1 /*
2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "gc_implementation/g1/concurrentMark.inline.hpp"
28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
32 #include "gc_implementation/g1/g1Log.hpp"
33 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
34 #include "gc_implementation/g1/g1RemSet.hpp"
35 #include "gc_implementation/g1/heapRegion.inline.hpp"
36 #include "gc_implementation/g1/heapRegionRemSet.hpp"
37 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
38 #include "gc_implementation/shared/vmGCOperations.hpp"
39 #include "gc_implementation/shared/gcTimer.hpp"
40 #include "gc_implementation/shared/gcTrace.hpp"
41 #include "gc_implementation/shared/gcTraceTime.hpp"
42 #include "memory/genOopClosures.inline.hpp"
43 #include "memory/referencePolicy.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "oops/oop.inline.hpp"
46 #include "runtime/handles.inline.hpp"
47 #include "runtime/java.hpp"
48 #include "services/memTracker.hpp"
49
50 // Concurrent marking bit map wrapper
51
52 CMBitMapRO::CMBitMapRO(int shifter) :
53 _bm(),
54 _shifter(shifter) {
55 _bmStartWord = 0;
56 _bmWordSize = 0;
57 }
58
59 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr,
60 HeapWord* limit) const {
61 // First we must round addr *up* to a possible object boundary.
62 addr = (HeapWord*)align_size_up((intptr_t)addr,
63 HeapWordSize << _shifter);
64 size_t addrOffset = heapWordToOffset(addr);
65 if (limit == NULL) {
66 limit = _bmStartWord + _bmWordSize;
67 }
68 size_t limitOffset = heapWordToOffset(limit);
69 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
70 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
71 assert(nextAddr >= addr, "get_next_one postcondition");
72 assert(nextAddr == limit || isMarked(nextAddr),
73 "get_next_one postcondition");
74 return nextAddr;
75 }
76
77 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr,
78 HeapWord* limit) const {
79 size_t addrOffset = heapWordToOffset(addr);
80 if (limit == NULL) {
81 limit = _bmStartWord + _bmWordSize;
82 }
83 size_t limitOffset = heapWordToOffset(limit);
84 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
85 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
86 assert(nextAddr >= addr, "get_next_one postcondition");
87 assert(nextAddr == limit || !isMarked(nextAddr),
88 "get_next_one postcondition");
89 return nextAddr;
90 }
91
92 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
93 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
94 return (int) (diff >> _shifter);
95 }
96
97 #ifndef PRODUCT
98 bool CMBitMapRO::covers(ReservedSpace heap_rs) const {
99 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
100 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
101 "size inconsistency");
102 return _bmStartWord == (HeapWord*)(heap_rs.base()) &&
103 _bmWordSize == heap_rs.size()>>LogHeapWordSize;
104 }
105 #endif
106
107 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
108 _bm.print_on_error(st, prefix);
109 }
110
111 bool CMBitMap::allocate(ReservedSpace heap_rs) {
112 _bmStartWord = (HeapWord*)(heap_rs.base());
113 _bmWordSize = heap_rs.size()/HeapWordSize; // heap_rs.size() is in bytes
114 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
115 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
116 if (!brs.is_reserved()) {
117 warning("ConcurrentMark marking bit map allocation failure");
118 return false;
119 }
120 MemTracker::record_virtual_memory_type((address)brs.base(), mtGC);
121 // For now we'll just commit all of the bit map up front.
122 // Later on we'll try to be more parsimonious with swap.
123 if (!_virtual_space.initialize(brs, brs.size())) {
124 warning("ConcurrentMark marking bit map backing store failure");
125 return false;
126 }
127 assert(_virtual_space.committed_size() == brs.size(),
128 "didn't reserve backing store for all of concurrent marking bit map?");
129 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
130 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
131 _bmWordSize, "inconsistency in bit map sizing");
132 _bm.set_size(_bmWordSize >> _shifter);
133 return true;
134 }
135
136 void CMBitMap::clearAll() {
137 _bm.clear();
138 return;
139 }
140
141 void CMBitMap::markRange(MemRegion mr) {
142 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
143 assert(!mr.is_empty(), "unexpected empty region");
144 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
145 ((HeapWord *) mr.end())),
146 "markRange memory region end is not card aligned");
147 // convert address range into offset range
148 _bm.at_put_range(heapWordToOffset(mr.start()),
149 heapWordToOffset(mr.end()), true);
150 }
151
152 void CMBitMap::clearRange(MemRegion mr) {
153 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
154 assert(!mr.is_empty(), "unexpected empty region");
155 // convert address range into offset range
156 _bm.at_put_range(heapWordToOffset(mr.start()),
157 heapWordToOffset(mr.end()), false);
158 }
159
160 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
161 HeapWord* end_addr) {
162 HeapWord* start = getNextMarkedWordAddress(addr);
163 start = MIN2(start, end_addr);
164 HeapWord* end = getNextUnmarkedWordAddress(start);
165 end = MIN2(end, end_addr);
166 assert(start <= end, "Consistency check");
167 MemRegion mr(start, end);
168 if (!mr.is_empty()) {
169 clearRange(mr);
170 }
171 return mr;
172 }
173
174 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
175 _base(NULL), _cm(cm)
176 #ifdef ASSERT
177 , _drain_in_progress(false)
178 , _drain_in_progress_yields(false)
179 #endif
180 {}
181
182 bool CMMarkStack::allocate(size_t capacity) {
183 // allocate a stack of the requisite depth
184 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
185 if (!rs.is_reserved()) {
186 warning("ConcurrentMark MarkStack allocation failure");
187 return false;
188 }
189 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
190 if (!_virtual_space.initialize(rs, rs.size())) {
191 warning("ConcurrentMark MarkStack backing store failure");
192 // Release the virtual memory reserved for the marking stack
193 rs.release();
194 return false;
195 }
196 assert(_virtual_space.committed_size() == rs.size(),
197 "Didn't reserve backing store for all of ConcurrentMark stack?");
198 _base = (oop*) _virtual_space.low();
199 setEmpty();
200 _capacity = (jint) capacity;
201 _saved_index = -1;
202 _should_expand = false;
203 NOT_PRODUCT(_max_depth = 0);
204 return true;
205 }
206
207 void CMMarkStack::expand() {
208 // Called, during remark, if we've overflown the marking stack during marking.
209 assert(isEmpty(), "stack should been emptied while handling overflow");
210 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
211 // Clear expansion flag
212 _should_expand = false;
213 if (_capacity == (jint) MarkStackSizeMax) {
214 if (PrintGCDetails && Verbose) {
215 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
216 }
217 return;
218 }
219 // Double capacity if possible
220 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
221 // Do not give up existing stack until we have managed to
222 // get the double capacity that we desired.
223 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
224 sizeof(oop)));
225 if (rs.is_reserved()) {
226 // Release the backing store associated with old stack
227 _virtual_space.release();
228 // Reinitialize virtual space for new stack
229 if (!_virtual_space.initialize(rs, rs.size())) {
230 fatal("Not enough swap for expanded marking stack capacity");
231 }
232 _base = (oop*)(_virtual_space.low());
233 _index = 0;
234 _capacity = new_capacity;
235 } else {
236 if (PrintGCDetails && Verbose) {
237 // Failed to double capacity, continue;
238 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
239 SIZE_FORMAT"K to " SIZE_FORMAT"K",
240 _capacity / K, new_capacity / K);
241 }
242 }
243 }
244
245 void CMMarkStack::set_should_expand() {
246 // If we're resetting the marking state because of an
247 // marking stack overflow, record that we should, if
248 // possible, expand the stack.
249 _should_expand = _cm->has_overflown();
250 }
251
252 CMMarkStack::~CMMarkStack() {
253 if (_base != NULL) {
254 _base = NULL;
255 _virtual_space.release();
256 }
257 }
258
259 void CMMarkStack::par_push(oop ptr) {
260 while (true) {
261 if (isFull()) {
262 _overflow = true;
263 return;
264 }
265 // Otherwise...
266 jint index = _index;
267 jint next_index = index+1;
268 jint res = Atomic::cmpxchg(next_index, &_index, index);
269 if (res == index) {
270 _base[index] = ptr;
271 // Note that we don't maintain this atomically. We could, but it
272 // doesn't seem necessary.
273 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
274 return;
275 }
276 // Otherwise, we need to try again.
277 }
278 }
279
280 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
281 while (true) {
282 if (isFull()) {
283 _overflow = true;
284 return;
285 }
286 // Otherwise...
287 jint index = _index;
288 jint next_index = index + n;
289 if (next_index > _capacity) {
290 _overflow = true;
291 return;
292 }
293 jint res = Atomic::cmpxchg(next_index, &_index, index);
294 if (res == index) {
295 for (int i = 0; i < n; i++) {
296 int ind = index + i;
297 assert(ind < _capacity, "By overflow test above.");
298 _base[ind] = ptr_arr[i];
299 }
300 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
301 return;
302 }
303 // Otherwise, we need to try again.
304 }
305 }
306
307 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
308 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
309 jint start = _index;
310 jint next_index = start + n;
311 if (next_index > _capacity) {
312 _overflow = true;
313 return;
314 }
315 // Otherwise.
316 _index = next_index;
317 for (int i = 0; i < n; i++) {
318 int ind = start + i;
319 assert(ind < _capacity, "By overflow test above.");
320 _base[ind] = ptr_arr[i];
321 }
322 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
323 }
324
325 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
326 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
327 jint index = _index;
328 if (index == 0) {
329 *n = 0;
330 return false;
331 } else {
332 int k = MIN2(max, index);
333 jint new_ind = index - k;
334 for (int j = 0; j < k; j++) {
335 ptr_arr[j] = _base[new_ind + j];
336 }
337 _index = new_ind;
338 *n = k;
339 return true;
340 }
341 }
342
343 template<class OopClosureClass>
344 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
345 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
346 || SafepointSynchronize::is_at_safepoint(),
347 "Drain recursion must be yield-safe.");
348 bool res = true;
349 debug_only(_drain_in_progress = true);
350 debug_only(_drain_in_progress_yields = yield_after);
351 while (!isEmpty()) {
352 oop newOop = pop();
353 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
354 assert(newOop->is_oop(), "Expected an oop");
355 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
356 "only grey objects on this stack");
357 newOop->oop_iterate(cl);
358 if (yield_after && _cm->do_yield_check()) {
359 res = false;
360 break;
361 }
362 }
363 debug_only(_drain_in_progress = false);
364 return res;
365 }
366
367 void CMMarkStack::note_start_of_gc() {
368 assert(_saved_index == -1,
369 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
370 _saved_index = _index;
371 }
372
373 void CMMarkStack::note_end_of_gc() {
374 // This is intentionally a guarantee, instead of an assert. If we
375 // accidentally add something to the mark stack during GC, it
376 // will be a correctness issue so it's better if we crash. we'll
377 // only check this once per GC anyway, so it won't be a performance
378 // issue in any way.
379 guarantee(_saved_index == _index,
380 err_msg("saved index: %d index: %d", _saved_index, _index));
381 _saved_index = -1;
382 }
383
384 void CMMarkStack::oops_do(OopClosure* f) {
385 assert(_saved_index == _index,
386 err_msg("saved index: %d index: %d", _saved_index, _index));
387 for (int i = 0; i < _index; i += 1) {
388 f->do_oop(&_base[i]);
389 }
390 }
391
392 bool ConcurrentMark::not_yet_marked(oop obj) const {
393 return _g1h->is_obj_ill(obj);
394 }
395
396 CMRootRegions::CMRootRegions() :
397 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
398 _should_abort(false), _next_survivor(NULL) { }
399
400 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
401 _young_list = g1h->young_list();
402 _cm = cm;
403 }
404
405 void CMRootRegions::prepare_for_scan() {
406 assert(!scan_in_progress(), "pre-condition");
407
408 // Currently, only survivors can be root regions.
409 assert(_next_survivor == NULL, "pre-condition");
410 _next_survivor = _young_list->first_survivor_region();
411 _scan_in_progress = (_next_survivor != NULL);
412 _should_abort = false;
413 }
414
415 HeapRegion* CMRootRegions::claim_next() {
416 if (_should_abort) {
417 // If someone has set the should_abort flag, we return NULL to
418 // force the caller to bail out of their loop.
419 return NULL;
420 }
421
422 // Currently, only survivors can be root regions.
423 HeapRegion* res = _next_survivor;
424 if (res != NULL) {
425 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
426 // Read it again in case it changed while we were waiting for the lock.
427 res = _next_survivor;
428 if (res != NULL) {
429 if (res == _young_list->last_survivor_region()) {
430 // We just claimed the last survivor so store NULL to indicate
431 // that we're done.
432 _next_survivor = NULL;
433 } else {
434 _next_survivor = res->get_next_young_region();
435 }
436 } else {
437 // Someone else claimed the last survivor while we were trying
438 // to take the lock so nothing else to do.
439 }
440 }
441 assert(res == NULL || res->is_survivor(), "post-condition");
442
443 return res;
444 }
445
446 void CMRootRegions::scan_finished() {
447 assert(scan_in_progress(), "pre-condition");
448
449 // Currently, only survivors can be root regions.
450 if (!_should_abort) {
451 assert(_next_survivor == NULL, "we should have claimed all survivors");
452 }
453 _next_survivor = NULL;
454
455 {
456 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
457 _scan_in_progress = false;
458 RootRegionScan_lock->notify_all();
459 }
460 }
461
462 bool CMRootRegions::wait_until_scan_finished() {
463 if (!scan_in_progress()) return false;
464
465 {
466 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
467 while (scan_in_progress()) {
468 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
469 }
470 }
471 return true;
472 }
473
474 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
475 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
476 #endif // _MSC_VER
477
478 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
479 return MAX2((n_par_threads + 2) / 4, 1U);
480 }
481
482 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs) :
483 _g1h(g1h),
484 _markBitMap1(log2_intptr(MinObjAlignment)),
485 _markBitMap2(log2_intptr(MinObjAlignment)),
486 _parallel_marking_threads(0),
487 _max_parallel_marking_threads(0),
488 _sleep_factor(0.0),
489 _marking_task_overhead(1.0),
490 _cleanup_sleep_factor(0.0),
491 _cleanup_task_overhead(1.0),
492 _cleanup_list("Cleanup List"),
493 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
494 _card_bm((heap_rs.size() + CardTableModRefBS::card_size - 1) >>
495 CardTableModRefBS::card_shift,
496 false /* in_resource_area*/),
497
498 _prevMarkBitMap(&_markBitMap1),
499 _nextMarkBitMap(&_markBitMap2),
500
501 _markStack(this),
502 // _finger set in set_non_marking_state
503
504 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
505 // _active_tasks set in set_non_marking_state
506 // _tasks set inside the constructor
507 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
508 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
509
510 _has_overflown(false),
511 _concurrent(false),
512 _has_aborted(false),
513 _restart_for_overflow(false),
514 _concurrent_marking_in_progress(false),
515
516 // _verbose_level set below
517
518 _init_times(),
519 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
520 _cleanup_times(),
521 _total_counting_time(0.0),
522 _total_rs_scrub_time(0.0),
523
524 _parallel_workers(NULL),
525
526 _count_card_bitmaps(NULL),
527 _count_marked_bytes(NULL),
528 _completed_initialization(false) {
529 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
530 if (verbose_level < no_verbose) {
531 verbose_level = no_verbose;
532 }
533 if (verbose_level > high_verbose) {
534 verbose_level = high_verbose;
535 }
536 _verbose_level = verbose_level;
537
538 if (verbose_low()) {
539 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
540 "heap end = " PTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
541 }
542
543 if (!_markBitMap1.allocate(heap_rs)) {
544 warning("Failed to allocate first CM bit map");
545 return;
546 }
547 if (!_markBitMap2.allocate(heap_rs)) {
548 warning("Failed to allocate second CM bit map");
549 return;
550 }
551
552 // Create & start a ConcurrentMark thread.
553 _cmThread = new ConcurrentMarkThread(this);
554 assert(cmThread() != NULL, "CM Thread should have been created");
555 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
556 if (_cmThread->osthread() == NULL) {
557 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
558 }
559
560 assert(CGC_lock != NULL, "Where's the CGC_lock?");
561 assert(_markBitMap1.covers(heap_rs), "_markBitMap1 inconsistency");
562 assert(_markBitMap2.covers(heap_rs), "_markBitMap2 inconsistency");
563
564 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
565 satb_qs.set_buffer_size(G1SATBBufferSize);
566
567 _root_regions.init(_g1h, this);
568
569 if (ConcGCThreads > ParallelGCThreads) {
570 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
571 "than ParallelGCThreads (" UINTX_FORMAT ").",
572 ConcGCThreads, ParallelGCThreads);
573 return;
574 }
575 if (ParallelGCThreads == 0) {
576 // if we are not running with any parallel GC threads we will not
577 // spawn any marking threads either
578 _parallel_marking_threads = 0;
579 _max_parallel_marking_threads = 0;
580 _sleep_factor = 0.0;
581 _marking_task_overhead = 1.0;
582 } else {
583 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
584 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
585 // if both are set
586 _sleep_factor = 0.0;
587 _marking_task_overhead = 1.0;
588 } else if (G1MarkingOverheadPercent > 0) {
589 // We will calculate the number of parallel marking threads based
590 // on a target overhead with respect to the soft real-time goal
591 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
592 double overall_cm_overhead =
593 (double) MaxGCPauseMillis * marking_overhead /
594 (double) GCPauseIntervalMillis;
595 double cpu_ratio = 1.0 / (double) os::processor_count();
596 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
597 double marking_task_overhead =
598 overall_cm_overhead / marking_thread_num *
599 (double) os::processor_count();
600 double sleep_factor =
601 (1.0 - marking_task_overhead) / marking_task_overhead;
602
603 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
604 _sleep_factor = sleep_factor;
605 _marking_task_overhead = marking_task_overhead;
606 } else {
607 // Calculate the number of parallel marking threads by scaling
608 // the number of parallel GC threads.
609 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
610 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
611 _sleep_factor = 0.0;
612 _marking_task_overhead = 1.0;
613 }
614
615 assert(ConcGCThreads > 0, "Should have been set");
616 _parallel_marking_threads = (uint) ConcGCThreads;
617 _max_parallel_marking_threads = _parallel_marking_threads;
618
619 if (parallel_marking_threads() > 1) {
620 _cleanup_task_overhead = 1.0;
621 } else {
622 _cleanup_task_overhead = marking_task_overhead();
623 }
624 _cleanup_sleep_factor =
625 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
626
627 #if 0
628 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
629 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
630 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
631 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
632 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
633 #endif
634
635 guarantee(parallel_marking_threads() > 0, "peace of mind");
636 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
637 _max_parallel_marking_threads, false, true);
638 if (_parallel_workers == NULL) {
639 vm_exit_during_initialization("Failed necessary allocation.");
640 } else {
641 _parallel_workers->initialize_workers();
642 }
643 }
644
645 if (FLAG_IS_DEFAULT(MarkStackSize)) {
646 uintx mark_stack_size =
647 MIN2(MarkStackSizeMax,
648 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
649 // Verify that the calculated value for MarkStackSize is in range.
650 // It would be nice to use the private utility routine from Arguments.
651 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
652 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
653 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
654 mark_stack_size, (uintx) 1, MarkStackSizeMax);
655 return;
656 }
657 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
658 } else {
659 // Verify MarkStackSize is in range.
660 if (FLAG_IS_CMDLINE(MarkStackSize)) {
661 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
662 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
663 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
664 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
665 MarkStackSize, (uintx) 1, MarkStackSizeMax);
666 return;
667 }
668 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
669 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
670 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
671 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
672 MarkStackSize, MarkStackSizeMax);
673 return;
674 }
675 }
676 }
677 }
678
679 if (!_markStack.allocate(MarkStackSize)) {
680 warning("Failed to allocate CM marking stack");
681 return;
682 }
683
684 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
685 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
686
687 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
688 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
689
690 BitMap::idx_t card_bm_size = _card_bm.size();
691
692 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
693 _active_tasks = _max_worker_id;
694
695 size_t max_regions = (size_t) _g1h->max_regions();
696 for (uint i = 0; i < _max_worker_id; ++i) {
697 CMTaskQueue* task_queue = new CMTaskQueue();
698 task_queue->initialize();
699 _task_queues->register_queue(i, task_queue);
700
701 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
702 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
703
704 _tasks[i] = new CMTask(i, this,
705 _count_marked_bytes[i],
706 &_count_card_bitmaps[i],
707 task_queue, _task_queues);
708
709 _accum_task_vtime[i] = 0.0;
710 }
711
712 // Calculate the card number for the bottom of the heap. Used
713 // in biasing indexes into the accounting card bitmaps.
714 _heap_bottom_card_num =
715 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
716 CardTableModRefBS::card_shift);
717
718 // Clear all the liveness counting data
719 clear_all_count_data();
720
721 // so that the call below can read a sensible value
722 _heap_start = (HeapWord*) heap_rs.base();
723 set_non_marking_state();
724 _completed_initialization = true;
725 }
726
727 void ConcurrentMark::update_g1_committed(bool force) {
728 // If concurrent marking is not in progress, then we do not need to
729 // update _heap_end.
730 if (!concurrent_marking_in_progress() && !force) return;
731
732 MemRegion committed = _g1h->g1_committed();
733 assert(committed.start() == _heap_start, "start shouldn't change");
734 HeapWord* new_end = committed.end();
735 if (new_end > _heap_end) {
736 // The heap has been expanded.
737
738 _heap_end = new_end;
739 }
740 // Notice that the heap can also shrink. However, this only happens
741 // during a Full GC (at least currently) and the entire marking
742 // phase will bail out and the task will not be restarted. So, let's
743 // do nothing.
744 }
745
746 void ConcurrentMark::reset() {
747 // Starting values for these two. This should be called in a STW
748 // phase. CM will be notified of any future g1_committed expansions
749 // will be at the end of evacuation pauses, when tasks are
750 // inactive.
751 MemRegion committed = _g1h->g1_committed();
752 _heap_start = committed.start();
753 _heap_end = committed.end();
754
755 // Separated the asserts so that we know which one fires.
756 assert(_heap_start != NULL, "heap bounds should look ok");
757 assert(_heap_end != NULL, "heap bounds should look ok");
758 assert(_heap_start < _heap_end, "heap bounds should look ok");
759
760 // Reset all the marking data structures and any necessary flags
761 reset_marking_state();
762
763 if (verbose_low()) {
764 gclog_or_tty->print_cr("[global] resetting");
765 }
766
767 // We do reset all of them, since different phases will use
768 // different number of active threads. So, it's easiest to have all
769 // of them ready.
770 for (uint i = 0; i < _max_worker_id; ++i) {
771 _tasks[i]->reset(_nextMarkBitMap);
772 }
773
774 // we need this to make sure that the flag is on during the evac
775 // pause with initial mark piggy-backed
776 set_concurrent_marking_in_progress();
777 }
778
779
780 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
781 _markStack.set_should_expand();
782 _markStack.setEmpty(); // Also clears the _markStack overflow flag
783 if (clear_overflow) {
784 clear_has_overflown();
785 } else {
786 assert(has_overflown(), "pre-condition");
787 }
788 _finger = _heap_start;
789
790 for (uint i = 0; i < _max_worker_id; ++i) {
791 CMTaskQueue* queue = _task_queues->queue(i);
792 queue->set_empty();
793 }
794 }
795
796 void ConcurrentMark::set_concurrency(uint active_tasks) {
797 assert(active_tasks <= _max_worker_id, "we should not have more");
798
799 _active_tasks = active_tasks;
800 // Need to update the three data structures below according to the
801 // number of active threads for this phase.
802 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
803 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
804 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
805 }
806
807 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
808 set_concurrency(active_tasks);
809
810 _concurrent = concurrent;
811 // We propagate this to all tasks, not just the active ones.
812 for (uint i = 0; i < _max_worker_id; ++i)
813 _tasks[i]->set_concurrent(concurrent);
814
815 if (concurrent) {
816 set_concurrent_marking_in_progress();
817 } else {
818 // We currently assume that the concurrent flag has been set to
819 // false before we start remark. At this point we should also be
820 // in a STW phase.
821 assert(!concurrent_marking_in_progress(), "invariant");
822 assert(_finger == _heap_end,
823 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
824 p2i(_finger), p2i(_heap_end)));
825 update_g1_committed(true);
826 }
827 }
828
829 void ConcurrentMark::set_non_marking_state() {
830 // We set the global marking state to some default values when we're
831 // not doing marking.
832 reset_marking_state();
833 _active_tasks = 0;
834 clear_concurrent_marking_in_progress();
835 }
836
837 ConcurrentMark::~ConcurrentMark() {
838 // The ConcurrentMark instance is never freed.
839 ShouldNotReachHere();
840 }
841
842 void ConcurrentMark::clearNextBitmap() {
843 G1CollectedHeap* g1h = G1CollectedHeap::heap();
844 G1CollectorPolicy* g1p = g1h->g1_policy();
845
846 // Make sure that the concurrent mark thread looks to still be in
847 // the current cycle.
848 guarantee(cmThread()->during_cycle(), "invariant");
849
850 // We are finishing up the current cycle by clearing the next
851 // marking bitmap and getting it ready for the next cycle. During
852 // this time no other cycle can start. So, let's make sure that this
853 // is the case.
854 guarantee(!g1h->mark_in_progress(), "invariant");
855
856 // clear the mark bitmap (no grey objects to start with).
857 // We need to do this in chunks and offer to yield in between
858 // each chunk.
859 HeapWord* start = _nextMarkBitMap->startWord();
860 HeapWord* end = _nextMarkBitMap->endWord();
861 HeapWord* cur = start;
862 size_t chunkSize = M;
863 while (cur < end) {
864 HeapWord* next = cur + chunkSize;
865 if (next > end) {
866 next = end;
867 }
868 MemRegion mr(cur,next);
869 _nextMarkBitMap->clearRange(mr);
870 cur = next;
871 do_yield_check();
872
873 // Repeat the asserts from above. We'll do them as asserts here to
874 // minimize their overhead on the product. However, we'll have
875 // them as guarantees at the beginning / end of the bitmap
876 // clearing to get some checking in the product.
877 assert(cmThread()->during_cycle(), "invariant");
878 assert(!g1h->mark_in_progress(), "invariant");
879 }
880
881 // Clear the liveness counting data
882 clear_all_count_data();
883
884 // Repeat the asserts from above.
885 guarantee(cmThread()->during_cycle(), "invariant");
886 guarantee(!g1h->mark_in_progress(), "invariant");
887 }
888
889 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
890 public:
891 bool doHeapRegion(HeapRegion* r) {
892 if (!r->continuesHumongous()) {
893 r->note_start_of_marking();
894 }
895 return false;
896 }
897 };
898
899 void ConcurrentMark::checkpointRootsInitialPre() {
900 G1CollectedHeap* g1h = G1CollectedHeap::heap();
901 G1CollectorPolicy* g1p = g1h->g1_policy();
902
903 _has_aborted = false;
904
905 #ifndef PRODUCT
906 if (G1PrintReachableAtInitialMark) {
907 print_reachable("at-cycle-start",
908 VerifyOption_G1UsePrevMarking, true /* all */);
909 }
910 #endif
911
912 // Initialize marking structures. This has to be done in a STW phase.
913 reset();
914
915 // For each region note start of marking.
916 NoteStartOfMarkHRClosure startcl;
917 g1h->heap_region_iterate(&startcl);
918 }
919
920
921 void ConcurrentMark::checkpointRootsInitialPost() {
922 G1CollectedHeap* g1h = G1CollectedHeap::heap();
923
924 // If we force an overflow during remark, the remark operation will
925 // actually abort and we'll restart concurrent marking. If we always
926 // force an overflow during remark we'll never actually complete the
927 // marking phase. So, we initialize this here, at the start of the
928 // cycle, so that at the remaining overflow number will decrease at
929 // every remark and we'll eventually not need to cause one.
930 force_overflow_stw()->init();
931
932 // Start Concurrent Marking weak-reference discovery.
933 ReferenceProcessor* rp = g1h->ref_processor_cm();
934 // enable ("weak") refs discovery
935 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
936 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
937
938 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
939 // This is the start of the marking cycle, we're expected all
940 // threads to have SATB queues with active set to false.
941 satb_mq_set.set_active_all_threads(true, /* new active value */
942 false /* expected_active */);
943
944 _root_regions.prepare_for_scan();
945
946 // update_g1_committed() will be called at the end of an evac pause
947 // when marking is on. So, it's also called at the end of the
948 // initial-mark pause to update the heap end, if the heap expands
949 // during it. No need to call it here.
950 }
951
952 /*
953 * Notice that in the next two methods, we actually leave the STS
954 * during the barrier sync and join it immediately afterwards. If we
955 * do not do this, the following deadlock can occur: one thread could
956 * be in the barrier sync code, waiting for the other thread to also
957 * sync up, whereas another one could be trying to yield, while also
958 * waiting for the other threads to sync up too.
959 *
960 * Note, however, that this code is also used during remark and in
961 * this case we should not attempt to leave / enter the STS, otherwise
962 * we'll either hit an assert (debug / fastdebug) or deadlock
963 * (product). So we should only leave / enter the STS if we are
964 * operating concurrently.
965 *
966 * Because the thread that does the sync barrier has left the STS, it
967 * is possible to be suspended for a Full GC or an evacuation pause
968 * could occur. This is actually safe, since the entering the sync
969 * barrier is one of the last things do_marking_step() does, and it
970 * doesn't manipulate any data structures afterwards.
971 */
972
973 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
974 if (verbose_low()) {
975 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
976 }
977
978 if (concurrent()) {
979 SuspendibleThreadSet::leave();
980 }
981 _first_overflow_barrier_sync.enter();
982 if (concurrent()) {
983 SuspendibleThreadSet::join();
984 }
985 // at this point everyone should have synced up and not be doing any
986 // more work
987
988 if (verbose_low()) {
989 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
990 }
991
992 // If we're executing the concurrent phase of marking, reset the marking
993 // state; otherwise the marking state is reset after reference processing,
994 // during the remark pause.
995 // If we reset here as a result of an overflow during the remark we will
996 // see assertion failures from any subsequent set_concurrency_and_phase()
997 // calls.
998 if (concurrent()) {
999 // let the task associated with with worker 0 do this
1000 if (worker_id == 0) {
1001 // task 0 is responsible for clearing the global data structures
1002 // We should be here because of an overflow. During STW we should
1003 // not clear the overflow flag since we rely on it being true when
1004 // we exit this method to abort the pause and restart concurrent
1005 // marking.
1006 reset_marking_state(true /* clear_overflow */);
1007 force_overflow()->update();
1008
1009 if (G1Log::fine()) {
1010 gclog_or_tty->date_stamp(PrintGCDateStamps);
1011 gclog_or_tty->stamp(PrintGCTimeStamps);
1012 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1013 }
1014 }
1015 }
1016
1017 // after this, each task should reset its own data structures then
1018 // then go into the second barrier
1019 }
1020
1021 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1022 if (verbose_low()) {
1023 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1024 }
1025
1026 if (concurrent()) {
1027 SuspendibleThreadSet::leave();
1028 }
1029 _second_overflow_barrier_sync.enter();
1030 if (concurrent()) {
1031 SuspendibleThreadSet::join();
1032 }
1033 // at this point everything should be re-initialized and ready to go
1034
1035 if (verbose_low()) {
1036 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1037 }
1038 }
1039
1040 #ifndef PRODUCT
1041 void ForceOverflowSettings::init() {
1042 _num_remaining = G1ConcMarkForceOverflow;
1043 _force = false;
1044 update();
1045 }
1046
1047 void ForceOverflowSettings::update() {
1048 if (_num_remaining > 0) {
1049 _num_remaining -= 1;
1050 _force = true;
1051 } else {
1052 _force = false;
1053 }
1054 }
1055
1056 bool ForceOverflowSettings::should_force() {
1057 if (_force) {
1058 _force = false;
1059 return true;
1060 } else {
1061 return false;
1062 }
1063 }
1064 #endif // !PRODUCT
1065
1066 class CMConcurrentMarkingTask: public AbstractGangTask {
1067 private:
1068 ConcurrentMark* _cm;
1069 ConcurrentMarkThread* _cmt;
1070
1071 public:
1072 void work(uint worker_id) {
1073 assert(Thread::current()->is_ConcurrentGC_thread(),
1074 "this should only be done by a conc GC thread");
1075 ResourceMark rm;
1076
1077 double start_vtime = os::elapsedVTime();
1078
1079 SuspendibleThreadSet::join();
1080
1081 assert(worker_id < _cm->active_tasks(), "invariant");
1082 CMTask* the_task = _cm->task(worker_id);
1083 the_task->record_start_time();
1084 if (!_cm->has_aborted()) {
1085 do {
1086 double start_vtime_sec = os::elapsedVTime();
1087 double start_time_sec = os::elapsedTime();
1088 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1089
1090 the_task->do_marking_step(mark_step_duration_ms,
1091 true /* do_termination */,
1092 false /* is_serial*/);
1093
1094 double end_time_sec = os::elapsedTime();
1095 double end_vtime_sec = os::elapsedVTime();
1096 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1097 double elapsed_time_sec = end_time_sec - start_time_sec;
1098 _cm->clear_has_overflown();
1099
1100 bool ret = _cm->do_yield_check(worker_id);
1101
1102 jlong sleep_time_ms;
1103 if (!_cm->has_aborted() && the_task->has_aborted()) {
1104 sleep_time_ms =
1105 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1106 SuspendibleThreadSet::leave();
1107 os::sleep(Thread::current(), sleep_time_ms, false);
1108 SuspendibleThreadSet::join();
1109 }
1110 double end_time2_sec = os::elapsedTime();
1111 double elapsed_time2_sec = end_time2_sec - start_time_sec;
1112
1113 #if 0
1114 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, "
1115 "overhead %1.4lf",
1116 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms,
1117 the_task->conc_overhead(os::elapsedTime()) * 8.0);
1118 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms",
1119 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0);
1120 #endif
1121 } while (!_cm->has_aborted() && the_task->has_aborted());
1122 }
1123 the_task->record_end_time();
1124 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1125
1126 SuspendibleThreadSet::leave();
1127
1128 double end_vtime = os::elapsedVTime();
1129 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1130 }
1131
1132 CMConcurrentMarkingTask(ConcurrentMark* cm,
1133 ConcurrentMarkThread* cmt) :
1134 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1135
1136 ~CMConcurrentMarkingTask() { }
1137 };
1138
1139 // Calculates the number of active workers for a concurrent
1140 // phase.
1141 uint ConcurrentMark::calc_parallel_marking_threads() {
1142 if (G1CollectedHeap::use_parallel_gc_threads()) {
1143 uint n_conc_workers = 0;
1144 if (!UseDynamicNumberOfGCThreads ||
1145 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1146 !ForceDynamicNumberOfGCThreads)) {
1147 n_conc_workers = max_parallel_marking_threads();
1148 } else {
1149 n_conc_workers =
1150 AdaptiveSizePolicy::calc_default_active_workers(
1151 max_parallel_marking_threads(),
1152 1, /* Minimum workers */
1153 parallel_marking_threads(),
1154 Threads::number_of_non_daemon_threads());
1155 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1156 // that scaling has already gone into "_max_parallel_marking_threads".
1157 }
1158 assert(n_conc_workers > 0, "Always need at least 1");
1159 return n_conc_workers;
1160 }
1161 // If we are not running with any parallel GC threads we will not
1162 // have spawned any marking threads either. Hence the number of
1163 // concurrent workers should be 0.
1164 return 0;
1165 }
1166
1167 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1168 // Currently, only survivors can be root regions.
1169 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1170 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1171
1172 const uintx interval = PrefetchScanIntervalInBytes;
1173 HeapWord* curr = hr->bottom();
1174 const HeapWord* end = hr->top();
1175 while (curr < end) {
1176 Prefetch::read(curr, interval);
1177 oop obj = oop(curr);
1178 int size = obj->oop_iterate(&cl);
1179 assert(size == obj->size(), "sanity");
1180 curr += size;
1181 }
1182 }
1183
1184 class CMRootRegionScanTask : public AbstractGangTask {
1185 private:
1186 ConcurrentMark* _cm;
1187
1188 public:
1189 CMRootRegionScanTask(ConcurrentMark* cm) :
1190 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1191
1192 void work(uint worker_id) {
1193 assert(Thread::current()->is_ConcurrentGC_thread(),
1194 "this should only be done by a conc GC thread");
1195
1196 CMRootRegions* root_regions = _cm->root_regions();
1197 HeapRegion* hr = root_regions->claim_next();
1198 while (hr != NULL) {
1199 _cm->scanRootRegion(hr, worker_id);
1200 hr = root_regions->claim_next();
1201 }
1202 }
1203 };
1204
1205 void ConcurrentMark::scanRootRegions() {
1206 // scan_in_progress() will have been set to true only if there was
1207 // at least one root region to scan. So, if it's false, we
1208 // should not attempt to do any further work.
1209 if (root_regions()->scan_in_progress()) {
1210 _parallel_marking_threads = calc_parallel_marking_threads();
1211 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1212 "Maximum number of marking threads exceeded");
1213 uint active_workers = MAX2(1U, parallel_marking_threads());
1214
1215 CMRootRegionScanTask task(this);
1216 if (use_parallel_marking_threads()) {
1217 _parallel_workers->set_active_workers((int) active_workers);
1218 _parallel_workers->run_task(&task);
1219 } else {
1220 task.work(0);
1221 }
1222
1223 // It's possible that has_aborted() is true here without actually
1224 // aborting the survivor scan earlier. This is OK as it's
1225 // mainly used for sanity checking.
1226 root_regions()->scan_finished();
1227 }
1228 }
1229
1230 void ConcurrentMark::markFromRoots() {
1231 // we might be tempted to assert that:
1232 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1233 // "inconsistent argument?");
1234 // However that wouldn't be right, because it's possible that
1235 // a safepoint is indeed in progress as a younger generation
1236 // stop-the-world GC happens even as we mark in this generation.
1237
1238 _restart_for_overflow = false;
1239 force_overflow_conc()->init();
1240
1241 // _g1h has _n_par_threads
1242 _parallel_marking_threads = calc_parallel_marking_threads();
1243 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1244 "Maximum number of marking threads exceeded");
1245
1246 uint active_workers = MAX2(1U, parallel_marking_threads());
1247
1248 // Parallel task terminator is set in "set_concurrency_and_phase()"
1249 set_concurrency_and_phase(active_workers, true /* concurrent */);
1250
1251 CMConcurrentMarkingTask markingTask(this, cmThread());
1252 if (use_parallel_marking_threads()) {
1253 _parallel_workers->set_active_workers((int)active_workers);
1254 // Don't set _n_par_threads because it affects MT in process_strong_roots()
1255 // and the decisions on that MT processing is made elsewhere.
1256 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1257 _parallel_workers->run_task(&markingTask);
1258 } else {
1259 markingTask.work(0);
1260 }
1261 print_stats();
1262 }
1263
1264 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1265 // world is stopped at this checkpoint
1266 assert(SafepointSynchronize::is_at_safepoint(),
1267 "world should be stopped");
1268
1269 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1270
1271 // If a full collection has happened, we shouldn't do this.
1272 if (has_aborted()) {
1273 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1274 return;
1275 }
1276
1277 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1278
1279 if (VerifyDuringGC) {
1280 HandleMark hm; // handle scope
1281 Universe::heap()->prepare_for_verify();
1282 Universe::verify(VerifyOption_G1UsePrevMarking,
1283 " VerifyDuringGC:(before)");
1284 }
1285 g1h->check_bitmaps("Remark Start");
1286
1287 G1CollectorPolicy* g1p = g1h->g1_policy();
1288 g1p->record_concurrent_mark_remark_start();
1289
1290 double start = os::elapsedTime();
1291
1292 checkpointRootsFinalWork();
1293
1294 double mark_work_end = os::elapsedTime();
1295
1296 weakRefsWork(clear_all_soft_refs);
1297
1298 if (has_overflown()) {
1299 // Oops. We overflowed. Restart concurrent marking.
1300 _restart_for_overflow = true;
1301 if (G1TraceMarkStackOverflow) {
1302 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1303 }
1304
1305 // Verify the heap w.r.t. the previous marking bitmap.
1306 if (VerifyDuringGC) {
1307 HandleMark hm; // handle scope
1308 Universe::heap()->prepare_for_verify();
1309 Universe::verify(VerifyOption_G1UsePrevMarking,
1310 " VerifyDuringGC:(overflow)");
1311 }
1312
1313 // Clear the marking state because we will be restarting
1314 // marking due to overflowing the global mark stack.
1315 reset_marking_state();
1316 } else {
1317 // Aggregate the per-task counting data that we have accumulated
1318 // while marking.
1319 aggregate_count_data();
1320
1321 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1322 // We're done with marking.
1323 // This is the end of the marking cycle, we're expected all
1324 // threads to have SATB queues with active set to true.
1325 satb_mq_set.set_active_all_threads(false, /* new active value */
1326 true /* expected_active */);
1327
1328 if (VerifyDuringGC) {
1329 HandleMark hm; // handle scope
1330 Universe::heap()->prepare_for_verify();
1331 Universe::verify(VerifyOption_G1UseNextMarking,
1332 " VerifyDuringGC:(after)");
1333 }
1334 g1h->check_bitmaps("Remark End");
1335 assert(!restart_for_overflow(), "sanity");
1336 // Completely reset the marking state since marking completed
1337 set_non_marking_state();
1338 }
1339
1340 // Expand the marking stack, if we have to and if we can.
1341 if (_markStack.should_expand()) {
1342 _markStack.expand();
1343 }
1344
1345 // Statistics
1346 double now = os::elapsedTime();
1347 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1348 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1349 _remark_times.add((now - start) * 1000.0);
1350
1351 g1p->record_concurrent_mark_remark_end();
1352
1353 G1CMIsAliveClosure is_alive(g1h);
1354 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1355 }
1356
1357 // Base class of the closures that finalize and verify the
1358 // liveness counting data.
1359 class CMCountDataClosureBase: public HeapRegionClosure {
1360 protected:
1361 G1CollectedHeap* _g1h;
1362 ConcurrentMark* _cm;
1363 CardTableModRefBS* _ct_bs;
1364
1365 BitMap* _region_bm;
1366 BitMap* _card_bm;
1367
1368 // Takes a region that's not empty (i.e., it has at least one
1369 // live object in it and sets its corresponding bit on the region
1370 // bitmap to 1. If the region is "starts humongous" it will also set
1371 // to 1 the bits on the region bitmap that correspond to its
1372 // associated "continues humongous" regions.
1373 void set_bit_for_region(HeapRegion* hr) {
1374 assert(!hr->continuesHumongous(), "should have filtered those out");
1375
1376 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1377 if (!hr->startsHumongous()) {
1378 // Normal (non-humongous) case: just set the bit.
1379 _region_bm->par_at_put(index, true);
1380 } else {
1381 // Starts humongous case: calculate how many regions are part of
1382 // this humongous region and then set the bit range.
1383 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1384 _region_bm->par_at_put_range(index, end_index, true);
1385 }
1386 }
1387
1388 public:
1389 CMCountDataClosureBase(G1CollectedHeap* g1h,
1390 BitMap* region_bm, BitMap* card_bm):
1391 _g1h(g1h), _cm(g1h->concurrent_mark()),
1392 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1393 _region_bm(region_bm), _card_bm(card_bm) { }
1394 };
1395
1396 // Closure that calculates the # live objects per region. Used
1397 // for verification purposes during the cleanup pause.
1398 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1399 CMBitMapRO* _bm;
1400 size_t _region_marked_bytes;
1401
1402 public:
1403 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1404 BitMap* region_bm, BitMap* card_bm) :
1405 CMCountDataClosureBase(g1h, region_bm, card_bm),
1406 _bm(bm), _region_marked_bytes(0) { }
1407
1408 bool doHeapRegion(HeapRegion* hr) {
1409
1410 if (hr->continuesHumongous()) {
1411 // We will ignore these here and process them when their
1412 // associated "starts humongous" region is processed (see
1413 // set_bit_for_heap_region()). Note that we cannot rely on their
1414 // associated "starts humongous" region to have their bit set to
1415 // 1 since, due to the region chunking in the parallel region
1416 // iteration, a "continues humongous" region might be visited
1417 // before its associated "starts humongous".
1418 return false;
1419 }
1420
1421 HeapWord* ntams = hr->next_top_at_mark_start();
1422 HeapWord* start = hr->bottom();
1423
1424 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1425 err_msg("Preconditions not met - "
1426 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1427 p2i(start), p2i(ntams), p2i(hr->end())));
1428
1429 // Find the first marked object at or after "start".
1430 start = _bm->getNextMarkedWordAddress(start, ntams);
1431
1432 size_t marked_bytes = 0;
1433
1434 while (start < ntams) {
1435 oop obj = oop(start);
1436 int obj_sz = obj->size();
1437 HeapWord* obj_end = start + obj_sz;
1438
1439 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1440 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1441
1442 // Note: if we're looking at the last region in heap - obj_end
1443 // could be actually just beyond the end of the heap; end_idx
1444 // will then correspond to a (non-existent) card that is also
1445 // just beyond the heap.
1446 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1447 // end of object is not card aligned - increment to cover
1448 // all the cards spanned by the object
1449 end_idx += 1;
1450 }
1451
1452 // Set the bits in the card BM for the cards spanned by this object.
1453 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1454
1455 // Add the size of this object to the number of marked bytes.
1456 marked_bytes += (size_t)obj_sz * HeapWordSize;
1457
1458 // Find the next marked object after this one.
1459 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1460 }
1461
1462 // Mark the allocated-since-marking portion...
1463 HeapWord* top = hr->top();
1464 if (ntams < top) {
1465 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1466 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1467
1468 // Note: if we're looking at the last region in heap - top
1469 // could be actually just beyond the end of the heap; end_idx
1470 // will then correspond to a (non-existent) card that is also
1471 // just beyond the heap.
1472 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1473 // end of object is not card aligned - increment to cover
1474 // all the cards spanned by the object
1475 end_idx += 1;
1476 }
1477 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1478
1479 // This definitely means the region has live objects.
1480 set_bit_for_region(hr);
1481 }
1482
1483 // Update the live region bitmap.
1484 if (marked_bytes > 0) {
1485 set_bit_for_region(hr);
1486 }
1487
1488 // Set the marked bytes for the current region so that
1489 // it can be queried by a calling verification routine
1490 _region_marked_bytes = marked_bytes;
1491
1492 return false;
1493 }
1494
1495 size_t region_marked_bytes() const { return _region_marked_bytes; }
1496 };
1497
1498 // Heap region closure used for verifying the counting data
1499 // that was accumulated concurrently and aggregated during
1500 // the remark pause. This closure is applied to the heap
1501 // regions during the STW cleanup pause.
1502
1503 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1504 G1CollectedHeap* _g1h;
1505 ConcurrentMark* _cm;
1506 CalcLiveObjectsClosure _calc_cl;
1507 BitMap* _region_bm; // Region BM to be verified
1508 BitMap* _card_bm; // Card BM to be verified
1509 bool _verbose; // verbose output?
1510
1511 BitMap* _exp_region_bm; // Expected Region BM values
1512 BitMap* _exp_card_bm; // Expected card BM values
1513
1514 int _failures;
1515
1516 public:
1517 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1518 BitMap* region_bm,
1519 BitMap* card_bm,
1520 BitMap* exp_region_bm,
1521 BitMap* exp_card_bm,
1522 bool verbose) :
1523 _g1h(g1h), _cm(g1h->concurrent_mark()),
1524 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1525 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1526 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1527 _failures(0) { }
1528
1529 int failures() const { return _failures; }
1530
1531 bool doHeapRegion(HeapRegion* hr) {
1532 if (hr->continuesHumongous()) {
1533 // We will ignore these here and process them when their
1534 // associated "starts humongous" region is processed (see
1535 // set_bit_for_heap_region()). Note that we cannot rely on their
1536 // associated "starts humongous" region to have their bit set to
1537 // 1 since, due to the region chunking in the parallel region
1538 // iteration, a "continues humongous" region might be visited
1539 // before its associated "starts humongous".
1540 return false;
1541 }
1542
1543 int failures = 0;
1544
1545 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1546 // this region and set the corresponding bits in the expected region
1547 // and card bitmaps.
1548 bool res = _calc_cl.doHeapRegion(hr);
1549 assert(res == false, "should be continuing");
1550
1551 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1552 Mutex::_no_safepoint_check_flag);
1553
1554 // Verify the marked bytes for this region.
1555 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1556 size_t act_marked_bytes = hr->next_marked_bytes();
1557
1558 // We're not OK if expected marked bytes > actual marked bytes. It means
1559 // we have missed accounting some objects during the actual marking.
1560 if (exp_marked_bytes > act_marked_bytes) {
1561 if (_verbose) {
1562 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1563 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1564 hr->hrs_index(), exp_marked_bytes, act_marked_bytes);
1565 }
1566 failures += 1;
1567 }
1568
1569 // Verify the bit, for this region, in the actual and expected
1570 // (which was just calculated) region bit maps.
1571 // We're not OK if the bit in the calculated expected region
1572 // bitmap is set and the bit in the actual region bitmap is not.
1573 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1574
1575 bool expected = _exp_region_bm->at(index);
1576 bool actual = _region_bm->at(index);
1577 if (expected && !actual) {
1578 if (_verbose) {
1579 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1580 "expected: %s, actual: %s",
1581 hr->hrs_index(),
1582 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1583 }
1584 failures += 1;
1585 }
1586
1587 // Verify that the card bit maps for the cards spanned by the current
1588 // region match. We have an error if we have a set bit in the expected
1589 // bit map and the corresponding bit in the actual bitmap is not set.
1590
1591 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1592 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1593
1594 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1595 expected = _exp_card_bm->at(i);
1596 actual = _card_bm->at(i);
1597
1598 if (expected && !actual) {
1599 if (_verbose) {
1600 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1601 "expected: %s, actual: %s",
1602 hr->hrs_index(), i,
1603 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1604 }
1605 failures += 1;
1606 }
1607 }
1608
1609 if (failures > 0 && _verbose) {
1610 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1611 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1612 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1613 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1614 }
1615
1616 _failures += failures;
1617
1618 // We could stop iteration over the heap when we
1619 // find the first violating region by returning true.
1620 return false;
1621 }
1622 };
1623
1624 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1625 protected:
1626 G1CollectedHeap* _g1h;
1627 ConcurrentMark* _cm;
1628 BitMap* _actual_region_bm;
1629 BitMap* _actual_card_bm;
1630
1631 uint _n_workers;
1632
1633 BitMap* _expected_region_bm;
1634 BitMap* _expected_card_bm;
1635
1636 int _failures;
1637 bool _verbose;
1638
1639 public:
1640 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1641 BitMap* region_bm, BitMap* card_bm,
1642 BitMap* expected_region_bm, BitMap* expected_card_bm)
1643 : AbstractGangTask("G1 verify final counting"),
1644 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1645 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1646 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1647 _failures(0), _verbose(false),
1648 _n_workers(0) {
1649 assert(VerifyDuringGC, "don't call this otherwise");
1650
1651 // Use the value already set as the number of active threads
1652 // in the call to run_task().
1653 if (G1CollectedHeap::use_parallel_gc_threads()) {
1654 assert( _g1h->workers()->active_workers() > 0,
1655 "Should have been previously set");
1656 _n_workers = _g1h->workers()->active_workers();
1657 } else {
1658 _n_workers = 1;
1659 }
1660
1661 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1662 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1663
1664 _verbose = _cm->verbose_medium();
1665 }
1666
1667 void work(uint worker_id) {
1668 assert(worker_id < _n_workers, "invariant");
1669
1670 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1671 _actual_region_bm, _actual_card_bm,
1672 _expected_region_bm,
1673 _expected_card_bm,
1674 _verbose);
1675
1676 if (G1CollectedHeap::use_parallel_gc_threads()) {
1677 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1678 worker_id,
1679 _n_workers,
1680 HeapRegion::VerifyCountClaimValue);
1681 } else {
1682 _g1h->heap_region_iterate(&verify_cl);
1683 }
1684
1685 Atomic::add(verify_cl.failures(), &_failures);
1686 }
1687
1688 int failures() const { return _failures; }
1689 };
1690
1691 // Closure that finalizes the liveness counting data.
1692 // Used during the cleanup pause.
1693 // Sets the bits corresponding to the interval [NTAMS, top]
1694 // (which contains the implicitly live objects) in the
1695 // card liveness bitmap. Also sets the bit for each region,
1696 // containing live data, in the region liveness bitmap.
1697
1698 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1699 public:
1700 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1701 BitMap* region_bm,
1702 BitMap* card_bm) :
1703 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1704
1705 bool doHeapRegion(HeapRegion* hr) {
1706
1707 if (hr->continuesHumongous()) {
1708 // We will ignore these here and process them when their
1709 // associated "starts humongous" region is processed (see
1710 // set_bit_for_heap_region()). Note that we cannot rely on their
1711 // associated "starts humongous" region to have their bit set to
1712 // 1 since, due to the region chunking in the parallel region
1713 // iteration, a "continues humongous" region might be visited
1714 // before its associated "starts humongous".
1715 return false;
1716 }
1717
1718 HeapWord* ntams = hr->next_top_at_mark_start();
1719 HeapWord* top = hr->top();
1720
1721 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1722
1723 // Mark the allocated-since-marking portion...
1724 if (ntams < top) {
1725 // This definitely means the region has live objects.
1726 set_bit_for_region(hr);
1727
1728 // Now set the bits in the card bitmap for [ntams, top)
1729 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1730 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1731
1732 // Note: if we're looking at the last region in heap - top
1733 // could be actually just beyond the end of the heap; end_idx
1734 // will then correspond to a (non-existent) card that is also
1735 // just beyond the heap.
1736 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1737 // end of object is not card aligned - increment to cover
1738 // all the cards spanned by the object
1739 end_idx += 1;
1740 }
1741
1742 assert(end_idx <= _card_bm->size(),
1743 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1744 end_idx, _card_bm->size()));
1745 assert(start_idx < _card_bm->size(),
1746 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1747 start_idx, _card_bm->size()));
1748
1749 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1750 }
1751
1752 // Set the bit for the region if it contains live data
1753 if (hr->next_marked_bytes() > 0) {
1754 set_bit_for_region(hr);
1755 }
1756
1757 return false;
1758 }
1759 };
1760
1761 class G1ParFinalCountTask: public AbstractGangTask {
1762 protected:
1763 G1CollectedHeap* _g1h;
1764 ConcurrentMark* _cm;
1765 BitMap* _actual_region_bm;
1766 BitMap* _actual_card_bm;
1767
1768 uint _n_workers;
1769
1770 public:
1771 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1772 : AbstractGangTask("G1 final counting"),
1773 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1774 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1775 _n_workers(0) {
1776 // Use the value already set as the number of active threads
1777 // in the call to run_task().
1778 if (G1CollectedHeap::use_parallel_gc_threads()) {
1779 assert( _g1h->workers()->active_workers() > 0,
1780 "Should have been previously set");
1781 _n_workers = _g1h->workers()->active_workers();
1782 } else {
1783 _n_workers = 1;
1784 }
1785 }
1786
1787 void work(uint worker_id) {
1788 assert(worker_id < _n_workers, "invariant");
1789
1790 FinalCountDataUpdateClosure final_update_cl(_g1h,
1791 _actual_region_bm,
1792 _actual_card_bm);
1793
1794 if (G1CollectedHeap::use_parallel_gc_threads()) {
1795 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1796 worker_id,
1797 _n_workers,
1798 HeapRegion::FinalCountClaimValue);
1799 } else {
1800 _g1h->heap_region_iterate(&final_update_cl);
1801 }
1802 }
1803 };
1804
1805 class G1ParNoteEndTask;
1806
1807 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1808 G1CollectedHeap* _g1;
1809 size_t _max_live_bytes;
1810 uint _regions_claimed;
1811 size_t _freed_bytes;
1812 FreeRegionList* _local_cleanup_list;
1813 HeapRegionSetCount _old_regions_removed;
1814 HeapRegionSetCount _humongous_regions_removed;
1815 HRRSCleanupTask* _hrrs_cleanup_task;
1816 double _claimed_region_time;
1817 double _max_region_time;
1818
1819 public:
1820 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1821 FreeRegionList* local_cleanup_list,
1822 HRRSCleanupTask* hrrs_cleanup_task) :
1823 _g1(g1),
1824 _max_live_bytes(0), _regions_claimed(0),
1825 _freed_bytes(0),
1826 _claimed_region_time(0.0), _max_region_time(0.0),
1827 _local_cleanup_list(local_cleanup_list),
1828 _old_regions_removed(),
1829 _humongous_regions_removed(),
1830 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1831
1832 size_t freed_bytes() { return _freed_bytes; }
1833 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1834 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1835
1836 bool doHeapRegion(HeapRegion *hr) {
1837 if (hr->continuesHumongous()) {
1838 return false;
1839 }
1840 // We use a claim value of zero here because all regions
1841 // were claimed with value 1 in the FinalCount task.
1842 _g1->reset_gc_time_stamps(hr);
1843 double start = os::elapsedTime();
1844 _regions_claimed++;
1845 hr->note_end_of_marking();
1846 _max_live_bytes += hr->max_live_bytes();
1847
1848 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1849 _freed_bytes += hr->used();
1850 hr->set_containing_set(NULL);
1851 if (hr->isHumongous()) {
1852 assert(hr->startsHumongous(), "we should only see starts humongous");
1853 _humongous_regions_removed.increment(1u, hr->capacity());
1854 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1855 } else {
1856 _old_regions_removed.increment(1u, hr->capacity());
1857 _g1->free_region(hr, _local_cleanup_list, true);
1858 }
1859 } else {
1860 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1861 }
1862
1863 double region_time = (os::elapsedTime() - start);
1864 _claimed_region_time += region_time;
1865 if (region_time > _max_region_time) {
1866 _max_region_time = region_time;
1867 }
1868 return false;
1869 }
1870
1871 size_t max_live_bytes() { return _max_live_bytes; }
1872 uint regions_claimed() { return _regions_claimed; }
1873 double claimed_region_time_sec() { return _claimed_region_time; }
1874 double max_region_time_sec() { return _max_region_time; }
1875 };
1876
1877 class G1ParNoteEndTask: public AbstractGangTask {
1878 friend class G1NoteEndOfConcMarkClosure;
1879
1880 protected:
1881 G1CollectedHeap* _g1h;
1882 size_t _max_live_bytes;
1883 size_t _freed_bytes;
1884 FreeRegionList* _cleanup_list;
1885
1886 public:
1887 G1ParNoteEndTask(G1CollectedHeap* g1h,
1888 FreeRegionList* cleanup_list) :
1889 AbstractGangTask("G1 note end"), _g1h(g1h),
1890 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1891
1892 void work(uint worker_id) {
1893 double start = os::elapsedTime();
1894 FreeRegionList local_cleanup_list("Local Cleanup List");
1895 HRRSCleanupTask hrrs_cleanup_task;
1896 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1897 &hrrs_cleanup_task);
1898 if (G1CollectedHeap::use_parallel_gc_threads()) {
1899 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1900 _g1h->workers()->active_workers(),
1901 HeapRegion::NoteEndClaimValue);
1902 } else {
1903 _g1h->heap_region_iterate(&g1_note_end);
1904 }
1905 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1906
1907 // Now update the lists
1908 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1909 {
1910 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1911 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1912 _max_live_bytes += g1_note_end.max_live_bytes();
1913 _freed_bytes += g1_note_end.freed_bytes();
1914
1915 // If we iterate over the global cleanup list at the end of
1916 // cleanup to do this printing we will not guarantee to only
1917 // generate output for the newly-reclaimed regions (the list
1918 // might not be empty at the beginning of cleanup; we might
1919 // still be working on its previous contents). So we do the
1920 // printing here, before we append the new regions to the global
1921 // cleanup list.
1922
1923 G1HRPrinter* hr_printer = _g1h->hr_printer();
1924 if (hr_printer->is_active()) {
1925 FreeRegionListIterator iter(&local_cleanup_list);
1926 while (iter.more_available()) {
1927 HeapRegion* hr = iter.get_next();
1928 hr_printer->cleanup(hr);
1929 }
1930 }
1931
1932 _cleanup_list->add_ordered(&local_cleanup_list);
1933 assert(local_cleanup_list.is_empty(), "post-condition");
1934
1935 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1936 }
1937 }
1938 size_t max_live_bytes() { return _max_live_bytes; }
1939 size_t freed_bytes() { return _freed_bytes; }
1940 };
1941
1942 class G1ParScrubRemSetTask: public AbstractGangTask {
1943 protected:
1944 G1RemSet* _g1rs;
1945 BitMap* _region_bm;
1946 BitMap* _card_bm;
1947 public:
1948 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1949 BitMap* region_bm, BitMap* card_bm) :
1950 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1951 _region_bm(region_bm), _card_bm(card_bm) { }
1952
1953 void work(uint worker_id) {
1954 if (G1CollectedHeap::use_parallel_gc_threads()) {
1955 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1956 HeapRegion::ScrubRemSetClaimValue);
1957 } else {
1958 _g1rs->scrub(_region_bm, _card_bm);
1959 }
1960 }
1961
1962 };
1963
1964 void ConcurrentMark::cleanup() {
1965 // world is stopped at this checkpoint
1966 assert(SafepointSynchronize::is_at_safepoint(),
1967 "world should be stopped");
1968 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1969
1970 // If a full collection has happened, we shouldn't do this.
1971 if (has_aborted()) {
1972 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1973 return;
1974 }
1975
1976 g1h->verify_region_sets_optional();
1977
1978 if (VerifyDuringGC) {
1979 HandleMark hm; // handle scope
1980 Universe::heap()->prepare_for_verify();
1981 Universe::verify(VerifyOption_G1UsePrevMarking,
1982 " VerifyDuringGC:(before)");
1983 }
1984 g1h->check_bitmaps("Cleanup Start");
1985
1986 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
1987 g1p->record_concurrent_mark_cleanup_start();
1988
1989 double start = os::elapsedTime();
1990
1991 HeapRegionRemSet::reset_for_cleanup_tasks();
1992
1993 uint n_workers;
1994
1995 // Do counting once more with the world stopped for good measure.
1996 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1997
1998 if (G1CollectedHeap::use_parallel_gc_threads()) {
1999 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2000 "sanity check");
2001
2002 g1h->set_par_threads();
2003 n_workers = g1h->n_par_threads();
2004 assert(g1h->n_par_threads() == n_workers,
2005 "Should not have been reset");
2006 g1h->workers()->run_task(&g1_par_count_task);
2007 // Done with the parallel phase so reset to 0.
2008 g1h->set_par_threads(0);
2009
2010 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2011 "sanity check");
2012 } else {
2013 n_workers = 1;
2014 g1_par_count_task.work(0);
2015 }
2016
2017 if (VerifyDuringGC) {
2018 // Verify that the counting data accumulated during marking matches
2019 // that calculated by walking the marking bitmap.
2020
2021 // Bitmaps to hold expected values
2022 BitMap expected_region_bm(_region_bm.size(), true);
2023 BitMap expected_card_bm(_card_bm.size(), true);
2024
2025 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2026 &_region_bm,
2027 &_card_bm,
2028 &expected_region_bm,
2029 &expected_card_bm);
2030
2031 if (G1CollectedHeap::use_parallel_gc_threads()) {
2032 g1h->set_par_threads((int)n_workers);
2033 g1h->workers()->run_task(&g1_par_verify_task);
2034 // Done with the parallel phase so reset to 0.
2035 g1h->set_par_threads(0);
2036
2037 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2038 "sanity check");
2039 } else {
2040 g1_par_verify_task.work(0);
2041 }
2042
2043 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2044 }
2045
2046 size_t start_used_bytes = g1h->used();
2047 g1h->set_marking_complete();
2048
2049 double count_end = os::elapsedTime();
2050 double this_final_counting_time = (count_end - start);
2051 _total_counting_time += this_final_counting_time;
2052
2053 if (G1PrintRegionLivenessInfo) {
2054 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2055 _g1h->heap_region_iterate(&cl);
2056 }
2057
2058 // Install newly created mark bitMap as "prev".
2059 swapMarkBitMaps();
2060
2061 g1h->reset_gc_time_stamp();
2062
2063 // Note end of marking in all heap regions.
2064 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2065 if (G1CollectedHeap::use_parallel_gc_threads()) {
2066 g1h->set_par_threads((int)n_workers);
2067 g1h->workers()->run_task(&g1_par_note_end_task);
2068 g1h->set_par_threads(0);
2069
2070 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2071 "sanity check");
2072 } else {
2073 g1_par_note_end_task.work(0);
2074 }
2075 g1h->check_gc_time_stamps();
2076
2077 if (!cleanup_list_is_empty()) {
2078 // The cleanup list is not empty, so we'll have to process it
2079 // concurrently. Notify anyone else that might be wanting free
2080 // regions that there will be more free regions coming soon.
2081 g1h->set_free_regions_coming();
2082 }
2083
2084 // call below, since it affects the metric by which we sort the heap
2085 // regions.
2086 if (G1ScrubRemSets) {
2087 double rs_scrub_start = os::elapsedTime();
2088 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2089 if (G1CollectedHeap::use_parallel_gc_threads()) {
2090 g1h->set_par_threads((int)n_workers);
2091 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2092 g1h->set_par_threads(0);
2093
2094 assert(g1h->check_heap_region_claim_values(
2095 HeapRegion::ScrubRemSetClaimValue),
2096 "sanity check");
2097 } else {
2098 g1_par_scrub_rs_task.work(0);
2099 }
2100
2101 double rs_scrub_end = os::elapsedTime();
2102 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2103 _total_rs_scrub_time += this_rs_scrub_time;
2104 }
2105
2106 // this will also free any regions totally full of garbage objects,
2107 // and sort the regions.
2108 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2109
2110 // Statistics.
2111 double end = os::elapsedTime();
2112 _cleanup_times.add((end - start) * 1000.0);
2113
2114 if (G1Log::fine()) {
2115 g1h->print_size_transition(gclog_or_tty,
2116 start_used_bytes,
2117 g1h->used(),
2118 g1h->capacity());
2119 }
2120
2121 // Clean up will have freed any regions completely full of garbage.
2122 // Update the soft reference policy with the new heap occupancy.
2123 Universe::update_heap_info_at_gc();
2124
2125 // We need to make this be a "collection" so any collection pause that
2126 // races with it goes around and waits for completeCleanup to finish.
2127 g1h->increment_total_collections();
2128
2129 // We reclaimed old regions so we should calculate the sizes to make
2130 // sure we update the old gen/space data.
2131 g1h->g1mm()->update_sizes();
2132
2133 if (VerifyDuringGC) {
2134 HandleMark hm; // handle scope
2135 Universe::heap()->prepare_for_verify();
2136 Universe::verify(VerifyOption_G1UsePrevMarking,
2137 " VerifyDuringGC:(after)");
2138 }
2139 g1h->check_bitmaps("Cleanup End");
2140
2141 g1h->verify_region_sets_optional();
2142 g1h->trace_heap_after_concurrent_cycle();
2143 }
2144
2145 void ConcurrentMark::completeCleanup() {
2146 if (has_aborted()) return;
2147
2148 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2149
2150 _cleanup_list.verify_optional();
2151 FreeRegionList tmp_free_list("Tmp Free List");
2152
2153 if (G1ConcRegionFreeingVerbose) {
2154 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2155 "cleanup list has %u entries",
2156 _cleanup_list.length());
2157 }
2158
2159 // Noone else should be accessing the _cleanup_list at this point,
2160 // so it's not necessary to take any locks
2161 while (!_cleanup_list.is_empty()) {
2162 HeapRegion* hr = _cleanup_list.remove_head();
2163 assert(hr != NULL, "Got NULL from a non-empty list");
2164 hr->par_clear();
2165 tmp_free_list.add_ordered(hr);
2166
2167 // Instead of adding one region at a time to the secondary_free_list,
2168 // we accumulate them in the local list and move them a few at a
2169 // time. This also cuts down on the number of notify_all() calls
2170 // we do during this process. We'll also append the local list when
2171 // _cleanup_list is empty (which means we just removed the last
2172 // region from the _cleanup_list).
2173 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2174 _cleanup_list.is_empty()) {
2175 if (G1ConcRegionFreeingVerbose) {
2176 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2177 "appending %u entries to the secondary_free_list, "
2178 "cleanup list still has %u entries",
2179 tmp_free_list.length(),
2180 _cleanup_list.length());
2181 }
2182
2183 {
2184 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2185 g1h->secondary_free_list_add(&tmp_free_list);
2186 SecondaryFreeList_lock->notify_all();
2187 }
2188
2189 if (G1StressConcRegionFreeing) {
2190 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2191 os::sleep(Thread::current(), (jlong) 1, false);
2192 }
2193 }
2194 }
2195 }
2196 assert(tmp_free_list.is_empty(), "post-condition");
2197 }
2198
2199 // Supporting Object and Oop closures for reference discovery
2200 // and processing in during marking
2201
2202 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2203 HeapWord* addr = (HeapWord*)obj;
2204 return addr != NULL &&
2205 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2206 }
2207
2208 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2209 // Uses the CMTask associated with a worker thread (for serial reference
2210 // processing the CMTask for worker 0 is used) to preserve (mark) and
2211 // trace referent objects.
2212 //
2213 // Using the CMTask and embedded local queues avoids having the worker
2214 // threads operating on the global mark stack. This reduces the risk
2215 // of overflowing the stack - which we would rather avoid at this late
2216 // state. Also using the tasks' local queues removes the potential
2217 // of the workers interfering with each other that could occur if
2218 // operating on the global stack.
2219
2220 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2221 ConcurrentMark* _cm;
2222 CMTask* _task;
2223 int _ref_counter_limit;
2224 int _ref_counter;
2225 bool _is_serial;
2226 public:
2227 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2228 _cm(cm), _task(task), _is_serial(is_serial),
2229 _ref_counter_limit(G1RefProcDrainInterval) {
2230 assert(_ref_counter_limit > 0, "sanity");
2231 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2232 _ref_counter = _ref_counter_limit;
2233 }
2234
2235 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2236 virtual void do_oop( oop* p) { do_oop_work(p); }
2237
2238 template <class T> void do_oop_work(T* p) {
2239 if (!_cm->has_overflown()) {
2240 oop obj = oopDesc::load_decode_heap_oop(p);
2241 if (_cm->verbose_high()) {
2242 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2243 "*"PTR_FORMAT" = "PTR_FORMAT,
2244 _task->worker_id(), p2i(p), p2i((void*) obj));
2245 }
2246
2247 _task->deal_with_reference(obj);
2248 _ref_counter--;
2249
2250 if (_ref_counter == 0) {
2251 // We have dealt with _ref_counter_limit references, pushing them
2252 // and objects reachable from them on to the local stack (and
2253 // possibly the global stack). Call CMTask::do_marking_step() to
2254 // process these entries.
2255 //
2256 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2257 // there's nothing more to do (i.e. we're done with the entries that
2258 // were pushed as a result of the CMTask::deal_with_reference() calls
2259 // above) or we overflow.
2260 //
2261 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2262 // flag while there may still be some work to do. (See the comment at
2263 // the beginning of CMTask::do_marking_step() for those conditions -
2264 // one of which is reaching the specified time target.) It is only
2265 // when CMTask::do_marking_step() returns without setting the
2266 // has_aborted() flag that the marking step has completed.
2267 do {
2268 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2269 _task->do_marking_step(mark_step_duration_ms,
2270 false /* do_termination */,
2271 _is_serial);
2272 } while (_task->has_aborted() && !_cm->has_overflown());
2273 _ref_counter = _ref_counter_limit;
2274 }
2275 } else {
2276 if (_cm->verbose_high()) {
2277 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2278 }
2279 }
2280 }
2281 };
2282
2283 // 'Drain' oop closure used by both serial and parallel reference processing.
2284 // Uses the CMTask associated with a given worker thread (for serial
2285 // reference processing the CMtask for worker 0 is used). Calls the
2286 // do_marking_step routine, with an unbelievably large timeout value,
2287 // to drain the marking data structures of the remaining entries
2288 // added by the 'keep alive' oop closure above.
2289
2290 class G1CMDrainMarkingStackClosure: public VoidClosure {
2291 ConcurrentMark* _cm;
2292 CMTask* _task;
2293 bool _is_serial;
2294 public:
2295 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2296 _cm(cm), _task(task), _is_serial(is_serial) {
2297 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2298 }
2299
2300 void do_void() {
2301 do {
2302 if (_cm->verbose_high()) {
2303 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2304 _task->worker_id(), BOOL_TO_STR(_is_serial));
2305 }
2306
2307 // We call CMTask::do_marking_step() to completely drain the local
2308 // and global marking stacks of entries pushed by the 'keep alive'
2309 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2310 //
2311 // CMTask::do_marking_step() is called in a loop, which we'll exit
2312 // if there's nothing more to do (i.e. we've completely drained the
2313 // entries that were pushed as a a result of applying the 'keep alive'
2314 // closure to the entries on the discovered ref lists) or we overflow
2315 // the global marking stack.
2316 //
2317 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2318 // flag while there may still be some work to do. (See the comment at
2319 // the beginning of CMTask::do_marking_step() for those conditions -
2320 // one of which is reaching the specified time target.) It is only
2321 // when CMTask::do_marking_step() returns without setting the
2322 // has_aborted() flag that the marking step has completed.
2323
2324 _task->do_marking_step(1000000000.0 /* something very large */,
2325 true /* do_termination */,
2326 _is_serial);
2327 } while (_task->has_aborted() && !_cm->has_overflown());
2328 }
2329 };
2330
2331 // Implementation of AbstractRefProcTaskExecutor for parallel
2332 // reference processing at the end of G1 concurrent marking
2333
2334 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2335 private:
2336 G1CollectedHeap* _g1h;
2337 ConcurrentMark* _cm;
2338 WorkGang* _workers;
2339 int _active_workers;
2340
2341 public:
2342 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2343 ConcurrentMark* cm,
2344 WorkGang* workers,
2345 int n_workers) :
2346 _g1h(g1h), _cm(cm),
2347 _workers(workers), _active_workers(n_workers) { }
2348
2349 // Executes the given task using concurrent marking worker threads.
2350 virtual void execute(ProcessTask& task);
2351 virtual void execute(EnqueueTask& task);
2352 };
2353
2354 class G1CMRefProcTaskProxy: public AbstractGangTask {
2355 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2356 ProcessTask& _proc_task;
2357 G1CollectedHeap* _g1h;
2358 ConcurrentMark* _cm;
2359
2360 public:
2361 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2362 G1CollectedHeap* g1h,
2363 ConcurrentMark* cm) :
2364 AbstractGangTask("Process reference objects in parallel"),
2365 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2366 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2367 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2368 }
2369
2370 virtual void work(uint worker_id) {
2371 CMTask* task = _cm->task(worker_id);
2372 G1CMIsAliveClosure g1_is_alive(_g1h);
2373 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2374 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2375
2376 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2377 }
2378 };
2379
2380 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2381 assert(_workers != NULL, "Need parallel worker threads.");
2382 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2383
2384 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2385
2386 // We need to reset the concurrency level before each
2387 // proxy task execution, so that the termination protocol
2388 // and overflow handling in CMTask::do_marking_step() knows
2389 // how many workers to wait for.
2390 _cm->set_concurrency(_active_workers);
2391 _g1h->set_par_threads(_active_workers);
2392 _workers->run_task(&proc_task_proxy);
2393 _g1h->set_par_threads(0);
2394 }
2395
2396 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2397 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2398 EnqueueTask& _enq_task;
2399
2400 public:
2401 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2402 AbstractGangTask("Enqueue reference objects in parallel"),
2403 _enq_task(enq_task) { }
2404
2405 virtual void work(uint worker_id) {
2406 _enq_task.work(worker_id);
2407 }
2408 };
2409
2410 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2411 assert(_workers != NULL, "Need parallel worker threads.");
2412 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2413
2414 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2415
2416 // Not strictly necessary but...
2417 //
2418 // We need to reset the concurrency level before each
2419 // proxy task execution, so that the termination protocol
2420 // and overflow handling in CMTask::do_marking_step() knows
2421 // how many workers to wait for.
2422 _cm->set_concurrency(_active_workers);
2423 _g1h->set_par_threads(_active_workers);
2424 _workers->run_task(&enq_task_proxy);
2425 _g1h->set_par_threads(0);
2426 }
2427
2428 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2429 if (has_overflown()) {
2430 // Skip processing the discovered references if we have
2431 // overflown the global marking stack. Reference objects
2432 // only get discovered once so it is OK to not
2433 // de-populate the discovered reference lists. We could have,
2434 // but the only benefit would be that, when marking restarts,
2435 // less reference objects are discovered.
2436 return;
2437 }
2438
2439 ResourceMark rm;
2440 HandleMark hm;
2441
2442 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2443
2444 // Is alive closure.
2445 G1CMIsAliveClosure g1_is_alive(g1h);
2446
2447 // Inner scope to exclude the cleaning of the string and symbol
2448 // tables from the displayed time.
2449 {
2450 if (G1Log::finer()) {
2451 gclog_or_tty->put(' ');
2452 }
2453 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm());
2454
2455 ReferenceProcessor* rp = g1h->ref_processor_cm();
2456
2457 // See the comment in G1CollectedHeap::ref_processing_init()
2458 // about how reference processing currently works in G1.
2459
2460 // Set the soft reference policy
2461 rp->setup_policy(clear_all_soft_refs);
2462 assert(_markStack.isEmpty(), "mark stack should be empty");
2463
2464 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2465 // in serial reference processing. Note these closures are also
2466 // used for serially processing (by the the current thread) the
2467 // JNI references during parallel reference processing.
2468 //
2469 // These closures do not need to synchronize with the worker
2470 // threads involved in parallel reference processing as these
2471 // instances are executed serially by the current thread (e.g.
2472 // reference processing is not multi-threaded and is thus
2473 // performed by the current thread instead of a gang worker).
2474 //
2475 // The gang tasks involved in parallel reference processing create
2476 // their own instances of these closures, which do their own
2477 // synchronization among themselves.
2478 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2479 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2480
2481 // We need at least one active thread. If reference processing
2482 // is not multi-threaded we use the current (VMThread) thread,
2483 // otherwise we use the work gang from the G1CollectedHeap and
2484 // we utilize all the worker threads we can.
2485 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2486 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2487 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2488
2489 // Parallel processing task executor.
2490 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2491 g1h->workers(), active_workers);
2492 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2493
2494 // Set the concurrency level. The phase was already set prior to
2495 // executing the remark task.
2496 set_concurrency(active_workers);
2497
2498 // Set the degree of MT processing here. If the discovery was done MT,
2499 // the number of threads involved during discovery could differ from
2500 // the number of active workers. This is OK as long as the discovered
2501 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2502 rp->set_active_mt_degree(active_workers);
2503
2504 // Process the weak references.
2505 const ReferenceProcessorStats& stats =
2506 rp->process_discovered_references(&g1_is_alive,
2507 &g1_keep_alive,
2508 &g1_drain_mark_stack,
2509 executor,
2510 g1h->gc_timer_cm());
2511 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2512
2513 // The do_oop work routines of the keep_alive and drain_marking_stack
2514 // oop closures will set the has_overflown flag if we overflow the
2515 // global marking stack.
2516
2517 assert(_markStack.overflow() || _markStack.isEmpty(),
2518 "mark stack should be empty (unless it overflowed)");
2519
2520 if (_markStack.overflow()) {
2521 // This should have been done already when we tried to push an
2522 // entry on to the global mark stack. But let's do it again.
2523 set_has_overflown();
2524 }
2525
2526 assert(rp->num_q() == active_workers, "why not");
2527
2528 rp->enqueue_discovered_references(executor);
2529
2530 rp->verify_no_references_recorded();
2531 assert(!rp->discovery_enabled(), "Post condition");
2532 }
2533
2534 if (has_overflown()) {
2535 // We can not trust g1_is_alive if the marking stack overflowed
2536 return;
2537 }
2538
2539 g1h->unlink_string_and_symbol_table(&g1_is_alive,
2540 /* process_strings */ false, // currently strings are always roots
2541 /* process_symbols */ true);
2542 }
2543
2544 void ConcurrentMark::swapMarkBitMaps() {
2545 CMBitMapRO* temp = _prevMarkBitMap;
2546 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2547 _nextMarkBitMap = (CMBitMap*) temp;
2548 }
2549
2550 class CMRemarkTask: public AbstractGangTask {
2551 private:
2552 ConcurrentMark* _cm;
2553 bool _is_serial;
2554 public:
2555 void work(uint worker_id) {
2556 // Since all available tasks are actually started, we should
2557 // only proceed if we're supposed to be active.
2558 if (worker_id < _cm->active_tasks()) {
2559 CMTask* task = _cm->task(worker_id);
2560 task->record_start_time();
2561 do {
2562 task->do_marking_step(1000000000.0 /* something very large */,
2563 true /* do_termination */,
2564 _is_serial);
2565 } while (task->has_aborted() && !_cm->has_overflown());
2566 // If we overflow, then we do not want to restart. We instead
2567 // want to abort remark and do concurrent marking again.
2568 task->record_end_time();
2569 }
2570 }
2571
2572 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2573 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2574 _cm->terminator()->reset_for_reuse(active_workers);
2575 }
2576 };
2577
2578 void ConcurrentMark::checkpointRootsFinalWork() {
2579 ResourceMark rm;
2580 HandleMark hm;
2581 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2582
2583 g1h->ensure_parsability(false);
2584
2585 if (G1CollectedHeap::use_parallel_gc_threads()) {
2586 G1CollectedHeap::StrongRootsScope srs(g1h);
2587 // this is remark, so we'll use up all active threads
2588 uint active_workers = g1h->workers()->active_workers();
2589 if (active_workers == 0) {
2590 assert(active_workers > 0, "Should have been set earlier");
2591 active_workers = (uint) ParallelGCThreads;
2592 g1h->workers()->set_active_workers(active_workers);
2593 }
2594 set_concurrency_and_phase(active_workers, false /* concurrent */);
2595 // Leave _parallel_marking_threads at it's
2596 // value originally calculated in the ConcurrentMark
2597 // constructor and pass values of the active workers
2598 // through the gang in the task.
2599
2600 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2601 // We will start all available threads, even if we decide that the
2602 // active_workers will be fewer. The extra ones will just bail out
2603 // immediately.
2604 g1h->set_par_threads(active_workers);
2605 g1h->workers()->run_task(&remarkTask);
2606 g1h->set_par_threads(0);
2607 } else {
2608 G1CollectedHeap::StrongRootsScope srs(g1h);
2609 uint active_workers = 1;
2610 set_concurrency_and_phase(active_workers, false /* concurrent */);
2611
2612 // Note - if there's no work gang then the VMThread will be
2613 // the thread to execute the remark - serially. We have
2614 // to pass true for the is_serial parameter so that
2615 // CMTask::do_marking_step() doesn't enter the sync
2616 // barriers in the event of an overflow. Doing so will
2617 // cause an assert that the current thread is not a
2618 // concurrent GC thread.
2619 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2620 remarkTask.work(0);
2621 }
2622 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2623 guarantee(has_overflown() ||
2624 satb_mq_set.completed_buffers_num() == 0,
2625 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2626 BOOL_TO_STR(has_overflown()),
2627 satb_mq_set.completed_buffers_num()));
2628
2629 print_stats();
2630 }
2631
2632 #ifndef PRODUCT
2633
2634 class PrintReachableOopClosure: public OopClosure {
2635 private:
2636 G1CollectedHeap* _g1h;
2637 outputStream* _out;
2638 VerifyOption _vo;
2639 bool _all;
2640
2641 public:
2642 PrintReachableOopClosure(outputStream* out,
2643 VerifyOption vo,
2644 bool all) :
2645 _g1h(G1CollectedHeap::heap()),
2646 _out(out), _vo(vo), _all(all) { }
2647
2648 void do_oop(narrowOop* p) { do_oop_work(p); }
2649 void do_oop( oop* p) { do_oop_work(p); }
2650
2651 template <class T> void do_oop_work(T* p) {
2652 oop obj = oopDesc::load_decode_heap_oop(p);
2653 const char* str = NULL;
2654 const char* str2 = "";
2655
2656 if (obj == NULL) {
2657 str = "";
2658 } else if (!_g1h->is_in_g1_reserved(obj)) {
2659 str = " O";
2660 } else {
2661 HeapRegion* hr = _g1h->heap_region_containing(obj);
2662 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2663 bool marked = _g1h->is_marked(obj, _vo);
2664
2665 if (over_tams) {
2666 str = " >";
2667 if (marked) {
2668 str2 = " AND MARKED";
2669 }
2670 } else if (marked) {
2671 str = " M";
2672 } else {
2673 str = " NOT";
2674 }
2675 }
2676
2677 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2678 p2i(p), p2i((void*) obj), str, str2);
2679 }
2680 };
2681
2682 class PrintReachableObjectClosure : public ObjectClosure {
2683 private:
2684 G1CollectedHeap* _g1h;
2685 outputStream* _out;
2686 VerifyOption _vo;
2687 bool _all;
2688 HeapRegion* _hr;
2689
2690 public:
2691 PrintReachableObjectClosure(outputStream* out,
2692 VerifyOption vo,
2693 bool all,
2694 HeapRegion* hr) :
2695 _g1h(G1CollectedHeap::heap()),
2696 _out(out), _vo(vo), _all(all), _hr(hr) { }
2697
2698 void do_object(oop o) {
2699 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2700 bool marked = _g1h->is_marked(o, _vo);
2701 bool print_it = _all || over_tams || marked;
2702
2703 if (print_it) {
2704 _out->print_cr(" "PTR_FORMAT"%s",
2705 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2706 PrintReachableOopClosure oopCl(_out, _vo, _all);
2707 o->oop_iterate_no_header(&oopCl);
2708 }
2709 }
2710 };
2711
2712 class PrintReachableRegionClosure : public HeapRegionClosure {
2713 private:
2714 G1CollectedHeap* _g1h;
2715 outputStream* _out;
2716 VerifyOption _vo;
2717 bool _all;
2718
2719 public:
2720 bool doHeapRegion(HeapRegion* hr) {
2721 HeapWord* b = hr->bottom();
2722 HeapWord* e = hr->end();
2723 HeapWord* t = hr->top();
2724 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2725 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2726 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2727 _out->cr();
2728
2729 HeapWord* from = b;
2730 HeapWord* to = t;
2731
2732 if (to > from) {
2733 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2734 _out->cr();
2735 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2736 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2737 _out->cr();
2738 }
2739
2740 return false;
2741 }
2742
2743 PrintReachableRegionClosure(outputStream* out,
2744 VerifyOption vo,
2745 bool all) :
2746 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2747 };
2748
2749 void ConcurrentMark::print_reachable(const char* str,
2750 VerifyOption vo,
2751 bool all) {
2752 gclog_or_tty->cr();
2753 gclog_or_tty->print_cr("== Doing heap dump... ");
2754
2755 if (G1PrintReachableBaseFile == NULL) {
2756 gclog_or_tty->print_cr(" #### error: no base file defined");
2757 return;
2758 }
2759
2760 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2761 (JVM_MAXPATHLEN - 1)) {
2762 gclog_or_tty->print_cr(" #### error: file name too long");
2763 return;
2764 }
2765
2766 char file_name[JVM_MAXPATHLEN];
2767 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2768 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2769
2770 fileStream fout(file_name);
2771 if (!fout.is_open()) {
2772 gclog_or_tty->print_cr(" #### error: could not open file");
2773 return;
2774 }
2775
2776 outputStream* out = &fout;
2777 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2778 out->cr();
2779
2780 out->print_cr("--- ITERATING OVER REGIONS");
2781 out->cr();
2782 PrintReachableRegionClosure rcl(out, vo, all);
2783 _g1h->heap_region_iterate(&rcl);
2784 out->cr();
2785
2786 gclog_or_tty->print_cr(" done");
2787 gclog_or_tty->flush();
2788 }
2789
2790 #endif // PRODUCT
2791
2792 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2793 // Note we are overriding the read-only view of the prev map here, via
2794 // the cast.
2795 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2796 }
2797
2798 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2799 _nextMarkBitMap->clearRange(mr);
2800 }
2801
2802 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) {
2803 clearRangePrevBitmap(mr);
2804 clearRangeNextBitmap(mr);
2805 }
2806
2807 HeapRegion*
2808 ConcurrentMark::claim_region(uint worker_id) {
2809 // "checkpoint" the finger
2810 HeapWord* finger = _finger;
2811
2812 // _heap_end will not change underneath our feet; it only changes at
2813 // yield points.
2814 while (finger < _heap_end) {
2815 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2816
2817 // Note on how this code handles humongous regions. In the
2818 // normal case the finger will reach the start of a "starts
2819 // humongous" (SH) region. Its end will either be the end of the
2820 // last "continues humongous" (CH) region in the sequence, or the
2821 // standard end of the SH region (if the SH is the only region in
2822 // the sequence). That way claim_region() will skip over the CH
2823 // regions. However, there is a subtle race between a CM thread
2824 // executing this method and a mutator thread doing a humongous
2825 // object allocation. The two are not mutually exclusive as the CM
2826 // thread does not need to hold the Heap_lock when it gets
2827 // here. So there is a chance that claim_region() will come across
2828 // a free region that's in the progress of becoming a SH or a CH
2829 // region. In the former case, it will either
2830 // a) Miss the update to the region's end, in which case it will
2831 // visit every subsequent CH region, will find their bitmaps
2832 // empty, and do nothing, or
2833 // b) Will observe the update of the region's end (in which case
2834 // it will skip the subsequent CH regions).
2835 // If it comes across a region that suddenly becomes CH, the
2836 // scenario will be similar to b). So, the race between
2837 // claim_region() and a humongous object allocation might force us
2838 // to do a bit of unnecessary work (due to some unnecessary bitmap
2839 // iterations) but it should not introduce and correctness issues.
2840 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2841 HeapWord* bottom = curr_region->bottom();
2842 HeapWord* end = curr_region->end();
2843 HeapWord* limit = curr_region->next_top_at_mark_start();
2844
2845 if (verbose_low()) {
2846 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2847 "["PTR_FORMAT", "PTR_FORMAT"), "
2848 "limit = "PTR_FORMAT,
2849 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2850 }
2851
2852 // Is the gap between reading the finger and doing the CAS too long?
2853 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2854 if (res == finger) {
2855 // we succeeded
2856
2857 // notice that _finger == end cannot be guaranteed here since,
2858 // someone else might have moved the finger even further
2859 assert(_finger >= end, "the finger should have moved forward");
2860
2861 if (verbose_low()) {
2862 gclog_or_tty->print_cr("[%u] we were successful with region = "
2863 PTR_FORMAT, worker_id, p2i(curr_region));
2864 }
2865
2866 if (limit > bottom) {
2867 if (verbose_low()) {
2868 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2869 "returning it ", worker_id, p2i(curr_region));
2870 }
2871 return curr_region;
2872 } else {
2873 assert(limit == bottom,
2874 "the region limit should be at bottom");
2875 if (verbose_low()) {
2876 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
2877 "returning NULL", worker_id, p2i(curr_region));
2878 }
2879 // we return NULL and the caller should try calling
2880 // claim_region() again.
2881 return NULL;
2882 }
2883 } else {
2884 assert(_finger > finger, "the finger should have moved forward");
2885 if (verbose_low()) {
2886 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2887 "global finger = "PTR_FORMAT", "
2888 "our finger = "PTR_FORMAT,
2889 worker_id, p2i(_finger), p2i(finger));
2890 }
2891
2892 // read it again
2893 finger = _finger;
2894 }
2895 }
2896
2897 return NULL;
2898 }
2899
2900 #ifndef PRODUCT
2901 enum VerifyNoCSetOopsPhase {
2902 VerifyNoCSetOopsStack,
2903 VerifyNoCSetOopsQueues,
2904 VerifyNoCSetOopsSATBCompleted,
2905 VerifyNoCSetOopsSATBThread
2906 };
2907
2908 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
2909 private:
2910 G1CollectedHeap* _g1h;
2911 VerifyNoCSetOopsPhase _phase;
2912 int _info;
2913
2914 const char* phase_str() {
2915 switch (_phase) {
2916 case VerifyNoCSetOopsStack: return "Stack";
2917 case VerifyNoCSetOopsQueues: return "Queue";
2918 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2919 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
2920 default: ShouldNotReachHere();
2921 }
2922 return NULL;
2923 }
2924
2925 void do_object_work(oop obj) {
2926 guarantee(!_g1h->obj_in_cs(obj),
2927 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2928 p2i((void*) obj), phase_str(), _info));
2929 }
2930
2931 public:
2932 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2933
2934 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2935 _phase = phase;
2936 _info = info;
2937 }
2938
2939 virtual void do_oop(oop* p) {
2940 oop obj = oopDesc::load_decode_heap_oop(p);
2941 do_object_work(obj);
2942 }
2943
2944 virtual void do_oop(narrowOop* p) {
2945 // We should not come across narrow oops while scanning marking
2946 // stacks and SATB buffers.
2947 ShouldNotReachHere();
2948 }
2949
2950 virtual void do_object(oop obj) {
2951 do_object_work(obj);
2952 }
2953 };
2954
2955 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
2956 bool verify_enqueued_buffers,
2957 bool verify_thread_buffers,
2958 bool verify_fingers) {
2959 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2960 if (!G1CollectedHeap::heap()->mark_in_progress()) {
2961 return;
2962 }
2963
2964 VerifyNoCSetOopsClosure cl;
2965
2966 if (verify_stacks) {
2967 // Verify entries on the global mark stack
2968 cl.set_phase(VerifyNoCSetOopsStack);
2969 _markStack.oops_do(&cl);
2970
2971 // Verify entries on the task queues
2972 for (uint i = 0; i < _max_worker_id; i += 1) {
2973 cl.set_phase(VerifyNoCSetOopsQueues, i);
2974 CMTaskQueue* queue = _task_queues->queue(i);
2975 queue->oops_do(&cl);
2976 }
2977 }
2978
2979 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
2980
2981 // Verify entries on the enqueued SATB buffers
2982 if (verify_enqueued_buffers) {
2983 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
2984 satb_qs.iterate_completed_buffers_read_only(&cl);
2985 }
2986
2987 // Verify entries on the per-thread SATB buffers
2988 if (verify_thread_buffers) {
2989 cl.set_phase(VerifyNoCSetOopsSATBThread);
2990 satb_qs.iterate_thread_buffers_read_only(&cl);
2991 }
2992
2993 if (verify_fingers) {
2994 // Verify the global finger
2995 HeapWord* global_finger = finger();
2996 if (global_finger != NULL && global_finger < _heap_end) {
2997 // The global finger always points to a heap region boundary. We
2998 // use heap_region_containing_raw() to get the containing region
2999 // given that the global finger could be pointing to a free region
3000 // which subsequently becomes continues humongous. If that
3001 // happens, heap_region_containing() will return the bottom of the
3002 // corresponding starts humongous region and the check below will
3003 // not hold any more.
3004 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3005 guarantee(global_finger == global_hr->bottom(),
3006 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3007 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3008 }
3009
3010 // Verify the task fingers
3011 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3012 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3013 CMTask* task = _tasks[i];
3014 HeapWord* task_finger = task->finger();
3015 if (task_finger != NULL && task_finger < _heap_end) {
3016 // See above note on the global finger verification.
3017 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3018 guarantee(task_finger == task_hr->bottom() ||
3019 !task_hr->in_collection_set(),
3020 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3021 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3022 }
3023 }
3024 }
3025 }
3026 #endif // PRODUCT
3027
3028 // Aggregate the counting data that was constructed concurrently
3029 // with marking.
3030 class AggregateCountDataHRClosure: public HeapRegionClosure {
3031 G1CollectedHeap* _g1h;
3032 ConcurrentMark* _cm;
3033 CardTableModRefBS* _ct_bs;
3034 BitMap* _cm_card_bm;
3035 uint _max_worker_id;
3036
3037 public:
3038 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3039 BitMap* cm_card_bm,
3040 uint max_worker_id) :
3041 _g1h(g1h), _cm(g1h->concurrent_mark()),
3042 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3043 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3044
3045 bool doHeapRegion(HeapRegion* hr) {
3046 if (hr->continuesHumongous()) {
3047 // We will ignore these here and process them when their
3048 // associated "starts humongous" region is processed.
3049 // Note that we cannot rely on their associated
3050 // "starts humongous" region to have their bit set to 1
3051 // since, due to the region chunking in the parallel region
3052 // iteration, a "continues humongous" region might be visited
3053 // before its associated "starts humongous".
3054 return false;
3055 }
3056
3057 HeapWord* start = hr->bottom();
3058 HeapWord* limit = hr->next_top_at_mark_start();
3059 HeapWord* end = hr->end();
3060
3061 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3062 err_msg("Preconditions not met - "
3063 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3064 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3065 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3066
3067 assert(hr->next_marked_bytes() == 0, "Precondition");
3068
3069 if (start == limit) {
3070 // NTAMS of this region has not been set so nothing to do.
3071 return false;
3072 }
3073
3074 // 'start' should be in the heap.
3075 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3076 // 'end' *may* be just beyond the end of the heap (if hr is the last region)
3077 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3078
3079 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3080 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3081 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3082
3083 // If ntams is not card aligned then we bump card bitmap index
3084 // for limit so that we get the all the cards spanned by
3085 // the object ending at ntams.
3086 // Note: if this is the last region in the heap then ntams
3087 // could be actually just beyond the end of the the heap;
3088 // limit_idx will then correspond to a (non-existent) card
3089 // that is also outside the heap.
3090 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3091 limit_idx += 1;
3092 }
3093
3094 assert(limit_idx <= end_idx, "or else use atomics");
3095
3096 // Aggregate the "stripe" in the count data associated with hr.
3097 uint hrs_index = hr->hrs_index();
3098 size_t marked_bytes = 0;
3099
3100 for (uint i = 0; i < _max_worker_id; i += 1) {
3101 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3102 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3103
3104 // Fetch the marked_bytes in this region for task i and
3105 // add it to the running total for this region.
3106 marked_bytes += marked_bytes_array[hrs_index];
3107
3108 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3109 // into the global card bitmap.
3110 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3111
3112 while (scan_idx < limit_idx) {
3113 assert(task_card_bm->at(scan_idx) == true, "should be");
3114 _cm_card_bm->set_bit(scan_idx);
3115 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3116
3117 // BitMap::get_next_one_offset() can handle the case when
3118 // its left_offset parameter is greater than its right_offset
3119 // parameter. It does, however, have an early exit if
3120 // left_offset == right_offset. So let's limit the value
3121 // passed in for left offset here.
3122 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3123 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3124 }
3125 }
3126
3127 // Update the marked bytes for this region.
3128 hr->add_to_marked_bytes(marked_bytes);
3129
3130 // Next heap region
3131 return false;
3132 }
3133 };
3134
3135 class G1AggregateCountDataTask: public AbstractGangTask {
3136 protected:
3137 G1CollectedHeap* _g1h;
3138 ConcurrentMark* _cm;
3139 BitMap* _cm_card_bm;
3140 uint _max_worker_id;
3141 int _active_workers;
3142
3143 public:
3144 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3145 ConcurrentMark* cm,
3146 BitMap* cm_card_bm,
3147 uint max_worker_id,
3148 int n_workers) :
3149 AbstractGangTask("Count Aggregation"),
3150 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3151 _max_worker_id(max_worker_id),
3152 _active_workers(n_workers) { }
3153
3154 void work(uint worker_id) {
3155 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3156
3157 if (G1CollectedHeap::use_parallel_gc_threads()) {
3158 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3159 _active_workers,
3160 HeapRegion::AggregateCountClaimValue);
3161 } else {
3162 _g1h->heap_region_iterate(&cl);
3163 }
3164 }
3165 };
3166
3167
3168 void ConcurrentMark::aggregate_count_data() {
3169 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3170 _g1h->workers()->active_workers() :
3171 1);
3172
3173 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3174 _max_worker_id, n_workers);
3175
3176 if (G1CollectedHeap::use_parallel_gc_threads()) {
3177 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3178 "sanity check");
3179 _g1h->set_par_threads(n_workers);
3180 _g1h->workers()->run_task(&g1_par_agg_task);
3181 _g1h->set_par_threads(0);
3182
3183 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3184 "sanity check");
3185 _g1h->reset_heap_region_claim_values();
3186 } else {
3187 g1_par_agg_task.work(0);
3188 }
3189 }
3190
3191 // Clear the per-worker arrays used to store the per-region counting data
3192 void ConcurrentMark::clear_all_count_data() {
3193 // Clear the global card bitmap - it will be filled during
3194 // liveness count aggregation (during remark) and the
3195 // final counting task.
3196 _card_bm.clear();
3197
3198 // Clear the global region bitmap - it will be filled as part
3199 // of the final counting task.
3200 _region_bm.clear();
3201
3202 uint max_regions = _g1h->max_regions();
3203 assert(_max_worker_id > 0, "uninitialized");
3204
3205 for (uint i = 0; i < _max_worker_id; i += 1) {
3206 BitMap* task_card_bm = count_card_bitmap_for(i);
3207 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3208
3209 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3210 assert(marked_bytes_array != NULL, "uninitialized");
3211
3212 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3213 task_card_bm->clear();
3214 }
3215 }
3216
3217 void ConcurrentMark::print_stats() {
3218 if (verbose_stats()) {
3219 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3220 for (size_t i = 0; i < _active_tasks; ++i) {
3221 _tasks[i]->print_stats();
3222 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3223 }
3224 }
3225 }
3226
3227 // abandon current marking iteration due to a Full GC
3228 void ConcurrentMark::abort() {
3229 // Clear all marks to force marking thread to do nothing
3230 _nextMarkBitMap->clearAll();
3231
3232 // Note we cannot clear the previous marking bitmap here
3233 // since VerifyDuringGC verifies the objects marked during
3234 // a full GC against the previous bitmap.
3235
3236 // Clear the liveness counting data
3237 clear_all_count_data();
3238 // Empty mark stack
3239 reset_marking_state();
3240 for (uint i = 0; i < _max_worker_id; ++i) {
3241 _tasks[i]->clear_region_fields();
3242 }
3243 _has_aborted = true;
3244
3245 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3246 satb_mq_set.abandon_partial_marking();
3247 // This can be called either during or outside marking, we'll read
3248 // the expected_active value from the SATB queue set.
3249 satb_mq_set.set_active_all_threads(
3250 false, /* new active value */
3251 satb_mq_set.is_active() /* expected_active */);
3252
3253 _g1h->trace_heap_after_concurrent_cycle();
3254 _g1h->register_concurrent_cycle_end();
3255 }
3256
3257 static void print_ms_time_info(const char* prefix, const char* name,
3258 NumberSeq& ns) {
3259 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3260 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3261 if (ns.num() > 0) {
3262 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3263 prefix, ns.sd(), ns.maximum());
3264 }
3265 }
3266
3267 void ConcurrentMark::print_summary_info() {
3268 gclog_or_tty->print_cr(" Concurrent marking:");
3269 print_ms_time_info(" ", "init marks", _init_times);
3270 print_ms_time_info(" ", "remarks", _remark_times);
3271 {
3272 print_ms_time_info(" ", "final marks", _remark_mark_times);
3273 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3274
3275 }
3276 print_ms_time_info(" ", "cleanups", _cleanup_times);
3277 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3278 _total_counting_time,
3279 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3280 (double)_cleanup_times.num()
3281 : 0.0));
3282 if (G1ScrubRemSets) {
3283 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3284 _total_rs_scrub_time,
3285 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3286 (double)_cleanup_times.num()
3287 : 0.0));
3288 }
3289 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3290 (_init_times.sum() + _remark_times.sum() +
3291 _cleanup_times.sum())/1000.0);
3292 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3293 "(%8.2f s marking).",
3294 cmThread()->vtime_accum(),
3295 cmThread()->vtime_mark_accum());
3296 }
3297
3298 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3299 if (use_parallel_marking_threads()) {
3300 _parallel_workers->print_worker_threads_on(st);
3301 }
3302 }
3303
3304 void ConcurrentMark::print_on_error(outputStream* st) const {
3305 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3306 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3307 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3308 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3309 }
3310
3311 // We take a break if someone is trying to stop the world.
3312 bool ConcurrentMark::do_yield_check(uint worker_id) {
3313 if (SuspendibleThreadSet::should_yield()) {
3314 if (worker_id == 0) {
3315 _g1h->g1_policy()->record_concurrent_pause();
3316 }
3317 SuspendibleThreadSet::yield();
3318 return true;
3319 } else {
3320 return false;
3321 }
3322 }
3323
3324 bool ConcurrentMark::containing_card_is_marked(void* p) {
3325 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1);
3326 return _card_bm.at(offset >> CardTableModRefBS::card_shift);
3327 }
3328
3329 bool ConcurrentMark::containing_cards_are_marked(void* start,
3330 void* last) {
3331 return containing_card_is_marked(start) &&
3332 containing_card_is_marked(last);
3333 }
3334
3335 #ifndef PRODUCT
3336 // for debugging purposes
3337 void ConcurrentMark::print_finger() {
3338 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3339 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3340 for (uint i = 0; i < _max_worker_id; ++i) {
3341 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3342 }
3343 gclog_or_tty->cr();
3344 }
3345 #endif
3346
3347 void CMTask::scan_object(oop obj) {
3348 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3349
3350 if (_cm->verbose_high()) {
3351 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3352 _worker_id, p2i((void*) obj));
3353 }
3354
3355 size_t obj_size = obj->size();
3356 _words_scanned += obj_size;
3357
3358 obj->oop_iterate(_cm_oop_closure);
3359 statsOnly( ++_objs_scanned );
3360 check_limits();
3361 }
3362
3363 // Closure for iteration over bitmaps
3364 class CMBitMapClosure : public BitMapClosure {
3365 private:
3366 // the bitmap that is being iterated over
3367 CMBitMap* _nextMarkBitMap;
3368 ConcurrentMark* _cm;
3369 CMTask* _task;
3370
3371 public:
3372 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3373 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3374
3375 bool do_bit(size_t offset) {
3376 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3377 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3378 assert( addr < _cm->finger(), "invariant");
3379
3380 statsOnly( _task->increase_objs_found_on_bitmap() );
3381 assert(addr >= _task->finger(), "invariant");
3382
3383 // We move that task's local finger along.
3384 _task->move_finger_to(addr);
3385
3386 _task->scan_object(oop(addr));
3387 // we only partially drain the local queue and global stack
3388 _task->drain_local_queue(true);
3389 _task->drain_global_stack(true);
3390
3391 // if the has_aborted flag has been raised, we need to bail out of
3392 // the iteration
3393 return !_task->has_aborted();
3394 }
3395 };
3396
3397 // Closure for iterating over objects, currently only used for
3398 // processing SATB buffers.
3399 class CMObjectClosure : public ObjectClosure {
3400 private:
3401 CMTask* _task;
3402
3403 public:
3404 void do_object(oop obj) {
3405 _task->deal_with_reference(obj);
3406 }
3407
3408 CMObjectClosure(CMTask* task) : _task(task) { }
3409 };
3410
3411 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3412 ConcurrentMark* cm,
3413 CMTask* task)
3414 : _g1h(g1h), _cm(cm), _task(task) {
3415 assert(_ref_processor == NULL, "should be initialized to NULL");
3416
3417 if (G1UseConcMarkReferenceProcessing) {
3418 _ref_processor = g1h->ref_processor_cm();
3419 assert(_ref_processor != NULL, "should not be NULL");
3420 }
3421 }
3422
3423 void CMTask::setup_for_region(HeapRegion* hr) {
3424 assert(hr != NULL,
3425 "claim_region() should have filtered out NULL regions");
3426 assert(!hr->continuesHumongous(),
3427 "claim_region() should have filtered out continues humongous regions");
3428
3429 if (_cm->verbose_low()) {
3430 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3431 _worker_id, p2i(hr));
3432 }
3433
3434 _curr_region = hr;
3435 _finger = hr->bottom();
3436 update_region_limit();
3437 }
3438
3439 void CMTask::update_region_limit() {
3440 HeapRegion* hr = _curr_region;
3441 HeapWord* bottom = hr->bottom();
3442 HeapWord* limit = hr->next_top_at_mark_start();
3443
3444 if (limit == bottom) {
3445 if (_cm->verbose_low()) {
3446 gclog_or_tty->print_cr("[%u] found an empty region "
3447 "["PTR_FORMAT", "PTR_FORMAT")",
3448 _worker_id, p2i(bottom), p2i(limit));
3449 }
3450 // The region was collected underneath our feet.
3451 // We set the finger to bottom to ensure that the bitmap
3452 // iteration that will follow this will not do anything.
3453 // (this is not a condition that holds when we set the region up,
3454 // as the region is not supposed to be empty in the first place)
3455 _finger = bottom;
3456 } else if (limit >= _region_limit) {
3457 assert(limit >= _finger, "peace of mind");
3458 } else {
3459 assert(limit < _region_limit, "only way to get here");
3460 // This can happen under some pretty unusual circumstances. An
3461 // evacuation pause empties the region underneath our feet (NTAMS
3462 // at bottom). We then do some allocation in the region (NTAMS
3463 // stays at bottom), followed by the region being used as a GC
3464 // alloc region (NTAMS will move to top() and the objects
3465 // originally below it will be grayed). All objects now marked in
3466 // the region are explicitly grayed, if below the global finger,
3467 // and we do not need in fact to scan anything else. So, we simply
3468 // set _finger to be limit to ensure that the bitmap iteration
3469 // doesn't do anything.
3470 _finger = limit;
3471 }
3472
3473 _region_limit = limit;
3474 }
3475
3476 void CMTask::giveup_current_region() {
3477 assert(_curr_region != NULL, "invariant");
3478 if (_cm->verbose_low()) {
3479 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3480 _worker_id, p2i(_curr_region));
3481 }
3482 clear_region_fields();
3483 }
3484
3485 void CMTask::clear_region_fields() {
3486 // Values for these three fields that indicate that we're not
3487 // holding on to a region.
3488 _curr_region = NULL;
3489 _finger = NULL;
3490 _region_limit = NULL;
3491 }
3492
3493 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3494 if (cm_oop_closure == NULL) {
3495 assert(_cm_oop_closure != NULL, "invariant");
3496 } else {
3497 assert(_cm_oop_closure == NULL, "invariant");
3498 }
3499 _cm_oop_closure = cm_oop_closure;
3500 }
3501
3502 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3503 guarantee(nextMarkBitMap != NULL, "invariant");
3504
3505 if (_cm->verbose_low()) {
3506 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3507 }
3508
3509 _nextMarkBitMap = nextMarkBitMap;
3510 clear_region_fields();
3511
3512 _calls = 0;
3513 _elapsed_time_ms = 0.0;
3514 _termination_time_ms = 0.0;
3515 _termination_start_time_ms = 0.0;
3516
3517 #if _MARKING_STATS_
3518 _local_pushes = 0;
3519 _local_pops = 0;
3520 _local_max_size = 0;
3521 _objs_scanned = 0;
3522 _global_pushes = 0;
3523 _global_pops = 0;
3524 _global_max_size = 0;
3525 _global_transfers_to = 0;
3526 _global_transfers_from = 0;
3527 _regions_claimed = 0;
3528 _objs_found_on_bitmap = 0;
3529 _satb_buffers_processed = 0;
3530 _steal_attempts = 0;
3531 _steals = 0;
3532 _aborted = 0;
3533 _aborted_overflow = 0;
3534 _aborted_cm_aborted = 0;
3535 _aborted_yield = 0;
3536 _aborted_timed_out = 0;
3537 _aborted_satb = 0;
3538 _aborted_termination = 0;
3539 #endif // _MARKING_STATS_
3540 }
3541
3542 bool CMTask::should_exit_termination() {
3543 regular_clock_call();
3544 // This is called when we are in the termination protocol. We should
3545 // quit if, for some reason, this task wants to abort or the global
3546 // stack is not empty (this means that we can get work from it).
3547 return !_cm->mark_stack_empty() || has_aborted();
3548 }
3549
3550 void CMTask::reached_limit() {
3551 assert(_words_scanned >= _words_scanned_limit ||
3552 _refs_reached >= _refs_reached_limit ,
3553 "shouldn't have been called otherwise");
3554 regular_clock_call();
3555 }
3556
3557 void CMTask::regular_clock_call() {
3558 if (has_aborted()) return;
3559
3560 // First, we need to recalculate the words scanned and refs reached
3561 // limits for the next clock call.
3562 recalculate_limits();
3563
3564 // During the regular clock call we do the following
3565
3566 // (1) If an overflow has been flagged, then we abort.
3567 if (_cm->has_overflown()) {
3568 set_has_aborted();
3569 return;
3570 }
3571
3572 // If we are not concurrent (i.e. we're doing remark) we don't need
3573 // to check anything else. The other steps are only needed during
3574 // the concurrent marking phase.
3575 if (!concurrent()) return;
3576
3577 // (2) If marking has been aborted for Full GC, then we also abort.
3578 if (_cm->has_aborted()) {
3579 set_has_aborted();
3580 statsOnly( ++_aborted_cm_aborted );
3581 return;
3582 }
3583
3584 double curr_time_ms = os::elapsedVTime() * 1000.0;
3585
3586 // (3) If marking stats are enabled, then we update the step history.
3587 #if _MARKING_STATS_
3588 if (_words_scanned >= _words_scanned_limit) {
3589 ++_clock_due_to_scanning;
3590 }
3591 if (_refs_reached >= _refs_reached_limit) {
3592 ++_clock_due_to_marking;
3593 }
3594
3595 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3596 _interval_start_time_ms = curr_time_ms;
3597 _all_clock_intervals_ms.add(last_interval_ms);
3598
3599 if (_cm->verbose_medium()) {
3600 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3601 "scanned = %d%s, refs reached = %d%s",
3602 _worker_id, last_interval_ms,
3603 _words_scanned,
3604 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3605 _refs_reached,
3606 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3607 }
3608 #endif // _MARKING_STATS_
3609
3610 // (4) We check whether we should yield. If we have to, then we abort.
3611 if (SuspendibleThreadSet::should_yield()) {
3612 // We should yield. To do this we abort the task. The caller is
3613 // responsible for yielding.
3614 set_has_aborted();
3615 statsOnly( ++_aborted_yield );
3616 return;
3617 }
3618
3619 // (5) We check whether we've reached our time quota. If we have,
3620 // then we abort.
3621 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3622 if (elapsed_time_ms > _time_target_ms) {
3623 set_has_aborted();
3624 _has_timed_out = true;
3625 statsOnly( ++_aborted_timed_out );
3626 return;
3627 }
3628
3629 // (6) Finally, we check whether there are enough completed STAB
3630 // buffers available for processing. If there are, we abort.
3631 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3632 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3633 if (_cm->verbose_low()) {
3634 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3635 _worker_id);
3636 }
3637 // we do need to process SATB buffers, we'll abort and restart
3638 // the marking task to do so
3639 set_has_aborted();
3640 statsOnly( ++_aborted_satb );
3641 return;
3642 }
3643 }
3644
3645 void CMTask::recalculate_limits() {
3646 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3647 _words_scanned_limit = _real_words_scanned_limit;
3648
3649 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3650 _refs_reached_limit = _real_refs_reached_limit;
3651 }
3652
3653 void CMTask::decrease_limits() {
3654 // This is called when we believe that we're going to do an infrequent
3655 // operation which will increase the per byte scanned cost (i.e. move
3656 // entries to/from the global stack). It basically tries to decrease the
3657 // scanning limit so that the clock is called earlier.
3658
3659 if (_cm->verbose_medium()) {
3660 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3661 }
3662
3663 _words_scanned_limit = _real_words_scanned_limit -
3664 3 * words_scanned_period / 4;
3665 _refs_reached_limit = _real_refs_reached_limit -
3666 3 * refs_reached_period / 4;
3667 }
3668
3669 void CMTask::move_entries_to_global_stack() {
3670 // local array where we'll store the entries that will be popped
3671 // from the local queue
3672 oop buffer[global_stack_transfer_size];
3673
3674 int n = 0;
3675 oop obj;
3676 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3677 buffer[n] = obj;
3678 ++n;
3679 }
3680
3681 if (n > 0) {
3682 // we popped at least one entry from the local queue
3683
3684 statsOnly( ++_global_transfers_to; _local_pops += n );
3685
3686 if (!_cm->mark_stack_push(buffer, n)) {
3687 if (_cm->verbose_low()) {
3688 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3689 _worker_id);
3690 }
3691 set_has_aborted();
3692 } else {
3693 // the transfer was successful
3694
3695 if (_cm->verbose_medium()) {
3696 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3697 _worker_id, n);
3698 }
3699 statsOnly( int tmp_size = _cm->mark_stack_size();
3700 if (tmp_size > _global_max_size) {
3701 _global_max_size = tmp_size;
3702 }
3703 _global_pushes += n );
3704 }
3705 }
3706
3707 // this operation was quite expensive, so decrease the limits
3708 decrease_limits();
3709 }
3710
3711 void CMTask::get_entries_from_global_stack() {
3712 // local array where we'll store the entries that will be popped
3713 // from the global stack.
3714 oop buffer[global_stack_transfer_size];
3715 int n;
3716 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3717 assert(n <= global_stack_transfer_size,
3718 "we should not pop more than the given limit");
3719 if (n > 0) {
3720 // yes, we did actually pop at least one entry
3721
3722 statsOnly( ++_global_transfers_from; _global_pops += n );
3723 if (_cm->verbose_medium()) {
3724 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3725 _worker_id, n);
3726 }
3727 for (int i = 0; i < n; ++i) {
3728 bool success = _task_queue->push(buffer[i]);
3729 // We only call this when the local queue is empty or under a
3730 // given target limit. So, we do not expect this push to fail.
3731 assert(success, "invariant");
3732 }
3733
3734 statsOnly( int tmp_size = _task_queue->size();
3735 if (tmp_size > _local_max_size) {
3736 _local_max_size = tmp_size;
3737 }
3738 _local_pushes += n );
3739 }
3740
3741 // this operation was quite expensive, so decrease the limits
3742 decrease_limits();
3743 }
3744
3745 void CMTask::drain_local_queue(bool partially) {
3746 if (has_aborted()) return;
3747
3748 // Decide what the target size is, depending whether we're going to
3749 // drain it partially (so that other tasks can steal if they run out
3750 // of things to do) or totally (at the very end).
3751 size_t target_size;
3752 if (partially) {
3753 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3754 } else {
3755 target_size = 0;
3756 }
3757
3758 if (_task_queue->size() > target_size) {
3759 if (_cm->verbose_high()) {
3760 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3761 _worker_id, target_size);
3762 }
3763
3764 oop obj;
3765 bool ret = _task_queue->pop_local(obj);
3766 while (ret) {
3767 statsOnly( ++_local_pops );
3768
3769 if (_cm->verbose_high()) {
3770 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3771 p2i((void*) obj));
3772 }
3773
3774 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3775 assert(!_g1h->is_on_master_free_list(
3776 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3777
3778 scan_object(obj);
3779
3780 if (_task_queue->size() <= target_size || has_aborted()) {
3781 ret = false;
3782 } else {
3783 ret = _task_queue->pop_local(obj);
3784 }
3785 }
3786
3787 if (_cm->verbose_high()) {
3788 gclog_or_tty->print_cr("[%u] drained local queue, size = %u",
3789 _worker_id, _task_queue->size());
3790 }
3791 }
3792 }
3793
3794 void CMTask::drain_global_stack(bool partially) {
3795 if (has_aborted()) return;
3796
3797 // We have a policy to drain the local queue before we attempt to
3798 // drain the global stack.
3799 assert(partially || _task_queue->size() == 0, "invariant");
3800
3801 // Decide what the target size is, depending whether we're going to
3802 // drain it partially (so that other tasks can steal if they run out
3803 // of things to do) or totally (at the very end). Notice that,
3804 // because we move entries from the global stack in chunks or
3805 // because another task might be doing the same, we might in fact
3806 // drop below the target. But, this is not a problem.
3807 size_t target_size;
3808 if (partially) {
3809 target_size = _cm->partial_mark_stack_size_target();
3810 } else {
3811 target_size = 0;
3812 }
3813
3814 if (_cm->mark_stack_size() > target_size) {
3815 if (_cm->verbose_low()) {
3816 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3817 _worker_id, target_size);
3818 }
3819
3820 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3821 get_entries_from_global_stack();
3822 drain_local_queue(partially);
3823 }
3824
3825 if (_cm->verbose_low()) {
3826 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3827 _worker_id, _cm->mark_stack_size());
3828 }
3829 }
3830 }
3831
3832 // SATB Queue has several assumptions on whether to call the par or
3833 // non-par versions of the methods. this is why some of the code is
3834 // replicated. We should really get rid of the single-threaded version
3835 // of the code to simplify things.
3836 void CMTask::drain_satb_buffers() {
3837 if (has_aborted()) return;
3838
3839 // We set this so that the regular clock knows that we're in the
3840 // middle of draining buffers and doesn't set the abort flag when it
3841 // notices that SATB buffers are available for draining. It'd be
3842 // very counter productive if it did that. :-)
3843 _draining_satb_buffers = true;
3844
3845 CMObjectClosure oc(this);
3846 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3847 if (G1CollectedHeap::use_parallel_gc_threads()) {
3848 satb_mq_set.set_par_closure(_worker_id, &oc);
3849 } else {
3850 satb_mq_set.set_closure(&oc);
3851 }
3852
3853 // This keeps claiming and applying the closure to completed buffers
3854 // until we run out of buffers or we need to abort.
3855 if (G1CollectedHeap::use_parallel_gc_threads()) {
3856 while (!has_aborted() &&
3857 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
3858 if (_cm->verbose_medium()) {
3859 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3860 }
3861 statsOnly( ++_satb_buffers_processed );
3862 regular_clock_call();
3863 }
3864 } else {
3865 while (!has_aborted() &&
3866 satb_mq_set.apply_closure_to_completed_buffer()) {
3867 if (_cm->verbose_medium()) {
3868 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3869 }
3870 statsOnly( ++_satb_buffers_processed );
3871 regular_clock_call();
3872 }
3873 }
3874
3875 if (!concurrent() && !has_aborted()) {
3876 // We should only do this during remark.
3877 if (G1CollectedHeap::use_parallel_gc_threads()) {
3878 satb_mq_set.par_iterate_closure_all_threads(_worker_id);
3879 } else {
3880 satb_mq_set.iterate_closure_all_threads();
3881 }
3882 }
3883
3884 _draining_satb_buffers = false;
3885
3886 assert(has_aborted() ||
3887 concurrent() ||
3888 satb_mq_set.completed_buffers_num() == 0, "invariant");
3889
3890 if (G1CollectedHeap::use_parallel_gc_threads()) {
3891 satb_mq_set.set_par_closure(_worker_id, NULL);
3892 } else {
3893 satb_mq_set.set_closure(NULL);
3894 }
3895
3896 // again, this was a potentially expensive operation, decrease the
3897 // limits to get the regular clock call early
3898 decrease_limits();
3899 }
3900
3901 void CMTask::print_stats() {
3902 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3903 _worker_id, _calls);
3904 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3905 _elapsed_time_ms, _termination_time_ms);
3906 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3907 _step_times_ms.num(), _step_times_ms.avg(),
3908 _step_times_ms.sd());
3909 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3910 _step_times_ms.maximum(), _step_times_ms.sum());
3911
3912 #if _MARKING_STATS_
3913 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3914 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3915 _all_clock_intervals_ms.sd());
3916 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3917 _all_clock_intervals_ms.maximum(),
3918 _all_clock_intervals_ms.sum());
3919 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
3920 _clock_due_to_scanning, _clock_due_to_marking);
3921 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
3922 _objs_scanned, _objs_found_on_bitmap);
3923 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
3924 _local_pushes, _local_pops, _local_max_size);
3925 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
3926 _global_pushes, _global_pops, _global_max_size);
3927 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
3928 _global_transfers_to,_global_transfers_from);
3929 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
3930 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
3931 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
3932 _steal_attempts, _steals);
3933 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
3934 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
3935 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3936 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
3937 _aborted_timed_out, _aborted_satb, _aborted_termination);
3938 #endif // _MARKING_STATS_
3939 }
3940
3941 /*****************************************************************************
3942
3943 The do_marking_step(time_target_ms, ...) method is the building
3944 block of the parallel marking framework. It can be called in parallel
3945 with other invocations of do_marking_step() on different tasks
3946 (but only one per task, obviously) and concurrently with the
3947 mutator threads, or during remark, hence it eliminates the need
3948 for two versions of the code. When called during remark, it will
3949 pick up from where the task left off during the concurrent marking
3950 phase. Interestingly, tasks are also claimable during evacuation
3951 pauses too, since do_marking_step() ensures that it aborts before
3952 it needs to yield.
3953
3954 The data structures that it uses to do marking work are the
3955 following:
3956
3957 (1) Marking Bitmap. If there are gray objects that appear only
3958 on the bitmap (this happens either when dealing with an overflow
3959 or when the initial marking phase has simply marked the roots
3960 and didn't push them on the stack), then tasks claim heap
3961 regions whose bitmap they then scan to find gray objects. A
3962 global finger indicates where the end of the last claimed region
3963 is. A local finger indicates how far into the region a task has
3964 scanned. The two fingers are used to determine how to gray an
3965 object (i.e. whether simply marking it is OK, as it will be
3966 visited by a task in the future, or whether it needs to be also
3967 pushed on a stack).
3968
3969 (2) Local Queue. The local queue of the task which is accessed
3970 reasonably efficiently by the task. Other tasks can steal from
3971 it when they run out of work. Throughout the marking phase, a
3972 task attempts to keep its local queue short but not totally
3973 empty, so that entries are available for stealing by other
3974 tasks. Only when there is no more work, a task will totally
3975 drain its local queue.
3976
3977 (3) Global Mark Stack. This handles local queue overflow. During
3978 marking only sets of entries are moved between it and the local
3979 queues, as access to it requires a mutex and more fine-grain
3980 interaction with it which might cause contention. If it
3981 overflows, then the marking phase should restart and iterate
3982 over the bitmap to identify gray objects. Throughout the marking
3983 phase, tasks attempt to keep the global mark stack at a small
3984 length but not totally empty, so that entries are available for
3985 popping by other tasks. Only when there is no more work, tasks
3986 will totally drain the global mark stack.
3987
3988 (4) SATB Buffer Queue. This is where completed SATB buffers are
3989 made available. Buffers are regularly removed from this queue
3990 and scanned for roots, so that the queue doesn't get too
3991 long. During remark, all completed buffers are processed, as
3992 well as the filled in parts of any uncompleted buffers.
3993
3994 The do_marking_step() method tries to abort when the time target
3995 has been reached. There are a few other cases when the
3996 do_marking_step() method also aborts:
3997
3998 (1) When the marking phase has been aborted (after a Full GC).
3999
4000 (2) When a global overflow (on the global stack) has been
4001 triggered. Before the task aborts, it will actually sync up with
4002 the other tasks to ensure that all the marking data structures
4003 (local queues, stacks, fingers etc.) are re-initialized so that
4004 when do_marking_step() completes, the marking phase can
4005 immediately restart.
4006
4007 (3) When enough completed SATB buffers are available. The
4008 do_marking_step() method only tries to drain SATB buffers right
4009 at the beginning. So, if enough buffers are available, the
4010 marking step aborts and the SATB buffers are processed at
4011 the beginning of the next invocation.
4012
4013 (4) To yield. when we have to yield then we abort and yield
4014 right at the end of do_marking_step(). This saves us from a lot
4015 of hassle as, by yielding we might allow a Full GC. If this
4016 happens then objects will be compacted underneath our feet, the
4017 heap might shrink, etc. We save checking for this by just
4018 aborting and doing the yield right at the end.
4019
4020 From the above it follows that the do_marking_step() method should
4021 be called in a loop (or, otherwise, regularly) until it completes.
4022
4023 If a marking step completes without its has_aborted() flag being
4024 true, it means it has completed the current marking phase (and
4025 also all other marking tasks have done so and have all synced up).
4026
4027 A method called regular_clock_call() is invoked "regularly" (in
4028 sub ms intervals) throughout marking. It is this clock method that
4029 checks all the abort conditions which were mentioned above and
4030 decides when the task should abort. A work-based scheme is used to
4031 trigger this clock method: when the number of object words the
4032 marking phase has scanned or the number of references the marking
4033 phase has visited reach a given limit. Additional invocations to
4034 the method clock have been planted in a few other strategic places
4035 too. The initial reason for the clock method was to avoid calling
4036 vtime too regularly, as it is quite expensive. So, once it was in
4037 place, it was natural to piggy-back all the other conditions on it
4038 too and not constantly check them throughout the code.
4039
4040 If do_termination is true then do_marking_step will enter its
4041 termination protocol.
4042
4043 The value of is_serial must be true when do_marking_step is being
4044 called serially (i.e. by the VMThread) and do_marking_step should
4045 skip any synchronization in the termination and overflow code.
4046 Examples include the serial remark code and the serial reference
4047 processing closures.
4048
4049 The value of is_serial must be false when do_marking_step is
4050 being called by any of the worker threads in a work gang.
4051 Examples include the concurrent marking code (CMMarkingTask),
4052 the MT remark code, and the MT reference processing closures.
4053
4054 *****************************************************************************/
4055
4056 void CMTask::do_marking_step(double time_target_ms,
4057 bool do_termination,
4058 bool is_serial) {
4059 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4060 assert(concurrent() == _cm->concurrent(), "they should be the same");
4061
4062 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4063 assert(_task_queues != NULL, "invariant");
4064 assert(_task_queue != NULL, "invariant");
4065 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4066
4067 assert(!_claimed,
4068 "only one thread should claim this task at any one time");
4069
4070 // OK, this doesn't safeguard again all possible scenarios, as it is
4071 // possible for two threads to set the _claimed flag at the same
4072 // time. But it is only for debugging purposes anyway and it will
4073 // catch most problems.
4074 _claimed = true;
4075
4076 _start_time_ms = os::elapsedVTime() * 1000.0;
4077 statsOnly( _interval_start_time_ms = _start_time_ms );
4078
4079 // If do_stealing is true then do_marking_step will attempt to
4080 // steal work from the other CMTasks. It only makes sense to
4081 // enable stealing when the termination protocol is enabled
4082 // and do_marking_step() is not being called serially.
4083 bool do_stealing = do_termination && !is_serial;
4084
4085 double diff_prediction_ms =
4086 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4087 _time_target_ms = time_target_ms - diff_prediction_ms;
4088
4089 // set up the variables that are used in the work-based scheme to
4090 // call the regular clock method
4091 _words_scanned = 0;
4092 _refs_reached = 0;
4093 recalculate_limits();
4094
4095 // clear all flags
4096 clear_has_aborted();
4097 _has_timed_out = false;
4098 _draining_satb_buffers = false;
4099
4100 ++_calls;
4101
4102 if (_cm->verbose_low()) {
4103 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4104 "target = %1.2lfms >>>>>>>>>>",
4105 _worker_id, _calls, _time_target_ms);
4106 }
4107
4108 // Set up the bitmap and oop closures. Anything that uses them is
4109 // eventually called from this method, so it is OK to allocate these
4110 // statically.
4111 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4112 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4113 set_cm_oop_closure(&cm_oop_closure);
4114
4115 if (_cm->has_overflown()) {
4116 // This can happen if the mark stack overflows during a GC pause
4117 // and this task, after a yield point, restarts. We have to abort
4118 // as we need to get into the overflow protocol which happens
4119 // right at the end of this task.
4120 set_has_aborted();
4121 }
4122
4123 // First drain any available SATB buffers. After this, we will not
4124 // look at SATB buffers before the next invocation of this method.
4125 // If enough completed SATB buffers are queued up, the regular clock
4126 // will abort this task so that it restarts.
4127 drain_satb_buffers();
4128 // ...then partially drain the local queue and the global stack
4129 drain_local_queue(true);
4130 drain_global_stack(true);
4131
4132 do {
4133 if (!has_aborted() && _curr_region != NULL) {
4134 // This means that we're already holding on to a region.
4135 assert(_finger != NULL, "if region is not NULL, then the finger "
4136 "should not be NULL either");
4137
4138 // We might have restarted this task after an evacuation pause
4139 // which might have evacuated the region we're holding on to
4140 // underneath our feet. Let's read its limit again to make sure
4141 // that we do not iterate over a region of the heap that
4142 // contains garbage (update_region_limit() will also move
4143 // _finger to the start of the region if it is found empty).
4144 update_region_limit();
4145 // We will start from _finger not from the start of the region,
4146 // as we might be restarting this task after aborting half-way
4147 // through scanning this region. In this case, _finger points to
4148 // the address where we last found a marked object. If this is a
4149 // fresh region, _finger points to start().
4150 MemRegion mr = MemRegion(_finger, _region_limit);
4151
4152 if (_cm->verbose_low()) {
4153 gclog_or_tty->print_cr("[%u] we're scanning part "
4154 "["PTR_FORMAT", "PTR_FORMAT") "
4155 "of region "HR_FORMAT,
4156 _worker_id, p2i(_finger), p2i(_region_limit),
4157 HR_FORMAT_PARAMS(_curr_region));
4158 }
4159
4160 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4161 "humongous regions should go around loop once only");
4162
4163 // Some special cases:
4164 // If the memory region is empty, we can just give up the region.
4165 // If the current region is humongous then we only need to check
4166 // the bitmap for the bit associated with the start of the object,
4167 // scan the object if it's live, and give up the region.
4168 // Otherwise, let's iterate over the bitmap of the part of the region
4169 // that is left.
4170 // If the iteration is successful, give up the region.
4171 if (mr.is_empty()) {
4172 giveup_current_region();
4173 regular_clock_call();
4174 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4175 if (_nextMarkBitMap->isMarked(mr.start())) {
4176 // The object is marked - apply the closure
4177 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4178 bitmap_closure.do_bit(offset);
4179 }
4180 // Even if this task aborted while scanning the humongous object
4181 // we can (and should) give up the current region.
4182 giveup_current_region();
4183 regular_clock_call();
4184 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4185 giveup_current_region();
4186 regular_clock_call();
4187 } else {
4188 assert(has_aborted(), "currently the only way to do so");
4189 // The only way to abort the bitmap iteration is to return
4190 // false from the do_bit() method. However, inside the
4191 // do_bit() method we move the _finger to point to the
4192 // object currently being looked at. So, if we bail out, we
4193 // have definitely set _finger to something non-null.
4194 assert(_finger != NULL, "invariant");
4195
4196 // Region iteration was actually aborted. So now _finger
4197 // points to the address of the object we last scanned. If we
4198 // leave it there, when we restart this task, we will rescan
4199 // the object. It is easy to avoid this. We move the finger by
4200 // enough to point to the next possible object header (the
4201 // bitmap knows by how much we need to move it as it knows its
4202 // granularity).
4203 assert(_finger < _region_limit, "invariant");
4204 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4205 // Check if bitmap iteration was aborted while scanning the last object
4206 if (new_finger >= _region_limit) {
4207 giveup_current_region();
4208 } else {
4209 move_finger_to(new_finger);
4210 }
4211 }
4212 }
4213 // At this point we have either completed iterating over the
4214 // region we were holding on to, or we have aborted.
4215
4216 // We then partially drain the local queue and the global stack.
4217 // (Do we really need this?)
4218 drain_local_queue(true);
4219 drain_global_stack(true);
4220
4221 // Read the note on the claim_region() method on why it might
4222 // return NULL with potentially more regions available for
4223 // claiming and why we have to check out_of_regions() to determine
4224 // whether we're done or not.
4225 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4226 // We are going to try to claim a new region. We should have
4227 // given up on the previous one.
4228 // Separated the asserts so that we know which one fires.
4229 assert(_curr_region == NULL, "invariant");
4230 assert(_finger == NULL, "invariant");
4231 assert(_region_limit == NULL, "invariant");
4232 if (_cm->verbose_low()) {
4233 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4234 }
4235 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4236 if (claimed_region != NULL) {
4237 // Yes, we managed to claim one
4238 statsOnly( ++_regions_claimed );
4239
4240 if (_cm->verbose_low()) {
4241 gclog_or_tty->print_cr("[%u] we successfully claimed "
4242 "region "PTR_FORMAT,
4243 _worker_id, p2i(claimed_region));
4244 }
4245
4246 setup_for_region(claimed_region);
4247 assert(_curr_region == claimed_region, "invariant");
4248 }
4249 // It is important to call the regular clock here. It might take
4250 // a while to claim a region if, for example, we hit a large
4251 // block of empty regions. So we need to call the regular clock
4252 // method once round the loop to make sure it's called
4253 // frequently enough.
4254 regular_clock_call();
4255 }
4256
4257 if (!has_aborted() && _curr_region == NULL) {
4258 assert(_cm->out_of_regions(),
4259 "at this point we should be out of regions");
4260 }
4261 } while ( _curr_region != NULL && !has_aborted());
4262
4263 if (!has_aborted()) {
4264 // We cannot check whether the global stack is empty, since other
4265 // tasks might be pushing objects to it concurrently.
4266 assert(_cm->out_of_regions(),
4267 "at this point we should be out of regions");
4268
4269 if (_cm->verbose_low()) {
4270 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4271 }
4272
4273 // Try to reduce the number of available SATB buffers so that
4274 // remark has less work to do.
4275 drain_satb_buffers();
4276 }
4277
4278 // Since we've done everything else, we can now totally drain the
4279 // local queue and global stack.
4280 drain_local_queue(false);
4281 drain_global_stack(false);
4282
4283 // Attempt at work stealing from other task's queues.
4284 if (do_stealing && !has_aborted()) {
4285 // We have not aborted. This means that we have finished all that
4286 // we could. Let's try to do some stealing...
4287
4288 // We cannot check whether the global stack is empty, since other
4289 // tasks might be pushing objects to it concurrently.
4290 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4291 "only way to reach here");
4292
4293 if (_cm->verbose_low()) {
4294 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4295 }
4296
4297 while (!has_aborted()) {
4298 oop obj;
4299 statsOnly( ++_steal_attempts );
4300
4301 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4302 if (_cm->verbose_medium()) {
4303 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4304 _worker_id, p2i((void*) obj));
4305 }
4306
4307 statsOnly( ++_steals );
4308
4309 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4310 "any stolen object should be marked");
4311 scan_object(obj);
4312
4313 // And since we're towards the end, let's totally drain the
4314 // local queue and global stack.
4315 drain_local_queue(false);
4316 drain_global_stack(false);
4317 } else {
4318 break;
4319 }
4320 }
4321 }
4322
4323 // If we are about to wrap up and go into termination, check if we
4324 // should raise the overflow flag.
4325 if (do_termination && !has_aborted()) {
4326 if (_cm->force_overflow()->should_force()) {
4327 _cm->set_has_overflown();
4328 regular_clock_call();
4329 }
4330 }
4331
4332 // We still haven't aborted. Now, let's try to get into the
4333 // termination protocol.
4334 if (do_termination && !has_aborted()) {
4335 // We cannot check whether the global stack is empty, since other
4336 // tasks might be concurrently pushing objects on it.
4337 // Separated the asserts so that we know which one fires.
4338 assert(_cm->out_of_regions(), "only way to reach here");
4339 assert(_task_queue->size() == 0, "only way to reach here");
4340
4341 if (_cm->verbose_low()) {
4342 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4343 }
4344
4345 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4346
4347 // The CMTask class also extends the TerminatorTerminator class,
4348 // hence its should_exit_termination() method will also decide
4349 // whether to exit the termination protocol or not.
4350 bool finished = (is_serial ||
4351 _cm->terminator()->offer_termination(this));
4352 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4353 _termination_time_ms +=
4354 termination_end_time_ms - _termination_start_time_ms;
4355
4356 if (finished) {
4357 // We're all done.
4358
4359 if (_worker_id == 0) {
4360 // let's allow task 0 to do this
4361 if (concurrent()) {
4362 assert(_cm->concurrent_marking_in_progress(), "invariant");
4363 // we need to set this to false before the next
4364 // safepoint. This way we ensure that the marking phase
4365 // doesn't observe any more heap expansions.
4366 _cm->clear_concurrent_marking_in_progress();
4367 }
4368 }
4369
4370 // We can now guarantee that the global stack is empty, since
4371 // all other tasks have finished. We separated the guarantees so
4372 // that, if a condition is false, we can immediately find out
4373 // which one.
4374 guarantee(_cm->out_of_regions(), "only way to reach here");
4375 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4376 guarantee(_task_queue->size() == 0, "only way to reach here");
4377 guarantee(!_cm->has_overflown(), "only way to reach here");
4378 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4379
4380 if (_cm->verbose_low()) {
4381 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4382 }
4383 } else {
4384 // Apparently there's more work to do. Let's abort this task. It
4385 // will restart it and we can hopefully find more things to do.
4386
4387 if (_cm->verbose_low()) {
4388 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4389 _worker_id);
4390 }
4391
4392 set_has_aborted();
4393 statsOnly( ++_aborted_termination );
4394 }
4395 }
4396
4397 // Mainly for debugging purposes to make sure that a pointer to the
4398 // closure which was statically allocated in this frame doesn't
4399 // escape it by accident.
4400 set_cm_oop_closure(NULL);
4401 double end_time_ms = os::elapsedVTime() * 1000.0;
4402 double elapsed_time_ms = end_time_ms - _start_time_ms;
4403 // Update the step history.
4404 _step_times_ms.add(elapsed_time_ms);
4405
4406 if (has_aborted()) {
4407 // The task was aborted for some reason.
4408
4409 statsOnly( ++_aborted );
4410
4411 if (_has_timed_out) {
4412 double diff_ms = elapsed_time_ms - _time_target_ms;
4413 // Keep statistics of how well we did with respect to hitting
4414 // our target only if we actually timed out (if we aborted for
4415 // other reasons, then the results might get skewed).
4416 _marking_step_diffs_ms.add(diff_ms);
4417 }
4418
4419 if (_cm->has_overflown()) {
4420 // This is the interesting one. We aborted because a global
4421 // overflow was raised. This means we have to restart the
4422 // marking phase and start iterating over regions. However, in
4423 // order to do this we have to make sure that all tasks stop
4424 // what they are doing and re-initialize in a safe manner. We
4425 // will achieve this with the use of two barrier sync points.
4426
4427 if (_cm->verbose_low()) {
4428 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4429 }
4430
4431 if (!is_serial) {
4432 // We only need to enter the sync barrier if being called
4433 // from a parallel context
4434 _cm->enter_first_sync_barrier(_worker_id);
4435
4436 // When we exit this sync barrier we know that all tasks have
4437 // stopped doing marking work. So, it's now safe to
4438 // re-initialize our data structures. At the end of this method,
4439 // task 0 will clear the global data structures.
4440 }
4441
4442 statsOnly( ++_aborted_overflow );
4443
4444 // We clear the local state of this task...
4445 clear_region_fields();
4446
4447 if (!is_serial) {
4448 // ...and enter the second barrier.
4449 _cm->enter_second_sync_barrier(_worker_id);
4450 }
4451 // At this point, if we're during the concurrent phase of
4452 // marking, everything has been re-initialized and we're
4453 // ready to restart.
4454 }
4455
4456 if (_cm->verbose_low()) {
4457 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4458 "elapsed = %1.2lfms <<<<<<<<<<",
4459 _worker_id, _time_target_ms, elapsed_time_ms);
4460 if (_cm->has_aborted()) {
4461 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4462 _worker_id);
4463 }
4464 }
4465 } else {
4466 if (_cm->verbose_low()) {
4467 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4468 "elapsed = %1.2lfms <<<<<<<<<<",
4469 _worker_id, _time_target_ms, elapsed_time_ms);
4470 }
4471 }
4472
4473 _claimed = false;
4474 }
4475
4476 CMTask::CMTask(uint worker_id,
4477 ConcurrentMark* cm,
4478 size_t* marked_bytes,
4479 BitMap* card_bm,
4480 CMTaskQueue* task_queue,
4481 CMTaskQueueSet* task_queues)
4482 : _g1h(G1CollectedHeap::heap()),
4483 _worker_id(worker_id), _cm(cm),
4484 _claimed(false),
4485 _nextMarkBitMap(NULL), _hash_seed(17),
4486 _task_queue(task_queue),
4487 _task_queues(task_queues),
4488 _cm_oop_closure(NULL),
4489 _marked_bytes_array(marked_bytes),
4490 _card_bm(card_bm) {
4491 guarantee(task_queue != NULL, "invariant");
4492 guarantee(task_queues != NULL, "invariant");
4493
4494 statsOnly( _clock_due_to_scanning = 0;
4495 _clock_due_to_marking = 0 );
4496
4497 _marking_step_diffs_ms.add(0.5);
4498 }
4499
4500 // These are formatting macros that are used below to ensure
4501 // consistent formatting. The *_H_* versions are used to format the
4502 // header for a particular value and they should be kept consistent
4503 // with the corresponding macro. Also note that most of the macros add
4504 // the necessary white space (as a prefix) which makes them a bit
4505 // easier to compose.
4506
4507 // All the output lines are prefixed with this string to be able to
4508 // identify them easily in a large log file.
4509 #define G1PPRL_LINE_PREFIX "###"
4510
4511 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4512 #ifdef _LP64
4513 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4514 #else // _LP64
4515 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4516 #endif // _LP64
4517
4518 // For per-region info
4519 #define G1PPRL_TYPE_FORMAT " %-4s"
4520 #define G1PPRL_TYPE_H_FORMAT " %4s"
4521 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4522 #define G1PPRL_BYTE_H_FORMAT " %9s"
4523 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4524 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4525
4526 // For summary info
4527 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4528 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4529 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4530 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4531
4532 G1PrintRegionLivenessInfoClosure::
4533 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4534 : _out(out),
4535 _total_used_bytes(0), _total_capacity_bytes(0),
4536 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4537 _hum_used_bytes(0), _hum_capacity_bytes(0),
4538 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4539 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4540 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4541 MemRegion g1_committed = g1h->g1_committed();
4542 MemRegion g1_reserved = g1h->g1_reserved();
4543 double now = os::elapsedTime();
4544
4545 // Print the header of the output.
4546 _out->cr();
4547 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4548 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4549 G1PPRL_SUM_ADDR_FORMAT("committed")
4550 G1PPRL_SUM_ADDR_FORMAT("reserved")
4551 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4552 p2i(g1_committed.start()), p2i(g1_committed.end()),
4553 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4554 HeapRegion::GrainBytes);
4555 _out->print_cr(G1PPRL_LINE_PREFIX);
4556 _out->print_cr(G1PPRL_LINE_PREFIX
4557 G1PPRL_TYPE_H_FORMAT
4558 G1PPRL_ADDR_BASE_H_FORMAT
4559 G1PPRL_BYTE_H_FORMAT
4560 G1PPRL_BYTE_H_FORMAT
4561 G1PPRL_BYTE_H_FORMAT
4562 G1PPRL_DOUBLE_H_FORMAT
4563 G1PPRL_BYTE_H_FORMAT
4564 G1PPRL_BYTE_H_FORMAT,
4565 "type", "address-range",
4566 "used", "prev-live", "next-live", "gc-eff",
4567 "remset", "code-roots");
4568 _out->print_cr(G1PPRL_LINE_PREFIX
4569 G1PPRL_TYPE_H_FORMAT
4570 G1PPRL_ADDR_BASE_H_FORMAT
4571 G1PPRL_BYTE_H_FORMAT
4572 G1PPRL_BYTE_H_FORMAT
4573 G1PPRL_BYTE_H_FORMAT
4574 G1PPRL_DOUBLE_H_FORMAT
4575 G1PPRL_BYTE_H_FORMAT
4576 G1PPRL_BYTE_H_FORMAT,
4577 "", "",
4578 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4579 "(bytes)", "(bytes)");
4580 }
4581
4582 // It takes as a parameter a reference to one of the _hum_* fields, it
4583 // deduces the corresponding value for a region in a humongous region
4584 // series (either the region size, or what's left if the _hum_* field
4585 // is < the region size), and updates the _hum_* field accordingly.
4586 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4587 size_t bytes = 0;
4588 // The > 0 check is to deal with the prev and next live bytes which
4589 // could be 0.
4590 if (*hum_bytes > 0) {
4591 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4592 *hum_bytes -= bytes;
4593 }
4594 return bytes;
4595 }
4596
4597 // It deduces the values for a region in a humongous region series
4598 // from the _hum_* fields and updates those accordingly. It assumes
4599 // that that _hum_* fields have already been set up from the "starts
4600 // humongous" region and we visit the regions in address order.
4601 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4602 size_t* capacity_bytes,
4603 size_t* prev_live_bytes,
4604 size_t* next_live_bytes) {
4605 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4606 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4607 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4608 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4609 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4610 }
4611
4612 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4613 const char* type = "";
4614 HeapWord* bottom = r->bottom();
4615 HeapWord* end = r->end();
4616 size_t capacity_bytes = r->capacity();
4617 size_t used_bytes = r->used();
4618 size_t prev_live_bytes = r->live_bytes();
4619 size_t next_live_bytes = r->next_live_bytes();
4620 double gc_eff = r->gc_efficiency();
4621 size_t remset_bytes = r->rem_set()->mem_size();
4622 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4623
4624 if (r->used() == 0) {
4625 type = "FREE";
4626 } else if (r->is_survivor()) {
4627 type = "SURV";
4628 } else if (r->is_young()) {
4629 type = "EDEN";
4630 } else if (r->startsHumongous()) {
4631 type = "HUMS";
4632
4633 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4634 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4635 "they should have been zeroed after the last time we used them");
4636 // Set up the _hum_* fields.
4637 _hum_capacity_bytes = capacity_bytes;
4638 _hum_used_bytes = used_bytes;
4639 _hum_prev_live_bytes = prev_live_bytes;
4640 _hum_next_live_bytes = next_live_bytes;
4641 get_hum_bytes(&used_bytes, &capacity_bytes,
4642 &prev_live_bytes, &next_live_bytes);
4643 end = bottom + HeapRegion::GrainWords;
4644 } else if (r->continuesHumongous()) {
4645 type = "HUMC";
4646 get_hum_bytes(&used_bytes, &capacity_bytes,
4647 &prev_live_bytes, &next_live_bytes);
4648 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4649 } else {
4650 type = "OLD";
4651 }
4652
4653 _total_used_bytes += used_bytes;
4654 _total_capacity_bytes += capacity_bytes;
4655 _total_prev_live_bytes += prev_live_bytes;
4656 _total_next_live_bytes += next_live_bytes;
4657 _total_remset_bytes += remset_bytes;
4658 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4659
4660 // Print a line for this particular region.
4661 _out->print_cr(G1PPRL_LINE_PREFIX
4662 G1PPRL_TYPE_FORMAT
4663 G1PPRL_ADDR_BASE_FORMAT
4664 G1PPRL_BYTE_FORMAT
4665 G1PPRL_BYTE_FORMAT
4666 G1PPRL_BYTE_FORMAT
4667 G1PPRL_DOUBLE_FORMAT
4668 G1PPRL_BYTE_FORMAT
4669 G1PPRL_BYTE_FORMAT,
4670 type, p2i(bottom), p2i(end),
4671 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4672 remset_bytes, strong_code_roots_bytes);
4673
4674 return false;
4675 }
4676
4677 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4678 // add static memory usages to remembered set sizes
4679 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4680 // Print the footer of the output.
4681 _out->print_cr(G1PPRL_LINE_PREFIX);
4682 _out->print_cr(G1PPRL_LINE_PREFIX
4683 " SUMMARY"
4684 G1PPRL_SUM_MB_FORMAT("capacity")
4685 G1PPRL_SUM_MB_PERC_FORMAT("used")
4686 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4687 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4688 G1PPRL_SUM_MB_FORMAT("remset")
4689 G1PPRL_SUM_MB_FORMAT("code-roots"),
4690 bytes_to_mb(_total_capacity_bytes),
4691 bytes_to_mb(_total_used_bytes),
4692 perc(_total_used_bytes, _total_capacity_bytes),
4693 bytes_to_mb(_total_prev_live_bytes),
4694 perc(_total_prev_live_bytes, _total_capacity_bytes),
4695 bytes_to_mb(_total_next_live_bytes),
4696 perc(_total_next_live_bytes, _total_capacity_bytes),
4697 bytes_to_mb(_total_remset_bytes),
4698 bytes_to_mb(_total_strong_code_roots_bytes));
4699 _out->cr();
4700 }