Wed, 10 Sep 2014 13:01:13 -0700
8057818: collect allocation context statistics at gc pauses
Reviewed-by: mikael, jmasa
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/symbolTable.hpp"
27 #include "code/codeCache.hpp"
28 #include "gc_implementation/g1/concurrentMark.inline.hpp"
29 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
30 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
31 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
32 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
33 #include "gc_implementation/g1/g1Log.hpp"
34 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
35 #include "gc_implementation/g1/g1RemSet.hpp"
36 #include "gc_implementation/g1/heapRegion.inline.hpp"
37 #include "gc_implementation/g1/heapRegionManager.inline.hpp"
38 #include "gc_implementation/g1/heapRegionRemSet.hpp"
39 #include "gc_implementation/g1/heapRegionSet.inline.hpp"
40 #include "gc_implementation/shared/vmGCOperations.hpp"
41 #include "gc_implementation/shared/gcTimer.hpp"
42 #include "gc_implementation/shared/gcTrace.hpp"
43 #include "gc_implementation/shared/gcTraceTime.hpp"
44 #include "memory/allocation.hpp"
45 #include "memory/genOopClosures.inline.hpp"
46 #include "memory/referencePolicy.hpp"
47 #include "memory/resourceArea.hpp"
48 #include "oops/oop.inline.hpp"
49 #include "runtime/handles.inline.hpp"
50 #include "runtime/java.hpp"
51 #include "runtime/prefetch.inline.hpp"
52 #include "services/memTracker.hpp"
54 // Concurrent marking bit map wrapper
56 CMBitMapRO::CMBitMapRO(int shifter) :
57 _bm(),
58 _shifter(shifter) {
59 _bmStartWord = 0;
60 _bmWordSize = 0;
61 }
63 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
64 const HeapWord* limit) const {
65 // First we must round addr *up* to a possible object boundary.
66 addr = (HeapWord*)align_size_up((intptr_t)addr,
67 HeapWordSize << _shifter);
68 size_t addrOffset = heapWordToOffset(addr);
69 if (limit == NULL) {
70 limit = _bmStartWord + _bmWordSize;
71 }
72 size_t limitOffset = heapWordToOffset(limit);
73 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
74 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
75 assert(nextAddr >= addr, "get_next_one postcondition");
76 assert(nextAddr == limit || isMarked(nextAddr),
77 "get_next_one postcondition");
78 return nextAddr;
79 }
81 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
82 const HeapWord* limit) const {
83 size_t addrOffset = heapWordToOffset(addr);
84 if (limit == NULL) {
85 limit = _bmStartWord + _bmWordSize;
86 }
87 size_t limitOffset = heapWordToOffset(limit);
88 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
89 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
90 assert(nextAddr >= addr, "get_next_one postcondition");
91 assert(nextAddr == limit || !isMarked(nextAddr),
92 "get_next_one postcondition");
93 return nextAddr;
94 }
96 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
97 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
98 return (int) (diff >> _shifter);
99 }
101 #ifndef PRODUCT
102 bool CMBitMapRO::covers(MemRegion heap_rs) const {
103 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
104 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
105 "size inconsistency");
106 return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
107 _bmWordSize == heap_rs.word_size();
108 }
109 #endif
111 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
112 _bm.print_on_error(st, prefix);
113 }
115 size_t CMBitMap::compute_size(size_t heap_size) {
116 return heap_size / mark_distance();
117 }
119 size_t CMBitMap::mark_distance() {
120 return MinObjAlignmentInBytes * BitsPerByte;
121 }
123 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
124 _bmStartWord = heap.start();
125 _bmWordSize = heap.word_size();
127 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
128 _bm.set_size(_bmWordSize >> _shifter);
130 storage->set_mapping_changed_listener(&_listener);
131 }
133 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions) {
134 // We need to clear the bitmap on commit, removing any existing information.
135 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
136 _bm->clearRange(mr);
137 }
139 // Closure used for clearing the given mark bitmap.
140 class ClearBitmapHRClosure : public HeapRegionClosure {
141 private:
142 ConcurrentMark* _cm;
143 CMBitMap* _bitmap;
144 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration.
145 public:
146 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
147 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
148 }
150 virtual bool doHeapRegion(HeapRegion* r) {
151 size_t const chunk_size_in_words = M / HeapWordSize;
153 HeapWord* cur = r->bottom();
154 HeapWord* const end = r->end();
156 while (cur < end) {
157 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
158 _bitmap->clearRange(mr);
160 cur += chunk_size_in_words;
162 // Abort iteration if after yielding the marking has been aborted.
163 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
164 return true;
165 }
166 // Repeat the asserts from before the start of the closure. We will do them
167 // as asserts here to minimize their overhead on the product. However, we
168 // will have them as guarantees at the beginning / end of the bitmap
169 // clearing to get some checking in the product.
170 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
171 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
172 }
174 return false;
175 }
176 };
178 void CMBitMap::clearAll() {
179 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
180 G1CollectedHeap::heap()->heap_region_iterate(&cl);
181 guarantee(cl.complete(), "Must have completed iteration.");
182 return;
183 }
185 void CMBitMap::markRange(MemRegion mr) {
186 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
187 assert(!mr.is_empty(), "unexpected empty region");
188 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
189 ((HeapWord *) mr.end())),
190 "markRange memory region end is not card aligned");
191 // convert address range into offset range
192 _bm.at_put_range(heapWordToOffset(mr.start()),
193 heapWordToOffset(mr.end()), true);
194 }
196 void CMBitMap::clearRange(MemRegion mr) {
197 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
198 assert(!mr.is_empty(), "unexpected empty region");
199 // convert address range into offset range
200 _bm.at_put_range(heapWordToOffset(mr.start()),
201 heapWordToOffset(mr.end()), false);
202 }
204 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
205 HeapWord* end_addr) {
206 HeapWord* start = getNextMarkedWordAddress(addr);
207 start = MIN2(start, end_addr);
208 HeapWord* end = getNextUnmarkedWordAddress(start);
209 end = MIN2(end, end_addr);
210 assert(start <= end, "Consistency check");
211 MemRegion mr(start, end);
212 if (!mr.is_empty()) {
213 clearRange(mr);
214 }
215 return mr;
216 }
218 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
219 _base(NULL), _cm(cm)
220 #ifdef ASSERT
221 , _drain_in_progress(false)
222 , _drain_in_progress_yields(false)
223 #endif
224 {}
226 bool CMMarkStack::allocate(size_t capacity) {
227 // allocate a stack of the requisite depth
228 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
229 if (!rs.is_reserved()) {
230 warning("ConcurrentMark MarkStack allocation failure");
231 return false;
232 }
233 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
234 if (!_virtual_space.initialize(rs, rs.size())) {
235 warning("ConcurrentMark MarkStack backing store failure");
236 // Release the virtual memory reserved for the marking stack
237 rs.release();
238 return false;
239 }
240 assert(_virtual_space.committed_size() == rs.size(),
241 "Didn't reserve backing store for all of ConcurrentMark stack?");
242 _base = (oop*) _virtual_space.low();
243 setEmpty();
244 _capacity = (jint) capacity;
245 _saved_index = -1;
246 _should_expand = false;
247 NOT_PRODUCT(_max_depth = 0);
248 return true;
249 }
251 void CMMarkStack::expand() {
252 // Called, during remark, if we've overflown the marking stack during marking.
253 assert(isEmpty(), "stack should been emptied while handling overflow");
254 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
255 // Clear expansion flag
256 _should_expand = false;
257 if (_capacity == (jint) MarkStackSizeMax) {
258 if (PrintGCDetails && Verbose) {
259 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
260 }
261 return;
262 }
263 // Double capacity if possible
264 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
265 // Do not give up existing stack until we have managed to
266 // get the double capacity that we desired.
267 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
268 sizeof(oop)));
269 if (rs.is_reserved()) {
270 // Release the backing store associated with old stack
271 _virtual_space.release();
272 // Reinitialize virtual space for new stack
273 if (!_virtual_space.initialize(rs, rs.size())) {
274 fatal("Not enough swap for expanded marking stack capacity");
275 }
276 _base = (oop*)(_virtual_space.low());
277 _index = 0;
278 _capacity = new_capacity;
279 } else {
280 if (PrintGCDetails && Verbose) {
281 // Failed to double capacity, continue;
282 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
283 SIZE_FORMAT"K to " SIZE_FORMAT"K",
284 _capacity / K, new_capacity / K);
285 }
286 }
287 }
289 void CMMarkStack::set_should_expand() {
290 // If we're resetting the marking state because of an
291 // marking stack overflow, record that we should, if
292 // possible, expand the stack.
293 _should_expand = _cm->has_overflown();
294 }
296 CMMarkStack::~CMMarkStack() {
297 if (_base != NULL) {
298 _base = NULL;
299 _virtual_space.release();
300 }
301 }
303 void CMMarkStack::par_push(oop ptr) {
304 while (true) {
305 if (isFull()) {
306 _overflow = true;
307 return;
308 }
309 // Otherwise...
310 jint index = _index;
311 jint next_index = index+1;
312 jint res = Atomic::cmpxchg(next_index, &_index, index);
313 if (res == index) {
314 _base[index] = ptr;
315 // Note that we don't maintain this atomically. We could, but it
316 // doesn't seem necessary.
317 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
318 return;
319 }
320 // Otherwise, we need to try again.
321 }
322 }
324 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
325 while (true) {
326 if (isFull()) {
327 _overflow = true;
328 return;
329 }
330 // Otherwise...
331 jint index = _index;
332 jint next_index = index + n;
333 if (next_index > _capacity) {
334 _overflow = true;
335 return;
336 }
337 jint res = Atomic::cmpxchg(next_index, &_index, index);
338 if (res == index) {
339 for (int i = 0; i < n; i++) {
340 int ind = index + i;
341 assert(ind < _capacity, "By overflow test above.");
342 _base[ind] = ptr_arr[i];
343 }
344 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
345 return;
346 }
347 // Otherwise, we need to try again.
348 }
349 }
351 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
352 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
353 jint start = _index;
354 jint next_index = start + n;
355 if (next_index > _capacity) {
356 _overflow = true;
357 return;
358 }
359 // Otherwise.
360 _index = next_index;
361 for (int i = 0; i < n; i++) {
362 int ind = start + i;
363 assert(ind < _capacity, "By overflow test above.");
364 _base[ind] = ptr_arr[i];
365 }
366 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
367 }
369 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
370 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
371 jint index = _index;
372 if (index == 0) {
373 *n = 0;
374 return false;
375 } else {
376 int k = MIN2(max, index);
377 jint new_ind = index - k;
378 for (int j = 0; j < k; j++) {
379 ptr_arr[j] = _base[new_ind + j];
380 }
381 _index = new_ind;
382 *n = k;
383 return true;
384 }
385 }
387 template<class OopClosureClass>
388 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
389 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
390 || SafepointSynchronize::is_at_safepoint(),
391 "Drain recursion must be yield-safe.");
392 bool res = true;
393 debug_only(_drain_in_progress = true);
394 debug_only(_drain_in_progress_yields = yield_after);
395 while (!isEmpty()) {
396 oop newOop = pop();
397 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
398 assert(newOop->is_oop(), "Expected an oop");
399 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
400 "only grey objects on this stack");
401 newOop->oop_iterate(cl);
402 if (yield_after && _cm->do_yield_check()) {
403 res = false;
404 break;
405 }
406 }
407 debug_only(_drain_in_progress = false);
408 return res;
409 }
411 void CMMarkStack::note_start_of_gc() {
412 assert(_saved_index == -1,
413 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
414 _saved_index = _index;
415 }
417 void CMMarkStack::note_end_of_gc() {
418 // This is intentionally a guarantee, instead of an assert. If we
419 // accidentally add something to the mark stack during GC, it
420 // will be a correctness issue so it's better if we crash. we'll
421 // only check this once per GC anyway, so it won't be a performance
422 // issue in any way.
423 guarantee(_saved_index == _index,
424 err_msg("saved index: %d index: %d", _saved_index, _index));
425 _saved_index = -1;
426 }
428 void CMMarkStack::oops_do(OopClosure* f) {
429 assert(_saved_index == _index,
430 err_msg("saved index: %d index: %d", _saved_index, _index));
431 for (int i = 0; i < _index; i += 1) {
432 f->do_oop(&_base[i]);
433 }
434 }
436 CMRootRegions::CMRootRegions() :
437 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
438 _should_abort(false), _next_survivor(NULL) { }
440 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
441 _young_list = g1h->young_list();
442 _cm = cm;
443 }
445 void CMRootRegions::prepare_for_scan() {
446 assert(!scan_in_progress(), "pre-condition");
448 // Currently, only survivors can be root regions.
449 assert(_next_survivor == NULL, "pre-condition");
450 _next_survivor = _young_list->first_survivor_region();
451 _scan_in_progress = (_next_survivor != NULL);
452 _should_abort = false;
453 }
455 HeapRegion* CMRootRegions::claim_next() {
456 if (_should_abort) {
457 // If someone has set the should_abort flag, we return NULL to
458 // force the caller to bail out of their loop.
459 return NULL;
460 }
462 // Currently, only survivors can be root regions.
463 HeapRegion* res = _next_survivor;
464 if (res != NULL) {
465 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
466 // Read it again in case it changed while we were waiting for the lock.
467 res = _next_survivor;
468 if (res != NULL) {
469 if (res == _young_list->last_survivor_region()) {
470 // We just claimed the last survivor so store NULL to indicate
471 // that we're done.
472 _next_survivor = NULL;
473 } else {
474 _next_survivor = res->get_next_young_region();
475 }
476 } else {
477 // Someone else claimed the last survivor while we were trying
478 // to take the lock so nothing else to do.
479 }
480 }
481 assert(res == NULL || res->is_survivor(), "post-condition");
483 return res;
484 }
486 void CMRootRegions::scan_finished() {
487 assert(scan_in_progress(), "pre-condition");
489 // Currently, only survivors can be root regions.
490 if (!_should_abort) {
491 assert(_next_survivor == NULL, "we should have claimed all survivors");
492 }
493 _next_survivor = NULL;
495 {
496 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
497 _scan_in_progress = false;
498 RootRegionScan_lock->notify_all();
499 }
500 }
502 bool CMRootRegions::wait_until_scan_finished() {
503 if (!scan_in_progress()) return false;
505 {
506 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
507 while (scan_in_progress()) {
508 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
509 }
510 }
511 return true;
512 }
514 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
515 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
516 #endif // _MSC_VER
518 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
519 return MAX2((n_par_threads + 2) / 4, 1U);
520 }
522 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
523 _g1h(g1h),
524 _markBitMap1(),
525 _markBitMap2(),
526 _parallel_marking_threads(0),
527 _max_parallel_marking_threads(0),
528 _sleep_factor(0.0),
529 _marking_task_overhead(1.0),
530 _cleanup_sleep_factor(0.0),
531 _cleanup_task_overhead(1.0),
532 _cleanup_list("Cleanup List"),
533 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
534 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
535 CardTableModRefBS::card_shift,
536 false /* in_resource_area*/),
538 _prevMarkBitMap(&_markBitMap1),
539 _nextMarkBitMap(&_markBitMap2),
541 _markStack(this),
542 // _finger set in set_non_marking_state
544 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
545 // _active_tasks set in set_non_marking_state
546 // _tasks set inside the constructor
547 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
548 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
550 _has_overflown(false),
551 _concurrent(false),
552 _has_aborted(false),
553 _aborted_gc_id(GCId::undefined()),
554 _restart_for_overflow(false),
555 _concurrent_marking_in_progress(false),
557 // _verbose_level set below
559 _init_times(),
560 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
561 _cleanup_times(),
562 _total_counting_time(0.0),
563 _total_rs_scrub_time(0.0),
565 _parallel_workers(NULL),
567 _count_card_bitmaps(NULL),
568 _count_marked_bytes(NULL),
569 _completed_initialization(false) {
570 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
571 if (verbose_level < no_verbose) {
572 verbose_level = no_verbose;
573 }
574 if (verbose_level > high_verbose) {
575 verbose_level = high_verbose;
576 }
577 _verbose_level = verbose_level;
579 if (verbose_low()) {
580 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
581 "heap end = " INTPTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
582 }
584 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
585 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
587 // Create & start a ConcurrentMark thread.
588 _cmThread = new ConcurrentMarkThread(this);
589 assert(cmThread() != NULL, "CM Thread should have been created");
590 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
591 if (_cmThread->osthread() == NULL) {
592 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
593 }
595 assert(CGC_lock != NULL, "Where's the CGC_lock?");
596 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
597 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
599 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
600 satb_qs.set_buffer_size(G1SATBBufferSize);
602 _root_regions.init(_g1h, this);
604 if (ConcGCThreads > ParallelGCThreads) {
605 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
606 "than ParallelGCThreads (" UINTX_FORMAT ").",
607 ConcGCThreads, ParallelGCThreads);
608 return;
609 }
610 if (ParallelGCThreads == 0) {
611 // if we are not running with any parallel GC threads we will not
612 // spawn any marking threads either
613 _parallel_marking_threads = 0;
614 _max_parallel_marking_threads = 0;
615 _sleep_factor = 0.0;
616 _marking_task_overhead = 1.0;
617 } else {
618 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
619 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
620 // if both are set
621 _sleep_factor = 0.0;
622 _marking_task_overhead = 1.0;
623 } else if (G1MarkingOverheadPercent > 0) {
624 // We will calculate the number of parallel marking threads based
625 // on a target overhead with respect to the soft real-time goal
626 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
627 double overall_cm_overhead =
628 (double) MaxGCPauseMillis * marking_overhead /
629 (double) GCPauseIntervalMillis;
630 double cpu_ratio = 1.0 / (double) os::processor_count();
631 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
632 double marking_task_overhead =
633 overall_cm_overhead / marking_thread_num *
634 (double) os::processor_count();
635 double sleep_factor =
636 (1.0 - marking_task_overhead) / marking_task_overhead;
638 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
639 _sleep_factor = sleep_factor;
640 _marking_task_overhead = marking_task_overhead;
641 } else {
642 // Calculate the number of parallel marking threads by scaling
643 // the number of parallel GC threads.
644 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
645 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
646 _sleep_factor = 0.0;
647 _marking_task_overhead = 1.0;
648 }
650 assert(ConcGCThreads > 0, "Should have been set");
651 _parallel_marking_threads = (uint) ConcGCThreads;
652 _max_parallel_marking_threads = _parallel_marking_threads;
654 if (parallel_marking_threads() > 1) {
655 _cleanup_task_overhead = 1.0;
656 } else {
657 _cleanup_task_overhead = marking_task_overhead();
658 }
659 _cleanup_sleep_factor =
660 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
662 #if 0
663 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
664 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
665 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
666 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
667 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
668 #endif
670 guarantee(parallel_marking_threads() > 0, "peace of mind");
671 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
672 _max_parallel_marking_threads, false, true);
673 if (_parallel_workers == NULL) {
674 vm_exit_during_initialization("Failed necessary allocation.");
675 } else {
676 _parallel_workers->initialize_workers();
677 }
678 }
680 if (FLAG_IS_DEFAULT(MarkStackSize)) {
681 uintx mark_stack_size =
682 MIN2(MarkStackSizeMax,
683 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
684 // Verify that the calculated value for MarkStackSize is in range.
685 // It would be nice to use the private utility routine from Arguments.
686 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
687 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
688 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
689 mark_stack_size, (uintx) 1, MarkStackSizeMax);
690 return;
691 }
692 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
693 } else {
694 // Verify MarkStackSize is in range.
695 if (FLAG_IS_CMDLINE(MarkStackSize)) {
696 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
697 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
698 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
699 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
700 MarkStackSize, (uintx) 1, MarkStackSizeMax);
701 return;
702 }
703 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
704 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
705 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
706 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
707 MarkStackSize, MarkStackSizeMax);
708 return;
709 }
710 }
711 }
712 }
714 if (!_markStack.allocate(MarkStackSize)) {
715 warning("Failed to allocate CM marking stack");
716 return;
717 }
719 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
720 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
722 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
723 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
725 BitMap::idx_t card_bm_size = _card_bm.size();
727 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
728 _active_tasks = _max_worker_id;
730 size_t max_regions = (size_t) _g1h->max_regions();
731 for (uint i = 0; i < _max_worker_id; ++i) {
732 CMTaskQueue* task_queue = new CMTaskQueue();
733 task_queue->initialize();
734 _task_queues->register_queue(i, task_queue);
736 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
737 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
739 _tasks[i] = new CMTask(i, this,
740 _count_marked_bytes[i],
741 &_count_card_bitmaps[i],
742 task_queue, _task_queues);
744 _accum_task_vtime[i] = 0.0;
745 }
747 // Calculate the card number for the bottom of the heap. Used
748 // in biasing indexes into the accounting card bitmaps.
749 _heap_bottom_card_num =
750 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
751 CardTableModRefBS::card_shift);
753 // Clear all the liveness counting data
754 clear_all_count_data();
756 // so that the call below can read a sensible value
757 _heap_start = g1h->reserved_region().start();
758 set_non_marking_state();
759 _completed_initialization = true;
760 }
762 void ConcurrentMark::reset() {
763 // Starting values for these two. This should be called in a STW
764 // phase.
765 MemRegion reserved = _g1h->g1_reserved();
766 _heap_start = reserved.start();
767 _heap_end = reserved.end();
769 // Separated the asserts so that we know which one fires.
770 assert(_heap_start != NULL, "heap bounds should look ok");
771 assert(_heap_end != NULL, "heap bounds should look ok");
772 assert(_heap_start < _heap_end, "heap bounds should look ok");
774 // Reset all the marking data structures and any necessary flags
775 reset_marking_state();
777 if (verbose_low()) {
778 gclog_or_tty->print_cr("[global] resetting");
779 }
781 // We do reset all of them, since different phases will use
782 // different number of active threads. So, it's easiest to have all
783 // of them ready.
784 for (uint i = 0; i < _max_worker_id; ++i) {
785 _tasks[i]->reset(_nextMarkBitMap);
786 }
788 // we need this to make sure that the flag is on during the evac
789 // pause with initial mark piggy-backed
790 set_concurrent_marking_in_progress();
791 }
794 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
795 _markStack.set_should_expand();
796 _markStack.setEmpty(); // Also clears the _markStack overflow flag
797 if (clear_overflow) {
798 clear_has_overflown();
799 } else {
800 assert(has_overflown(), "pre-condition");
801 }
802 _finger = _heap_start;
804 for (uint i = 0; i < _max_worker_id; ++i) {
805 CMTaskQueue* queue = _task_queues->queue(i);
806 queue->set_empty();
807 }
808 }
810 void ConcurrentMark::set_concurrency(uint active_tasks) {
811 assert(active_tasks <= _max_worker_id, "we should not have more");
813 _active_tasks = active_tasks;
814 // Need to update the three data structures below according to the
815 // number of active threads for this phase.
816 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
817 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
818 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
819 }
821 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
822 set_concurrency(active_tasks);
824 _concurrent = concurrent;
825 // We propagate this to all tasks, not just the active ones.
826 for (uint i = 0; i < _max_worker_id; ++i)
827 _tasks[i]->set_concurrent(concurrent);
829 if (concurrent) {
830 set_concurrent_marking_in_progress();
831 } else {
832 // We currently assume that the concurrent flag has been set to
833 // false before we start remark. At this point we should also be
834 // in a STW phase.
835 assert(!concurrent_marking_in_progress(), "invariant");
836 assert(out_of_regions(),
837 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
838 p2i(_finger), p2i(_heap_end)));
839 }
840 }
842 void ConcurrentMark::set_non_marking_state() {
843 // We set the global marking state to some default values when we're
844 // not doing marking.
845 reset_marking_state();
846 _active_tasks = 0;
847 clear_concurrent_marking_in_progress();
848 }
850 ConcurrentMark::~ConcurrentMark() {
851 // The ConcurrentMark instance is never freed.
852 ShouldNotReachHere();
853 }
855 void ConcurrentMark::clearNextBitmap() {
856 G1CollectedHeap* g1h = G1CollectedHeap::heap();
858 // Make sure that the concurrent mark thread looks to still be in
859 // the current cycle.
860 guarantee(cmThread()->during_cycle(), "invariant");
862 // We are finishing up the current cycle by clearing the next
863 // marking bitmap and getting it ready for the next cycle. During
864 // this time no other cycle can start. So, let's make sure that this
865 // is the case.
866 guarantee(!g1h->mark_in_progress(), "invariant");
868 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
869 g1h->heap_region_iterate(&cl);
871 // Clear the liveness counting data. If the marking has been aborted, the abort()
872 // call already did that.
873 if (cl.complete()) {
874 clear_all_count_data();
875 }
877 // Repeat the asserts from above.
878 guarantee(cmThread()->during_cycle(), "invariant");
879 guarantee(!g1h->mark_in_progress(), "invariant");
880 }
882 class CheckBitmapClearHRClosure : public HeapRegionClosure {
883 CMBitMap* _bitmap;
884 bool _error;
885 public:
886 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
887 }
889 virtual bool doHeapRegion(HeapRegion* r) {
890 // This closure can be called concurrently to the mutator, so we must make sure
891 // that the result of the getNextMarkedWordAddress() call is compared to the
892 // value passed to it as limit to detect any found bits.
893 // We can use the region's orig_end() for the limit and the comparison value
894 // as it always contains the "real" end of the region that never changes and
895 // has no side effects.
896 // Due to the latter, there can also be no problem with the compiler generating
897 // reloads of the orig_end() call.
898 HeapWord* end = r->orig_end();
899 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
900 }
901 };
903 bool ConcurrentMark::nextMarkBitmapIsClear() {
904 CheckBitmapClearHRClosure cl(_nextMarkBitMap);
905 _g1h->heap_region_iterate(&cl);
906 return cl.complete();
907 }
909 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
910 public:
911 bool doHeapRegion(HeapRegion* r) {
912 if (!r->continuesHumongous()) {
913 r->note_start_of_marking();
914 }
915 return false;
916 }
917 };
919 void ConcurrentMark::checkpointRootsInitialPre() {
920 G1CollectedHeap* g1h = G1CollectedHeap::heap();
921 G1CollectorPolicy* g1p = g1h->g1_policy();
923 _has_aborted = false;
925 #ifndef PRODUCT
926 if (G1PrintReachableAtInitialMark) {
927 print_reachable("at-cycle-start",
928 VerifyOption_G1UsePrevMarking, true /* all */);
929 }
930 #endif
932 // Initialise marking structures. This has to be done in a STW phase.
933 reset();
935 // For each region note start of marking.
936 NoteStartOfMarkHRClosure startcl;
937 g1h->heap_region_iterate(&startcl);
938 }
941 void ConcurrentMark::checkpointRootsInitialPost() {
942 G1CollectedHeap* g1h = G1CollectedHeap::heap();
944 // If we force an overflow during remark, the remark operation will
945 // actually abort and we'll restart concurrent marking. If we always
946 // force an oveflow during remark we'll never actually complete the
947 // marking phase. So, we initilize this here, at the start of the
948 // cycle, so that at the remaining overflow number will decrease at
949 // every remark and we'll eventually not need to cause one.
950 force_overflow_stw()->init();
952 // Start Concurrent Marking weak-reference discovery.
953 ReferenceProcessor* rp = g1h->ref_processor_cm();
954 // enable ("weak") refs discovery
955 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
956 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
958 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
959 // This is the start of the marking cycle, we're expected all
960 // threads to have SATB queues with active set to false.
961 satb_mq_set.set_active_all_threads(true, /* new active value */
962 false /* expected_active */);
964 _root_regions.prepare_for_scan();
966 // update_g1_committed() will be called at the end of an evac pause
967 // when marking is on. So, it's also called at the end of the
968 // initial-mark pause to update the heap end, if the heap expands
969 // during it. No need to call it here.
970 }
972 /*
973 * Notice that in the next two methods, we actually leave the STS
974 * during the barrier sync and join it immediately afterwards. If we
975 * do not do this, the following deadlock can occur: one thread could
976 * be in the barrier sync code, waiting for the other thread to also
977 * sync up, whereas another one could be trying to yield, while also
978 * waiting for the other threads to sync up too.
979 *
980 * Note, however, that this code is also used during remark and in
981 * this case we should not attempt to leave / enter the STS, otherwise
982 * we'll either hit an asseert (debug / fastdebug) or deadlock
983 * (product). So we should only leave / enter the STS if we are
984 * operating concurrently.
985 *
986 * Because the thread that does the sync barrier has left the STS, it
987 * is possible to be suspended for a Full GC or an evacuation pause
988 * could occur. This is actually safe, since the entering the sync
989 * barrier is one of the last things do_marking_step() does, and it
990 * doesn't manipulate any data structures afterwards.
991 */
993 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
994 if (verbose_low()) {
995 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
996 }
998 if (concurrent()) {
999 SuspendibleThreadSet::leave();
1000 }
1002 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
1004 if (concurrent()) {
1005 SuspendibleThreadSet::join();
1006 }
1007 // at this point everyone should have synced up and not be doing any
1008 // more work
1010 if (verbose_low()) {
1011 if (barrier_aborted) {
1012 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1013 } else {
1014 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1015 }
1016 }
1018 if (barrier_aborted) {
1019 // If the barrier aborted we ignore the overflow condition and
1020 // just abort the whole marking phase as quickly as possible.
1021 return;
1022 }
1024 // If we're executing the concurrent phase of marking, reset the marking
1025 // state; otherwise the marking state is reset after reference processing,
1026 // during the remark pause.
1027 // If we reset here as a result of an overflow during the remark we will
1028 // see assertion failures from any subsequent set_concurrency_and_phase()
1029 // calls.
1030 if (concurrent()) {
1031 // let the task associated with with worker 0 do this
1032 if (worker_id == 0) {
1033 // task 0 is responsible for clearing the global data structures
1034 // We should be here because of an overflow. During STW we should
1035 // not clear the overflow flag since we rely on it being true when
1036 // we exit this method to abort the pause and restart concurent
1037 // marking.
1038 reset_marking_state(true /* clear_overflow */);
1039 force_overflow()->update();
1041 if (G1Log::fine()) {
1042 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1043 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1044 }
1045 }
1046 }
1048 // after this, each task should reset its own data structures then
1049 // then go into the second barrier
1050 }
1052 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1053 if (verbose_low()) {
1054 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1055 }
1057 if (concurrent()) {
1058 SuspendibleThreadSet::leave();
1059 }
1061 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1063 if (concurrent()) {
1064 SuspendibleThreadSet::join();
1065 }
1066 // at this point everything should be re-initialized and ready to go
1068 if (verbose_low()) {
1069 if (barrier_aborted) {
1070 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1071 } else {
1072 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1073 }
1074 }
1075 }
1077 #ifndef PRODUCT
1078 void ForceOverflowSettings::init() {
1079 _num_remaining = G1ConcMarkForceOverflow;
1080 _force = false;
1081 update();
1082 }
1084 void ForceOverflowSettings::update() {
1085 if (_num_remaining > 0) {
1086 _num_remaining -= 1;
1087 _force = true;
1088 } else {
1089 _force = false;
1090 }
1091 }
1093 bool ForceOverflowSettings::should_force() {
1094 if (_force) {
1095 _force = false;
1096 return true;
1097 } else {
1098 return false;
1099 }
1100 }
1101 #endif // !PRODUCT
1103 class CMConcurrentMarkingTask: public AbstractGangTask {
1104 private:
1105 ConcurrentMark* _cm;
1106 ConcurrentMarkThread* _cmt;
1108 public:
1109 void work(uint worker_id) {
1110 assert(Thread::current()->is_ConcurrentGC_thread(),
1111 "this should only be done by a conc GC thread");
1112 ResourceMark rm;
1114 double start_vtime = os::elapsedVTime();
1116 SuspendibleThreadSet::join();
1118 assert(worker_id < _cm->active_tasks(), "invariant");
1119 CMTask* the_task = _cm->task(worker_id);
1120 the_task->record_start_time();
1121 if (!_cm->has_aborted()) {
1122 do {
1123 double start_vtime_sec = os::elapsedVTime();
1124 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1126 the_task->do_marking_step(mark_step_duration_ms,
1127 true /* do_termination */,
1128 false /* is_serial*/);
1130 double end_vtime_sec = os::elapsedVTime();
1131 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1132 _cm->clear_has_overflown();
1134 _cm->do_yield_check(worker_id);
1136 jlong sleep_time_ms;
1137 if (!_cm->has_aborted() && the_task->has_aborted()) {
1138 sleep_time_ms =
1139 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1140 SuspendibleThreadSet::leave();
1141 os::sleep(Thread::current(), sleep_time_ms, false);
1142 SuspendibleThreadSet::join();
1143 }
1144 } while (!_cm->has_aborted() && the_task->has_aborted());
1145 }
1146 the_task->record_end_time();
1147 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1149 SuspendibleThreadSet::leave();
1151 double end_vtime = os::elapsedVTime();
1152 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1153 }
1155 CMConcurrentMarkingTask(ConcurrentMark* cm,
1156 ConcurrentMarkThread* cmt) :
1157 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1159 ~CMConcurrentMarkingTask() { }
1160 };
1162 // Calculates the number of active workers for a concurrent
1163 // phase.
1164 uint ConcurrentMark::calc_parallel_marking_threads() {
1165 if (G1CollectedHeap::use_parallel_gc_threads()) {
1166 uint n_conc_workers = 0;
1167 if (!UseDynamicNumberOfGCThreads ||
1168 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1169 !ForceDynamicNumberOfGCThreads)) {
1170 n_conc_workers = max_parallel_marking_threads();
1171 } else {
1172 n_conc_workers =
1173 AdaptiveSizePolicy::calc_default_active_workers(
1174 max_parallel_marking_threads(),
1175 1, /* Minimum workers */
1176 parallel_marking_threads(),
1177 Threads::number_of_non_daemon_threads());
1178 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1179 // that scaling has already gone into "_max_parallel_marking_threads".
1180 }
1181 assert(n_conc_workers > 0, "Always need at least 1");
1182 return n_conc_workers;
1183 }
1184 // If we are not running with any parallel GC threads we will not
1185 // have spawned any marking threads either. Hence the number of
1186 // concurrent workers should be 0.
1187 return 0;
1188 }
1190 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1191 // Currently, only survivors can be root regions.
1192 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1193 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1195 const uintx interval = PrefetchScanIntervalInBytes;
1196 HeapWord* curr = hr->bottom();
1197 const HeapWord* end = hr->top();
1198 while (curr < end) {
1199 Prefetch::read(curr, interval);
1200 oop obj = oop(curr);
1201 int size = obj->oop_iterate(&cl);
1202 assert(size == obj->size(), "sanity");
1203 curr += size;
1204 }
1205 }
1207 class CMRootRegionScanTask : public AbstractGangTask {
1208 private:
1209 ConcurrentMark* _cm;
1211 public:
1212 CMRootRegionScanTask(ConcurrentMark* cm) :
1213 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1215 void work(uint worker_id) {
1216 assert(Thread::current()->is_ConcurrentGC_thread(),
1217 "this should only be done by a conc GC thread");
1219 CMRootRegions* root_regions = _cm->root_regions();
1220 HeapRegion* hr = root_regions->claim_next();
1221 while (hr != NULL) {
1222 _cm->scanRootRegion(hr, worker_id);
1223 hr = root_regions->claim_next();
1224 }
1225 }
1226 };
1228 void ConcurrentMark::scanRootRegions() {
1229 // Start of concurrent marking.
1230 ClassLoaderDataGraph::clear_claimed_marks();
1232 // scan_in_progress() will have been set to true only if there was
1233 // at least one root region to scan. So, if it's false, we
1234 // should not attempt to do any further work.
1235 if (root_regions()->scan_in_progress()) {
1236 _parallel_marking_threads = calc_parallel_marking_threads();
1237 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1238 "Maximum number of marking threads exceeded");
1239 uint active_workers = MAX2(1U, parallel_marking_threads());
1241 CMRootRegionScanTask task(this);
1242 if (use_parallel_marking_threads()) {
1243 _parallel_workers->set_active_workers((int) active_workers);
1244 _parallel_workers->run_task(&task);
1245 } else {
1246 task.work(0);
1247 }
1249 // It's possible that has_aborted() is true here without actually
1250 // aborting the survivor scan earlier. This is OK as it's
1251 // mainly used for sanity checking.
1252 root_regions()->scan_finished();
1253 }
1254 }
1256 void ConcurrentMark::markFromRoots() {
1257 // we might be tempted to assert that:
1258 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1259 // "inconsistent argument?");
1260 // However that wouldn't be right, because it's possible that
1261 // a safepoint is indeed in progress as a younger generation
1262 // stop-the-world GC happens even as we mark in this generation.
1264 _restart_for_overflow = false;
1265 force_overflow_conc()->init();
1267 // _g1h has _n_par_threads
1268 _parallel_marking_threads = calc_parallel_marking_threads();
1269 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1270 "Maximum number of marking threads exceeded");
1272 uint active_workers = MAX2(1U, parallel_marking_threads());
1274 // Parallel task terminator is set in "set_concurrency_and_phase()"
1275 set_concurrency_and_phase(active_workers, true /* concurrent */);
1277 CMConcurrentMarkingTask markingTask(this, cmThread());
1278 if (use_parallel_marking_threads()) {
1279 _parallel_workers->set_active_workers((int)active_workers);
1280 // Don't set _n_par_threads because it affects MT in process_roots()
1281 // and the decisions on that MT processing is made elsewhere.
1282 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1283 _parallel_workers->run_task(&markingTask);
1284 } else {
1285 markingTask.work(0);
1286 }
1287 print_stats();
1288 }
1290 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1291 // world is stopped at this checkpoint
1292 assert(SafepointSynchronize::is_at_safepoint(),
1293 "world should be stopped");
1295 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1297 // If a full collection has happened, we shouldn't do this.
1298 if (has_aborted()) {
1299 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1300 return;
1301 }
1303 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1305 if (VerifyDuringGC) {
1306 HandleMark hm; // handle scope
1307 Universe::heap()->prepare_for_verify();
1308 Universe::verify(VerifyOption_G1UsePrevMarking,
1309 " VerifyDuringGC:(before)");
1310 }
1311 g1h->check_bitmaps("Remark Start");
1313 G1CollectorPolicy* g1p = g1h->g1_policy();
1314 g1p->record_concurrent_mark_remark_start();
1316 double start = os::elapsedTime();
1318 checkpointRootsFinalWork();
1320 double mark_work_end = os::elapsedTime();
1322 weakRefsWork(clear_all_soft_refs);
1324 if (has_overflown()) {
1325 // Oops. We overflowed. Restart concurrent marking.
1326 _restart_for_overflow = true;
1327 if (G1TraceMarkStackOverflow) {
1328 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1329 }
1331 // Verify the heap w.r.t. the previous marking bitmap.
1332 if (VerifyDuringGC) {
1333 HandleMark hm; // handle scope
1334 Universe::heap()->prepare_for_verify();
1335 Universe::verify(VerifyOption_G1UsePrevMarking,
1336 " VerifyDuringGC:(overflow)");
1337 }
1339 // Clear the marking state because we will be restarting
1340 // marking due to overflowing the global mark stack.
1341 reset_marking_state();
1342 } else {
1343 // Aggregate the per-task counting data that we have accumulated
1344 // while marking.
1345 aggregate_count_data();
1347 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1348 // We're done with marking.
1349 // This is the end of the marking cycle, we're expected all
1350 // threads to have SATB queues with active set to true.
1351 satb_mq_set.set_active_all_threads(false, /* new active value */
1352 true /* expected_active */);
1354 if (VerifyDuringGC) {
1355 HandleMark hm; // handle scope
1356 Universe::heap()->prepare_for_verify();
1357 Universe::verify(VerifyOption_G1UseNextMarking,
1358 " VerifyDuringGC:(after)");
1359 }
1360 g1h->check_bitmaps("Remark End");
1361 assert(!restart_for_overflow(), "sanity");
1362 // Completely reset the marking state since marking completed
1363 set_non_marking_state();
1364 }
1366 // Expand the marking stack, if we have to and if we can.
1367 if (_markStack.should_expand()) {
1368 _markStack.expand();
1369 }
1371 // Statistics
1372 double now = os::elapsedTime();
1373 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1374 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1375 _remark_times.add((now - start) * 1000.0);
1377 g1p->record_concurrent_mark_remark_end();
1379 G1CMIsAliveClosure is_alive(g1h);
1380 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1381 }
1383 // Base class of the closures that finalize and verify the
1384 // liveness counting data.
1385 class CMCountDataClosureBase: public HeapRegionClosure {
1386 protected:
1387 G1CollectedHeap* _g1h;
1388 ConcurrentMark* _cm;
1389 CardTableModRefBS* _ct_bs;
1391 BitMap* _region_bm;
1392 BitMap* _card_bm;
1394 // Takes a region that's not empty (i.e., it has at least one
1395 // live object in it and sets its corresponding bit on the region
1396 // bitmap to 1. If the region is "starts humongous" it will also set
1397 // to 1 the bits on the region bitmap that correspond to its
1398 // associated "continues humongous" regions.
1399 void set_bit_for_region(HeapRegion* hr) {
1400 assert(!hr->continuesHumongous(), "should have filtered those out");
1402 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1403 if (!hr->startsHumongous()) {
1404 // Normal (non-humongous) case: just set the bit.
1405 _region_bm->par_at_put(index, true);
1406 } else {
1407 // Starts humongous case: calculate how many regions are part of
1408 // this humongous region and then set the bit range.
1409 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1410 _region_bm->par_at_put_range(index, end_index, true);
1411 }
1412 }
1414 public:
1415 CMCountDataClosureBase(G1CollectedHeap* g1h,
1416 BitMap* region_bm, BitMap* card_bm):
1417 _g1h(g1h), _cm(g1h->concurrent_mark()),
1418 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1419 _region_bm(region_bm), _card_bm(card_bm) { }
1420 };
1422 // Closure that calculates the # live objects per region. Used
1423 // for verification purposes during the cleanup pause.
1424 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1425 CMBitMapRO* _bm;
1426 size_t _region_marked_bytes;
1428 public:
1429 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1430 BitMap* region_bm, BitMap* card_bm) :
1431 CMCountDataClosureBase(g1h, region_bm, card_bm),
1432 _bm(bm), _region_marked_bytes(0) { }
1434 bool doHeapRegion(HeapRegion* hr) {
1436 if (hr->continuesHumongous()) {
1437 // We will ignore these here and process them when their
1438 // associated "starts humongous" region is processed (see
1439 // set_bit_for_heap_region()). Note that we cannot rely on their
1440 // associated "starts humongous" region to have their bit set to
1441 // 1 since, due to the region chunking in the parallel region
1442 // iteration, a "continues humongous" region might be visited
1443 // before its associated "starts humongous".
1444 return false;
1445 }
1447 HeapWord* ntams = hr->next_top_at_mark_start();
1448 HeapWord* start = hr->bottom();
1450 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1451 err_msg("Preconditions not met - "
1452 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1453 p2i(start), p2i(ntams), p2i(hr->end())));
1455 // Find the first marked object at or after "start".
1456 start = _bm->getNextMarkedWordAddress(start, ntams);
1458 size_t marked_bytes = 0;
1460 while (start < ntams) {
1461 oop obj = oop(start);
1462 int obj_sz = obj->size();
1463 HeapWord* obj_end = start + obj_sz;
1465 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1466 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1468 // Note: if we're looking at the last region in heap - obj_end
1469 // could be actually just beyond the end of the heap; end_idx
1470 // will then correspond to a (non-existent) card that is also
1471 // just beyond the heap.
1472 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1473 // end of object is not card aligned - increment to cover
1474 // all the cards spanned by the object
1475 end_idx += 1;
1476 }
1478 // Set the bits in the card BM for the cards spanned by this object.
1479 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1481 // Add the size of this object to the number of marked bytes.
1482 marked_bytes += (size_t)obj_sz * HeapWordSize;
1484 // Find the next marked object after this one.
1485 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1486 }
1488 // Mark the allocated-since-marking portion...
1489 HeapWord* top = hr->top();
1490 if (ntams < top) {
1491 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1492 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1494 // Note: if we're looking at the last region in heap - top
1495 // could be actually just beyond the end of the heap; end_idx
1496 // will then correspond to a (non-existent) card that is also
1497 // just beyond the heap.
1498 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1499 // end of object is not card aligned - increment to cover
1500 // all the cards spanned by the object
1501 end_idx += 1;
1502 }
1503 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1505 // This definitely means the region has live objects.
1506 set_bit_for_region(hr);
1507 }
1509 // Update the live region bitmap.
1510 if (marked_bytes > 0) {
1511 set_bit_for_region(hr);
1512 }
1514 // Set the marked bytes for the current region so that
1515 // it can be queried by a calling verificiation routine
1516 _region_marked_bytes = marked_bytes;
1518 return false;
1519 }
1521 size_t region_marked_bytes() const { return _region_marked_bytes; }
1522 };
1524 // Heap region closure used for verifying the counting data
1525 // that was accumulated concurrently and aggregated during
1526 // the remark pause. This closure is applied to the heap
1527 // regions during the STW cleanup pause.
1529 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1530 G1CollectedHeap* _g1h;
1531 ConcurrentMark* _cm;
1532 CalcLiveObjectsClosure _calc_cl;
1533 BitMap* _region_bm; // Region BM to be verified
1534 BitMap* _card_bm; // Card BM to be verified
1535 bool _verbose; // verbose output?
1537 BitMap* _exp_region_bm; // Expected Region BM values
1538 BitMap* _exp_card_bm; // Expected card BM values
1540 int _failures;
1542 public:
1543 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1544 BitMap* region_bm,
1545 BitMap* card_bm,
1546 BitMap* exp_region_bm,
1547 BitMap* exp_card_bm,
1548 bool verbose) :
1549 _g1h(g1h), _cm(g1h->concurrent_mark()),
1550 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1551 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1552 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1553 _failures(0) { }
1555 int failures() const { return _failures; }
1557 bool doHeapRegion(HeapRegion* hr) {
1558 if (hr->continuesHumongous()) {
1559 // We will ignore these here and process them when their
1560 // associated "starts humongous" region is processed (see
1561 // set_bit_for_heap_region()). Note that we cannot rely on their
1562 // associated "starts humongous" region to have their bit set to
1563 // 1 since, due to the region chunking in the parallel region
1564 // iteration, a "continues humongous" region might be visited
1565 // before its associated "starts humongous".
1566 return false;
1567 }
1569 int failures = 0;
1571 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1572 // this region and set the corresponding bits in the expected region
1573 // and card bitmaps.
1574 bool res = _calc_cl.doHeapRegion(hr);
1575 assert(res == false, "should be continuing");
1577 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1578 Mutex::_no_safepoint_check_flag);
1580 // Verify the marked bytes for this region.
1581 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1582 size_t act_marked_bytes = hr->next_marked_bytes();
1584 // We're not OK if expected marked bytes > actual marked bytes. It means
1585 // we have missed accounting some objects during the actual marking.
1586 if (exp_marked_bytes > act_marked_bytes) {
1587 if (_verbose) {
1588 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1589 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1590 hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1591 }
1592 failures += 1;
1593 }
1595 // Verify the bit, for this region, in the actual and expected
1596 // (which was just calculated) region bit maps.
1597 // We're not OK if the bit in the calculated expected region
1598 // bitmap is set and the bit in the actual region bitmap is not.
1599 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1601 bool expected = _exp_region_bm->at(index);
1602 bool actual = _region_bm->at(index);
1603 if (expected && !actual) {
1604 if (_verbose) {
1605 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1606 "expected: %s, actual: %s",
1607 hr->hrm_index(),
1608 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1609 }
1610 failures += 1;
1611 }
1613 // Verify that the card bit maps for the cards spanned by the current
1614 // region match. We have an error if we have a set bit in the expected
1615 // bit map and the corresponding bit in the actual bitmap is not set.
1617 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1618 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1620 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1621 expected = _exp_card_bm->at(i);
1622 actual = _card_bm->at(i);
1624 if (expected && !actual) {
1625 if (_verbose) {
1626 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1627 "expected: %s, actual: %s",
1628 hr->hrm_index(), i,
1629 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1630 }
1631 failures += 1;
1632 }
1633 }
1635 if (failures > 0 && _verbose) {
1636 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1637 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1638 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1639 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1640 }
1642 _failures += failures;
1644 // We could stop iteration over the heap when we
1645 // find the first violating region by returning true.
1646 return false;
1647 }
1648 };
1650 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1651 protected:
1652 G1CollectedHeap* _g1h;
1653 ConcurrentMark* _cm;
1654 BitMap* _actual_region_bm;
1655 BitMap* _actual_card_bm;
1657 uint _n_workers;
1659 BitMap* _expected_region_bm;
1660 BitMap* _expected_card_bm;
1662 int _failures;
1663 bool _verbose;
1665 public:
1666 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1667 BitMap* region_bm, BitMap* card_bm,
1668 BitMap* expected_region_bm, BitMap* expected_card_bm)
1669 : AbstractGangTask("G1 verify final counting"),
1670 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1671 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1672 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1673 _failures(0), _verbose(false),
1674 _n_workers(0) {
1675 assert(VerifyDuringGC, "don't call this otherwise");
1677 // Use the value already set as the number of active threads
1678 // in the call to run_task().
1679 if (G1CollectedHeap::use_parallel_gc_threads()) {
1680 assert( _g1h->workers()->active_workers() > 0,
1681 "Should have been previously set");
1682 _n_workers = _g1h->workers()->active_workers();
1683 } else {
1684 _n_workers = 1;
1685 }
1687 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1688 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1690 _verbose = _cm->verbose_medium();
1691 }
1693 void work(uint worker_id) {
1694 assert(worker_id < _n_workers, "invariant");
1696 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1697 _actual_region_bm, _actual_card_bm,
1698 _expected_region_bm,
1699 _expected_card_bm,
1700 _verbose);
1702 if (G1CollectedHeap::use_parallel_gc_threads()) {
1703 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1704 worker_id,
1705 _n_workers,
1706 HeapRegion::VerifyCountClaimValue);
1707 } else {
1708 _g1h->heap_region_iterate(&verify_cl);
1709 }
1711 Atomic::add(verify_cl.failures(), &_failures);
1712 }
1714 int failures() const { return _failures; }
1715 };
1717 // Closure that finalizes the liveness counting data.
1718 // Used during the cleanup pause.
1719 // Sets the bits corresponding to the interval [NTAMS, top]
1720 // (which contains the implicitly live objects) in the
1721 // card liveness bitmap. Also sets the bit for each region,
1722 // containing live data, in the region liveness bitmap.
1724 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1725 public:
1726 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1727 BitMap* region_bm,
1728 BitMap* card_bm) :
1729 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1731 bool doHeapRegion(HeapRegion* hr) {
1733 if (hr->continuesHumongous()) {
1734 // We will ignore these here and process them when their
1735 // associated "starts humongous" region is processed (see
1736 // set_bit_for_heap_region()). Note that we cannot rely on their
1737 // associated "starts humongous" region to have their bit set to
1738 // 1 since, due to the region chunking in the parallel region
1739 // iteration, a "continues humongous" region might be visited
1740 // before its associated "starts humongous".
1741 return false;
1742 }
1744 HeapWord* ntams = hr->next_top_at_mark_start();
1745 HeapWord* top = hr->top();
1747 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1749 // Mark the allocated-since-marking portion...
1750 if (ntams < top) {
1751 // This definitely means the region has live objects.
1752 set_bit_for_region(hr);
1754 // Now set the bits in the card bitmap for [ntams, top)
1755 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1756 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1758 // Note: if we're looking at the last region in heap - top
1759 // could be actually just beyond the end of the heap; end_idx
1760 // will then correspond to a (non-existent) card that is also
1761 // just beyond the heap.
1762 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1763 // end of object is not card aligned - increment to cover
1764 // all the cards spanned by the object
1765 end_idx += 1;
1766 }
1768 assert(end_idx <= _card_bm->size(),
1769 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1770 end_idx, _card_bm->size()));
1771 assert(start_idx < _card_bm->size(),
1772 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1773 start_idx, _card_bm->size()));
1775 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1776 }
1778 // Set the bit for the region if it contains live data
1779 if (hr->next_marked_bytes() > 0) {
1780 set_bit_for_region(hr);
1781 }
1783 return false;
1784 }
1785 };
1787 class G1ParFinalCountTask: public AbstractGangTask {
1788 protected:
1789 G1CollectedHeap* _g1h;
1790 ConcurrentMark* _cm;
1791 BitMap* _actual_region_bm;
1792 BitMap* _actual_card_bm;
1794 uint _n_workers;
1796 public:
1797 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1798 : AbstractGangTask("G1 final counting"),
1799 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1800 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1801 _n_workers(0) {
1802 // Use the value already set as the number of active threads
1803 // in the call to run_task().
1804 if (G1CollectedHeap::use_parallel_gc_threads()) {
1805 assert( _g1h->workers()->active_workers() > 0,
1806 "Should have been previously set");
1807 _n_workers = _g1h->workers()->active_workers();
1808 } else {
1809 _n_workers = 1;
1810 }
1811 }
1813 void work(uint worker_id) {
1814 assert(worker_id < _n_workers, "invariant");
1816 FinalCountDataUpdateClosure final_update_cl(_g1h,
1817 _actual_region_bm,
1818 _actual_card_bm);
1820 if (G1CollectedHeap::use_parallel_gc_threads()) {
1821 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1822 worker_id,
1823 _n_workers,
1824 HeapRegion::FinalCountClaimValue);
1825 } else {
1826 _g1h->heap_region_iterate(&final_update_cl);
1827 }
1828 }
1829 };
1831 class G1ParNoteEndTask;
1833 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1834 G1CollectedHeap* _g1;
1835 size_t _max_live_bytes;
1836 uint _regions_claimed;
1837 size_t _freed_bytes;
1838 FreeRegionList* _local_cleanup_list;
1839 HeapRegionSetCount _old_regions_removed;
1840 HeapRegionSetCount _humongous_regions_removed;
1841 HRRSCleanupTask* _hrrs_cleanup_task;
1842 double _claimed_region_time;
1843 double _max_region_time;
1845 public:
1846 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1847 FreeRegionList* local_cleanup_list,
1848 HRRSCleanupTask* hrrs_cleanup_task) :
1849 _g1(g1),
1850 _max_live_bytes(0), _regions_claimed(0),
1851 _freed_bytes(0),
1852 _claimed_region_time(0.0), _max_region_time(0.0),
1853 _local_cleanup_list(local_cleanup_list),
1854 _old_regions_removed(),
1855 _humongous_regions_removed(),
1856 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1858 size_t freed_bytes() { return _freed_bytes; }
1859 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1860 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1862 bool doHeapRegion(HeapRegion *hr) {
1863 if (hr->continuesHumongous()) {
1864 return false;
1865 }
1866 // We use a claim value of zero here because all regions
1867 // were claimed with value 1 in the FinalCount task.
1868 _g1->reset_gc_time_stamps(hr);
1869 double start = os::elapsedTime();
1870 _regions_claimed++;
1871 hr->note_end_of_marking();
1872 _max_live_bytes += hr->max_live_bytes();
1874 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1875 _freed_bytes += hr->used();
1876 hr->set_containing_set(NULL);
1877 if (hr->isHumongous()) {
1878 assert(hr->startsHumongous(), "we should only see starts humongous");
1879 _humongous_regions_removed.increment(1u, hr->capacity());
1880 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1881 } else {
1882 _old_regions_removed.increment(1u, hr->capacity());
1883 _g1->free_region(hr, _local_cleanup_list, true);
1884 }
1885 } else {
1886 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1887 }
1889 double region_time = (os::elapsedTime() - start);
1890 _claimed_region_time += region_time;
1891 if (region_time > _max_region_time) {
1892 _max_region_time = region_time;
1893 }
1894 return false;
1895 }
1897 size_t max_live_bytes() { return _max_live_bytes; }
1898 uint regions_claimed() { return _regions_claimed; }
1899 double claimed_region_time_sec() { return _claimed_region_time; }
1900 double max_region_time_sec() { return _max_region_time; }
1901 };
1903 class G1ParNoteEndTask: public AbstractGangTask {
1904 friend class G1NoteEndOfConcMarkClosure;
1906 protected:
1907 G1CollectedHeap* _g1h;
1908 size_t _max_live_bytes;
1909 size_t _freed_bytes;
1910 FreeRegionList* _cleanup_list;
1912 public:
1913 G1ParNoteEndTask(G1CollectedHeap* g1h,
1914 FreeRegionList* cleanup_list) :
1915 AbstractGangTask("G1 note end"), _g1h(g1h),
1916 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1918 void work(uint worker_id) {
1919 double start = os::elapsedTime();
1920 FreeRegionList local_cleanup_list("Local Cleanup List");
1921 HRRSCleanupTask hrrs_cleanup_task;
1922 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1923 &hrrs_cleanup_task);
1924 if (G1CollectedHeap::use_parallel_gc_threads()) {
1925 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1926 _g1h->workers()->active_workers(),
1927 HeapRegion::NoteEndClaimValue);
1928 } else {
1929 _g1h->heap_region_iterate(&g1_note_end);
1930 }
1931 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1933 // Now update the lists
1934 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1935 {
1936 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1937 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1938 _max_live_bytes += g1_note_end.max_live_bytes();
1939 _freed_bytes += g1_note_end.freed_bytes();
1941 // If we iterate over the global cleanup list at the end of
1942 // cleanup to do this printing we will not guarantee to only
1943 // generate output for the newly-reclaimed regions (the list
1944 // might not be empty at the beginning of cleanup; we might
1945 // still be working on its previous contents). So we do the
1946 // printing here, before we append the new regions to the global
1947 // cleanup list.
1949 G1HRPrinter* hr_printer = _g1h->hr_printer();
1950 if (hr_printer->is_active()) {
1951 FreeRegionListIterator iter(&local_cleanup_list);
1952 while (iter.more_available()) {
1953 HeapRegion* hr = iter.get_next();
1954 hr_printer->cleanup(hr);
1955 }
1956 }
1958 _cleanup_list->add_ordered(&local_cleanup_list);
1959 assert(local_cleanup_list.is_empty(), "post-condition");
1961 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1962 }
1963 }
1964 size_t max_live_bytes() { return _max_live_bytes; }
1965 size_t freed_bytes() { return _freed_bytes; }
1966 };
1968 class G1ParScrubRemSetTask: public AbstractGangTask {
1969 protected:
1970 G1RemSet* _g1rs;
1971 BitMap* _region_bm;
1972 BitMap* _card_bm;
1973 public:
1974 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1975 BitMap* region_bm, BitMap* card_bm) :
1976 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1977 _region_bm(region_bm), _card_bm(card_bm) { }
1979 void work(uint worker_id) {
1980 if (G1CollectedHeap::use_parallel_gc_threads()) {
1981 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1982 HeapRegion::ScrubRemSetClaimValue);
1983 } else {
1984 _g1rs->scrub(_region_bm, _card_bm);
1985 }
1986 }
1988 };
1990 void ConcurrentMark::cleanup() {
1991 // world is stopped at this checkpoint
1992 assert(SafepointSynchronize::is_at_safepoint(),
1993 "world should be stopped");
1994 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1996 // If a full collection has happened, we shouldn't do this.
1997 if (has_aborted()) {
1998 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1999 return;
2000 }
2002 g1h->verify_region_sets_optional();
2004 if (VerifyDuringGC) {
2005 HandleMark hm; // handle scope
2006 Universe::heap()->prepare_for_verify();
2007 Universe::verify(VerifyOption_G1UsePrevMarking,
2008 " VerifyDuringGC:(before)");
2009 }
2010 g1h->check_bitmaps("Cleanup Start");
2012 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
2013 g1p->record_concurrent_mark_cleanup_start();
2015 double start = os::elapsedTime();
2017 HeapRegionRemSet::reset_for_cleanup_tasks();
2019 uint n_workers;
2021 // Do counting once more with the world stopped for good measure.
2022 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2024 if (G1CollectedHeap::use_parallel_gc_threads()) {
2025 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2026 "sanity check");
2028 g1h->set_par_threads();
2029 n_workers = g1h->n_par_threads();
2030 assert(g1h->n_par_threads() == n_workers,
2031 "Should not have been reset");
2032 g1h->workers()->run_task(&g1_par_count_task);
2033 // Done with the parallel phase so reset to 0.
2034 g1h->set_par_threads(0);
2036 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2037 "sanity check");
2038 } else {
2039 n_workers = 1;
2040 g1_par_count_task.work(0);
2041 }
2043 if (VerifyDuringGC) {
2044 // Verify that the counting data accumulated during marking matches
2045 // that calculated by walking the marking bitmap.
2047 // Bitmaps to hold expected values
2048 BitMap expected_region_bm(_region_bm.size(), true);
2049 BitMap expected_card_bm(_card_bm.size(), true);
2051 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2052 &_region_bm,
2053 &_card_bm,
2054 &expected_region_bm,
2055 &expected_card_bm);
2057 if (G1CollectedHeap::use_parallel_gc_threads()) {
2058 g1h->set_par_threads((int)n_workers);
2059 g1h->workers()->run_task(&g1_par_verify_task);
2060 // Done with the parallel phase so reset to 0.
2061 g1h->set_par_threads(0);
2063 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2064 "sanity check");
2065 } else {
2066 g1_par_verify_task.work(0);
2067 }
2069 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2070 }
2072 size_t start_used_bytes = g1h->used();
2073 g1h->set_marking_complete();
2075 double count_end = os::elapsedTime();
2076 double this_final_counting_time = (count_end - start);
2077 _total_counting_time += this_final_counting_time;
2079 if (G1PrintRegionLivenessInfo) {
2080 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2081 _g1h->heap_region_iterate(&cl);
2082 }
2084 // Install newly created mark bitMap as "prev".
2085 swapMarkBitMaps();
2087 g1h->reset_gc_time_stamp();
2089 // Note end of marking in all heap regions.
2090 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2091 if (G1CollectedHeap::use_parallel_gc_threads()) {
2092 g1h->set_par_threads((int)n_workers);
2093 g1h->workers()->run_task(&g1_par_note_end_task);
2094 g1h->set_par_threads(0);
2096 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2097 "sanity check");
2098 } else {
2099 g1_par_note_end_task.work(0);
2100 }
2101 g1h->check_gc_time_stamps();
2103 if (!cleanup_list_is_empty()) {
2104 // The cleanup list is not empty, so we'll have to process it
2105 // concurrently. Notify anyone else that might be wanting free
2106 // regions that there will be more free regions coming soon.
2107 g1h->set_free_regions_coming();
2108 }
2110 // call below, since it affects the metric by which we sort the heap
2111 // regions.
2112 if (G1ScrubRemSets) {
2113 double rs_scrub_start = os::elapsedTime();
2114 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2115 if (G1CollectedHeap::use_parallel_gc_threads()) {
2116 g1h->set_par_threads((int)n_workers);
2117 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2118 g1h->set_par_threads(0);
2120 assert(g1h->check_heap_region_claim_values(
2121 HeapRegion::ScrubRemSetClaimValue),
2122 "sanity check");
2123 } else {
2124 g1_par_scrub_rs_task.work(0);
2125 }
2127 double rs_scrub_end = os::elapsedTime();
2128 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2129 _total_rs_scrub_time += this_rs_scrub_time;
2130 }
2132 // this will also free any regions totally full of garbage objects,
2133 // and sort the regions.
2134 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2136 // Statistics.
2137 double end = os::elapsedTime();
2138 _cleanup_times.add((end - start) * 1000.0);
2140 if (G1Log::fine()) {
2141 g1h->print_size_transition(gclog_or_tty,
2142 start_used_bytes,
2143 g1h->used(),
2144 g1h->capacity());
2145 }
2147 // Clean up will have freed any regions completely full of garbage.
2148 // Update the soft reference policy with the new heap occupancy.
2149 Universe::update_heap_info_at_gc();
2151 if (VerifyDuringGC) {
2152 HandleMark hm; // handle scope
2153 Universe::heap()->prepare_for_verify();
2154 Universe::verify(VerifyOption_G1UsePrevMarking,
2155 " VerifyDuringGC:(after)");
2156 }
2157 g1h->check_bitmaps("Cleanup End");
2159 g1h->verify_region_sets_optional();
2161 // We need to make this be a "collection" so any collection pause that
2162 // races with it goes around and waits for completeCleanup to finish.
2163 g1h->increment_total_collections();
2165 // Clean out dead classes and update Metaspace sizes.
2166 if (ClassUnloadingWithConcurrentMark) {
2167 ClassLoaderDataGraph::purge();
2168 }
2169 MetaspaceGC::compute_new_size();
2171 // We reclaimed old regions so we should calculate the sizes to make
2172 // sure we update the old gen/space data.
2173 g1h->g1mm()->update_sizes();
2175 g1h->trace_heap_after_concurrent_cycle();
2176 }
2178 void ConcurrentMark::completeCleanup() {
2179 if (has_aborted()) return;
2181 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2183 _cleanup_list.verify_optional();
2184 FreeRegionList tmp_free_list("Tmp Free List");
2186 if (G1ConcRegionFreeingVerbose) {
2187 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2188 "cleanup list has %u entries",
2189 _cleanup_list.length());
2190 }
2192 // No one else should be accessing the _cleanup_list at this point,
2193 // so it is not necessary to take any locks
2194 while (!_cleanup_list.is_empty()) {
2195 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2196 assert(hr != NULL, "Got NULL from a non-empty list");
2197 hr->par_clear();
2198 tmp_free_list.add_ordered(hr);
2200 // Instead of adding one region at a time to the secondary_free_list,
2201 // we accumulate them in the local list and move them a few at a
2202 // time. This also cuts down on the number of notify_all() calls
2203 // we do during this process. We'll also append the local list when
2204 // _cleanup_list is empty (which means we just removed the last
2205 // region from the _cleanup_list).
2206 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2207 _cleanup_list.is_empty()) {
2208 if (G1ConcRegionFreeingVerbose) {
2209 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2210 "appending %u entries to the secondary_free_list, "
2211 "cleanup list still has %u entries",
2212 tmp_free_list.length(),
2213 _cleanup_list.length());
2214 }
2216 {
2217 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2218 g1h->secondary_free_list_add(&tmp_free_list);
2219 SecondaryFreeList_lock->notify_all();
2220 }
2222 if (G1StressConcRegionFreeing) {
2223 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2224 os::sleep(Thread::current(), (jlong) 1, false);
2225 }
2226 }
2227 }
2228 }
2229 assert(tmp_free_list.is_empty(), "post-condition");
2230 }
2232 // Supporting Object and Oop closures for reference discovery
2233 // and processing in during marking
2235 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2236 HeapWord* addr = (HeapWord*)obj;
2237 return addr != NULL &&
2238 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2239 }
2241 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2242 // Uses the CMTask associated with a worker thread (for serial reference
2243 // processing the CMTask for worker 0 is used) to preserve (mark) and
2244 // trace referent objects.
2245 //
2246 // Using the CMTask and embedded local queues avoids having the worker
2247 // threads operating on the global mark stack. This reduces the risk
2248 // of overflowing the stack - which we would rather avoid at this late
2249 // state. Also using the tasks' local queues removes the potential
2250 // of the workers interfering with each other that could occur if
2251 // operating on the global stack.
2253 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2254 ConcurrentMark* _cm;
2255 CMTask* _task;
2256 int _ref_counter_limit;
2257 int _ref_counter;
2258 bool _is_serial;
2259 public:
2260 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2261 _cm(cm), _task(task), _is_serial(is_serial),
2262 _ref_counter_limit(G1RefProcDrainInterval) {
2263 assert(_ref_counter_limit > 0, "sanity");
2264 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2265 _ref_counter = _ref_counter_limit;
2266 }
2268 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2269 virtual void do_oop( oop* p) { do_oop_work(p); }
2271 template <class T> void do_oop_work(T* p) {
2272 if (!_cm->has_overflown()) {
2273 oop obj = oopDesc::load_decode_heap_oop(p);
2274 if (_cm->verbose_high()) {
2275 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2276 "*"PTR_FORMAT" = "PTR_FORMAT,
2277 _task->worker_id(), p2i(p), p2i((void*) obj));
2278 }
2280 _task->deal_with_reference(obj);
2281 _ref_counter--;
2283 if (_ref_counter == 0) {
2284 // We have dealt with _ref_counter_limit references, pushing them
2285 // and objects reachable from them on to the local stack (and
2286 // possibly the global stack). Call CMTask::do_marking_step() to
2287 // process these entries.
2288 //
2289 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2290 // there's nothing more to do (i.e. we're done with the entries that
2291 // were pushed as a result of the CMTask::deal_with_reference() calls
2292 // above) or we overflow.
2293 //
2294 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2295 // flag while there may still be some work to do. (See the comment at
2296 // the beginning of CMTask::do_marking_step() for those conditions -
2297 // one of which is reaching the specified time target.) It is only
2298 // when CMTask::do_marking_step() returns without setting the
2299 // has_aborted() flag that the marking step has completed.
2300 do {
2301 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2302 _task->do_marking_step(mark_step_duration_ms,
2303 false /* do_termination */,
2304 _is_serial);
2305 } while (_task->has_aborted() && !_cm->has_overflown());
2306 _ref_counter = _ref_counter_limit;
2307 }
2308 } else {
2309 if (_cm->verbose_high()) {
2310 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2311 }
2312 }
2313 }
2314 };
2316 // 'Drain' oop closure used by both serial and parallel reference processing.
2317 // Uses the CMTask associated with a given worker thread (for serial
2318 // reference processing the CMtask for worker 0 is used). Calls the
2319 // do_marking_step routine, with an unbelievably large timeout value,
2320 // to drain the marking data structures of the remaining entries
2321 // added by the 'keep alive' oop closure above.
2323 class G1CMDrainMarkingStackClosure: public VoidClosure {
2324 ConcurrentMark* _cm;
2325 CMTask* _task;
2326 bool _is_serial;
2327 public:
2328 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2329 _cm(cm), _task(task), _is_serial(is_serial) {
2330 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2331 }
2333 void do_void() {
2334 do {
2335 if (_cm->verbose_high()) {
2336 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2337 _task->worker_id(), BOOL_TO_STR(_is_serial));
2338 }
2340 // We call CMTask::do_marking_step() to completely drain the local
2341 // and global marking stacks of entries pushed by the 'keep alive'
2342 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2343 //
2344 // CMTask::do_marking_step() is called in a loop, which we'll exit
2345 // if there's nothing more to do (i.e. we'completely drained the
2346 // entries that were pushed as a a result of applying the 'keep alive'
2347 // closure to the entries on the discovered ref lists) or we overflow
2348 // the global marking stack.
2349 //
2350 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2351 // flag while there may still be some work to do. (See the comment at
2352 // the beginning of CMTask::do_marking_step() for those conditions -
2353 // one of which is reaching the specified time target.) It is only
2354 // when CMTask::do_marking_step() returns without setting the
2355 // has_aborted() flag that the marking step has completed.
2357 _task->do_marking_step(1000000000.0 /* something very large */,
2358 true /* do_termination */,
2359 _is_serial);
2360 } while (_task->has_aborted() && !_cm->has_overflown());
2361 }
2362 };
2364 // Implementation of AbstractRefProcTaskExecutor for parallel
2365 // reference processing at the end of G1 concurrent marking
2367 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2368 private:
2369 G1CollectedHeap* _g1h;
2370 ConcurrentMark* _cm;
2371 WorkGang* _workers;
2372 int _active_workers;
2374 public:
2375 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2376 ConcurrentMark* cm,
2377 WorkGang* workers,
2378 int n_workers) :
2379 _g1h(g1h), _cm(cm),
2380 _workers(workers), _active_workers(n_workers) { }
2382 // Executes the given task using concurrent marking worker threads.
2383 virtual void execute(ProcessTask& task);
2384 virtual void execute(EnqueueTask& task);
2385 };
2387 class G1CMRefProcTaskProxy: public AbstractGangTask {
2388 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2389 ProcessTask& _proc_task;
2390 G1CollectedHeap* _g1h;
2391 ConcurrentMark* _cm;
2393 public:
2394 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2395 G1CollectedHeap* g1h,
2396 ConcurrentMark* cm) :
2397 AbstractGangTask("Process reference objects in parallel"),
2398 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2399 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2400 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2401 }
2403 virtual void work(uint worker_id) {
2404 ResourceMark rm;
2405 HandleMark hm;
2406 CMTask* task = _cm->task(worker_id);
2407 G1CMIsAliveClosure g1_is_alive(_g1h);
2408 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2409 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2411 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2412 }
2413 };
2415 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2416 assert(_workers != NULL, "Need parallel worker threads.");
2417 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2419 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2421 // We need to reset the concurrency level before each
2422 // proxy task execution, so that the termination protocol
2423 // and overflow handling in CMTask::do_marking_step() knows
2424 // how many workers to wait for.
2425 _cm->set_concurrency(_active_workers);
2426 _g1h->set_par_threads(_active_workers);
2427 _workers->run_task(&proc_task_proxy);
2428 _g1h->set_par_threads(0);
2429 }
2431 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2432 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2433 EnqueueTask& _enq_task;
2435 public:
2436 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2437 AbstractGangTask("Enqueue reference objects in parallel"),
2438 _enq_task(enq_task) { }
2440 virtual void work(uint worker_id) {
2441 _enq_task.work(worker_id);
2442 }
2443 };
2445 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2446 assert(_workers != NULL, "Need parallel worker threads.");
2447 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2449 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2451 // Not strictly necessary but...
2452 //
2453 // We need to reset the concurrency level before each
2454 // proxy task execution, so that the termination protocol
2455 // and overflow handling in CMTask::do_marking_step() knows
2456 // how many workers to wait for.
2457 _cm->set_concurrency(_active_workers);
2458 _g1h->set_par_threads(_active_workers);
2459 _workers->run_task(&enq_task_proxy);
2460 _g1h->set_par_threads(0);
2461 }
2463 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2464 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2465 }
2467 // Helper class to get rid of some boilerplate code.
2468 class G1RemarkGCTraceTime : public GCTraceTime {
2469 static bool doit_and_prepend(bool doit) {
2470 if (doit) {
2471 gclog_or_tty->put(' ');
2472 }
2473 return doit;
2474 }
2476 public:
2477 G1RemarkGCTraceTime(const char* title, bool doit)
2478 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
2479 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
2480 }
2481 };
2483 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2484 if (has_overflown()) {
2485 // Skip processing the discovered references if we have
2486 // overflown the global marking stack. Reference objects
2487 // only get discovered once so it is OK to not
2488 // de-populate the discovered reference lists. We could have,
2489 // but the only benefit would be that, when marking restarts,
2490 // less reference objects are discovered.
2491 return;
2492 }
2494 ResourceMark rm;
2495 HandleMark hm;
2497 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2499 // Is alive closure.
2500 G1CMIsAliveClosure g1_is_alive(g1h);
2502 // Inner scope to exclude the cleaning of the string and symbol
2503 // tables from the displayed time.
2504 {
2505 if (G1Log::finer()) {
2506 gclog_or_tty->put(' ');
2507 }
2508 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm(), concurrent_gc_id());
2510 ReferenceProcessor* rp = g1h->ref_processor_cm();
2512 // See the comment in G1CollectedHeap::ref_processing_init()
2513 // about how reference processing currently works in G1.
2515 // Set the soft reference policy
2516 rp->setup_policy(clear_all_soft_refs);
2517 assert(_markStack.isEmpty(), "mark stack should be empty");
2519 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2520 // in serial reference processing. Note these closures are also
2521 // used for serially processing (by the the current thread) the
2522 // JNI references during parallel reference processing.
2523 //
2524 // These closures do not need to synchronize with the worker
2525 // threads involved in parallel reference processing as these
2526 // instances are executed serially by the current thread (e.g.
2527 // reference processing is not multi-threaded and is thus
2528 // performed by the current thread instead of a gang worker).
2529 //
2530 // The gang tasks involved in parallel reference procssing create
2531 // their own instances of these closures, which do their own
2532 // synchronization among themselves.
2533 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2534 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2536 // We need at least one active thread. If reference processing
2537 // is not multi-threaded we use the current (VMThread) thread,
2538 // otherwise we use the work gang from the G1CollectedHeap and
2539 // we utilize all the worker threads we can.
2540 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2541 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2542 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2544 // Parallel processing task executor.
2545 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2546 g1h->workers(), active_workers);
2547 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2549 // Set the concurrency level. The phase was already set prior to
2550 // executing the remark task.
2551 set_concurrency(active_workers);
2553 // Set the degree of MT processing here. If the discovery was done MT,
2554 // the number of threads involved during discovery could differ from
2555 // the number of active workers. This is OK as long as the discovered
2556 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2557 rp->set_active_mt_degree(active_workers);
2559 // Process the weak references.
2560 const ReferenceProcessorStats& stats =
2561 rp->process_discovered_references(&g1_is_alive,
2562 &g1_keep_alive,
2563 &g1_drain_mark_stack,
2564 executor,
2565 g1h->gc_timer_cm(),
2566 concurrent_gc_id());
2567 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2569 // The do_oop work routines of the keep_alive and drain_marking_stack
2570 // oop closures will set the has_overflown flag if we overflow the
2571 // global marking stack.
2573 assert(_markStack.overflow() || _markStack.isEmpty(),
2574 "mark stack should be empty (unless it overflowed)");
2576 if (_markStack.overflow()) {
2577 // This should have been done already when we tried to push an
2578 // entry on to the global mark stack. But let's do it again.
2579 set_has_overflown();
2580 }
2582 assert(rp->num_q() == active_workers, "why not");
2584 rp->enqueue_discovered_references(executor);
2586 rp->verify_no_references_recorded();
2587 assert(!rp->discovery_enabled(), "Post condition");
2588 }
2590 if (has_overflown()) {
2591 // We can not trust g1_is_alive if the marking stack overflowed
2592 return;
2593 }
2595 assert(_markStack.isEmpty(), "Marking should have completed");
2597 // Unload Klasses, String, Symbols, Code Cache, etc.
2598 {
2599 G1RemarkGCTraceTime trace("Unloading", G1Log::finer());
2601 if (ClassUnloadingWithConcurrentMark) {
2602 bool purged_classes;
2604 {
2605 G1RemarkGCTraceTime trace("System Dictionary Unloading", G1Log::finest());
2606 purged_classes = SystemDictionary::do_unloading(&g1_is_alive);
2607 }
2609 {
2610 G1RemarkGCTraceTime trace("Parallel Unloading", G1Log::finest());
2611 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2612 }
2613 }
2615 if (G1StringDedup::is_enabled()) {
2616 G1RemarkGCTraceTime trace("String Deduplication Unlink", G1Log::finest());
2617 G1StringDedup::unlink(&g1_is_alive);
2618 }
2619 }
2620 }
2622 void ConcurrentMark::swapMarkBitMaps() {
2623 CMBitMapRO* temp = _prevMarkBitMap;
2624 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2625 _nextMarkBitMap = (CMBitMap*) temp;
2626 }
2628 class CMObjectClosure;
2630 // Closure for iterating over objects, currently only used for
2631 // processing SATB buffers.
2632 class CMObjectClosure : public ObjectClosure {
2633 private:
2634 CMTask* _task;
2636 public:
2637 void do_object(oop obj) {
2638 _task->deal_with_reference(obj);
2639 }
2641 CMObjectClosure(CMTask* task) : _task(task) { }
2642 };
2644 class G1RemarkThreadsClosure : public ThreadClosure {
2645 CMObjectClosure _cm_obj;
2646 G1CMOopClosure _cm_cl;
2647 MarkingCodeBlobClosure _code_cl;
2648 int _thread_parity;
2649 bool _is_par;
2651 public:
2652 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) :
2653 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2654 _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {}
2656 void do_thread(Thread* thread) {
2657 if (thread->is_Java_thread()) {
2658 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2659 JavaThread* jt = (JavaThread*)thread;
2661 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2662 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2663 // * Alive if on the stack of an executing method
2664 // * Weakly reachable otherwise
2665 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2666 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2667 jt->nmethods_do(&_code_cl);
2669 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2670 }
2671 } else if (thread->is_VM_thread()) {
2672 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2673 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2674 }
2675 }
2676 }
2677 };
2679 class CMRemarkTask: public AbstractGangTask {
2680 private:
2681 ConcurrentMark* _cm;
2682 bool _is_serial;
2683 public:
2684 void work(uint worker_id) {
2685 // Since all available tasks are actually started, we should
2686 // only proceed if we're supposed to be actived.
2687 if (worker_id < _cm->active_tasks()) {
2688 CMTask* task = _cm->task(worker_id);
2689 task->record_start_time();
2690 {
2691 ResourceMark rm;
2692 HandleMark hm;
2694 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial);
2695 Threads::threads_do(&threads_f);
2696 }
2698 do {
2699 task->do_marking_step(1000000000.0 /* something very large */,
2700 true /* do_termination */,
2701 _is_serial);
2702 } while (task->has_aborted() && !_cm->has_overflown());
2703 // If we overflow, then we do not want to restart. We instead
2704 // want to abort remark and do concurrent marking again.
2705 task->record_end_time();
2706 }
2707 }
2709 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2710 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2711 _cm->terminator()->reset_for_reuse(active_workers);
2712 }
2713 };
2715 void ConcurrentMark::checkpointRootsFinalWork() {
2716 ResourceMark rm;
2717 HandleMark hm;
2718 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2720 G1RemarkGCTraceTime trace("Finalize Marking", G1Log::finer());
2722 g1h->ensure_parsability(false);
2724 if (G1CollectedHeap::use_parallel_gc_threads()) {
2725 G1CollectedHeap::StrongRootsScope srs(g1h);
2726 // this is remark, so we'll use up all active threads
2727 uint active_workers = g1h->workers()->active_workers();
2728 if (active_workers == 0) {
2729 assert(active_workers > 0, "Should have been set earlier");
2730 active_workers = (uint) ParallelGCThreads;
2731 g1h->workers()->set_active_workers(active_workers);
2732 }
2733 set_concurrency_and_phase(active_workers, false /* concurrent */);
2734 // Leave _parallel_marking_threads at it's
2735 // value originally calculated in the ConcurrentMark
2736 // constructor and pass values of the active workers
2737 // through the gang in the task.
2739 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2740 // We will start all available threads, even if we decide that the
2741 // active_workers will be fewer. The extra ones will just bail out
2742 // immediately.
2743 g1h->set_par_threads(active_workers);
2744 g1h->workers()->run_task(&remarkTask);
2745 g1h->set_par_threads(0);
2746 } else {
2747 G1CollectedHeap::StrongRootsScope srs(g1h);
2748 uint active_workers = 1;
2749 set_concurrency_and_phase(active_workers, false /* concurrent */);
2751 // Note - if there's no work gang then the VMThread will be
2752 // the thread to execute the remark - serially. We have
2753 // to pass true for the is_serial parameter so that
2754 // CMTask::do_marking_step() doesn't enter the sync
2755 // barriers in the event of an overflow. Doing so will
2756 // cause an assert that the current thread is not a
2757 // concurrent GC thread.
2758 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2759 remarkTask.work(0);
2760 }
2761 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2762 guarantee(has_overflown() ||
2763 satb_mq_set.completed_buffers_num() == 0,
2764 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2765 BOOL_TO_STR(has_overflown()),
2766 satb_mq_set.completed_buffers_num()));
2768 print_stats();
2769 }
2771 #ifndef PRODUCT
2773 class PrintReachableOopClosure: public OopClosure {
2774 private:
2775 G1CollectedHeap* _g1h;
2776 outputStream* _out;
2777 VerifyOption _vo;
2778 bool _all;
2780 public:
2781 PrintReachableOopClosure(outputStream* out,
2782 VerifyOption vo,
2783 bool all) :
2784 _g1h(G1CollectedHeap::heap()),
2785 _out(out), _vo(vo), _all(all) { }
2787 void do_oop(narrowOop* p) { do_oop_work(p); }
2788 void do_oop( oop* p) { do_oop_work(p); }
2790 template <class T> void do_oop_work(T* p) {
2791 oop obj = oopDesc::load_decode_heap_oop(p);
2792 const char* str = NULL;
2793 const char* str2 = "";
2795 if (obj == NULL) {
2796 str = "";
2797 } else if (!_g1h->is_in_g1_reserved(obj)) {
2798 str = " O";
2799 } else {
2800 HeapRegion* hr = _g1h->heap_region_containing(obj);
2801 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2802 bool marked = _g1h->is_marked(obj, _vo);
2804 if (over_tams) {
2805 str = " >";
2806 if (marked) {
2807 str2 = " AND MARKED";
2808 }
2809 } else if (marked) {
2810 str = " M";
2811 } else {
2812 str = " NOT";
2813 }
2814 }
2816 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2817 p2i(p), p2i((void*) obj), str, str2);
2818 }
2819 };
2821 class PrintReachableObjectClosure : public ObjectClosure {
2822 private:
2823 G1CollectedHeap* _g1h;
2824 outputStream* _out;
2825 VerifyOption _vo;
2826 bool _all;
2827 HeapRegion* _hr;
2829 public:
2830 PrintReachableObjectClosure(outputStream* out,
2831 VerifyOption vo,
2832 bool all,
2833 HeapRegion* hr) :
2834 _g1h(G1CollectedHeap::heap()),
2835 _out(out), _vo(vo), _all(all), _hr(hr) { }
2837 void do_object(oop o) {
2838 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2839 bool marked = _g1h->is_marked(o, _vo);
2840 bool print_it = _all || over_tams || marked;
2842 if (print_it) {
2843 _out->print_cr(" "PTR_FORMAT"%s",
2844 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2845 PrintReachableOopClosure oopCl(_out, _vo, _all);
2846 o->oop_iterate_no_header(&oopCl);
2847 }
2848 }
2849 };
2851 class PrintReachableRegionClosure : public HeapRegionClosure {
2852 private:
2853 G1CollectedHeap* _g1h;
2854 outputStream* _out;
2855 VerifyOption _vo;
2856 bool _all;
2858 public:
2859 bool doHeapRegion(HeapRegion* hr) {
2860 HeapWord* b = hr->bottom();
2861 HeapWord* e = hr->end();
2862 HeapWord* t = hr->top();
2863 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2864 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2865 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2866 _out->cr();
2868 HeapWord* from = b;
2869 HeapWord* to = t;
2871 if (to > from) {
2872 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2873 _out->cr();
2874 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2875 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2876 _out->cr();
2877 }
2879 return false;
2880 }
2882 PrintReachableRegionClosure(outputStream* out,
2883 VerifyOption vo,
2884 bool all) :
2885 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2886 };
2888 void ConcurrentMark::print_reachable(const char* str,
2889 VerifyOption vo,
2890 bool all) {
2891 gclog_or_tty->cr();
2892 gclog_or_tty->print_cr("== Doing heap dump... ");
2894 if (G1PrintReachableBaseFile == NULL) {
2895 gclog_or_tty->print_cr(" #### error: no base file defined");
2896 return;
2897 }
2899 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2900 (JVM_MAXPATHLEN - 1)) {
2901 gclog_or_tty->print_cr(" #### error: file name too long");
2902 return;
2903 }
2905 char file_name[JVM_MAXPATHLEN];
2906 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2907 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2909 fileStream fout(file_name);
2910 if (!fout.is_open()) {
2911 gclog_or_tty->print_cr(" #### error: could not open file");
2912 return;
2913 }
2915 outputStream* out = &fout;
2916 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2917 out->cr();
2919 out->print_cr("--- ITERATING OVER REGIONS");
2920 out->cr();
2921 PrintReachableRegionClosure rcl(out, vo, all);
2922 _g1h->heap_region_iterate(&rcl);
2923 out->cr();
2925 gclog_or_tty->print_cr(" done");
2926 gclog_or_tty->flush();
2927 }
2929 #endif // PRODUCT
2931 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2932 // Note we are overriding the read-only view of the prev map here, via
2933 // the cast.
2934 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2935 }
2937 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2938 _nextMarkBitMap->clearRange(mr);
2939 }
2941 HeapRegion*
2942 ConcurrentMark::claim_region(uint worker_id) {
2943 // "checkpoint" the finger
2944 HeapWord* finger = _finger;
2946 // _heap_end will not change underneath our feet; it only changes at
2947 // yield points.
2948 while (finger < _heap_end) {
2949 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2951 // Note on how this code handles humongous regions. In the
2952 // normal case the finger will reach the start of a "starts
2953 // humongous" (SH) region. Its end will either be the end of the
2954 // last "continues humongous" (CH) region in the sequence, or the
2955 // standard end of the SH region (if the SH is the only region in
2956 // the sequence). That way claim_region() will skip over the CH
2957 // regions. However, there is a subtle race between a CM thread
2958 // executing this method and a mutator thread doing a humongous
2959 // object allocation. The two are not mutually exclusive as the CM
2960 // thread does not need to hold the Heap_lock when it gets
2961 // here. So there is a chance that claim_region() will come across
2962 // a free region that's in the progress of becoming a SH or a CH
2963 // region. In the former case, it will either
2964 // a) Miss the update to the region's end, in which case it will
2965 // visit every subsequent CH region, will find their bitmaps
2966 // empty, and do nothing, or
2967 // b) Will observe the update of the region's end (in which case
2968 // it will skip the subsequent CH regions).
2969 // If it comes across a region that suddenly becomes CH, the
2970 // scenario will be similar to b). So, the race between
2971 // claim_region() and a humongous object allocation might force us
2972 // to do a bit of unnecessary work (due to some unnecessary bitmap
2973 // iterations) but it should not introduce and correctness issues.
2974 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2976 // Above heap_region_containing_raw may return NULL as we always scan claim
2977 // until the end of the heap. In this case, just jump to the next region.
2978 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2980 // Is the gap between reading the finger and doing the CAS too long?
2981 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2982 if (res == finger && curr_region != NULL) {
2983 // we succeeded
2984 HeapWord* bottom = curr_region->bottom();
2985 HeapWord* limit = curr_region->next_top_at_mark_start();
2987 if (verbose_low()) {
2988 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2989 "["PTR_FORMAT", "PTR_FORMAT"), "
2990 "limit = "PTR_FORMAT,
2991 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2992 }
2994 // notice that _finger == end cannot be guaranteed here since,
2995 // someone else might have moved the finger even further
2996 assert(_finger >= end, "the finger should have moved forward");
2998 if (verbose_low()) {
2999 gclog_or_tty->print_cr("[%u] we were successful with region = "
3000 PTR_FORMAT, worker_id, p2i(curr_region));
3001 }
3003 if (limit > bottom) {
3004 if (verbose_low()) {
3005 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
3006 "returning it ", worker_id, p2i(curr_region));
3007 }
3008 return curr_region;
3009 } else {
3010 assert(limit == bottom,
3011 "the region limit should be at bottom");
3012 if (verbose_low()) {
3013 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
3014 "returning NULL", worker_id, p2i(curr_region));
3015 }
3016 // we return NULL and the caller should try calling
3017 // claim_region() again.
3018 return NULL;
3019 }
3020 } else {
3021 assert(_finger > finger, "the finger should have moved forward");
3022 if (verbose_low()) {
3023 if (curr_region == NULL) {
3024 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
3025 "global finger = "PTR_FORMAT", "
3026 "our finger = "PTR_FORMAT,
3027 worker_id, p2i(_finger), p2i(finger));
3028 } else {
3029 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
3030 "global finger = "PTR_FORMAT", "
3031 "our finger = "PTR_FORMAT,
3032 worker_id, p2i(_finger), p2i(finger));
3033 }
3034 }
3036 // read it again
3037 finger = _finger;
3038 }
3039 }
3041 return NULL;
3042 }
3044 #ifndef PRODUCT
3045 enum VerifyNoCSetOopsPhase {
3046 VerifyNoCSetOopsStack,
3047 VerifyNoCSetOopsQueues,
3048 VerifyNoCSetOopsSATBCompleted,
3049 VerifyNoCSetOopsSATBThread
3050 };
3052 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
3053 private:
3054 G1CollectedHeap* _g1h;
3055 VerifyNoCSetOopsPhase _phase;
3056 int _info;
3058 const char* phase_str() {
3059 switch (_phase) {
3060 case VerifyNoCSetOopsStack: return "Stack";
3061 case VerifyNoCSetOopsQueues: return "Queue";
3062 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
3063 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
3064 default: ShouldNotReachHere();
3065 }
3066 return NULL;
3067 }
3069 void do_object_work(oop obj) {
3070 guarantee(!_g1h->obj_in_cs(obj),
3071 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
3072 p2i((void*) obj), phase_str(), _info));
3073 }
3075 public:
3076 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3078 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3079 _phase = phase;
3080 _info = info;
3081 }
3083 virtual void do_oop(oop* p) {
3084 oop obj = oopDesc::load_decode_heap_oop(p);
3085 do_object_work(obj);
3086 }
3088 virtual void do_oop(narrowOop* p) {
3089 // We should not come across narrow oops while scanning marking
3090 // stacks and SATB buffers.
3091 ShouldNotReachHere();
3092 }
3094 virtual void do_object(oop obj) {
3095 do_object_work(obj);
3096 }
3097 };
3099 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3100 bool verify_enqueued_buffers,
3101 bool verify_thread_buffers,
3102 bool verify_fingers) {
3103 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3104 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3105 return;
3106 }
3108 VerifyNoCSetOopsClosure cl;
3110 if (verify_stacks) {
3111 // Verify entries on the global mark stack
3112 cl.set_phase(VerifyNoCSetOopsStack);
3113 _markStack.oops_do(&cl);
3115 // Verify entries on the task queues
3116 for (uint i = 0; i < _max_worker_id; i += 1) {
3117 cl.set_phase(VerifyNoCSetOopsQueues, i);
3118 CMTaskQueue* queue = _task_queues->queue(i);
3119 queue->oops_do(&cl);
3120 }
3121 }
3123 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3125 // Verify entries on the enqueued SATB buffers
3126 if (verify_enqueued_buffers) {
3127 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3128 satb_qs.iterate_completed_buffers_read_only(&cl);
3129 }
3131 // Verify entries on the per-thread SATB buffers
3132 if (verify_thread_buffers) {
3133 cl.set_phase(VerifyNoCSetOopsSATBThread);
3134 satb_qs.iterate_thread_buffers_read_only(&cl);
3135 }
3137 if (verify_fingers) {
3138 // Verify the global finger
3139 HeapWord* global_finger = finger();
3140 if (global_finger != NULL && global_finger < _heap_end) {
3141 // The global finger always points to a heap region boundary. We
3142 // use heap_region_containing_raw() to get the containing region
3143 // given that the global finger could be pointing to a free region
3144 // which subsequently becomes continues humongous. If that
3145 // happens, heap_region_containing() will return the bottom of the
3146 // corresponding starts humongous region and the check below will
3147 // not hold any more.
3148 // Since we always iterate over all regions, we might get a NULL HeapRegion
3149 // here.
3150 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3151 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3152 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3153 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3154 }
3156 // Verify the task fingers
3157 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3158 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3159 CMTask* task = _tasks[i];
3160 HeapWord* task_finger = task->finger();
3161 if (task_finger != NULL && task_finger < _heap_end) {
3162 // See above note on the global finger verification.
3163 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3164 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3165 !task_hr->in_collection_set(),
3166 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3167 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3168 }
3169 }
3170 }
3171 }
3172 #endif // PRODUCT
3174 // Aggregate the counting data that was constructed concurrently
3175 // with marking.
3176 class AggregateCountDataHRClosure: public HeapRegionClosure {
3177 G1CollectedHeap* _g1h;
3178 ConcurrentMark* _cm;
3179 CardTableModRefBS* _ct_bs;
3180 BitMap* _cm_card_bm;
3181 uint _max_worker_id;
3183 public:
3184 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3185 BitMap* cm_card_bm,
3186 uint max_worker_id) :
3187 _g1h(g1h), _cm(g1h->concurrent_mark()),
3188 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3189 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3191 bool doHeapRegion(HeapRegion* hr) {
3192 if (hr->continuesHumongous()) {
3193 // We will ignore these here and process them when their
3194 // associated "starts humongous" region is processed.
3195 // Note that we cannot rely on their associated
3196 // "starts humongous" region to have their bit set to 1
3197 // since, due to the region chunking in the parallel region
3198 // iteration, a "continues humongous" region might be visited
3199 // before its associated "starts humongous".
3200 return false;
3201 }
3203 HeapWord* start = hr->bottom();
3204 HeapWord* limit = hr->next_top_at_mark_start();
3205 HeapWord* end = hr->end();
3207 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3208 err_msg("Preconditions not met - "
3209 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3210 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3211 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3213 assert(hr->next_marked_bytes() == 0, "Precondition");
3215 if (start == limit) {
3216 // NTAMS of this region has not been set so nothing to do.
3217 return false;
3218 }
3220 // 'start' should be in the heap.
3221 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3222 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3223 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3225 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3226 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3227 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3229 // If ntams is not card aligned then we bump card bitmap index
3230 // for limit so that we get the all the cards spanned by
3231 // the object ending at ntams.
3232 // Note: if this is the last region in the heap then ntams
3233 // could be actually just beyond the end of the the heap;
3234 // limit_idx will then correspond to a (non-existent) card
3235 // that is also outside the heap.
3236 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3237 limit_idx += 1;
3238 }
3240 assert(limit_idx <= end_idx, "or else use atomics");
3242 // Aggregate the "stripe" in the count data associated with hr.
3243 uint hrm_index = hr->hrm_index();
3244 size_t marked_bytes = 0;
3246 for (uint i = 0; i < _max_worker_id; i += 1) {
3247 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3248 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3250 // Fetch the marked_bytes in this region for task i and
3251 // add it to the running total for this region.
3252 marked_bytes += marked_bytes_array[hrm_index];
3254 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3255 // into the global card bitmap.
3256 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3258 while (scan_idx < limit_idx) {
3259 assert(task_card_bm->at(scan_idx) == true, "should be");
3260 _cm_card_bm->set_bit(scan_idx);
3261 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3263 // BitMap::get_next_one_offset() can handle the case when
3264 // its left_offset parameter is greater than its right_offset
3265 // parameter. It does, however, have an early exit if
3266 // left_offset == right_offset. So let's limit the value
3267 // passed in for left offset here.
3268 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3269 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3270 }
3271 }
3273 // Update the marked bytes for this region.
3274 hr->add_to_marked_bytes(marked_bytes);
3276 // Next heap region
3277 return false;
3278 }
3279 };
3281 class G1AggregateCountDataTask: public AbstractGangTask {
3282 protected:
3283 G1CollectedHeap* _g1h;
3284 ConcurrentMark* _cm;
3285 BitMap* _cm_card_bm;
3286 uint _max_worker_id;
3287 int _active_workers;
3289 public:
3290 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3291 ConcurrentMark* cm,
3292 BitMap* cm_card_bm,
3293 uint max_worker_id,
3294 int n_workers) :
3295 AbstractGangTask("Count Aggregation"),
3296 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3297 _max_worker_id(max_worker_id),
3298 _active_workers(n_workers) { }
3300 void work(uint worker_id) {
3301 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3303 if (G1CollectedHeap::use_parallel_gc_threads()) {
3304 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3305 _active_workers,
3306 HeapRegion::AggregateCountClaimValue);
3307 } else {
3308 _g1h->heap_region_iterate(&cl);
3309 }
3310 }
3311 };
3314 void ConcurrentMark::aggregate_count_data() {
3315 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3316 _g1h->workers()->active_workers() :
3317 1);
3319 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3320 _max_worker_id, n_workers);
3322 if (G1CollectedHeap::use_parallel_gc_threads()) {
3323 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3324 "sanity check");
3325 _g1h->set_par_threads(n_workers);
3326 _g1h->workers()->run_task(&g1_par_agg_task);
3327 _g1h->set_par_threads(0);
3329 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3330 "sanity check");
3331 _g1h->reset_heap_region_claim_values();
3332 } else {
3333 g1_par_agg_task.work(0);
3334 }
3335 _g1h->allocation_context_stats().update_at_remark();
3336 }
3338 // Clear the per-worker arrays used to store the per-region counting data
3339 void ConcurrentMark::clear_all_count_data() {
3340 // Clear the global card bitmap - it will be filled during
3341 // liveness count aggregation (during remark) and the
3342 // final counting task.
3343 _card_bm.clear();
3345 // Clear the global region bitmap - it will be filled as part
3346 // of the final counting task.
3347 _region_bm.clear();
3349 uint max_regions = _g1h->max_regions();
3350 assert(_max_worker_id > 0, "uninitialized");
3352 for (uint i = 0; i < _max_worker_id; i += 1) {
3353 BitMap* task_card_bm = count_card_bitmap_for(i);
3354 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3356 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3357 assert(marked_bytes_array != NULL, "uninitialized");
3359 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3360 task_card_bm->clear();
3361 }
3362 }
3364 void ConcurrentMark::print_stats() {
3365 if (verbose_stats()) {
3366 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3367 for (size_t i = 0; i < _active_tasks; ++i) {
3368 _tasks[i]->print_stats();
3369 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3370 }
3371 }
3372 }
3374 // abandon current marking iteration due to a Full GC
3375 void ConcurrentMark::abort() {
3376 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3377 // concurrent bitmap clearing.
3378 _nextMarkBitMap->clearAll();
3380 // Note we cannot clear the previous marking bitmap here
3381 // since VerifyDuringGC verifies the objects marked during
3382 // a full GC against the previous bitmap.
3384 // Clear the liveness counting data
3385 clear_all_count_data();
3386 // Empty mark stack
3387 reset_marking_state();
3388 for (uint i = 0; i < _max_worker_id; ++i) {
3389 _tasks[i]->clear_region_fields();
3390 }
3391 _first_overflow_barrier_sync.abort();
3392 _second_overflow_barrier_sync.abort();
3393 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3394 if (!gc_id.is_undefined()) {
3395 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3396 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3397 _aborted_gc_id = gc_id;
3398 }
3399 _has_aborted = true;
3401 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3402 satb_mq_set.abandon_partial_marking();
3403 // This can be called either during or outside marking, we'll read
3404 // the expected_active value from the SATB queue set.
3405 satb_mq_set.set_active_all_threads(
3406 false, /* new active value */
3407 satb_mq_set.is_active() /* expected_active */);
3409 _g1h->trace_heap_after_concurrent_cycle();
3410 _g1h->register_concurrent_cycle_end();
3411 }
3413 const GCId& ConcurrentMark::concurrent_gc_id() {
3414 if (has_aborted()) {
3415 return _aborted_gc_id;
3416 }
3417 return _g1h->gc_tracer_cm()->gc_id();
3418 }
3420 static void print_ms_time_info(const char* prefix, const char* name,
3421 NumberSeq& ns) {
3422 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3423 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3424 if (ns.num() > 0) {
3425 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3426 prefix, ns.sd(), ns.maximum());
3427 }
3428 }
3430 void ConcurrentMark::print_summary_info() {
3431 gclog_or_tty->print_cr(" Concurrent marking:");
3432 print_ms_time_info(" ", "init marks", _init_times);
3433 print_ms_time_info(" ", "remarks", _remark_times);
3434 {
3435 print_ms_time_info(" ", "final marks", _remark_mark_times);
3436 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3438 }
3439 print_ms_time_info(" ", "cleanups", _cleanup_times);
3440 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3441 _total_counting_time,
3442 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3443 (double)_cleanup_times.num()
3444 : 0.0));
3445 if (G1ScrubRemSets) {
3446 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3447 _total_rs_scrub_time,
3448 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3449 (double)_cleanup_times.num()
3450 : 0.0));
3451 }
3452 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3453 (_init_times.sum() + _remark_times.sum() +
3454 _cleanup_times.sum())/1000.0);
3455 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3456 "(%8.2f s marking).",
3457 cmThread()->vtime_accum(),
3458 cmThread()->vtime_mark_accum());
3459 }
3461 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3462 if (use_parallel_marking_threads()) {
3463 _parallel_workers->print_worker_threads_on(st);
3464 }
3465 }
3467 void ConcurrentMark::print_on_error(outputStream* st) const {
3468 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3469 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3470 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3471 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3472 }
3474 // We take a break if someone is trying to stop the world.
3475 bool ConcurrentMark::do_yield_check(uint worker_id) {
3476 if (SuspendibleThreadSet::should_yield()) {
3477 if (worker_id == 0) {
3478 _g1h->g1_policy()->record_concurrent_pause();
3479 }
3480 SuspendibleThreadSet::yield();
3481 return true;
3482 } else {
3483 return false;
3484 }
3485 }
3487 #ifndef PRODUCT
3488 // for debugging purposes
3489 void ConcurrentMark::print_finger() {
3490 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3491 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3492 for (uint i = 0; i < _max_worker_id; ++i) {
3493 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3494 }
3495 gclog_or_tty->cr();
3496 }
3497 #endif
3499 void CMTask::scan_object(oop obj) {
3500 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3502 if (_cm->verbose_high()) {
3503 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3504 _worker_id, p2i((void*) obj));
3505 }
3507 size_t obj_size = obj->size();
3508 _words_scanned += obj_size;
3510 obj->oop_iterate(_cm_oop_closure);
3511 statsOnly( ++_objs_scanned );
3512 check_limits();
3513 }
3515 // Closure for iteration over bitmaps
3516 class CMBitMapClosure : public BitMapClosure {
3517 private:
3518 // the bitmap that is being iterated over
3519 CMBitMap* _nextMarkBitMap;
3520 ConcurrentMark* _cm;
3521 CMTask* _task;
3523 public:
3524 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3525 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3527 bool do_bit(size_t offset) {
3528 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3529 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3530 assert( addr < _cm->finger(), "invariant");
3532 statsOnly( _task->increase_objs_found_on_bitmap() );
3533 assert(addr >= _task->finger(), "invariant");
3535 // We move that task's local finger along.
3536 _task->move_finger_to(addr);
3538 _task->scan_object(oop(addr));
3539 // we only partially drain the local queue and global stack
3540 _task->drain_local_queue(true);
3541 _task->drain_global_stack(true);
3543 // if the has_aborted flag has been raised, we need to bail out of
3544 // the iteration
3545 return !_task->has_aborted();
3546 }
3547 };
3549 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3550 ConcurrentMark* cm,
3551 CMTask* task)
3552 : _g1h(g1h), _cm(cm), _task(task) {
3553 assert(_ref_processor == NULL, "should be initialized to NULL");
3555 if (G1UseConcMarkReferenceProcessing) {
3556 _ref_processor = g1h->ref_processor_cm();
3557 assert(_ref_processor != NULL, "should not be NULL");
3558 }
3559 }
3561 void CMTask::setup_for_region(HeapRegion* hr) {
3562 assert(hr != NULL,
3563 "claim_region() should have filtered out NULL regions");
3564 assert(!hr->continuesHumongous(),
3565 "claim_region() should have filtered out continues humongous regions");
3567 if (_cm->verbose_low()) {
3568 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3569 _worker_id, p2i(hr));
3570 }
3572 _curr_region = hr;
3573 _finger = hr->bottom();
3574 update_region_limit();
3575 }
3577 void CMTask::update_region_limit() {
3578 HeapRegion* hr = _curr_region;
3579 HeapWord* bottom = hr->bottom();
3580 HeapWord* limit = hr->next_top_at_mark_start();
3582 if (limit == bottom) {
3583 if (_cm->verbose_low()) {
3584 gclog_or_tty->print_cr("[%u] found an empty region "
3585 "["PTR_FORMAT", "PTR_FORMAT")",
3586 _worker_id, p2i(bottom), p2i(limit));
3587 }
3588 // The region was collected underneath our feet.
3589 // We set the finger to bottom to ensure that the bitmap
3590 // iteration that will follow this will not do anything.
3591 // (this is not a condition that holds when we set the region up,
3592 // as the region is not supposed to be empty in the first place)
3593 _finger = bottom;
3594 } else if (limit >= _region_limit) {
3595 assert(limit >= _finger, "peace of mind");
3596 } else {
3597 assert(limit < _region_limit, "only way to get here");
3598 // This can happen under some pretty unusual circumstances. An
3599 // evacuation pause empties the region underneath our feet (NTAMS
3600 // at bottom). We then do some allocation in the region (NTAMS
3601 // stays at bottom), followed by the region being used as a GC
3602 // alloc region (NTAMS will move to top() and the objects
3603 // originally below it will be grayed). All objects now marked in
3604 // the region are explicitly grayed, if below the global finger,
3605 // and we do not need in fact to scan anything else. So, we simply
3606 // set _finger to be limit to ensure that the bitmap iteration
3607 // doesn't do anything.
3608 _finger = limit;
3609 }
3611 _region_limit = limit;
3612 }
3614 void CMTask::giveup_current_region() {
3615 assert(_curr_region != NULL, "invariant");
3616 if (_cm->verbose_low()) {
3617 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3618 _worker_id, p2i(_curr_region));
3619 }
3620 clear_region_fields();
3621 }
3623 void CMTask::clear_region_fields() {
3624 // Values for these three fields that indicate that we're not
3625 // holding on to a region.
3626 _curr_region = NULL;
3627 _finger = NULL;
3628 _region_limit = NULL;
3629 }
3631 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3632 if (cm_oop_closure == NULL) {
3633 assert(_cm_oop_closure != NULL, "invariant");
3634 } else {
3635 assert(_cm_oop_closure == NULL, "invariant");
3636 }
3637 _cm_oop_closure = cm_oop_closure;
3638 }
3640 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3641 guarantee(nextMarkBitMap != NULL, "invariant");
3643 if (_cm->verbose_low()) {
3644 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3645 }
3647 _nextMarkBitMap = nextMarkBitMap;
3648 clear_region_fields();
3650 _calls = 0;
3651 _elapsed_time_ms = 0.0;
3652 _termination_time_ms = 0.0;
3653 _termination_start_time_ms = 0.0;
3655 #if _MARKING_STATS_
3656 _local_pushes = 0;
3657 _local_pops = 0;
3658 _local_max_size = 0;
3659 _objs_scanned = 0;
3660 _global_pushes = 0;
3661 _global_pops = 0;
3662 _global_max_size = 0;
3663 _global_transfers_to = 0;
3664 _global_transfers_from = 0;
3665 _regions_claimed = 0;
3666 _objs_found_on_bitmap = 0;
3667 _satb_buffers_processed = 0;
3668 _steal_attempts = 0;
3669 _steals = 0;
3670 _aborted = 0;
3671 _aborted_overflow = 0;
3672 _aborted_cm_aborted = 0;
3673 _aborted_yield = 0;
3674 _aborted_timed_out = 0;
3675 _aborted_satb = 0;
3676 _aborted_termination = 0;
3677 #endif // _MARKING_STATS_
3678 }
3680 bool CMTask::should_exit_termination() {
3681 regular_clock_call();
3682 // This is called when we are in the termination protocol. We should
3683 // quit if, for some reason, this task wants to abort or the global
3684 // stack is not empty (this means that we can get work from it).
3685 return !_cm->mark_stack_empty() || has_aborted();
3686 }
3688 void CMTask::reached_limit() {
3689 assert(_words_scanned >= _words_scanned_limit ||
3690 _refs_reached >= _refs_reached_limit ,
3691 "shouldn't have been called otherwise");
3692 regular_clock_call();
3693 }
3695 void CMTask::regular_clock_call() {
3696 if (has_aborted()) return;
3698 // First, we need to recalculate the words scanned and refs reached
3699 // limits for the next clock call.
3700 recalculate_limits();
3702 // During the regular clock call we do the following
3704 // (1) If an overflow has been flagged, then we abort.
3705 if (_cm->has_overflown()) {
3706 set_has_aborted();
3707 return;
3708 }
3710 // If we are not concurrent (i.e. we're doing remark) we don't need
3711 // to check anything else. The other steps are only needed during
3712 // the concurrent marking phase.
3713 if (!concurrent()) return;
3715 // (2) If marking has been aborted for Full GC, then we also abort.
3716 if (_cm->has_aborted()) {
3717 set_has_aborted();
3718 statsOnly( ++_aborted_cm_aborted );
3719 return;
3720 }
3722 double curr_time_ms = os::elapsedVTime() * 1000.0;
3724 // (3) If marking stats are enabled, then we update the step history.
3725 #if _MARKING_STATS_
3726 if (_words_scanned >= _words_scanned_limit) {
3727 ++_clock_due_to_scanning;
3728 }
3729 if (_refs_reached >= _refs_reached_limit) {
3730 ++_clock_due_to_marking;
3731 }
3733 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3734 _interval_start_time_ms = curr_time_ms;
3735 _all_clock_intervals_ms.add(last_interval_ms);
3737 if (_cm->verbose_medium()) {
3738 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3739 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3740 _worker_id, last_interval_ms,
3741 _words_scanned,
3742 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3743 _refs_reached,
3744 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3745 }
3746 #endif // _MARKING_STATS_
3748 // (4) We check whether we should yield. If we have to, then we abort.
3749 if (SuspendibleThreadSet::should_yield()) {
3750 // We should yield. To do this we abort the task. The caller is
3751 // responsible for yielding.
3752 set_has_aborted();
3753 statsOnly( ++_aborted_yield );
3754 return;
3755 }
3757 // (5) We check whether we've reached our time quota. If we have,
3758 // then we abort.
3759 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3760 if (elapsed_time_ms > _time_target_ms) {
3761 set_has_aborted();
3762 _has_timed_out = true;
3763 statsOnly( ++_aborted_timed_out );
3764 return;
3765 }
3767 // (6) Finally, we check whether there are enough completed STAB
3768 // buffers available for processing. If there are, we abort.
3769 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3770 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3771 if (_cm->verbose_low()) {
3772 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3773 _worker_id);
3774 }
3775 // we do need to process SATB buffers, we'll abort and restart
3776 // the marking task to do so
3777 set_has_aborted();
3778 statsOnly( ++_aborted_satb );
3779 return;
3780 }
3781 }
3783 void CMTask::recalculate_limits() {
3784 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3785 _words_scanned_limit = _real_words_scanned_limit;
3787 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3788 _refs_reached_limit = _real_refs_reached_limit;
3789 }
3791 void CMTask::decrease_limits() {
3792 // This is called when we believe that we're going to do an infrequent
3793 // operation which will increase the per byte scanned cost (i.e. move
3794 // entries to/from the global stack). It basically tries to decrease the
3795 // scanning limit so that the clock is called earlier.
3797 if (_cm->verbose_medium()) {
3798 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3799 }
3801 _words_scanned_limit = _real_words_scanned_limit -
3802 3 * words_scanned_period / 4;
3803 _refs_reached_limit = _real_refs_reached_limit -
3804 3 * refs_reached_period / 4;
3805 }
3807 void CMTask::move_entries_to_global_stack() {
3808 // local array where we'll store the entries that will be popped
3809 // from the local queue
3810 oop buffer[global_stack_transfer_size];
3812 int n = 0;
3813 oop obj;
3814 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3815 buffer[n] = obj;
3816 ++n;
3817 }
3819 if (n > 0) {
3820 // we popped at least one entry from the local queue
3822 statsOnly( ++_global_transfers_to; _local_pops += n );
3824 if (!_cm->mark_stack_push(buffer, n)) {
3825 if (_cm->verbose_low()) {
3826 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3827 _worker_id);
3828 }
3829 set_has_aborted();
3830 } else {
3831 // the transfer was successful
3833 if (_cm->verbose_medium()) {
3834 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3835 _worker_id, n);
3836 }
3837 statsOnly( int tmp_size = _cm->mark_stack_size();
3838 if (tmp_size > _global_max_size) {
3839 _global_max_size = tmp_size;
3840 }
3841 _global_pushes += n );
3842 }
3843 }
3845 // this operation was quite expensive, so decrease the limits
3846 decrease_limits();
3847 }
3849 void CMTask::get_entries_from_global_stack() {
3850 // local array where we'll store the entries that will be popped
3851 // from the global stack.
3852 oop buffer[global_stack_transfer_size];
3853 int n;
3854 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3855 assert(n <= global_stack_transfer_size,
3856 "we should not pop more than the given limit");
3857 if (n > 0) {
3858 // yes, we did actually pop at least one entry
3860 statsOnly( ++_global_transfers_from; _global_pops += n );
3861 if (_cm->verbose_medium()) {
3862 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3863 _worker_id, n);
3864 }
3865 for (int i = 0; i < n; ++i) {
3866 bool success = _task_queue->push(buffer[i]);
3867 // We only call this when the local queue is empty or under a
3868 // given target limit. So, we do not expect this push to fail.
3869 assert(success, "invariant");
3870 }
3872 statsOnly( int tmp_size = _task_queue->size();
3873 if (tmp_size > _local_max_size) {
3874 _local_max_size = tmp_size;
3875 }
3876 _local_pushes += n );
3877 }
3879 // this operation was quite expensive, so decrease the limits
3880 decrease_limits();
3881 }
3883 void CMTask::drain_local_queue(bool partially) {
3884 if (has_aborted()) return;
3886 // Decide what the target size is, depending whether we're going to
3887 // drain it partially (so that other tasks can steal if they run out
3888 // of things to do) or totally (at the very end).
3889 size_t target_size;
3890 if (partially) {
3891 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3892 } else {
3893 target_size = 0;
3894 }
3896 if (_task_queue->size() > target_size) {
3897 if (_cm->verbose_high()) {
3898 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3899 _worker_id, target_size);
3900 }
3902 oop obj;
3903 bool ret = _task_queue->pop_local(obj);
3904 while (ret) {
3905 statsOnly( ++_local_pops );
3907 if (_cm->verbose_high()) {
3908 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3909 p2i((void*) obj));
3910 }
3912 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3913 assert(!_g1h->is_on_master_free_list(
3914 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3916 scan_object(obj);
3918 if (_task_queue->size() <= target_size || has_aborted()) {
3919 ret = false;
3920 } else {
3921 ret = _task_queue->pop_local(obj);
3922 }
3923 }
3925 if (_cm->verbose_high()) {
3926 gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3927 _worker_id, _task_queue->size());
3928 }
3929 }
3930 }
3932 void CMTask::drain_global_stack(bool partially) {
3933 if (has_aborted()) return;
3935 // We have a policy to drain the local queue before we attempt to
3936 // drain the global stack.
3937 assert(partially || _task_queue->size() == 0, "invariant");
3939 // Decide what the target size is, depending whether we're going to
3940 // drain it partially (so that other tasks can steal if they run out
3941 // of things to do) or totally (at the very end). Notice that,
3942 // because we move entries from the global stack in chunks or
3943 // because another task might be doing the same, we might in fact
3944 // drop below the target. But, this is not a problem.
3945 size_t target_size;
3946 if (partially) {
3947 target_size = _cm->partial_mark_stack_size_target();
3948 } else {
3949 target_size = 0;
3950 }
3952 if (_cm->mark_stack_size() > target_size) {
3953 if (_cm->verbose_low()) {
3954 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3955 _worker_id, target_size);
3956 }
3958 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3959 get_entries_from_global_stack();
3960 drain_local_queue(partially);
3961 }
3963 if (_cm->verbose_low()) {
3964 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3965 _worker_id, _cm->mark_stack_size());
3966 }
3967 }
3968 }
3970 // SATB Queue has several assumptions on whether to call the par or
3971 // non-par versions of the methods. this is why some of the code is
3972 // replicated. We should really get rid of the single-threaded version
3973 // of the code to simplify things.
3974 void CMTask::drain_satb_buffers() {
3975 if (has_aborted()) return;
3977 // We set this so that the regular clock knows that we're in the
3978 // middle of draining buffers and doesn't set the abort flag when it
3979 // notices that SATB buffers are available for draining. It'd be
3980 // very counter productive if it did that. :-)
3981 _draining_satb_buffers = true;
3983 CMObjectClosure oc(this);
3984 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3985 if (G1CollectedHeap::use_parallel_gc_threads()) {
3986 satb_mq_set.set_par_closure(_worker_id, &oc);
3987 } else {
3988 satb_mq_set.set_closure(&oc);
3989 }
3991 // This keeps claiming and applying the closure to completed buffers
3992 // until we run out of buffers or we need to abort.
3993 if (G1CollectedHeap::use_parallel_gc_threads()) {
3994 while (!has_aborted() &&
3995 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
3996 if (_cm->verbose_medium()) {
3997 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3998 }
3999 statsOnly( ++_satb_buffers_processed );
4000 regular_clock_call();
4001 }
4002 } else {
4003 while (!has_aborted() &&
4004 satb_mq_set.apply_closure_to_completed_buffer()) {
4005 if (_cm->verbose_medium()) {
4006 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4007 }
4008 statsOnly( ++_satb_buffers_processed );
4009 regular_clock_call();
4010 }
4011 }
4013 _draining_satb_buffers = false;
4015 assert(has_aborted() ||
4016 concurrent() ||
4017 satb_mq_set.completed_buffers_num() == 0, "invariant");
4019 if (G1CollectedHeap::use_parallel_gc_threads()) {
4020 satb_mq_set.set_par_closure(_worker_id, NULL);
4021 } else {
4022 satb_mq_set.set_closure(NULL);
4023 }
4025 // again, this was a potentially expensive operation, decrease the
4026 // limits to get the regular clock call early
4027 decrease_limits();
4028 }
4030 void CMTask::print_stats() {
4031 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
4032 _worker_id, _calls);
4033 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
4034 _elapsed_time_ms, _termination_time_ms);
4035 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4036 _step_times_ms.num(), _step_times_ms.avg(),
4037 _step_times_ms.sd());
4038 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4039 _step_times_ms.maximum(), _step_times_ms.sum());
4041 #if _MARKING_STATS_
4042 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4043 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
4044 _all_clock_intervals_ms.sd());
4045 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4046 _all_clock_intervals_ms.maximum(),
4047 _all_clock_intervals_ms.sum());
4048 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
4049 _clock_due_to_scanning, _clock_due_to_marking);
4050 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
4051 _objs_scanned, _objs_found_on_bitmap);
4052 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
4053 _local_pushes, _local_pops, _local_max_size);
4054 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
4055 _global_pushes, _global_pops, _global_max_size);
4056 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
4057 _global_transfers_to,_global_transfers_from);
4058 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
4059 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
4060 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
4061 _steal_attempts, _steals);
4062 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
4063 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
4064 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
4065 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
4066 _aborted_timed_out, _aborted_satb, _aborted_termination);
4067 #endif // _MARKING_STATS_
4068 }
4070 /*****************************************************************************
4072 The do_marking_step(time_target_ms, ...) method is the building
4073 block of the parallel marking framework. It can be called in parallel
4074 with other invocations of do_marking_step() on different tasks
4075 (but only one per task, obviously) and concurrently with the
4076 mutator threads, or during remark, hence it eliminates the need
4077 for two versions of the code. When called during remark, it will
4078 pick up from where the task left off during the concurrent marking
4079 phase. Interestingly, tasks are also claimable during evacuation
4080 pauses too, since do_marking_step() ensures that it aborts before
4081 it needs to yield.
4083 The data structures that it uses to do marking work are the
4084 following:
4086 (1) Marking Bitmap. If there are gray objects that appear only
4087 on the bitmap (this happens either when dealing with an overflow
4088 or when the initial marking phase has simply marked the roots
4089 and didn't push them on the stack), then tasks claim heap
4090 regions whose bitmap they then scan to find gray objects. A
4091 global finger indicates where the end of the last claimed region
4092 is. A local finger indicates how far into the region a task has
4093 scanned. The two fingers are used to determine how to gray an
4094 object (i.e. whether simply marking it is OK, as it will be
4095 visited by a task in the future, or whether it needs to be also
4096 pushed on a stack).
4098 (2) Local Queue. The local queue of the task which is accessed
4099 reasonably efficiently by the task. Other tasks can steal from
4100 it when they run out of work. Throughout the marking phase, a
4101 task attempts to keep its local queue short but not totally
4102 empty, so that entries are available for stealing by other
4103 tasks. Only when there is no more work, a task will totally
4104 drain its local queue.
4106 (3) Global Mark Stack. This handles local queue overflow. During
4107 marking only sets of entries are moved between it and the local
4108 queues, as access to it requires a mutex and more fine-grain
4109 interaction with it which might cause contention. If it
4110 overflows, then the marking phase should restart and iterate
4111 over the bitmap to identify gray objects. Throughout the marking
4112 phase, tasks attempt to keep the global mark stack at a small
4113 length but not totally empty, so that entries are available for
4114 popping by other tasks. Only when there is no more work, tasks
4115 will totally drain the global mark stack.
4117 (4) SATB Buffer Queue. This is where completed SATB buffers are
4118 made available. Buffers are regularly removed from this queue
4119 and scanned for roots, so that the queue doesn't get too
4120 long. During remark, all completed buffers are processed, as
4121 well as the filled in parts of any uncompleted buffers.
4123 The do_marking_step() method tries to abort when the time target
4124 has been reached. There are a few other cases when the
4125 do_marking_step() method also aborts:
4127 (1) When the marking phase has been aborted (after a Full GC).
4129 (2) When a global overflow (on the global stack) has been
4130 triggered. Before the task aborts, it will actually sync up with
4131 the other tasks to ensure that all the marking data structures
4132 (local queues, stacks, fingers etc.) are re-initialized so that
4133 when do_marking_step() completes, the marking phase can
4134 immediately restart.
4136 (3) When enough completed SATB buffers are available. The
4137 do_marking_step() method only tries to drain SATB buffers right
4138 at the beginning. So, if enough buffers are available, the
4139 marking step aborts and the SATB buffers are processed at
4140 the beginning of the next invocation.
4142 (4) To yield. when we have to yield then we abort and yield
4143 right at the end of do_marking_step(). This saves us from a lot
4144 of hassle as, by yielding we might allow a Full GC. If this
4145 happens then objects will be compacted underneath our feet, the
4146 heap might shrink, etc. We save checking for this by just
4147 aborting and doing the yield right at the end.
4149 From the above it follows that the do_marking_step() method should
4150 be called in a loop (or, otherwise, regularly) until it completes.
4152 If a marking step completes without its has_aborted() flag being
4153 true, it means it has completed the current marking phase (and
4154 also all other marking tasks have done so and have all synced up).
4156 A method called regular_clock_call() is invoked "regularly" (in
4157 sub ms intervals) throughout marking. It is this clock method that
4158 checks all the abort conditions which were mentioned above and
4159 decides when the task should abort. A work-based scheme is used to
4160 trigger this clock method: when the number of object words the
4161 marking phase has scanned or the number of references the marking
4162 phase has visited reach a given limit. Additional invocations to
4163 the method clock have been planted in a few other strategic places
4164 too. The initial reason for the clock method was to avoid calling
4165 vtime too regularly, as it is quite expensive. So, once it was in
4166 place, it was natural to piggy-back all the other conditions on it
4167 too and not constantly check them throughout the code.
4169 If do_termination is true then do_marking_step will enter its
4170 termination protocol.
4172 The value of is_serial must be true when do_marking_step is being
4173 called serially (i.e. by the VMThread) and do_marking_step should
4174 skip any synchronization in the termination and overflow code.
4175 Examples include the serial remark code and the serial reference
4176 processing closures.
4178 The value of is_serial must be false when do_marking_step is
4179 being called by any of the worker threads in a work gang.
4180 Examples include the concurrent marking code (CMMarkingTask),
4181 the MT remark code, and the MT reference processing closures.
4183 *****************************************************************************/
4185 void CMTask::do_marking_step(double time_target_ms,
4186 bool do_termination,
4187 bool is_serial) {
4188 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4189 assert(concurrent() == _cm->concurrent(), "they should be the same");
4191 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4192 assert(_task_queues != NULL, "invariant");
4193 assert(_task_queue != NULL, "invariant");
4194 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4196 assert(!_claimed,
4197 "only one thread should claim this task at any one time");
4199 // OK, this doesn't safeguard again all possible scenarios, as it is
4200 // possible for two threads to set the _claimed flag at the same
4201 // time. But it is only for debugging purposes anyway and it will
4202 // catch most problems.
4203 _claimed = true;
4205 _start_time_ms = os::elapsedVTime() * 1000.0;
4206 statsOnly( _interval_start_time_ms = _start_time_ms );
4208 // If do_stealing is true then do_marking_step will attempt to
4209 // steal work from the other CMTasks. It only makes sense to
4210 // enable stealing when the termination protocol is enabled
4211 // and do_marking_step() is not being called serially.
4212 bool do_stealing = do_termination && !is_serial;
4214 double diff_prediction_ms =
4215 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4216 _time_target_ms = time_target_ms - diff_prediction_ms;
4218 // set up the variables that are used in the work-based scheme to
4219 // call the regular clock method
4220 _words_scanned = 0;
4221 _refs_reached = 0;
4222 recalculate_limits();
4224 // clear all flags
4225 clear_has_aborted();
4226 _has_timed_out = false;
4227 _draining_satb_buffers = false;
4229 ++_calls;
4231 if (_cm->verbose_low()) {
4232 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4233 "target = %1.2lfms >>>>>>>>>>",
4234 _worker_id, _calls, _time_target_ms);
4235 }
4237 // Set up the bitmap and oop closures. Anything that uses them is
4238 // eventually called from this method, so it is OK to allocate these
4239 // statically.
4240 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4241 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4242 set_cm_oop_closure(&cm_oop_closure);
4244 if (_cm->has_overflown()) {
4245 // This can happen if the mark stack overflows during a GC pause
4246 // and this task, after a yield point, restarts. We have to abort
4247 // as we need to get into the overflow protocol which happens
4248 // right at the end of this task.
4249 set_has_aborted();
4250 }
4252 // First drain any available SATB buffers. After this, we will not
4253 // look at SATB buffers before the next invocation of this method.
4254 // If enough completed SATB buffers are queued up, the regular clock
4255 // will abort this task so that it restarts.
4256 drain_satb_buffers();
4257 // ...then partially drain the local queue and the global stack
4258 drain_local_queue(true);
4259 drain_global_stack(true);
4261 do {
4262 if (!has_aborted() && _curr_region != NULL) {
4263 // This means that we're already holding on to a region.
4264 assert(_finger != NULL, "if region is not NULL, then the finger "
4265 "should not be NULL either");
4267 // We might have restarted this task after an evacuation pause
4268 // which might have evacuated the region we're holding on to
4269 // underneath our feet. Let's read its limit again to make sure
4270 // that we do not iterate over a region of the heap that
4271 // contains garbage (update_region_limit() will also move
4272 // _finger to the start of the region if it is found empty).
4273 update_region_limit();
4274 // We will start from _finger not from the start of the region,
4275 // as we might be restarting this task after aborting half-way
4276 // through scanning this region. In this case, _finger points to
4277 // the address where we last found a marked object. If this is a
4278 // fresh region, _finger points to start().
4279 MemRegion mr = MemRegion(_finger, _region_limit);
4281 if (_cm->verbose_low()) {
4282 gclog_or_tty->print_cr("[%u] we're scanning part "
4283 "["PTR_FORMAT", "PTR_FORMAT") "
4284 "of region "HR_FORMAT,
4285 _worker_id, p2i(_finger), p2i(_region_limit),
4286 HR_FORMAT_PARAMS(_curr_region));
4287 }
4289 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4290 "humongous regions should go around loop once only");
4292 // Some special cases:
4293 // If the memory region is empty, we can just give up the region.
4294 // If the current region is humongous then we only need to check
4295 // the bitmap for the bit associated with the start of the object,
4296 // scan the object if it's live, and give up the region.
4297 // Otherwise, let's iterate over the bitmap of the part of the region
4298 // that is left.
4299 // If the iteration is successful, give up the region.
4300 if (mr.is_empty()) {
4301 giveup_current_region();
4302 regular_clock_call();
4303 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4304 if (_nextMarkBitMap->isMarked(mr.start())) {
4305 // The object is marked - apply the closure
4306 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4307 bitmap_closure.do_bit(offset);
4308 }
4309 // Even if this task aborted while scanning the humongous object
4310 // we can (and should) give up the current region.
4311 giveup_current_region();
4312 regular_clock_call();
4313 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4314 giveup_current_region();
4315 regular_clock_call();
4316 } else {
4317 assert(has_aborted(), "currently the only way to do so");
4318 // The only way to abort the bitmap iteration is to return
4319 // false from the do_bit() method. However, inside the
4320 // do_bit() method we move the _finger to point to the
4321 // object currently being looked at. So, if we bail out, we
4322 // have definitely set _finger to something non-null.
4323 assert(_finger != NULL, "invariant");
4325 // Region iteration was actually aborted. So now _finger
4326 // points to the address of the object we last scanned. If we
4327 // leave it there, when we restart this task, we will rescan
4328 // the object. It is easy to avoid this. We move the finger by
4329 // enough to point to the next possible object header (the
4330 // bitmap knows by how much we need to move it as it knows its
4331 // granularity).
4332 assert(_finger < _region_limit, "invariant");
4333 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4334 // Check if bitmap iteration was aborted while scanning the last object
4335 if (new_finger >= _region_limit) {
4336 giveup_current_region();
4337 } else {
4338 move_finger_to(new_finger);
4339 }
4340 }
4341 }
4342 // At this point we have either completed iterating over the
4343 // region we were holding on to, or we have aborted.
4345 // We then partially drain the local queue and the global stack.
4346 // (Do we really need this?)
4347 drain_local_queue(true);
4348 drain_global_stack(true);
4350 // Read the note on the claim_region() method on why it might
4351 // return NULL with potentially more regions available for
4352 // claiming and why we have to check out_of_regions() to determine
4353 // whether we're done or not.
4354 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4355 // We are going to try to claim a new region. We should have
4356 // given up on the previous one.
4357 // Separated the asserts so that we know which one fires.
4358 assert(_curr_region == NULL, "invariant");
4359 assert(_finger == NULL, "invariant");
4360 assert(_region_limit == NULL, "invariant");
4361 if (_cm->verbose_low()) {
4362 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4363 }
4364 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4365 if (claimed_region != NULL) {
4366 // Yes, we managed to claim one
4367 statsOnly( ++_regions_claimed );
4369 if (_cm->verbose_low()) {
4370 gclog_or_tty->print_cr("[%u] we successfully claimed "
4371 "region "PTR_FORMAT,
4372 _worker_id, p2i(claimed_region));
4373 }
4375 setup_for_region(claimed_region);
4376 assert(_curr_region == claimed_region, "invariant");
4377 }
4378 // It is important to call the regular clock here. It might take
4379 // a while to claim a region if, for example, we hit a large
4380 // block of empty regions. So we need to call the regular clock
4381 // method once round the loop to make sure it's called
4382 // frequently enough.
4383 regular_clock_call();
4384 }
4386 if (!has_aborted() && _curr_region == NULL) {
4387 assert(_cm->out_of_regions(),
4388 "at this point we should be out of regions");
4389 }
4390 } while ( _curr_region != NULL && !has_aborted());
4392 if (!has_aborted()) {
4393 // We cannot check whether the global stack is empty, since other
4394 // tasks might be pushing objects to it concurrently.
4395 assert(_cm->out_of_regions(),
4396 "at this point we should be out of regions");
4398 if (_cm->verbose_low()) {
4399 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4400 }
4402 // Try to reduce the number of available SATB buffers so that
4403 // remark has less work to do.
4404 drain_satb_buffers();
4405 }
4407 // Since we've done everything else, we can now totally drain the
4408 // local queue and global stack.
4409 drain_local_queue(false);
4410 drain_global_stack(false);
4412 // Attempt at work stealing from other task's queues.
4413 if (do_stealing && !has_aborted()) {
4414 // We have not aborted. This means that we have finished all that
4415 // we could. Let's try to do some stealing...
4417 // We cannot check whether the global stack is empty, since other
4418 // tasks might be pushing objects to it concurrently.
4419 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4420 "only way to reach here");
4422 if (_cm->verbose_low()) {
4423 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4424 }
4426 while (!has_aborted()) {
4427 oop obj;
4428 statsOnly( ++_steal_attempts );
4430 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4431 if (_cm->verbose_medium()) {
4432 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4433 _worker_id, p2i((void*) obj));
4434 }
4436 statsOnly( ++_steals );
4438 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4439 "any stolen object should be marked");
4440 scan_object(obj);
4442 // And since we're towards the end, let's totally drain the
4443 // local queue and global stack.
4444 drain_local_queue(false);
4445 drain_global_stack(false);
4446 } else {
4447 break;
4448 }
4449 }
4450 }
4452 // If we are about to wrap up and go into termination, check if we
4453 // should raise the overflow flag.
4454 if (do_termination && !has_aborted()) {
4455 if (_cm->force_overflow()->should_force()) {
4456 _cm->set_has_overflown();
4457 regular_clock_call();
4458 }
4459 }
4461 // We still haven't aborted. Now, let's try to get into the
4462 // termination protocol.
4463 if (do_termination && !has_aborted()) {
4464 // We cannot check whether the global stack is empty, since other
4465 // tasks might be concurrently pushing objects on it.
4466 // Separated the asserts so that we know which one fires.
4467 assert(_cm->out_of_regions(), "only way to reach here");
4468 assert(_task_queue->size() == 0, "only way to reach here");
4470 if (_cm->verbose_low()) {
4471 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4472 }
4474 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4476 // The CMTask class also extends the TerminatorTerminator class,
4477 // hence its should_exit_termination() method will also decide
4478 // whether to exit the termination protocol or not.
4479 bool finished = (is_serial ||
4480 _cm->terminator()->offer_termination(this));
4481 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4482 _termination_time_ms +=
4483 termination_end_time_ms - _termination_start_time_ms;
4485 if (finished) {
4486 // We're all done.
4488 if (_worker_id == 0) {
4489 // let's allow task 0 to do this
4490 if (concurrent()) {
4491 assert(_cm->concurrent_marking_in_progress(), "invariant");
4492 // we need to set this to false before the next
4493 // safepoint. This way we ensure that the marking phase
4494 // doesn't observe any more heap expansions.
4495 _cm->clear_concurrent_marking_in_progress();
4496 }
4497 }
4499 // We can now guarantee that the global stack is empty, since
4500 // all other tasks have finished. We separated the guarantees so
4501 // that, if a condition is false, we can immediately find out
4502 // which one.
4503 guarantee(_cm->out_of_regions(), "only way to reach here");
4504 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4505 guarantee(_task_queue->size() == 0, "only way to reach here");
4506 guarantee(!_cm->has_overflown(), "only way to reach here");
4507 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4509 if (_cm->verbose_low()) {
4510 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4511 }
4512 } else {
4513 // Apparently there's more work to do. Let's abort this task. It
4514 // will restart it and we can hopefully find more things to do.
4516 if (_cm->verbose_low()) {
4517 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4518 _worker_id);
4519 }
4521 set_has_aborted();
4522 statsOnly( ++_aborted_termination );
4523 }
4524 }
4526 // Mainly for debugging purposes to make sure that a pointer to the
4527 // closure which was statically allocated in this frame doesn't
4528 // escape it by accident.
4529 set_cm_oop_closure(NULL);
4530 double end_time_ms = os::elapsedVTime() * 1000.0;
4531 double elapsed_time_ms = end_time_ms - _start_time_ms;
4532 // Update the step history.
4533 _step_times_ms.add(elapsed_time_ms);
4535 if (has_aborted()) {
4536 // The task was aborted for some reason.
4538 statsOnly( ++_aborted );
4540 if (_has_timed_out) {
4541 double diff_ms = elapsed_time_ms - _time_target_ms;
4542 // Keep statistics of how well we did with respect to hitting
4543 // our target only if we actually timed out (if we aborted for
4544 // other reasons, then the results might get skewed).
4545 _marking_step_diffs_ms.add(diff_ms);
4546 }
4548 if (_cm->has_overflown()) {
4549 // This is the interesting one. We aborted because a global
4550 // overflow was raised. This means we have to restart the
4551 // marking phase and start iterating over regions. However, in
4552 // order to do this we have to make sure that all tasks stop
4553 // what they are doing and re-initialise in a safe manner. We
4554 // will achieve this with the use of two barrier sync points.
4556 if (_cm->verbose_low()) {
4557 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4558 }
4560 if (!is_serial) {
4561 // We only need to enter the sync barrier if being called
4562 // from a parallel context
4563 _cm->enter_first_sync_barrier(_worker_id);
4565 // When we exit this sync barrier we know that all tasks have
4566 // stopped doing marking work. So, it's now safe to
4567 // re-initialise our data structures. At the end of this method,
4568 // task 0 will clear the global data structures.
4569 }
4571 statsOnly( ++_aborted_overflow );
4573 // We clear the local state of this task...
4574 clear_region_fields();
4576 if (!is_serial) {
4577 // ...and enter the second barrier.
4578 _cm->enter_second_sync_barrier(_worker_id);
4579 }
4580 // At this point, if we're during the concurrent phase of
4581 // marking, everything has been re-initialized and we're
4582 // ready to restart.
4583 }
4585 if (_cm->verbose_low()) {
4586 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4587 "elapsed = %1.2lfms <<<<<<<<<<",
4588 _worker_id, _time_target_ms, elapsed_time_ms);
4589 if (_cm->has_aborted()) {
4590 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4591 _worker_id);
4592 }
4593 }
4594 } else {
4595 if (_cm->verbose_low()) {
4596 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4597 "elapsed = %1.2lfms <<<<<<<<<<",
4598 _worker_id, _time_target_ms, elapsed_time_ms);
4599 }
4600 }
4602 _claimed = false;
4603 }
4605 CMTask::CMTask(uint worker_id,
4606 ConcurrentMark* cm,
4607 size_t* marked_bytes,
4608 BitMap* card_bm,
4609 CMTaskQueue* task_queue,
4610 CMTaskQueueSet* task_queues)
4611 : _g1h(G1CollectedHeap::heap()),
4612 _worker_id(worker_id), _cm(cm),
4613 _claimed(false),
4614 _nextMarkBitMap(NULL), _hash_seed(17),
4615 _task_queue(task_queue),
4616 _task_queues(task_queues),
4617 _cm_oop_closure(NULL),
4618 _marked_bytes_array(marked_bytes),
4619 _card_bm(card_bm) {
4620 guarantee(task_queue != NULL, "invariant");
4621 guarantee(task_queues != NULL, "invariant");
4623 statsOnly( _clock_due_to_scanning = 0;
4624 _clock_due_to_marking = 0 );
4626 _marking_step_diffs_ms.add(0.5);
4627 }
4629 // These are formatting macros that are used below to ensure
4630 // consistent formatting. The *_H_* versions are used to format the
4631 // header for a particular value and they should be kept consistent
4632 // with the corresponding macro. Also note that most of the macros add
4633 // the necessary white space (as a prefix) which makes them a bit
4634 // easier to compose.
4636 // All the output lines are prefixed with this string to be able to
4637 // identify them easily in a large log file.
4638 #define G1PPRL_LINE_PREFIX "###"
4640 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4641 #ifdef _LP64
4642 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4643 #else // _LP64
4644 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4645 #endif // _LP64
4647 // For per-region info
4648 #define G1PPRL_TYPE_FORMAT " %-4s"
4649 #define G1PPRL_TYPE_H_FORMAT " %4s"
4650 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4651 #define G1PPRL_BYTE_H_FORMAT " %9s"
4652 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4653 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4655 // For summary info
4656 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4657 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4658 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4659 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4661 G1PrintRegionLivenessInfoClosure::
4662 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4663 : _out(out),
4664 _total_used_bytes(0), _total_capacity_bytes(0),
4665 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4666 _hum_used_bytes(0), _hum_capacity_bytes(0),
4667 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4668 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4669 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4670 MemRegion g1_reserved = g1h->g1_reserved();
4671 double now = os::elapsedTime();
4673 // Print the header of the output.
4674 _out->cr();
4675 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4676 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4677 G1PPRL_SUM_ADDR_FORMAT("reserved")
4678 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4679 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4680 HeapRegion::GrainBytes);
4681 _out->print_cr(G1PPRL_LINE_PREFIX);
4682 _out->print_cr(G1PPRL_LINE_PREFIX
4683 G1PPRL_TYPE_H_FORMAT
4684 G1PPRL_ADDR_BASE_H_FORMAT
4685 G1PPRL_BYTE_H_FORMAT
4686 G1PPRL_BYTE_H_FORMAT
4687 G1PPRL_BYTE_H_FORMAT
4688 G1PPRL_DOUBLE_H_FORMAT
4689 G1PPRL_BYTE_H_FORMAT
4690 G1PPRL_BYTE_H_FORMAT,
4691 "type", "address-range",
4692 "used", "prev-live", "next-live", "gc-eff",
4693 "remset", "code-roots");
4694 _out->print_cr(G1PPRL_LINE_PREFIX
4695 G1PPRL_TYPE_H_FORMAT
4696 G1PPRL_ADDR_BASE_H_FORMAT
4697 G1PPRL_BYTE_H_FORMAT
4698 G1PPRL_BYTE_H_FORMAT
4699 G1PPRL_BYTE_H_FORMAT
4700 G1PPRL_DOUBLE_H_FORMAT
4701 G1PPRL_BYTE_H_FORMAT
4702 G1PPRL_BYTE_H_FORMAT,
4703 "", "",
4704 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4705 "(bytes)", "(bytes)");
4706 }
4708 // It takes as a parameter a reference to one of the _hum_* fields, it
4709 // deduces the corresponding value for a region in a humongous region
4710 // series (either the region size, or what's left if the _hum_* field
4711 // is < the region size), and updates the _hum_* field accordingly.
4712 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4713 size_t bytes = 0;
4714 // The > 0 check is to deal with the prev and next live bytes which
4715 // could be 0.
4716 if (*hum_bytes > 0) {
4717 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4718 *hum_bytes -= bytes;
4719 }
4720 return bytes;
4721 }
4723 // It deduces the values for a region in a humongous region series
4724 // from the _hum_* fields and updates those accordingly. It assumes
4725 // that that _hum_* fields have already been set up from the "starts
4726 // humongous" region and we visit the regions in address order.
4727 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4728 size_t* capacity_bytes,
4729 size_t* prev_live_bytes,
4730 size_t* next_live_bytes) {
4731 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4732 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4733 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4734 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4735 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4736 }
4738 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4739 const char* type = "";
4740 HeapWord* bottom = r->bottom();
4741 HeapWord* end = r->end();
4742 size_t capacity_bytes = r->capacity();
4743 size_t used_bytes = r->used();
4744 size_t prev_live_bytes = r->live_bytes();
4745 size_t next_live_bytes = r->next_live_bytes();
4746 double gc_eff = r->gc_efficiency();
4747 size_t remset_bytes = r->rem_set()->mem_size();
4748 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4750 if (r->used() == 0) {
4751 type = "FREE";
4752 } else if (r->is_survivor()) {
4753 type = "SURV";
4754 } else if (r->is_young()) {
4755 type = "EDEN";
4756 } else if (r->startsHumongous()) {
4757 type = "HUMS";
4759 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4760 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4761 "they should have been zeroed after the last time we used them");
4762 // Set up the _hum_* fields.
4763 _hum_capacity_bytes = capacity_bytes;
4764 _hum_used_bytes = used_bytes;
4765 _hum_prev_live_bytes = prev_live_bytes;
4766 _hum_next_live_bytes = next_live_bytes;
4767 get_hum_bytes(&used_bytes, &capacity_bytes,
4768 &prev_live_bytes, &next_live_bytes);
4769 end = bottom + HeapRegion::GrainWords;
4770 } else if (r->continuesHumongous()) {
4771 type = "HUMC";
4772 get_hum_bytes(&used_bytes, &capacity_bytes,
4773 &prev_live_bytes, &next_live_bytes);
4774 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4775 } else {
4776 type = "OLD";
4777 }
4779 _total_used_bytes += used_bytes;
4780 _total_capacity_bytes += capacity_bytes;
4781 _total_prev_live_bytes += prev_live_bytes;
4782 _total_next_live_bytes += next_live_bytes;
4783 _total_remset_bytes += remset_bytes;
4784 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4786 // Print a line for this particular region.
4787 _out->print_cr(G1PPRL_LINE_PREFIX
4788 G1PPRL_TYPE_FORMAT
4789 G1PPRL_ADDR_BASE_FORMAT
4790 G1PPRL_BYTE_FORMAT
4791 G1PPRL_BYTE_FORMAT
4792 G1PPRL_BYTE_FORMAT
4793 G1PPRL_DOUBLE_FORMAT
4794 G1PPRL_BYTE_FORMAT
4795 G1PPRL_BYTE_FORMAT,
4796 type, p2i(bottom), p2i(end),
4797 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4798 remset_bytes, strong_code_roots_bytes);
4800 return false;
4801 }
4803 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4804 // add static memory usages to remembered set sizes
4805 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4806 // Print the footer of the output.
4807 _out->print_cr(G1PPRL_LINE_PREFIX);
4808 _out->print_cr(G1PPRL_LINE_PREFIX
4809 " SUMMARY"
4810 G1PPRL_SUM_MB_FORMAT("capacity")
4811 G1PPRL_SUM_MB_PERC_FORMAT("used")
4812 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4813 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4814 G1PPRL_SUM_MB_FORMAT("remset")
4815 G1PPRL_SUM_MB_FORMAT("code-roots"),
4816 bytes_to_mb(_total_capacity_bytes),
4817 bytes_to_mb(_total_used_bytes),
4818 perc(_total_used_bytes, _total_capacity_bytes),
4819 bytes_to_mb(_total_prev_live_bytes),
4820 perc(_total_prev_live_bytes, _total_capacity_bytes),
4821 bytes_to_mb(_total_next_live_bytes),
4822 perc(_total_next_live_bytes, _total_capacity_bytes),
4823 bytes_to_mb(_total_remset_bytes),
4824 bytes_to_mb(_total_strong_code_roots_bytes));
4825 _out->cr();
4826 }