Mon, 19 Aug 2019 10:11:31 +0200
8229401: Fix JFR code cache test failures
8223689: Add JFR Thread Sampling Support
8223690: Add JFR BiasedLock Event Support
8223691: Add JFR G1 Region Type Change Event Support
8223692: Add JFR G1 Heap Summary Event Support
Summary: Backport JFR from JDK11, additional fixes
Reviewed-by: neugens, apetushkov
Contributed-by: denghui.ddh@alibaba-inc.com
1 /*
2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "classfile/metadataOnStackMark.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/g1/concurrentMark.inline.hpp"
30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
32 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
33 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
34 #include "gc_implementation/g1/g1Log.hpp"
35 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
36 #include "gc_implementation/g1/g1RemSet.hpp"
37 #include "gc_implementation/g1/heapRegion.inline.hpp"
38 #include "gc_implementation/g1/heapRegionManager.inline.hpp"
39 #include "gc_implementation/g1/heapRegionRemSet.hpp"
40 #include "gc_implementation/g1/heapRegionSet.inline.hpp"
41 #include "gc_implementation/shared/vmGCOperations.hpp"
42 #include "gc_implementation/shared/gcTimer.hpp"
43 #include "gc_implementation/shared/gcTrace.hpp"
44 #include "gc_implementation/shared/gcTraceTime.hpp"
45 #include "memory/allocation.hpp"
46 #include "memory/genOopClosures.inline.hpp"
47 #include "memory/referencePolicy.hpp"
48 #include "memory/resourceArea.hpp"
49 #include "oops/oop.inline.hpp"
50 #include "runtime/handles.inline.hpp"
51 #include "runtime/java.hpp"
52 #include "runtime/prefetch.inline.hpp"
53 #include "services/memTracker.hpp"
55 // Concurrent marking bit map wrapper
57 CMBitMapRO::CMBitMapRO(int shifter) :
58 _bm(),
59 _shifter(shifter) {
60 _bmStartWord = 0;
61 _bmWordSize = 0;
62 }
64 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
65 const HeapWord* limit) const {
66 // First we must round addr *up* to a possible object boundary.
67 addr = (HeapWord*)align_size_up((intptr_t)addr,
68 HeapWordSize << _shifter);
69 size_t addrOffset = heapWordToOffset(addr);
70 if (limit == NULL) {
71 limit = _bmStartWord + _bmWordSize;
72 }
73 size_t limitOffset = heapWordToOffset(limit);
74 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
75 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
76 assert(nextAddr >= addr, "get_next_one postcondition");
77 assert(nextAddr == limit || isMarked(nextAddr),
78 "get_next_one postcondition");
79 return nextAddr;
80 }
82 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
83 const HeapWord* limit) const {
84 size_t addrOffset = heapWordToOffset(addr);
85 if (limit == NULL) {
86 limit = _bmStartWord + _bmWordSize;
87 }
88 size_t limitOffset = heapWordToOffset(limit);
89 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
90 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
91 assert(nextAddr >= addr, "get_next_one postcondition");
92 assert(nextAddr == limit || !isMarked(nextAddr),
93 "get_next_one postcondition");
94 return nextAddr;
95 }
97 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
98 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
99 return (int) (diff >> _shifter);
100 }
102 #ifndef PRODUCT
103 bool CMBitMapRO::covers(MemRegion heap_rs) const {
104 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
105 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
106 "size inconsistency");
107 return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
108 _bmWordSize == heap_rs.word_size();
109 }
110 #endif
112 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
113 _bm.print_on_error(st, prefix);
114 }
116 size_t CMBitMap::compute_size(size_t heap_size) {
117 return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
118 }
120 size_t CMBitMap::mark_distance() {
121 return MinObjAlignmentInBytes * BitsPerByte;
122 }
124 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
125 _bmStartWord = heap.start();
126 _bmWordSize = heap.word_size();
128 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
129 _bm.set_size(_bmWordSize >> _shifter);
131 storage->set_mapping_changed_listener(&_listener);
132 }
134 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
135 if (zero_filled) {
136 return;
137 }
138 // We need to clear the bitmap on commit, removing any existing information.
139 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
140 _bm->clearRange(mr);
141 }
143 // Closure used for clearing the given mark bitmap.
144 class ClearBitmapHRClosure : public HeapRegionClosure {
145 private:
146 ConcurrentMark* _cm;
147 CMBitMap* _bitmap;
148 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration.
149 public:
150 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
151 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
152 }
154 virtual bool doHeapRegion(HeapRegion* r) {
155 size_t const chunk_size_in_words = M / HeapWordSize;
157 HeapWord* cur = r->bottom();
158 HeapWord* const end = r->end();
160 while (cur < end) {
161 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
162 _bitmap->clearRange(mr);
164 cur += chunk_size_in_words;
166 // Abort iteration if after yielding the marking has been aborted.
167 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
168 return true;
169 }
170 // Repeat the asserts from before the start of the closure. We will do them
171 // as asserts here to minimize their overhead on the product. However, we
172 // will have them as guarantees at the beginning / end of the bitmap
173 // clearing to get some checking in the product.
174 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
175 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
176 }
178 return false;
179 }
180 };
182 void CMBitMap::clearAll() {
183 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
184 G1CollectedHeap::heap()->heap_region_iterate(&cl);
185 guarantee(cl.complete(), "Must have completed iteration.");
186 return;
187 }
189 void CMBitMap::markRange(MemRegion mr) {
190 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
191 assert(!mr.is_empty(), "unexpected empty region");
192 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
193 ((HeapWord *) mr.end())),
194 "markRange memory region end is not card aligned");
195 // convert address range into offset range
196 _bm.at_put_range(heapWordToOffset(mr.start()),
197 heapWordToOffset(mr.end()), true);
198 }
200 void CMBitMap::clearRange(MemRegion mr) {
201 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
202 assert(!mr.is_empty(), "unexpected empty region");
203 // convert address range into offset range
204 _bm.at_put_range(heapWordToOffset(mr.start()),
205 heapWordToOffset(mr.end()), false);
206 }
208 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
209 HeapWord* end_addr) {
210 HeapWord* start = getNextMarkedWordAddress(addr);
211 start = MIN2(start, end_addr);
212 HeapWord* end = getNextUnmarkedWordAddress(start);
213 end = MIN2(end, end_addr);
214 assert(start <= end, "Consistency check");
215 MemRegion mr(start, end);
216 if (!mr.is_empty()) {
217 clearRange(mr);
218 }
219 return mr;
220 }
222 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
223 _base(NULL), _cm(cm)
224 #ifdef ASSERT
225 , _drain_in_progress(false)
226 , _drain_in_progress_yields(false)
227 #endif
228 {}
230 bool CMMarkStack::allocate(size_t capacity) {
231 // allocate a stack of the requisite depth
232 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
233 if (!rs.is_reserved()) {
234 warning("ConcurrentMark MarkStack allocation failure");
235 return false;
236 }
237 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
238 if (!_virtual_space.initialize(rs, rs.size())) {
239 warning("ConcurrentMark MarkStack backing store failure");
240 // Release the virtual memory reserved for the marking stack
241 rs.release();
242 return false;
243 }
244 assert(_virtual_space.committed_size() == rs.size(),
245 "Didn't reserve backing store for all of ConcurrentMark stack?");
246 _base = (oop*) _virtual_space.low();
247 setEmpty();
248 _capacity = (jint) capacity;
249 _saved_index = -1;
250 _should_expand = false;
251 NOT_PRODUCT(_max_depth = 0);
252 return true;
253 }
255 void CMMarkStack::expand() {
256 // Called, during remark, if we've overflown the marking stack during marking.
257 assert(isEmpty(), "stack should been emptied while handling overflow");
258 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
259 // Clear expansion flag
260 _should_expand = false;
261 if (_capacity == (jint) MarkStackSizeMax) {
262 if (PrintGCDetails && Verbose) {
263 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
264 }
265 return;
266 }
267 // Double capacity if possible
268 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
269 // Do not give up existing stack until we have managed to
270 // get the double capacity that we desired.
271 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
272 sizeof(oop)));
273 if (rs.is_reserved()) {
274 // Release the backing store associated with old stack
275 _virtual_space.release();
276 // Reinitialize virtual space for new stack
277 if (!_virtual_space.initialize(rs, rs.size())) {
278 fatal("Not enough swap for expanded marking stack capacity");
279 }
280 _base = (oop*)(_virtual_space.low());
281 _index = 0;
282 _capacity = new_capacity;
283 } else {
284 if (PrintGCDetails && Verbose) {
285 // Failed to double capacity, continue;
286 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
287 SIZE_FORMAT "K to " SIZE_FORMAT "K",
288 _capacity / K, new_capacity / K);
289 }
290 }
291 }
293 void CMMarkStack::set_should_expand() {
294 // If we're resetting the marking state because of an
295 // marking stack overflow, record that we should, if
296 // possible, expand the stack.
297 _should_expand = _cm->has_overflown();
298 }
300 CMMarkStack::~CMMarkStack() {
301 if (_base != NULL) {
302 _base = NULL;
303 _virtual_space.release();
304 }
305 }
307 void CMMarkStack::par_push(oop ptr) {
308 while (true) {
309 if (isFull()) {
310 _overflow = true;
311 return;
312 }
313 // Otherwise...
314 jint index = _index;
315 jint next_index = index+1;
316 jint res = Atomic::cmpxchg(next_index, &_index, index);
317 if (res == index) {
318 _base[index] = ptr;
319 // Note that we don't maintain this atomically. We could, but it
320 // doesn't seem necessary.
321 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
322 return;
323 }
324 // Otherwise, we need to try again.
325 }
326 }
328 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
329 while (true) {
330 if (isFull()) {
331 _overflow = true;
332 return;
333 }
334 // Otherwise...
335 jint index = _index;
336 jint next_index = index + n;
337 if (next_index > _capacity) {
338 _overflow = true;
339 return;
340 }
341 jint res = Atomic::cmpxchg(next_index, &_index, index);
342 if (res == index) {
343 for (int i = 0; i < n; i++) {
344 int ind = index + i;
345 assert(ind < _capacity, "By overflow test above.");
346 _base[ind] = ptr_arr[i];
347 }
348 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
349 return;
350 }
351 // Otherwise, we need to try again.
352 }
353 }
355 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
356 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
357 jint start = _index;
358 jint next_index = start + n;
359 if (next_index > _capacity) {
360 _overflow = true;
361 return;
362 }
363 // Otherwise.
364 _index = next_index;
365 for (int i = 0; i < n; i++) {
366 int ind = start + i;
367 assert(ind < _capacity, "By overflow test above.");
368 _base[ind] = ptr_arr[i];
369 }
370 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
371 }
373 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
374 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
375 jint index = _index;
376 if (index == 0) {
377 *n = 0;
378 return false;
379 } else {
380 int k = MIN2(max, index);
381 jint new_ind = index - k;
382 for (int j = 0; j < k; j++) {
383 ptr_arr[j] = _base[new_ind + j];
384 }
385 _index = new_ind;
386 *n = k;
387 return true;
388 }
389 }
391 template<class OopClosureClass>
392 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
393 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
394 || SafepointSynchronize::is_at_safepoint(),
395 "Drain recursion must be yield-safe.");
396 bool res = true;
397 debug_only(_drain_in_progress = true);
398 debug_only(_drain_in_progress_yields = yield_after);
399 while (!isEmpty()) {
400 oop newOop = pop();
401 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
402 assert(newOop->is_oop(), "Expected an oop");
403 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
404 "only grey objects on this stack");
405 newOop->oop_iterate(cl);
406 if (yield_after && _cm->do_yield_check()) {
407 res = false;
408 break;
409 }
410 }
411 debug_only(_drain_in_progress = false);
412 return res;
413 }
415 void CMMarkStack::note_start_of_gc() {
416 assert(_saved_index == -1,
417 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
418 _saved_index = _index;
419 }
421 void CMMarkStack::note_end_of_gc() {
422 // This is intentionally a guarantee, instead of an assert. If we
423 // accidentally add something to the mark stack during GC, it
424 // will be a correctness issue so it's better if we crash. we'll
425 // only check this once per GC anyway, so it won't be a performance
426 // issue in any way.
427 guarantee(_saved_index == _index,
428 err_msg("saved index: %d index: %d", _saved_index, _index));
429 _saved_index = -1;
430 }
432 void CMMarkStack::oops_do(OopClosure* f) {
433 assert(_saved_index == _index,
434 err_msg("saved index: %d index: %d", _saved_index, _index));
435 for (int i = 0; i < _index; i += 1) {
436 f->do_oop(&_base[i]);
437 }
438 }
440 CMRootRegions::CMRootRegions() :
441 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
442 _should_abort(false), _next_survivor(NULL) { }
444 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
445 _young_list = g1h->young_list();
446 _cm = cm;
447 }
449 void CMRootRegions::prepare_for_scan() {
450 assert(!scan_in_progress(), "pre-condition");
452 // Currently, only survivors can be root regions.
453 assert(_next_survivor == NULL, "pre-condition");
454 _next_survivor = _young_list->first_survivor_region();
455 _scan_in_progress = (_next_survivor != NULL);
456 _should_abort = false;
457 }
459 HeapRegion* CMRootRegions::claim_next() {
460 if (_should_abort) {
461 // If someone has set the should_abort flag, we return NULL to
462 // force the caller to bail out of their loop.
463 return NULL;
464 }
466 // Currently, only survivors can be root regions.
467 HeapRegion* res = _next_survivor;
468 if (res != NULL) {
469 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
470 // Read it again in case it changed while we were waiting for the lock.
471 res = _next_survivor;
472 if (res != NULL) {
473 if (res == _young_list->last_survivor_region()) {
474 // We just claimed the last survivor so store NULL to indicate
475 // that we're done.
476 _next_survivor = NULL;
477 } else {
478 _next_survivor = res->get_next_young_region();
479 }
480 } else {
481 // Someone else claimed the last survivor while we were trying
482 // to take the lock so nothing else to do.
483 }
484 }
485 assert(res == NULL || res->is_survivor(), "post-condition");
487 return res;
488 }
490 void CMRootRegions::scan_finished() {
491 assert(scan_in_progress(), "pre-condition");
493 // Currently, only survivors can be root regions.
494 if (!_should_abort) {
495 assert(_next_survivor == NULL, "we should have claimed all survivors");
496 }
497 _next_survivor = NULL;
499 {
500 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
501 _scan_in_progress = false;
502 RootRegionScan_lock->notify_all();
503 }
504 }
506 bool CMRootRegions::wait_until_scan_finished() {
507 if (!scan_in_progress()) return false;
509 {
510 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
511 while (scan_in_progress()) {
512 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
513 }
514 }
515 return true;
516 }
518 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
519 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
520 #endif // _MSC_VER
522 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
523 return MAX2((n_par_threads + 2) / 4, 1U);
524 }
526 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
527 _g1h(g1h),
528 _markBitMap1(),
529 _markBitMap2(),
530 _parallel_marking_threads(0),
531 _max_parallel_marking_threads(0),
532 _sleep_factor(0.0),
533 _marking_task_overhead(1.0),
534 _cleanup_sleep_factor(0.0),
535 _cleanup_task_overhead(1.0),
536 _cleanup_list("Cleanup List"),
537 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
538 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
539 CardTableModRefBS::card_shift,
540 false /* in_resource_area*/),
542 _prevMarkBitMap(&_markBitMap1),
543 _nextMarkBitMap(&_markBitMap2),
545 _markStack(this),
546 // _finger set in set_non_marking_state
548 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
549 // _active_tasks set in set_non_marking_state
550 // _tasks set inside the constructor
551 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
552 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
554 _has_overflown(false),
555 _concurrent(false),
556 _has_aborted(false),
557 _aborted_gc_id(GCId::undefined()),
558 _restart_for_overflow(false),
559 _concurrent_marking_in_progress(false),
561 // _verbose_level set below
563 _init_times(),
564 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
565 _cleanup_times(),
566 _total_counting_time(0.0),
567 _total_rs_scrub_time(0.0),
569 _parallel_workers(NULL),
571 _count_card_bitmaps(NULL),
572 _count_marked_bytes(NULL),
573 _completed_initialization(false) {
574 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
575 if (verbose_level < no_verbose) {
576 verbose_level = no_verbose;
577 }
578 if (verbose_level > high_verbose) {
579 verbose_level = high_verbose;
580 }
581 _verbose_level = verbose_level;
583 if (verbose_low()) {
584 gclog_or_tty->print_cr("[global] init, heap start = " PTR_FORMAT", "
585 "heap end = " INTPTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
586 }
588 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
589 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
591 // Create & start a ConcurrentMark thread.
592 _cmThread = new ConcurrentMarkThread(this);
593 assert(cmThread() != NULL, "CM Thread should have been created");
594 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
595 if (_cmThread->osthread() == NULL) {
596 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
597 }
599 assert(CGC_lock != NULL, "Where's the CGC_lock?");
600 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
601 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
603 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
604 satb_qs.set_buffer_size(G1SATBBufferSize);
606 _root_regions.init(_g1h, this);
608 if (ConcGCThreads > ParallelGCThreads) {
609 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
610 "than ParallelGCThreads (" UINTX_FORMAT ").",
611 ConcGCThreads, ParallelGCThreads);
612 return;
613 }
614 if (ParallelGCThreads == 0) {
615 // if we are not running with any parallel GC threads we will not
616 // spawn any marking threads either
617 _parallel_marking_threads = 0;
618 _max_parallel_marking_threads = 0;
619 _sleep_factor = 0.0;
620 _marking_task_overhead = 1.0;
621 } else {
622 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
623 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
624 // if both are set
625 _sleep_factor = 0.0;
626 _marking_task_overhead = 1.0;
627 } else if (G1MarkingOverheadPercent > 0) {
628 // We will calculate the number of parallel marking threads based
629 // on a target overhead with respect to the soft real-time goal
630 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
631 double overall_cm_overhead =
632 (double) MaxGCPauseMillis * marking_overhead /
633 (double) GCPauseIntervalMillis;
634 double cpu_ratio = 1.0 / os::initial_active_processor_count();
635 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
636 double marking_task_overhead =
637 overall_cm_overhead / marking_thread_num * os::initial_active_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();
2177 g1h->allocation_context_stats().update_after_mark();
2179 g1h->trace_heap_after_concurrent_cycle();
2180 }
2182 void ConcurrentMark::completeCleanup() {
2183 if (has_aborted()) return;
2185 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2187 _cleanup_list.verify_optional();
2188 FreeRegionList tmp_free_list("Tmp Free List");
2190 if (G1ConcRegionFreeingVerbose) {
2191 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2192 "cleanup list has %u entries",
2193 _cleanup_list.length());
2194 }
2196 // No one else should be accessing the _cleanup_list at this point,
2197 // so it is not necessary to take any locks
2198 while (!_cleanup_list.is_empty()) {
2199 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2200 assert(hr != NULL, "Got NULL from a non-empty list");
2201 hr->par_clear();
2202 tmp_free_list.add_ordered(hr);
2204 // Instead of adding one region at a time to the secondary_free_list,
2205 // we accumulate them in the local list and move them a few at a
2206 // time. This also cuts down on the number of notify_all() calls
2207 // we do during this process. We'll also append the local list when
2208 // _cleanup_list is empty (which means we just removed the last
2209 // region from the _cleanup_list).
2210 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2211 _cleanup_list.is_empty()) {
2212 if (G1ConcRegionFreeingVerbose) {
2213 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2214 "appending %u entries to the secondary_free_list, "
2215 "cleanup list still has %u entries",
2216 tmp_free_list.length(),
2217 _cleanup_list.length());
2218 }
2220 {
2221 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2222 g1h->secondary_free_list_add(&tmp_free_list);
2223 SecondaryFreeList_lock->notify_all();
2224 }
2226 if (G1StressConcRegionFreeing) {
2227 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2228 os::sleep(Thread::current(), (jlong) 1, false);
2229 }
2230 }
2231 }
2232 }
2233 assert(tmp_free_list.is_empty(), "post-condition");
2234 }
2236 // Supporting Object and Oop closures for reference discovery
2237 // and processing in during marking
2239 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2240 HeapWord* addr = (HeapWord*)obj;
2241 return addr != NULL &&
2242 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2243 }
2245 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2246 // Uses the CMTask associated with a worker thread (for serial reference
2247 // processing the CMTask for worker 0 is used) to preserve (mark) and
2248 // trace referent objects.
2249 //
2250 // Using the CMTask and embedded local queues avoids having the worker
2251 // threads operating on the global mark stack. This reduces the risk
2252 // of overflowing the stack - which we would rather avoid at this late
2253 // state. Also using the tasks' local queues removes the potential
2254 // of the workers interfering with each other that could occur if
2255 // operating on the global stack.
2257 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2258 ConcurrentMark* _cm;
2259 CMTask* _task;
2260 int _ref_counter_limit;
2261 int _ref_counter;
2262 bool _is_serial;
2263 public:
2264 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2265 _cm(cm), _task(task), _is_serial(is_serial),
2266 _ref_counter_limit(G1RefProcDrainInterval) {
2267 assert(_ref_counter_limit > 0, "sanity");
2268 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2269 _ref_counter = _ref_counter_limit;
2270 }
2272 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2273 virtual void do_oop( oop* p) { do_oop_work(p); }
2275 template <class T> void do_oop_work(T* p) {
2276 if (!_cm->has_overflown()) {
2277 oop obj = oopDesc::load_decode_heap_oop(p);
2278 if (_cm->verbose_high()) {
2279 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2280 "*" PTR_FORMAT " = " PTR_FORMAT,
2281 _task->worker_id(), p2i(p), p2i((void*) obj));
2282 }
2284 _task->deal_with_reference(obj);
2285 _ref_counter--;
2287 if (_ref_counter == 0) {
2288 // We have dealt with _ref_counter_limit references, pushing them
2289 // and objects reachable from them on to the local stack (and
2290 // possibly the global stack). Call CMTask::do_marking_step() to
2291 // process these entries.
2292 //
2293 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2294 // there's nothing more to do (i.e. we're done with the entries that
2295 // were pushed as a result of the CMTask::deal_with_reference() calls
2296 // above) or we overflow.
2297 //
2298 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2299 // flag while there may still be some work to do. (See the comment at
2300 // the beginning of CMTask::do_marking_step() for those conditions -
2301 // one of which is reaching the specified time target.) It is only
2302 // when CMTask::do_marking_step() returns without setting the
2303 // has_aborted() flag that the marking step has completed.
2304 do {
2305 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2306 _task->do_marking_step(mark_step_duration_ms,
2307 false /* do_termination */,
2308 _is_serial);
2309 } while (_task->has_aborted() && !_cm->has_overflown());
2310 _ref_counter = _ref_counter_limit;
2311 }
2312 } else {
2313 if (_cm->verbose_high()) {
2314 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2315 }
2316 }
2317 }
2318 };
2320 // 'Drain' oop closure used by both serial and parallel reference processing.
2321 // Uses the CMTask associated with a given worker thread (for serial
2322 // reference processing the CMtask for worker 0 is used). Calls the
2323 // do_marking_step routine, with an unbelievably large timeout value,
2324 // to drain the marking data structures of the remaining entries
2325 // added by the 'keep alive' oop closure above.
2327 class G1CMDrainMarkingStackClosure: public VoidClosure {
2328 ConcurrentMark* _cm;
2329 CMTask* _task;
2330 bool _is_serial;
2331 public:
2332 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2333 _cm(cm), _task(task), _is_serial(is_serial) {
2334 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2335 }
2337 void do_void() {
2338 do {
2339 if (_cm->verbose_high()) {
2340 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2341 _task->worker_id(), BOOL_TO_STR(_is_serial));
2342 }
2344 // We call CMTask::do_marking_step() to completely drain the local
2345 // and global marking stacks of entries pushed by the 'keep alive'
2346 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2347 //
2348 // CMTask::do_marking_step() is called in a loop, which we'll exit
2349 // if there's nothing more to do (i.e. we'completely drained the
2350 // entries that were pushed as a a result of applying the 'keep alive'
2351 // closure to the entries on the discovered ref lists) or we overflow
2352 // the global marking stack.
2353 //
2354 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2355 // flag while there may still be some work to do. (See the comment at
2356 // the beginning of CMTask::do_marking_step() for those conditions -
2357 // one of which is reaching the specified time target.) It is only
2358 // when CMTask::do_marking_step() returns without setting the
2359 // has_aborted() flag that the marking step has completed.
2361 _task->do_marking_step(1000000000.0 /* something very large */,
2362 true /* do_termination */,
2363 _is_serial);
2364 } while (_task->has_aborted() && !_cm->has_overflown());
2365 }
2366 };
2368 // Implementation of AbstractRefProcTaskExecutor for parallel
2369 // reference processing at the end of G1 concurrent marking
2371 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2372 private:
2373 G1CollectedHeap* _g1h;
2374 ConcurrentMark* _cm;
2375 WorkGang* _workers;
2376 int _active_workers;
2378 public:
2379 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2380 ConcurrentMark* cm,
2381 WorkGang* workers,
2382 int n_workers) :
2383 _g1h(g1h), _cm(cm),
2384 _workers(workers), _active_workers(n_workers) { }
2386 // Executes the given task using concurrent marking worker threads.
2387 virtual void execute(ProcessTask& task);
2388 virtual void execute(EnqueueTask& task);
2389 };
2391 class G1CMRefProcTaskProxy: public AbstractGangTask {
2392 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2393 ProcessTask& _proc_task;
2394 G1CollectedHeap* _g1h;
2395 ConcurrentMark* _cm;
2397 public:
2398 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2399 G1CollectedHeap* g1h,
2400 ConcurrentMark* cm) :
2401 AbstractGangTask("Process reference objects in parallel"),
2402 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2403 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2404 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2405 }
2407 virtual void work(uint worker_id) {
2408 ResourceMark rm;
2409 HandleMark hm;
2410 CMTask* task = _cm->task(worker_id);
2411 G1CMIsAliveClosure g1_is_alive(_g1h);
2412 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2413 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2415 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2416 }
2417 };
2419 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2420 assert(_workers != NULL, "Need parallel worker threads.");
2421 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2423 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2425 // We need to reset the concurrency level before each
2426 // proxy task execution, so that the termination protocol
2427 // and overflow handling in CMTask::do_marking_step() knows
2428 // how many workers to wait for.
2429 _cm->set_concurrency(_active_workers);
2430 _g1h->set_par_threads(_active_workers);
2431 _workers->run_task(&proc_task_proxy);
2432 _g1h->set_par_threads(0);
2433 }
2435 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2436 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2437 EnqueueTask& _enq_task;
2439 public:
2440 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2441 AbstractGangTask("Enqueue reference objects in parallel"),
2442 _enq_task(enq_task) { }
2444 virtual void work(uint worker_id) {
2445 _enq_task.work(worker_id);
2446 }
2447 };
2449 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2450 assert(_workers != NULL, "Need parallel worker threads.");
2451 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2453 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2455 // Not strictly necessary but...
2456 //
2457 // We need to reset the concurrency level before each
2458 // proxy task execution, so that the termination protocol
2459 // and overflow handling in CMTask::do_marking_step() knows
2460 // how many workers to wait for.
2461 _cm->set_concurrency(_active_workers);
2462 _g1h->set_par_threads(_active_workers);
2463 _workers->run_task(&enq_task_proxy);
2464 _g1h->set_par_threads(0);
2465 }
2467 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2468 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2469 }
2471 // Helper class to get rid of some boilerplate code.
2472 class G1RemarkGCTraceTime : public GCTraceTime {
2473 static bool doit_and_prepend(bool doit) {
2474 if (doit) {
2475 gclog_or_tty->put(' ');
2476 }
2477 return doit;
2478 }
2480 public:
2481 G1RemarkGCTraceTime(const char* title, bool doit)
2482 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
2483 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
2484 }
2485 };
2487 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2488 if (has_overflown()) {
2489 // Skip processing the discovered references if we have
2490 // overflown the global marking stack. Reference objects
2491 // only get discovered once so it is OK to not
2492 // de-populate the discovered reference lists. We could have,
2493 // but the only benefit would be that, when marking restarts,
2494 // less reference objects are discovered.
2495 return;
2496 }
2498 ResourceMark rm;
2499 HandleMark hm;
2501 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2503 // Is alive closure.
2504 G1CMIsAliveClosure g1_is_alive(g1h);
2506 // Inner scope to exclude the cleaning of the string and symbol
2507 // tables from the displayed time.
2508 {
2509 if (G1Log::finer()) {
2510 gclog_or_tty->put(' ');
2511 }
2512 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm(), concurrent_gc_id());
2514 ReferenceProcessor* rp = g1h->ref_processor_cm();
2516 // See the comment in G1CollectedHeap::ref_processing_init()
2517 // about how reference processing currently works in G1.
2519 // Set the soft reference policy
2520 rp->setup_policy(clear_all_soft_refs);
2521 assert(_markStack.isEmpty(), "mark stack should be empty");
2523 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2524 // in serial reference processing. Note these closures are also
2525 // used for serially processing (by the the current thread) the
2526 // JNI references during parallel reference processing.
2527 //
2528 // These closures do not need to synchronize with the worker
2529 // threads involved in parallel reference processing as these
2530 // instances are executed serially by the current thread (e.g.
2531 // reference processing is not multi-threaded and is thus
2532 // performed by the current thread instead of a gang worker).
2533 //
2534 // The gang tasks involved in parallel reference procssing create
2535 // their own instances of these closures, which do their own
2536 // synchronization among themselves.
2537 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2538 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2540 // We need at least one active thread. If reference processing
2541 // is not multi-threaded we use the current (VMThread) thread,
2542 // otherwise we use the work gang from the G1CollectedHeap and
2543 // we utilize all the worker threads we can.
2544 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2545 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2546 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2548 // Parallel processing task executor.
2549 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2550 g1h->workers(), active_workers);
2551 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2553 // Set the concurrency level. The phase was already set prior to
2554 // executing the remark task.
2555 set_concurrency(active_workers);
2557 // Set the degree of MT processing here. If the discovery was done MT,
2558 // the number of threads involved during discovery could differ from
2559 // the number of active workers. This is OK as long as the discovered
2560 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2561 rp->set_active_mt_degree(active_workers);
2563 // Process the weak references.
2564 const ReferenceProcessorStats& stats =
2565 rp->process_discovered_references(&g1_is_alive,
2566 &g1_keep_alive,
2567 &g1_drain_mark_stack,
2568 executor,
2569 g1h->gc_timer_cm(),
2570 concurrent_gc_id());
2571 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2573 // The do_oop work routines of the keep_alive and drain_marking_stack
2574 // oop closures will set the has_overflown flag if we overflow the
2575 // global marking stack.
2577 assert(_markStack.overflow() || _markStack.isEmpty(),
2578 "mark stack should be empty (unless it overflowed)");
2580 if (_markStack.overflow()) {
2581 // This should have been done already when we tried to push an
2582 // entry on to the global mark stack. But let's do it again.
2583 set_has_overflown();
2584 }
2586 assert(rp->num_q() == active_workers, "why not");
2588 rp->enqueue_discovered_references(executor);
2590 rp->verify_no_references_recorded();
2591 assert(!rp->discovery_enabled(), "Post condition");
2592 }
2594 if (has_overflown()) {
2595 // We can not trust g1_is_alive if the marking stack overflowed
2596 return;
2597 }
2599 assert(_markStack.isEmpty(), "Marking should have completed");
2601 // Unload Klasses, String, Symbols, Code Cache, etc.
2602 {
2603 G1RemarkGCTraceTime trace("Unloading", G1Log::finer());
2605 if (ClassUnloadingWithConcurrentMark) {
2606 // Cleaning of klasses depends on correct information from MetadataMarkOnStack. The CodeCache::mark_on_stack
2607 // part is too slow to be done serially, so it is handled during the weakRefsWorkParallelPart phase.
2608 // Defer the cleaning until we have complete on_stack data.
2609 MetadataOnStackMark md_on_stack(false /* Don't visit the code cache at this point */);
2611 bool purged_classes;
2613 {
2614 G1RemarkGCTraceTime trace("System Dictionary Unloading", G1Log::finest());
2615 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2616 }
2618 {
2619 G1RemarkGCTraceTime trace("Parallel Unloading", G1Log::finest());
2620 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2621 }
2623 {
2624 G1RemarkGCTraceTime trace("Deallocate Metadata", G1Log::finest());
2625 ClassLoaderDataGraph::free_deallocate_lists();
2626 }
2627 }
2629 if (G1StringDedup::is_enabled()) {
2630 G1RemarkGCTraceTime trace("String Deduplication Unlink", G1Log::finest());
2631 G1StringDedup::unlink(&g1_is_alive);
2632 }
2633 }
2634 }
2636 void ConcurrentMark::swapMarkBitMaps() {
2637 CMBitMapRO* temp = _prevMarkBitMap;
2638 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2639 _nextMarkBitMap = (CMBitMap*) temp;
2640 }
2642 // Closure for marking entries in SATB buffers.
2643 class CMSATBBufferClosure : public SATBBufferClosure {
2644 private:
2645 CMTask* _task;
2646 G1CollectedHeap* _g1h;
2648 // This is very similar to CMTask::deal_with_reference, but with
2649 // more relaxed requirements for the argument, so this must be more
2650 // circumspect about treating the argument as an object.
2651 void do_entry(void* entry) const {
2652 _task->increment_refs_reached();
2653 HeapRegion* hr = _g1h->heap_region_containing_raw(entry);
2654 if (entry < hr->next_top_at_mark_start()) {
2655 // Until we get here, we don't know whether entry refers to a valid
2656 // object; it could instead have been a stale reference.
2657 oop obj = static_cast<oop>(entry);
2658 assert(obj->is_oop(true /* ignore mark word */),
2659 err_msg("Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj)));
2660 _task->make_reference_grey(obj, hr);
2661 }
2662 }
2664 public:
2665 CMSATBBufferClosure(CMTask* task, G1CollectedHeap* g1h)
2666 : _task(task), _g1h(g1h) { }
2668 virtual void do_buffer(void** buffer, size_t size) {
2669 for (size_t i = 0; i < size; ++i) {
2670 do_entry(buffer[i]);
2671 }
2672 }
2673 };
2675 class G1RemarkThreadsClosure : public ThreadClosure {
2676 CMSATBBufferClosure _cm_satb_cl;
2677 G1CMOopClosure _cm_cl;
2678 MarkingCodeBlobClosure _code_cl;
2679 int _thread_parity;
2680 bool _is_par;
2682 public:
2683 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) :
2684 _cm_satb_cl(task, g1h),
2685 _cm_cl(g1h, g1h->concurrent_mark(), task),
2686 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2687 _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {}
2689 void do_thread(Thread* thread) {
2690 if (thread->is_Java_thread()) {
2691 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2692 JavaThread* jt = (JavaThread*)thread;
2694 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2695 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2696 // * Alive if on the stack of an executing method
2697 // * Weakly reachable otherwise
2698 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2699 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2700 jt->nmethods_do(&_code_cl);
2702 jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
2703 }
2704 } else if (thread->is_VM_thread()) {
2705 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2706 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
2707 }
2708 }
2709 }
2710 };
2712 class CMRemarkTask: public AbstractGangTask {
2713 private:
2714 ConcurrentMark* _cm;
2715 bool _is_serial;
2716 public:
2717 void work(uint worker_id) {
2718 // Since all available tasks are actually started, we should
2719 // only proceed if we're supposed to be actived.
2720 if (worker_id < _cm->active_tasks()) {
2721 CMTask* task = _cm->task(worker_id);
2722 task->record_start_time();
2723 {
2724 ResourceMark rm;
2725 HandleMark hm;
2727 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial);
2728 Threads::threads_do(&threads_f);
2729 }
2731 do {
2732 task->do_marking_step(1000000000.0 /* something very large */,
2733 true /* do_termination */,
2734 _is_serial);
2735 } while (task->has_aborted() && !_cm->has_overflown());
2736 // If we overflow, then we do not want to restart. We instead
2737 // want to abort remark and do concurrent marking again.
2738 task->record_end_time();
2739 }
2740 }
2742 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2743 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2744 _cm->terminator()->reset_for_reuse(active_workers);
2745 }
2746 };
2748 void ConcurrentMark::checkpointRootsFinalWork() {
2749 ResourceMark rm;
2750 HandleMark hm;
2751 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2753 G1RemarkGCTraceTime trace("Finalize Marking", G1Log::finer());
2755 g1h->ensure_parsability(false);
2757 if (G1CollectedHeap::use_parallel_gc_threads()) {
2758 G1CollectedHeap::StrongRootsScope srs(g1h);
2759 // this is remark, so we'll use up all active threads
2760 uint active_workers = g1h->workers()->active_workers();
2761 if (active_workers == 0) {
2762 assert(active_workers > 0, "Should have been set earlier");
2763 active_workers = (uint) ParallelGCThreads;
2764 g1h->workers()->set_active_workers(active_workers);
2765 }
2766 set_concurrency_and_phase(active_workers, false /* concurrent */);
2767 // Leave _parallel_marking_threads at it's
2768 // value originally calculated in the ConcurrentMark
2769 // constructor and pass values of the active workers
2770 // through the gang in the task.
2772 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2773 // We will start all available threads, even if we decide that the
2774 // active_workers will be fewer. The extra ones will just bail out
2775 // immediately.
2776 g1h->set_par_threads(active_workers);
2777 g1h->workers()->run_task(&remarkTask);
2778 g1h->set_par_threads(0);
2779 } else {
2780 G1CollectedHeap::StrongRootsScope srs(g1h);
2781 uint active_workers = 1;
2782 set_concurrency_and_phase(active_workers, false /* concurrent */);
2784 // Note - if there's no work gang then the VMThread will be
2785 // the thread to execute the remark - serially. We have
2786 // to pass true for the is_serial parameter so that
2787 // CMTask::do_marking_step() doesn't enter the sync
2788 // barriers in the event of an overflow. Doing so will
2789 // cause an assert that the current thread is not a
2790 // concurrent GC thread.
2791 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2792 remarkTask.work(0);
2793 }
2794 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2795 guarantee(has_overflown() ||
2796 satb_mq_set.completed_buffers_num() == 0,
2797 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2798 BOOL_TO_STR(has_overflown()),
2799 satb_mq_set.completed_buffers_num()));
2801 print_stats();
2802 }
2804 #ifndef PRODUCT
2806 class PrintReachableOopClosure: public OopClosure {
2807 private:
2808 G1CollectedHeap* _g1h;
2809 outputStream* _out;
2810 VerifyOption _vo;
2811 bool _all;
2813 public:
2814 PrintReachableOopClosure(outputStream* out,
2815 VerifyOption vo,
2816 bool all) :
2817 _g1h(G1CollectedHeap::heap()),
2818 _out(out), _vo(vo), _all(all) { }
2820 void do_oop(narrowOop* p) { do_oop_work(p); }
2821 void do_oop( oop* p) { do_oop_work(p); }
2823 template <class T> void do_oop_work(T* p) {
2824 oop obj = oopDesc::load_decode_heap_oop(p);
2825 const char* str = NULL;
2826 const char* str2 = "";
2828 if (obj == NULL) {
2829 str = "";
2830 } else if (!_g1h->is_in_g1_reserved(obj)) {
2831 str = " O";
2832 } else {
2833 HeapRegion* hr = _g1h->heap_region_containing(obj);
2834 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2835 bool marked = _g1h->is_marked(obj, _vo);
2837 if (over_tams) {
2838 str = " >";
2839 if (marked) {
2840 str2 = " AND MARKED";
2841 }
2842 } else if (marked) {
2843 str = " M";
2844 } else {
2845 str = " NOT";
2846 }
2847 }
2849 _out->print_cr(" " PTR_FORMAT ": " PTR_FORMAT "%s%s",
2850 p2i(p), p2i((void*) obj), str, str2);
2851 }
2852 };
2854 class PrintReachableObjectClosure : public ObjectClosure {
2855 private:
2856 G1CollectedHeap* _g1h;
2857 outputStream* _out;
2858 VerifyOption _vo;
2859 bool _all;
2860 HeapRegion* _hr;
2862 public:
2863 PrintReachableObjectClosure(outputStream* out,
2864 VerifyOption vo,
2865 bool all,
2866 HeapRegion* hr) :
2867 _g1h(G1CollectedHeap::heap()),
2868 _out(out), _vo(vo), _all(all), _hr(hr) { }
2870 void do_object(oop o) {
2871 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2872 bool marked = _g1h->is_marked(o, _vo);
2873 bool print_it = _all || over_tams || marked;
2875 if (print_it) {
2876 _out->print_cr(" " PTR_FORMAT "%s",
2877 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2878 PrintReachableOopClosure oopCl(_out, _vo, _all);
2879 o->oop_iterate_no_header(&oopCl);
2880 }
2881 }
2882 };
2884 class PrintReachableRegionClosure : public HeapRegionClosure {
2885 private:
2886 G1CollectedHeap* _g1h;
2887 outputStream* _out;
2888 VerifyOption _vo;
2889 bool _all;
2891 public:
2892 bool doHeapRegion(HeapRegion* hr) {
2893 HeapWord* b = hr->bottom();
2894 HeapWord* e = hr->end();
2895 HeapWord* t = hr->top();
2896 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2897 _out->print_cr("** [" PTR_FORMAT ", " PTR_FORMAT "] top: " PTR_FORMAT " "
2898 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2899 _out->cr();
2901 HeapWord* from = b;
2902 HeapWord* to = t;
2904 if (to > from) {
2905 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2906 _out->cr();
2907 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2908 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2909 _out->cr();
2910 }
2912 return false;
2913 }
2915 PrintReachableRegionClosure(outputStream* out,
2916 VerifyOption vo,
2917 bool all) :
2918 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2919 };
2921 void ConcurrentMark::print_reachable(const char* str,
2922 VerifyOption vo,
2923 bool all) {
2924 gclog_or_tty->cr();
2925 gclog_or_tty->print_cr("== Doing heap dump... ");
2927 if (G1PrintReachableBaseFile == NULL) {
2928 gclog_or_tty->print_cr(" #### error: no base file defined");
2929 return;
2930 }
2932 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2933 (JVM_MAXPATHLEN - 1)) {
2934 gclog_or_tty->print_cr(" #### error: file name too long");
2935 return;
2936 }
2938 char file_name[JVM_MAXPATHLEN];
2939 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2940 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2942 fileStream fout(file_name);
2943 if (!fout.is_open()) {
2944 gclog_or_tty->print_cr(" #### error: could not open file");
2945 return;
2946 }
2948 outputStream* out = &fout;
2949 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2950 out->cr();
2952 out->print_cr("--- ITERATING OVER REGIONS");
2953 out->cr();
2954 PrintReachableRegionClosure rcl(out, vo, all);
2955 _g1h->heap_region_iterate(&rcl);
2956 out->cr();
2958 gclog_or_tty->print_cr(" done");
2959 gclog_or_tty->flush();
2960 }
2962 #endif // PRODUCT
2964 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2965 // Note we are overriding the read-only view of the prev map here, via
2966 // the cast.
2967 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2968 }
2970 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2971 _nextMarkBitMap->clearRange(mr);
2972 }
2974 HeapRegion*
2975 ConcurrentMark::claim_region(uint worker_id) {
2976 // "checkpoint" the finger
2977 HeapWord* finger = _finger;
2979 // _heap_end will not change underneath our feet; it only changes at
2980 // yield points.
2981 while (finger < _heap_end) {
2982 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2984 // Note on how this code handles humongous regions. In the
2985 // normal case the finger will reach the start of a "starts
2986 // humongous" (SH) region. Its end will either be the end of the
2987 // last "continues humongous" (CH) region in the sequence, or the
2988 // standard end of the SH region (if the SH is the only region in
2989 // the sequence). That way claim_region() will skip over the CH
2990 // regions. However, there is a subtle race between a CM thread
2991 // executing this method and a mutator thread doing a humongous
2992 // object allocation. The two are not mutually exclusive as the CM
2993 // thread does not need to hold the Heap_lock when it gets
2994 // here. So there is a chance that claim_region() will come across
2995 // a free region that's in the progress of becoming a SH or a CH
2996 // region. In the former case, it will either
2997 // a) Miss the update to the region's end, in which case it will
2998 // visit every subsequent CH region, will find their bitmaps
2999 // empty, and do nothing, or
3000 // b) Will observe the update of the region's end (in which case
3001 // it will skip the subsequent CH regions).
3002 // If it comes across a region that suddenly becomes CH, the
3003 // scenario will be similar to b). So, the race between
3004 // claim_region() and a humongous object allocation might force us
3005 // to do a bit of unnecessary work (due to some unnecessary bitmap
3006 // iterations) but it should not introduce and correctness issues.
3007 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
3009 // Above heap_region_containing_raw may return NULL as we always scan claim
3010 // until the end of the heap. In this case, just jump to the next region.
3011 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
3013 // Is the gap between reading the finger and doing the CAS too long?
3014 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
3015 if (res == finger && curr_region != NULL) {
3016 // we succeeded
3017 HeapWord* bottom = curr_region->bottom();
3018 HeapWord* limit = curr_region->next_top_at_mark_start();
3020 if (verbose_low()) {
3021 gclog_or_tty->print_cr("[%u] curr_region = " PTR_FORMAT " "
3022 "[" PTR_FORMAT ", " PTR_FORMAT "), "
3023 "limit = " PTR_FORMAT,
3024 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
3025 }
3027 // notice that _finger == end cannot be guaranteed here since,
3028 // someone else might have moved the finger even further
3029 assert(_finger >= end, "the finger should have moved forward");
3031 if (verbose_low()) {
3032 gclog_or_tty->print_cr("[%u] we were successful with region = "
3033 PTR_FORMAT, worker_id, p2i(curr_region));
3034 }
3036 if (limit > bottom) {
3037 if (verbose_low()) {
3038 gclog_or_tty->print_cr("[%u] region " PTR_FORMAT " is not empty, "
3039 "returning it ", worker_id, p2i(curr_region));
3040 }
3041 return curr_region;
3042 } else {
3043 assert(limit == bottom,
3044 "the region limit should be at bottom");
3045 if (verbose_low()) {
3046 gclog_or_tty->print_cr("[%u] region " PTR_FORMAT " is empty, "
3047 "returning NULL", worker_id, p2i(curr_region));
3048 }
3049 // we return NULL and the caller should try calling
3050 // claim_region() again.
3051 return NULL;
3052 }
3053 } else {
3054 assert(_finger > finger, "the finger should have moved forward");
3055 if (verbose_low()) {
3056 if (curr_region == NULL) {
3057 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
3058 "global finger = " PTR_FORMAT ", "
3059 "our finger = " PTR_FORMAT,
3060 worker_id, p2i(_finger), p2i(finger));
3061 } else {
3062 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
3063 "global finger = " PTR_FORMAT ", "
3064 "our finger = " PTR_FORMAT,
3065 worker_id, p2i(_finger), p2i(finger));
3066 }
3067 }
3069 // read it again
3070 finger = _finger;
3071 }
3072 }
3074 return NULL;
3075 }
3077 #ifndef PRODUCT
3078 enum VerifyNoCSetOopsPhase {
3079 VerifyNoCSetOopsStack,
3080 VerifyNoCSetOopsQueues
3081 };
3083 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
3084 private:
3085 G1CollectedHeap* _g1h;
3086 VerifyNoCSetOopsPhase _phase;
3087 int _info;
3089 const char* phase_str() {
3090 switch (_phase) {
3091 case VerifyNoCSetOopsStack: return "Stack";
3092 case VerifyNoCSetOopsQueues: return "Queue";
3093 default: ShouldNotReachHere();
3094 }
3095 return NULL;
3096 }
3098 void do_object_work(oop obj) {
3099 guarantee(!_g1h->obj_in_cs(obj),
3100 err_msg("obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
3101 p2i((void*) obj), phase_str(), _info));
3102 }
3104 public:
3105 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3107 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3108 _phase = phase;
3109 _info = info;
3110 }
3112 virtual void do_oop(oop* p) {
3113 oop obj = oopDesc::load_decode_heap_oop(p);
3114 do_object_work(obj);
3115 }
3117 virtual void do_oop(narrowOop* p) {
3118 // We should not come across narrow oops while scanning marking
3119 // stacks
3120 ShouldNotReachHere();
3121 }
3123 virtual void do_object(oop obj) {
3124 do_object_work(obj);
3125 }
3126 };
3128 void ConcurrentMark::verify_no_cset_oops() {
3129 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3130 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3131 return;
3132 }
3134 VerifyNoCSetOopsClosure cl;
3136 // Verify entries on the global mark stack
3137 cl.set_phase(VerifyNoCSetOopsStack);
3138 _markStack.oops_do(&cl);
3140 // Verify entries on the task queues
3141 for (uint i = 0; i < _max_worker_id; i += 1) {
3142 cl.set_phase(VerifyNoCSetOopsQueues, i);
3143 CMTaskQueue* queue = _task_queues->queue(i);
3144 queue->oops_do(&cl);
3145 }
3147 // Verify the global finger
3148 HeapWord* global_finger = finger();
3149 if (global_finger != NULL && global_finger < _heap_end) {
3150 // The global finger always points to a heap region boundary. We
3151 // use heap_region_containing_raw() to get the containing region
3152 // given that the global finger could be pointing to a free region
3153 // which subsequently becomes continues humongous. If that
3154 // happens, heap_region_containing() will return the bottom of the
3155 // corresponding starts humongous region and the check below will
3156 // not hold any more.
3157 // Since we always iterate over all regions, we might get a NULL HeapRegion
3158 // here.
3159 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3160 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3161 err_msg("global finger: " PTR_FORMAT " region: " HR_FORMAT,
3162 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3163 }
3165 // Verify the task fingers
3166 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3167 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3168 CMTask* task = _tasks[i];
3169 HeapWord* task_finger = task->finger();
3170 if (task_finger != NULL && task_finger < _heap_end) {
3171 // See above note on the global finger verification.
3172 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3173 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3174 !task_hr->in_collection_set(),
3175 err_msg("task finger: " PTR_FORMAT " region: " HR_FORMAT,
3176 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3177 }
3178 }
3179 }
3180 #endif // PRODUCT
3182 // Aggregate the counting data that was constructed concurrently
3183 // with marking.
3184 class AggregateCountDataHRClosure: public HeapRegionClosure {
3185 G1CollectedHeap* _g1h;
3186 ConcurrentMark* _cm;
3187 CardTableModRefBS* _ct_bs;
3188 BitMap* _cm_card_bm;
3189 uint _max_worker_id;
3191 public:
3192 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3193 BitMap* cm_card_bm,
3194 uint max_worker_id) :
3195 _g1h(g1h), _cm(g1h->concurrent_mark()),
3196 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3197 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3199 bool doHeapRegion(HeapRegion* hr) {
3200 if (hr->continuesHumongous()) {
3201 // We will ignore these here and process them when their
3202 // associated "starts humongous" region is processed.
3203 // Note that we cannot rely on their associated
3204 // "starts humongous" region to have their bit set to 1
3205 // since, due to the region chunking in the parallel region
3206 // iteration, a "continues humongous" region might be visited
3207 // before its associated "starts humongous".
3208 return false;
3209 }
3211 HeapWord* start = hr->bottom();
3212 HeapWord* limit = hr->next_top_at_mark_start();
3213 HeapWord* end = hr->end();
3215 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3216 err_msg("Preconditions not met - "
3217 "start: " PTR_FORMAT ", limit: " PTR_FORMAT ", "
3218 "top: " PTR_FORMAT ", end: " PTR_FORMAT,
3219 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3221 assert(hr->next_marked_bytes() == 0, "Precondition");
3223 if (start == limit) {
3224 // NTAMS of this region has not been set so nothing to do.
3225 return false;
3226 }
3228 // 'start' should be in the heap.
3229 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3230 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3231 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3233 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3234 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3235 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3237 // If ntams is not card aligned then we bump card bitmap index
3238 // for limit so that we get the all the cards spanned by
3239 // the object ending at ntams.
3240 // Note: if this is the last region in the heap then ntams
3241 // could be actually just beyond the end of the the heap;
3242 // limit_idx will then correspond to a (non-existent) card
3243 // that is also outside the heap.
3244 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3245 limit_idx += 1;
3246 }
3248 assert(limit_idx <= end_idx, "or else use atomics");
3250 // Aggregate the "stripe" in the count data associated with hr.
3251 uint hrm_index = hr->hrm_index();
3252 size_t marked_bytes = 0;
3254 for (uint i = 0; i < _max_worker_id; i += 1) {
3255 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3256 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3258 // Fetch the marked_bytes in this region for task i and
3259 // add it to the running total for this region.
3260 marked_bytes += marked_bytes_array[hrm_index];
3262 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3263 // into the global card bitmap.
3264 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3266 while (scan_idx < limit_idx) {
3267 assert(task_card_bm->at(scan_idx) == true, "should be");
3268 _cm_card_bm->set_bit(scan_idx);
3269 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3271 // BitMap::get_next_one_offset() can handle the case when
3272 // its left_offset parameter is greater than its right_offset
3273 // parameter. It does, however, have an early exit if
3274 // left_offset == right_offset. So let's limit the value
3275 // passed in for left offset here.
3276 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3277 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3278 }
3279 }
3281 // Update the marked bytes for this region.
3282 hr->add_to_marked_bytes(marked_bytes);
3284 // Next heap region
3285 return false;
3286 }
3287 };
3289 class G1AggregateCountDataTask: public AbstractGangTask {
3290 protected:
3291 G1CollectedHeap* _g1h;
3292 ConcurrentMark* _cm;
3293 BitMap* _cm_card_bm;
3294 uint _max_worker_id;
3295 int _active_workers;
3297 public:
3298 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3299 ConcurrentMark* cm,
3300 BitMap* cm_card_bm,
3301 uint max_worker_id,
3302 int n_workers) :
3303 AbstractGangTask("Count Aggregation"),
3304 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3305 _max_worker_id(max_worker_id),
3306 _active_workers(n_workers) { }
3308 void work(uint worker_id) {
3309 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3311 if (G1CollectedHeap::use_parallel_gc_threads()) {
3312 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3313 _active_workers,
3314 HeapRegion::AggregateCountClaimValue);
3315 } else {
3316 _g1h->heap_region_iterate(&cl);
3317 }
3318 }
3319 };
3322 void ConcurrentMark::aggregate_count_data() {
3323 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3324 _g1h->workers()->active_workers() :
3325 1);
3327 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3328 _max_worker_id, n_workers);
3330 if (G1CollectedHeap::use_parallel_gc_threads()) {
3331 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3332 "sanity check");
3333 _g1h->set_par_threads(n_workers);
3334 _g1h->workers()->run_task(&g1_par_agg_task);
3335 _g1h->set_par_threads(0);
3337 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3338 "sanity check");
3339 _g1h->reset_heap_region_claim_values();
3340 } else {
3341 g1_par_agg_task.work(0);
3342 }
3343 }
3345 // Clear the per-worker arrays used to store the per-region counting data
3346 void ConcurrentMark::clear_all_count_data() {
3347 // Clear the global card bitmap - it will be filled during
3348 // liveness count aggregation (during remark) and the
3349 // final counting task.
3350 _card_bm.clear();
3352 // Clear the global region bitmap - it will be filled as part
3353 // of the final counting task.
3354 _region_bm.clear();
3356 uint max_regions = _g1h->max_regions();
3357 assert(_max_worker_id > 0, "uninitialized");
3359 for (uint i = 0; i < _max_worker_id; i += 1) {
3360 BitMap* task_card_bm = count_card_bitmap_for(i);
3361 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3363 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3364 assert(marked_bytes_array != NULL, "uninitialized");
3366 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3367 task_card_bm->clear();
3368 }
3369 }
3371 void ConcurrentMark::print_stats() {
3372 if (verbose_stats()) {
3373 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3374 for (size_t i = 0; i < _active_tasks; ++i) {
3375 _tasks[i]->print_stats();
3376 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3377 }
3378 }
3379 }
3381 // abandon current marking iteration due to a Full GC
3382 void ConcurrentMark::abort() {
3383 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3384 // concurrent bitmap clearing.
3385 _nextMarkBitMap->clearAll();
3387 // Note we cannot clear the previous marking bitmap here
3388 // since VerifyDuringGC verifies the objects marked during
3389 // a full GC against the previous bitmap.
3391 // Clear the liveness counting data
3392 clear_all_count_data();
3393 // Empty mark stack
3394 reset_marking_state();
3395 for (uint i = 0; i < _max_worker_id; ++i) {
3396 _tasks[i]->clear_region_fields();
3397 }
3398 _first_overflow_barrier_sync.abort();
3399 _second_overflow_barrier_sync.abort();
3400 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3401 if (!gc_id.is_undefined()) {
3402 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3403 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3404 _aborted_gc_id = gc_id;
3405 }
3406 _has_aborted = true;
3408 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3409 satb_mq_set.abandon_partial_marking();
3410 // This can be called either during or outside marking, we'll read
3411 // the expected_active value from the SATB queue set.
3412 satb_mq_set.set_active_all_threads(
3413 false, /* new active value */
3414 satb_mq_set.is_active() /* expected_active */);
3416 _g1h->trace_heap_after_concurrent_cycle();
3417 _g1h->register_concurrent_cycle_end();
3418 }
3420 const GCId& ConcurrentMark::concurrent_gc_id() {
3421 if (has_aborted()) {
3422 return _aborted_gc_id;
3423 }
3424 return _g1h->gc_tracer_cm()->gc_id();
3425 }
3427 static void print_ms_time_info(const char* prefix, const char* name,
3428 NumberSeq& ns) {
3429 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3430 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3431 if (ns.num() > 0) {
3432 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3433 prefix, ns.sd(), ns.maximum());
3434 }
3435 }
3437 void ConcurrentMark::print_summary_info() {
3438 gclog_or_tty->print_cr(" Concurrent marking:");
3439 print_ms_time_info(" ", "init marks", _init_times);
3440 print_ms_time_info(" ", "remarks", _remark_times);
3441 {
3442 print_ms_time_info(" ", "final marks", _remark_mark_times);
3443 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3445 }
3446 print_ms_time_info(" ", "cleanups", _cleanup_times);
3447 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3448 _total_counting_time,
3449 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3450 (double)_cleanup_times.num()
3451 : 0.0));
3452 if (G1ScrubRemSets) {
3453 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3454 _total_rs_scrub_time,
3455 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3456 (double)_cleanup_times.num()
3457 : 0.0));
3458 }
3459 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3460 (_init_times.sum() + _remark_times.sum() +
3461 _cleanup_times.sum())/1000.0);
3462 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3463 "(%8.2f s marking).",
3464 cmThread()->vtime_accum(),
3465 cmThread()->vtime_mark_accum());
3466 }
3468 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3469 if (use_parallel_marking_threads()) {
3470 _parallel_workers->print_worker_threads_on(st);
3471 }
3472 }
3474 void ConcurrentMark::print_on_error(outputStream* st) const {
3475 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3476 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3477 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3478 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3479 }
3481 // We take a break if someone is trying to stop the world.
3482 bool ConcurrentMark::do_yield_check(uint worker_id) {
3483 if (SuspendibleThreadSet::should_yield()) {
3484 if (worker_id == 0) {
3485 _g1h->g1_policy()->record_concurrent_pause();
3486 }
3487 SuspendibleThreadSet::yield();
3488 return true;
3489 } else {
3490 return false;
3491 }
3492 }
3494 #ifndef PRODUCT
3495 // for debugging purposes
3496 void ConcurrentMark::print_finger() {
3497 gclog_or_tty->print_cr("heap [" PTR_FORMAT ", " PTR_FORMAT "), global finger = " PTR_FORMAT,
3498 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3499 for (uint i = 0; i < _max_worker_id; ++i) {
3500 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3501 }
3502 gclog_or_tty->cr();
3503 }
3504 #endif
3506 template<bool scan>
3507 inline void CMTask::process_grey_object(oop obj) {
3508 assert(scan || obj->is_typeArray(), "Skipping scan of grey non-typeArray");
3509 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3511 if (_cm->verbose_high()) {
3512 gclog_or_tty->print_cr("[%u] processing grey object " PTR_FORMAT,
3513 _worker_id, p2i((void*) obj));
3514 }
3516 size_t obj_size = obj->size();
3517 _words_scanned += obj_size;
3519 if (scan) {
3520 obj->oop_iterate(_cm_oop_closure);
3521 }
3522 statsOnly( ++_objs_scanned );
3523 check_limits();
3524 }
3526 template void CMTask::process_grey_object<true>(oop);
3527 template void CMTask::process_grey_object<false>(oop);
3529 // Closure for iteration over bitmaps
3530 class CMBitMapClosure : public BitMapClosure {
3531 private:
3532 // the bitmap that is being iterated over
3533 CMBitMap* _nextMarkBitMap;
3534 ConcurrentMark* _cm;
3535 CMTask* _task;
3537 public:
3538 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3539 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3541 bool do_bit(size_t offset) {
3542 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3543 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3544 assert( addr < _cm->finger(), "invariant");
3546 statsOnly( _task->increase_objs_found_on_bitmap() );
3547 assert(addr >= _task->finger(), "invariant");
3549 // We move that task's local finger along.
3550 _task->move_finger_to(addr);
3552 _task->scan_object(oop(addr));
3553 // we only partially drain the local queue and global stack
3554 _task->drain_local_queue(true);
3555 _task->drain_global_stack(true);
3557 // if the has_aborted flag has been raised, we need to bail out of
3558 // the iteration
3559 return !_task->has_aborted();
3560 }
3561 };
3563 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3564 ConcurrentMark* cm,
3565 CMTask* task)
3566 : _g1h(g1h), _cm(cm), _task(task) {
3567 assert(_ref_processor == NULL, "should be initialized to NULL");
3569 if (G1UseConcMarkReferenceProcessing) {
3570 _ref_processor = g1h->ref_processor_cm();
3571 assert(_ref_processor != NULL, "should not be NULL");
3572 }
3573 }
3575 void CMTask::setup_for_region(HeapRegion* hr) {
3576 assert(hr != NULL,
3577 "claim_region() should have filtered out NULL regions");
3578 assert(!hr->continuesHumongous(),
3579 "claim_region() should have filtered out continues humongous regions");
3581 if (_cm->verbose_low()) {
3582 gclog_or_tty->print_cr("[%u] setting up for region " PTR_FORMAT,
3583 _worker_id, p2i(hr));
3584 }
3586 _curr_region = hr;
3587 _finger = hr->bottom();
3588 update_region_limit();
3589 }
3591 void CMTask::update_region_limit() {
3592 HeapRegion* hr = _curr_region;
3593 HeapWord* bottom = hr->bottom();
3594 HeapWord* limit = hr->next_top_at_mark_start();
3596 if (limit == bottom) {
3597 if (_cm->verbose_low()) {
3598 gclog_or_tty->print_cr("[%u] found an empty region "
3599 "[" PTR_FORMAT ", " PTR_FORMAT ")",
3600 _worker_id, p2i(bottom), p2i(limit));
3601 }
3602 // The region was collected underneath our feet.
3603 // We set the finger to bottom to ensure that the bitmap
3604 // iteration that will follow this will not do anything.
3605 // (this is not a condition that holds when we set the region up,
3606 // as the region is not supposed to be empty in the first place)
3607 _finger = bottom;
3608 } else if (limit >= _region_limit) {
3609 assert(limit >= _finger, "peace of mind");
3610 } else {
3611 assert(limit < _region_limit, "only way to get here");
3612 // This can happen under some pretty unusual circumstances. An
3613 // evacuation pause empties the region underneath our feet (NTAMS
3614 // at bottom). We then do some allocation in the region (NTAMS
3615 // stays at bottom), followed by the region being used as a GC
3616 // alloc region (NTAMS will move to top() and the objects
3617 // originally below it will be grayed). All objects now marked in
3618 // the region are explicitly grayed, if below the global finger,
3619 // and we do not need in fact to scan anything else. So, we simply
3620 // set _finger to be limit to ensure that the bitmap iteration
3621 // doesn't do anything.
3622 _finger = limit;
3623 }
3625 _region_limit = limit;
3626 }
3628 void CMTask::giveup_current_region() {
3629 assert(_curr_region != NULL, "invariant");
3630 if (_cm->verbose_low()) {
3631 gclog_or_tty->print_cr("[%u] giving up region " PTR_FORMAT,
3632 _worker_id, p2i(_curr_region));
3633 }
3634 clear_region_fields();
3635 }
3637 void CMTask::clear_region_fields() {
3638 // Values for these three fields that indicate that we're not
3639 // holding on to a region.
3640 _curr_region = NULL;
3641 _finger = NULL;
3642 _region_limit = NULL;
3643 }
3645 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3646 if (cm_oop_closure == NULL) {
3647 assert(_cm_oop_closure != NULL, "invariant");
3648 } else {
3649 assert(_cm_oop_closure == NULL, "invariant");
3650 }
3651 _cm_oop_closure = cm_oop_closure;
3652 }
3654 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3655 guarantee(nextMarkBitMap != NULL, "invariant");
3657 if (_cm->verbose_low()) {
3658 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3659 }
3661 _nextMarkBitMap = nextMarkBitMap;
3662 clear_region_fields();
3664 _calls = 0;
3665 _elapsed_time_ms = 0.0;
3666 _termination_time_ms = 0.0;
3667 _termination_start_time_ms = 0.0;
3669 #if _MARKING_STATS_
3670 _local_pushes = 0;
3671 _local_pops = 0;
3672 _local_max_size = 0;
3673 _objs_scanned = 0;
3674 _global_pushes = 0;
3675 _global_pops = 0;
3676 _global_max_size = 0;
3677 _global_transfers_to = 0;
3678 _global_transfers_from = 0;
3679 _regions_claimed = 0;
3680 _objs_found_on_bitmap = 0;
3681 _satb_buffers_processed = 0;
3682 _steal_attempts = 0;
3683 _steals = 0;
3684 _aborted = 0;
3685 _aborted_overflow = 0;
3686 _aborted_cm_aborted = 0;
3687 _aborted_yield = 0;
3688 _aborted_timed_out = 0;
3689 _aborted_satb = 0;
3690 _aborted_termination = 0;
3691 #endif // _MARKING_STATS_
3692 }
3694 bool CMTask::should_exit_termination() {
3695 regular_clock_call();
3696 // This is called when we are in the termination protocol. We should
3697 // quit if, for some reason, this task wants to abort or the global
3698 // stack is not empty (this means that we can get work from it).
3699 return !_cm->mark_stack_empty() || has_aborted();
3700 }
3702 void CMTask::reached_limit() {
3703 assert(_words_scanned >= _words_scanned_limit ||
3704 _refs_reached >= _refs_reached_limit ,
3705 "shouldn't have been called otherwise");
3706 regular_clock_call();
3707 }
3709 void CMTask::regular_clock_call() {
3710 if (has_aborted()) return;
3712 // First, we need to recalculate the words scanned and refs reached
3713 // limits for the next clock call.
3714 recalculate_limits();
3716 // During the regular clock call we do the following
3718 // (1) If an overflow has been flagged, then we abort.
3719 if (_cm->has_overflown()) {
3720 set_has_aborted();
3721 return;
3722 }
3724 // If we are not concurrent (i.e. we're doing remark) we don't need
3725 // to check anything else. The other steps are only needed during
3726 // the concurrent marking phase.
3727 if (!concurrent()) return;
3729 // (2) If marking has been aborted for Full GC, then we also abort.
3730 if (_cm->has_aborted()) {
3731 set_has_aborted();
3732 statsOnly( ++_aborted_cm_aborted );
3733 return;
3734 }
3736 double curr_time_ms = os::elapsedVTime() * 1000.0;
3738 // (3) If marking stats are enabled, then we update the step history.
3739 #if _MARKING_STATS_
3740 if (_words_scanned >= _words_scanned_limit) {
3741 ++_clock_due_to_scanning;
3742 }
3743 if (_refs_reached >= _refs_reached_limit) {
3744 ++_clock_due_to_marking;
3745 }
3747 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3748 _interval_start_time_ms = curr_time_ms;
3749 _all_clock_intervals_ms.add(last_interval_ms);
3751 if (_cm->verbose_medium()) {
3752 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3753 "scanned = " SIZE_FORMAT "%s, refs reached = " SIZE_FORMAT "%s",
3754 _worker_id, last_interval_ms,
3755 _words_scanned,
3756 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3757 _refs_reached,
3758 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3759 }
3760 #endif // _MARKING_STATS_
3762 // (4) We check whether we should yield. If we have to, then we abort.
3763 if (SuspendibleThreadSet::should_yield()) {
3764 // We should yield. To do this we abort the task. The caller is
3765 // responsible for yielding.
3766 set_has_aborted();
3767 statsOnly( ++_aborted_yield );
3768 return;
3769 }
3771 // (5) We check whether we've reached our time quota. If we have,
3772 // then we abort.
3773 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3774 if (elapsed_time_ms > _time_target_ms) {
3775 set_has_aborted();
3776 _has_timed_out = true;
3777 statsOnly( ++_aborted_timed_out );
3778 return;
3779 }
3781 // (6) Finally, we check whether there are enough completed STAB
3782 // buffers available for processing. If there are, we abort.
3783 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3784 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3785 if (_cm->verbose_low()) {
3786 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3787 _worker_id);
3788 }
3789 // we do need to process SATB buffers, we'll abort and restart
3790 // the marking task to do so
3791 set_has_aborted();
3792 statsOnly( ++_aborted_satb );
3793 return;
3794 }
3795 }
3797 void CMTask::recalculate_limits() {
3798 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3799 _words_scanned_limit = _real_words_scanned_limit;
3801 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3802 _refs_reached_limit = _real_refs_reached_limit;
3803 }
3805 void CMTask::decrease_limits() {
3806 // This is called when we believe that we're going to do an infrequent
3807 // operation which will increase the per byte scanned cost (i.e. move
3808 // entries to/from the global stack). It basically tries to decrease the
3809 // scanning limit so that the clock is called earlier.
3811 if (_cm->verbose_medium()) {
3812 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3813 }
3815 _words_scanned_limit = _real_words_scanned_limit -
3816 3 * words_scanned_period / 4;
3817 _refs_reached_limit = _real_refs_reached_limit -
3818 3 * refs_reached_period / 4;
3819 }
3821 void CMTask::move_entries_to_global_stack() {
3822 // local array where we'll store the entries that will be popped
3823 // from the local queue
3824 oop buffer[global_stack_transfer_size];
3826 int n = 0;
3827 oop obj;
3828 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3829 buffer[n] = obj;
3830 ++n;
3831 }
3833 if (n > 0) {
3834 // we popped at least one entry from the local queue
3836 statsOnly( ++_global_transfers_to; _local_pops += n );
3838 if (!_cm->mark_stack_push(buffer, n)) {
3839 if (_cm->verbose_low()) {
3840 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3841 _worker_id);
3842 }
3843 set_has_aborted();
3844 } else {
3845 // the transfer was successful
3847 if (_cm->verbose_medium()) {
3848 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3849 _worker_id, n);
3850 }
3851 statsOnly( int tmp_size = _cm->mark_stack_size();
3852 if (tmp_size > _global_max_size) {
3853 _global_max_size = tmp_size;
3854 }
3855 _global_pushes += n );
3856 }
3857 }
3859 // this operation was quite expensive, so decrease the limits
3860 decrease_limits();
3861 }
3863 void CMTask::get_entries_from_global_stack() {
3864 // local array where we'll store the entries that will be popped
3865 // from the global stack.
3866 oop buffer[global_stack_transfer_size];
3867 int n;
3868 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3869 assert(n <= global_stack_transfer_size,
3870 "we should not pop more than the given limit");
3871 if (n > 0) {
3872 // yes, we did actually pop at least one entry
3874 statsOnly( ++_global_transfers_from; _global_pops += n );
3875 if (_cm->verbose_medium()) {
3876 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3877 _worker_id, n);
3878 }
3879 for (int i = 0; i < n; ++i) {
3880 bool success = _task_queue->push(buffer[i]);
3881 // We only call this when the local queue is empty or under a
3882 // given target limit. So, we do not expect this push to fail.
3883 assert(success, "invariant");
3884 }
3886 statsOnly( int tmp_size = _task_queue->size();
3887 if (tmp_size > _local_max_size) {
3888 _local_max_size = tmp_size;
3889 }
3890 _local_pushes += n );
3891 }
3893 // this operation was quite expensive, so decrease the limits
3894 decrease_limits();
3895 }
3897 void CMTask::drain_local_queue(bool partially) {
3898 if (has_aborted()) return;
3900 // Decide what the target size is, depending whether we're going to
3901 // drain it partially (so that other tasks can steal if they run out
3902 // of things to do) or totally (at the very end).
3903 size_t target_size;
3904 if (partially) {
3905 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3906 } else {
3907 target_size = 0;
3908 }
3910 if (_task_queue->size() > target_size) {
3911 if (_cm->verbose_high()) {
3912 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3913 _worker_id, target_size);
3914 }
3916 oop obj;
3917 bool ret = _task_queue->pop_local(obj);
3918 while (ret) {
3919 statsOnly( ++_local_pops );
3921 if (_cm->verbose_high()) {
3922 gclog_or_tty->print_cr("[%u] popped " PTR_FORMAT, _worker_id,
3923 p2i((void*) obj));
3924 }
3926 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3927 assert(!_g1h->is_on_master_free_list(
3928 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3930 scan_object(obj);
3932 if (_task_queue->size() <= target_size || has_aborted()) {
3933 ret = false;
3934 } else {
3935 ret = _task_queue->pop_local(obj);
3936 }
3937 }
3939 if (_cm->verbose_high()) {
3940 gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3941 _worker_id, _task_queue->size());
3942 }
3943 }
3944 }
3946 void CMTask::drain_global_stack(bool partially) {
3947 if (has_aborted()) return;
3949 // We have a policy to drain the local queue before we attempt to
3950 // drain the global stack.
3951 assert(partially || _task_queue->size() == 0, "invariant");
3953 // Decide what the target size is, depending whether we're going to
3954 // drain it partially (so that other tasks can steal if they run out
3955 // of things to do) or totally (at the very end). Notice that,
3956 // because we move entries from the global stack in chunks or
3957 // because another task might be doing the same, we might in fact
3958 // drop below the target. But, this is not a problem.
3959 size_t target_size;
3960 if (partially) {
3961 target_size = _cm->partial_mark_stack_size_target();
3962 } else {
3963 target_size = 0;
3964 }
3966 if (_cm->mark_stack_size() > target_size) {
3967 if (_cm->verbose_low()) {
3968 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3969 _worker_id, target_size);
3970 }
3972 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3973 get_entries_from_global_stack();
3974 drain_local_queue(partially);
3975 }
3977 if (_cm->verbose_low()) {
3978 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3979 _worker_id, _cm->mark_stack_size());
3980 }
3981 }
3982 }
3984 // SATB Queue has several assumptions on whether to call the par or
3985 // non-par versions of the methods. this is why some of the code is
3986 // replicated. We should really get rid of the single-threaded version
3987 // of the code to simplify things.
3988 void CMTask::drain_satb_buffers() {
3989 if (has_aborted()) return;
3991 // We set this so that the regular clock knows that we're in the
3992 // middle of draining buffers and doesn't set the abort flag when it
3993 // notices that SATB buffers are available for draining. It'd be
3994 // very counter productive if it did that. :-)
3995 _draining_satb_buffers = true;
3997 CMSATBBufferClosure satb_cl(this, _g1h);
3998 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
4000 // This keeps claiming and applying the closure to completed buffers
4001 // until we run out of buffers or we need to abort.
4002 while (!has_aborted() &&
4003 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
4004 if (_cm->verbose_medium()) {
4005 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4006 }
4007 statsOnly( ++_satb_buffers_processed );
4008 regular_clock_call();
4009 }
4011 _draining_satb_buffers = false;
4013 assert(has_aborted() ||
4014 concurrent() ||
4015 satb_mq_set.completed_buffers_num() == 0, "invariant");
4017 // again, this was a potentially expensive operation, decrease the
4018 // limits to get the regular clock call early
4019 decrease_limits();
4020 }
4022 void CMTask::print_stats() {
4023 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
4024 _worker_id, _calls);
4025 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
4026 _elapsed_time_ms, _termination_time_ms);
4027 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4028 _step_times_ms.num(), _step_times_ms.avg(),
4029 _step_times_ms.sd());
4030 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4031 _step_times_ms.maximum(), _step_times_ms.sum());
4033 #if _MARKING_STATS_
4034 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4035 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
4036 _all_clock_intervals_ms.sd());
4037 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4038 _all_clock_intervals_ms.maximum(),
4039 _all_clock_intervals_ms.sum());
4040 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
4041 _clock_due_to_scanning, _clock_due_to_marking);
4042 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
4043 _objs_scanned, _objs_found_on_bitmap);
4044 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
4045 _local_pushes, _local_pops, _local_max_size);
4046 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
4047 _global_pushes, _global_pops, _global_max_size);
4048 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
4049 _global_transfers_to,_global_transfers_from);
4050 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
4051 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
4052 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
4053 _steal_attempts, _steals);
4054 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
4055 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
4056 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
4057 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
4058 _aborted_timed_out, _aborted_satb, _aborted_termination);
4059 #endif // _MARKING_STATS_
4060 }
4062 /*****************************************************************************
4064 The do_marking_step(time_target_ms, ...) method is the building
4065 block of the parallel marking framework. It can be called in parallel
4066 with other invocations of do_marking_step() on different tasks
4067 (but only one per task, obviously) and concurrently with the
4068 mutator threads, or during remark, hence it eliminates the need
4069 for two versions of the code. When called during remark, it will
4070 pick up from where the task left off during the concurrent marking
4071 phase. Interestingly, tasks are also claimable during evacuation
4072 pauses too, since do_marking_step() ensures that it aborts before
4073 it needs to yield.
4075 The data structures that it uses to do marking work are the
4076 following:
4078 (1) Marking Bitmap. If there are gray objects that appear only
4079 on the bitmap (this happens either when dealing with an overflow
4080 or when the initial marking phase has simply marked the roots
4081 and didn't push them on the stack), then tasks claim heap
4082 regions whose bitmap they then scan to find gray objects. A
4083 global finger indicates where the end of the last claimed region
4084 is. A local finger indicates how far into the region a task has
4085 scanned. The two fingers are used to determine how to gray an
4086 object (i.e. whether simply marking it is OK, as it will be
4087 visited by a task in the future, or whether it needs to be also
4088 pushed on a stack).
4090 (2) Local Queue. The local queue of the task which is accessed
4091 reasonably efficiently by the task. Other tasks can steal from
4092 it when they run out of work. Throughout the marking phase, a
4093 task attempts to keep its local queue short but not totally
4094 empty, so that entries are available for stealing by other
4095 tasks. Only when there is no more work, a task will totally
4096 drain its local queue.
4098 (3) Global Mark Stack. This handles local queue overflow. During
4099 marking only sets of entries are moved between it and the local
4100 queues, as access to it requires a mutex and more fine-grain
4101 interaction with it which might cause contention. If it
4102 overflows, then the marking phase should restart and iterate
4103 over the bitmap to identify gray objects. Throughout the marking
4104 phase, tasks attempt to keep the global mark stack at a small
4105 length but not totally empty, so that entries are available for
4106 popping by other tasks. Only when there is no more work, tasks
4107 will totally drain the global mark stack.
4109 (4) SATB Buffer Queue. This is where completed SATB buffers are
4110 made available. Buffers are regularly removed from this queue
4111 and scanned for roots, so that the queue doesn't get too
4112 long. During remark, all completed buffers are processed, as
4113 well as the filled in parts of any uncompleted buffers.
4115 The do_marking_step() method tries to abort when the time target
4116 has been reached. There are a few other cases when the
4117 do_marking_step() method also aborts:
4119 (1) When the marking phase has been aborted (after a Full GC).
4121 (2) When a global overflow (on the global stack) has been
4122 triggered. Before the task aborts, it will actually sync up with
4123 the other tasks to ensure that all the marking data structures
4124 (local queues, stacks, fingers etc.) are re-initialized so that
4125 when do_marking_step() completes, the marking phase can
4126 immediately restart.
4128 (3) When enough completed SATB buffers are available. The
4129 do_marking_step() method only tries to drain SATB buffers right
4130 at the beginning. So, if enough buffers are available, the
4131 marking step aborts and the SATB buffers are processed at
4132 the beginning of the next invocation.
4134 (4) To yield. when we have to yield then we abort and yield
4135 right at the end of do_marking_step(). This saves us from a lot
4136 of hassle as, by yielding we might allow a Full GC. If this
4137 happens then objects will be compacted underneath our feet, the
4138 heap might shrink, etc. We save checking for this by just
4139 aborting and doing the yield right at the end.
4141 From the above it follows that the do_marking_step() method should
4142 be called in a loop (or, otherwise, regularly) until it completes.
4144 If a marking step completes without its has_aborted() flag being
4145 true, it means it has completed the current marking phase (and
4146 also all other marking tasks have done so and have all synced up).
4148 A method called regular_clock_call() is invoked "regularly" (in
4149 sub ms intervals) throughout marking. It is this clock method that
4150 checks all the abort conditions which were mentioned above and
4151 decides when the task should abort. A work-based scheme is used to
4152 trigger this clock method: when the number of object words the
4153 marking phase has scanned or the number of references the marking
4154 phase has visited reach a given limit. Additional invocations to
4155 the method clock have been planted in a few other strategic places
4156 too. The initial reason for the clock method was to avoid calling
4157 vtime too regularly, as it is quite expensive. So, once it was in
4158 place, it was natural to piggy-back all the other conditions on it
4159 too and not constantly check them throughout the code.
4161 If do_termination is true then do_marking_step will enter its
4162 termination protocol.
4164 The value of is_serial must be true when do_marking_step is being
4165 called serially (i.e. by the VMThread) and do_marking_step should
4166 skip any synchronization in the termination and overflow code.
4167 Examples include the serial remark code and the serial reference
4168 processing closures.
4170 The value of is_serial must be false when do_marking_step is
4171 being called by any of the worker threads in a work gang.
4172 Examples include the concurrent marking code (CMMarkingTask),
4173 the MT remark code, and the MT reference processing closures.
4175 *****************************************************************************/
4177 void CMTask::do_marking_step(double time_target_ms,
4178 bool do_termination,
4179 bool is_serial) {
4180 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4181 assert(concurrent() == _cm->concurrent(), "they should be the same");
4183 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4184 assert(_task_queues != NULL, "invariant");
4185 assert(_task_queue != NULL, "invariant");
4186 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4188 assert(!_claimed,
4189 "only one thread should claim this task at any one time");
4191 // OK, this doesn't safeguard again all possible scenarios, as it is
4192 // possible for two threads to set the _claimed flag at the same
4193 // time. But it is only for debugging purposes anyway and it will
4194 // catch most problems.
4195 _claimed = true;
4197 _start_time_ms = os::elapsedVTime() * 1000.0;
4198 statsOnly( _interval_start_time_ms = _start_time_ms );
4200 // If do_stealing is true then do_marking_step will attempt to
4201 // steal work from the other CMTasks. It only makes sense to
4202 // enable stealing when the termination protocol is enabled
4203 // and do_marking_step() is not being called serially.
4204 bool do_stealing = do_termination && !is_serial;
4206 double diff_prediction_ms =
4207 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4208 _time_target_ms = time_target_ms - diff_prediction_ms;
4210 // set up the variables that are used in the work-based scheme to
4211 // call the regular clock method
4212 _words_scanned = 0;
4213 _refs_reached = 0;
4214 recalculate_limits();
4216 // clear all flags
4217 clear_has_aborted();
4218 _has_timed_out = false;
4219 _draining_satb_buffers = false;
4221 ++_calls;
4223 if (_cm->verbose_low()) {
4224 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4225 "target = %1.2lfms >>>>>>>>>>",
4226 _worker_id, _calls, _time_target_ms);
4227 }
4229 // Set up the bitmap and oop closures. Anything that uses them is
4230 // eventually called from this method, so it is OK to allocate these
4231 // statically.
4232 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4233 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4234 set_cm_oop_closure(&cm_oop_closure);
4236 if (_cm->has_overflown()) {
4237 // This can happen if the mark stack overflows during a GC pause
4238 // and this task, after a yield point, restarts. We have to abort
4239 // as we need to get into the overflow protocol which happens
4240 // right at the end of this task.
4241 set_has_aborted();
4242 }
4244 // First drain any available SATB buffers. After this, we will not
4245 // look at SATB buffers before the next invocation of this method.
4246 // If enough completed SATB buffers are queued up, the regular clock
4247 // will abort this task so that it restarts.
4248 drain_satb_buffers();
4249 // ...then partially drain the local queue and the global stack
4250 drain_local_queue(true);
4251 drain_global_stack(true);
4253 do {
4254 if (!has_aborted() && _curr_region != NULL) {
4255 // This means that we're already holding on to a region.
4256 assert(_finger != NULL, "if region is not NULL, then the finger "
4257 "should not be NULL either");
4259 // We might have restarted this task after an evacuation pause
4260 // which might have evacuated the region we're holding on to
4261 // underneath our feet. Let's read its limit again to make sure
4262 // that we do not iterate over a region of the heap that
4263 // contains garbage (update_region_limit() will also move
4264 // _finger to the start of the region if it is found empty).
4265 update_region_limit();
4266 // We will start from _finger not from the start of the region,
4267 // as we might be restarting this task after aborting half-way
4268 // through scanning this region. In this case, _finger points to
4269 // the address where we last found a marked object. If this is a
4270 // fresh region, _finger points to start().
4271 MemRegion mr = MemRegion(_finger, _region_limit);
4273 if (_cm->verbose_low()) {
4274 gclog_or_tty->print_cr("[%u] we're scanning part "
4275 "[" PTR_FORMAT ", " PTR_FORMAT ") "
4276 "of region " HR_FORMAT,
4277 _worker_id, p2i(_finger), p2i(_region_limit),
4278 HR_FORMAT_PARAMS(_curr_region));
4279 }
4281 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4282 "humongous regions should go around loop once only");
4284 // Some special cases:
4285 // If the memory region is empty, we can just give up the region.
4286 // If the current region is humongous then we only need to check
4287 // the bitmap for the bit associated with the start of the object,
4288 // scan the object if it's live, and give up the region.
4289 // Otherwise, let's iterate over the bitmap of the part of the region
4290 // that is left.
4291 // If the iteration is successful, give up the region.
4292 if (mr.is_empty()) {
4293 giveup_current_region();
4294 regular_clock_call();
4295 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4296 if (_nextMarkBitMap->isMarked(mr.start())) {
4297 // The object is marked - apply the closure
4298 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4299 bitmap_closure.do_bit(offset);
4300 }
4301 // Even if this task aborted while scanning the humongous object
4302 // we can (and should) give up the current region.
4303 giveup_current_region();
4304 regular_clock_call();
4305 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4306 giveup_current_region();
4307 regular_clock_call();
4308 } else {
4309 assert(has_aborted(), "currently the only way to do so");
4310 // The only way to abort the bitmap iteration is to return
4311 // false from the do_bit() method. However, inside the
4312 // do_bit() method we move the _finger to point to the
4313 // object currently being looked at. So, if we bail out, we
4314 // have definitely set _finger to something non-null.
4315 assert(_finger != NULL, "invariant");
4317 // Region iteration was actually aborted. So now _finger
4318 // points to the address of the object we last scanned. If we
4319 // leave it there, when we restart this task, we will rescan
4320 // the object. It is easy to avoid this. We move the finger by
4321 // enough to point to the next possible object header (the
4322 // bitmap knows by how much we need to move it as it knows its
4323 // granularity).
4324 assert(_finger < _region_limit, "invariant");
4325 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4326 // Check if bitmap iteration was aborted while scanning the last object
4327 if (new_finger >= _region_limit) {
4328 giveup_current_region();
4329 } else {
4330 move_finger_to(new_finger);
4331 }
4332 }
4333 }
4334 // At this point we have either completed iterating over the
4335 // region we were holding on to, or we have aborted.
4337 // We then partially drain the local queue and the global stack.
4338 // (Do we really need this?)
4339 drain_local_queue(true);
4340 drain_global_stack(true);
4342 // Read the note on the claim_region() method on why it might
4343 // return NULL with potentially more regions available for
4344 // claiming and why we have to check out_of_regions() to determine
4345 // whether we're done or not.
4346 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4347 // We are going to try to claim a new region. We should have
4348 // given up on the previous one.
4349 // Separated the asserts so that we know which one fires.
4350 assert(_curr_region == NULL, "invariant");
4351 assert(_finger == NULL, "invariant");
4352 assert(_region_limit == NULL, "invariant");
4353 if (_cm->verbose_low()) {
4354 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4355 }
4356 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4357 if (claimed_region != NULL) {
4358 // Yes, we managed to claim one
4359 statsOnly( ++_regions_claimed );
4361 if (_cm->verbose_low()) {
4362 gclog_or_tty->print_cr("[%u] we successfully claimed "
4363 "region " PTR_FORMAT,
4364 _worker_id, p2i(claimed_region));
4365 }
4367 setup_for_region(claimed_region);
4368 assert(_curr_region == claimed_region, "invariant");
4369 }
4370 // It is important to call the regular clock here. It might take
4371 // a while to claim a region if, for example, we hit a large
4372 // block of empty regions. So we need to call the regular clock
4373 // method once round the loop to make sure it's called
4374 // frequently enough.
4375 regular_clock_call();
4376 }
4378 if (!has_aborted() && _curr_region == NULL) {
4379 assert(_cm->out_of_regions(),
4380 "at this point we should be out of regions");
4381 }
4382 } while ( _curr_region != NULL && !has_aborted());
4384 if (!has_aborted()) {
4385 // We cannot check whether the global stack is empty, since other
4386 // tasks might be pushing objects to it concurrently.
4387 assert(_cm->out_of_regions(),
4388 "at this point we should be out of regions");
4390 if (_cm->verbose_low()) {
4391 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4392 }
4394 // Try to reduce the number of available SATB buffers so that
4395 // remark has less work to do.
4396 drain_satb_buffers();
4397 }
4399 // Since we've done everything else, we can now totally drain the
4400 // local queue and global stack.
4401 drain_local_queue(false);
4402 drain_global_stack(false);
4404 // Attempt at work stealing from other task's queues.
4405 if (do_stealing && !has_aborted()) {
4406 // We have not aborted. This means that we have finished all that
4407 // we could. Let's try to do some stealing...
4409 // We cannot check whether the global stack is empty, since other
4410 // tasks might be pushing objects to it concurrently.
4411 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4412 "only way to reach here");
4414 if (_cm->verbose_low()) {
4415 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4416 }
4418 while (!has_aborted()) {
4419 oop obj;
4420 statsOnly( ++_steal_attempts );
4422 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4423 if (_cm->verbose_medium()) {
4424 gclog_or_tty->print_cr("[%u] stolen " PTR_FORMAT " successfully",
4425 _worker_id, p2i((void*) obj));
4426 }
4428 statsOnly( ++_steals );
4430 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4431 "any stolen object should be marked");
4432 scan_object(obj);
4434 // And since we're towards the end, let's totally drain the
4435 // local queue and global stack.
4436 drain_local_queue(false);
4437 drain_global_stack(false);
4438 } else {
4439 break;
4440 }
4441 }
4442 }
4444 // If we are about to wrap up and go into termination, check if we
4445 // should raise the overflow flag.
4446 if (do_termination && !has_aborted()) {
4447 if (_cm->force_overflow()->should_force()) {
4448 _cm->set_has_overflown();
4449 regular_clock_call();
4450 }
4451 }
4453 // We still haven't aborted. Now, let's try to get into the
4454 // termination protocol.
4455 if (do_termination && !has_aborted()) {
4456 // We cannot check whether the global stack is empty, since other
4457 // tasks might be concurrently pushing objects on it.
4458 // Separated the asserts so that we know which one fires.
4459 assert(_cm->out_of_regions(), "only way to reach here");
4460 assert(_task_queue->size() == 0, "only way to reach here");
4462 if (_cm->verbose_low()) {
4463 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4464 }
4466 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4468 // The CMTask class also extends the TerminatorTerminator class,
4469 // hence its should_exit_termination() method will also decide
4470 // whether to exit the termination protocol or not.
4471 bool finished = (is_serial ||
4472 _cm->terminator()->offer_termination(this));
4473 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4474 _termination_time_ms +=
4475 termination_end_time_ms - _termination_start_time_ms;
4477 if (finished) {
4478 // We're all done.
4480 if (_worker_id == 0) {
4481 // let's allow task 0 to do this
4482 if (concurrent()) {
4483 assert(_cm->concurrent_marking_in_progress(), "invariant");
4484 // we need to set this to false before the next
4485 // safepoint. This way we ensure that the marking phase
4486 // doesn't observe any more heap expansions.
4487 _cm->clear_concurrent_marking_in_progress();
4488 }
4489 }
4491 // We can now guarantee that the global stack is empty, since
4492 // all other tasks have finished. We separated the guarantees so
4493 // that, if a condition is false, we can immediately find out
4494 // which one.
4495 guarantee(_cm->out_of_regions(), "only way to reach here");
4496 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4497 guarantee(_task_queue->size() == 0, "only way to reach here");
4498 guarantee(!_cm->has_overflown(), "only way to reach here");
4499 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4501 if (_cm->verbose_low()) {
4502 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4503 }
4504 } else {
4505 // Apparently there's more work to do. Let's abort this task. It
4506 // will restart it and we can hopefully find more things to do.
4508 if (_cm->verbose_low()) {
4509 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4510 _worker_id);
4511 }
4513 set_has_aborted();
4514 statsOnly( ++_aborted_termination );
4515 }
4516 }
4518 // Mainly for debugging purposes to make sure that a pointer to the
4519 // closure which was statically allocated in this frame doesn't
4520 // escape it by accident.
4521 set_cm_oop_closure(NULL);
4522 double end_time_ms = os::elapsedVTime() * 1000.0;
4523 double elapsed_time_ms = end_time_ms - _start_time_ms;
4524 // Update the step history.
4525 _step_times_ms.add(elapsed_time_ms);
4527 if (has_aborted()) {
4528 // The task was aborted for some reason.
4530 statsOnly( ++_aborted );
4532 if (_has_timed_out) {
4533 double diff_ms = elapsed_time_ms - _time_target_ms;
4534 // Keep statistics of how well we did with respect to hitting
4535 // our target only if we actually timed out (if we aborted for
4536 // other reasons, then the results might get skewed).
4537 _marking_step_diffs_ms.add(diff_ms);
4538 }
4540 if (_cm->has_overflown()) {
4541 // This is the interesting one. We aborted because a global
4542 // overflow was raised. This means we have to restart the
4543 // marking phase and start iterating over regions. However, in
4544 // order to do this we have to make sure that all tasks stop
4545 // what they are doing and re-initialise in a safe manner. We
4546 // will achieve this with the use of two barrier sync points.
4548 if (_cm->verbose_low()) {
4549 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4550 }
4552 if (!is_serial) {
4553 // We only need to enter the sync barrier if being called
4554 // from a parallel context
4555 _cm->enter_first_sync_barrier(_worker_id);
4557 // When we exit this sync barrier we know that all tasks have
4558 // stopped doing marking work. So, it's now safe to
4559 // re-initialise our data structures. At the end of this method,
4560 // task 0 will clear the global data structures.
4561 }
4563 statsOnly( ++_aborted_overflow );
4565 // We clear the local state of this task...
4566 clear_region_fields();
4568 if (!is_serial) {
4569 // ...and enter the second barrier.
4570 _cm->enter_second_sync_barrier(_worker_id);
4571 }
4572 // At this point, if we're during the concurrent phase of
4573 // marking, everything has been re-initialized and we're
4574 // ready to restart.
4575 }
4577 if (_cm->verbose_low()) {
4578 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4579 "elapsed = %1.2lfms <<<<<<<<<<",
4580 _worker_id, _time_target_ms, elapsed_time_ms);
4581 if (_cm->has_aborted()) {
4582 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4583 _worker_id);
4584 }
4585 }
4586 } else {
4587 if (_cm->verbose_low()) {
4588 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4589 "elapsed = %1.2lfms <<<<<<<<<<",
4590 _worker_id, _time_target_ms, elapsed_time_ms);
4591 }
4592 }
4594 _claimed = false;
4595 }
4597 CMTask::CMTask(uint worker_id,
4598 ConcurrentMark* cm,
4599 size_t* marked_bytes,
4600 BitMap* card_bm,
4601 CMTaskQueue* task_queue,
4602 CMTaskQueueSet* task_queues)
4603 : _g1h(G1CollectedHeap::heap()),
4604 _worker_id(worker_id), _cm(cm),
4605 _claimed(false),
4606 _nextMarkBitMap(NULL), _hash_seed(17),
4607 _task_queue(task_queue),
4608 _task_queues(task_queues),
4609 _cm_oop_closure(NULL),
4610 _marked_bytes_array(marked_bytes),
4611 _card_bm(card_bm) {
4612 guarantee(task_queue != NULL, "invariant");
4613 guarantee(task_queues != NULL, "invariant");
4615 statsOnly( _clock_due_to_scanning = 0;
4616 _clock_due_to_marking = 0 );
4618 _marking_step_diffs_ms.add(0.5);
4619 }
4621 // These are formatting macros that are used below to ensure
4622 // consistent formatting. The *_H_* versions are used to format the
4623 // header for a particular value and they should be kept consistent
4624 // with the corresponding macro. Also note that most of the macros add
4625 // the necessary white space (as a prefix) which makes them a bit
4626 // easier to compose.
4628 // All the output lines are prefixed with this string to be able to
4629 // identify them easily in a large log file.
4630 #define G1PPRL_LINE_PREFIX "###"
4632 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT
4633 #ifdef _LP64
4634 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4635 #else // _LP64
4636 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4637 #endif // _LP64
4639 // For per-region info
4640 #define G1PPRL_TYPE_FORMAT " %-4s"
4641 #define G1PPRL_TYPE_H_FORMAT " %4s"
4642 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9)
4643 #define G1PPRL_BYTE_H_FORMAT " %9s"
4644 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4645 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4647 // For summary info
4648 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT
4649 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT
4650 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB"
4651 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
4653 G1PrintRegionLivenessInfoClosure::
4654 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4655 : _out(out),
4656 _total_used_bytes(0), _total_capacity_bytes(0),
4657 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4658 _hum_used_bytes(0), _hum_capacity_bytes(0),
4659 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4660 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4661 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4662 MemRegion g1_reserved = g1h->g1_reserved();
4663 double now = os::elapsedTime();
4665 // Print the header of the output.
4666 _out->cr();
4667 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4668 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4669 G1PPRL_SUM_ADDR_FORMAT("reserved")
4670 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4671 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4672 HeapRegion::GrainBytes);
4673 _out->print_cr(G1PPRL_LINE_PREFIX);
4674 _out->print_cr(G1PPRL_LINE_PREFIX
4675 G1PPRL_TYPE_H_FORMAT
4676 G1PPRL_ADDR_BASE_H_FORMAT
4677 G1PPRL_BYTE_H_FORMAT
4678 G1PPRL_BYTE_H_FORMAT
4679 G1PPRL_BYTE_H_FORMAT
4680 G1PPRL_DOUBLE_H_FORMAT
4681 G1PPRL_BYTE_H_FORMAT
4682 G1PPRL_BYTE_H_FORMAT,
4683 "type", "address-range",
4684 "used", "prev-live", "next-live", "gc-eff",
4685 "remset", "code-roots");
4686 _out->print_cr(G1PPRL_LINE_PREFIX
4687 G1PPRL_TYPE_H_FORMAT
4688 G1PPRL_ADDR_BASE_H_FORMAT
4689 G1PPRL_BYTE_H_FORMAT
4690 G1PPRL_BYTE_H_FORMAT
4691 G1PPRL_BYTE_H_FORMAT
4692 G1PPRL_DOUBLE_H_FORMAT
4693 G1PPRL_BYTE_H_FORMAT
4694 G1PPRL_BYTE_H_FORMAT,
4695 "", "",
4696 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4697 "(bytes)", "(bytes)");
4698 }
4700 // It takes as a parameter a reference to one of the _hum_* fields, it
4701 // deduces the corresponding value for a region in a humongous region
4702 // series (either the region size, or what's left if the _hum_* field
4703 // is < the region size), and updates the _hum_* field accordingly.
4704 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4705 size_t bytes = 0;
4706 // The > 0 check is to deal with the prev and next live bytes which
4707 // could be 0.
4708 if (*hum_bytes > 0) {
4709 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4710 *hum_bytes -= bytes;
4711 }
4712 return bytes;
4713 }
4715 // It deduces the values for a region in a humongous region series
4716 // from the _hum_* fields and updates those accordingly. It assumes
4717 // that that _hum_* fields have already been set up from the "starts
4718 // humongous" region and we visit the regions in address order.
4719 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4720 size_t* capacity_bytes,
4721 size_t* prev_live_bytes,
4722 size_t* next_live_bytes) {
4723 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4724 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4725 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4726 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4727 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4728 }
4730 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4731 const char* type = r->get_type_str();
4732 HeapWord* bottom = r->bottom();
4733 HeapWord* end = r->end();
4734 size_t capacity_bytes = r->capacity();
4735 size_t used_bytes = r->used();
4736 size_t prev_live_bytes = r->live_bytes();
4737 size_t next_live_bytes = r->next_live_bytes();
4738 double gc_eff = r->gc_efficiency();
4739 size_t remset_bytes = r->rem_set()->mem_size();
4740 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4742 if (r->startsHumongous()) {
4743 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4744 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4745 "they should have been zeroed after the last time we used them");
4746 // Set up the _hum_* fields.
4747 _hum_capacity_bytes = capacity_bytes;
4748 _hum_used_bytes = used_bytes;
4749 _hum_prev_live_bytes = prev_live_bytes;
4750 _hum_next_live_bytes = next_live_bytes;
4751 get_hum_bytes(&used_bytes, &capacity_bytes,
4752 &prev_live_bytes, &next_live_bytes);
4753 end = bottom + HeapRegion::GrainWords;
4754 } else if (r->continuesHumongous()) {
4755 get_hum_bytes(&used_bytes, &capacity_bytes,
4756 &prev_live_bytes, &next_live_bytes);
4757 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4758 }
4760 _total_used_bytes += used_bytes;
4761 _total_capacity_bytes += capacity_bytes;
4762 _total_prev_live_bytes += prev_live_bytes;
4763 _total_next_live_bytes += next_live_bytes;
4764 _total_remset_bytes += remset_bytes;
4765 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4767 // Print a line for this particular region.
4768 _out->print_cr(G1PPRL_LINE_PREFIX
4769 G1PPRL_TYPE_FORMAT
4770 G1PPRL_ADDR_BASE_FORMAT
4771 G1PPRL_BYTE_FORMAT
4772 G1PPRL_BYTE_FORMAT
4773 G1PPRL_BYTE_FORMAT
4774 G1PPRL_DOUBLE_FORMAT
4775 G1PPRL_BYTE_FORMAT
4776 G1PPRL_BYTE_FORMAT,
4777 type, p2i(bottom), p2i(end),
4778 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4779 remset_bytes, strong_code_roots_bytes);
4781 return false;
4782 }
4784 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4785 // add static memory usages to remembered set sizes
4786 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4787 // Print the footer of the output.
4788 _out->print_cr(G1PPRL_LINE_PREFIX);
4789 _out->print_cr(G1PPRL_LINE_PREFIX
4790 " SUMMARY"
4791 G1PPRL_SUM_MB_FORMAT("capacity")
4792 G1PPRL_SUM_MB_PERC_FORMAT("used")
4793 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4794 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4795 G1PPRL_SUM_MB_FORMAT("remset")
4796 G1PPRL_SUM_MB_FORMAT("code-roots"),
4797 bytes_to_mb(_total_capacity_bytes),
4798 bytes_to_mb(_total_used_bytes),
4799 perc(_total_used_bytes, _total_capacity_bytes),
4800 bytes_to_mb(_total_prev_live_bytes),
4801 perc(_total_prev_live_bytes, _total_capacity_bytes),
4802 bytes_to_mb(_total_next_live_bytes),
4803 perc(_total_next_live_bytes, _total_capacity_bytes),
4804 bytes_to_mb(_total_remset_bytes),
4805 bytes_to_mb(_total_strong_code_roots_bytes));
4806 _out->cr();
4807 }