Thu, 23 Oct 2014 12:02:08 -0700
Merge
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, bool zero_filled) {
134 if (zero_filled) {
135 return;
136 }
137 // We need to clear the bitmap on commit, removing any existing information.
138 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
139 _bm->clearRange(mr);
140 }
142 // Closure used for clearing the given mark bitmap.
143 class ClearBitmapHRClosure : public HeapRegionClosure {
144 private:
145 ConcurrentMark* _cm;
146 CMBitMap* _bitmap;
147 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration.
148 public:
149 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
150 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
151 }
153 virtual bool doHeapRegion(HeapRegion* r) {
154 size_t const chunk_size_in_words = M / HeapWordSize;
156 HeapWord* cur = r->bottom();
157 HeapWord* const end = r->end();
159 while (cur < end) {
160 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
161 _bitmap->clearRange(mr);
163 cur += chunk_size_in_words;
165 // Abort iteration if after yielding the marking has been aborted.
166 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
167 return true;
168 }
169 // Repeat the asserts from before the start of the closure. We will do them
170 // as asserts here to minimize their overhead on the product. However, we
171 // will have them as guarantees at the beginning / end of the bitmap
172 // clearing to get some checking in the product.
173 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
174 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
175 }
177 return false;
178 }
179 };
181 void CMBitMap::clearAll() {
182 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
183 G1CollectedHeap::heap()->heap_region_iterate(&cl);
184 guarantee(cl.complete(), "Must have completed iteration.");
185 return;
186 }
188 void CMBitMap::markRange(MemRegion mr) {
189 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
190 assert(!mr.is_empty(), "unexpected empty region");
191 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
192 ((HeapWord *) mr.end())),
193 "markRange memory region end is not card aligned");
194 // convert address range into offset range
195 _bm.at_put_range(heapWordToOffset(mr.start()),
196 heapWordToOffset(mr.end()), true);
197 }
199 void CMBitMap::clearRange(MemRegion mr) {
200 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
201 assert(!mr.is_empty(), "unexpected empty region");
202 // convert address range into offset range
203 _bm.at_put_range(heapWordToOffset(mr.start()),
204 heapWordToOffset(mr.end()), false);
205 }
207 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
208 HeapWord* end_addr) {
209 HeapWord* start = getNextMarkedWordAddress(addr);
210 start = MIN2(start, end_addr);
211 HeapWord* end = getNextUnmarkedWordAddress(start);
212 end = MIN2(end, end_addr);
213 assert(start <= end, "Consistency check");
214 MemRegion mr(start, end);
215 if (!mr.is_empty()) {
216 clearRange(mr);
217 }
218 return mr;
219 }
221 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
222 _base(NULL), _cm(cm)
223 #ifdef ASSERT
224 , _drain_in_progress(false)
225 , _drain_in_progress_yields(false)
226 #endif
227 {}
229 bool CMMarkStack::allocate(size_t capacity) {
230 // allocate a stack of the requisite depth
231 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
232 if (!rs.is_reserved()) {
233 warning("ConcurrentMark MarkStack allocation failure");
234 return false;
235 }
236 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
237 if (!_virtual_space.initialize(rs, rs.size())) {
238 warning("ConcurrentMark MarkStack backing store failure");
239 // Release the virtual memory reserved for the marking stack
240 rs.release();
241 return false;
242 }
243 assert(_virtual_space.committed_size() == rs.size(),
244 "Didn't reserve backing store for all of ConcurrentMark stack?");
245 _base = (oop*) _virtual_space.low();
246 setEmpty();
247 _capacity = (jint) capacity;
248 _saved_index = -1;
249 _should_expand = false;
250 NOT_PRODUCT(_max_depth = 0);
251 return true;
252 }
254 void CMMarkStack::expand() {
255 // Called, during remark, if we've overflown the marking stack during marking.
256 assert(isEmpty(), "stack should been emptied while handling overflow");
257 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
258 // Clear expansion flag
259 _should_expand = false;
260 if (_capacity == (jint) MarkStackSizeMax) {
261 if (PrintGCDetails && Verbose) {
262 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
263 }
264 return;
265 }
266 // Double capacity if possible
267 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
268 // Do not give up existing stack until we have managed to
269 // get the double capacity that we desired.
270 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
271 sizeof(oop)));
272 if (rs.is_reserved()) {
273 // Release the backing store associated with old stack
274 _virtual_space.release();
275 // Reinitialize virtual space for new stack
276 if (!_virtual_space.initialize(rs, rs.size())) {
277 fatal("Not enough swap for expanded marking stack capacity");
278 }
279 _base = (oop*)(_virtual_space.low());
280 _index = 0;
281 _capacity = new_capacity;
282 } else {
283 if (PrintGCDetails && Verbose) {
284 // Failed to double capacity, continue;
285 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
286 SIZE_FORMAT"K to " SIZE_FORMAT"K",
287 _capacity / K, new_capacity / K);
288 }
289 }
290 }
292 void CMMarkStack::set_should_expand() {
293 // If we're resetting the marking state because of an
294 // marking stack overflow, record that we should, if
295 // possible, expand the stack.
296 _should_expand = _cm->has_overflown();
297 }
299 CMMarkStack::~CMMarkStack() {
300 if (_base != NULL) {
301 _base = NULL;
302 _virtual_space.release();
303 }
304 }
306 void CMMarkStack::par_push(oop ptr) {
307 while (true) {
308 if (isFull()) {
309 _overflow = true;
310 return;
311 }
312 // Otherwise...
313 jint index = _index;
314 jint next_index = index+1;
315 jint res = Atomic::cmpxchg(next_index, &_index, index);
316 if (res == index) {
317 _base[index] = ptr;
318 // Note that we don't maintain this atomically. We could, but it
319 // doesn't seem necessary.
320 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
321 return;
322 }
323 // Otherwise, we need to try again.
324 }
325 }
327 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
328 while (true) {
329 if (isFull()) {
330 _overflow = true;
331 return;
332 }
333 // Otherwise...
334 jint index = _index;
335 jint next_index = index + n;
336 if (next_index > _capacity) {
337 _overflow = true;
338 return;
339 }
340 jint res = Atomic::cmpxchg(next_index, &_index, index);
341 if (res == index) {
342 for (int i = 0; i < n; i++) {
343 int ind = index + i;
344 assert(ind < _capacity, "By overflow test above.");
345 _base[ind] = ptr_arr[i];
346 }
347 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
348 return;
349 }
350 // Otherwise, we need to try again.
351 }
352 }
354 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
355 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
356 jint start = _index;
357 jint next_index = start + n;
358 if (next_index > _capacity) {
359 _overflow = true;
360 return;
361 }
362 // Otherwise.
363 _index = next_index;
364 for (int i = 0; i < n; i++) {
365 int ind = start + i;
366 assert(ind < _capacity, "By overflow test above.");
367 _base[ind] = ptr_arr[i];
368 }
369 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
370 }
372 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
373 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
374 jint index = _index;
375 if (index == 0) {
376 *n = 0;
377 return false;
378 } else {
379 int k = MIN2(max, index);
380 jint new_ind = index - k;
381 for (int j = 0; j < k; j++) {
382 ptr_arr[j] = _base[new_ind + j];
383 }
384 _index = new_ind;
385 *n = k;
386 return true;
387 }
388 }
390 template<class OopClosureClass>
391 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
392 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
393 || SafepointSynchronize::is_at_safepoint(),
394 "Drain recursion must be yield-safe.");
395 bool res = true;
396 debug_only(_drain_in_progress = true);
397 debug_only(_drain_in_progress_yields = yield_after);
398 while (!isEmpty()) {
399 oop newOop = pop();
400 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
401 assert(newOop->is_oop(), "Expected an oop");
402 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
403 "only grey objects on this stack");
404 newOop->oop_iterate(cl);
405 if (yield_after && _cm->do_yield_check()) {
406 res = false;
407 break;
408 }
409 }
410 debug_only(_drain_in_progress = false);
411 return res;
412 }
414 void CMMarkStack::note_start_of_gc() {
415 assert(_saved_index == -1,
416 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
417 _saved_index = _index;
418 }
420 void CMMarkStack::note_end_of_gc() {
421 // This is intentionally a guarantee, instead of an assert. If we
422 // accidentally add something to the mark stack during GC, it
423 // will be a correctness issue so it's better if we crash. we'll
424 // only check this once per GC anyway, so it won't be a performance
425 // issue in any way.
426 guarantee(_saved_index == _index,
427 err_msg("saved index: %d index: %d", _saved_index, _index));
428 _saved_index = -1;
429 }
431 void CMMarkStack::oops_do(OopClosure* f) {
432 assert(_saved_index == _index,
433 err_msg("saved index: %d index: %d", _saved_index, _index));
434 for (int i = 0; i < _index; i += 1) {
435 f->do_oop(&_base[i]);
436 }
437 }
439 CMRootRegions::CMRootRegions() :
440 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
441 _should_abort(false), _next_survivor(NULL) { }
443 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
444 _young_list = g1h->young_list();
445 _cm = cm;
446 }
448 void CMRootRegions::prepare_for_scan() {
449 assert(!scan_in_progress(), "pre-condition");
451 // Currently, only survivors can be root regions.
452 assert(_next_survivor == NULL, "pre-condition");
453 _next_survivor = _young_list->first_survivor_region();
454 _scan_in_progress = (_next_survivor != NULL);
455 _should_abort = false;
456 }
458 HeapRegion* CMRootRegions::claim_next() {
459 if (_should_abort) {
460 // If someone has set the should_abort flag, we return NULL to
461 // force the caller to bail out of their loop.
462 return NULL;
463 }
465 // Currently, only survivors can be root regions.
466 HeapRegion* res = _next_survivor;
467 if (res != NULL) {
468 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
469 // Read it again in case it changed while we were waiting for the lock.
470 res = _next_survivor;
471 if (res != NULL) {
472 if (res == _young_list->last_survivor_region()) {
473 // We just claimed the last survivor so store NULL to indicate
474 // that we're done.
475 _next_survivor = NULL;
476 } else {
477 _next_survivor = res->get_next_young_region();
478 }
479 } else {
480 // Someone else claimed the last survivor while we were trying
481 // to take the lock so nothing else to do.
482 }
483 }
484 assert(res == NULL || res->is_survivor(), "post-condition");
486 return res;
487 }
489 void CMRootRegions::scan_finished() {
490 assert(scan_in_progress(), "pre-condition");
492 // Currently, only survivors can be root regions.
493 if (!_should_abort) {
494 assert(_next_survivor == NULL, "we should have claimed all survivors");
495 }
496 _next_survivor = NULL;
498 {
499 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
500 _scan_in_progress = false;
501 RootRegionScan_lock->notify_all();
502 }
503 }
505 bool CMRootRegions::wait_until_scan_finished() {
506 if (!scan_in_progress()) return false;
508 {
509 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
510 while (scan_in_progress()) {
511 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
512 }
513 }
514 return true;
515 }
517 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
518 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
519 #endif // _MSC_VER
521 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
522 return MAX2((n_par_threads + 2) / 4, 1U);
523 }
525 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
526 _g1h(g1h),
527 _markBitMap1(),
528 _markBitMap2(),
529 _parallel_marking_threads(0),
530 _max_parallel_marking_threads(0),
531 _sleep_factor(0.0),
532 _marking_task_overhead(1.0),
533 _cleanup_sleep_factor(0.0),
534 _cleanup_task_overhead(1.0),
535 _cleanup_list("Cleanup List"),
536 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
537 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
538 CardTableModRefBS::card_shift,
539 false /* in_resource_area*/),
541 _prevMarkBitMap(&_markBitMap1),
542 _nextMarkBitMap(&_markBitMap2),
544 _markStack(this),
545 // _finger set in set_non_marking_state
547 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
548 // _active_tasks set in set_non_marking_state
549 // _tasks set inside the constructor
550 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
551 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
553 _has_overflown(false),
554 _concurrent(false),
555 _has_aborted(false),
556 _aborted_gc_id(GCId::undefined()),
557 _restart_for_overflow(false),
558 _concurrent_marking_in_progress(false),
560 // _verbose_level set below
562 _init_times(),
563 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
564 _cleanup_times(),
565 _total_counting_time(0.0),
566 _total_rs_scrub_time(0.0),
568 _parallel_workers(NULL),
570 _count_card_bitmaps(NULL),
571 _count_marked_bytes(NULL),
572 _completed_initialization(false) {
573 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
574 if (verbose_level < no_verbose) {
575 verbose_level = no_verbose;
576 }
577 if (verbose_level > high_verbose) {
578 verbose_level = high_verbose;
579 }
580 _verbose_level = verbose_level;
582 if (verbose_low()) {
583 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
584 "heap end = " INTPTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
585 }
587 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
588 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
590 // Create & start a ConcurrentMark thread.
591 _cmThread = new ConcurrentMarkThread(this);
592 assert(cmThread() != NULL, "CM Thread should have been created");
593 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
594 if (_cmThread->osthread() == NULL) {
595 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
596 }
598 assert(CGC_lock != NULL, "Where's the CGC_lock?");
599 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
600 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
602 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
603 satb_qs.set_buffer_size(G1SATBBufferSize);
605 _root_regions.init(_g1h, this);
607 if (ConcGCThreads > ParallelGCThreads) {
608 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
609 "than ParallelGCThreads (" UINTX_FORMAT ").",
610 ConcGCThreads, ParallelGCThreads);
611 return;
612 }
613 if (ParallelGCThreads == 0) {
614 // if we are not running with any parallel GC threads we will not
615 // spawn any marking threads either
616 _parallel_marking_threads = 0;
617 _max_parallel_marking_threads = 0;
618 _sleep_factor = 0.0;
619 _marking_task_overhead = 1.0;
620 } else {
621 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
622 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
623 // if both are set
624 _sleep_factor = 0.0;
625 _marking_task_overhead = 1.0;
626 } else if (G1MarkingOverheadPercent > 0) {
627 // We will calculate the number of parallel marking threads based
628 // on a target overhead with respect to the soft real-time goal
629 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
630 double overall_cm_overhead =
631 (double) MaxGCPauseMillis * marking_overhead /
632 (double) GCPauseIntervalMillis;
633 double cpu_ratio = 1.0 / (double) os::processor_count();
634 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
635 double marking_task_overhead =
636 overall_cm_overhead / marking_thread_num *
637 (double) os::processor_count();
638 double sleep_factor =
639 (1.0 - marking_task_overhead) / marking_task_overhead;
641 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
642 _sleep_factor = sleep_factor;
643 _marking_task_overhead = marking_task_overhead;
644 } else {
645 // Calculate the number of parallel marking threads by scaling
646 // the number of parallel GC threads.
647 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
648 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
649 _sleep_factor = 0.0;
650 _marking_task_overhead = 1.0;
651 }
653 assert(ConcGCThreads > 0, "Should have been set");
654 _parallel_marking_threads = (uint) ConcGCThreads;
655 _max_parallel_marking_threads = _parallel_marking_threads;
657 if (parallel_marking_threads() > 1) {
658 _cleanup_task_overhead = 1.0;
659 } else {
660 _cleanup_task_overhead = marking_task_overhead();
661 }
662 _cleanup_sleep_factor =
663 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
665 #if 0
666 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
667 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
668 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
669 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
670 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
671 #endif
673 guarantee(parallel_marking_threads() > 0, "peace of mind");
674 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
675 _max_parallel_marking_threads, false, true);
676 if (_parallel_workers == NULL) {
677 vm_exit_during_initialization("Failed necessary allocation.");
678 } else {
679 _parallel_workers->initialize_workers();
680 }
681 }
683 if (FLAG_IS_DEFAULT(MarkStackSize)) {
684 uintx mark_stack_size =
685 MIN2(MarkStackSizeMax,
686 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
687 // Verify that the calculated value for MarkStackSize is in range.
688 // It would be nice to use the private utility routine from Arguments.
689 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
690 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
691 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
692 mark_stack_size, (uintx) 1, MarkStackSizeMax);
693 return;
694 }
695 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
696 } else {
697 // Verify MarkStackSize is in range.
698 if (FLAG_IS_CMDLINE(MarkStackSize)) {
699 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
700 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
701 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
702 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
703 MarkStackSize, (uintx) 1, MarkStackSizeMax);
704 return;
705 }
706 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
707 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
708 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
709 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
710 MarkStackSize, MarkStackSizeMax);
711 return;
712 }
713 }
714 }
715 }
717 if (!_markStack.allocate(MarkStackSize)) {
718 warning("Failed to allocate CM marking stack");
719 return;
720 }
722 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
723 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
725 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
726 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
728 BitMap::idx_t card_bm_size = _card_bm.size();
730 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
731 _active_tasks = _max_worker_id;
733 size_t max_regions = (size_t) _g1h->max_regions();
734 for (uint i = 0; i < _max_worker_id; ++i) {
735 CMTaskQueue* task_queue = new CMTaskQueue();
736 task_queue->initialize();
737 _task_queues->register_queue(i, task_queue);
739 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
740 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
742 _tasks[i] = new CMTask(i, this,
743 _count_marked_bytes[i],
744 &_count_card_bitmaps[i],
745 task_queue, _task_queues);
747 _accum_task_vtime[i] = 0.0;
748 }
750 // Calculate the card number for the bottom of the heap. Used
751 // in biasing indexes into the accounting card bitmaps.
752 _heap_bottom_card_num =
753 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
754 CardTableModRefBS::card_shift);
756 // Clear all the liveness counting data
757 clear_all_count_data();
759 // so that the call below can read a sensible value
760 _heap_start = g1h->reserved_region().start();
761 set_non_marking_state();
762 _completed_initialization = true;
763 }
765 void ConcurrentMark::reset() {
766 // Starting values for these two. This should be called in a STW
767 // phase.
768 MemRegion reserved = _g1h->g1_reserved();
769 _heap_start = reserved.start();
770 _heap_end = reserved.end();
772 // Separated the asserts so that we know which one fires.
773 assert(_heap_start != NULL, "heap bounds should look ok");
774 assert(_heap_end != NULL, "heap bounds should look ok");
775 assert(_heap_start < _heap_end, "heap bounds should look ok");
777 // Reset all the marking data structures and any necessary flags
778 reset_marking_state();
780 if (verbose_low()) {
781 gclog_or_tty->print_cr("[global] resetting");
782 }
784 // We do reset all of them, since different phases will use
785 // different number of active threads. So, it's easiest to have all
786 // of them ready.
787 for (uint i = 0; i < _max_worker_id; ++i) {
788 _tasks[i]->reset(_nextMarkBitMap);
789 }
791 // we need this to make sure that the flag is on during the evac
792 // pause with initial mark piggy-backed
793 set_concurrent_marking_in_progress();
794 }
797 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
798 _markStack.set_should_expand();
799 _markStack.setEmpty(); // Also clears the _markStack overflow flag
800 if (clear_overflow) {
801 clear_has_overflown();
802 } else {
803 assert(has_overflown(), "pre-condition");
804 }
805 _finger = _heap_start;
807 for (uint i = 0; i < _max_worker_id; ++i) {
808 CMTaskQueue* queue = _task_queues->queue(i);
809 queue->set_empty();
810 }
811 }
813 void ConcurrentMark::set_concurrency(uint active_tasks) {
814 assert(active_tasks <= _max_worker_id, "we should not have more");
816 _active_tasks = active_tasks;
817 // Need to update the three data structures below according to the
818 // number of active threads for this phase.
819 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
820 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
821 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
822 }
824 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
825 set_concurrency(active_tasks);
827 _concurrent = concurrent;
828 // We propagate this to all tasks, not just the active ones.
829 for (uint i = 0; i < _max_worker_id; ++i)
830 _tasks[i]->set_concurrent(concurrent);
832 if (concurrent) {
833 set_concurrent_marking_in_progress();
834 } else {
835 // We currently assume that the concurrent flag has been set to
836 // false before we start remark. At this point we should also be
837 // in a STW phase.
838 assert(!concurrent_marking_in_progress(), "invariant");
839 assert(out_of_regions(),
840 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
841 p2i(_finger), p2i(_heap_end)));
842 }
843 }
845 void ConcurrentMark::set_non_marking_state() {
846 // We set the global marking state to some default values when we're
847 // not doing marking.
848 reset_marking_state();
849 _active_tasks = 0;
850 clear_concurrent_marking_in_progress();
851 }
853 ConcurrentMark::~ConcurrentMark() {
854 // The ConcurrentMark instance is never freed.
855 ShouldNotReachHere();
856 }
858 void ConcurrentMark::clearNextBitmap() {
859 G1CollectedHeap* g1h = G1CollectedHeap::heap();
861 // Make sure that the concurrent mark thread looks to still be in
862 // the current cycle.
863 guarantee(cmThread()->during_cycle(), "invariant");
865 // We are finishing up the current cycle by clearing the next
866 // marking bitmap and getting it ready for the next cycle. During
867 // this time no other cycle can start. So, let's make sure that this
868 // is the case.
869 guarantee(!g1h->mark_in_progress(), "invariant");
871 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
872 g1h->heap_region_iterate(&cl);
874 // Clear the liveness counting data. If the marking has been aborted, the abort()
875 // call already did that.
876 if (cl.complete()) {
877 clear_all_count_data();
878 }
880 // Repeat the asserts from above.
881 guarantee(cmThread()->during_cycle(), "invariant");
882 guarantee(!g1h->mark_in_progress(), "invariant");
883 }
885 class CheckBitmapClearHRClosure : public HeapRegionClosure {
886 CMBitMap* _bitmap;
887 bool _error;
888 public:
889 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
890 }
892 virtual bool doHeapRegion(HeapRegion* r) {
893 // This closure can be called concurrently to the mutator, so we must make sure
894 // that the result of the getNextMarkedWordAddress() call is compared to the
895 // value passed to it as limit to detect any found bits.
896 // We can use the region's orig_end() for the limit and the comparison value
897 // as it always contains the "real" end of the region that never changes and
898 // has no side effects.
899 // Due to the latter, there can also be no problem with the compiler generating
900 // reloads of the orig_end() call.
901 HeapWord* end = r->orig_end();
902 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
903 }
904 };
906 bool ConcurrentMark::nextMarkBitmapIsClear() {
907 CheckBitmapClearHRClosure cl(_nextMarkBitMap);
908 _g1h->heap_region_iterate(&cl);
909 return cl.complete();
910 }
912 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
913 public:
914 bool doHeapRegion(HeapRegion* r) {
915 if (!r->continuesHumongous()) {
916 r->note_start_of_marking();
917 }
918 return false;
919 }
920 };
922 void ConcurrentMark::checkpointRootsInitialPre() {
923 G1CollectedHeap* g1h = G1CollectedHeap::heap();
924 G1CollectorPolicy* g1p = g1h->g1_policy();
926 _has_aborted = false;
928 #ifndef PRODUCT
929 if (G1PrintReachableAtInitialMark) {
930 print_reachable("at-cycle-start",
931 VerifyOption_G1UsePrevMarking, true /* all */);
932 }
933 #endif
935 // Initialise marking structures. This has to be done in a STW phase.
936 reset();
938 // For each region note start of marking.
939 NoteStartOfMarkHRClosure startcl;
940 g1h->heap_region_iterate(&startcl);
941 }
944 void ConcurrentMark::checkpointRootsInitialPost() {
945 G1CollectedHeap* g1h = G1CollectedHeap::heap();
947 // If we force an overflow during remark, the remark operation will
948 // actually abort and we'll restart concurrent marking. If we always
949 // force an oveflow during remark we'll never actually complete the
950 // marking phase. So, we initilize this here, at the start of the
951 // cycle, so that at the remaining overflow number will decrease at
952 // every remark and we'll eventually not need to cause one.
953 force_overflow_stw()->init();
955 // Start Concurrent Marking weak-reference discovery.
956 ReferenceProcessor* rp = g1h->ref_processor_cm();
957 // enable ("weak") refs discovery
958 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
959 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
961 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
962 // This is the start of the marking cycle, we're expected all
963 // threads to have SATB queues with active set to false.
964 satb_mq_set.set_active_all_threads(true, /* new active value */
965 false /* expected_active */);
967 _root_regions.prepare_for_scan();
969 // update_g1_committed() will be called at the end of an evac pause
970 // when marking is on. So, it's also called at the end of the
971 // initial-mark pause to update the heap end, if the heap expands
972 // during it. No need to call it here.
973 }
975 /*
976 * Notice that in the next two methods, we actually leave the STS
977 * during the barrier sync and join it immediately afterwards. If we
978 * do not do this, the following deadlock can occur: one thread could
979 * be in the barrier sync code, waiting for the other thread to also
980 * sync up, whereas another one could be trying to yield, while also
981 * waiting for the other threads to sync up too.
982 *
983 * Note, however, that this code is also used during remark and in
984 * this case we should not attempt to leave / enter the STS, otherwise
985 * we'll either hit an asseert (debug / fastdebug) or deadlock
986 * (product). So we should only leave / enter the STS if we are
987 * operating concurrently.
988 *
989 * Because the thread that does the sync barrier has left the STS, it
990 * is possible to be suspended for a Full GC or an evacuation pause
991 * could occur. This is actually safe, since the entering the sync
992 * barrier is one of the last things do_marking_step() does, and it
993 * doesn't manipulate any data structures afterwards.
994 */
996 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
997 if (verbose_low()) {
998 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
999 }
1001 if (concurrent()) {
1002 SuspendibleThreadSet::leave();
1003 }
1005 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
1007 if (concurrent()) {
1008 SuspendibleThreadSet::join();
1009 }
1010 // at this point everyone should have synced up and not be doing any
1011 // more work
1013 if (verbose_low()) {
1014 if (barrier_aborted) {
1015 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1016 } else {
1017 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1018 }
1019 }
1021 if (barrier_aborted) {
1022 // If the barrier aborted we ignore the overflow condition and
1023 // just abort the whole marking phase as quickly as possible.
1024 return;
1025 }
1027 // If we're executing the concurrent phase of marking, reset the marking
1028 // state; otherwise the marking state is reset after reference processing,
1029 // during the remark pause.
1030 // If we reset here as a result of an overflow during the remark we will
1031 // see assertion failures from any subsequent set_concurrency_and_phase()
1032 // calls.
1033 if (concurrent()) {
1034 // let the task associated with with worker 0 do this
1035 if (worker_id == 0) {
1036 // task 0 is responsible for clearing the global data structures
1037 // We should be here because of an overflow. During STW we should
1038 // not clear the overflow flag since we rely on it being true when
1039 // we exit this method to abort the pause and restart concurent
1040 // marking.
1041 reset_marking_state(true /* clear_overflow */);
1042 force_overflow()->update();
1044 if (G1Log::fine()) {
1045 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1046 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1047 }
1048 }
1049 }
1051 // after this, each task should reset its own data structures then
1052 // then go into the second barrier
1053 }
1055 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1056 if (verbose_low()) {
1057 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1058 }
1060 if (concurrent()) {
1061 SuspendibleThreadSet::leave();
1062 }
1064 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1066 if (concurrent()) {
1067 SuspendibleThreadSet::join();
1068 }
1069 // at this point everything should be re-initialized and ready to go
1071 if (verbose_low()) {
1072 if (barrier_aborted) {
1073 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1074 } else {
1075 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1076 }
1077 }
1078 }
1080 #ifndef PRODUCT
1081 void ForceOverflowSettings::init() {
1082 _num_remaining = G1ConcMarkForceOverflow;
1083 _force = false;
1084 update();
1085 }
1087 void ForceOverflowSettings::update() {
1088 if (_num_remaining > 0) {
1089 _num_remaining -= 1;
1090 _force = true;
1091 } else {
1092 _force = false;
1093 }
1094 }
1096 bool ForceOverflowSettings::should_force() {
1097 if (_force) {
1098 _force = false;
1099 return true;
1100 } else {
1101 return false;
1102 }
1103 }
1104 #endif // !PRODUCT
1106 class CMConcurrentMarkingTask: public AbstractGangTask {
1107 private:
1108 ConcurrentMark* _cm;
1109 ConcurrentMarkThread* _cmt;
1111 public:
1112 void work(uint worker_id) {
1113 assert(Thread::current()->is_ConcurrentGC_thread(),
1114 "this should only be done by a conc GC thread");
1115 ResourceMark rm;
1117 double start_vtime = os::elapsedVTime();
1119 SuspendibleThreadSet::join();
1121 assert(worker_id < _cm->active_tasks(), "invariant");
1122 CMTask* the_task = _cm->task(worker_id);
1123 the_task->record_start_time();
1124 if (!_cm->has_aborted()) {
1125 do {
1126 double start_vtime_sec = os::elapsedVTime();
1127 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1129 the_task->do_marking_step(mark_step_duration_ms,
1130 true /* do_termination */,
1131 false /* is_serial*/);
1133 double end_vtime_sec = os::elapsedVTime();
1134 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1135 _cm->clear_has_overflown();
1137 _cm->do_yield_check(worker_id);
1139 jlong sleep_time_ms;
1140 if (!_cm->has_aborted() && the_task->has_aborted()) {
1141 sleep_time_ms =
1142 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1143 SuspendibleThreadSet::leave();
1144 os::sleep(Thread::current(), sleep_time_ms, false);
1145 SuspendibleThreadSet::join();
1146 }
1147 } while (!_cm->has_aborted() && the_task->has_aborted());
1148 }
1149 the_task->record_end_time();
1150 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1152 SuspendibleThreadSet::leave();
1154 double end_vtime = os::elapsedVTime();
1155 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1156 }
1158 CMConcurrentMarkingTask(ConcurrentMark* cm,
1159 ConcurrentMarkThread* cmt) :
1160 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1162 ~CMConcurrentMarkingTask() { }
1163 };
1165 // Calculates the number of active workers for a concurrent
1166 // phase.
1167 uint ConcurrentMark::calc_parallel_marking_threads() {
1168 if (G1CollectedHeap::use_parallel_gc_threads()) {
1169 uint n_conc_workers = 0;
1170 if (!UseDynamicNumberOfGCThreads ||
1171 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1172 !ForceDynamicNumberOfGCThreads)) {
1173 n_conc_workers = max_parallel_marking_threads();
1174 } else {
1175 n_conc_workers =
1176 AdaptiveSizePolicy::calc_default_active_workers(
1177 max_parallel_marking_threads(),
1178 1, /* Minimum workers */
1179 parallel_marking_threads(),
1180 Threads::number_of_non_daemon_threads());
1181 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1182 // that scaling has already gone into "_max_parallel_marking_threads".
1183 }
1184 assert(n_conc_workers > 0, "Always need at least 1");
1185 return n_conc_workers;
1186 }
1187 // If we are not running with any parallel GC threads we will not
1188 // have spawned any marking threads either. Hence the number of
1189 // concurrent workers should be 0.
1190 return 0;
1191 }
1193 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1194 // Currently, only survivors can be root regions.
1195 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1196 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1198 const uintx interval = PrefetchScanIntervalInBytes;
1199 HeapWord* curr = hr->bottom();
1200 const HeapWord* end = hr->top();
1201 while (curr < end) {
1202 Prefetch::read(curr, interval);
1203 oop obj = oop(curr);
1204 int size = obj->oop_iterate(&cl);
1205 assert(size == obj->size(), "sanity");
1206 curr += size;
1207 }
1208 }
1210 class CMRootRegionScanTask : public AbstractGangTask {
1211 private:
1212 ConcurrentMark* _cm;
1214 public:
1215 CMRootRegionScanTask(ConcurrentMark* cm) :
1216 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1218 void work(uint worker_id) {
1219 assert(Thread::current()->is_ConcurrentGC_thread(),
1220 "this should only be done by a conc GC thread");
1222 CMRootRegions* root_regions = _cm->root_regions();
1223 HeapRegion* hr = root_regions->claim_next();
1224 while (hr != NULL) {
1225 _cm->scanRootRegion(hr, worker_id);
1226 hr = root_regions->claim_next();
1227 }
1228 }
1229 };
1231 void ConcurrentMark::scanRootRegions() {
1232 // Start of concurrent marking.
1233 ClassLoaderDataGraph::clear_claimed_marks();
1235 // scan_in_progress() will have been set to true only if there was
1236 // at least one root region to scan. So, if it's false, we
1237 // should not attempt to do any further work.
1238 if (root_regions()->scan_in_progress()) {
1239 _parallel_marking_threads = calc_parallel_marking_threads();
1240 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1241 "Maximum number of marking threads exceeded");
1242 uint active_workers = MAX2(1U, parallel_marking_threads());
1244 CMRootRegionScanTask task(this);
1245 if (use_parallel_marking_threads()) {
1246 _parallel_workers->set_active_workers((int) active_workers);
1247 _parallel_workers->run_task(&task);
1248 } else {
1249 task.work(0);
1250 }
1252 // It's possible that has_aborted() is true here without actually
1253 // aborting the survivor scan earlier. This is OK as it's
1254 // mainly used for sanity checking.
1255 root_regions()->scan_finished();
1256 }
1257 }
1259 void ConcurrentMark::markFromRoots() {
1260 // we might be tempted to assert that:
1261 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1262 // "inconsistent argument?");
1263 // However that wouldn't be right, because it's possible that
1264 // a safepoint is indeed in progress as a younger generation
1265 // stop-the-world GC happens even as we mark in this generation.
1267 _restart_for_overflow = false;
1268 force_overflow_conc()->init();
1270 // _g1h has _n_par_threads
1271 _parallel_marking_threads = calc_parallel_marking_threads();
1272 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1273 "Maximum number of marking threads exceeded");
1275 uint active_workers = MAX2(1U, parallel_marking_threads());
1277 // Parallel task terminator is set in "set_concurrency_and_phase()"
1278 set_concurrency_and_phase(active_workers, true /* concurrent */);
1280 CMConcurrentMarkingTask markingTask(this, cmThread());
1281 if (use_parallel_marking_threads()) {
1282 _parallel_workers->set_active_workers((int)active_workers);
1283 // Don't set _n_par_threads because it affects MT in process_roots()
1284 // and the decisions on that MT processing is made elsewhere.
1285 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1286 _parallel_workers->run_task(&markingTask);
1287 } else {
1288 markingTask.work(0);
1289 }
1290 print_stats();
1291 }
1293 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1294 // world is stopped at this checkpoint
1295 assert(SafepointSynchronize::is_at_safepoint(),
1296 "world should be stopped");
1298 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1300 // If a full collection has happened, we shouldn't do this.
1301 if (has_aborted()) {
1302 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1303 return;
1304 }
1306 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1308 if (VerifyDuringGC) {
1309 HandleMark hm; // handle scope
1310 Universe::heap()->prepare_for_verify();
1311 Universe::verify(VerifyOption_G1UsePrevMarking,
1312 " VerifyDuringGC:(before)");
1313 }
1314 g1h->check_bitmaps("Remark Start");
1316 G1CollectorPolicy* g1p = g1h->g1_policy();
1317 g1p->record_concurrent_mark_remark_start();
1319 double start = os::elapsedTime();
1321 checkpointRootsFinalWork();
1323 double mark_work_end = os::elapsedTime();
1325 weakRefsWork(clear_all_soft_refs);
1327 if (has_overflown()) {
1328 // Oops. We overflowed. Restart concurrent marking.
1329 _restart_for_overflow = true;
1330 if (G1TraceMarkStackOverflow) {
1331 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1332 }
1334 // Verify the heap w.r.t. the previous marking bitmap.
1335 if (VerifyDuringGC) {
1336 HandleMark hm; // handle scope
1337 Universe::heap()->prepare_for_verify();
1338 Universe::verify(VerifyOption_G1UsePrevMarking,
1339 " VerifyDuringGC:(overflow)");
1340 }
1342 // Clear the marking state because we will be restarting
1343 // marking due to overflowing the global mark stack.
1344 reset_marking_state();
1345 } else {
1346 // Aggregate the per-task counting data that we have accumulated
1347 // while marking.
1348 aggregate_count_data();
1350 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1351 // We're done with marking.
1352 // This is the end of the marking cycle, we're expected all
1353 // threads to have SATB queues with active set to true.
1354 satb_mq_set.set_active_all_threads(false, /* new active value */
1355 true /* expected_active */);
1357 if (VerifyDuringGC) {
1358 HandleMark hm; // handle scope
1359 Universe::heap()->prepare_for_verify();
1360 Universe::verify(VerifyOption_G1UseNextMarking,
1361 " VerifyDuringGC:(after)");
1362 }
1363 g1h->check_bitmaps("Remark End");
1364 assert(!restart_for_overflow(), "sanity");
1365 // Completely reset the marking state since marking completed
1366 set_non_marking_state();
1367 }
1369 // Expand the marking stack, if we have to and if we can.
1370 if (_markStack.should_expand()) {
1371 _markStack.expand();
1372 }
1374 // Statistics
1375 double now = os::elapsedTime();
1376 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1377 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1378 _remark_times.add((now - start) * 1000.0);
1380 g1p->record_concurrent_mark_remark_end();
1382 G1CMIsAliveClosure is_alive(g1h);
1383 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1384 }
1386 // Base class of the closures that finalize and verify the
1387 // liveness counting data.
1388 class CMCountDataClosureBase: public HeapRegionClosure {
1389 protected:
1390 G1CollectedHeap* _g1h;
1391 ConcurrentMark* _cm;
1392 CardTableModRefBS* _ct_bs;
1394 BitMap* _region_bm;
1395 BitMap* _card_bm;
1397 // Takes a region that's not empty (i.e., it has at least one
1398 // live object in it and sets its corresponding bit on the region
1399 // bitmap to 1. If the region is "starts humongous" it will also set
1400 // to 1 the bits on the region bitmap that correspond to its
1401 // associated "continues humongous" regions.
1402 void set_bit_for_region(HeapRegion* hr) {
1403 assert(!hr->continuesHumongous(), "should have filtered those out");
1405 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1406 if (!hr->startsHumongous()) {
1407 // Normal (non-humongous) case: just set the bit.
1408 _region_bm->par_at_put(index, true);
1409 } else {
1410 // Starts humongous case: calculate how many regions are part of
1411 // this humongous region and then set the bit range.
1412 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1413 _region_bm->par_at_put_range(index, end_index, true);
1414 }
1415 }
1417 public:
1418 CMCountDataClosureBase(G1CollectedHeap* g1h,
1419 BitMap* region_bm, BitMap* card_bm):
1420 _g1h(g1h), _cm(g1h->concurrent_mark()),
1421 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1422 _region_bm(region_bm), _card_bm(card_bm) { }
1423 };
1425 // Closure that calculates the # live objects per region. Used
1426 // for verification purposes during the cleanup pause.
1427 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1428 CMBitMapRO* _bm;
1429 size_t _region_marked_bytes;
1431 public:
1432 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1433 BitMap* region_bm, BitMap* card_bm) :
1434 CMCountDataClosureBase(g1h, region_bm, card_bm),
1435 _bm(bm), _region_marked_bytes(0) { }
1437 bool doHeapRegion(HeapRegion* hr) {
1439 if (hr->continuesHumongous()) {
1440 // We will ignore these here and process them when their
1441 // associated "starts humongous" region is processed (see
1442 // set_bit_for_heap_region()). Note that we cannot rely on their
1443 // associated "starts humongous" region to have their bit set to
1444 // 1 since, due to the region chunking in the parallel region
1445 // iteration, a "continues humongous" region might be visited
1446 // before its associated "starts humongous".
1447 return false;
1448 }
1450 HeapWord* ntams = hr->next_top_at_mark_start();
1451 HeapWord* start = hr->bottom();
1453 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1454 err_msg("Preconditions not met - "
1455 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1456 p2i(start), p2i(ntams), p2i(hr->end())));
1458 // Find the first marked object at or after "start".
1459 start = _bm->getNextMarkedWordAddress(start, ntams);
1461 size_t marked_bytes = 0;
1463 while (start < ntams) {
1464 oop obj = oop(start);
1465 int obj_sz = obj->size();
1466 HeapWord* obj_end = start + obj_sz;
1468 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1469 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1471 // Note: if we're looking at the last region in heap - obj_end
1472 // could be actually just beyond the end of the heap; end_idx
1473 // will then correspond to a (non-existent) card that is also
1474 // just beyond the heap.
1475 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1476 // end of object is not card aligned - increment to cover
1477 // all the cards spanned by the object
1478 end_idx += 1;
1479 }
1481 // Set the bits in the card BM for the cards spanned by this object.
1482 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1484 // Add the size of this object to the number of marked bytes.
1485 marked_bytes += (size_t)obj_sz * HeapWordSize;
1487 // Find the next marked object after this one.
1488 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1489 }
1491 // Mark the allocated-since-marking portion...
1492 HeapWord* top = hr->top();
1493 if (ntams < top) {
1494 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1495 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1497 // Note: if we're looking at the last region in heap - top
1498 // could be actually just beyond the end of the heap; end_idx
1499 // will then correspond to a (non-existent) card that is also
1500 // just beyond the heap.
1501 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1502 // end of object is not card aligned - increment to cover
1503 // all the cards spanned by the object
1504 end_idx += 1;
1505 }
1506 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1508 // This definitely means the region has live objects.
1509 set_bit_for_region(hr);
1510 }
1512 // Update the live region bitmap.
1513 if (marked_bytes > 0) {
1514 set_bit_for_region(hr);
1515 }
1517 // Set the marked bytes for the current region so that
1518 // it can be queried by a calling verificiation routine
1519 _region_marked_bytes = marked_bytes;
1521 return false;
1522 }
1524 size_t region_marked_bytes() const { return _region_marked_bytes; }
1525 };
1527 // Heap region closure used for verifying the counting data
1528 // that was accumulated concurrently and aggregated during
1529 // the remark pause. This closure is applied to the heap
1530 // regions during the STW cleanup pause.
1532 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1533 G1CollectedHeap* _g1h;
1534 ConcurrentMark* _cm;
1535 CalcLiveObjectsClosure _calc_cl;
1536 BitMap* _region_bm; // Region BM to be verified
1537 BitMap* _card_bm; // Card BM to be verified
1538 bool _verbose; // verbose output?
1540 BitMap* _exp_region_bm; // Expected Region BM values
1541 BitMap* _exp_card_bm; // Expected card BM values
1543 int _failures;
1545 public:
1546 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1547 BitMap* region_bm,
1548 BitMap* card_bm,
1549 BitMap* exp_region_bm,
1550 BitMap* exp_card_bm,
1551 bool verbose) :
1552 _g1h(g1h), _cm(g1h->concurrent_mark()),
1553 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1554 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1555 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1556 _failures(0) { }
1558 int failures() const { return _failures; }
1560 bool doHeapRegion(HeapRegion* hr) {
1561 if (hr->continuesHumongous()) {
1562 // We will ignore these here and process them when their
1563 // associated "starts humongous" region is processed (see
1564 // set_bit_for_heap_region()). Note that we cannot rely on their
1565 // associated "starts humongous" region to have their bit set to
1566 // 1 since, due to the region chunking in the parallel region
1567 // iteration, a "continues humongous" region might be visited
1568 // before its associated "starts humongous".
1569 return false;
1570 }
1572 int failures = 0;
1574 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1575 // this region and set the corresponding bits in the expected region
1576 // and card bitmaps.
1577 bool res = _calc_cl.doHeapRegion(hr);
1578 assert(res == false, "should be continuing");
1580 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1581 Mutex::_no_safepoint_check_flag);
1583 // Verify the marked bytes for this region.
1584 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1585 size_t act_marked_bytes = hr->next_marked_bytes();
1587 // We're not OK if expected marked bytes > actual marked bytes. It means
1588 // we have missed accounting some objects during the actual marking.
1589 if (exp_marked_bytes > act_marked_bytes) {
1590 if (_verbose) {
1591 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1592 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1593 hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1594 }
1595 failures += 1;
1596 }
1598 // Verify the bit, for this region, in the actual and expected
1599 // (which was just calculated) region bit maps.
1600 // We're not OK if the bit in the calculated expected region
1601 // bitmap is set and the bit in the actual region bitmap is not.
1602 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1604 bool expected = _exp_region_bm->at(index);
1605 bool actual = _region_bm->at(index);
1606 if (expected && !actual) {
1607 if (_verbose) {
1608 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1609 "expected: %s, actual: %s",
1610 hr->hrm_index(),
1611 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1612 }
1613 failures += 1;
1614 }
1616 // Verify that the card bit maps for the cards spanned by the current
1617 // region match. We have an error if we have a set bit in the expected
1618 // bit map and the corresponding bit in the actual bitmap is not set.
1620 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1621 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1623 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1624 expected = _exp_card_bm->at(i);
1625 actual = _card_bm->at(i);
1627 if (expected && !actual) {
1628 if (_verbose) {
1629 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1630 "expected: %s, actual: %s",
1631 hr->hrm_index(), i,
1632 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1633 }
1634 failures += 1;
1635 }
1636 }
1638 if (failures > 0 && _verbose) {
1639 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1640 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1641 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1642 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1643 }
1645 _failures += failures;
1647 // We could stop iteration over the heap when we
1648 // find the first violating region by returning true.
1649 return false;
1650 }
1651 };
1653 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1654 protected:
1655 G1CollectedHeap* _g1h;
1656 ConcurrentMark* _cm;
1657 BitMap* _actual_region_bm;
1658 BitMap* _actual_card_bm;
1660 uint _n_workers;
1662 BitMap* _expected_region_bm;
1663 BitMap* _expected_card_bm;
1665 int _failures;
1666 bool _verbose;
1668 public:
1669 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1670 BitMap* region_bm, BitMap* card_bm,
1671 BitMap* expected_region_bm, BitMap* expected_card_bm)
1672 : AbstractGangTask("G1 verify final counting"),
1673 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1674 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1675 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1676 _failures(0), _verbose(false),
1677 _n_workers(0) {
1678 assert(VerifyDuringGC, "don't call this otherwise");
1680 // Use the value already set as the number of active threads
1681 // in the call to run_task().
1682 if (G1CollectedHeap::use_parallel_gc_threads()) {
1683 assert( _g1h->workers()->active_workers() > 0,
1684 "Should have been previously set");
1685 _n_workers = _g1h->workers()->active_workers();
1686 } else {
1687 _n_workers = 1;
1688 }
1690 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1691 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1693 _verbose = _cm->verbose_medium();
1694 }
1696 void work(uint worker_id) {
1697 assert(worker_id < _n_workers, "invariant");
1699 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1700 _actual_region_bm, _actual_card_bm,
1701 _expected_region_bm,
1702 _expected_card_bm,
1703 _verbose);
1705 if (G1CollectedHeap::use_parallel_gc_threads()) {
1706 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1707 worker_id,
1708 _n_workers,
1709 HeapRegion::VerifyCountClaimValue);
1710 } else {
1711 _g1h->heap_region_iterate(&verify_cl);
1712 }
1714 Atomic::add(verify_cl.failures(), &_failures);
1715 }
1717 int failures() const { return _failures; }
1718 };
1720 // Closure that finalizes the liveness counting data.
1721 // Used during the cleanup pause.
1722 // Sets the bits corresponding to the interval [NTAMS, top]
1723 // (which contains the implicitly live objects) in the
1724 // card liveness bitmap. Also sets the bit for each region,
1725 // containing live data, in the region liveness bitmap.
1727 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1728 public:
1729 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1730 BitMap* region_bm,
1731 BitMap* card_bm) :
1732 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1734 bool doHeapRegion(HeapRegion* hr) {
1736 if (hr->continuesHumongous()) {
1737 // We will ignore these here and process them when their
1738 // associated "starts humongous" region is processed (see
1739 // set_bit_for_heap_region()). Note that we cannot rely on their
1740 // associated "starts humongous" region to have their bit set to
1741 // 1 since, due to the region chunking in the parallel region
1742 // iteration, a "continues humongous" region might be visited
1743 // before its associated "starts humongous".
1744 return false;
1745 }
1747 HeapWord* ntams = hr->next_top_at_mark_start();
1748 HeapWord* top = hr->top();
1750 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1752 // Mark the allocated-since-marking portion...
1753 if (ntams < top) {
1754 // This definitely means the region has live objects.
1755 set_bit_for_region(hr);
1757 // Now set the bits in the card bitmap for [ntams, top)
1758 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1759 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1761 // Note: if we're looking at the last region in heap - top
1762 // could be actually just beyond the end of the heap; end_idx
1763 // will then correspond to a (non-existent) card that is also
1764 // just beyond the heap.
1765 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1766 // end of object is not card aligned - increment to cover
1767 // all the cards spanned by the object
1768 end_idx += 1;
1769 }
1771 assert(end_idx <= _card_bm->size(),
1772 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1773 end_idx, _card_bm->size()));
1774 assert(start_idx < _card_bm->size(),
1775 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1776 start_idx, _card_bm->size()));
1778 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1779 }
1781 // Set the bit for the region if it contains live data
1782 if (hr->next_marked_bytes() > 0) {
1783 set_bit_for_region(hr);
1784 }
1786 return false;
1787 }
1788 };
1790 class G1ParFinalCountTask: public AbstractGangTask {
1791 protected:
1792 G1CollectedHeap* _g1h;
1793 ConcurrentMark* _cm;
1794 BitMap* _actual_region_bm;
1795 BitMap* _actual_card_bm;
1797 uint _n_workers;
1799 public:
1800 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1801 : AbstractGangTask("G1 final counting"),
1802 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1803 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1804 _n_workers(0) {
1805 // Use the value already set as the number of active threads
1806 // in the call to run_task().
1807 if (G1CollectedHeap::use_parallel_gc_threads()) {
1808 assert( _g1h->workers()->active_workers() > 0,
1809 "Should have been previously set");
1810 _n_workers = _g1h->workers()->active_workers();
1811 } else {
1812 _n_workers = 1;
1813 }
1814 }
1816 void work(uint worker_id) {
1817 assert(worker_id < _n_workers, "invariant");
1819 FinalCountDataUpdateClosure final_update_cl(_g1h,
1820 _actual_region_bm,
1821 _actual_card_bm);
1823 if (G1CollectedHeap::use_parallel_gc_threads()) {
1824 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1825 worker_id,
1826 _n_workers,
1827 HeapRegion::FinalCountClaimValue);
1828 } else {
1829 _g1h->heap_region_iterate(&final_update_cl);
1830 }
1831 }
1832 };
1834 class G1ParNoteEndTask;
1836 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1837 G1CollectedHeap* _g1;
1838 size_t _max_live_bytes;
1839 uint _regions_claimed;
1840 size_t _freed_bytes;
1841 FreeRegionList* _local_cleanup_list;
1842 HeapRegionSetCount _old_regions_removed;
1843 HeapRegionSetCount _humongous_regions_removed;
1844 HRRSCleanupTask* _hrrs_cleanup_task;
1845 double _claimed_region_time;
1846 double _max_region_time;
1848 public:
1849 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1850 FreeRegionList* local_cleanup_list,
1851 HRRSCleanupTask* hrrs_cleanup_task) :
1852 _g1(g1),
1853 _max_live_bytes(0), _regions_claimed(0),
1854 _freed_bytes(0),
1855 _claimed_region_time(0.0), _max_region_time(0.0),
1856 _local_cleanup_list(local_cleanup_list),
1857 _old_regions_removed(),
1858 _humongous_regions_removed(),
1859 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1861 size_t freed_bytes() { return _freed_bytes; }
1862 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1863 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1865 bool doHeapRegion(HeapRegion *hr) {
1866 if (hr->continuesHumongous()) {
1867 return false;
1868 }
1869 // We use a claim value of zero here because all regions
1870 // were claimed with value 1 in the FinalCount task.
1871 _g1->reset_gc_time_stamps(hr);
1872 double start = os::elapsedTime();
1873 _regions_claimed++;
1874 hr->note_end_of_marking();
1875 _max_live_bytes += hr->max_live_bytes();
1877 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1878 _freed_bytes += hr->used();
1879 hr->set_containing_set(NULL);
1880 if (hr->isHumongous()) {
1881 assert(hr->startsHumongous(), "we should only see starts humongous");
1882 _humongous_regions_removed.increment(1u, hr->capacity());
1883 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1884 } else {
1885 _old_regions_removed.increment(1u, hr->capacity());
1886 _g1->free_region(hr, _local_cleanup_list, true);
1887 }
1888 } else {
1889 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1890 }
1892 double region_time = (os::elapsedTime() - start);
1893 _claimed_region_time += region_time;
1894 if (region_time > _max_region_time) {
1895 _max_region_time = region_time;
1896 }
1897 return false;
1898 }
1900 size_t max_live_bytes() { return _max_live_bytes; }
1901 uint regions_claimed() { return _regions_claimed; }
1902 double claimed_region_time_sec() { return _claimed_region_time; }
1903 double max_region_time_sec() { return _max_region_time; }
1904 };
1906 class G1ParNoteEndTask: public AbstractGangTask {
1907 friend class G1NoteEndOfConcMarkClosure;
1909 protected:
1910 G1CollectedHeap* _g1h;
1911 size_t _max_live_bytes;
1912 size_t _freed_bytes;
1913 FreeRegionList* _cleanup_list;
1915 public:
1916 G1ParNoteEndTask(G1CollectedHeap* g1h,
1917 FreeRegionList* cleanup_list) :
1918 AbstractGangTask("G1 note end"), _g1h(g1h),
1919 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1921 void work(uint worker_id) {
1922 double start = os::elapsedTime();
1923 FreeRegionList local_cleanup_list("Local Cleanup List");
1924 HRRSCleanupTask hrrs_cleanup_task;
1925 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1926 &hrrs_cleanup_task);
1927 if (G1CollectedHeap::use_parallel_gc_threads()) {
1928 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1929 _g1h->workers()->active_workers(),
1930 HeapRegion::NoteEndClaimValue);
1931 } else {
1932 _g1h->heap_region_iterate(&g1_note_end);
1933 }
1934 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1936 // Now update the lists
1937 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1938 {
1939 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1940 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1941 _max_live_bytes += g1_note_end.max_live_bytes();
1942 _freed_bytes += g1_note_end.freed_bytes();
1944 // If we iterate over the global cleanup list at the end of
1945 // cleanup to do this printing we will not guarantee to only
1946 // generate output for the newly-reclaimed regions (the list
1947 // might not be empty at the beginning of cleanup; we might
1948 // still be working on its previous contents). So we do the
1949 // printing here, before we append the new regions to the global
1950 // cleanup list.
1952 G1HRPrinter* hr_printer = _g1h->hr_printer();
1953 if (hr_printer->is_active()) {
1954 FreeRegionListIterator iter(&local_cleanup_list);
1955 while (iter.more_available()) {
1956 HeapRegion* hr = iter.get_next();
1957 hr_printer->cleanup(hr);
1958 }
1959 }
1961 _cleanup_list->add_ordered(&local_cleanup_list);
1962 assert(local_cleanup_list.is_empty(), "post-condition");
1964 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1965 }
1966 }
1967 size_t max_live_bytes() { return _max_live_bytes; }
1968 size_t freed_bytes() { return _freed_bytes; }
1969 };
1971 class G1ParScrubRemSetTask: public AbstractGangTask {
1972 protected:
1973 G1RemSet* _g1rs;
1974 BitMap* _region_bm;
1975 BitMap* _card_bm;
1976 public:
1977 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1978 BitMap* region_bm, BitMap* card_bm) :
1979 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1980 _region_bm(region_bm), _card_bm(card_bm) { }
1982 void work(uint worker_id) {
1983 if (G1CollectedHeap::use_parallel_gc_threads()) {
1984 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1985 HeapRegion::ScrubRemSetClaimValue);
1986 } else {
1987 _g1rs->scrub(_region_bm, _card_bm);
1988 }
1989 }
1991 };
1993 void ConcurrentMark::cleanup() {
1994 // world is stopped at this checkpoint
1995 assert(SafepointSynchronize::is_at_safepoint(),
1996 "world should be stopped");
1997 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1999 // If a full collection has happened, we shouldn't do this.
2000 if (has_aborted()) {
2001 g1h->set_marking_complete(); // So bitmap clearing isn't confused
2002 return;
2003 }
2005 g1h->verify_region_sets_optional();
2007 if (VerifyDuringGC) {
2008 HandleMark hm; // handle scope
2009 Universe::heap()->prepare_for_verify();
2010 Universe::verify(VerifyOption_G1UsePrevMarking,
2011 " VerifyDuringGC:(before)");
2012 }
2013 g1h->check_bitmaps("Cleanup Start");
2015 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
2016 g1p->record_concurrent_mark_cleanup_start();
2018 double start = os::elapsedTime();
2020 HeapRegionRemSet::reset_for_cleanup_tasks();
2022 uint n_workers;
2024 // Do counting once more with the world stopped for good measure.
2025 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2027 if (G1CollectedHeap::use_parallel_gc_threads()) {
2028 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2029 "sanity check");
2031 g1h->set_par_threads();
2032 n_workers = g1h->n_par_threads();
2033 assert(g1h->n_par_threads() == n_workers,
2034 "Should not have been reset");
2035 g1h->workers()->run_task(&g1_par_count_task);
2036 // Done with the parallel phase so reset to 0.
2037 g1h->set_par_threads(0);
2039 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2040 "sanity check");
2041 } else {
2042 n_workers = 1;
2043 g1_par_count_task.work(0);
2044 }
2046 if (VerifyDuringGC) {
2047 // Verify that the counting data accumulated during marking matches
2048 // that calculated by walking the marking bitmap.
2050 // Bitmaps to hold expected values
2051 BitMap expected_region_bm(_region_bm.size(), true);
2052 BitMap expected_card_bm(_card_bm.size(), true);
2054 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2055 &_region_bm,
2056 &_card_bm,
2057 &expected_region_bm,
2058 &expected_card_bm);
2060 if (G1CollectedHeap::use_parallel_gc_threads()) {
2061 g1h->set_par_threads((int)n_workers);
2062 g1h->workers()->run_task(&g1_par_verify_task);
2063 // Done with the parallel phase so reset to 0.
2064 g1h->set_par_threads(0);
2066 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2067 "sanity check");
2068 } else {
2069 g1_par_verify_task.work(0);
2070 }
2072 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2073 }
2075 size_t start_used_bytes = g1h->used();
2076 g1h->set_marking_complete();
2078 double count_end = os::elapsedTime();
2079 double this_final_counting_time = (count_end - start);
2080 _total_counting_time += this_final_counting_time;
2082 if (G1PrintRegionLivenessInfo) {
2083 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2084 _g1h->heap_region_iterate(&cl);
2085 }
2087 // Install newly created mark bitMap as "prev".
2088 swapMarkBitMaps();
2090 g1h->reset_gc_time_stamp();
2092 // Note end of marking in all heap regions.
2093 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2094 if (G1CollectedHeap::use_parallel_gc_threads()) {
2095 g1h->set_par_threads((int)n_workers);
2096 g1h->workers()->run_task(&g1_par_note_end_task);
2097 g1h->set_par_threads(0);
2099 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2100 "sanity check");
2101 } else {
2102 g1_par_note_end_task.work(0);
2103 }
2104 g1h->check_gc_time_stamps();
2106 if (!cleanup_list_is_empty()) {
2107 // The cleanup list is not empty, so we'll have to process it
2108 // concurrently. Notify anyone else that might be wanting free
2109 // regions that there will be more free regions coming soon.
2110 g1h->set_free_regions_coming();
2111 }
2113 // call below, since it affects the metric by which we sort the heap
2114 // regions.
2115 if (G1ScrubRemSets) {
2116 double rs_scrub_start = os::elapsedTime();
2117 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2118 if (G1CollectedHeap::use_parallel_gc_threads()) {
2119 g1h->set_par_threads((int)n_workers);
2120 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2121 g1h->set_par_threads(0);
2123 assert(g1h->check_heap_region_claim_values(
2124 HeapRegion::ScrubRemSetClaimValue),
2125 "sanity check");
2126 } else {
2127 g1_par_scrub_rs_task.work(0);
2128 }
2130 double rs_scrub_end = os::elapsedTime();
2131 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2132 _total_rs_scrub_time += this_rs_scrub_time;
2133 }
2135 // this will also free any regions totally full of garbage objects,
2136 // and sort the regions.
2137 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2139 // Statistics.
2140 double end = os::elapsedTime();
2141 _cleanup_times.add((end - start) * 1000.0);
2143 if (G1Log::fine()) {
2144 g1h->print_size_transition(gclog_or_tty,
2145 start_used_bytes,
2146 g1h->used(),
2147 g1h->capacity());
2148 }
2150 // Clean up will have freed any regions completely full of garbage.
2151 // Update the soft reference policy with the new heap occupancy.
2152 Universe::update_heap_info_at_gc();
2154 if (VerifyDuringGC) {
2155 HandleMark hm; // handle scope
2156 Universe::heap()->prepare_for_verify();
2157 Universe::verify(VerifyOption_G1UsePrevMarking,
2158 " VerifyDuringGC:(after)");
2159 }
2160 g1h->check_bitmaps("Cleanup End");
2162 g1h->verify_region_sets_optional();
2164 // We need to make this be a "collection" so any collection pause that
2165 // races with it goes around and waits for completeCleanup to finish.
2166 g1h->increment_total_collections();
2168 // Clean out dead classes and update Metaspace sizes.
2169 if (ClassUnloadingWithConcurrentMark) {
2170 ClassLoaderDataGraph::purge();
2171 }
2172 MetaspaceGC::compute_new_size();
2174 // We reclaimed old regions so we should calculate the sizes to make
2175 // sure we update the old gen/space data.
2176 g1h->g1mm()->update_sizes();
2178 g1h->trace_heap_after_concurrent_cycle();
2179 }
2181 void ConcurrentMark::completeCleanup() {
2182 if (has_aborted()) return;
2184 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2186 _cleanup_list.verify_optional();
2187 FreeRegionList tmp_free_list("Tmp Free List");
2189 if (G1ConcRegionFreeingVerbose) {
2190 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2191 "cleanup list has %u entries",
2192 _cleanup_list.length());
2193 }
2195 // No one else should be accessing the _cleanup_list at this point,
2196 // so it is not necessary to take any locks
2197 while (!_cleanup_list.is_empty()) {
2198 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2199 assert(hr != NULL, "Got NULL from a non-empty list");
2200 hr->par_clear();
2201 tmp_free_list.add_ordered(hr);
2203 // Instead of adding one region at a time to the secondary_free_list,
2204 // we accumulate them in the local list and move them a few at a
2205 // time. This also cuts down on the number of notify_all() calls
2206 // we do during this process. We'll also append the local list when
2207 // _cleanup_list is empty (which means we just removed the last
2208 // region from the _cleanup_list).
2209 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2210 _cleanup_list.is_empty()) {
2211 if (G1ConcRegionFreeingVerbose) {
2212 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2213 "appending %u entries to the secondary_free_list, "
2214 "cleanup list still has %u entries",
2215 tmp_free_list.length(),
2216 _cleanup_list.length());
2217 }
2219 {
2220 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2221 g1h->secondary_free_list_add(&tmp_free_list);
2222 SecondaryFreeList_lock->notify_all();
2223 }
2225 if (G1StressConcRegionFreeing) {
2226 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2227 os::sleep(Thread::current(), (jlong) 1, false);
2228 }
2229 }
2230 }
2231 }
2232 assert(tmp_free_list.is_empty(), "post-condition");
2233 }
2235 // Supporting Object and Oop closures for reference discovery
2236 // and processing in during marking
2238 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2239 HeapWord* addr = (HeapWord*)obj;
2240 return addr != NULL &&
2241 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2242 }
2244 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2245 // Uses the CMTask associated with a worker thread (for serial reference
2246 // processing the CMTask for worker 0 is used) to preserve (mark) and
2247 // trace referent objects.
2248 //
2249 // Using the CMTask and embedded local queues avoids having the worker
2250 // threads operating on the global mark stack. This reduces the risk
2251 // of overflowing the stack - which we would rather avoid at this late
2252 // state. Also using the tasks' local queues removes the potential
2253 // of the workers interfering with each other that could occur if
2254 // operating on the global stack.
2256 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2257 ConcurrentMark* _cm;
2258 CMTask* _task;
2259 int _ref_counter_limit;
2260 int _ref_counter;
2261 bool _is_serial;
2262 public:
2263 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2264 _cm(cm), _task(task), _is_serial(is_serial),
2265 _ref_counter_limit(G1RefProcDrainInterval) {
2266 assert(_ref_counter_limit > 0, "sanity");
2267 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2268 _ref_counter = _ref_counter_limit;
2269 }
2271 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2272 virtual void do_oop( oop* p) { do_oop_work(p); }
2274 template <class T> void do_oop_work(T* p) {
2275 if (!_cm->has_overflown()) {
2276 oop obj = oopDesc::load_decode_heap_oop(p);
2277 if (_cm->verbose_high()) {
2278 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2279 "*"PTR_FORMAT" = "PTR_FORMAT,
2280 _task->worker_id(), p2i(p), p2i((void*) obj));
2281 }
2283 _task->deal_with_reference(obj);
2284 _ref_counter--;
2286 if (_ref_counter == 0) {
2287 // We have dealt with _ref_counter_limit references, pushing them
2288 // and objects reachable from them on to the local stack (and
2289 // possibly the global stack). Call CMTask::do_marking_step() to
2290 // process these entries.
2291 //
2292 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2293 // there's nothing more to do (i.e. we're done with the entries that
2294 // were pushed as a result of the CMTask::deal_with_reference() calls
2295 // above) or we overflow.
2296 //
2297 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2298 // flag while there may still be some work to do. (See the comment at
2299 // the beginning of CMTask::do_marking_step() for those conditions -
2300 // one of which is reaching the specified time target.) It is only
2301 // when CMTask::do_marking_step() returns without setting the
2302 // has_aborted() flag that the marking step has completed.
2303 do {
2304 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2305 _task->do_marking_step(mark_step_duration_ms,
2306 false /* do_termination */,
2307 _is_serial);
2308 } while (_task->has_aborted() && !_cm->has_overflown());
2309 _ref_counter = _ref_counter_limit;
2310 }
2311 } else {
2312 if (_cm->verbose_high()) {
2313 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2314 }
2315 }
2316 }
2317 };
2319 // 'Drain' oop closure used by both serial and parallel reference processing.
2320 // Uses the CMTask associated with a given worker thread (for serial
2321 // reference processing the CMtask for worker 0 is used). Calls the
2322 // do_marking_step routine, with an unbelievably large timeout value,
2323 // to drain the marking data structures of the remaining entries
2324 // added by the 'keep alive' oop closure above.
2326 class G1CMDrainMarkingStackClosure: public VoidClosure {
2327 ConcurrentMark* _cm;
2328 CMTask* _task;
2329 bool _is_serial;
2330 public:
2331 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2332 _cm(cm), _task(task), _is_serial(is_serial) {
2333 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2334 }
2336 void do_void() {
2337 do {
2338 if (_cm->verbose_high()) {
2339 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2340 _task->worker_id(), BOOL_TO_STR(_is_serial));
2341 }
2343 // We call CMTask::do_marking_step() to completely drain the local
2344 // and global marking stacks of entries pushed by the 'keep alive'
2345 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2346 //
2347 // CMTask::do_marking_step() is called in a loop, which we'll exit
2348 // if there's nothing more to do (i.e. we'completely drained the
2349 // entries that were pushed as a a result of applying the 'keep alive'
2350 // closure to the entries on the discovered ref lists) or we overflow
2351 // the global marking stack.
2352 //
2353 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2354 // flag while there may still be some work to do. (See the comment at
2355 // the beginning of CMTask::do_marking_step() for those conditions -
2356 // one of which is reaching the specified time target.) It is only
2357 // when CMTask::do_marking_step() returns without setting the
2358 // has_aborted() flag that the marking step has completed.
2360 _task->do_marking_step(1000000000.0 /* something very large */,
2361 true /* do_termination */,
2362 _is_serial);
2363 } while (_task->has_aborted() && !_cm->has_overflown());
2364 }
2365 };
2367 // Implementation of AbstractRefProcTaskExecutor for parallel
2368 // reference processing at the end of G1 concurrent marking
2370 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2371 private:
2372 G1CollectedHeap* _g1h;
2373 ConcurrentMark* _cm;
2374 WorkGang* _workers;
2375 int _active_workers;
2377 public:
2378 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2379 ConcurrentMark* cm,
2380 WorkGang* workers,
2381 int n_workers) :
2382 _g1h(g1h), _cm(cm),
2383 _workers(workers), _active_workers(n_workers) { }
2385 // Executes the given task using concurrent marking worker threads.
2386 virtual void execute(ProcessTask& task);
2387 virtual void execute(EnqueueTask& task);
2388 };
2390 class G1CMRefProcTaskProxy: public AbstractGangTask {
2391 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2392 ProcessTask& _proc_task;
2393 G1CollectedHeap* _g1h;
2394 ConcurrentMark* _cm;
2396 public:
2397 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2398 G1CollectedHeap* g1h,
2399 ConcurrentMark* cm) :
2400 AbstractGangTask("Process reference objects in parallel"),
2401 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2402 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2403 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2404 }
2406 virtual void work(uint worker_id) {
2407 ResourceMark rm;
2408 HandleMark hm;
2409 CMTask* task = _cm->task(worker_id);
2410 G1CMIsAliveClosure g1_is_alive(_g1h);
2411 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2412 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2414 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2415 }
2416 };
2418 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2419 assert(_workers != NULL, "Need parallel worker threads.");
2420 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2422 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2424 // We need to reset the concurrency level before each
2425 // proxy task execution, so that the termination protocol
2426 // and overflow handling in CMTask::do_marking_step() knows
2427 // how many workers to wait for.
2428 _cm->set_concurrency(_active_workers);
2429 _g1h->set_par_threads(_active_workers);
2430 _workers->run_task(&proc_task_proxy);
2431 _g1h->set_par_threads(0);
2432 }
2434 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2435 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2436 EnqueueTask& _enq_task;
2438 public:
2439 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2440 AbstractGangTask("Enqueue reference objects in parallel"),
2441 _enq_task(enq_task) { }
2443 virtual void work(uint worker_id) {
2444 _enq_task.work(worker_id);
2445 }
2446 };
2448 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2449 assert(_workers != NULL, "Need parallel worker threads.");
2450 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2452 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2454 // Not strictly necessary but...
2455 //
2456 // We need to reset the concurrency level before each
2457 // proxy task execution, so that the termination protocol
2458 // and overflow handling in CMTask::do_marking_step() knows
2459 // how many workers to wait for.
2460 _cm->set_concurrency(_active_workers);
2461 _g1h->set_par_threads(_active_workers);
2462 _workers->run_task(&enq_task_proxy);
2463 _g1h->set_par_threads(0);
2464 }
2466 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2467 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2468 }
2470 // Helper class to get rid of some boilerplate code.
2471 class G1RemarkGCTraceTime : public GCTraceTime {
2472 static bool doit_and_prepend(bool doit) {
2473 if (doit) {
2474 gclog_or_tty->put(' ');
2475 }
2476 return doit;
2477 }
2479 public:
2480 G1RemarkGCTraceTime(const char* title, bool doit)
2481 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
2482 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
2483 }
2484 };
2486 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2487 if (has_overflown()) {
2488 // Skip processing the discovered references if we have
2489 // overflown the global marking stack. Reference objects
2490 // only get discovered once so it is OK to not
2491 // de-populate the discovered reference lists. We could have,
2492 // but the only benefit would be that, when marking restarts,
2493 // less reference objects are discovered.
2494 return;
2495 }
2497 ResourceMark rm;
2498 HandleMark hm;
2500 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2502 // Is alive closure.
2503 G1CMIsAliveClosure g1_is_alive(g1h);
2505 // Inner scope to exclude the cleaning of the string and symbol
2506 // tables from the displayed time.
2507 {
2508 if (G1Log::finer()) {
2509 gclog_or_tty->put(' ');
2510 }
2511 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm(), concurrent_gc_id());
2513 ReferenceProcessor* rp = g1h->ref_processor_cm();
2515 // See the comment in G1CollectedHeap::ref_processing_init()
2516 // about how reference processing currently works in G1.
2518 // Set the soft reference policy
2519 rp->setup_policy(clear_all_soft_refs);
2520 assert(_markStack.isEmpty(), "mark stack should be empty");
2522 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2523 // in serial reference processing. Note these closures are also
2524 // used for serially processing (by the the current thread) the
2525 // JNI references during parallel reference processing.
2526 //
2527 // These closures do not need to synchronize with the worker
2528 // threads involved in parallel reference processing as these
2529 // instances are executed serially by the current thread (e.g.
2530 // reference processing is not multi-threaded and is thus
2531 // performed by the current thread instead of a gang worker).
2532 //
2533 // The gang tasks involved in parallel reference procssing create
2534 // their own instances of these closures, which do their own
2535 // synchronization among themselves.
2536 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2537 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2539 // We need at least one active thread. If reference processing
2540 // is not multi-threaded we use the current (VMThread) thread,
2541 // otherwise we use the work gang from the G1CollectedHeap and
2542 // we utilize all the worker threads we can.
2543 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2544 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2545 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2547 // Parallel processing task executor.
2548 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2549 g1h->workers(), active_workers);
2550 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2552 // Set the concurrency level. The phase was already set prior to
2553 // executing the remark task.
2554 set_concurrency(active_workers);
2556 // Set the degree of MT processing here. If the discovery was done MT,
2557 // the number of threads involved during discovery could differ from
2558 // the number of active workers. This is OK as long as the discovered
2559 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2560 rp->set_active_mt_degree(active_workers);
2562 // Process the weak references.
2563 const ReferenceProcessorStats& stats =
2564 rp->process_discovered_references(&g1_is_alive,
2565 &g1_keep_alive,
2566 &g1_drain_mark_stack,
2567 executor,
2568 g1h->gc_timer_cm(),
2569 concurrent_gc_id());
2570 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2572 // The do_oop work routines of the keep_alive and drain_marking_stack
2573 // oop closures will set the has_overflown flag if we overflow the
2574 // global marking stack.
2576 assert(_markStack.overflow() || _markStack.isEmpty(),
2577 "mark stack should be empty (unless it overflowed)");
2579 if (_markStack.overflow()) {
2580 // This should have been done already when we tried to push an
2581 // entry on to the global mark stack. But let's do it again.
2582 set_has_overflown();
2583 }
2585 assert(rp->num_q() == active_workers, "why not");
2587 rp->enqueue_discovered_references(executor);
2589 rp->verify_no_references_recorded();
2590 assert(!rp->discovery_enabled(), "Post condition");
2591 }
2593 if (has_overflown()) {
2594 // We can not trust g1_is_alive if the marking stack overflowed
2595 return;
2596 }
2598 assert(_markStack.isEmpty(), "Marking should have completed");
2600 // Unload Klasses, String, Symbols, Code Cache, etc.
2601 {
2602 G1RemarkGCTraceTime trace("Unloading", G1Log::finer());
2604 if (ClassUnloadingWithConcurrentMark) {
2605 bool purged_classes;
2607 {
2608 G1RemarkGCTraceTime trace("System Dictionary Unloading", G1Log::finest());
2609 purged_classes = SystemDictionary::do_unloading(&g1_is_alive);
2610 }
2612 {
2613 G1RemarkGCTraceTime trace("Parallel Unloading", G1Log::finest());
2614 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2615 }
2616 }
2618 if (G1StringDedup::is_enabled()) {
2619 G1RemarkGCTraceTime trace("String Deduplication Unlink", G1Log::finest());
2620 G1StringDedup::unlink(&g1_is_alive);
2621 }
2622 }
2623 }
2625 void ConcurrentMark::swapMarkBitMaps() {
2626 CMBitMapRO* temp = _prevMarkBitMap;
2627 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2628 _nextMarkBitMap = (CMBitMap*) temp;
2629 }
2631 class CMObjectClosure;
2633 // Closure for iterating over objects, currently only used for
2634 // processing SATB buffers.
2635 class CMObjectClosure : public ObjectClosure {
2636 private:
2637 CMTask* _task;
2639 public:
2640 void do_object(oop obj) {
2641 _task->deal_with_reference(obj);
2642 }
2644 CMObjectClosure(CMTask* task) : _task(task) { }
2645 };
2647 class G1RemarkThreadsClosure : public ThreadClosure {
2648 CMObjectClosure _cm_obj;
2649 G1CMOopClosure _cm_cl;
2650 MarkingCodeBlobClosure _code_cl;
2651 int _thread_parity;
2652 bool _is_par;
2654 public:
2655 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) :
2656 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2657 _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {}
2659 void do_thread(Thread* thread) {
2660 if (thread->is_Java_thread()) {
2661 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2662 JavaThread* jt = (JavaThread*)thread;
2664 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2665 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2666 // * Alive if on the stack of an executing method
2667 // * Weakly reachable otherwise
2668 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2669 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2670 jt->nmethods_do(&_code_cl);
2672 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2673 }
2674 } else if (thread->is_VM_thread()) {
2675 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2676 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2677 }
2678 }
2679 }
2680 };
2682 class CMRemarkTask: public AbstractGangTask {
2683 private:
2684 ConcurrentMark* _cm;
2685 bool _is_serial;
2686 public:
2687 void work(uint worker_id) {
2688 // Since all available tasks are actually started, we should
2689 // only proceed if we're supposed to be actived.
2690 if (worker_id < _cm->active_tasks()) {
2691 CMTask* task = _cm->task(worker_id);
2692 task->record_start_time();
2693 {
2694 ResourceMark rm;
2695 HandleMark hm;
2697 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial);
2698 Threads::threads_do(&threads_f);
2699 }
2701 do {
2702 task->do_marking_step(1000000000.0 /* something very large */,
2703 true /* do_termination */,
2704 _is_serial);
2705 } while (task->has_aborted() && !_cm->has_overflown());
2706 // If we overflow, then we do not want to restart. We instead
2707 // want to abort remark and do concurrent marking again.
2708 task->record_end_time();
2709 }
2710 }
2712 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2713 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2714 _cm->terminator()->reset_for_reuse(active_workers);
2715 }
2716 };
2718 void ConcurrentMark::checkpointRootsFinalWork() {
2719 ResourceMark rm;
2720 HandleMark hm;
2721 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2723 G1RemarkGCTraceTime trace("Finalize Marking", G1Log::finer());
2725 g1h->ensure_parsability(false);
2727 if (G1CollectedHeap::use_parallel_gc_threads()) {
2728 G1CollectedHeap::StrongRootsScope srs(g1h);
2729 // this is remark, so we'll use up all active threads
2730 uint active_workers = g1h->workers()->active_workers();
2731 if (active_workers == 0) {
2732 assert(active_workers > 0, "Should have been set earlier");
2733 active_workers = (uint) ParallelGCThreads;
2734 g1h->workers()->set_active_workers(active_workers);
2735 }
2736 set_concurrency_and_phase(active_workers, false /* concurrent */);
2737 // Leave _parallel_marking_threads at it's
2738 // value originally calculated in the ConcurrentMark
2739 // constructor and pass values of the active workers
2740 // through the gang in the task.
2742 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2743 // We will start all available threads, even if we decide that the
2744 // active_workers will be fewer. The extra ones will just bail out
2745 // immediately.
2746 g1h->set_par_threads(active_workers);
2747 g1h->workers()->run_task(&remarkTask);
2748 g1h->set_par_threads(0);
2749 } else {
2750 G1CollectedHeap::StrongRootsScope srs(g1h);
2751 uint active_workers = 1;
2752 set_concurrency_and_phase(active_workers, false /* concurrent */);
2754 // Note - if there's no work gang then the VMThread will be
2755 // the thread to execute the remark - serially. We have
2756 // to pass true for the is_serial parameter so that
2757 // CMTask::do_marking_step() doesn't enter the sync
2758 // barriers in the event of an overflow. Doing so will
2759 // cause an assert that the current thread is not a
2760 // concurrent GC thread.
2761 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2762 remarkTask.work(0);
2763 }
2764 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2765 guarantee(has_overflown() ||
2766 satb_mq_set.completed_buffers_num() == 0,
2767 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2768 BOOL_TO_STR(has_overflown()),
2769 satb_mq_set.completed_buffers_num()));
2771 print_stats();
2772 }
2774 #ifndef PRODUCT
2776 class PrintReachableOopClosure: public OopClosure {
2777 private:
2778 G1CollectedHeap* _g1h;
2779 outputStream* _out;
2780 VerifyOption _vo;
2781 bool _all;
2783 public:
2784 PrintReachableOopClosure(outputStream* out,
2785 VerifyOption vo,
2786 bool all) :
2787 _g1h(G1CollectedHeap::heap()),
2788 _out(out), _vo(vo), _all(all) { }
2790 void do_oop(narrowOop* p) { do_oop_work(p); }
2791 void do_oop( oop* p) { do_oop_work(p); }
2793 template <class T> void do_oop_work(T* p) {
2794 oop obj = oopDesc::load_decode_heap_oop(p);
2795 const char* str = NULL;
2796 const char* str2 = "";
2798 if (obj == NULL) {
2799 str = "";
2800 } else if (!_g1h->is_in_g1_reserved(obj)) {
2801 str = " O";
2802 } else {
2803 HeapRegion* hr = _g1h->heap_region_containing(obj);
2804 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2805 bool marked = _g1h->is_marked(obj, _vo);
2807 if (over_tams) {
2808 str = " >";
2809 if (marked) {
2810 str2 = " AND MARKED";
2811 }
2812 } else if (marked) {
2813 str = " M";
2814 } else {
2815 str = " NOT";
2816 }
2817 }
2819 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2820 p2i(p), p2i((void*) obj), str, str2);
2821 }
2822 };
2824 class PrintReachableObjectClosure : public ObjectClosure {
2825 private:
2826 G1CollectedHeap* _g1h;
2827 outputStream* _out;
2828 VerifyOption _vo;
2829 bool _all;
2830 HeapRegion* _hr;
2832 public:
2833 PrintReachableObjectClosure(outputStream* out,
2834 VerifyOption vo,
2835 bool all,
2836 HeapRegion* hr) :
2837 _g1h(G1CollectedHeap::heap()),
2838 _out(out), _vo(vo), _all(all), _hr(hr) { }
2840 void do_object(oop o) {
2841 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2842 bool marked = _g1h->is_marked(o, _vo);
2843 bool print_it = _all || over_tams || marked;
2845 if (print_it) {
2846 _out->print_cr(" "PTR_FORMAT"%s",
2847 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2848 PrintReachableOopClosure oopCl(_out, _vo, _all);
2849 o->oop_iterate_no_header(&oopCl);
2850 }
2851 }
2852 };
2854 class PrintReachableRegionClosure : public HeapRegionClosure {
2855 private:
2856 G1CollectedHeap* _g1h;
2857 outputStream* _out;
2858 VerifyOption _vo;
2859 bool _all;
2861 public:
2862 bool doHeapRegion(HeapRegion* hr) {
2863 HeapWord* b = hr->bottom();
2864 HeapWord* e = hr->end();
2865 HeapWord* t = hr->top();
2866 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2867 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2868 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2869 _out->cr();
2871 HeapWord* from = b;
2872 HeapWord* to = t;
2874 if (to > from) {
2875 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2876 _out->cr();
2877 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2878 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2879 _out->cr();
2880 }
2882 return false;
2883 }
2885 PrintReachableRegionClosure(outputStream* out,
2886 VerifyOption vo,
2887 bool all) :
2888 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2889 };
2891 void ConcurrentMark::print_reachable(const char* str,
2892 VerifyOption vo,
2893 bool all) {
2894 gclog_or_tty->cr();
2895 gclog_or_tty->print_cr("== Doing heap dump... ");
2897 if (G1PrintReachableBaseFile == NULL) {
2898 gclog_or_tty->print_cr(" #### error: no base file defined");
2899 return;
2900 }
2902 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2903 (JVM_MAXPATHLEN - 1)) {
2904 gclog_or_tty->print_cr(" #### error: file name too long");
2905 return;
2906 }
2908 char file_name[JVM_MAXPATHLEN];
2909 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2910 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2912 fileStream fout(file_name);
2913 if (!fout.is_open()) {
2914 gclog_or_tty->print_cr(" #### error: could not open file");
2915 return;
2916 }
2918 outputStream* out = &fout;
2919 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2920 out->cr();
2922 out->print_cr("--- ITERATING OVER REGIONS");
2923 out->cr();
2924 PrintReachableRegionClosure rcl(out, vo, all);
2925 _g1h->heap_region_iterate(&rcl);
2926 out->cr();
2928 gclog_or_tty->print_cr(" done");
2929 gclog_or_tty->flush();
2930 }
2932 #endif // PRODUCT
2934 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2935 // Note we are overriding the read-only view of the prev map here, via
2936 // the cast.
2937 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2938 }
2940 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2941 _nextMarkBitMap->clearRange(mr);
2942 }
2944 HeapRegion*
2945 ConcurrentMark::claim_region(uint worker_id) {
2946 // "checkpoint" the finger
2947 HeapWord* finger = _finger;
2949 // _heap_end will not change underneath our feet; it only changes at
2950 // yield points.
2951 while (finger < _heap_end) {
2952 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2954 // Note on how this code handles humongous regions. In the
2955 // normal case the finger will reach the start of a "starts
2956 // humongous" (SH) region. Its end will either be the end of the
2957 // last "continues humongous" (CH) region in the sequence, or the
2958 // standard end of the SH region (if the SH is the only region in
2959 // the sequence). That way claim_region() will skip over the CH
2960 // regions. However, there is a subtle race between a CM thread
2961 // executing this method and a mutator thread doing a humongous
2962 // object allocation. The two are not mutually exclusive as the CM
2963 // thread does not need to hold the Heap_lock when it gets
2964 // here. So there is a chance that claim_region() will come across
2965 // a free region that's in the progress of becoming a SH or a CH
2966 // region. In the former case, it will either
2967 // a) Miss the update to the region's end, in which case it will
2968 // visit every subsequent CH region, will find their bitmaps
2969 // empty, and do nothing, or
2970 // b) Will observe the update of the region's end (in which case
2971 // it will skip the subsequent CH regions).
2972 // If it comes across a region that suddenly becomes CH, the
2973 // scenario will be similar to b). So, the race between
2974 // claim_region() and a humongous object allocation might force us
2975 // to do a bit of unnecessary work (due to some unnecessary bitmap
2976 // iterations) but it should not introduce and correctness issues.
2977 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2979 // Above heap_region_containing_raw may return NULL as we always scan claim
2980 // until the end of the heap. In this case, just jump to the next region.
2981 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2983 // Is the gap between reading the finger and doing the CAS too long?
2984 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2985 if (res == finger && curr_region != NULL) {
2986 // we succeeded
2987 HeapWord* bottom = curr_region->bottom();
2988 HeapWord* limit = curr_region->next_top_at_mark_start();
2990 if (verbose_low()) {
2991 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2992 "["PTR_FORMAT", "PTR_FORMAT"), "
2993 "limit = "PTR_FORMAT,
2994 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2995 }
2997 // notice that _finger == end cannot be guaranteed here since,
2998 // someone else might have moved the finger even further
2999 assert(_finger >= end, "the finger should have moved forward");
3001 if (verbose_low()) {
3002 gclog_or_tty->print_cr("[%u] we were successful with region = "
3003 PTR_FORMAT, worker_id, p2i(curr_region));
3004 }
3006 if (limit > bottom) {
3007 if (verbose_low()) {
3008 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
3009 "returning it ", worker_id, p2i(curr_region));
3010 }
3011 return curr_region;
3012 } else {
3013 assert(limit == bottom,
3014 "the region limit should be at bottom");
3015 if (verbose_low()) {
3016 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
3017 "returning NULL", worker_id, p2i(curr_region));
3018 }
3019 // we return NULL and the caller should try calling
3020 // claim_region() again.
3021 return NULL;
3022 }
3023 } else {
3024 assert(_finger > finger, "the finger should have moved forward");
3025 if (verbose_low()) {
3026 if (curr_region == NULL) {
3027 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
3028 "global finger = "PTR_FORMAT", "
3029 "our finger = "PTR_FORMAT,
3030 worker_id, p2i(_finger), p2i(finger));
3031 } else {
3032 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
3033 "global finger = "PTR_FORMAT", "
3034 "our finger = "PTR_FORMAT,
3035 worker_id, p2i(_finger), p2i(finger));
3036 }
3037 }
3039 // read it again
3040 finger = _finger;
3041 }
3042 }
3044 return NULL;
3045 }
3047 #ifndef PRODUCT
3048 enum VerifyNoCSetOopsPhase {
3049 VerifyNoCSetOopsStack,
3050 VerifyNoCSetOopsQueues,
3051 VerifyNoCSetOopsSATBCompleted,
3052 VerifyNoCSetOopsSATBThread
3053 };
3055 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
3056 private:
3057 G1CollectedHeap* _g1h;
3058 VerifyNoCSetOopsPhase _phase;
3059 int _info;
3061 const char* phase_str() {
3062 switch (_phase) {
3063 case VerifyNoCSetOopsStack: return "Stack";
3064 case VerifyNoCSetOopsQueues: return "Queue";
3065 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
3066 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
3067 default: ShouldNotReachHere();
3068 }
3069 return NULL;
3070 }
3072 void do_object_work(oop obj) {
3073 guarantee(!_g1h->obj_in_cs(obj),
3074 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
3075 p2i((void*) obj), phase_str(), _info));
3076 }
3078 public:
3079 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3081 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3082 _phase = phase;
3083 _info = info;
3084 }
3086 virtual void do_oop(oop* p) {
3087 oop obj = oopDesc::load_decode_heap_oop(p);
3088 do_object_work(obj);
3089 }
3091 virtual void do_oop(narrowOop* p) {
3092 // We should not come across narrow oops while scanning marking
3093 // stacks and SATB buffers.
3094 ShouldNotReachHere();
3095 }
3097 virtual void do_object(oop obj) {
3098 do_object_work(obj);
3099 }
3100 };
3102 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3103 bool verify_enqueued_buffers,
3104 bool verify_thread_buffers,
3105 bool verify_fingers) {
3106 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3107 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3108 return;
3109 }
3111 VerifyNoCSetOopsClosure cl;
3113 if (verify_stacks) {
3114 // Verify entries on the global mark stack
3115 cl.set_phase(VerifyNoCSetOopsStack);
3116 _markStack.oops_do(&cl);
3118 // Verify entries on the task queues
3119 for (uint i = 0; i < _max_worker_id; i += 1) {
3120 cl.set_phase(VerifyNoCSetOopsQueues, i);
3121 CMTaskQueue* queue = _task_queues->queue(i);
3122 queue->oops_do(&cl);
3123 }
3124 }
3126 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3128 // Verify entries on the enqueued SATB buffers
3129 if (verify_enqueued_buffers) {
3130 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3131 satb_qs.iterate_completed_buffers_read_only(&cl);
3132 }
3134 // Verify entries on the per-thread SATB buffers
3135 if (verify_thread_buffers) {
3136 cl.set_phase(VerifyNoCSetOopsSATBThread);
3137 satb_qs.iterate_thread_buffers_read_only(&cl);
3138 }
3140 if (verify_fingers) {
3141 // Verify the global finger
3142 HeapWord* global_finger = finger();
3143 if (global_finger != NULL && global_finger < _heap_end) {
3144 // The global finger always points to a heap region boundary. We
3145 // use heap_region_containing_raw() to get the containing region
3146 // given that the global finger could be pointing to a free region
3147 // which subsequently becomes continues humongous. If that
3148 // happens, heap_region_containing() will return the bottom of the
3149 // corresponding starts humongous region and the check below will
3150 // not hold any more.
3151 // Since we always iterate over all regions, we might get a NULL HeapRegion
3152 // here.
3153 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3154 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3155 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3156 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3157 }
3159 // Verify the task fingers
3160 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3161 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3162 CMTask* task = _tasks[i];
3163 HeapWord* task_finger = task->finger();
3164 if (task_finger != NULL && task_finger < _heap_end) {
3165 // See above note on the global finger verification.
3166 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3167 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3168 !task_hr->in_collection_set(),
3169 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3170 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3171 }
3172 }
3173 }
3174 }
3175 #endif // PRODUCT
3177 // Aggregate the counting data that was constructed concurrently
3178 // with marking.
3179 class AggregateCountDataHRClosure: public HeapRegionClosure {
3180 G1CollectedHeap* _g1h;
3181 ConcurrentMark* _cm;
3182 CardTableModRefBS* _ct_bs;
3183 BitMap* _cm_card_bm;
3184 uint _max_worker_id;
3186 public:
3187 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3188 BitMap* cm_card_bm,
3189 uint max_worker_id) :
3190 _g1h(g1h), _cm(g1h->concurrent_mark()),
3191 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3192 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3194 bool doHeapRegion(HeapRegion* hr) {
3195 if (hr->continuesHumongous()) {
3196 // We will ignore these here and process them when their
3197 // associated "starts humongous" region is processed.
3198 // Note that we cannot rely on their associated
3199 // "starts humongous" region to have their bit set to 1
3200 // since, due to the region chunking in the parallel region
3201 // iteration, a "continues humongous" region might be visited
3202 // before its associated "starts humongous".
3203 return false;
3204 }
3206 HeapWord* start = hr->bottom();
3207 HeapWord* limit = hr->next_top_at_mark_start();
3208 HeapWord* end = hr->end();
3210 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3211 err_msg("Preconditions not met - "
3212 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3213 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3214 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3216 assert(hr->next_marked_bytes() == 0, "Precondition");
3218 if (start == limit) {
3219 // NTAMS of this region has not been set so nothing to do.
3220 return false;
3221 }
3223 // 'start' should be in the heap.
3224 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3225 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3226 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3228 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3229 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3230 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3232 // If ntams is not card aligned then we bump card bitmap index
3233 // for limit so that we get the all the cards spanned by
3234 // the object ending at ntams.
3235 // Note: if this is the last region in the heap then ntams
3236 // could be actually just beyond the end of the the heap;
3237 // limit_idx will then correspond to a (non-existent) card
3238 // that is also outside the heap.
3239 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3240 limit_idx += 1;
3241 }
3243 assert(limit_idx <= end_idx, "or else use atomics");
3245 // Aggregate the "stripe" in the count data associated with hr.
3246 uint hrm_index = hr->hrm_index();
3247 size_t marked_bytes = 0;
3249 for (uint i = 0; i < _max_worker_id; i += 1) {
3250 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3251 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3253 // Fetch the marked_bytes in this region for task i and
3254 // add it to the running total for this region.
3255 marked_bytes += marked_bytes_array[hrm_index];
3257 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3258 // into the global card bitmap.
3259 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3261 while (scan_idx < limit_idx) {
3262 assert(task_card_bm->at(scan_idx) == true, "should be");
3263 _cm_card_bm->set_bit(scan_idx);
3264 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3266 // BitMap::get_next_one_offset() can handle the case when
3267 // its left_offset parameter is greater than its right_offset
3268 // parameter. It does, however, have an early exit if
3269 // left_offset == right_offset. So let's limit the value
3270 // passed in for left offset here.
3271 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3272 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3273 }
3274 }
3276 // Update the marked bytes for this region.
3277 hr->add_to_marked_bytes(marked_bytes);
3279 // Next heap region
3280 return false;
3281 }
3282 };
3284 class G1AggregateCountDataTask: public AbstractGangTask {
3285 protected:
3286 G1CollectedHeap* _g1h;
3287 ConcurrentMark* _cm;
3288 BitMap* _cm_card_bm;
3289 uint _max_worker_id;
3290 int _active_workers;
3292 public:
3293 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3294 ConcurrentMark* cm,
3295 BitMap* cm_card_bm,
3296 uint max_worker_id,
3297 int n_workers) :
3298 AbstractGangTask("Count Aggregation"),
3299 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3300 _max_worker_id(max_worker_id),
3301 _active_workers(n_workers) { }
3303 void work(uint worker_id) {
3304 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3306 if (G1CollectedHeap::use_parallel_gc_threads()) {
3307 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3308 _active_workers,
3309 HeapRegion::AggregateCountClaimValue);
3310 } else {
3311 _g1h->heap_region_iterate(&cl);
3312 }
3313 }
3314 };
3317 void ConcurrentMark::aggregate_count_data() {
3318 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3319 _g1h->workers()->active_workers() :
3320 1);
3322 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3323 _max_worker_id, n_workers);
3325 if (G1CollectedHeap::use_parallel_gc_threads()) {
3326 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3327 "sanity check");
3328 _g1h->set_par_threads(n_workers);
3329 _g1h->workers()->run_task(&g1_par_agg_task);
3330 _g1h->set_par_threads(0);
3332 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3333 "sanity check");
3334 _g1h->reset_heap_region_claim_values();
3335 } else {
3336 g1_par_agg_task.work(0);
3337 }
3338 _g1h->allocation_context_stats().update_at_remark();
3339 }
3341 // Clear the per-worker arrays used to store the per-region counting data
3342 void ConcurrentMark::clear_all_count_data() {
3343 // Clear the global card bitmap - it will be filled during
3344 // liveness count aggregation (during remark) and the
3345 // final counting task.
3346 _card_bm.clear();
3348 // Clear the global region bitmap - it will be filled as part
3349 // of the final counting task.
3350 _region_bm.clear();
3352 uint max_regions = _g1h->max_regions();
3353 assert(_max_worker_id > 0, "uninitialized");
3355 for (uint i = 0; i < _max_worker_id; i += 1) {
3356 BitMap* task_card_bm = count_card_bitmap_for(i);
3357 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3359 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3360 assert(marked_bytes_array != NULL, "uninitialized");
3362 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3363 task_card_bm->clear();
3364 }
3365 }
3367 void ConcurrentMark::print_stats() {
3368 if (verbose_stats()) {
3369 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3370 for (size_t i = 0; i < _active_tasks; ++i) {
3371 _tasks[i]->print_stats();
3372 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3373 }
3374 }
3375 }
3377 // abandon current marking iteration due to a Full GC
3378 void ConcurrentMark::abort() {
3379 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3380 // concurrent bitmap clearing.
3381 _nextMarkBitMap->clearAll();
3383 // Note we cannot clear the previous marking bitmap here
3384 // since VerifyDuringGC verifies the objects marked during
3385 // a full GC against the previous bitmap.
3387 // Clear the liveness counting data
3388 clear_all_count_data();
3389 // Empty mark stack
3390 reset_marking_state();
3391 for (uint i = 0; i < _max_worker_id; ++i) {
3392 _tasks[i]->clear_region_fields();
3393 }
3394 _first_overflow_barrier_sync.abort();
3395 _second_overflow_barrier_sync.abort();
3396 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3397 if (!gc_id.is_undefined()) {
3398 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3399 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3400 _aborted_gc_id = gc_id;
3401 }
3402 _has_aborted = true;
3404 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3405 satb_mq_set.abandon_partial_marking();
3406 // This can be called either during or outside marking, we'll read
3407 // the expected_active value from the SATB queue set.
3408 satb_mq_set.set_active_all_threads(
3409 false, /* new active value */
3410 satb_mq_set.is_active() /* expected_active */);
3412 _g1h->trace_heap_after_concurrent_cycle();
3413 _g1h->register_concurrent_cycle_end();
3414 }
3416 const GCId& ConcurrentMark::concurrent_gc_id() {
3417 if (has_aborted()) {
3418 return _aborted_gc_id;
3419 }
3420 return _g1h->gc_tracer_cm()->gc_id();
3421 }
3423 static void print_ms_time_info(const char* prefix, const char* name,
3424 NumberSeq& ns) {
3425 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3426 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3427 if (ns.num() > 0) {
3428 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3429 prefix, ns.sd(), ns.maximum());
3430 }
3431 }
3433 void ConcurrentMark::print_summary_info() {
3434 gclog_or_tty->print_cr(" Concurrent marking:");
3435 print_ms_time_info(" ", "init marks", _init_times);
3436 print_ms_time_info(" ", "remarks", _remark_times);
3437 {
3438 print_ms_time_info(" ", "final marks", _remark_mark_times);
3439 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3441 }
3442 print_ms_time_info(" ", "cleanups", _cleanup_times);
3443 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3444 _total_counting_time,
3445 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3446 (double)_cleanup_times.num()
3447 : 0.0));
3448 if (G1ScrubRemSets) {
3449 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3450 _total_rs_scrub_time,
3451 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3452 (double)_cleanup_times.num()
3453 : 0.0));
3454 }
3455 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3456 (_init_times.sum() + _remark_times.sum() +
3457 _cleanup_times.sum())/1000.0);
3458 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3459 "(%8.2f s marking).",
3460 cmThread()->vtime_accum(),
3461 cmThread()->vtime_mark_accum());
3462 }
3464 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3465 if (use_parallel_marking_threads()) {
3466 _parallel_workers->print_worker_threads_on(st);
3467 }
3468 }
3470 void ConcurrentMark::print_on_error(outputStream* st) const {
3471 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3472 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3473 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3474 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3475 }
3477 // We take a break if someone is trying to stop the world.
3478 bool ConcurrentMark::do_yield_check(uint worker_id) {
3479 if (SuspendibleThreadSet::should_yield()) {
3480 if (worker_id == 0) {
3481 _g1h->g1_policy()->record_concurrent_pause();
3482 }
3483 SuspendibleThreadSet::yield();
3484 return true;
3485 } else {
3486 return false;
3487 }
3488 }
3490 #ifndef PRODUCT
3491 // for debugging purposes
3492 void ConcurrentMark::print_finger() {
3493 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3494 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3495 for (uint i = 0; i < _max_worker_id; ++i) {
3496 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3497 }
3498 gclog_or_tty->cr();
3499 }
3500 #endif
3502 void CMTask::scan_object(oop obj) {
3503 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3505 if (_cm->verbose_high()) {
3506 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3507 _worker_id, p2i((void*) obj));
3508 }
3510 size_t obj_size = obj->size();
3511 _words_scanned += obj_size;
3513 obj->oop_iterate(_cm_oop_closure);
3514 statsOnly( ++_objs_scanned );
3515 check_limits();
3516 }
3518 // Closure for iteration over bitmaps
3519 class CMBitMapClosure : public BitMapClosure {
3520 private:
3521 // the bitmap that is being iterated over
3522 CMBitMap* _nextMarkBitMap;
3523 ConcurrentMark* _cm;
3524 CMTask* _task;
3526 public:
3527 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3528 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3530 bool do_bit(size_t offset) {
3531 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3532 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3533 assert( addr < _cm->finger(), "invariant");
3535 statsOnly( _task->increase_objs_found_on_bitmap() );
3536 assert(addr >= _task->finger(), "invariant");
3538 // We move that task's local finger along.
3539 _task->move_finger_to(addr);
3541 _task->scan_object(oop(addr));
3542 // we only partially drain the local queue and global stack
3543 _task->drain_local_queue(true);
3544 _task->drain_global_stack(true);
3546 // if the has_aborted flag has been raised, we need to bail out of
3547 // the iteration
3548 return !_task->has_aborted();
3549 }
3550 };
3552 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3553 ConcurrentMark* cm,
3554 CMTask* task)
3555 : _g1h(g1h), _cm(cm), _task(task) {
3556 assert(_ref_processor == NULL, "should be initialized to NULL");
3558 if (G1UseConcMarkReferenceProcessing) {
3559 _ref_processor = g1h->ref_processor_cm();
3560 assert(_ref_processor != NULL, "should not be NULL");
3561 }
3562 }
3564 void CMTask::setup_for_region(HeapRegion* hr) {
3565 assert(hr != NULL,
3566 "claim_region() should have filtered out NULL regions");
3567 assert(!hr->continuesHumongous(),
3568 "claim_region() should have filtered out continues humongous regions");
3570 if (_cm->verbose_low()) {
3571 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3572 _worker_id, p2i(hr));
3573 }
3575 _curr_region = hr;
3576 _finger = hr->bottom();
3577 update_region_limit();
3578 }
3580 void CMTask::update_region_limit() {
3581 HeapRegion* hr = _curr_region;
3582 HeapWord* bottom = hr->bottom();
3583 HeapWord* limit = hr->next_top_at_mark_start();
3585 if (limit == bottom) {
3586 if (_cm->verbose_low()) {
3587 gclog_or_tty->print_cr("[%u] found an empty region "
3588 "["PTR_FORMAT", "PTR_FORMAT")",
3589 _worker_id, p2i(bottom), p2i(limit));
3590 }
3591 // The region was collected underneath our feet.
3592 // We set the finger to bottom to ensure that the bitmap
3593 // iteration that will follow this will not do anything.
3594 // (this is not a condition that holds when we set the region up,
3595 // as the region is not supposed to be empty in the first place)
3596 _finger = bottom;
3597 } else if (limit >= _region_limit) {
3598 assert(limit >= _finger, "peace of mind");
3599 } else {
3600 assert(limit < _region_limit, "only way to get here");
3601 // This can happen under some pretty unusual circumstances. An
3602 // evacuation pause empties the region underneath our feet (NTAMS
3603 // at bottom). We then do some allocation in the region (NTAMS
3604 // stays at bottom), followed by the region being used as a GC
3605 // alloc region (NTAMS will move to top() and the objects
3606 // originally below it will be grayed). All objects now marked in
3607 // the region are explicitly grayed, if below the global finger,
3608 // and we do not need in fact to scan anything else. So, we simply
3609 // set _finger to be limit to ensure that the bitmap iteration
3610 // doesn't do anything.
3611 _finger = limit;
3612 }
3614 _region_limit = limit;
3615 }
3617 void CMTask::giveup_current_region() {
3618 assert(_curr_region != NULL, "invariant");
3619 if (_cm->verbose_low()) {
3620 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3621 _worker_id, p2i(_curr_region));
3622 }
3623 clear_region_fields();
3624 }
3626 void CMTask::clear_region_fields() {
3627 // Values for these three fields that indicate that we're not
3628 // holding on to a region.
3629 _curr_region = NULL;
3630 _finger = NULL;
3631 _region_limit = NULL;
3632 }
3634 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3635 if (cm_oop_closure == NULL) {
3636 assert(_cm_oop_closure != NULL, "invariant");
3637 } else {
3638 assert(_cm_oop_closure == NULL, "invariant");
3639 }
3640 _cm_oop_closure = cm_oop_closure;
3641 }
3643 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3644 guarantee(nextMarkBitMap != NULL, "invariant");
3646 if (_cm->verbose_low()) {
3647 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3648 }
3650 _nextMarkBitMap = nextMarkBitMap;
3651 clear_region_fields();
3653 _calls = 0;
3654 _elapsed_time_ms = 0.0;
3655 _termination_time_ms = 0.0;
3656 _termination_start_time_ms = 0.0;
3658 #if _MARKING_STATS_
3659 _local_pushes = 0;
3660 _local_pops = 0;
3661 _local_max_size = 0;
3662 _objs_scanned = 0;
3663 _global_pushes = 0;
3664 _global_pops = 0;
3665 _global_max_size = 0;
3666 _global_transfers_to = 0;
3667 _global_transfers_from = 0;
3668 _regions_claimed = 0;
3669 _objs_found_on_bitmap = 0;
3670 _satb_buffers_processed = 0;
3671 _steal_attempts = 0;
3672 _steals = 0;
3673 _aborted = 0;
3674 _aborted_overflow = 0;
3675 _aborted_cm_aborted = 0;
3676 _aborted_yield = 0;
3677 _aborted_timed_out = 0;
3678 _aborted_satb = 0;
3679 _aborted_termination = 0;
3680 #endif // _MARKING_STATS_
3681 }
3683 bool CMTask::should_exit_termination() {
3684 regular_clock_call();
3685 // This is called when we are in the termination protocol. We should
3686 // quit if, for some reason, this task wants to abort or the global
3687 // stack is not empty (this means that we can get work from it).
3688 return !_cm->mark_stack_empty() || has_aborted();
3689 }
3691 void CMTask::reached_limit() {
3692 assert(_words_scanned >= _words_scanned_limit ||
3693 _refs_reached >= _refs_reached_limit ,
3694 "shouldn't have been called otherwise");
3695 regular_clock_call();
3696 }
3698 void CMTask::regular_clock_call() {
3699 if (has_aborted()) return;
3701 // First, we need to recalculate the words scanned and refs reached
3702 // limits for the next clock call.
3703 recalculate_limits();
3705 // During the regular clock call we do the following
3707 // (1) If an overflow has been flagged, then we abort.
3708 if (_cm->has_overflown()) {
3709 set_has_aborted();
3710 return;
3711 }
3713 // If we are not concurrent (i.e. we're doing remark) we don't need
3714 // to check anything else. The other steps are only needed during
3715 // the concurrent marking phase.
3716 if (!concurrent()) return;
3718 // (2) If marking has been aborted for Full GC, then we also abort.
3719 if (_cm->has_aborted()) {
3720 set_has_aborted();
3721 statsOnly( ++_aborted_cm_aborted );
3722 return;
3723 }
3725 double curr_time_ms = os::elapsedVTime() * 1000.0;
3727 // (3) If marking stats are enabled, then we update the step history.
3728 #if _MARKING_STATS_
3729 if (_words_scanned >= _words_scanned_limit) {
3730 ++_clock_due_to_scanning;
3731 }
3732 if (_refs_reached >= _refs_reached_limit) {
3733 ++_clock_due_to_marking;
3734 }
3736 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3737 _interval_start_time_ms = curr_time_ms;
3738 _all_clock_intervals_ms.add(last_interval_ms);
3740 if (_cm->verbose_medium()) {
3741 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3742 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3743 _worker_id, last_interval_ms,
3744 _words_scanned,
3745 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3746 _refs_reached,
3747 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3748 }
3749 #endif // _MARKING_STATS_
3751 // (4) We check whether we should yield. If we have to, then we abort.
3752 if (SuspendibleThreadSet::should_yield()) {
3753 // We should yield. To do this we abort the task. The caller is
3754 // responsible for yielding.
3755 set_has_aborted();
3756 statsOnly( ++_aborted_yield );
3757 return;
3758 }
3760 // (5) We check whether we've reached our time quota. If we have,
3761 // then we abort.
3762 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3763 if (elapsed_time_ms > _time_target_ms) {
3764 set_has_aborted();
3765 _has_timed_out = true;
3766 statsOnly( ++_aborted_timed_out );
3767 return;
3768 }
3770 // (6) Finally, we check whether there are enough completed STAB
3771 // buffers available for processing. If there are, we abort.
3772 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3773 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3774 if (_cm->verbose_low()) {
3775 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3776 _worker_id);
3777 }
3778 // we do need to process SATB buffers, we'll abort and restart
3779 // the marking task to do so
3780 set_has_aborted();
3781 statsOnly( ++_aborted_satb );
3782 return;
3783 }
3784 }
3786 void CMTask::recalculate_limits() {
3787 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3788 _words_scanned_limit = _real_words_scanned_limit;
3790 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3791 _refs_reached_limit = _real_refs_reached_limit;
3792 }
3794 void CMTask::decrease_limits() {
3795 // This is called when we believe that we're going to do an infrequent
3796 // operation which will increase the per byte scanned cost (i.e. move
3797 // entries to/from the global stack). It basically tries to decrease the
3798 // scanning limit so that the clock is called earlier.
3800 if (_cm->verbose_medium()) {
3801 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3802 }
3804 _words_scanned_limit = _real_words_scanned_limit -
3805 3 * words_scanned_period / 4;
3806 _refs_reached_limit = _real_refs_reached_limit -
3807 3 * refs_reached_period / 4;
3808 }
3810 void CMTask::move_entries_to_global_stack() {
3811 // local array where we'll store the entries that will be popped
3812 // from the local queue
3813 oop buffer[global_stack_transfer_size];
3815 int n = 0;
3816 oop obj;
3817 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3818 buffer[n] = obj;
3819 ++n;
3820 }
3822 if (n > 0) {
3823 // we popped at least one entry from the local queue
3825 statsOnly( ++_global_transfers_to; _local_pops += n );
3827 if (!_cm->mark_stack_push(buffer, n)) {
3828 if (_cm->verbose_low()) {
3829 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3830 _worker_id);
3831 }
3832 set_has_aborted();
3833 } else {
3834 // the transfer was successful
3836 if (_cm->verbose_medium()) {
3837 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3838 _worker_id, n);
3839 }
3840 statsOnly( int tmp_size = _cm->mark_stack_size();
3841 if (tmp_size > _global_max_size) {
3842 _global_max_size = tmp_size;
3843 }
3844 _global_pushes += n );
3845 }
3846 }
3848 // this operation was quite expensive, so decrease the limits
3849 decrease_limits();
3850 }
3852 void CMTask::get_entries_from_global_stack() {
3853 // local array where we'll store the entries that will be popped
3854 // from the global stack.
3855 oop buffer[global_stack_transfer_size];
3856 int n;
3857 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3858 assert(n <= global_stack_transfer_size,
3859 "we should not pop more than the given limit");
3860 if (n > 0) {
3861 // yes, we did actually pop at least one entry
3863 statsOnly( ++_global_transfers_from; _global_pops += n );
3864 if (_cm->verbose_medium()) {
3865 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3866 _worker_id, n);
3867 }
3868 for (int i = 0; i < n; ++i) {
3869 bool success = _task_queue->push(buffer[i]);
3870 // We only call this when the local queue is empty or under a
3871 // given target limit. So, we do not expect this push to fail.
3872 assert(success, "invariant");
3873 }
3875 statsOnly( int tmp_size = _task_queue->size();
3876 if (tmp_size > _local_max_size) {
3877 _local_max_size = tmp_size;
3878 }
3879 _local_pushes += n );
3880 }
3882 // this operation was quite expensive, so decrease the limits
3883 decrease_limits();
3884 }
3886 void CMTask::drain_local_queue(bool partially) {
3887 if (has_aborted()) return;
3889 // Decide what the target size is, depending whether we're going to
3890 // drain it partially (so that other tasks can steal if they run out
3891 // of things to do) or totally (at the very end).
3892 size_t target_size;
3893 if (partially) {
3894 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3895 } else {
3896 target_size = 0;
3897 }
3899 if (_task_queue->size() > target_size) {
3900 if (_cm->verbose_high()) {
3901 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3902 _worker_id, target_size);
3903 }
3905 oop obj;
3906 bool ret = _task_queue->pop_local(obj);
3907 while (ret) {
3908 statsOnly( ++_local_pops );
3910 if (_cm->verbose_high()) {
3911 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3912 p2i((void*) obj));
3913 }
3915 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3916 assert(!_g1h->is_on_master_free_list(
3917 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3919 scan_object(obj);
3921 if (_task_queue->size() <= target_size || has_aborted()) {
3922 ret = false;
3923 } else {
3924 ret = _task_queue->pop_local(obj);
3925 }
3926 }
3928 if (_cm->verbose_high()) {
3929 gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3930 _worker_id, _task_queue->size());
3931 }
3932 }
3933 }
3935 void CMTask::drain_global_stack(bool partially) {
3936 if (has_aborted()) return;
3938 // We have a policy to drain the local queue before we attempt to
3939 // drain the global stack.
3940 assert(partially || _task_queue->size() == 0, "invariant");
3942 // Decide what the target size is, depending whether we're going to
3943 // drain it partially (so that other tasks can steal if they run out
3944 // of things to do) or totally (at the very end). Notice that,
3945 // because we move entries from the global stack in chunks or
3946 // because another task might be doing the same, we might in fact
3947 // drop below the target. But, this is not a problem.
3948 size_t target_size;
3949 if (partially) {
3950 target_size = _cm->partial_mark_stack_size_target();
3951 } else {
3952 target_size = 0;
3953 }
3955 if (_cm->mark_stack_size() > target_size) {
3956 if (_cm->verbose_low()) {
3957 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3958 _worker_id, target_size);
3959 }
3961 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3962 get_entries_from_global_stack();
3963 drain_local_queue(partially);
3964 }
3966 if (_cm->verbose_low()) {
3967 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3968 _worker_id, _cm->mark_stack_size());
3969 }
3970 }
3971 }
3973 // SATB Queue has several assumptions on whether to call the par or
3974 // non-par versions of the methods. this is why some of the code is
3975 // replicated. We should really get rid of the single-threaded version
3976 // of the code to simplify things.
3977 void CMTask::drain_satb_buffers() {
3978 if (has_aborted()) return;
3980 // We set this so that the regular clock knows that we're in the
3981 // middle of draining buffers and doesn't set the abort flag when it
3982 // notices that SATB buffers are available for draining. It'd be
3983 // very counter productive if it did that. :-)
3984 _draining_satb_buffers = true;
3986 CMObjectClosure oc(this);
3987 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3988 if (G1CollectedHeap::use_parallel_gc_threads()) {
3989 satb_mq_set.set_par_closure(_worker_id, &oc);
3990 } else {
3991 satb_mq_set.set_closure(&oc);
3992 }
3994 // This keeps claiming and applying the closure to completed buffers
3995 // until we run out of buffers or we need to abort.
3996 if (G1CollectedHeap::use_parallel_gc_threads()) {
3997 while (!has_aborted() &&
3998 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
3999 if (_cm->verbose_medium()) {
4000 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4001 }
4002 statsOnly( ++_satb_buffers_processed );
4003 regular_clock_call();
4004 }
4005 } else {
4006 while (!has_aborted() &&
4007 satb_mq_set.apply_closure_to_completed_buffer()) {
4008 if (_cm->verbose_medium()) {
4009 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4010 }
4011 statsOnly( ++_satb_buffers_processed );
4012 regular_clock_call();
4013 }
4014 }
4016 _draining_satb_buffers = false;
4018 assert(has_aborted() ||
4019 concurrent() ||
4020 satb_mq_set.completed_buffers_num() == 0, "invariant");
4022 if (G1CollectedHeap::use_parallel_gc_threads()) {
4023 satb_mq_set.set_par_closure(_worker_id, NULL);
4024 } else {
4025 satb_mq_set.set_closure(NULL);
4026 }
4028 // again, this was a potentially expensive operation, decrease the
4029 // limits to get the regular clock call early
4030 decrease_limits();
4031 }
4033 void CMTask::print_stats() {
4034 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
4035 _worker_id, _calls);
4036 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
4037 _elapsed_time_ms, _termination_time_ms);
4038 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4039 _step_times_ms.num(), _step_times_ms.avg(),
4040 _step_times_ms.sd());
4041 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4042 _step_times_ms.maximum(), _step_times_ms.sum());
4044 #if _MARKING_STATS_
4045 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4046 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
4047 _all_clock_intervals_ms.sd());
4048 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4049 _all_clock_intervals_ms.maximum(),
4050 _all_clock_intervals_ms.sum());
4051 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
4052 _clock_due_to_scanning, _clock_due_to_marking);
4053 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
4054 _objs_scanned, _objs_found_on_bitmap);
4055 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
4056 _local_pushes, _local_pops, _local_max_size);
4057 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
4058 _global_pushes, _global_pops, _global_max_size);
4059 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
4060 _global_transfers_to,_global_transfers_from);
4061 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
4062 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
4063 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
4064 _steal_attempts, _steals);
4065 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
4066 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
4067 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
4068 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
4069 _aborted_timed_out, _aborted_satb, _aborted_termination);
4070 #endif // _MARKING_STATS_
4071 }
4073 /*****************************************************************************
4075 The do_marking_step(time_target_ms, ...) method is the building
4076 block of the parallel marking framework. It can be called in parallel
4077 with other invocations of do_marking_step() on different tasks
4078 (but only one per task, obviously) and concurrently with the
4079 mutator threads, or during remark, hence it eliminates the need
4080 for two versions of the code. When called during remark, it will
4081 pick up from where the task left off during the concurrent marking
4082 phase. Interestingly, tasks are also claimable during evacuation
4083 pauses too, since do_marking_step() ensures that it aborts before
4084 it needs to yield.
4086 The data structures that it uses to do marking work are the
4087 following:
4089 (1) Marking Bitmap. If there are gray objects that appear only
4090 on the bitmap (this happens either when dealing with an overflow
4091 or when the initial marking phase has simply marked the roots
4092 and didn't push them on the stack), then tasks claim heap
4093 regions whose bitmap they then scan to find gray objects. A
4094 global finger indicates where the end of the last claimed region
4095 is. A local finger indicates how far into the region a task has
4096 scanned. The two fingers are used to determine how to gray an
4097 object (i.e. whether simply marking it is OK, as it will be
4098 visited by a task in the future, or whether it needs to be also
4099 pushed on a stack).
4101 (2) Local Queue. The local queue of the task which is accessed
4102 reasonably efficiently by the task. Other tasks can steal from
4103 it when they run out of work. Throughout the marking phase, a
4104 task attempts to keep its local queue short but not totally
4105 empty, so that entries are available for stealing by other
4106 tasks. Only when there is no more work, a task will totally
4107 drain its local queue.
4109 (3) Global Mark Stack. This handles local queue overflow. During
4110 marking only sets of entries are moved between it and the local
4111 queues, as access to it requires a mutex and more fine-grain
4112 interaction with it which might cause contention. If it
4113 overflows, then the marking phase should restart and iterate
4114 over the bitmap to identify gray objects. Throughout the marking
4115 phase, tasks attempt to keep the global mark stack at a small
4116 length but not totally empty, so that entries are available for
4117 popping by other tasks. Only when there is no more work, tasks
4118 will totally drain the global mark stack.
4120 (4) SATB Buffer Queue. This is where completed SATB buffers are
4121 made available. Buffers are regularly removed from this queue
4122 and scanned for roots, so that the queue doesn't get too
4123 long. During remark, all completed buffers are processed, as
4124 well as the filled in parts of any uncompleted buffers.
4126 The do_marking_step() method tries to abort when the time target
4127 has been reached. There are a few other cases when the
4128 do_marking_step() method also aborts:
4130 (1) When the marking phase has been aborted (after a Full GC).
4132 (2) When a global overflow (on the global stack) has been
4133 triggered. Before the task aborts, it will actually sync up with
4134 the other tasks to ensure that all the marking data structures
4135 (local queues, stacks, fingers etc.) are re-initialized so that
4136 when do_marking_step() completes, the marking phase can
4137 immediately restart.
4139 (3) When enough completed SATB buffers are available. The
4140 do_marking_step() method only tries to drain SATB buffers right
4141 at the beginning. So, if enough buffers are available, the
4142 marking step aborts and the SATB buffers are processed at
4143 the beginning of the next invocation.
4145 (4) To yield. when we have to yield then we abort and yield
4146 right at the end of do_marking_step(). This saves us from a lot
4147 of hassle as, by yielding we might allow a Full GC. If this
4148 happens then objects will be compacted underneath our feet, the
4149 heap might shrink, etc. We save checking for this by just
4150 aborting and doing the yield right at the end.
4152 From the above it follows that the do_marking_step() method should
4153 be called in a loop (or, otherwise, regularly) until it completes.
4155 If a marking step completes without its has_aborted() flag being
4156 true, it means it has completed the current marking phase (and
4157 also all other marking tasks have done so and have all synced up).
4159 A method called regular_clock_call() is invoked "regularly" (in
4160 sub ms intervals) throughout marking. It is this clock method that
4161 checks all the abort conditions which were mentioned above and
4162 decides when the task should abort. A work-based scheme is used to
4163 trigger this clock method: when the number of object words the
4164 marking phase has scanned or the number of references the marking
4165 phase has visited reach a given limit. Additional invocations to
4166 the method clock have been planted in a few other strategic places
4167 too. The initial reason for the clock method was to avoid calling
4168 vtime too regularly, as it is quite expensive. So, once it was in
4169 place, it was natural to piggy-back all the other conditions on it
4170 too and not constantly check them throughout the code.
4172 If do_termination is true then do_marking_step will enter its
4173 termination protocol.
4175 The value of is_serial must be true when do_marking_step is being
4176 called serially (i.e. by the VMThread) and do_marking_step should
4177 skip any synchronization in the termination and overflow code.
4178 Examples include the serial remark code and the serial reference
4179 processing closures.
4181 The value of is_serial must be false when do_marking_step is
4182 being called by any of the worker threads in a work gang.
4183 Examples include the concurrent marking code (CMMarkingTask),
4184 the MT remark code, and the MT reference processing closures.
4186 *****************************************************************************/
4188 void CMTask::do_marking_step(double time_target_ms,
4189 bool do_termination,
4190 bool is_serial) {
4191 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4192 assert(concurrent() == _cm->concurrent(), "they should be the same");
4194 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4195 assert(_task_queues != NULL, "invariant");
4196 assert(_task_queue != NULL, "invariant");
4197 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4199 assert(!_claimed,
4200 "only one thread should claim this task at any one time");
4202 // OK, this doesn't safeguard again all possible scenarios, as it is
4203 // possible for two threads to set the _claimed flag at the same
4204 // time. But it is only for debugging purposes anyway and it will
4205 // catch most problems.
4206 _claimed = true;
4208 _start_time_ms = os::elapsedVTime() * 1000.0;
4209 statsOnly( _interval_start_time_ms = _start_time_ms );
4211 // If do_stealing is true then do_marking_step will attempt to
4212 // steal work from the other CMTasks. It only makes sense to
4213 // enable stealing when the termination protocol is enabled
4214 // and do_marking_step() is not being called serially.
4215 bool do_stealing = do_termination && !is_serial;
4217 double diff_prediction_ms =
4218 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4219 _time_target_ms = time_target_ms - diff_prediction_ms;
4221 // set up the variables that are used in the work-based scheme to
4222 // call the regular clock method
4223 _words_scanned = 0;
4224 _refs_reached = 0;
4225 recalculate_limits();
4227 // clear all flags
4228 clear_has_aborted();
4229 _has_timed_out = false;
4230 _draining_satb_buffers = false;
4232 ++_calls;
4234 if (_cm->verbose_low()) {
4235 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4236 "target = %1.2lfms >>>>>>>>>>",
4237 _worker_id, _calls, _time_target_ms);
4238 }
4240 // Set up the bitmap and oop closures. Anything that uses them is
4241 // eventually called from this method, so it is OK to allocate these
4242 // statically.
4243 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4244 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4245 set_cm_oop_closure(&cm_oop_closure);
4247 if (_cm->has_overflown()) {
4248 // This can happen if the mark stack overflows during a GC pause
4249 // and this task, after a yield point, restarts. We have to abort
4250 // as we need to get into the overflow protocol which happens
4251 // right at the end of this task.
4252 set_has_aborted();
4253 }
4255 // First drain any available SATB buffers. After this, we will not
4256 // look at SATB buffers before the next invocation of this method.
4257 // If enough completed SATB buffers are queued up, the regular clock
4258 // will abort this task so that it restarts.
4259 drain_satb_buffers();
4260 // ...then partially drain the local queue and the global stack
4261 drain_local_queue(true);
4262 drain_global_stack(true);
4264 do {
4265 if (!has_aborted() && _curr_region != NULL) {
4266 // This means that we're already holding on to a region.
4267 assert(_finger != NULL, "if region is not NULL, then the finger "
4268 "should not be NULL either");
4270 // We might have restarted this task after an evacuation pause
4271 // which might have evacuated the region we're holding on to
4272 // underneath our feet. Let's read its limit again to make sure
4273 // that we do not iterate over a region of the heap that
4274 // contains garbage (update_region_limit() will also move
4275 // _finger to the start of the region if it is found empty).
4276 update_region_limit();
4277 // We will start from _finger not from the start of the region,
4278 // as we might be restarting this task after aborting half-way
4279 // through scanning this region. In this case, _finger points to
4280 // the address where we last found a marked object. If this is a
4281 // fresh region, _finger points to start().
4282 MemRegion mr = MemRegion(_finger, _region_limit);
4284 if (_cm->verbose_low()) {
4285 gclog_or_tty->print_cr("[%u] we're scanning part "
4286 "["PTR_FORMAT", "PTR_FORMAT") "
4287 "of region "HR_FORMAT,
4288 _worker_id, p2i(_finger), p2i(_region_limit),
4289 HR_FORMAT_PARAMS(_curr_region));
4290 }
4292 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4293 "humongous regions should go around loop once only");
4295 // Some special cases:
4296 // If the memory region is empty, we can just give up the region.
4297 // If the current region is humongous then we only need to check
4298 // the bitmap for the bit associated with the start of the object,
4299 // scan the object if it's live, and give up the region.
4300 // Otherwise, let's iterate over the bitmap of the part of the region
4301 // that is left.
4302 // If the iteration is successful, give up the region.
4303 if (mr.is_empty()) {
4304 giveup_current_region();
4305 regular_clock_call();
4306 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4307 if (_nextMarkBitMap->isMarked(mr.start())) {
4308 // The object is marked - apply the closure
4309 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4310 bitmap_closure.do_bit(offset);
4311 }
4312 // Even if this task aborted while scanning the humongous object
4313 // we can (and should) give up the current region.
4314 giveup_current_region();
4315 regular_clock_call();
4316 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4317 giveup_current_region();
4318 regular_clock_call();
4319 } else {
4320 assert(has_aborted(), "currently the only way to do so");
4321 // The only way to abort the bitmap iteration is to return
4322 // false from the do_bit() method. However, inside the
4323 // do_bit() method we move the _finger to point to the
4324 // object currently being looked at. So, if we bail out, we
4325 // have definitely set _finger to something non-null.
4326 assert(_finger != NULL, "invariant");
4328 // Region iteration was actually aborted. So now _finger
4329 // points to the address of the object we last scanned. If we
4330 // leave it there, when we restart this task, we will rescan
4331 // the object. It is easy to avoid this. We move the finger by
4332 // enough to point to the next possible object header (the
4333 // bitmap knows by how much we need to move it as it knows its
4334 // granularity).
4335 assert(_finger < _region_limit, "invariant");
4336 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4337 // Check if bitmap iteration was aborted while scanning the last object
4338 if (new_finger >= _region_limit) {
4339 giveup_current_region();
4340 } else {
4341 move_finger_to(new_finger);
4342 }
4343 }
4344 }
4345 // At this point we have either completed iterating over the
4346 // region we were holding on to, or we have aborted.
4348 // We then partially drain the local queue and the global stack.
4349 // (Do we really need this?)
4350 drain_local_queue(true);
4351 drain_global_stack(true);
4353 // Read the note on the claim_region() method on why it might
4354 // return NULL with potentially more regions available for
4355 // claiming and why we have to check out_of_regions() to determine
4356 // whether we're done or not.
4357 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4358 // We are going to try to claim a new region. We should have
4359 // given up on the previous one.
4360 // Separated the asserts so that we know which one fires.
4361 assert(_curr_region == NULL, "invariant");
4362 assert(_finger == NULL, "invariant");
4363 assert(_region_limit == NULL, "invariant");
4364 if (_cm->verbose_low()) {
4365 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4366 }
4367 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4368 if (claimed_region != NULL) {
4369 // Yes, we managed to claim one
4370 statsOnly( ++_regions_claimed );
4372 if (_cm->verbose_low()) {
4373 gclog_or_tty->print_cr("[%u] we successfully claimed "
4374 "region "PTR_FORMAT,
4375 _worker_id, p2i(claimed_region));
4376 }
4378 setup_for_region(claimed_region);
4379 assert(_curr_region == claimed_region, "invariant");
4380 }
4381 // It is important to call the regular clock here. It might take
4382 // a while to claim a region if, for example, we hit a large
4383 // block of empty regions. So we need to call the regular clock
4384 // method once round the loop to make sure it's called
4385 // frequently enough.
4386 regular_clock_call();
4387 }
4389 if (!has_aborted() && _curr_region == NULL) {
4390 assert(_cm->out_of_regions(),
4391 "at this point we should be out of regions");
4392 }
4393 } while ( _curr_region != NULL && !has_aborted());
4395 if (!has_aborted()) {
4396 // We cannot check whether the global stack is empty, since other
4397 // tasks might be pushing objects to it concurrently.
4398 assert(_cm->out_of_regions(),
4399 "at this point we should be out of regions");
4401 if (_cm->verbose_low()) {
4402 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4403 }
4405 // Try to reduce the number of available SATB buffers so that
4406 // remark has less work to do.
4407 drain_satb_buffers();
4408 }
4410 // Since we've done everything else, we can now totally drain the
4411 // local queue and global stack.
4412 drain_local_queue(false);
4413 drain_global_stack(false);
4415 // Attempt at work stealing from other task's queues.
4416 if (do_stealing && !has_aborted()) {
4417 // We have not aborted. This means that we have finished all that
4418 // we could. Let's try to do some stealing...
4420 // We cannot check whether the global stack is empty, since other
4421 // tasks might be pushing objects to it concurrently.
4422 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4423 "only way to reach here");
4425 if (_cm->verbose_low()) {
4426 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4427 }
4429 while (!has_aborted()) {
4430 oop obj;
4431 statsOnly( ++_steal_attempts );
4433 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4434 if (_cm->verbose_medium()) {
4435 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4436 _worker_id, p2i((void*) obj));
4437 }
4439 statsOnly( ++_steals );
4441 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4442 "any stolen object should be marked");
4443 scan_object(obj);
4445 // And since we're towards the end, let's totally drain the
4446 // local queue and global stack.
4447 drain_local_queue(false);
4448 drain_global_stack(false);
4449 } else {
4450 break;
4451 }
4452 }
4453 }
4455 // If we are about to wrap up and go into termination, check if we
4456 // should raise the overflow flag.
4457 if (do_termination && !has_aborted()) {
4458 if (_cm->force_overflow()->should_force()) {
4459 _cm->set_has_overflown();
4460 regular_clock_call();
4461 }
4462 }
4464 // We still haven't aborted. Now, let's try to get into the
4465 // termination protocol.
4466 if (do_termination && !has_aborted()) {
4467 // We cannot check whether the global stack is empty, since other
4468 // tasks might be concurrently pushing objects on it.
4469 // Separated the asserts so that we know which one fires.
4470 assert(_cm->out_of_regions(), "only way to reach here");
4471 assert(_task_queue->size() == 0, "only way to reach here");
4473 if (_cm->verbose_low()) {
4474 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4475 }
4477 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4479 // The CMTask class also extends the TerminatorTerminator class,
4480 // hence its should_exit_termination() method will also decide
4481 // whether to exit the termination protocol or not.
4482 bool finished = (is_serial ||
4483 _cm->terminator()->offer_termination(this));
4484 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4485 _termination_time_ms +=
4486 termination_end_time_ms - _termination_start_time_ms;
4488 if (finished) {
4489 // We're all done.
4491 if (_worker_id == 0) {
4492 // let's allow task 0 to do this
4493 if (concurrent()) {
4494 assert(_cm->concurrent_marking_in_progress(), "invariant");
4495 // we need to set this to false before the next
4496 // safepoint. This way we ensure that the marking phase
4497 // doesn't observe any more heap expansions.
4498 _cm->clear_concurrent_marking_in_progress();
4499 }
4500 }
4502 // We can now guarantee that the global stack is empty, since
4503 // all other tasks have finished. We separated the guarantees so
4504 // that, if a condition is false, we can immediately find out
4505 // which one.
4506 guarantee(_cm->out_of_regions(), "only way to reach here");
4507 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4508 guarantee(_task_queue->size() == 0, "only way to reach here");
4509 guarantee(!_cm->has_overflown(), "only way to reach here");
4510 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4512 if (_cm->verbose_low()) {
4513 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4514 }
4515 } else {
4516 // Apparently there's more work to do. Let's abort this task. It
4517 // will restart it and we can hopefully find more things to do.
4519 if (_cm->verbose_low()) {
4520 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4521 _worker_id);
4522 }
4524 set_has_aborted();
4525 statsOnly( ++_aborted_termination );
4526 }
4527 }
4529 // Mainly for debugging purposes to make sure that a pointer to the
4530 // closure which was statically allocated in this frame doesn't
4531 // escape it by accident.
4532 set_cm_oop_closure(NULL);
4533 double end_time_ms = os::elapsedVTime() * 1000.0;
4534 double elapsed_time_ms = end_time_ms - _start_time_ms;
4535 // Update the step history.
4536 _step_times_ms.add(elapsed_time_ms);
4538 if (has_aborted()) {
4539 // The task was aborted for some reason.
4541 statsOnly( ++_aborted );
4543 if (_has_timed_out) {
4544 double diff_ms = elapsed_time_ms - _time_target_ms;
4545 // Keep statistics of how well we did with respect to hitting
4546 // our target only if we actually timed out (if we aborted for
4547 // other reasons, then the results might get skewed).
4548 _marking_step_diffs_ms.add(diff_ms);
4549 }
4551 if (_cm->has_overflown()) {
4552 // This is the interesting one. We aborted because a global
4553 // overflow was raised. This means we have to restart the
4554 // marking phase and start iterating over regions. However, in
4555 // order to do this we have to make sure that all tasks stop
4556 // what they are doing and re-initialise in a safe manner. We
4557 // will achieve this with the use of two barrier sync points.
4559 if (_cm->verbose_low()) {
4560 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4561 }
4563 if (!is_serial) {
4564 // We only need to enter the sync barrier if being called
4565 // from a parallel context
4566 _cm->enter_first_sync_barrier(_worker_id);
4568 // When we exit this sync barrier we know that all tasks have
4569 // stopped doing marking work. So, it's now safe to
4570 // re-initialise our data structures. At the end of this method,
4571 // task 0 will clear the global data structures.
4572 }
4574 statsOnly( ++_aborted_overflow );
4576 // We clear the local state of this task...
4577 clear_region_fields();
4579 if (!is_serial) {
4580 // ...and enter the second barrier.
4581 _cm->enter_second_sync_barrier(_worker_id);
4582 }
4583 // At this point, if we're during the concurrent phase of
4584 // marking, everything has been re-initialized and we're
4585 // ready to restart.
4586 }
4588 if (_cm->verbose_low()) {
4589 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4590 "elapsed = %1.2lfms <<<<<<<<<<",
4591 _worker_id, _time_target_ms, elapsed_time_ms);
4592 if (_cm->has_aborted()) {
4593 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4594 _worker_id);
4595 }
4596 }
4597 } else {
4598 if (_cm->verbose_low()) {
4599 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4600 "elapsed = %1.2lfms <<<<<<<<<<",
4601 _worker_id, _time_target_ms, elapsed_time_ms);
4602 }
4603 }
4605 _claimed = false;
4606 }
4608 CMTask::CMTask(uint worker_id,
4609 ConcurrentMark* cm,
4610 size_t* marked_bytes,
4611 BitMap* card_bm,
4612 CMTaskQueue* task_queue,
4613 CMTaskQueueSet* task_queues)
4614 : _g1h(G1CollectedHeap::heap()),
4615 _worker_id(worker_id), _cm(cm),
4616 _claimed(false),
4617 _nextMarkBitMap(NULL), _hash_seed(17),
4618 _task_queue(task_queue),
4619 _task_queues(task_queues),
4620 _cm_oop_closure(NULL),
4621 _marked_bytes_array(marked_bytes),
4622 _card_bm(card_bm) {
4623 guarantee(task_queue != NULL, "invariant");
4624 guarantee(task_queues != NULL, "invariant");
4626 statsOnly( _clock_due_to_scanning = 0;
4627 _clock_due_to_marking = 0 );
4629 _marking_step_diffs_ms.add(0.5);
4630 }
4632 // These are formatting macros that are used below to ensure
4633 // consistent formatting. The *_H_* versions are used to format the
4634 // header for a particular value and they should be kept consistent
4635 // with the corresponding macro. Also note that most of the macros add
4636 // the necessary white space (as a prefix) which makes them a bit
4637 // easier to compose.
4639 // All the output lines are prefixed with this string to be able to
4640 // identify them easily in a large log file.
4641 #define G1PPRL_LINE_PREFIX "###"
4643 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4644 #ifdef _LP64
4645 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4646 #else // _LP64
4647 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4648 #endif // _LP64
4650 // For per-region info
4651 #define G1PPRL_TYPE_FORMAT " %-4s"
4652 #define G1PPRL_TYPE_H_FORMAT " %4s"
4653 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4654 #define G1PPRL_BYTE_H_FORMAT " %9s"
4655 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4656 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4658 // For summary info
4659 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4660 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4661 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4662 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4664 G1PrintRegionLivenessInfoClosure::
4665 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4666 : _out(out),
4667 _total_used_bytes(0), _total_capacity_bytes(0),
4668 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4669 _hum_used_bytes(0), _hum_capacity_bytes(0),
4670 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4671 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4672 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4673 MemRegion g1_reserved = g1h->g1_reserved();
4674 double now = os::elapsedTime();
4676 // Print the header of the output.
4677 _out->cr();
4678 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4679 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4680 G1PPRL_SUM_ADDR_FORMAT("reserved")
4681 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4682 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4683 HeapRegion::GrainBytes);
4684 _out->print_cr(G1PPRL_LINE_PREFIX);
4685 _out->print_cr(G1PPRL_LINE_PREFIX
4686 G1PPRL_TYPE_H_FORMAT
4687 G1PPRL_ADDR_BASE_H_FORMAT
4688 G1PPRL_BYTE_H_FORMAT
4689 G1PPRL_BYTE_H_FORMAT
4690 G1PPRL_BYTE_H_FORMAT
4691 G1PPRL_DOUBLE_H_FORMAT
4692 G1PPRL_BYTE_H_FORMAT
4693 G1PPRL_BYTE_H_FORMAT,
4694 "type", "address-range",
4695 "used", "prev-live", "next-live", "gc-eff",
4696 "remset", "code-roots");
4697 _out->print_cr(G1PPRL_LINE_PREFIX
4698 G1PPRL_TYPE_H_FORMAT
4699 G1PPRL_ADDR_BASE_H_FORMAT
4700 G1PPRL_BYTE_H_FORMAT
4701 G1PPRL_BYTE_H_FORMAT
4702 G1PPRL_BYTE_H_FORMAT
4703 G1PPRL_DOUBLE_H_FORMAT
4704 G1PPRL_BYTE_H_FORMAT
4705 G1PPRL_BYTE_H_FORMAT,
4706 "", "",
4707 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4708 "(bytes)", "(bytes)");
4709 }
4711 // It takes as a parameter a reference to one of the _hum_* fields, it
4712 // deduces the corresponding value for a region in a humongous region
4713 // series (either the region size, or what's left if the _hum_* field
4714 // is < the region size), and updates the _hum_* field accordingly.
4715 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4716 size_t bytes = 0;
4717 // The > 0 check is to deal with the prev and next live bytes which
4718 // could be 0.
4719 if (*hum_bytes > 0) {
4720 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4721 *hum_bytes -= bytes;
4722 }
4723 return bytes;
4724 }
4726 // It deduces the values for a region in a humongous region series
4727 // from the _hum_* fields and updates those accordingly. It assumes
4728 // that that _hum_* fields have already been set up from the "starts
4729 // humongous" region and we visit the regions in address order.
4730 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4731 size_t* capacity_bytes,
4732 size_t* prev_live_bytes,
4733 size_t* next_live_bytes) {
4734 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4735 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4736 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4737 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4738 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4739 }
4741 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4742 const char* type = r->get_type_str();
4743 HeapWord* bottom = r->bottom();
4744 HeapWord* end = r->end();
4745 size_t capacity_bytes = r->capacity();
4746 size_t used_bytes = r->used();
4747 size_t prev_live_bytes = r->live_bytes();
4748 size_t next_live_bytes = r->next_live_bytes();
4749 double gc_eff = r->gc_efficiency();
4750 size_t remset_bytes = r->rem_set()->mem_size();
4751 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4753 if (r->startsHumongous()) {
4754 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4755 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4756 "they should have been zeroed after the last time we used them");
4757 // Set up the _hum_* fields.
4758 _hum_capacity_bytes = capacity_bytes;
4759 _hum_used_bytes = used_bytes;
4760 _hum_prev_live_bytes = prev_live_bytes;
4761 _hum_next_live_bytes = next_live_bytes;
4762 get_hum_bytes(&used_bytes, &capacity_bytes,
4763 &prev_live_bytes, &next_live_bytes);
4764 end = bottom + HeapRegion::GrainWords;
4765 } else if (r->continuesHumongous()) {
4766 get_hum_bytes(&used_bytes, &capacity_bytes,
4767 &prev_live_bytes, &next_live_bytes);
4768 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4769 }
4771 _total_used_bytes += used_bytes;
4772 _total_capacity_bytes += capacity_bytes;
4773 _total_prev_live_bytes += prev_live_bytes;
4774 _total_next_live_bytes += next_live_bytes;
4775 _total_remset_bytes += remset_bytes;
4776 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4778 // Print a line for this particular region.
4779 _out->print_cr(G1PPRL_LINE_PREFIX
4780 G1PPRL_TYPE_FORMAT
4781 G1PPRL_ADDR_BASE_FORMAT
4782 G1PPRL_BYTE_FORMAT
4783 G1PPRL_BYTE_FORMAT
4784 G1PPRL_BYTE_FORMAT
4785 G1PPRL_DOUBLE_FORMAT
4786 G1PPRL_BYTE_FORMAT
4787 G1PPRL_BYTE_FORMAT,
4788 type, p2i(bottom), p2i(end),
4789 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4790 remset_bytes, strong_code_roots_bytes);
4792 return false;
4793 }
4795 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4796 // add static memory usages to remembered set sizes
4797 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4798 // Print the footer of the output.
4799 _out->print_cr(G1PPRL_LINE_PREFIX);
4800 _out->print_cr(G1PPRL_LINE_PREFIX
4801 " SUMMARY"
4802 G1PPRL_SUM_MB_FORMAT("capacity")
4803 G1PPRL_SUM_MB_PERC_FORMAT("used")
4804 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4805 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4806 G1PPRL_SUM_MB_FORMAT("remset")
4807 G1PPRL_SUM_MB_FORMAT("code-roots"),
4808 bytes_to_mb(_total_capacity_bytes),
4809 bytes_to_mb(_total_used_bytes),
4810 perc(_total_used_bytes, _total_capacity_bytes),
4811 bytes_to_mb(_total_prev_live_bytes),
4812 perc(_total_prev_live_bytes, _total_capacity_bytes),
4813 bytes_to_mb(_total_next_live_bytes),
4814 perc(_total_next_live_bytes, _total_capacity_bytes),
4815 bytes_to_mb(_total_remset_bytes),
4816 bytes_to_mb(_total_strong_code_roots_bytes));
4817 _out->cr();
4818 }