, we need to try again.
325 }
326 }
328 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
329 while (true) {
330 if (isFull()) {
331 _overflow = true;
332 return;
333 }
334 // Otherwise...
335 jint index = _index;
336 jint next_index = index + n;
337 if (next_index > _capacity) {
338 _overflow = true;
339 return;
340 }
341 jint res = Atomic::cmpxchg(next_index, &_index, index);
342 if (res == index) {
343 for (int i = 0; i < n; i++) {
344 int ind = index + i;
345 assert(ind < _capacity, "By overflow test above.");
346 _base[ind] = ptr_arr[i];
347 }
348 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
349 return;
350 }
351 // Otherwise, we need to try again.
352 }
353 }
355 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
356 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
357 jint start = _index;
358 jint next_index = start + n;
359 if (next_index > _capacity) {
360 _overflow = true;
361 return;
362 }
363 // Otherwise.
364 _index = next_index;
365 for (int i = 0; i < n; i++) {
366 int ind = start + i;
367 assert(ind < _capacity, "By overflow test above.");
368 _base[ind] = ptr_arr[i];
369 }
370 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
371 }
373 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
374 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
375 jint index = _index;
376 if (index == 0) {
377 *n = 0;
378 return false;
379 } else {
380 int k = MIN2(max, index);
381 jint new_ind = index - k;
382 for (int j = 0; j < k; j++) {
383 ptr_arr[j] = _base[new_ind + j];
384 }
385 _index = new_ind;
386 *n = k;
387 return true;
388 }
389 }
391 template<class OopClosureClass>
392 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
393 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
394 || SafepointSynchronize::is_at_safepoint(),
395 "Drain recursion must be yield-safe.");
396 bool res = true;
397 debug_only(_drain_in_progress = true);
398 debug_only(_drain_in_progress_yields = yield_after);
399 while (!isEmpty()) {
400 oop newOop = pop();
401 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
402 assert(newOop->is_oop(), "Expected an oop");
403 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
404 "only grey objects on this stack");
405 newOop->oop_iterate(cl);
406 if (yield_after && _cm->do_yield_check()) {
407 res = false;
408 break;
409 }
410 }
411 debug_only(_drain_in_progress = false);
412 return res;
413 }
415 void CMMarkStack::note_start_of_gc() {
416 assert(_saved_index == -1,
417 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
418 _saved_index = _index;
419 }
421 void CMMarkStack::note_end_of_gc() {
422 // This is intentionally a guarantee, instead of an assert. If we
423 // accidentally add something to the mark stack during GC, it
424 // will be a correctness issue so it's better if we crash. we'll
425 // only check this once per GC anyway, so it won't be a performance
426 // issue in any way.
427 guarantee(_saved_index == _index,
428 err_msg("saved index: %d index: %d", _saved_index, _index));
429 _saved_index = -1;
430 }
432 void CMMarkStack::oops_do(OopClosure* f) {
433 assert(_saved_index == _index,
434 err_msg("saved index: %d index: %d", _saved_index, _index));
435 for (int i = 0; i < _index; i += 1) {
436 f->do_oop(&_base[i]);
437 }
438 }
440 CMRootRegions::CMRootRegions() :
441 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
442 _should_abort(false), _next_survivor(NULL) { }
444 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
445 _young_list = g1h->young_list();
446 _cm = cm;
447 }
449 void CMRootRegions::prepare_for_scan() {
450 assert(!scan_in_progress(), "pre-condition");
452 // Currently, only survivors can be root regions.
453 assert(_next_survivor == NULL, "pre-condition");
454 _next_survivor = _young_list->first_survivor_region();
455 _scan_in_progress = (_next_survivor != NULL);
456 _should_abort = false;
457 }
459 HeapRegion* CMRootRegions::claim_next() {
460 if (_should_abort) {
461 // If someone has set the should_abort flag, we return NULL to
462 // force the caller to bail out of their loop.
463 return NULL;
464 }
466 // Currently, only survivors can be root regions.
467 HeapRegion* res = _next_survivor;
468 if (res != NULL) {
469 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
470 // Read it again in case it changed while we were waiting for the lock.
471 res = _next_survivor;
472 if (res != NULL) {
473 if (res == _young_list->last_survivor_region()) {
474 // We just claimed the last survivor so store NULL to indicate
475 // that we're done.
476 _next_survivor = NULL;
477 } else {
478 _next_survivor = res->get_next_young_region();
479 }
480 } else {
481 // Someone else claimed the last survivor while we were trying
482 // to take the lock so nothing else to do.
483 }
484 }
485 assert(res == NULL || res->is_survivor(), "post-condition");
487 return res;
488 }
490 void CMRootRegions::scan_finished() {
491 assert(scan_in_progress(), "pre-condition");
493 // Currently, only survivors can be root regions.
494 if (!_should_abort) {
495 assert(_next_survivor == NULL, "we should have claimed all survivors");
496 }
497 _next_survivor = NULL;
499 {
500 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
501 _scan_in_progress = false;
502 RootRegionScan_lock->notify_all();
503 }
504 }
506 bool CMRootRegions::wait_until_scan_finished() {
507 if (!scan_in_progress()) return false;
509 {
510 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
511 while (scan_in_progress()) {
512 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
513 }
514 }
515 return true;
516 }
518 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
519 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
520 #endif // _MSC_VER
522 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
523 return MAX2((n_par_threads + 2) / 4, 1U);
524 }
526 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
527 _g1h(g1h),
528 _markBitMap1(),
529 _markBitMap2(),
530 _parallel_marking_threads(0),
531 _max_parallel_marking_threads(0),
532 _sleep_factor(0.0),
533 _marking_task_overhead(1.0),
534 _cleanup_sleep_factor(0.0),
535 _cleanup_task_overhead(1.0),
536 _cleanup_list("Cleanup List"),
537 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
538 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
539 CardTableModRefBS::card_shift,
540 false /* in_resource_area*/),
542 _prevMarkBitMap(&_markBitMap1),
543 _nextMarkBitMap(&_markBitMap2),
545 _markStack(this),
546 // _finger set in set_non_marking_state
548 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
549 // _active_tasks set in set_non_marking_state
550 // _tasks set inside the constructor
551 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
552 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
554 _has_overflown(false),
555 _concurrent(false),
556 _has_aborted(false),
557 _aborted_gc_id(GCId::undefined()),
558 _restart_for_overflow(false),
559 _concurrent_marking_in_progress(false),
561 // _verbose_level set below
563 _init_times(),
564 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
565 _cleanup_times(),
566 _total_counting_time(0.0),
567 _total_rs_scrub_time(0.0),
569 _parallel_workers(NULL),
571 _count_card_bitmaps(NULL),
572 _count_marked_bytes(NULL),
573 _completed_initialization(false) {
574 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
575 if (verbose_level < no_verbose) {
576 verbose_level = no_verbose;
577 }
578 if (verbose_level > high_verbose) {
579 verbose_level = high_verbose;
580 }
581 _verbose_level = verbose_level;
583 if (verbose_low()) {
584 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
585 "heap end = " INTPTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
586 }
588 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
589 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
591 // Create & start a ConcurrentMark thread.
592 _cmThread = new ConcurrentMarkThread(this);
593 assert(cmThread() != NULL, "CM Thread should have been created");
594 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
595 if (_cmThread->osthread() == NULL) {
596 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
597 }
599 assert(CGC_lock != NULL, "Where's the CGC_lock?");
600 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
601 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
603 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
604 satb_qs.set_buffer_size(G1SATBBufferSize);
606 _root_regions.init(_g1h, this);
608 if (ConcGCThreads > ParallelGCThreads) {
609 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
610 "than ParallelGCThreads (" UINTX_FORMAT ").",
611 ConcGCThreads, ParallelGCThreads);
612 return;
613 }
614 if (ParallelGCThreads == 0) {
615 // if we are not running with any parallel GC threads we will not
616 // spawn any marking threads either
617 _parallel_marking_threads = 0;
618 _max_parallel_marking_threads = 0;
619 _sleep_factor = 0.0;
620 _marking_task_overhead = 1.0;
621 } else {
622 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
623 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
624 // if both are set
625 _sleep_factor = 0.0;
626 _marking_task_overhead = 1.0;
627 } else if (G1MarkingOverheadPercent > 0) {
628 // We will calculate the number of parallel marking threads based
629 // on a target overhead with respect to the soft real-time goal
630 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
631 double overall_cm_overhead =
632 (double) MaxGCPauseMillis * marking_overhead /
633 (double) GCPauseIntervalMillis;
634 double cpu_ratio = 1.0 / (double) os::processor_count();
635 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
636 double marking_task_overhead =
637 overall_cm_overhead / marking_thread_num *
638 (double) os::processor_count();
639 double sleep_factor =
640 (1.0 - marking_task_overhead) / marking_task_overhead;
642 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
643 _sleep_factor = sleep_factor;
644 _marking_task_overhead = marking_task_overhead;
645 } else {
646 // Calculate the number of parallel marking threads by scaling
647 // the number of parallel GC threads.
648 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
649 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
650 _sleep_factor = 0.0;
651 _marking_task_overhead = 1.0;
652 }
654 assert(ConcGCThreads > 0, "Should have been set");
655 _parallel_marking_threads = (uint) ConcGCThreads;
656 _max_parallel_marking_threads = _parallel_marking_threads;
658 if (parallel_marking_threads() > 1) {
659 _cleanup_task_overhead = 1.0;
660 } else {
661 _cleanup_task_overhead = marking_task_overhead();
662 }
663 _cleanup_sleep_factor =
664 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
666 #if 0
667 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
668 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
669 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
670 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
671 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
672 #endif
674 guarantee(parallel_marking_threads() > 0, "peace of mind");
675 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
676 _max_parallel_marking_threads, false, true);
677 if (_parallel_workers == NULL) {
678 vm_exit_during_initialization("Failed necessary allocation.");
679 } else {
680 _parallel_workers->initialize_workers();
681 }
682 }
684 if (FLAG_IS_DEFAULT(MarkStackSize)) {
685 uintx mark_stack_size =
686 MIN2(MarkStackSizeMax,
687 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
688 // Verify that the calculated value for MarkStackSize is in range.
689 // It would be nice to use the private utility routine from Arguments.
690 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
691 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
692 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
693 mark_stack_size, (uintx) 1, MarkStackSizeMax);
694 return;
695 }
696 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
697 } else {
698 // Verify MarkStackSize is in range.
699 if (FLAG_IS_CMDLINE(MarkStackSize)) {
700 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
701 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
702 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
703 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
704 MarkStackSize, (uintx) 1, MarkStackSizeMax);
705 return;
706 }
707 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
708 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
709 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
710 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
711 MarkStackSize, MarkStackSizeMax);
712 return;
713 }
714 }
715 }
716 }
718 if (!_markStack.allocate(MarkStackSize)) {
719 warning("Failed to allocate CM marking stack");
720 return;
721 }
723 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
724 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
726 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
727 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
729 BitMap::idx_t card_bm_size = _card_bm.size();
731 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
732 _active_tasks = _max_worker_id;
734 size_t max_regions = (size_t) _g1h->max_regions();
735 for (uint i = 0; i < _max_worker_id; ++i) {
736 CMTaskQueue* task_queue = new CMTaskQueue();
737 task_queue->initialize();
738 _task_queues->register_queue(i, task_queue);
740 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
741 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
743 _tasks[i] = new CMTask(i, this,
744 _count_marked_bytes[i],
745 &_count_card_bitmaps[i],
746 task_queue, _task_queues);
748 _accum_task_vtime[i] = 0.0;
749 }
751 // Calculate the card number for the bottom of the heap. Used
752 // in biasing indexes into the accounting card bitmaps.
753 _heap_bottom_card_num =
754 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
755 CardTableModRefBS::card_shift);
757 // Clear all the liveness counting data
758 clear_all_count_data();
760 // so that the call below can read a sensible value
761 _heap_start = g1h->reserved_region().start();
762 set_non_marking_state();
763 _completed_initialization = true;
764 }
766 void ConcurrentMark::reset() {
767 // Starting values for these two. This should be called in a STW
768 // phase.
769 MemRegion reserved = _g1h->g1_reserved();
770 _heap_start = reserved.start();
771 _heap_end = reserved.end();
773 // Separated the asserts so that we know which one fires.
774 assert(_heap_start != NULL, "heap bounds should look ok");
775 assert(_heap_end != NULL, "heap bounds should look ok");
776 assert(_heap_start < _heap_end, "heap bounds should look ok");
778 // Reset all the marking data structures and any necessary flags
779 reset_marking_state();
781 if (verbose_low()) {
782 gclog_or_tty->print_cr("[global] resetting");
783 }
785 // We do reset all of them, since different phases will use
786 // different number of active threads. So, it's easiest to have all
787 // of them ready.
788 for (uint i = 0; i < _max_worker_id; ++i) {
789 _tasks[i]->reset(_nextMarkBitMap);
790 }
792 // we need this to make sure that the flag is on during the evac
793 // pause with initial mark piggy-backed
794 set_concurrent_marking_in_progress();
795 }
798 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
799 _markStack.set_should_expand();
800 _markStack.setEmpty(); // Also clears the _markStack overflow flag
801 if (clear_overflow) {
802 clear_has_overflown();
803 } else {
804 assert(has_overflown(), "pre-condition");
805 }
806 _finger = _heap_start;
808 for (uint i = 0; i < _max_worker_id; ++i) {
809 CMTaskQueue* queue = _task_queues->queue(i);
810 queue->set_empty();
811 }
812 }
814 void ConcurrentMark::set_concurrency(uint active_tasks) {
815 assert(active_tasks <= _max_worker_id, "we should not have more");
817 _active_tasks = active_tasks;
818 // Need to update the three data structures below according to the
819 // number of active threads for this phase.
820 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
821 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
822 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
823 }
825 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
826 set_concurrency(active_tasks);
828 _concurrent = concurrent;
829 // We propagate this to all tasks, not just the active ones.
830 for (uint i = 0; i < _max_worker_id; ++i)
831 _tasks[i]->set_concurrent(concurrent);
833 if (concurrent) {
834 set_concurrent_marking_in_progress();
835 } else {
836 // We currently assume that the concurrent flag has been set to
837 // false before we start remark. At this point we should also be
838 // in a STW phase.
839 assert(!concurrent_marking_in_progress(), "invariant");
840 assert(out_of_regions(),
841 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
842 p2i(_finger), p2i(_heap_end)));
843 }
844 }
846 void ConcurrentMark::set_non_marking_state() {
847 // We set the global marking state to some default values when we're
848 // not doing marking.
849 reset_marking_state();
850 _active_tasks = 0;
851 clear_concurrent_marking_in_progress();
852 }
854 ConcurrentMark::~ConcurrentMark() {
855 // The ConcurrentMark instance is never freed.
856 ShouldNotReachHere();
857 }
859 void ConcurrentMark::clearNextBitmap() {
860 G1CollectedHeap* g1h = G1CollectedHeap::heap();
862 // Make sure that the concurrent mark thread looks to still be in
863 // the current cycle.
864 guarantee(cmThread()->during_cycle(), "invariant");
866 // We are finishing up the current cycle by clearing the next
867 // marking bitmap and getting it ready for the next cycle. During
868 // this time no other cycle can start. So, let's make sure that this
869 // is the case.
870 guarantee(!g1h->mark_in_progress(), "invariant");
872 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
873 g1h->heap_region_iterate(&cl);
875 // Clear the liveness counting data. If the marking has been aborted, the abort()
876 // call already did that.
877 if (cl.complete()) {
878 clear_all_count_data();
879 }
881 // Repeat the asserts from above.
882 guarantee(cmThread()->during_cycle(), "invariant");
883 guarantee(!g1h->mark_in_progress(), "invariant");
884 }
886 class CheckBitmapClearHRClosure : public HeapRegionClosure {
887 CMBitMap* _bitmap;
888 bool _error;
889 public:
890 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
891 }
893 virtual bool doHeapRegion(HeapRegion* r) {
894 // This closure can be called concurrently to the mutator, so we must make sure
895 // that the result of the getNextMarkedWordAddress() call is compared to the
896 // value passed to it as limit to detect any found bits.
897 // We can use the region's orig_end() for the limit and the comparison value
898 // as it always contains the "real" end of the region that never changes and
899 // has no side effects.
900 // Due to the latter, there can also be no problem with the compiler generating
901 // reloads of the orig_end() call.
902 HeapWord* end = r->orig_end();
903 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
904 }
905 };
907 bool ConcurrentMark::nextMarkBitmapIsClear() {
908 CheckBitmapClearHRClosure cl(_nextMarkBitMap);
909 _g1h->heap_region_iterate(&cl);
910 return cl.complete();
911 }
913 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
914 public:
915 bool doHeapRegion(HeapRegion* r) {
916 if (!r->continuesHumongous()) {
917 r->note_start_of_marking();
918 }
919 return false;
920 }
921 };
923 void ConcurrentMark::checkpointRootsInitialPre() {
924 G1CollectedHeap* g1h = G1CollectedHeap::heap();
925 G1CollectorPolicy* g1p = g1h->g1_policy();
927 _has_aborted = false;
929 #ifndef PRODUCT
930 if (G1PrintReachableAtInitialMark) {
931 print_reachable("at-cycle-start",
932 VerifyOption_G1UsePrevMarking, true /* all */);
933 }
934 #endif
936 // Initialise marking structures. This has to be done in a STW phase.
937 reset();
939 // For each region note start of marking.
940 NoteStartOfMarkHRClosure startcl;
941 g1h->heap_region_iterate(&startcl);
942 }
945 void ConcurrentMark::checkpointRootsInitialPost() {
946 G1CollectedHeap* g1h = G1CollectedHeap::heap();
948 // If we force an overflow during remark, the remark operation will
949 // actually abort and we'll restart concurrent marking. If we always
950 // force an oveflow during remark we'll never actually complete the
951 // marking phase. So, we initilize this here, at the start of the
952 // cycle, so that at the remaining overflow number will decrease at
953 // every remark and we'll eventually not need to cause one.
954 force_overflow_stw()->init();
956 // Start Concurrent Marking weak-reference discovery.
957 ReferenceProcessor* rp = g1h->ref_processor_cm();
958 // enable ("weak") refs discovery
959 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
960 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
962 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
963 // This is the start of the marking cycle, we're expected all
964 // threads to have SATB queues with active set to false.
965 satb_mq_set.set_active_all_threads(true, /* new active value */
966 false /* expected_active */);
968 _root_regions.prepare_for_scan();
970 // update_g1_committed() will be called at the end of an evac pause
971 // when marking is on. So, it's also called at the end of the
972 // initial-mark pause to update the heap end, if the heap expands
973 // during it. No need to call it here.
974 }
976 /*
977 * Notice that in the next two methods, we actually leave the STS
978 * during the barrier sync and join it immediately afterwards. If we
979 * do not do this, the following deadlock can occur: one thread could
980 * be in the barrier sync code, waiting for the other thread to also
981 * sync up, whereas another one could be trying to yield, while also
982 * waiting for the other threads to sync up too.
983 *
984 * Note, however, that this code is also used during remark and in
985 * this case we should not attempt to leave / enter the STS, otherwise
986 * we'll either hit an asseert (debug / fastdebug) or deadlock
987 * (product). So we should only leave / enter the STS if we are
988 * operating concurrently.
989 *
990 * Because the thread that does the sync barrier has left the STS, it
991 * is possible to be suspended for a Full GC or an evacuation pause
992 * could occur. This is actually safe, since the entering the sync
993 * barrier is one of the last things do_marking_step() does, and it
994 * doesn't manipulate any data structures afterwards.
995 */
997 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
998 if (verbose_low()) {
999 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
1000 }
1002 if (concurrent()) {
1003 SuspendibleThreadSet::leave();
1004 }
1006 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
1008 if (concurrent()) {
1009 SuspendibleThreadSet::join();
1010 }
1011 // at this point everyone should have synced up and not be doing any
1012 // more work
1014 if (verbose_low()) {
1015 if (barrier_aborted) {
1016 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1017 } else {
1018 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1019 }
1020 }
1022 if (barrier_aborted) {
1023 // If the barrier aborted we ignore the overflow condition and
1024 // just abort the whole marking phase as quickly as possible.
1025 return;
1026 }
1028 // If we're executing the concurrent phase of marking, reset the marking
1029 // state; otherwise the marking state is reset after reference processing,
1030 // during the remark pause.
1031 // If we reset here as a result of an overflow during the remark we will
1032 // see assertion failures from any subsequent set_concurrency_and_phase()
1033 // calls.
1034 if (concurrent()) {
1035 // let the task associated with with worker 0 do this
1036 if (worker_id == 0) {
1037 // task 0 is responsible for clearing the global data structures
1038 // We should be here because of an overflow. During STW we should
1039 // not clear the overflow flag since we rely on it being true when
1040 // we exit this method to abort the pause and restart concurent
1041 // marking.
1042 reset_marking_state(true /* clear_overflow */);
1043 force_overflow()->update();
1045 if (G1Log::fine()) {
1046 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1047 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1048 }
1049 }
1050 }
1052 // after this, each task should reset its own data structures then
1053 // then go into the second barrier
1054 }
1056 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1057 if (verbose_low()) {
1058 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1059 }
1061 if (concurrent()) {
1062 SuspendibleThreadSet::leave();
1063 }
1065 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1067 if (concurrent()) {
1068 SuspendibleThreadSet::join();
1069 }
1070 // at this point everything should be re-initialized and ready to go
1072 if (verbose_low()) {
1073 if (barrier_aborted) {
1074 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1075 } else {
1076 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1077 }
1078 }
1079 }
1081 #ifndef PRODUCT
1082 void ForceOverflowSettings::init() {
1083 _num_remaining = G1ConcMarkForceOverflow;
1084 _force = false;
1085 update();
1086 }
1088 void ForceOverflowSettings::update() {
1089 if (_num_remaining > 0) {
1090 _num_remaining -= 1;
1091 _force = true;
1092 } else {
1093 _force = false;
1094 }
1095 }
1097 bool ForceOverflowSettings::should_force() {
1098 if (_force) {
1099 _force = false;
1100 return true;
1101 } else {
1102 return false;
1103 }
1104 }
1105 #endif // !PRODUCT
1107 class CMConcurrentMarkingTask: public AbstractGangTask {
1108 private:
1109 ConcurrentMark* _cm;
1110 ConcurrentMarkThread* _cmt;
1112 public:
1113 void work(uint worker_id) {
1114 assert(Thread::current()->is_ConcurrentGC_thread(),
1115 "this should only be done by a conc GC thread");
1116 ResourceMark rm;
1118 double start_vtime = os::elapsedVTime();
1120 SuspendibleThreadSet::join();
1122 assert(worker_id < _cm->active_tasks(), "invariant");
1123 CMTask* the_task = _cm->task(worker_id);
1124 the_task->record_start_time();
1125 if (!_cm->has_aborted()) {
1126 do {
1127 double start_vtime_sec = os::elapsedVTime();
1128 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1130 the_task->do_marking_step(mark_step_duration_ms,
1131 true /* do_termination */,
1132 false /* is_serial*/);
1134 double end_vtime_sec = os::elapsedVTime();
1135 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1136 _cm->clear_has_overflown();
1138 _cm->do_yield_check(worker_id);
1140 jlong sleep_time_ms;
1141 if (!_cm->has_aborted() && the_task->has_aborted()) {
1142 sleep_time_ms =
1143 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1144 SuspendibleThreadSet::leave();
1145 os::sleep(Thread::current(), sleep_time_ms, false);
1146 SuspendibleThreadSet::join();
1147 }
1148 } while (!_cm->has_aborted() && the_task->has_aborted());
1149 }
1150 the_task->record_end_time();
1151 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1153 SuspendibleThreadSet::leave();
1155 double end_vtime = os::elapsedVTime();
1156 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1157 }
1159 CMConcurrentMarkingTask(ConcurrentMark* cm,
1160 ConcurrentMarkThread* cmt) :
1161 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1163 ~CMConcurrentMarkingTask() { }
1164 };
1166 // Calculates the number of active workers for a concurrent
1167 // phase.
1168 uint ConcurrentMark::calc_parallel_marking_threads() {
1169 if (G1CollectedHeap::use_parallel_gc_threads()) {
1170 uint n_conc_workers = 0;
1171 if (!UseDynamicNumberOfGCThreads ||
1172 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1173 !ForceDynamicNumberOfGCThreads)) {
1174 n_conc_workers = max_parallel_marking_threads();
1175 } else {
1176 n_conc_workers =
1177 AdaptiveSizePolicy::calc_default_active_workers(
1178 max_parallel_marking_threads(),
1179 1, /* Minimum workers */
1180 parallel_marking_threads(),
1181 Threads::number_of_non_daemon_threads());
1182 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1183 // that scaling has already gone into "_max_parallel_marking_threads".
1184 }
1185 assert(n_conc_workers > 0, "Always need at least 1");
1186 return n_conc_workers;
1187 }
1188 // If we are not running with any parallel GC threads we will not
1189 // have spawned any marking threads either. Hence the number of
1190 // concurrent workers should be 0.
1191 return 0;
1192 }
1194 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1195 // Currently, only survivors can be root regions.
1196 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1197 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1199 const uintx interval = PrefetchScanIntervalInBytes;
1200 HeapWord* curr = hr->bottom();
1201 const HeapWord* end = hr->top();
1202 while (curr < end) {
1203 Prefetch::read(curr, interval);
1204 oop obj = oop(curr);
1205 int size = obj->oop_iterate(&cl);
1206 assert(size == obj->size(), "sanity");
1207 curr += size;
1208 }
1209 }
1211 class CMRootRegionScanTask : public AbstractGangTask {
1212 private:
1213 ConcurrentMark* _cm;
1215 public:
1216 CMRootRegionScanTask(ConcurrentMark* cm) :
1217 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1219 void work(uint worker_id) {
1220 assert(Thread::current()->is_ConcurrentGC_thread(),
1221 "this should only be done by a conc GC thread");
1223 CMRootRegions* root_regions = _cm->root_regions();
1224 HeapRegion* hr = root_regions->claim_next();
1225 while (hr != NULL) {
1226 _cm->scanRootRegion(hr, worker_id);
1227 hr = root_regions->claim_next();
1228 }
1229 }
1230 };
1232 void ConcurrentMark::scanRootRegions() {
1233 // Start of concurrent marking.
1234 ClassLoaderDataGraph::clear_claimed_marks();
1236 // scan_in_progress() will have been set to true only if there was
1237 // at least one root region to scan. So, if it's false, we
1238 // should not attempt to do any further work.
1239 if (root_regions()->scan_in_progress()) {
1240 _parallel_marking_threads = calc_parallel_marking_threads();
1241 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1242 "Maximum number of marking threads exceeded");
1243 uint active_workers = MAX2(1U, parallel_marking_threads());
1245 CMRootRegionScanTask task(this);
1246 if (use_parallel_marking_threads()) {
1247 _parallel_workers->set_active_workers((int) active_workers);
1248 _parallel_workers->run_task(&task);
1249 } else {
1250 task.work(0);
1251 }
1253 // It's possible that has_aborted() is true here without actually
1254 // aborting the survivor scan earlier. This is OK as it's
1255 // mainly used for sanity checking.
1256 root_regions()->scan_finished();
1257 }
1258 }
1260 void ConcurrentMark::markFromRoots() {
1261 // we might be tempted to assert that:
1262 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1263 // "inconsistent argument?");
1264 // However that wouldn't be right, because it's possible that
1265 // a safepoint is indeed in progress as a younger generation
1266 // stop-the-world GC happens even as we mark in this generation.
1268 _restart_for_overflow = false;
1269 force_overflow_conc()->init();
1271 // _g1h has _n_par_threads
1272 _parallel_marking_threads = calc_parallel_marking_threads();
1273 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1274 "Maximum number of marking threads exceeded");
1276 uint active_workers = MAX2(1U, parallel_marking_threads());
1278 // Parallel task terminator is set in "set_concurrency_and_phase()"
1279 set_concurrency_and_phase(active_workers, true /* concurrent */);
1281 CMConcurrentMarkingTask markingTask(this, cmThread());
1282 if (use_parallel_marking_threads()) {
1283 _parallel_workers->set_active_workers((int)active_workers);
1284 // Don't set _n_par_threads because it affects MT in process_roots()
1285 // and the decisions on that MT processing is made elsewhere.
1286 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1287 _parallel_workers->run_task(&markingTask);
1288 } else {
1289 markingTask.work(0);
1290 }
1291 print_stats();
1292 }
1294 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1295 // world is stopped at this checkpoint
1296 assert(SafepointSynchronize::is_at_safepoint(),
1297 "world should be stopped");
1299 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1301 // If a full collection has happened, we shouldn't do this.
1302 if (has_aborted()) {
1303 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1304 return;
1305 }
1307 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1309 if (VerifyDuringGC) {
1310 HandleMark hm; // handle scope
1311 Universe::heap()->prepare_for_verify();
1312 Universe::verify(VerifyOption_G1UsePrevMarking,
1313 " VerifyDuringGC:(before)");
1314 }
1315 g1h->check_bitmaps("Remark Start");
1317 G1CollectorPolicy* g1p = g1h->g1_policy();
1318 g1p->record_concurrent_mark_remark_start();
1320 double start = os::elapsedTime();
1322 checkpointRootsFinalWork();
1324 double mark_work_end = os::elapsedTime();
1326 weakRefsWork(clear_all_soft_refs);
1328 if (has_overflown()) {
1329 // Oops. We overflowed. Restart concurrent marking.
1330 _restart_for_overflow = true;
1331 if (G1TraceMarkStackOverflow) {
1332 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1333 }
1335 // Verify the heap w.r.t. the previous marking bitmap.
1336 if (VerifyDuringGC) {
1337 HandleMark hm; // handle scope
1338 Universe::heap()->prepare_for_verify();
1339 Universe::verify(VerifyOption_G1UsePrevMarking,
1340 " VerifyDuringGC:(overflow)");
1341 }
1343 // Clear the marking state because we will be restarting
1344 // marking due to overflowing the global mark stack.
1345 reset_marking_state();
1346 } else {
1347 // Aggregate the per-task counting data that we have accumulated
1348 // while marking.
1349 aggregate_count_data();
1351 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1352 // We're done with marking.
1353 // This is the end of the marking cycle, we're expected all
1354 // threads to have SATB queues with active set to true.
1355 satb_mq_set.set_active_all_threads(false, /* new active value */
1356 true /* expected_active */);
1358 if (VerifyDuringGC) {
1359 HandleMark hm; // handle scope
1360 Universe::heap()->prepare_for_verify();
1361 Universe::verify(VerifyOption_G1UseNextMarking,
1362 " VerifyDuringGC:(after)");
1363 }
1364 g1h->check_bitmaps("Remark End");
1365 assert(!restart_for_overflow(), "sanity");
1366 // Completely reset the marking state since marking completed
1367 set_non_marking_state();
1368 }
1370 // Expand the marking stack, if we have to and if we can.
1371 if (_markStack.should_expand()) {
1372 _markStack.expand();
1373 }
1375 // Statistics
1376 double now = os::elapsedTime();
1377 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1378 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1379 _remark_times.add((now - start) * 1000.0);
1381 g1p->record_concurrent_mark_remark_end();
1383 G1CMIsAliveClosure is_alive(g1h);
1384 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1385 }
1387 // Base class of the closures that finalize and verify the
1388 // liveness counting data.
1389 class CMCountDataClosureBase: public HeapRegionClosure {
1390 protected:
1391 G1CollectedHeap* _g1h;
1392 ConcurrentMark* _cm;
1393 CardTableModRefBS* _ct_bs;
1395 BitMap* _region_bm;
1396 BitMap* _card_bm;
1398 // Takes a region that's not empty (i.e., it has at least one
1399 // live object in it and sets its corresponding bit on the region
1400 // bitmap to 1. If the region is "starts humongous" it will also set
1401 // to 1 the bits on the region bitmap that correspond to its
1402 // associated "continues humongous" regions.
1403 void set_bit_for_region(HeapRegion* hr) {
1404 assert(!hr->continuesHumongous(), "should have filtered those out");
1406 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1407 if (!hr->startsHumongous()) {
1408 // Normal (non-humongous) case: just set the bit.
1409 _region_bm->par_at_put(index, true);
1410 } else {
1411 // Starts humongous case: calculate how many regions are part of
1412 // this humongous region and then set the bit range.
1413 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1414 _region_bm->par_at_put_range(index, end_index, true);
1415 }
1416 }
1418 public:
1419 CMCountDataClosureBase(G1CollectedHeap* g1h,
1420 BitMap* region_bm, BitMap* card_bm):
1421 _g1h(g1h), _cm(g1h->concurrent_mark()),
1422 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1423 _region_bm(region_bm), _card_bm(card_bm) { }
1424 };
1426 // Closure that calculates the # live objects per region. Used
1427 // for verification purposes during the cleanup pause.
1428 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1429 CMBitMapRO* _bm;
1430 size_t _region_marked_bytes;
1432 public:
1433 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1434 BitMap* region_bm, BitMap* card_bm) :
1435 CMCountDataClosureBase(g1h, region_bm, card_bm),
1436 _bm(bm), _region_marked_bytes(0) { }
1438 bool doHeapRegion(HeapRegion* hr) {
1440 if (hr->continuesHumongous()) {
1441 // We will ignore these here and process them when their
1442 // associated "starts humongous" region is processed (see
1443 // set_bit_for_heap_region()). Note that we cannot rely on their
1444 // associated "starts humongous" region to have their bit set to
1445 // 1 since, due to the region chunking in the parallel region
1446 // iteration, a "continues humongous" region might be visited
1447 // before its associated "starts humongous".
1448 return false;
1449 }
1451 HeapWord* ntams = hr->next_top_at_mark_start();
1452 HeapWord* start = hr->bottom();
1454 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1455 err_msg("Preconditions not met - "
1456 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1457 p2i(start), p2i(ntams), p2i(hr->end())));
1459 // Find the first marked object at or after "start".
1460 start = _bm->getNextMarkedWordAddress(start, ntams);
1462 size_t marked_bytes = 0;
1464 while (start < ntams) {
1465 oop obj = oop(start);
1466 int obj_sz = obj->size();
1467 HeapWord* obj_end = start + obj_sz;
1469 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1470 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1472 // Note: if we're looking at the last region in heap - obj_end
1473 // could be actually just beyond the end of the heap; end_idx
1474 // will then correspond to a (non-existent) card that is also
1475 // just beyond the heap.
1476 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1477 // end of object is not card aligned - increment to cover
1478 // all the cards spanned by the object
1479 end_idx += 1;
1480 }
1482 // Set the bits in the card BM for the cards spanned by this object.
1483 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1485 // Add the size of this object to the number of marked bytes.
1486 marked_bytes += (size_t)obj_sz * HeapWordSize;
1488 // Find the next marked object after this one.
1489 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1490 }
1492 // Mark the allocated-since-marking portion...
1493 HeapWord* top = hr->top();
1494 if (ntams < top) {
1495 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1496 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1498 // Note: if we're looking at the last region in heap - top
1499 // could be actually just beyond the end of the heap; end_idx
1500 // will then correspond to a (non-existent) card that is also
1501 // just beyond the heap.
1502 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1503 // end of object is not card aligned - increment to cover
1504 // all the cards spanned by the object
1505 end_idx += 1;
1506 }
1507 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1509 // This definitely means the region has live objects.
1510 set_bit_for_region(hr);
1511 }
1513 // Update the live region bitmap.
1514 if (marked_bytes > 0) {
1515 set_bit_for_region(hr);
1516 }
1518 // Set the marked bytes for the current region so that
1519 // it can be queried by a calling verificiation routine
1520 _region_marked_bytes = marked_bytes;
1522 return false;
1523 }
1525 size_t region_marked_bytes() const { return _region_marked_bytes; }
1526 };
1528 // Heap region closure used for verifying the counting data
1529 // that was accumulated concurrently and aggregated during
1530 // the remark pause. This closure is applied to the heap
1531 // regions during the STW cleanup pause.
1533 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1534 G1CollectedHeap* _g1h;
1535 ConcurrentMark* _cm;
1536 CalcLiveObjectsClosure _calc_cl;
1537 BitMap* _region_bm; // Region BM to be verified
1538 BitMap* _card_bm; // Card BM to be verified
1539 bool _verbose; // verbose output?
1541 BitMap* _exp_region_bm; // Expected Region BM values
1542 BitMap* _exp_card_bm; // Expected card BM values
1544 int _failures;
1546 public:
1547 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1548 BitMap* region_bm,
1549 BitMap* card_bm,
1550 BitMap* exp_region_bm,
1551 BitMap* exp_card_bm,
1552 bool verbose) :
1553 _g1h(g1h), _cm(g1h->concurrent_mark()),
1554 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1555 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1556 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1557 _failures(0) { }
1559 int failures() const { return _failures; }
1561 bool doHeapRegion(HeapRegion* hr) {
1562 if (hr->continuesHumongous()) {
1563 // We will ignore these here and process them when their
1564 // associated "starts humongous" region is processed (see
1565 // set_bit_for_heap_region()). Note that we cannot rely on their
1566 // associated "starts humongous" region to have their bit set to
1567 // 1 since, due to the region chunking in the parallel region
1568 // iteration, a "continues humongous" region might be visited
1569 // before its associated "starts humongous".
1570 return false;
1571 }
1573 int failures = 0;
1575 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1576 // this region and set the corresponding bits in the expected region
1577 // and card bitmaps.
1578 bool res = _calc_cl.doHeapRegion(hr);
1579 assert(res == false, "should be continuing");
1581 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1582 Mutex::_no_safepoint_check_flag);
1584 // Verify the marked bytes for this region.
1585 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1586 size_t act_marked_bytes = hr->next_marked_bytes();
1588 // We're not OK if expected marked bytes > actual marked bytes. It means
1589 // we have missed accounting some objects during the actual marking.
1590 if (exp_marked_bytes > act_marked_bytes) {
1591 if (_verbose) {
1592 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1593 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1594 hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1595 }
1596 failures += 1;
1597 }
1599 // Verify the bit, for this region, in the actual and expected
1600 // (which was just calculated) region bit maps.
1601 // We're not OK if the bit in the calculated expected region
1602 // bitmap is set and the bit in the actual region bitmap is not.
1603 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1605 bool expected = _exp_region_bm->at(index);
1606 bool actual = _region_bm->at(index);
1607 if (expected && !actual) {
1608 if (_verbose) {
1609 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1610 "expected: %s, actual: %s",
1611 hr->hrm_index(),
1612 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1613 }
1614 failures += 1;
1615 }
1617 // Verify that the card bit maps for the cards spanned by the current
1618 // region match. We have an error if we have a set bit in the expected
1619 // bit map and the corresponding bit in the actual bitmap is not set.
1621 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1622 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1624 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1625 expected = _exp_card_bm->at(i);
1626 actual = _card_bm->at(i);
1628 if (expected && !actual) {
1629 if (_verbose) {
1630 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1631 "expected: %s, actual: %s",
1632 hr->hrm_index(), i,
1633 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1634 }
1635 failures += 1;
1636 }
1637 }
1639 if (failures > 0 && _verbose) {
1640 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1641 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1642 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1643 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1644 }
1646 _failures += failures;
1648 // We could stop iteration over the heap when we
1649 // find the first violating region by returning true.
1650 return false;
1651 }
1652 };
1654 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1655 protected:
1656 G1CollectedHeap* _g1h;
1657 ConcurrentMark* _cm;
1658 BitMap* _actual_region_bm;
1659 BitMap* _actual_card_bm;
1661 uint _n_workers;
1663 BitMap* _expected_region_bm;
1664 BitMap* _expected_card_bm;
1666 int _failures;
1667 bool _verbose;
1669 public:
1670 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1671 BitMap* region_bm, BitMap* card_bm,
1672 BitMap* expected_region_bm, BitMap* expected_card_bm)
1673 : AbstractGangTask("G1 verify final counting"),
1674 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1675 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1676 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1677 _failures(0), _verbose(false),
1678 _n_workers(0) {
1679 assert(VerifyDuringGC, "don't call this otherwise");
1681 // Use the value already set as the number of active threads
1682 // in the call to run_task().
1683 if (G1CollectedHeap::use_parallel_gc_threads()) {
1684 assert( _g1h->workers()->active_workers() > 0,
1685 "Should have been previously set");
1686 _n_workers = _g1h->workers()->active_workers();
1687 } else {
1688 _n_workers = 1;
1689 }
1691 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1692 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1694 _verbose = _cm->verbose_medium();
1695 }
1697 void work(uint worker_id) {
1698 assert(worker_id < _n_workers, "invariant");
1700 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1701 _actual_region_bm, _actual_card_bm,
1702 _expected_region_bm,
1703 _expected_card_bm,
1704 _verbose);
1706 if (G1CollectedHeap::use_parallel_gc_threads()) {
1707 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1708 worker_id,
1709 _n_workers,
1710 HeapRegion::VerifyCountClaimValue);
1711 } else {
1712 _g1h->heap_region_iterate(&verify_cl);
1713 }
1715 Atomic::add(verify_cl.failures(), &_failures);
1716 }
1718 int failures() const { return _failures; }
1719 };
1721 // Closure that finalizes the liveness counting data.
1722 // Used during the cleanup pause.
1723 // Sets the bits corresponding to the interval [NTAMS, top]
1724 // (which contains the implicitly live objects) in the
1725 // card liveness bitmap. Also sets the bit for each region,
1726 // containing live data, in the region liveness bitmap.
1728 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1729 public:
1730 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1731 BitMap* region_bm,
1732 BitMap* card_bm) :
1733 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1735 bool doHeapRegion(HeapRegion* hr) {
1737 if (hr->continuesHumongous()) {
1738 // We will ignore these here and process them when their
1739 // associated "starts humongous" region is processed (see
1740 // set_bit_for_heap_region()). Note that we cannot rely on their
1741 // associated "starts humongous" region to have their bit set to
1742 // 1 since, due to the region chunking in the parallel region
1743 // iteration, a "continues humongous" region might be visited
1744 // before its associated "starts humongous".
1745 return false;
1746 }
1748 HeapWord* ntams = hr->next_top_at_mark_start();
1749 HeapWord* top = hr->top();
1751 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1753 // Mark the allocated-since-marking portion...
1754 if (ntams < top) {
1755 // This definitely means the region has live objects.
1756 set_bit_for_region(hr);
1758 // Now set the bits in the card bitmap for [ntams, top)
1759 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1760 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1762 // Note: if we're looking at the last region in heap - top
1763 // could be actually just beyond the end of the heap; end_idx
1764 // will then correspond to a (non-existent) card that is also
1765 // just beyond the heap.
1766 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1767 // end of object is not card aligned - increment to cover
1768 // all the cards spanned by the object
1769 end_idx += 1;
1770 }
1772 assert(end_idx <= _card_bm->size(),
1773 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1774 end_idx, _card_bm->size()));
1775 assert(start_idx < _card_bm->size(),
1776 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1777 start_idx, _card_bm->size()));
1779 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1780 }
1782 // Set the bit for the region if it contains live data
1783 if (hr->next_marked_bytes() > 0) {
1784 set_bit_for_region(hr);
1785 }
1787 return false;
1788 }
1789 };
1791 class G1ParFinalCountTask: public AbstractGangTask {
1792 protected:
1793 G1CollectedHeap* _g1h;
1794 ConcurrentMark* _cm;
1795 BitMap* _actual_region_bm;
1796 BitMap* _actual_card_bm;
1798 uint _n_workers;
1800 public:
1801 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1802 : AbstractGangTask("G1 final counting"),
1803 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1804 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1805 _n_workers(0) {
1806 // Use the value already set as the number of active threads
1807 // in the call to run_task().
1808 if (G1CollectedHeap::use_parallel_gc_threads()) {
1809 assert( _g1h->workers()->active_workers() > 0,
1810 "Should have been previously set");
1811 _n_workers = _g1h->workers()->active_workers();
1812 } else {
1813 _n_workers = 1;
1814 }
1815 }
1817 void work(uint worker_id) {
1818 assert(worker_id < _n_workers, "invariant");
1820 FinalCountDataUpdateClosure final_update_cl(_g1h,
1821 _actual_region_bm,
1822 _actual_card_bm);
1824 if (G1CollectedHeap::use_parallel_gc_threads()) {
1825 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1826 worker_id,
1827 _n_workers,
1828 HeapRegion::FinalCountClaimValue);
1829 } else {
1830 _g1h->heap_region_iterate(&final_update_cl);
1831 }
1832 }
1833 };
1835 class G1ParNoteEndTask;
1837 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1838 G1CollectedHeap* _g1;
1839 size_t _max_live_bytes;
1840 uint _regions_claimed;
1841 size_t _freed_bytes;
1842 FreeRegionList* _local_cleanup_list;
1843 HeapRegionSetCount _old_regions_removed;
1844 HeapRegionSetCount _humongous_regions_removed;
1845 HRRSCleanupTask* _hrrs_cleanup_task;
1846 double _claimed_region_time;
1847 double _max_region_time;
1849 public:
1850 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1851 FreeRegionList* local_cleanup_list,
1852 HRRSCleanupTask* hrrs_cleanup_task) :
1853 _g1(g1),
1854 _max_live_bytes(0), _regions_claimed(0),
1855 _freed_bytes(0),
1856 _claimed_region_time(0.0), _max_region_time(0.0),
1857 _local_cleanup_list(local_cleanup_list),
1858 _old_regions_removed(),
1859 _humongous_regions_removed(),
1860 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1862 size_t freed_bytes() { return _freed_bytes; }
1863 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1864 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1866 bool doHeapRegion(HeapRegion *hr) {
1867 if (hr->continuesHumongous()) {
1868 return false;
1869 }
1870 // We use a claim value of zero here because all regions
1871 // were claimed with value 1 in the FinalCount task.
1872 _g1->reset_gc_time_stamps(hr);
1873 double start = os::elapsedTime();
1874 _regions_claimed++;
1875 hr->note_end_of_marking();
1876 _max_live_bytes += hr->max_live_bytes();
1878 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1879 _freed_bytes += hr->used();
1880 hr->set_containing_set(NULL);
1881 if (hr->isHumongous()) {
1882 assert(hr->startsHumongous(), "we should only see starts humongous");
1883 _humongous_regions_removed.increment(1u, hr->capacity());
1884 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1885 } else {
1886 _old_regions_removed.increment(1u, hr->capacity());
1887 _g1->free_region(hr, _local_cleanup_list, true);
1888 }
1889 } else {
1890 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1891 }
1893 double region_time = (os::elapsedTime() - start);
1894 _claimed_region_time += region_time;
1895 if (region_time > _max_region_time) {
1896 _max_region_time = region_time;
1897 }
1898 return false;
1899 }
1901 size_t max_live_bytes() { return _max_live_bytes; }
1902 uint regions_claimed() { return _regions_claimed; }
1903 double claimed_region_time_sec() { return _claimed_region_time; }
1904 double max_region_time_sec() { return _max_region_time; }
1905 };
1907 class G1ParNoteEndTask: public AbstractGangTask {
1908 friend class G1NoteEndOfConcMarkClosure;
1910 protected:
1911 G1CollectedHeap* _g1h;
1912 size_t _max_live_bytes;
1913 size_t _freed_bytes;
1914 FreeRegionList* _cleanup_list;
1916 public:
1917 G1ParNoteEndTask(G1CollectedHeap* g1h,
1918 FreeRegionList* cleanup_list) :
1919 AbstractGangTask("G1 note end"), _g1h(g1h),
1920 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1922 void work(uint worker_id) {
1923 double start = os::elapsedTime();
1924 FreeRegionList local_cleanup_list("Local Cleanup List");
1925 HRRSCleanupTask hrrs_cleanup_task;
1926 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1927 &hrrs_cleanup_task);
1928 if (G1CollectedHeap::use_parallel_gc_threads()) {
1929 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1930 _g1h->workers()->active_workers(),
1931 HeapRegion::NoteEndClaimValue);
1932 } else {
1933 _g1h->heap_region_iterate(&g1_note_end);
1934 }
1935 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1937 // Now update the lists
1938 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1939 {
1940 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1941 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1942 _max_live_bytes += g1_note_end.max_live_bytes();
1943 _freed_bytes += g1_note_end.freed_bytes();
1945 // If we iterate over the global cleanup list at the end of
1946 // cleanup to do this printing we will not guarantee to only
1947 // generate output for the newly-reclaimed regions (the list
1948 // might not be empty at the beginning of cleanup; we might
1949 // still be working on its previous contents). So we do the
1950 // printing here, before we append the new regions to the global
1951 // cleanup list.
1953 G1HRPrinter* hr_printer = _g1h->hr_printer();
1954 if (hr_printer->is_active()) {
1955 FreeRegionListIterator iter(&local_cleanup_list);
1956 while (iter.more_available()) {
1957 HeapRegion* hr = iter.get_next();
1958 hr_printer->cleanup(hr);
1959 }
1960 }
1962 _cleanup_list->add_ordered(&local_cleanup_list);
1963 assert(local_cleanup_list.is_empty(), "post-condition");
1965 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1966 }
1967 }
1968 size_t max_live_bytes() { return _max_live_bytes; }
1969 size_t freed_bytes() { return _freed_bytes; }
1970 };
1972 class G1ParScrubRemSetTask: public AbstractGangTask {
1973 protected:
1974 G1RemSet* _g1rs;
1975 BitMap* _region_bm;
1976 BitMap* _card_bm;
1977 public:
1978 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1979 BitMap* region_bm, BitMap* card_bm) :
1980 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1981 _region_bm(region_bm), _card_bm(card_bm) { }
1983 void work(uint worker_id) {
1984 if (G1CollectedHeap::use_parallel_gc_threads()) {
1985 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1986 HeapRegion::ScrubRemSetClaimValue);
1987 } else {
1988 _g1rs->scrub(_region_bm, _card_bm);
1989 }
1990 }
1992 };
1994 void ConcurrentMark::cleanup() {
1995 // world is stopped at this checkpoint
1996 assert(SafepointSynchronize::is_at_safepoint(),
1997 "world should be stopped");
1998 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2000 // If a full collection has happened, we shouldn't do this.
2001 if (has_aborted()) {
2002 g1h->set_marking_complete(); // So bitmap clearing isn't confused
2003 return;
2004 }
2006 g1h->verify_region_sets_optional();
2008 if (VerifyDuringGC) {
2009 HandleMark hm; // handle scope
2010 Universe::heap()->prepare_for_verify();
2011 Universe::verify(VerifyOption_G1UsePrevMarking,
2012 " VerifyDuringGC:(before)");
2013 }
2014 g1h->check_bitmaps("Cleanup Start");
2016 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
2017 g1p->record_concurrent_mark_cleanup_start();
2019 double start = os::elapsedTime();
2021 HeapRegionRemSet::reset_for_cleanup_tasks();
2023 uint n_workers;
2025 // Do counting once more with the world stopped for good measure.
2026 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2028 if (G1CollectedHeap::use_parallel_gc_threads()) {
2029 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2030 "sanity check");
2032 g1h->set_par_threads();
2033 n_workers = g1h->n_par_threads();
2034 assert(g1h->n_par_threads() == n_workers,
2035 "Should not have been reset");
2036 g1h->workers()->run_task(&g1_par_count_task);
2037 // Done with the parallel phase so reset to 0.
2038 g1h->set_par_threads(0);
2040 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2041 "sanity check");
2042 } else {
2043 n_workers = 1;
2044 g1_par_count_task.work(0);
2045 }
2047 if (VerifyDuringGC) {
2048 // Verify that the counting data accumulated during marking matches
2049 // that calculated by walking the marking bitmap.
2051 // Bitmaps to hold expected values
2052 BitMap expected_region_bm(_region_bm.size(), true);
2053 BitMap expected_card_bm(_card_bm.size(), true);
2055 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2056 &_region_bm,
2057 &_card_bm,
2058 &expected_region_bm,
2059 &expected_card_bm);
2061 if (G1CollectedHeap::use_parallel_gc_threads()) {
2062 g1h->set_par_threads((int)n_workers);
2063 g1h->workers()->run_task(&g1_par_verify_task);
2064 // Done with the parallel phase so reset to 0.
2065 g1h->set_par_threads(0);
2067 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2068 "sanity check");
2069 } else {
2070 g1_par_verify_task.work(0);
2071 }
2073 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2074 }
2076 size_t start_used_bytes = g1h->used();
2077 g1h->set_marking_complete();
2079 double count_end = os::elapsedTime();
2080 double this_final_counting_time = (count_end - start);
2081 _total_counting_time += this_final_counting_time;
2083 if (G1PrintRegionLivenessInfo) {
2084 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2085 _g1h->heap_region_iterate(&cl);
2086 }
2088 // Install newly created mark bitMap as "prev".
2089 swapMarkBitMaps();
2091 g1h->reset_gc_time_stamp();
2093 // Note end of marking in all heap regions.
2094 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2095 if (G1CollectedHeap::use_parallel_gc_threads()) {
2096 g1h->set_par_threads((int)n_workers);
2097 g1h->workers()->run_task(&g1_par_note_end_task);
2098 g1h->set_par_threads(0);
2100 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2101 "sanity check");
2102 } else {
2103 g1_par_note_end_task.work(0);
2104 }
2105 g1h->check_gc_time_stamps();
2107 if (!cleanup_list_is_empty()) {
2108 // The cleanup list is not empty, so we'll have to process it
2109 // concurrently. Notify anyone else that might be wanting free
2110 // regions that there will be more free regions coming soon.
2111 g1h->set_free_regions_coming();
2112 }
2114 // call below, since it affects the metric by which we sort the heap
2115 // regions.
2116 if (G1ScrubRemSets) {
2117 double rs_scrub_start = os::elapsedTime();
2118 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2119 if (G1CollectedHeap::use_parallel_gc_threads()) {
2120 g1h->set_par_threads((int)n_workers);
2121 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2122 g1h->set_par_threads(0);
2124 assert(g1h->check_heap_region_claim_values(
2125 HeapRegion::ScrubRemSetClaimValue),
2126 "sanity check");
2127 } else {
2128 g1_par_scrub_rs_task.work(0);
2129 }
2131 double rs_scrub_end = os::elapsedTime();
2132 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2133 _total_rs_scrub_time += this_rs_scrub_time;
2134 }
2136 // this will also free any regions totally full of garbage objects,
2137 // and sort the regions.
2138 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2140 // Statistics.
2141 double end = os::elapsedTime();
2142 _cleanup_times.add((end - start) * 1000.0);
2144 if (G1Log::fine()) {
2145 g1h->print_size_transition(gclog_or_tty,
2146 start_used_bytes,
2147 g1h->used(),
2148 g1h->capacity());
2149 }
2151 // Clean up will have freed any regions completely full of garbage.
2152 // Update the soft reference policy with the new heap occupancy.
2153 Universe::update_heap_info_at_gc();
2155 if (VerifyDuringGC) {
2156 HandleMark hm; // handle scope
2157 Universe::heap()->prepare_for_verify();
2158 Universe::verify(VerifyOption_G1UsePrevMarking,
2159 " VerifyDuringGC:(after)");
2160 }
2161 g1h->check_bitmaps("Cleanup End");
2163 g1h->verify_region_sets_optional();
2165 // We need to make this be a "collection" so any collection pause that
2166 // races with it goes around and waits for completeCleanup to finish.
2167 g1h->increment_total_collections();
2169 // Clean out dead classes and update Metaspace sizes.
2170 if (ClassUnloadingWithConcurrentMark) {
2171 ClassLoaderDataGraph::purge();
2172 }
2173 MetaspaceGC::compute_new_size();
2175 // We reclaimed old regions so we should calculate the sizes to make
2176 // sure we update the old gen/space data.
2177 g1h->g1mm()->update_sizes();
2178 g1h->allocation_context_stats().update_after_mark();
2180 g1h->trace_heap_after_concurrent_cycle();
2181 }
2183 void ConcurrentMark::completeCleanup() {
2184 if (has_aborted()) return;
2186 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2188 _cleanup_list.verify_optional();
2189 FreeRegionList tmp_free_list("Tmp Free List");
2191 if (G1ConcRegionFreeingVerbose) {
2192 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2193 "cleanup list has %u entries",
2194 _cleanup_list.length());
2195 }
2197 // No one else should be accessing the _cleanup_list at this point,
2198 // so it is not necessary to take any locks
2199 while (!_cleanup_list.is_empty()) {
2200 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2201 assert(hr != NULL, "Got NULL from a non-empty list");
2202 hr->par_clear();
2203 tmp_free_list.add_ordered(hr);
2205 // Instead of adding one region at a time to the secondary_free_list,
2206 // we accumulate them in the local list and move them a few at a
2207 // time. This also cuts down on the number of notify_all() calls
2208 // we do during this process. We'll also append the local list when
2209 // _cleanup_list is empty (which means we just removed the last
2210 // region from the _cleanup_list).
2211 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2212 _cleanup_list.is_empty()) {
2213 if (G1ConcRegionFreeingVerbose) {
2214 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2215 "appending %u entries to the secondary_free_list, "
2216 "cleanup list still has %u entries",
2217 tmp_free_list.length(),
2218 _cleanup_list.length());
2219 }
2221 {
2222 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2223 g1h->secondary_free_list_add(&tmp_free_list);
2224 SecondaryFreeList_lock->notify_all();
2225 }
2227 if (G1StressConcRegionFreeing) {
2228 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2229 os::sleep(Thread::current(), (jlong) 1, false);
2230 }
2231 }
2232 }
2233 }
2234 assert(tmp_free_list.is_empty(), "post-condition");
2235 }
2237 // Supporting Object and Oop closures for reference discovery
2238 // and processing in during marking
2240 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2241 HeapWord* addr = (HeapWord*)obj;
2242 return addr != NULL &&
2243 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2244 }
2246 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2247 // Uses the CMTask associated with a worker thread (for serial reference
2248 // processing the CMTask for worker 0 is used) to preserve (mark) and
2249 // trace referent objects.
2250 //
2251 // Using the CMTask and embedded local queues avoids having the worker
2252 // threads operating on the global mark stack. This reduces the risk
2253 // of overflowing the stack - which we would rather avoid at this late
2254 // state. Also using the tasks' local queues removes the potential
2255 // of the workers interfering with each other that could occur if
2256 // operating on the global stack.
2258 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2259 ConcurrentMark* _cm;
2260 CMTask* _task;
2261 int _ref_counter_limit;
2262 int _ref_counter;
2263 bool _is_serial;
2264 public:
2265 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2266 _cm(cm), _task(task), _is_serial(is_serial),
2267 _ref_counter_limit(G1RefProcDrainInterval) {
2268 assert(_ref_counter_limit > 0, "sanity");
2269 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2270 _ref_counter = _ref_counter_limit;
2271 }
2273 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2274 virtual void do_oop( oop* p) { do_oop_work(p); }
2276 template <class T> void do_oop_work(T* p) {
2277 if (!_cm->has_overflown()) {
2278 oop obj = oopDesc::load_decode_heap_oop(p);
2279 if (_cm->verbose_high()) {
2280 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2281 "*"PTR_FORMAT" = "PTR_FORMAT,
2282 _task->worker_id(), p2i(p), p2i((void*) obj));
2283 }
2285 _task->deal_with_reference(obj);
2286 _ref_counter--;
2288 if (_ref_counter == 0) {
2289 // We have dealt with _ref_counter_limit references, pushing them
2290 // and objects reachable from them on to the local stack (and
2291 // possibly the global stack). Call CMTask::do_marking_step() to
2292 // process these entries.
2293 //
2294 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2295 // there's nothing more to do (i.e. we're done with the entries that
2296 // were pushed as a result of the CMTask::deal_with_reference() calls
2297 // above) or we overflow.
2298 //
2299 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2300 // flag while there may still be some work to do. (See the comment at
2301 // the beginning of CMTask::do_marking_step() for those conditions -
2302 // one of which is reaching the specified time target.) It is only
2303 // when CMTask::do_marking_step() returns without setting the
2304 // has_aborted() flag that the marking step has completed.
2305 do {
2306 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2307 _task->do_marking_step(mark_step_duration_ms,
2308 false /* do_termination */,
2309 _is_serial);
2310 } while (_task->has_aborted() && !_cm->has_overflown());
2311 _ref_counter = _ref_counter_limit;
2312 }
2313 } else {
2314 if (_cm->verbose_high()) {
2315 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2316 }
2317 }
2318 }
2319 };
2321 // 'Drain' oop closure used by both serial and parallel reference processing.
2322 // Uses the CMTask associated with a given worker thread (for serial
2323 // reference processing the CMtask for worker 0 is used). Calls the
2324 // do_marking_step routine, with an unbelievably large timeout value,
2325 // to drain the marking data structures of the remaining entries
2326 // added by the 'keep alive' oop closure above.
2328 class G1CMDrainMarkingStackClosure: public VoidClosure {
2329 ConcurrentMark* _cm;
2330 CMTask* _task;
2331 bool _is_serial;
2332 public:
2333 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2334 _cm(cm), _task(task), _is_serial(is_serial) {
2335 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2336 }
2338 void do_void() {
2339 do {
2340 if (_cm->verbose_high()) {
2341 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2342 _task->worker_id(), BOOL_TO_STR(_is_serial));
2343 }
2345 // We call CMTask::do_marking_step() to completely drain the local
2346 // and global marking stacks of entries pushed by the 'keep alive'
2347 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2348 //
2349 // CMTask::do_marking_step() is called in a loop, which we'll exit
2350 // if there's nothing more to do (i.e. we'completely drained the
2351 // entries that were pushed as a a result of applying the 'keep alive'
2352 // closure to the entries on the discovered ref lists) or we overflow
2353 // the global marking stack.
2354 //
2355 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2356 // flag while there may still be some work to do. (See the comment at
2357 // the beginning of CMTask::do_marking_step() for those conditions -
2358 // one of which is reaching the specified time target.) It is only
2359 // when CMTask::do_marking_step() returns without setting the
2360 // has_aborted() flag that the marking step has completed.
2362 _task->do_marking_step(1000000000.0 /* something very large */,
2363 true /* do_termination */,
2364 _is_serial);
2365 } while (_task->has_aborted() && !_cm->has_overflown());
2366 }
2367 };
2369 // Implementation of AbstractRefProcTaskExecutor for parallel
2370 // reference processing at the end of G1 concurrent marking
2372 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2373 private:
2374 G1CollectedHeap* _g1h;
2375 ConcurrentMark* _cm;
2376 WorkGang* _workers;
2377 int _active_workers;
2379 public:
2380 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2381 ConcurrentMark* cm,
2382 WorkGang* workers,
2383 int n_workers) :
2384 _g1h(g1h), _cm(cm),
2385 _workers(workers), _active_workers(n_workers) { }
2387 // Executes the given task using concurrent marking worker threads.
2388 virtual void execute(ProcessTask& task);
2389 virtual void execute(EnqueueTask& task);
2390 };
2392 class G1CMRefProcTaskProxy: public AbstractGangTask {
2393 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2394 ProcessTask& _proc_task;
2395 G1CollectedHeap* _g1h;
2396 ConcurrentMark* _cm;
2398 public:
2399 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2400 G1CollectedHeap* g1h,
2401 ConcurrentMark* cm) :
2402 AbstractGangTask("Process reference objects in parallel"),
2403 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2404 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2405 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2406 }
2408 virtual void work(uint worker_id) {
2409 ResourceMark rm;
2410 HandleMark hm;
2411 CMTask* task = _cm->task(worker_id);
2412 G1CMIsAliveClosure g1_is_alive(_g1h);
2413 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2414 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2416 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2417 }
2418 };
2420 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2421 assert(_workers != NULL, "Need parallel worker threads.");
2422 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2424 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2426 // We need to reset the concurrency level before each
2427 // proxy task execution, so that the termination protocol
2428 // and overflow handling in CMTask::do_marking_step() knows
2429 // how many workers to wait for.
2430 _cm->set_concurrency(_active_workers);
2431 _g1h->set_par_threads(_active_workers);
2432 _workers->run_task(&proc_task_proxy);
2433 _g1h->set_par_threads(0);
2434 }
2436 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2437 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2438 EnqueueTask& _enq_task;
2440 public:
2441 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2442 AbstractGangTask("Enqueue reference objects in parallel"),
2443 _enq_task(enq_task) { }
2445 virtual void work(uint worker_id) {
2446 _enq_task.work(worker_id);
2447 }
2448 };
2450 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2451 assert(_workers != NULL, "Need parallel worker threads.");
2452 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2454 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2456 // Not strictly necessary but...
2457 //
2458 // We need to reset the concurrency level before each
2459 // proxy task execution, so that the termination protocol
2460 // and overflow handling in CMTask::do_marking_step() knows
2461 // how many workers to wait for.
2462 _cm->set_concurrency(_active_workers);
2463 _g1h->set_par_threads(_active_workers);
2464 _workers->run_task(&enq_task_proxy);
2465 _g1h->set_par_threads(0);
2466 }
2468 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2469 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2470 }
2472 // Helper class to get rid of some boilerplate code.
2473 class G1RemarkGCTraceTime : public GCTraceTime {
2474 static bool doit_and_prepend(bool doit) {
2475 if (doit) {
2476 gclog_or_tty->put(' ');
2477 }
2478 return doit;
2479 }
2481 public:
2482 G1RemarkGCTraceTime(const char* title, bool doit)
2483 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
2484 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
2485 }
2486 };
2488 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2489 if (has_overflown()) {
2490 // Skip processing the discovered references if we have
2491 // overflown the global marking stack. Reference objects
2492 // only get discovered once so it is OK to not
2493 // de-populate the discovered reference lists. We could have,
2494 // but the only benefit would be that, when marking restarts,
2495 // less reference objects are discovered.
2496 return;
2497 }
2499 ResourceMark rm;
2500 HandleMark hm;
2502 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2504 // Is alive closure.
2505 G1CMIsAliveClosure g1_is_alive(g1h);
2507 // Inner scope to exclude the cleaning of the string and symbol
2508 // tables from the displayed time.
2509 {
2510 if (G1Log::finer()) {
2511 gclog_or_tty->put(' ');
2512 }
2513 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm(), concurrent_gc_id());
2515 ReferenceProcessor* rp = g1h->ref_processor_cm();
2517 // See the comment in G1CollectedHeap::ref_processing_init()
2518 // about how reference processing currently works in G1.
2520 // Set the soft reference policy
2521 rp->setup_policy(clear_all_soft_refs);
2522 assert(_markStack.isEmpty(), "mark stack should be empty");
2524 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2525 // in serial reference processing. Note these closures are also
2526 // used for serially processing (by the the current thread) the
2527 // JNI references during parallel reference processing.
2528 //
2529 // These closures do not need to synchronize with the worker
2530 // threads involved in parallel reference processing as these
2531 // instances are executed serially by the current thread (e.g.
2532 // reference processing is not multi-threaded and is thus
2533 // performed by the current thread instead of a gang worker).
2534 //
2535 // The gang tasks involved in parallel reference procssing create
2536 // their own instances of these closures, which do their own
2537 // synchronization among themselves.
2538 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2539 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2541 // We need at least one active thread. If reference processing
2542 // is not multi-threaded we use the current (VMThread) thread,
2543 // otherwise we use the work gang from the G1CollectedHeap and
2544 // we utilize all the worker threads we can.
2545 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2546 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2547 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2549 // Parallel processing task executor.
2550 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2551 g1h->workers(), active_workers);
2552 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2554 // Set the concurrency level. The phase was already set prior to
2555 // executing the remark task.
2556 set_concurrency(active_workers);
2558 // Set the degree of MT processing here. If the discovery was done MT,
2559 // the number of threads involved during discovery could differ from
2560 // the number of active workers. This is OK as long as the discovered
2561 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2562 rp->set_active_mt_degree(active_workers);
2564 // Process the weak references.
2565 const ReferenceProcessorStats& stats =
2566 rp->process_discovered_references(&g1_is_alive,
2567 &g1_keep_alive,
2568 &g1_drain_mark_stack,
2569 executor,
2570 g1h->gc_timer_cm(),
2571 concurrent_gc_id());
2572 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2574 // The do_oop work routines of the keep_alive and drain_marking_stack
2575 // oop closures will set the has_overflown flag if we overflow the
2576 // global marking stack.
2578 assert(_markStack.overflow() || _markStack.isEmpty(),
2579 "mark stack should be empty (unless it overflowed)");
2581 if (_markStack.overflow()) {
2582 // This should have been done already when we tried to push an
2583 // entry on to the global mark stack. But let's do it again.
2584 set_has_overflown();
2585 }
2587 assert(rp->num_q() == active_workers, "why not");
2589 rp->enqueue_discovered_references(executor);
2591 rp->verify_no_references_recorded();
2592 assert(!rp->discovery_enabled(), "Post condition");
2593 }
2595 if (has_overflown()) {
2596 // We can not trust g1_is_alive if the marking stack overflowed
2597 return;
2598 }
2600 assert(_markStack.isEmpty(), "Marking should have completed");
2602 // Unload Klasses, String, Symbols, Code Cache, etc.
2603 {
2604 G1RemarkGCTraceTime trace("Unloading", G1Log::finer());
2606 if (ClassUnloadingWithConcurrentMark) {
2607 // Cleaning of klasses depends on correct information from MetadataMarkOnStack. The CodeCache::mark_on_stack
2608 // part is too slow to be done serially, so it is handled during the weakRefsWorkParallelPart phase.
2609 // Defer the cleaning until we have complete on_stack data.
2610 MetadataOnStackMark md_on_stack(false /* Don't visit the code cache at this point */);
2612 bool purged_classes;
2614 {
2615 G1RemarkGCTraceTime trace("System Dictionary Unloading", G1Log::finest());
2616 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2617 }
2619 {
2620 G1RemarkGCTraceTime trace("Parallel Unloading", G1Log::finest());
2621 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2622 }
2624 {
2625 G1RemarkGCTraceTime trace("Deallocate Metadata", G1Log::finest());
2626 ClassLoaderDataGraph::free_deallocate_lists();
2627 }
2628 }
2630 if (G1StringDedup::is_enabled()) {
2631 G1RemarkGCTraceTime trace("String Deduplication Unlink", G1Log::finest());
2632 G1StringDedup::unlink(&g1_is_alive);
2633 }
2634 }
2635 }
2637 void ConcurrentMark::swapMarkBitMaps() {
2638 CMBitMapRO* temp = _prevMarkBitMap;
2639 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2640 _nextMarkBitMap = (CMBitMap*) temp;
2641 }
2643 class CMObjectClosure;
2645 // Closure for iterating over objects, currently only used for
2646 // processing SATB buffers.
2647 class CMObjectClosure : public ObjectClosure {
2648 private:
2649 CMTask* _task;
2651 public:
2652 void do_object(oop obj) {
2653 _task->deal_with_reference(obj);
2654 }
2656 CMObjectClosure(CMTask* task) : _task(task) { }
2657 };
2659 class G1RemarkThreadsClosure : public ThreadClosure {
2660 CMObjectClosure _cm_obj;
2661 G1CMOopClosure _cm_cl;
2662 MarkingCodeBlobClosure _code_cl;
2663 int _thread_parity;
2664 bool _is_par;
2666 public:
2667 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) :
2668 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2669 _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {}
2671 void do_thread(Thread* thread) {
2672 if (thread->is_Java_thread()) {
2673 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2674 JavaThread* jt = (JavaThread*)thread;
2676 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2677 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2678 // * Alive if on the stack of an executing method
2679 // * Weakly reachable otherwise
2680 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2681 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2682 jt->nmethods_do(&_code_cl);
2684 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2685 }
2686 } else if (thread->is_VM_thread()) {
2687 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2688 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2689 }
2690 }
2691 }
2692 };
2694 class CMRemarkTask: public AbstractGangTask {
2695 private:
2696 ConcurrentMark* _cm;
2697 bool _is_serial;
2698 public:
2699 void work(uint worker_id) {
2700 // Since all available tasks are actually started, we should
2701 // only proceed if we're supposed to be actived.
2702 if (worker_id < _cm->active_tasks()) {
2703 CMTask* task = _cm->task(worker_id);
2704 task->record_start_time();
2705 {
2706 ResourceMark rm;
2707 HandleMark hm;
2709 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial);
2710 Threads::threads_do(&threads_f);
2711 }
2713 do {
2714 task->do_marking_step(1000000000.0 /* something very large */,
2715 true /* do_termination */,
2716 _is_serial);
2717 } while (task->has_aborted() && !_cm->has_overflown());
2718 // If we overflow, then we do not want to restart. We instead
2719 // want to abort remark and do concurrent marking again.
2720 task->record_end_time();
2721 }
2722 }
2724 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2725 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2726 _cm->terminator()->reset_for_reuse(active_workers);
2727 }
2728 };
2730 void ConcurrentMark::checkpointRootsFinalWork() {
2731 ResourceMark rm;
2732 HandleMark hm;
2733 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2735 G1RemarkGCTraceTime trace("Finalize Marking", G1Log::finer());
2737 g1h->ensure_parsability(false);
2739 if (G1CollectedHeap::use_parallel_gc_threads()) {
2740 G1CollectedHeap::StrongRootsScope srs(g1h);
2741 // this is remark, so we'll use up all active threads
2742 uint active_workers = g1h->workers()->active_workers();
2743 if (active_workers == 0) {
2744 assert(active_workers > 0, "Should have been set earlier");
2745 active_workers = (uint) ParallelGCThreads;
2746 g1h->workers()->set_active_workers(active_workers);
2747 }
2748 set_concurrency_and_phase(active_workers, false /* concurrent */);
2749 // Leave _parallel_marking_threads at it's
2750 // value originally calculated in the ConcurrentMark
2751 // constructor and pass values of the active workers
2752 // through the gang in the task.
2754 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2755 // We will start all available threads, even if we decide that the
2756 // active_workers will be fewer. The extra ones will just bail out
2757 // immediately.
2758 g1h->set_par_threads(active_workers);
2759 g1h->workers()->run_task(&remarkTask);
2760 g1h->set_par_threads(0);
2761 } else {
2762 G1CollectedHeap::StrongRootsScope srs(g1h);
2763 uint active_workers = 1;
2764 set_concurrency_and_phase(active_workers, false /* concurrent */);
2766 // Note - if there's no work gang then the VMThread will be
2767 // the thread to execute the remark - serially. We have
2768 // to pass true for the is_serial parameter so that
2769 // CMTask::do_marking_step() doesn't enter the sync
2770 // barriers in the event of an overflow. Doing so will
2771 // cause an assert that the current thread is not a
2772 // concurrent GC thread.
2773 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2774 remarkTask.work(0);
2775 }
2776 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2777 guarantee(has_overflown() ||
2778 satb_mq_set.completed_buffers_num() == 0,
2779 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2780 BOOL_TO_STR(has_overflown()),
2781 satb_mq_set.completed_buffers_num()));
2783 print_stats();
2784 }
2786 #ifndef PRODUCT
2788 class PrintReachableOopClosure: public OopClosure {
2789 private:
2790 G1CollectedHeap* _g1h;
2791 outputStream* _out;
2792 VerifyOption _vo;
2793 bool _all;
2795 public:
2796 PrintReachableOopClosure(outputStream* out,
2797 VerifyOption vo,
2798 bool all) :
2799 _g1h(G1CollectedHeap::heap()),
2800 _out(out), _vo(vo), _all(all) { }
2802 void do_oop(narrowOop* p) { do_oop_work(p); }
2803 void do_oop( oop* p) { do_oop_work(p); }
2805 template <class T> void do_oop_work(T* p) {
2806 oop obj = oopDesc::load_decode_heap_oop(p);
2807 const char* str = NULL;
2808 const char* str2 = "";
2810 if (obj == NULL) {
2811 str = "";
2812 } else if (!_g1h->is_in_g1_reserved(obj)) {
2813 str = " O";
2814 } else {
2815 HeapRegion* hr = _g1h->heap_region_containing(obj);
2816 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2817 bool marked = _g1h->is_marked(obj, _vo);
2819 if (over_tams) {
2820 str = " >";
2821 if (marked) {
2822 str2 = " AND MARKED";
2823 }
2824 } else if (marked) {
2825 str = " M";
2826 } else {
2827 str = " NOT";
2828 }
2829 }
2831 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2832 p2i(p), p2i((void*) obj), str, str2);
2833 }
2834 };
2836 class PrintReachableObjectClosure : public ObjectClosure {
2837 private:
2838 G1CollectedHeap* _g1h;
2839 outputStream* _out;
2840 VerifyOption _vo;
2841 bool _all;
2842 HeapRegion* _hr;
2844 public:
2845 PrintReachableObjectClosure(outputStream* out,
2846 VerifyOption vo,
2847 bool all,
2848 HeapRegion* hr) :
2849 _g1h(G1CollectedHeap::heap()),
2850 _out(out), _vo(vo), _all(all), _hr(hr) { }
2852 void do_object(oop o) {
2853 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2854 bool marked = _g1h->is_marked(o, _vo);
2855 bool print_it = _all || over_tams || marked;
2857 if (print_it) {
2858 _out->print_cr(" "PTR_FORMAT"%s",
2859 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2860 PrintReachableOopClosure oopCl(_out, _vo, _all);
2861 o->oop_iterate_no_header(&oopCl);
2862 }
2863 }
2864 };
2866 class PrintReachableRegionClosure : public HeapRegionClosure {
2867 private:
2868 G1CollectedHeap* _g1h;
2869 outputStream* _out;
2870 VerifyOption _vo;
2871 bool _all;
2873 public:
2874 bool doHeapRegion(HeapRegion* hr) {
2875 HeapWord* b = hr->bottom();
2876 HeapWord* e = hr->end();
2877 HeapWord* t = hr->top();
2878 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2879 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2880 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2881 _out->cr();
2883 HeapWord* from = b;
2884 HeapWord* to = t;
2886 if (to > from) {
2887 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2888 _out->cr();
2889 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2890 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2891 _out->cr();
2892 }
2894 return false;
2895 }
2897 PrintReachableRegionClosure(outputStream* out,
2898 VerifyOption vo,
2899 bool all) :
2900 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2901 };
2903 void ConcurrentMark::print_reachable(const char* str,
2904 VerifyOption vo,
2905 bool all) {
2906 gclog_or_tty->cr();
2907 gclog_or_tty->print_cr("== Doing heap dump... ");
2909 if (G1PrintReachableBaseFile == NULL) {
2910 gclog_or_tty->print_cr(" #### error: no base file defined");
2911 return;
2912 }
2914 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2915 (JVM_MAXPATHLEN - 1)) {
2916 gclog_or_tty->print_cr(" #### error: file name too long");
2917 return;
2918 }
2920 char file_name[JVM_MAXPATHLEN];
2921 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2922 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2924 fileStream fout(file_name);
2925 if (!fout.is_open()) {
2926 gclog_or_tty->print_cr(" #### error: could not open file");
2927 return;
2928 }
2930 outputStream* out = &fout;
2931 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2932 out->cr();
2934 out->print_cr("--- ITERATING OVER REGIONS");
2935 out->cr();
2936 PrintReachableRegionClosure rcl(out, vo, all);
2937 _g1h->heap_region_iterate(&rcl);
2938 out->cr();
2940 gclog_or_tty->print_cr(" done");
2941 gclog_or_tty->flush();
2942 }
2944 #endif // PRODUCT
2946 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2947 // Note we are overriding the read-only view of the prev map here, via
2948 // the cast.
2949 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2950 }
2952 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2953 _nextMarkBitMap->clearRange(mr);
2954 }
2956 HeapRegion*
2957 ConcurrentMark::claim_region(uint worker_id) {
2958 // "checkpoint" the finger
2959 HeapWord* finger = _finger;
2961 // _heap_end will not change underneath our feet; it only changes at
2962 // yield points.
2963 while (finger < _heap_end) {
2964 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2966 // Note on how this code handles humongous regions. In the
2967 // normal case the finger will reach the start of a "starts
2968 // humongous" (SH) region. Its end will either be the end of the
2969 // last "continues humongous" (CH) region in the sequence, or the
2970 // standard end of the SH region (if the SH is the only region in
2971 // the sequence). That way claim_region() will skip over the CH
2972 // regions. However, there is a subtle race between a CM thread
2973 // executing this method and a mutator thread doing a humongous
2974 // object allocation. The two are not mutually exclusive as the CM
2975 // thread does not need to hold the Heap_lock when it gets
2976 // here. So there is a chance that claim_region() will come across
2977 // a free region that's in the progress of becoming a SH or a CH
2978 // region. In the former case, it will either
2979 // a) Miss the update to the region's end, in which case it will
2980 // visit every subsequent CH region, will find their bitmaps
2981 // empty, and do nothing, or
2982 // b) Will observe the update of the region's end (in which case
2983 // it will skip the subsequent CH regions).
2984 // If it comes across a region that suddenly becomes CH, the
2985 // scenario will be similar to b). So, the race between
2986 // claim_region() and a humongous object allocation might force us
2987 // to do a bit of unnecessary work (due to some unnecessary bitmap
2988 // iterations) but it should not introduce and correctness issues.
2989 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2991 // Above heap_region_containing_raw may return NULL as we always scan claim
2992 // until the end of the heap. In this case, just jump to the next region.
2993 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2995 // Is the gap between reading the finger and doing the CAS too long?
2996 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2997 if (res == finger && curr_region != NULL) {
2998 // we succeeded
2999 HeapWord* bottom = curr_region->bottom();
3000 HeapWord* limit = curr_region->next_top_at_mark_start();
3002 if (verbose_low()) {
3003 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
3004 "["PTR_FORMAT", "PTR_FORMAT"), "
3005 "limit = "PTR_FORMAT,
3006 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
3007 }
3009 // notice that _finger == end cannot be guaranteed here since,
3010 // someone else might have moved the finger even further
3011 assert(_finger >= end, "the finger should have moved forward");
3013 if (verbose_low()) {
3014 gclog_or_tty->print_cr("[%u] we were successful with region = "
3015 PTR_FORMAT, worker_id, p2i(curr_region));
3016 }
3018 if (limit > bottom) {
3019 if (verbose_low()) {
3020 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
3021 "returning it ", worker_id, p2i(curr_region));
3022 }
3023 return curr_region;
3024 } else {
3025 assert(limit == bottom,
3026 "the region limit should be at bottom");
3027 if (verbose_low()) {
3028 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
3029 "returning NULL", worker_id, p2i(curr_region));
3030 }
3031 // we return NULL and the caller should try calling
3032 // claim_region() again.
3033 return NULL;
3034 }
3035 } else {
3036 assert(_finger > finger, "the finger should have moved forward");
3037 if (verbose_low()) {
3038 if (curr_region == NULL) {
3039 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
3040 "global finger = "PTR_FORMAT", "
3041 "our finger = "PTR_FORMAT,
3042 worker_id, p2i(_finger), p2i(finger));
3043 } else {
3044 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
3045 "global finger = "PTR_FORMAT", "
3046 "our finger = "PTR_FORMAT,
3047 worker_id, p2i(_finger), p2i(finger));
3048 }
3049 }
3051 // read it again
3052 finger = _finger;
3053 }
3054 }
3056 return NULL;
3057 }
3059 #ifndef PRODUCT
3060 enum VerifyNoCSetOopsPhase {
3061 VerifyNoCSetOopsStack,
3062 VerifyNoCSetOopsQueues,
3063 VerifyNoCSetOopsSATBCompleted,
3064 VerifyNoCSetOopsSATBThread
3065 };
3067 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
3068 private:
3069 G1CollectedHeap* _g1h;
3070 VerifyNoCSetOopsPhase _phase;
3071 int _info;
3073 const char* phase_str() {
3074 switch (_phase) {
3075 case VerifyNoCSetOopsStack: return "Stack";
3076 case VerifyNoCSetOopsQueues: return "Queue";
3077 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
3078 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
3079 default: ShouldNotReachHere();
3080 }
3081 return NULL;
3082 }
3084 void do_object_work(oop obj) {
3085 guarantee(!_g1h->obj_in_cs(obj),
3086 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
3087 p2i((void*) obj), phase_str(), _info));
3088 }
3090 public:
3091 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3093 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3094 _phase = phase;
3095 _info = info;
3096 }
3098 virtual void do_oop(oop* p) {
3099 oop obj = oopDesc::load_decode_heap_oop(p);
3100 do_object_work(obj);
3101 }
3103 virtual void do_oop(narrowOop* p) {
3104 // We should not come across narrow oops while scanning marking
3105 // stacks and SATB buffers.
3106 ShouldNotReachHere();
3107 }
3109 virtual void do_object(oop obj) {
3110 do_object_work(obj);
3111 }
3112 };
3114 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3115 bool verify_enqueued_buffers,
3116 bool verify_thread_buffers,
3117 bool verify_fingers) {
3118 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3119 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3120 return;
3121 }
3123 VerifyNoCSetOopsClosure cl;
3125 if (verify_stacks) {
3126 // Verify entries on the global mark stack
3127 cl.set_phase(VerifyNoCSetOopsStack);
3128 _markStack.oops_do(&cl);
3130 // Verify entries on the task queues
3131 for (uint i = 0; i < _max_worker_id; i += 1) {
3132 cl.set_phase(VerifyNoCSetOopsQueues, i);
3133 CMTaskQueue* queue = _task_queues->queue(i);
3134 queue->oops_do(&cl);
3135 }
3136 }
3138 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3140 // Verify entries on the enqueued SATB buffers
3141 if (verify_enqueued_buffers) {
3142 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3143 satb_qs.iterate_completed_buffers_read_only(&cl);
3144 }
3146 // Verify entries on the per-thread SATB buffers
3147 if (verify_thread_buffers) {
3148 cl.set_phase(VerifyNoCSetOopsSATBThread);
3149 satb_qs.iterate_thread_buffers_read_only(&cl);
3150 }
3152 if (verify_fingers) {
3153 // Verify the global finger
3154 HeapWord* global_finger = finger();
3155 if (global_finger != NULL && global_finger < _heap_end) {
3156 // The global finger always points to a heap region boundary. We
3157 // use heap_region_containing_raw() to get the containing region
3158 // given that the global finger could be pointing to a free region
3159 // which subsequently becomes continues humongous. If that
3160 // happens, heap_region_containing() will return the bottom of the
3161 // corresponding starts humongous region and the check below will
3162 // not hold any more.
3163 // Since we always iterate over all regions, we might get a NULL HeapRegion
3164 // here.
3165 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3166 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3167 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3168 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3169 }
3171 // Verify the task fingers
3172 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3173 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3174 CMTask* task = _tasks[i];
3175 HeapWord* task_finger = task->finger();
3176 if (task_finger != NULL && task_finger < _heap_end) {
3177 // See above note on the global finger verification.
3178 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3179 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3180 !task_hr->in_collection_set(),
3181 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3182 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3183 }
3184 }
3185 }
3186 }
3187 #endif // PRODUCT
3189 // Aggregate the counting data that was constructed concurrently
3190 // with marking.
3191 class AggregateCountDataHRClosure: public HeapRegionClosure {
3192 G1CollectedHeap* _g1h;
3193 ConcurrentMark* _cm;
3194 CardTableModRefBS* _ct_bs;
3195 BitMap* _cm_card_bm;
3196 uint _max_worker_id;
3198 public:
3199 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3200 BitMap* cm_card_bm,
3201 uint max_worker_id) :
3202 _g1h(g1h), _cm(g1h->concurrent_mark()),
3203 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3204 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3206 bool doHeapRegion(HeapRegion* hr) {
3207 if (hr->continuesHumongous()) {
3208 // We will ignore these here and process them when their
3209 // associated "starts humongous" region is processed.
3210 // Note that we cannot rely on their associated
3211 // "starts humongous" region to have their bit set to 1
3212 // since, due to the region chunking in the parallel region
3213 // iteration, a "continues humongous" region might be visited
3214 // before its associated "starts humongous".
3215 return false;
3216 }
3218 HeapWord* start = hr->bottom();
3219 HeapWord* limit = hr->next_top_at_mark_start();
3220 HeapWord* end = hr->end();
3222 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3223 err_msg("Preconditions not met - "
3224 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3225 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3226 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3228 assert(hr->next_marked_bytes() == 0, "Precondition");
3230 if (start == limit) {
3231 // NTAMS of this region has not been set so nothing to do.
3232 return false;
3233 }
3235 // 'start' should be in the heap.
3236 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3237 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3238 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3240 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3241 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3242 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3244 // If ntams is not card aligned then we bump card bitmap index
3245 // for limit so that we get the all the cards spanned by
3246 // the object ending at ntams.
3247 // Note: if this is the last region in the heap then ntams
3248 // could be actually just beyond the end of the the heap;
3249 // limit_idx will then correspond to a (non-existent) card
3250 // that is also outside the heap.
3251 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3252 limit_idx += 1;
3253 }
3255 assert(limit_idx <= end_idx, "or else use atomics");
3257 // Aggregate the "stripe" in the count data associated with hr.
3258 uint hrm_index = hr->hrm_index();
3259 size_t marked_bytes = 0;
3261 for (uint i = 0; i < _max_worker_id; i += 1) {
3262 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3263 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3265 // Fetch the marked_bytes in this region for task i and
3266 // add it to the running total for this region.
3267 marked_bytes += marked_bytes_array[hrm_index];
3269 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3270 // into the global card bitmap.
3271 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3273 while (scan_idx < limit_idx) {
3274 assert(task_card_bm->at(scan_idx) == true, "should be");
3275 _cm_card_bm->set_bit(scan_idx);
3276 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3278 // BitMap::get_next_one_offset() can handle the case when
3279 // its left_offset parameter is greater than its right_offset
3280 // parameter. It does, however, have an early exit if
3281 // left_offset == right_offset. So let's limit the value
3282 // passed in for left offset here.
3283 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3284 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3285 }
3286 }
3288 // Update the marked bytes for this region.
3289 hr->add_to_marked_bytes(marked_bytes);
3291 // Next heap region
3292 return false;
3293 }
3294 };
3296 class G1AggregateCountDataTask: public AbstractGangTask {
3297 protected:
3298 G1CollectedHeap* _g1h;
3299 ConcurrentMark* _cm;
3300 BitMap* _cm_card_bm;
3301 uint _max_worker_id;
3302 int _active_workers;
3304 public:
3305 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3306 ConcurrentMark* cm,
3307 BitMap* cm_card_bm,
3308 uint max_worker_id,
3309 int n_workers) :
3310 AbstractGangTask("Count Aggregation"),
3311 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3312 _max_worker_id(max_worker_id),
3313 _active_workers(n_workers) { }
3315 void work(uint worker_id) {
3316 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3318 if (G1CollectedHeap::use_parallel_gc_threads()) {
3319 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3320 _active_workers,
3321 HeapRegion::AggregateCountClaimValue);
3322 } else {
3323 _g1h->heap_region_iterate(&cl);
3324 }
3325 }
3326 };
3329 void ConcurrentMark::aggregate_count_data() {
3330 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3331 _g1h->workers()->active_workers() :
3332 1);
3334 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3335 _max_worker_id, n_workers);
3337 if (G1CollectedHeap::use_parallel_gc_threads()) {
3338 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3339 "sanity check");
3340 _g1h->set_par_threads(n_workers);
3341 _g1h->workers()->run_task(&g1_par_agg_task);
3342 _g1h->set_par_threads(0);
3344 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3345 "sanity check");
3346 _g1h->reset_heap_region_claim_values();
3347 } else {
3348 g1_par_agg_task.work(0);
3349 }
3350 }
3352 // Clear the per-worker arrays used to store the per-region counting data
3353 void ConcurrentMark::clear_all_count_data() {
3354 // Clear the global card bitmap - it will be filled during
3355 // liveness count aggregation (during remark) and the
3356 // final counting task.
3357 _card_bm.clear();
3359 // Clear the global region bitmap - it will be filled as part
3360 // of the final counting task.
3361 _region_bm.clear();
3363 uint max_regions = _g1h->max_regions();
3364 assert(_max_worker_id > 0, "uninitialized");
3366 for (uint i = 0; i < _max_worker_id; i += 1) {
3367 BitMap* task_card_bm = count_card_bitmap_for(i);
3368 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3370 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3371 assert(marked_bytes_array != NULL, "uninitialized");
3373 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3374 task_card_bm->clear();
3375 }
3376 }
3378 void ConcurrentMark::print_stats() {
3379 if (verbose_stats()) {
3380 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3381 for (size_t i = 0; i < _active_tasks; ++i) {
3382 _tasks[i]->print_stats();
3383 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3384 }
3385 }
3386 }
3388 // abandon current marking iteration due to a Full GC
3389 void ConcurrentMark::abort() {
3390 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3391 // concurrent bitmap clearing.
3392 _nextMarkBitMap->clearAll();
3394 // Note we cannot clear the previous marking bitmap here
3395 // since VerifyDuringGC verifies the objects marked during
3396 // a full GC against the previous bitmap.
3398 // Clear the liveness counting data
3399 clear_all_count_data();
3400 // Empty mark stack
3401 reset_marking_state();
3402 for (uint i = 0; i < _max_worker_id; ++i) {
3403 _tasks[i]->clear_region_fields();
3404 }
3405 _first_overflow_barrier_sync.abort();
3406 _second_overflow_barrier_sync.abort();
3407 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3408 if (!gc_id.is_undefined()) {
3409 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3410 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3411 _aborted_gc_id = gc_id;
3412 }
3413 _has_aborted = true;
3415 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3416 satb_mq_set.abandon_partial_marking();
3417 // This can be called either during or outside marking, we'll read
3418 // the expected_active value from the SATB queue set.
3419 satb_mq_set.set_active_all_threads(
3420 false, /* new active value */
3421 satb_mq_set.is_active() /* expected_active */);
3423 _g1h->trace_heap_after_concurrent_cycle();
3424 _g1h->register_concurrent_cycle_end();
3425 }
3427 const GCId& ConcurrentMark::concurrent_gc_id() {
3428 if (has_aborted()) {
3429 return _aborted_gc_id;
3430 }
3431 return _g1h->gc_tracer_cm()->gc_id();
3432 }
3434 static void print_ms_time_info(const char* prefix, const char* name,
3435 NumberSeq& ns) {
3436 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3437 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3438 if (ns.num() > 0) {
3439 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3440 prefix, ns.sd(), ns.maximum());
3441 }
3442 }
3444 void ConcurrentMark::print_summary_info() {
3445 gclog_or_tty->print_cr(" Concurrent marking:");
3446 print_ms_time_info(" ", "init marks", _init_times);
3447 print_ms_time_info(" ", "remarks", _remark_times);
3448 {
3449 print_ms_time_info(" ", "final marks", _remark_mark_times);
3450 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3452 }
3453 print_ms_time_info(" ", "cleanups", _cleanup_times);
3454 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3455 _total_counting_time,
3456 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3457 (double)_cleanup_times.num()
3458 : 0.0));
3459 if (G1ScrubRemSets) {
3460 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3461 _total_rs_scrub_time,
3462 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3463 (double)_cleanup_times.num()
3464 : 0.0));
3465 }
3466 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3467 (_init_times.sum() + _remark_times.sum() +
3468 _cleanup_times.sum())/1000.0);
3469 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3470 "(%8.2f s marking).",
3471 cmThread()->vtime_accum(),
3472 cmThread()->vtime_mark_accum());
3473 }
3475 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3476 if (use_parallel_marking_threads()) {
3477 _parallel_workers->print_worker_threads_on(st);
3478 }
3479 }
3481 void ConcurrentMark::print_on_error(outputStream* st) const {
3482 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3483 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3484 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3485 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3486 }
3488 // We take a break if someone is trying to stop the world.
3489 bool ConcurrentMark::do_yield_check(uint worker_id) {
3490 if (SuspendibleThreadSet::should_yield()) {
3491 if (worker_id == 0) {
3492 _g1h->g1_policy()->record_concurrent_pause();
3493 }
3494 SuspendibleThreadSet::yield();
3495 return true;
3496 } else {
3497 return false;
3498 }
3499 }
3501 #ifndef PRODUCT
3502 // for debugging purposes
3503 void ConcurrentMark::print_finger() {
3504 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3505 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3506 for (uint i = 0; i < _max_worker_id; ++i) {
3507 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3508 }
3509 gclog_or_tty->cr();
3510 }
3511 #endif
3513 void CMTask::scan_object(oop obj) {
3514 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3516 if (_cm->verbose_high()) {
3517 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3518 _worker_id, p2i((void*) obj));
3519 }
3521 size_t obj_size = obj->size();
3522 _words_scanned += obj_size;
3524 obj->oop_iterate(_cm_oop_closure);
3525 statsOnly( ++_objs_scanned );
3526 check_limits();
3527 }
3529 // Closure for iteration over bitmaps
3530 class CMBitMapClosure : public BitMapClosure {
3531 private:
3532 // the bitmap that is being iterated over
3533 CMBitMap* _nextMarkBitMap;
3534 ConcurrentMark* _cm;
3535 CMTask* _task;
3537 public:
3538 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3539 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3541 bool do_bit(size_t offset) {
3542 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3543 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3544 assert( addr < _cm->finger(), "invariant");
3546 statsOnly( _task->increase_objs_found_on_bitmap() );
3547 assert(addr >= _task->finger(), "invariant");
3549 // We move that task's local finger along.
3550 _task->move_finger_to(addr);
3552 _task->scan_object(oop(addr));
3553 // we only partially drain the local queue and global stack
3554 _task->drain_local_queue(true);
3555 _task->drain_global_stack(true);
3557 // if the has_aborted flag has been raised, we need to bail out of
3558 // the iteration
3559 return !_task->has_aborted();
3560 }
3561 };
3563 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3564 ConcurrentMark* cm,
3565 CMTask* task)
3566 : _g1h(g1h), _cm(cm), _task(task) {
3567 assert(_ref_processor == NULL, "should be initialized to NULL");
3569 if (G1UseConcMarkReferenceProcessing) {
3570 _ref_processor = g1h->ref_processor_cm();
3571 assert(_ref_processor != NULL, "should not be NULL");
3572 }
3573 }
3575 void CMTask::setup_for_region(HeapRegion* hr) {
3576 assert(hr != NULL,
3577 "claim_region() should have filtered out NULL regions");
3578 assert(!hr->continuesHumongous(),
3579 "claim_region() should have filtered out continues humongous regions");
3581 if (_cm->verbose_low()) {
3582 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3583 _worker_id, p2i(hr));
3584 }
3586 _curr_region = hr;
3587 _finger = hr->bottom();
3588 update_region_limit();
3589 }
3591 void CMTask::update_region_limit() {
3592 HeapRegion* hr = _curr_region;
3593 HeapWord* bottom = hr->bottom();
3594 HeapWord* limit = hr->next_top_at_mark_start();
3596 if (limit == bottom) {
3597 if (_cm->verbose_low()) {
3598 gclog_or_tty->print_cr("[%u] found an empty region "
3599 "["PTR_FORMAT", "PTR_FORMAT")",
3600 _worker_id, p2i(bottom), p2i(limit));
3601 }
3602 // The region was collected underneath our feet.
3603 // We set the finger to bottom to ensure that the bitmap
3604 // iteration that will follow this will not do anything.
3605 // (this is not a condition that holds when we set the region up,
3606 // as the region is not supposed to be empty in the first place)
3607 _finger = bottom;
3608 } else if (limit >= _region_limit) {
3609 assert(limit >= _finger, "peace of mind");
3610 } else {
3611 assert(limit < _region_limit, "only way to get here");
3612 // This can happen under some pretty unusual circumstances. An
3613 // evacuation pause empties the region underneath our feet (NTAMS
3614 // at bottom). We then do some allocation in the region (NTAMS
3615 // stays at bottom), followed by the region being used as a GC
3616 // alloc region (NTAMS will move to top() and the objects
3617 // originally below it will be grayed). All objects now marked in
3618 // the region are explicitly grayed, if below the global finger,
3619 // and we do not need in fact to scan anything else. So, we simply
3620 // set _finger to be limit to ensure that the bitmap iteration
3621 // doesn't do anything.
3622 _finger = limit;
3623 }
3625 _region_limit = limit;
3626 }
3628 void CMTask::giveup_current_region() {
3629 assert(_curr_region != NULL, "invariant");
3630 if (_cm->verbose_low()) {
3631 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3632 _worker_id, p2i(_curr_region));
3633 }
3634 clear_region_fields();
3635 }
3637 void CMTask::clear_region_fields() {
3638 // Values for these three fields that indicate that we're not
3639 // holding on to a region.
3640 _curr_region = NULL;
3641 _finger = NULL;
3642 _region_limit = NULL;
3643 }
3645 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3646 if (cm_oop_closure == NULL) {
3647 assert(_cm_oop_closure != NULL, "invariant");
3648 } else {
3649 assert(_cm_oop_closure == NULL, "invariant");
3650 }
3651 _cm_oop_closure = cm_oop_closure;
3652 }
3654 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3655 guarantee(nextMarkBitMap != NULL, "invariant");
3657 if (_cm->verbose_low()) {
3658 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3659 }
3661 _nextMarkBitMap = nextMarkBitMap;
3662 clear_region_fields();
3664 _calls = 0;
3665 _elapsed_time_ms = 0.0;
3666 _termination_time_ms = 0.0;
3667 _termination_start_time_ms = 0.0;
3669 #if _MARKING_STATS_
3670 _local_pushes = 0;
3671 _local_pops = 0;
3672 _local_max_size = 0;
3673 _objs_scanned = 0;
3674 _global_pushes = 0;
3675 _global_pops = 0;
3676 _global_max_size = 0;
3677 _global_transfers_to = 0;
3678 _global_transfers_from = 0;
3679 _regions_claimed = 0;
3680 _objs_found_on_bitmap = 0;
3681 _satb_buffers_processed = 0;
3682 _steal_attempts = 0;
3683 _steals = 0;
3684 _aborted = 0;
3685 _aborted_overflow = 0;
3686 _aborted_cm_aborted = 0;
3687 _aborted_yield = 0;
3688 _aborted_timed_out = 0;
3689 _aborted_satb = 0;
3690 _aborted_termination = 0;
3691 #endif // _MARKING_STATS_
3692 }
3694 bool CMTask::should_exit_termination() {
3695 regular_clock_call();
3696 // This is called when we are in the termination protocol. We should
3697 // quit if, for some reason, this task wants to abort or the global
3698 // stack is not empty (this means that we can get work from it).
3699 return !_cm->mark_stack_empty() || has_aborted();
3700 }
3702 void CMTask::reached_limit() {
3703 assert(_words_scanned >= _words_scanned_limit ||
3704 _refs_reached >= _refs_reached_limit ,
3705 "shouldn't have been called otherwise");
3706 regular_clock_call();
3707 }
3709 void CMTask::regular_clock_call() {
3710 if (has_aborted()) return;
3712 // First, we need to recalculate the words scanned and refs reached
3713 // limits for the next clock call.
3714 recalculate_limits();
3716 // During the regular clock call we do the following
3718 // (1) If an overflow has been flagged, then we abort.
3719 if (_cm->has_overflown()) {
3720 set_has_aborted();
3721 return;
3722 }
3724 // If we are not concurrent (i.e. we're doing remark) we don't need
3725 // to check anything else. The other steps are only needed during
3726 // the concurrent marking phase.
3727 if (!concurrent()) return;
3729 // (2) If marking has been aborted for Full GC, then we also abort.
3730 if (_cm->has_aborted()) {
3731 set_has_aborted();
3732 statsOnly( ++_aborted_cm_aborted );
3733 return;
3734 }
3736 double curr_time_ms = os::elapsedVTime() * 1000.0;
3738 // (3) If marking stats are enabled, then we update the step history.
3739 #if _MARKING_STATS_
3740 if (_words_scanned >= _words_scanned_limit) {
3741 ++_clock_due_to_scanning;
3742 }
3743 if (_refs_reached >= _refs_reached_limit) {
3744 ++_clock_due_to_marking;
3745 }
3747 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3748 _interval_start_time_ms = curr_time_ms;
3749 _all_clock_intervals_ms.add(last_interval_ms);
3751 if (_cm->verbose_medium()) {
3752 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3753 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3754 _worker_id, last_interval_ms,
3755 _words_scanned,
3756 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3757 _refs_reached,
3758 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3759 }
3760 #endif // _MARKING_STATS_
3762 // (4) We check whether we should yield. If we have to, then we abort.
3763 if (SuspendibleThreadSet::should_yield()) {
3764 // We should yield. To do this we abort the task. The caller is
3765 // responsible for yielding.
3766 set_has_aborted();
3767 statsOnly( ++_aborted_yield );
3768 return;
3769 }
3771 // (5) We check whether we've reached our time quota. If we have,
3772 // then we abort.
3773 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3774 if (elapsed_time_ms > _time_target_ms) {
3775 set_has_aborted();
3776 _has_timed_out = true;
3777 statsOnly( ++_aborted_timed_out );
3778 return;
3779 }
3781 // (6) Finally, we check whether there are enough completed STAB
3782 // buffers available for processing. If there are, we abort.
3783 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3784 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3785 if (_cm->verbose_low()) {
3786 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3787 _worker_id);
3788 }
3789 // we do need to process SATB buffers, we'll abort and restart
3790 // the marking task to do so
3791 set_has_aborted();
3792 statsOnly( ++_aborted_satb );
3793 return;
3794 }
3795 }
3797 void CMTask::recalculate_limits() {
3798 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3799 _words_scanned_limit = _real_words_scanned_limit;
3801 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3802 _refs_reached_limit = _real_refs_reached_limit;
3803 }
3805 void CMTask::decrease_limits() {
3806 // This is called when we believe that we're going to do an infrequent
3807 // operation which will increase the per byte scanned cost (i.e. move
3808 // entries to/from the global stack). It basically tries to decrease the
3809 // scanning limit so that the clock is called earlier.
3811 if (_cm->verbose_medium()) {
3812 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3813 }
3815 _words_scanned_limit = _real_words_scanned_limit -
3816 3 * words_scanned_period / 4;
3817 _refs_reached_limit = _real_refs_reached_limit -
3818 3 * refs_reached_period / 4;
3819 }
3821 void CMTask::move_entries_to_global_stack() {
3822 // local array where we'll store the entries that will be popped
3823 // from the local queue
3824 oop buffer[global_stack_transfer_size];
3826 int n = 0;
3827 oop obj;
3828 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3829 buffer[n] = obj;
3830 ++n;
3831 }
3833 if (n > 0) {
3834 // we popped at least one entry from the local queue
3836 statsOnly( ++_global_transfers_to; _local_pops += n );
3838 if (!_cm->mark_stack_push(buffer, n)) {
3839 if (_cm->verbose_low()) {
3840 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3841 _worker_id);
3842 }
3843 set_has_aborted();
3844 } else {
3845 // the transfer was successful
3847 if (_cm->verbose_medium()) {
3848 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3849 _worker_id, n);
3850 }
3851 statsOnly( int tmp_size = _cm->mark_stack_size();
3852 if (tmp_size > _global_max_size) {
3853 _global_max_size = tmp_size;
3854 }
3855 _global_pushes += n );
3856 }
3857 }
3859 // this operation was quite expensive, so decrease the limits
3860 decrease_limits();
3861 }
3863 void CMTask::get_entries_from_global_stack() {
3864 // local array where we'll store the entries that will be popped
3865 // from the global stack.
3866 oop buffer[global_stack_transfer_size];
3867 int n;
3868 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3869 assert(n <= global_stack_transfer_size,
3870 "we should not pop more than the given limit");
3871 if (n > 0) {
3872 // yes, we did actually pop at least one entry
3874 statsOnly( ++_global_transfers_from; _global_pops += n );
3875 if (_cm->verbose_medium()) {
3876 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3877 _worker_id, n);
3878 }
3879 for (int i = 0; i < n; ++i) {
3880 bool success = _task_queue->push(buffer[i]);
3881 // We only call this when the local queue is empty or under a
3882 // given target limit. So, we do not expect this push to fail.
3883 assert(success, "invariant");
3884 }
3886 statsOnly( int tmp_size = _task_queue->size();
3887 if (tmp_size > _local_max_size) {
3888 _local_max_size = tmp_size;
3889 }
3890 _local_pushes += n );
3891 }
3893 // this operation was quite expensive, so decrease the limits
3894 decrease_limits();
3895 }
3897 void CMTask::drain_local_queue(bool partially) {
3898 if (has_aborted()) return;
3900 // Decide what the target size is, depending whether we're going to
3901 // drain it partially (so that other tasks can steal if they run out
3902 // of things to do) or totally (at the very end).
3903 size_t target_size;
3904 if (partially) {
3905 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3906 } else {
3907 target_size = 0;
3908 }
3910 if (_task_queue->size() > target_size) {
3911 if (_cm->verbose_high()) {
3912 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3913 _worker_id, target_size);
3914 }
3916 oop obj;
3917 bool ret = _task_queue->pop_local(obj);
3918 while (ret) {
3919 statsOnly( ++_local_pops );
3921 if (_cm->verbose_high()) {
3922 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3923 p2i((void*) obj));
3924 }
3926 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3927 assert(!_g1h->is_on_master_free_list(
3928 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3930 scan_object(obj);
3932 if (_task_queue->size() <= target_size || has_aborted()) {
3933 ret = false;
3934 } else {
3935 ret = _task_queue->pop_local(obj);
3936 }
3937 }
3939 if (_cm->verbose_high()) {
3940 gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3941 _worker_id, _task_queue->size());
3942 }
3943 }
3944 }
3946 void CMTask::drain_global_stack(bool partially) {
3947 if (has_aborted()) return;
3949 // We have a policy to drain the local queue before we attempt to
3950 // drain the global stack.
3951 assert(partially || _task_queue->size() == 0, "invariant");
3953 // Decide what the target size is, depending whether we're going to
3954 // drain it partially (so that other tasks can steal if they run out
3955 // of things to do) or totally (at the very end). Notice that,
3956 // because we move entries from the global stack in chunks or
3957 // because another task might be doing the same, we might in fact
3958 // drop below the target. But, this is not a problem.
3959 size_t target_size;
3960 if (partially) {
3961 target_size = _cm->partial_mark_stack_size_target();
3962 } else {
3963 target_size = 0;
3964 }
3966 if (_cm->mark_stack_size() > target_size) {
3967 if (_cm->verbose_low()) {
3968 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3969 _worker_id, target_size);
3970 }
3972 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3973 get_entries_from_global_stack();
3974 drain_local_queue(partially);
3975 }
3977 if (_cm->verbose_low()) {
3978 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3979 _worker_id, _cm->mark_stack_size());
3980 }
3981 }
3982 }
3984 // SATB Queue has several assumptions on whether to call the par or
3985 // non-par versions of the methods. this is why some of the code is
3986 // replicated. We should really get rid of the single-threaded version
3987 // of the code to simplify things.
3988 void CMTask::drain_satb_buffers() {
3989 if (has_aborted()) return;
3991 // We set this so that the regular clock knows that we're in the
3992 // middle of draining buffers and doesn't set the abort flag when it
3993 // notices that SATB buffers are available for draining. It'd be
3994 // very counter productive if it did that. :-)
3995 _draining_satb_buffers = true;
3997 CMObjectClosure oc(this);
3998 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3999 if (G1CollectedHeap::use_parallel_gc_threads()) {
4000 satb_mq_set.set_par_closure(_worker_id, &oc);
4001 } else {
4002 satb_mq_set.set_closure(&oc);
4003 }
4005 // This keeps claiming and applying the closure to completed buffers
4006 // until we run out of buffers or we need to abort.
4007 if (G1CollectedHeap::use_parallel_gc_threads()) {
4008 while (!has_aborted() &&
4009 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
4010 if (_cm->verbose_medium()) {
4011 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4012 }
4013 statsOnly( ++_satb_buffers_processed );
4014 regular_clock_call();
4015 }
4016 } else {
4017 while (!has_aborted() &&
4018 satb_mq_set.apply_closure_to_completed_buffer()) {
4019 if (_cm->verbose_medium()) {
4020 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4021 }
4022 statsOnly( ++_satb_buffers_processed );
4023 regular_clock_call();
4024 }
4025 }
4027 _draining_satb_buffers = false;
4029 assert(has_aborted() ||
4030 concurrent() ||
4031 satb_mq_set.completed_buffers_num() == 0, "invariant");
4033 if (G1CollectedHeap::use_parallel_gc_threads()) {
4034 satb_mq_set.set_par_closure(_worker_id, NULL);
4035 } else {
4036 satb_mq_set.set_closure(NULL);
4037 }
4039 // again, this was a potentially expensive operation, decrease the
4040 // limits to get the regular clock call early
4041 decrease_limits();
4042 }
4044 void CMTask::print_stats() {
4045 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
4046 _worker_id, _calls);
4047 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
4048 _elapsed_time_ms, _termination_time_ms);
4049 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4050 _step_times_ms.num(), _step_times_ms.avg(),
4051 _step_times_ms.sd());
4052 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4053 _step_times_ms.maximum(), _step_times_ms.sum());
4055 #if _MARKING_STATS_
4056 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4057 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
4058 _all_clock_intervals_ms.sd());
4059 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4060 _all_clock_intervals_ms.maximum(),
4061 _all_clock_intervals_ms.sum());
4062 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
4063 _clock_due_to_scanning, _clock_due_to_marking);
4064 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
4065 _objs_scanned, _objs_found_on_bitmap);
4066 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
4067 _local_pushes, _local_pops, _local_max_size);
4068 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
4069 _global_pushes, _global_pops, _global_max_size);
4070 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
4071 _global_transfers_to,_global_transfers_from);
4072 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
4073 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
4074 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
4075 _steal_attempts, _steals);
4076 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
4077 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
4078 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
4079 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
4080 _aborted_timed_out, _aborted_satb, _aborted_termination);
4081 #endif // _MARKING_STATS_
4082 }
4084 /*****************************************************************************
4086 The do_marking_step(time_target_ms, ...) method is the building
4087 block of the parallel marking framework. It can be called in parallel
4088 with other invocations of do_marking_step() on different tasks
4089 (but only one per task, obviously) and concurrently with the
4090 mutator threads, or during remark, hence it eliminates the need
4091 for two versions of the code. When called during remark, it will
4092 pick up from where the task left off during the concurrent marking
4093 phase. Interestingly, tasks are also claimable during evacuation
4094 pauses too, since do_marking_step() ensures that it aborts before
4095 it needs to yield.
4097 The data structures that it uses to do marking work are the
4098 following:
4100 (1) Marking Bitmap. If there are gray objects that appear only
4101 on the bitmap (this happens either when dealing with an overflow
4102 or when the initial marking phase has simply marked the roots
4103 and didn't push them on the stack), then tasks claim heap
4104 regions whose bitmap they then scan to find gray objects. A
4105 global finger indicates where the end of the last claimed region
4106 is. A local finger indicates how far into the region a task has
4107 scanned. The two fingers are used to determine how to gray an
4108 object (i.e. whether simply marking it is OK, as it will be
4109 visited by a task in the future, or whether it needs to be also
4110 pushed on a stack).
4112 (2) Local Queue. The local queue of the task which is accessed
4113 reasonably efficiently by the task. Other tasks can steal from
4114 it when they run out of work. Throughout the marking phase, a
4115 task attempts to keep its local queue short but not totally
4116 empty, so that entries are available for stealing by other
4117 tasks. Only when there is no more work, a task will totally
4118 drain its local queue.
4120 (3) Global Mark Stack. This handles local queue overflow. During
4121 marking only sets of entries are moved between it and the local
4122 queues, as access to it requires a mutex and more fine-grain
4123 interaction with it which might cause contention. If it
4124 overflows, then the marking phase should restart and iterate
4125 over the bitmap to identify gray objects. Throughout the marking
4126 phase, tasks attempt to keep the global mark stack at a small
4127 length but not totally empty, so that entries are available for
4128 popping by other tasks. Only when there is no more work, tasks
4129 will totally drain the global mark stack.
4131 (4) SATB Buffer Queue. This is where completed SATB buffers are
4132 made available. Buffers are regularly removed from this queue
4133 and scanned for roots, so that the queue doesn't get too
4134 long. During remark, all completed buffers are processed, as
4135 well as the filled in parts of any uncompleted buffers.
4137 The do_marking_step() method tries to abort when the time target
4138 has been reached. There are a few other cases when the
4139 do_marking_step() method also aborts:
4141 (1) When the marking phase has been aborted (after a Full GC).
4143 (2) When a global overflow (on the global stack) has been
4144 triggered. Before the task aborts, it will actually sync up with
4145 the other tasks to ensure that all the marking data structures
4146 (local queues, stacks, fingers etc.) are re-initialized so that
4147 when do_marking_step() completes, the marking phase can
4148 immediately restart.
4150 (3) When enough completed SATB buffers are available. The
4151 do_marking_step() method only tries to drain SATB buffers right
4152 at the beginning. So, if enough buffers are available, the
4153 marking step aborts and the SATB buffers are processed at
4154 the beginning of the next invocation.
4156 (4) To yield. when we have to yield then we abort and yield
4157 right at the end of do_marking_step(). This saves us from a lot
4158 of hassle as, by yielding we might allow a Full GC. If this
4159 happens then objects will be compacted underneath our feet, the
4160 heap might shrink, etc. We save checking for this by just
4161 aborting and doing the yield right at the end.
4163 From the above it follows that the do_marking_step() method should
4164 be called in a loop (or, otherwise, regularly) until it completes.
4166 If a marking step completes without its has_aborted() flag being
4167 true, it means it has completed the current marking phase (and
4168 also all other marking tasks have done so and have all synced up).
4170 A method called regular_clock_call() is invoked "regularly" (in
4171 sub ms intervals) throughout marking. It is this clock method that
4172 checks all the abort conditions which were mentioned above and
4173 decides when the task should abort. A work-based scheme is used to
4174 trigger this clock method: when the number of object words the
4175 marking phase has scanned or the number of references the marking
4176 phase has visited reach a given limit. Additional invocations to
4177 the method clock have been planted in a few other strategic places
4178 too. The initial reason for the clock method was to avoid calling
4179 vtime too regularly, as it is quite expensive. So, once it was in
4180 place, it was natural to piggy-back all the other conditions on it
4181 too and not constantly check them throughout the code.
4183 If do_termination is true then do_marking_step will enter its
4184 termination protocol.
4186 The value of is_serial must be true when do_marking_step is being
4187 called serially (i.e. by the VMThread) and do_marking_step should
4188 skip any synchronization in the termination and overflow code.
4189 Examples include the serial remark code and the serial reference
4190 processing closures.
4192 The value of is_serial must be false when do_marking_step is
4193 being called by any of the worker threads in a work gang.
4194 Examples include the concurrent marking code (CMMarkingTask),
4195 the MT remark code, and the MT reference processing closures.
4197 *****************************************************************************/
4199 void CMTask::do_marking_step(double time_target_ms,
4200 bool do_termination,
4201 bool is_serial) {
4202 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4203 assert(concurrent() == _cm->concurrent(), "they should be the same");
4205 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4206 assert(_task_queues != NULL, "invariant");
4207 assert(_task_queue != NULL, "invariant");
4208 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4210 assert(!_claimed,
4211 "only one thread should claim this task at any one time");
4213 // OK, this doesn't safeguard again all possible scenarios, as it is
4214 // possible for two threads to set the _claimed flag at the same
4215 // time. But it is only for debugging purposes anyway and it will
4216 // catch most problems.
4217 _claimed = true;
4219 _start_time_ms = os::elapsedVTime() * 1000.0;
4220 statsOnly( _interval_start_time_ms = _start_time_ms );
4222 // If do_stealing is true then do_marking_step will attempt to
4223 // steal work from the other CMTasks. It only makes sense to
4224 // enable stealing when the termination protocol is enabled
4225 // and do_marking_step() is not being called serially.
4226 bool do_stealing = do_termination && !is_serial;
4228 double diff_prediction_ms =
4229 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4230 _time_target_ms = time_target_ms - diff_prediction_ms;
4232 // set up the variables that are used in the work-based scheme to
4233 // call the regular clock method
4234 _words_scanned = 0;
4235 _refs_reached = 0;
4236 recalculate_limits();
4238 // clear all flags
4239 clear_has_aborted();
4240 _has_timed_out = false;
4241 _draining_satb_buffers = false;
4243 ++_calls;
4245 if (_cm->verbose_low()) {
4246 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4247 "target = %1.2lfms >>>>>>>>>>",
4248 _worker_id, _calls, _time_target_ms);
4249 }
4251 // Set up the bitmap and oop closures. Anything that uses them is
4252 // eventually called from this method, so it is OK to allocate these
4253 // statically.
4254 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4255 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4256 set_cm_oop_closure(&cm_oop_closure);
4258 if (_cm->has_overflown()) {
4259 // This can happen if the mark stack overflows during a GC pause
4260 // and this task, after a yield point, restarts. We have to abort
4261 // as we need to get into the overflow protocol which happens
4262 // right at the end of this task.
4263 set_has_aborted();
4264 }
4266 // First drain any available SATB buffers. After this, we will not
4267 // look at SATB buffers before the next invocation of this method.
4268 // If enough completed SATB buffers are queued up, the regular clock
4269 // will abort this task so that it restarts.
4270 drain_satb_buffers();
4271 // ...then partially drain the local queue and the global stack
4272 drain_local_queue(true);
4273 drain_global_stack(true);
4275 do {
4276 if (!has_aborted() && _curr_region != NULL) {
4277 // This means that we're already holding on to a region.
4278 assert(_finger != NULL, "if region is not NULL, then the finger "
4279 "should not be NULL either");
4281 // We might have restarted this task after an evacuation pause
4282 // which might have evacuated the region we're holding on to
4283 // underneath our feet. Let's read its limit again to make sure
4284 // that we do not iterate over a region of the heap that
4285 // contains garbage (update_region_limit() will also move
4286 // _finger to the start of the region if it is found empty).
4287 update_region_limit();
4288 // We will start from _finger not from the start of the region,
4289 // as we might be restarting this task after aborting half-way
4290 // through scanning this region. In this case, _finger points to
4291 // the address where we last found a marked object. If this is a
4292 // fresh region, _finger points to start().
4293 MemRegion mr = MemRegion(_finger, _region_limit);
4295 if (_cm->verbose_low()) {
4296 gclog_or_tty->print_cr("[%u] we're scanning part "
4297 "["PTR_FORMAT", "PTR_FORMAT") "
4298 "of region "HR_FORMAT,
4299 _worker_id, p2i(_finger), p2i(_region_limit),
4300 HR_FORMAT_PARAMS(_curr_region));
4301 }
4303 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4304 "humongous regions should go around loop once only");
4306 // Some special cases:
4307 // If the memory region is empty, we can just give up the region.
4308 // If the current region is humongous then we only need to check
4309 // the bitmap for the bit associated with the start of the object,
4310 // scan the object if it's live, and give up the region.
4311 // Otherwise, let's iterate over the bitmap of the part of the region
4312 // that is left.
4313 // If the iteration is successful, give up the region.
4314 if (mr.is_empty()) {
4315 giveup_current_region();
4316 regular_clock_call();
4317 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4318 if (_nextMarkBitMap->isMarked(mr.start())) {
4319 // The object is marked - apply the closure
4320 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4321 bitmap_closure.do_bit(offset);
4322 }
4323 // Even if this task aborted while scanning the humongous object
4324 // we can (and should) give up the current region.
4325 giveup_current_region();
4326 regular_clock_call();
4327 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4328 giveup_current_region();
4329 regular_clock_call();
4330 } else {
4331 assert(has_aborted(), "currently the only way to do so");
4332 // The only way to abort the bitmap iteration is to return
4333 // false from the do_bit() method. However, inside the
4334 // do_bit() method we move the _finger to point to the
4335 // object currently being looked at. So, if we bail out, we
4336 // have definitely set _finger to something non-null.
4337 assert(_finger != NULL, "invariant");
4339 // Region iteration was actually aborted. So now _finger
4340 // points to the address of the object we last scanned. If we
4341 // leave it there, when we restart this task, we will rescan
4342 // the object. It is easy to avoid this. We move the finger by
4343 // enough to point to the next possible object header (the
4344 // bitmap knows by how much we need to move it as it knows its
4345 // granularity).
4346 assert(_finger < _region_limit, "invariant");
4347 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4348 // Check if bitmap iteration was aborted while scanning the last object
4349 if (new_finger >= _region_limit) {
4350 giveup_current_region();
4351 } else {
4352 move_finger_to(new_finger);
4353 }
4354 }
4355 }
4356 // At this point we have either completed iterating over the
4357 // region we were holding on to, or we have aborted.
4359 // We then partially drain the local queue and the global stack.
4360 // (Do we really need this?)
4361 drain_local_queue(true);
4362 drain_global_stack(true);
4364 // Read the note on the claim_region() method on why it might
4365 // return NULL with potentially more regions available for
4366 // claiming and why we have to check out_of_regions() to determine
4367 // whether we're done or not.
4368 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4369 // We are going to try to claim a new region. We should have
4370 // given up on the previous one.
4371 // Separated the asserts so that we know which one fires.
4372 assert(_curr_region == NULL, "invariant");
4373 assert(_finger == NULL, "invariant");
4374 assert(_region_limit == NULL, "invariant");
4375 if (_cm->verbose_low()) {
4376 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4377 }
4378 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4379 if (claimed_region != NULL) {
4380 // Yes, we managed to claim one
4381 statsOnly( ++_regions_claimed );
4383 if (_cm->verbose_low()) {
4384 gclog_or_tty->print_cr("[%u] we successfully claimed "
4385 "region "PTR_FORMAT,
4386 _worker_id, p2i(claimed_region));
4387 }
4389 setup_for_region(claimed_region);
4390 assert(_curr_region == claimed_region, "invariant");
4391 }
4392 // It is important to call the regular clock here. It might take
4393 // a while to claim a region if, for example, we hit a large
4394 // block of empty regions. So we need to call the regular clock
4395 // method once round the loop to make sure it's called
4396 // frequently enough.
4397 regular_clock_call();
4398 }
4400 if (!has_aborted() && _curr_region == NULL) {
4401 assert(_cm->out_of_regions(),
4402 "at this point we should be out of regions");
4403 }
4404 } while ( _curr_region != NULL && !has_aborted());
4406 if (!has_aborted()) {
4407 // We cannot check whether the global stack is empty, since other
4408 // tasks might be pushing objects to it concurrently.
4409 assert(_cm->out_of_regions(),
4410 "at this point we should be out of regions");
4412 if (_cm->verbose_low()) {
4413 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4414 }
4416 // Try to reduce the number of available SATB buffers so that
4417 // remark has less work to do.
4418 drain_satb_buffers();
4419 }
4421 // Since we've done everything else, we can now totally drain the
4422 // local queue and global stack.
4423 drain_local_queue(false);
4424 drain_global_stack(false);
4426 // Attempt at work stealing from other task's queues.
4427 if (do_stealing && !has_aborted()) {
4428 // We have not aborted. This means that we have finished all that
4429 // we could. Let's try to do some stealing...
4431 // We cannot check whether the global stack is empty, since other
4432 // tasks might be pushing objects to it concurrently.
4433 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4434 "only way to reach here");
4436 if (_cm->verbose_low()) {
4437 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4438 }
4440 while (!has_aborted()) {
4441 oop obj;
4442 statsOnly( ++_steal_attempts );
4444 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4445 if (_cm->verbose_medium()) {
4446 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4447 _worker_id, p2i((void*) obj));
4448 }
4450 statsOnly( ++_steals );
4452 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4453 "any stolen object should be marked");
4454 scan_object(obj);
4456 // And since we're towards the end, let's totally drain the
4457 // local queue and global stack.
4458 drain_local_queue(false);
4459 drain_global_stack(false);
4460 } else {
4461 break;
4462 }
4463 }
4464 }
4466 // If we are about to wrap up and go into termination, check if we
4467 // should raise the overflow flag.
4468 if (do_termination && !has_aborted()) {
4469 if (_cm->force_overflow()->should_force()) {
4470 _cm->set_has_overflown();
4471 regular_clock_call();
4472 }
4473 }
4475 // We still haven't aborted. Now, let's try to get into the
4476 // termination protocol.
4477 if (do_termination && !has_aborted()) {
4478 // We cannot check whether the global stack is empty, since other
4479 // tasks might be concurrently pushing objects on it.
4480 // Separated the asserts so that we know which one fires.
4481 assert(_cm->out_of_regions(), "only way to reach here");
4482 assert(_task_queue->size() == 0, "only way to reach here");
4484 if (_cm->verbose_low()) {
4485 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4486 }
4488 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4490 // The CMTask class also extends the TerminatorTerminator class,
4491 // hence its should_exit_termination() method will also decide
4492 // whether to exit the termination protocol or not.
4493 bool finished = (is_serial ||
4494 _cm->terminator()->offer_termination(this));
4495 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4496 _termination_time_ms +=
4497 termination_end_time_ms - _termination_start_time_ms;
4499 if (finished) {
4500 // We're all done.
4502 if (_worker_id == 0) {
4503 // let's allow task 0 to do this
4504 if (concurrent()) {
4505 assert(_cm->concurrent_marking_in_progress(), "invariant");
4506 // we need to set this to false before the next
4507 // safepoint. This way we ensure that the marking phase
4508 // doesn't observe any more heap expansions.
4509 _cm->clear_concurrent_marking_in_progress();
4510 }
4511 }
4513 // We can now guarantee that the global stack is empty, since
4514 // all other tasks have finished. We separated the guarantees so
4515 // that, if a condition is false, we can immediately find out
4516 // which one.
4517 guarantee(_cm->out_of_regions(), "only way to reach here");
4518 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4519 guarantee(_task_queue->size() == 0, "only way to reach here");
4520 guarantee(!_cm->has_overflown(), "only way to reach here");
4521 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4523 if (_cm->verbose_low()) {
4524 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4525 }
4526 } else {
4527 // Apparently there's more work to do. Let's abort this task. It
4528 // will restart it and we can hopefully find more things to do.
4530 if (_cm->verbose_low()) {
4531 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4532 _worker_id);
4533 }
4535 set_has_aborted();
4536 statsOnly( ++_aborted_termination );
4537 }
4538 }
4540 // Mainly for debugging purposes to make sure that a pointer to the
4541 // closure which was statically allocated in this frame doesn't
4542 // escape it by accident.
4543 set_cm_oop_closure(NULL);
4544 double end_time_ms = os::elapsedVTime() * 1000.0;
4545 double elapsed_time_ms = end_time_ms - _start_time_ms;
4546 // Update the step history.
4547 _step_times_ms.add(elapsed_time_ms);
4549 if (has_aborted()) {
4550 // The task was aborted for some reason.
4552 statsOnly( ++_aborted );
4554 if (_has_timed_out) {
4555 double diff_ms = elapsed_time_ms - _time_target_ms;
4556 // Keep statistics of how well we did with respect to hitting
4557 // our target only if we actually timed out (if we aborted for
4558 // other reasons, then the results might get skewed).
4559 _marking_step_diffs_ms.add(diff_ms);
4560 }
4562 if (_cm->has_overflown()) {
4563 // This is the interesting one. We aborted because a global
4564 // overflow was raised. This means we have to restart the
4565 // marking phase and start iterating over regions. However, in
4566 // order to do this we have to make sure that all tasks stop
4567 // what they are doing and re-initialise in a safe manner. We
4568 // will achieve this with the use of two barrier sync points.
4570 if (_cm->verbose_low()) {
4571 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4572 }
4574 if (!is_serial) {
4575 // We only need to enter the sync barrier if being called
4576 // from a parallel context
4577 _cm->enter_first_sync_barrier(_worker_id);
4579 // When we exit this sync barrier we know that all tasks have
4580 // stopped doing marking work. So, it's now safe to
4581 // re-initialise our data structures. At the end of this method,
4582 // task 0 will clear the global data structures.
4583 }
4585 statsOnly( ++_aborted_overflow );
4587 // We clear the local state of this task...
4588 clear_region_fields();
4590 if (!is_serial) {
4591 // ...and enter the second barrier.
4592 _cm->enter_second_sync_barrier(_worker_id);
4593 }
4594 // At this point, if we're during the concurrent phase of
4595 // marking, everything has been re-initialized and we're
4596 // ready to restart.
4597 }
4599 if (_cm->verbose_low()) {
4600 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4601 "elapsed = %1.2lfms <<<<<<<<<<",
4602 _worker_id, _time_target_ms, elapsed_time_ms);
4603 if (_cm->has_aborted()) {
4604 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4605 _worker_id);
4606 }
4607 }
4608 } else {
4609 if (_cm->verbose_low()) {
4610 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4611 "elapsed = %1.2lfms <<<<<<<<<<",
4612 _worker_id, _time_target_ms, elapsed_time_ms);
4613 }
4614 }
4616 _claimed = false;
4617 }
4619 CMTask::CMTask(uint worker_id,
4620 ConcurrentMark* cm,
4621 size_t* marked_bytes,
4622 BitMap* card_bm,
4623 CMTaskQueue* task_queue,
4624 CMTaskQueueSet* task_queues)
4625 : _g1h(G1CollectedHeap::heap()),
4626 _worker_id(worker_id), _cm(cm),
4627 _claimed(false),
4628 _nextMarkBitMap(NULL), _hash_seed(17),
4629 _task_queue(task_queue),
4630 _task_queues(task_queues),
4631 _cm_oop_closure(NULL),
4632 _marked_bytes_array(marked_bytes),
4633 _card_bm(card_bm) {
4634 guarantee(task_queue != NULL, "invariant");
4635 guarantee(task_queues != NULL, "invariant");
4637 statsOnly( _clock_due_to_scanning = 0;
4638 _clock_due_to_marking = 0 );
4640 _marking_step_diffs_ms.add(0.5);
4641 }
4643 // These are formatting macros that are used below to ensure
4644 // consistent formatting. The *_H_* versions are used to format the
4645 // header for a particular value and they should be kept consistent
4646 // with the corresponding macro. Also note that most of the macros add
4647 // the necessary white space (as a prefix) which makes them a bit
4648 // easier to compose.
4650 // All the output lines are prefixed with this string to be able to
4651 // identify them easily in a large log file.
4652 #define G1PPRL_LINE_PREFIX "###"
4654 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4655 #ifdef _LP64
4656 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4657 #else // _LP64
4658 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4659 #endif // _LP64
4661 // For per-region info
4662 #define G1PPRL_TYPE_FORMAT " %-4s"
4663 #define G1PPRL_TYPE_H_FORMAT " %4s"
4664 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4665 #define G1PPRL_BYTE_H_FORMAT " %9s"
4666 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4667 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4669 // For summary info
4670 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4671 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4672 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4673 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4675 G1PrintRegionLivenessInfoClosure::
4676 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4677 : _out(out),
4678 _total_used_bytes(0), _total_capacity_bytes(0),
4679 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4680 _hum_used_bytes(0), _hum_capacity_bytes(0),
4681 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4682 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4683 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4684 MemRegion g1_reserved = g1h->g1_reserved();
4685 double now = os::elapsedTime();
4687 // Print the header of the output.
4688 _out->cr();
4689 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4690 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4691 G1PPRL_SUM_ADDR_FORMAT("reserved")
4692 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4693 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4694 HeapRegion::GrainBytes);
4695 _out->print_cr(G1PPRL_LINE_PREFIX);
4696 _out->print_cr(G1PPRL_LINE_PREFIX
4697 G1PPRL_TYPE_H_FORMAT
4698 G1PPRL_ADDR_BASE_H_FORMAT
4699 G1PPRL_BYTE_H_FORMAT
4700 G1PPRL_BYTE_H_FORMAT
4701 G1PPRL_BYTE_H_FORMAT
4702 G1PPRL_DOUBLE_H_FORMAT
4703 G1PPRL_BYTE_H_FORMAT
4704 G1PPRL_BYTE_H_FORMAT,
4705 "type", "address-range",
4706 "used", "prev-live", "next-live", "gc-eff",
4707 "remset", "code-roots");
4708 _out->print_cr(G1PPRL_LINE_PREFIX
4709 G1PPRL_TYPE_H_FORMAT
4710 G1PPRL_ADDR_BASE_H_FORMAT
4711 G1PPRL_BYTE_H_FORMAT
4712 G1PPRL_BYTE_H_FORMAT
4713 G1PPRL_BYTE_H_FORMAT
4714 G1PPRL_DOUBLE_H_FORMAT
4715 G1PPRL_BYTE_H_FORMAT
4716 G1PPRL_BYTE_H_FORMAT,
4717 "", "",
4718 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4719 "(bytes)", "(bytes)");
4720 }
4722 // It takes as a parameter a reference to one of the _hum_* fields, it
4723 // deduces the corresponding value for a region in a humongous region
4724 // series (either the region size, or what's left if the _hum_* field
4725 // is < the region size), and updates the _hum_* field accordingly.
4726 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4727 size_t bytes = 0;
4728 // The > 0 check is to deal with the prev and next live bytes which
4729 // could be 0.
4730 if (*hum_bytes > 0) {
4731 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4732 *hum_bytes -= bytes;
4733 }
4734 return bytes;
4735 }
4737 // It deduces the values for a region in a humongous region series
4738 // from the _hum_* fields and updates those accordingly. It assumes
4739 // that that _hum_* fields have already been set up from the "starts
4740 // humongous" region and we visit the regions in address order.
4741 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4742 size_t* capacity_bytes,
4743 size_t* prev_live_bytes,
4744 size_t* next_live_bytes) {
4745 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4746 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4747 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4748 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4749 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4750 }
4752 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4753 const char* type = r->get_type_str();
4754 HeapWord* bottom = r->bottom();
4755 HeapWord* end = r->end();
4756 size_t capacity_bytes = r->capacity();
4757 size_t used_bytes = r->used();
4758 size_t prev_live_bytes = r->live_bytes();
4759 size_t next_live_bytes = r->next_live_bytes();
4760 double gc_eff = r->gc_efficiency();
4761 size_t remset_bytes = r->rem_set()->mem_size();
4762 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4764 if (r->startsHumongous()) {
4765 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4766 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4767 "they should have been zeroed after the last time we used them");
4768 // Set up the _hum_* fields.
4769 _hum_capacity_bytes = capacity_bytes;
4770 _hum_used_bytes = used_bytes;
4771 _hum_prev_live_bytes = prev_live_bytes;
4772 _hum_next_live_bytes = next_live_bytes;
4773 get_hum_bytes(&used_bytes, &capacity_bytes,
4774 &prev_live_bytes, &next_live_bytes);
4775 end = bottom + HeapRegion::GrainWords;
4776 } else if (r->continuesHumongous()) {
4777 get_hum_bytes(&used_bytes, &capacity_bytes,
4778 &prev_live_bytes, &next_live_bytes);
4779 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4780 }
4782 _total_used_bytes += used_bytes;
4783 _total_capacity_bytes += capacity_bytes;
4784 _total_prev_live_bytes += prev_live_bytes;
4785 _total_next_live_bytes += next_live_bytes;
4786 _total_remset_bytes += remset_bytes;
4787 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4789 // Print a line for this particular region.
4790 _out->print_cr(G1PPRL_LINE_PREFIX
4791 G1PPRL_TYPE_FORMAT
4792 G1PPRL_ADDR_BASE_FORMAT
4793 G1PPRL_BYTE_FORMAT
4794 G1PPRL_BYTE_FORMAT
4795 G1PPRL_BYTE_FORMAT
4796 G1PPRL_DOUBLE_FORMAT
4797 G1PPRL_BYTE_FORMAT
4798 G1PPRL_BYTE_FORMAT,
4799 type, p2i(bottom), p2i(end),
4800 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4801 remset_bytes, strong_code_roots_bytes);
4803 return false;
4804 }
4806 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4807 // add static memory usages to remembered set sizes
4808 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4809 // Print the footer of the output.
4810 _out->print_cr(G1PPRL_LINE_PREFIX);
4811 _out->print_cr(G1PPRL_LINE_PREFIX
4812 " SUMMARY"
4813 G1PPRL_SUM_MB_FORMAT("capacity")
4814 G1PPRL_SUM_MB_PERC_FORMAT("used")
4815 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4816 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4817 G1PPRL_SUM_MB_FORMAT("remset")
4818 G1PPRL_SUM_MB_FORMAT("code-roots"),
4819 bytes_to_mb(_total_capacity_bytes),
4820 bytes_to_mb(_total_used_bytes),
4821 perc(_total_used_bytes, _total_capacity_bytes),
4822 bytes_to_mb(_total_prev_live_bytes),
4823 perc(_total_prev_live_bytes, _total_capacity_bytes),
4824 bytes_to_mb(_total_next_live_bytes),
4825 perc(_total_next_live_bytes, _total_capacity_bytes),
4826 bytes_to_mb(_total_remset_bytes),
4827 bytes_to_mb(_total_strong_code_roots_bytes));
4828 _out->cr();
4829 }