1 /*
2 * Copyright (c) 2001, 2013, 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((uintptr_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(MinObjAlignment - 1),
485 _markBitMap2(MinObjAlignment - 1),
486
487 _parallel_marking_threads(0),
488 _max_parallel_marking_threads(0),
489 _sleep_factor(0.0),
490 _marking_task_overhead(1.0),
491 _cleanup_sleep_factor(0.0),
492 _cleanup_task_overhead(1.0),
493 _cleanup_list("Cleanup List"),
494 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
495 _card_bm((heap_rs.size() + CardTableModRefBS::card_size - 1) >>
496 CardTableModRefBS::card_shift,
497 false /* in_resource_area*/),
498
499 _prevMarkBitMap(&_markBitMap1),
500 _nextMarkBitMap(&_markBitMap2),
501
502 _markStack(this),
503 // _finger set in set_non_marking_state
504
505 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
506 // _active_tasks set in set_non_marking_state
507 // _tasks set inside the constructor
508 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
509 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
510
511 _has_overflown(false),
512 _concurrent(false),
513 _has_aborted(false),
514 _restart_for_overflow(false),
515 _concurrent_marking_in_progress(false),
516
517 // _verbose_level set below
518
519 _init_times(),
520 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
521 _cleanup_times(),
522 _total_counting_time(0.0),
523 _total_rs_scrub_time(0.0),
524
525 _parallel_workers(NULL),
526
527 _count_card_bitmaps(NULL),
528 _count_marked_bytes(NULL),
529 _completed_initialization(false) {
530 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
531 if (verbose_level < no_verbose) {
532 verbose_level = no_verbose;
533 }
534 if (verbose_level > high_verbose) {
535 verbose_level = high_verbose;
536 }
537 _verbose_level = verbose_level;
538
539 if (verbose_low()) {
540 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
541 "heap end = "PTR_FORMAT, _heap_start, _heap_end);
542 }
543
544 if (!_markBitMap1.allocate(heap_rs)) {
545 warning("Failed to allocate first CM bit map");
546 return;
547 }
548 if (!_markBitMap2.allocate(heap_rs)) {
549 warning("Failed to allocate second CM bit map");
550 return;
551 }
552
553 // Create & start a ConcurrentMark thread.
554 _cmThread = new ConcurrentMarkThread(this);
555 assert(cmThread() != NULL, "CM Thread should have been created");
556 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
557
558 assert(CGC_lock != NULL, "Where's the CGC_lock?");
559 assert(_markBitMap1.covers(heap_rs), "_markBitMap1 inconsistency");
560 assert(_markBitMap2.covers(heap_rs), "_markBitMap2 inconsistency");
561
562 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
563 satb_qs.set_buffer_size(G1SATBBufferSize);
564
565 _root_regions.init(_g1h, this);
566
567 if (ConcGCThreads > ParallelGCThreads) {
568 warning("Can't have more ConcGCThreads (" UINT32_FORMAT ") "
569 "than ParallelGCThreads (" UINT32_FORMAT ").",
570 ConcGCThreads, ParallelGCThreads);
571 return;
572 }
573 if (ParallelGCThreads == 0) {
574 // if we are not running with any parallel GC threads we will not
575 // spawn any marking threads either
576 _parallel_marking_threads = 0;
577 _max_parallel_marking_threads = 0;
578 _sleep_factor = 0.0;
579 _marking_task_overhead = 1.0;
580 } else {
581 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
582 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
583 // if both are set
584 _sleep_factor = 0.0;
585 _marking_task_overhead = 1.0;
586 } else if (G1MarkingOverheadPercent > 0) {
587 // We will calculate the number of parallel marking threads based
588 // on a target overhead with respect to the soft real-time goal
589 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
590 double overall_cm_overhead =
591 (double) MaxGCPauseMillis * marking_overhead /
592 (double) GCPauseIntervalMillis;
593 double cpu_ratio = 1.0 / (double) os::processor_count();
594 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
595 double marking_task_overhead =
596 overall_cm_overhead / marking_thread_num *
597 (double) os::processor_count();
598 double sleep_factor =
599 (1.0 - marking_task_overhead) / marking_task_overhead;
600
601 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
602 _sleep_factor = sleep_factor;
603 _marking_task_overhead = marking_task_overhead;
604 } else {
605 // Calculate the number of parallel marking threads by scaling
606 // the number of parallel GC threads.
607 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
608 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
609 _sleep_factor = 0.0;
610 _marking_task_overhead = 1.0;
611 }
612
613 assert(ConcGCThreads > 0, "Should have been set");
614 _parallel_marking_threads = (uint) ConcGCThreads;
615 _max_parallel_marking_threads = _parallel_marking_threads;
616
617 if (parallel_marking_threads() > 1) {
618 _cleanup_task_overhead = 1.0;
619 } else {
620 _cleanup_task_overhead = marking_task_overhead();
621 }
622 _cleanup_sleep_factor =
623 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
624
625 #if 0
626 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
627 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
628 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
629 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
630 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
631 #endif
632
633 guarantee(parallel_marking_threads() > 0, "peace of mind");
634 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
635 _max_parallel_marking_threads, false, true);
636 if (_parallel_workers == NULL) {
637 vm_exit_during_initialization("Failed necessary allocation.");
638 } else {
639 _parallel_workers->initialize_workers();
640 }
641 }
642
643 if (FLAG_IS_DEFAULT(MarkStackSize)) {
644 uintx mark_stack_size =
645 MIN2(MarkStackSizeMax,
646 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
647 // Verify that the calculated value for MarkStackSize is in range.
648 // It would be nice to use the private utility routine from Arguments.
649 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
650 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
651 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
652 mark_stack_size, 1, MarkStackSizeMax);
653 return;
654 }
655 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
656 } else {
657 // Verify MarkStackSize is in range.
658 if (FLAG_IS_CMDLINE(MarkStackSize)) {
659 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
660 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
661 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
662 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
663 MarkStackSize, 1, MarkStackSizeMax);
664 return;
665 }
666 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
667 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
668 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
669 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
670 MarkStackSize, MarkStackSizeMax);
671 return;
672 }
673 }
674 }
675 }
676
677 if (!_markStack.allocate(MarkStackSize)) {
678 warning("Failed to allocate CM marking stack");
679 return;
680 }
681
682 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
683 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
684
685 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
686 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
687
688 BitMap::idx_t card_bm_size = _card_bm.size();
689
690 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
691 _active_tasks = _max_worker_id;
692
693 size_t max_regions = (size_t) _g1h->max_regions();
694 for (uint i = 0; i < _max_worker_id; ++i) {
695 CMTaskQueue* task_queue = new CMTaskQueue();
696 task_queue->initialize();
697 _task_queues->register_queue(i, task_queue);
698
699 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
700 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
701
702 _tasks[i] = new CMTask(i, this,
703 _count_marked_bytes[i],
704 &_count_card_bitmaps[i],
705 task_queue, _task_queues);
706
707 _accum_task_vtime[i] = 0.0;
708 }
709
710 // Calculate the card number for the bottom of the heap. Used
711 // in biasing indexes into the accounting card bitmaps.
712 _heap_bottom_card_num =
713 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
714 CardTableModRefBS::card_shift);
715
716 // Clear all the liveness counting data
717 clear_all_count_data();
718
719 // so that the call below can read a sensible value
720 _heap_start = (HeapWord*) heap_rs.base();
721 set_non_marking_state();
722 _completed_initialization = true;
723 }
724
725 void ConcurrentMark::update_g1_committed(bool force) {
726 // If concurrent marking is not in progress, then we do not need to
727 // update _heap_end.
728 if (!concurrent_marking_in_progress() && !force) return;
729
730 MemRegion committed = _g1h->g1_committed();
731 assert(committed.start() == _heap_start, "start shouldn't change");
732 HeapWord* new_end = committed.end();
733 if (new_end > _heap_end) {
734 // The heap has been expanded.
735
736 _heap_end = new_end;
737 }
738 // Notice that the heap can also shrink. However, this only happens
739 // during a Full GC (at least currently) and the entire marking
740 // phase will bail out and the task will not be restarted. So, let's
741 // do nothing.
742 }
743
744 void ConcurrentMark::reset() {
745 // Starting values for these two. This should be called in a STW
746 // phase. CM will be notified of any future g1_committed expansions
747 // will be at the end of evacuation pauses, when tasks are
748 // inactive.
749 MemRegion committed = _g1h->g1_committed();
750 _heap_start = committed.start();
751 _heap_end = committed.end();
752
753 // Separated the asserts so that we know which one fires.
754 assert(_heap_start != NULL, "heap bounds should look ok");
755 assert(_heap_end != NULL, "heap bounds should look ok");
756 assert(_heap_start < _heap_end, "heap bounds should look ok");
757
758 // Reset all the marking data structures and any necessary flags
759 reset_marking_state();
760
761 if (verbose_low()) {
762 gclog_or_tty->print_cr("[global] resetting");
763 }
764
765 // We do reset all of them, since different phases will use
766 // different number of active threads. So, it's easiest to have all
767 // of them ready.
768 for (uint i = 0; i < _max_worker_id; ++i) {
769 _tasks[i]->reset(_nextMarkBitMap);
770 }
771
772 // we need this to make sure that the flag is on during the evac
773 // pause with initial mark piggy-backed
774 set_concurrent_marking_in_progress();
775 }
776
777
778 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
779 _markStack.set_should_expand();
780 _markStack.setEmpty(); // Also clears the _markStack overflow flag
781 if (clear_overflow) {
782 clear_has_overflown();
783 } else {
784 assert(has_overflown(), "pre-condition");
785 }
786 _finger = _heap_start;
787
788 for (uint i = 0; i < _max_worker_id; ++i) {
789 CMTaskQueue* queue = _task_queues->queue(i);
790 queue->set_empty();
791 }
792 }
793
794 void ConcurrentMark::set_concurrency(uint active_tasks) {
795 assert(active_tasks <= _max_worker_id, "we should not have more");
796
797 _active_tasks = active_tasks;
798 // Need to update the three data structures below according to the
799 // number of active threads for this phase.
800 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
801 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
802 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
803 }
804
805 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
806 set_concurrency(active_tasks);
807
808 _concurrent = concurrent;
809 // We propagate this to all tasks, not just the active ones.
810 for (uint i = 0; i < _max_worker_id; ++i)
811 _tasks[i]->set_concurrent(concurrent);
812
813 if (concurrent) {
814 set_concurrent_marking_in_progress();
815 } else {
816 // We currently assume that the concurrent flag has been set to
817 // false before we start remark. At this point we should also be
818 // in a STW phase.
819 assert(!concurrent_marking_in_progress(), "invariant");
820 assert(_finger == _heap_end,
821 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
822 _finger, _heap_end));
823 update_g1_committed(true);
824 }
825 }
826
827 void ConcurrentMark::set_non_marking_state() {
828 // We set the global marking state to some default values when we're
829 // not doing marking.
830 reset_marking_state();
831 _active_tasks = 0;
832 clear_concurrent_marking_in_progress();
833 }
834
835 ConcurrentMark::~ConcurrentMark() {
836 // The ConcurrentMark instance is never freed.
837 ShouldNotReachHere();
838 }
839
840 void ConcurrentMark::clearNextBitmap() {
841 G1CollectedHeap* g1h = G1CollectedHeap::heap();
842 G1CollectorPolicy* g1p = g1h->g1_policy();
843
844 // Make sure that the concurrent mark thread looks to still be in
845 // the current cycle.
846 guarantee(cmThread()->during_cycle(), "invariant");
847
848 // We are finishing up the current cycle by clearing the next
849 // marking bitmap and getting it ready for the next cycle. During
850 // this time no other cycle can start. So, let's make sure that this
851 // is the case.
852 guarantee(!g1h->mark_in_progress(), "invariant");
853
854 // clear the mark bitmap (no grey objects to start with).
855 // We need to do this in chunks and offer to yield in between
856 // each chunk.
857 HeapWord* start = _nextMarkBitMap->startWord();
858 HeapWord* end = _nextMarkBitMap->endWord();
859 HeapWord* cur = start;
860 size_t chunkSize = M;
861 while (cur < end) {
862 HeapWord* next = cur + chunkSize;
863 if (next > end) {
864 next = end;
865 }
866 MemRegion mr(cur,next);
867 _nextMarkBitMap->clearRange(mr);
868 cur = next;
869 do_yield_check();
870
871 // Repeat the asserts from above. We'll do them as asserts here to
872 // minimize their overhead on the product. However, we'll have
873 // them as guarantees at the beginning / end of the bitmap
874 // clearing to get some checking in the product.
875 assert(cmThread()->during_cycle(), "invariant");
876 assert(!g1h->mark_in_progress(), "invariant");
877 }
878
879 // Clear the liveness counting data
880 clear_all_count_data();
881
882 // Repeat the asserts from above.
883 guarantee(cmThread()->during_cycle(), "invariant");
884 guarantee(!g1h->mark_in_progress(), "invariant");
885 }
886
887 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
888 public:
889 bool doHeapRegion(HeapRegion* r) {
890 if (!r->continuesHumongous()) {
891 r->note_start_of_marking();
892 }
893 return false;
894 }
895 };
896
897 void ConcurrentMark::checkpointRootsInitialPre() {
898 G1CollectedHeap* g1h = G1CollectedHeap::heap();
899 G1CollectorPolicy* g1p = g1h->g1_policy();
900
901 _has_aborted = false;
902
903 #ifndef PRODUCT
904 if (G1PrintReachableAtInitialMark) {
905 print_reachable("at-cycle-start",
906 VerifyOption_G1UsePrevMarking, true /* all */);
907 }
908 #endif
909
910 // Initialise marking structures. This has to be done in a STW phase.
911 reset();
912
913 // For each region note start of marking.
914 NoteStartOfMarkHRClosure startcl;
915 g1h->heap_region_iterate(&startcl);
916 }
917
918
919 void ConcurrentMark::checkpointRootsInitialPost() {
920 G1CollectedHeap* g1h = G1CollectedHeap::heap();
921
922 // If we force an overflow during remark, the remark operation will
923 // actually abort and we'll restart concurrent marking. If we always
924 // force an oveflow during remark we'll never actually complete the
925 // marking phase. So, we initilize this here, at the start of the
926 // cycle, so that at the remaining overflow number will decrease at
927 // every remark and we'll eventually not need to cause one.
928 force_overflow_stw()->init();
929
930 // Start Concurrent Marking weak-reference discovery.
931 ReferenceProcessor* rp = g1h->ref_processor_cm();
932 // enable ("weak") refs discovery
933 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
934 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
935
936 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
937 // This is the start of the marking cycle, we're expected all
938 // threads to have SATB queues with active set to false.
939 satb_mq_set.set_active_all_threads(true, /* new active value */
940 false /* expected_active */);
941
942 _root_regions.prepare_for_scan();
943
944 // update_g1_committed() will be called at the end of an evac pause
945 // when marking is on. So, it's also called at the end of the
946 // initial-mark pause to update the heap end, if the heap expands
947 // during it. No need to call it here.
948 }
949
950 /*
951 * Notice that in the next two methods, we actually leave the STS
952 * during the barrier sync and join it immediately afterwards. If we
953 * do not do this, the following deadlock can occur: one thread could
954 * be in the barrier sync code, waiting for the other thread to also
955 * sync up, whereas another one could be trying to yield, while also
956 * waiting for the other threads to sync up too.
957 *
958 * Note, however, that this code is also used during remark and in
959 * this case we should not attempt to leave / enter the STS, otherwise
960 * we'll either hit an asseert (debug / fastdebug) or deadlock
961 * (product). So we should only leave / enter the STS if we are
962 * operating concurrently.
963 *
964 * Because the thread that does the sync barrier has left the STS, it
965 * is possible to be suspended for a Full GC or an evacuation pause
966 * could occur. This is actually safe, since the entering the sync
967 * barrier is one of the last things do_marking_step() does, and it
968 * doesn't manipulate any data structures afterwards.
969 */
970
971 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
972 if (verbose_low()) {
973 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
974 }
975
976 if (concurrent()) {
977 ConcurrentGCThread::stsLeave();
978 }
979 _first_overflow_barrier_sync.enter();
980 if (concurrent()) {
981 ConcurrentGCThread::stsJoin();
982 }
983 // at this point everyone should have synced up and not be doing any
984 // more work
985
986 if (verbose_low()) {
987 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
988 }
989
990 // If we're executing the concurrent phase of marking, reset the marking
991 // state; otherwise the marking state is reset after reference processing,
992 // during the remark pause.
993 // If we reset here as a result of an overflow during the remark we will
994 // see assertion failures from any subsequent set_concurrency_and_phase()
995 // calls.
996 if (concurrent()) {
997 // let the task associated with with worker 0 do this
998 if (worker_id == 0) {
999 // task 0 is responsible for clearing the global data structures
1000 // We should be here because of an overflow. During STW we should
1001 // not clear the overflow flag since we rely on it being true when
1002 // we exit this method to abort the pause and restart concurent
1003 // marking.
1004 reset_marking_state(true /* clear_overflow */);
1005 force_overflow()->update();
1006
1007 if (G1Log::fine()) {
1008 gclog_or_tty->date_stamp(PrintGCDateStamps);
1009 gclog_or_tty->stamp(PrintGCTimeStamps);
1010 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1011 }
1012 }
1013 }
1014
1015 // after this, each task should reset its own data structures then
1016 // then go into the second barrier
1017 }
1018
1019 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1020 if (verbose_low()) {
1021 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1022 }
1023
1024 if (concurrent()) {
1025 ConcurrentGCThread::stsLeave();
1026 }
1027 _second_overflow_barrier_sync.enter();
1028 if (concurrent()) {
1029 ConcurrentGCThread::stsJoin();
1030 }
1031 // at this point everything should be re-initialized and ready to go
1032
1033 if (verbose_low()) {
1034 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1035 }
1036 }
1037
1038 #ifndef PRODUCT
1039 void ForceOverflowSettings::init() {
1040 _num_remaining = G1ConcMarkForceOverflow;
1041 _force = false;
1042 update();
1043 }
1044
1045 void ForceOverflowSettings::update() {
1046 if (_num_remaining > 0) {
1047 _num_remaining -= 1;
1048 _force = true;
1049 } else {
1050 _force = false;
1051 }
1052 }
1053
1054 bool ForceOverflowSettings::should_force() {
1055 if (_force) {
1056 _force = false;
1057 return true;
1058 } else {
1059 return false;
1060 }
1061 }
1062 #endif // !PRODUCT
1063
1064 class CMConcurrentMarkingTask: public AbstractGangTask {
1065 private:
1066 ConcurrentMark* _cm;
1067 ConcurrentMarkThread* _cmt;
1068
1069 public:
1070 void work(uint worker_id) {
1071 assert(Thread::current()->is_ConcurrentGC_thread(),
1072 "this should only be done by a conc GC thread");
1073 ResourceMark rm;
1074
1075 double start_vtime = os::elapsedVTime();
1076
1077 ConcurrentGCThread::stsJoin();
1078
1079 assert(worker_id < _cm->active_tasks(), "invariant");
1080 CMTask* the_task = _cm->task(worker_id);
1081 the_task->record_start_time();
1082 if (!_cm->has_aborted()) {
1083 do {
1084 double start_vtime_sec = os::elapsedVTime();
1085 double start_time_sec = os::elapsedTime();
1086 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1087
1088 the_task->do_marking_step(mark_step_duration_ms,
1089 true /* do_termination */,
1090 false /* is_serial*/);
1091
1092 double end_time_sec = os::elapsedTime();
1093 double end_vtime_sec = os::elapsedVTime();
1094 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1095 double elapsed_time_sec = end_time_sec - start_time_sec;
1096 _cm->clear_has_overflown();
1097
1098 bool ret = _cm->do_yield_check(worker_id);
1099
1100 jlong sleep_time_ms;
1101 if (!_cm->has_aborted() && the_task->has_aborted()) {
1102 sleep_time_ms =
1103 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1104 ConcurrentGCThread::stsLeave();
1105 os::sleep(Thread::current(), sleep_time_ms, false);
1106 ConcurrentGCThread::stsJoin();
1107 }
1108 double end_time2_sec = os::elapsedTime();
1109 double elapsed_time2_sec = end_time2_sec - start_time_sec;
1110
1111 #if 0
1112 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, "
1113 "overhead %1.4lf",
1114 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms,
1115 the_task->conc_overhead(os::elapsedTime()) * 8.0);
1116 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms",
1117 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0);
1118 #endif
1119 } while (!_cm->has_aborted() && the_task->has_aborted());
1120 }
1121 the_task->record_end_time();
1122 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1123
1124 ConcurrentGCThread::stsLeave();
1125
1126 double end_vtime = os::elapsedVTime();
1127 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1128 }
1129
1130 CMConcurrentMarkingTask(ConcurrentMark* cm,
1131 ConcurrentMarkThread* cmt) :
1132 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1133
1134 ~CMConcurrentMarkingTask() { }
1135 };
1136
1137 // Calculates the number of active workers for a concurrent
1138 // phase.
1139 uint ConcurrentMark::calc_parallel_marking_threads() {
1140 if (G1CollectedHeap::use_parallel_gc_threads()) {
1141 uint n_conc_workers = 0;
1142 if (!UseDynamicNumberOfGCThreads ||
1143 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1144 !ForceDynamicNumberOfGCThreads)) {
1145 n_conc_workers = max_parallel_marking_threads();
1146 } else {
1147 n_conc_workers =
1148 AdaptiveSizePolicy::calc_default_active_workers(
1149 max_parallel_marking_threads(),
1150 1, /* Minimum workers */
1151 parallel_marking_threads(),
1152 Threads::number_of_non_daemon_threads());
1153 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1154 // that scaling has already gone into "_max_parallel_marking_threads".
1155 }
1156 assert(n_conc_workers > 0, "Always need at least 1");
1157 return n_conc_workers;
1158 }
1159 // If we are not running with any parallel GC threads we will not
1160 // have spawned any marking threads either. Hence the number of
1161 // concurrent workers should be 0.
1162 return 0;
1163 }
1164
1165 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1166 // Currently, only survivors can be root regions.
1167 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1168 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1169
1170 const uintx interval = PrefetchScanIntervalInBytes;
1171 HeapWord* curr = hr->bottom();
1172 const HeapWord* end = hr->top();
1173 while (curr < end) {
1174 Prefetch::read(curr, interval);
1175 oop obj = oop(curr);
1176 int size = obj->oop_iterate(&cl);
1177 assert(size == obj->size(), "sanity");
1178 curr += size;
1179 }
1180 }
1181
1182 class CMRootRegionScanTask : public AbstractGangTask {
1183 private:
1184 ConcurrentMark* _cm;
1185
1186 public:
1187 CMRootRegionScanTask(ConcurrentMark* cm) :
1188 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1189
1190 void work(uint worker_id) {
1191 assert(Thread::current()->is_ConcurrentGC_thread(),
1192 "this should only be done by a conc GC thread");
1193
1194 CMRootRegions* root_regions = _cm->root_regions();
1195 HeapRegion* hr = root_regions->claim_next();
1196 while (hr != NULL) {
1197 _cm->scanRootRegion(hr, worker_id);
1198 hr = root_regions->claim_next();
1199 }
1200 }
1201 };
1202
1203 void ConcurrentMark::scanRootRegions() {
1204 // scan_in_progress() will have been set to true only if there was
1205 // at least one root region to scan. So, if it's false, we
1206 // should not attempt to do any further work.
1207 if (root_regions()->scan_in_progress()) {
1208 _parallel_marking_threads = calc_parallel_marking_threads();
1209 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1210 "Maximum number of marking threads exceeded");
1211 uint active_workers = MAX2(1U, parallel_marking_threads());
1212
1213 CMRootRegionScanTask task(this);
1214 if (use_parallel_marking_threads()) {
1215 _parallel_workers->set_active_workers((int) active_workers);
1216 _parallel_workers->run_task(&task);
1217 } else {
1218 task.work(0);
1219 }
1220
1221 // It's possible that has_aborted() is true here without actually
1222 // aborting the survivor scan earlier. This is OK as it's
1223 // mainly used for sanity checking.
1224 root_regions()->scan_finished();
1225 }
1226 }
1227
1228 void ConcurrentMark::markFromRoots() {
1229 // we might be tempted to assert that:
1230 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1231 // "inconsistent argument?");
1232 // However that wouldn't be right, because it's possible that
1233 // a safepoint is indeed in progress as a younger generation
1234 // stop-the-world GC happens even as we mark in this generation.
1235
1236 _restart_for_overflow = false;
1237 force_overflow_conc()->init();
1238
1239 // _g1h has _n_par_threads
1240 _parallel_marking_threads = calc_parallel_marking_threads();
1241 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1242 "Maximum number of marking threads exceeded");
1243
1244 uint active_workers = MAX2(1U, parallel_marking_threads());
1245
1246 // Parallel task terminator is set in "set_concurrency_and_phase()"
1247 set_concurrency_and_phase(active_workers, true /* concurrent */);
1248
1249 CMConcurrentMarkingTask markingTask(this, cmThread());
1250 if (use_parallel_marking_threads()) {
1251 _parallel_workers->set_active_workers((int)active_workers);
1252 // Don't set _n_par_threads because it affects MT in proceess_strong_roots()
1253 // and the decisions on that MT processing is made elsewhere.
1254 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1255 _parallel_workers->run_task(&markingTask);
1256 } else {
1257 markingTask.work(0);
1258 }
1259 print_stats();
1260 }
1261
1262 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1263 // world is stopped at this checkpoint
1264 assert(SafepointSynchronize::is_at_safepoint(),
1265 "world should be stopped");
1266
1267 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1268
1269 // If a full collection has happened, we shouldn't do this.
1270 if (has_aborted()) {
1271 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1272 return;
1273 }
1274
1275 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1276
1277 if (VerifyDuringGC) {
1278 HandleMark hm; // handle scope
1279 Universe::heap()->prepare_for_verify();
1280 Universe::verify(VerifyOption_G1UsePrevMarking,
1281 " VerifyDuringGC:(before)");
1282 }
1283
1284 G1CollectorPolicy* g1p = g1h->g1_policy();
1285 g1p->record_concurrent_mark_remark_start();
1286
1287 double start = os::elapsedTime();
1288
1289 checkpointRootsFinalWork();
1290
1291 double mark_work_end = os::elapsedTime();
1292
1293 weakRefsWork(clear_all_soft_refs);
1294
1295 if (has_overflown()) {
1296 // Oops. We overflowed. Restart concurrent marking.
1297 _restart_for_overflow = true;
1298 if (G1TraceMarkStackOverflow) {
1299 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1300 }
1301
1302 // Verify the heap w.r.t. the previous marking bitmap.
1303 if (VerifyDuringGC) {
1304 HandleMark hm; // handle scope
1305 Universe::heap()->prepare_for_verify();
1306 Universe::verify(VerifyOption_G1UsePrevMarking,
1307 " VerifyDuringGC:(overflow)");
1308 }
1309
1310 // Clear the marking state because we will be restarting
1311 // marking due to overflowing the global mark stack.
1312 reset_marking_state();
1313 } else {
1314 // Aggregate the per-task counting data that we have accumulated
1315 // while marking.
1316 aggregate_count_data();
1317
1318 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1319 // We're done with marking.
1320 // This is the end of the marking cycle, we're expected all
1321 // threads to have SATB queues with active set to true.
1322 satb_mq_set.set_active_all_threads(false, /* new active value */
1323 true /* expected_active */);
1324
1325 if (VerifyDuringGC) {
1326 HandleMark hm; // handle scope
1327 Universe::heap()->prepare_for_verify();
1328 Universe::verify(VerifyOption_G1UseNextMarking,
1329 " VerifyDuringGC:(after)");
1330 }
1331 assert(!restart_for_overflow(), "sanity");
1332 // Completely reset the marking state since marking completed
1333 set_non_marking_state();
1334 }
1335
1336 // Expand the marking stack, if we have to and if we can.
1337 if (_markStack.should_expand()) {
1338 _markStack.expand();
1339 }
1340
1341 // Statistics
1342 double now = os::elapsedTime();
1343 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1344 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1345 _remark_times.add((now - start) * 1000.0);
1346
1347 g1p->record_concurrent_mark_remark_end();
1348
1349 G1CMIsAliveClosure is_alive(g1h);
1350 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1351 }
1352
1353 // Base class of the closures that finalize and verify the
1354 // liveness counting data.
1355 class CMCountDataClosureBase: public HeapRegionClosure {
1356 protected:
1357 G1CollectedHeap* _g1h;
1358 ConcurrentMark* _cm;
1359 CardTableModRefBS* _ct_bs;
1360
1361 BitMap* _region_bm;
1362 BitMap* _card_bm;
1363
1364 // Takes a region that's not empty (i.e., it has at least one
1365 // live object in it and sets its corresponding bit on the region
1366 // bitmap to 1. If the region is "starts humongous" it will also set
1367 // to 1 the bits on the region bitmap that correspond to its
1368 // associated "continues humongous" regions.
1369 void set_bit_for_region(HeapRegion* hr) {
1370 assert(!hr->continuesHumongous(), "should have filtered those out");
1371
1372 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1373 if (!hr->startsHumongous()) {
1374 // Normal (non-humongous) case: just set the bit.
1375 _region_bm->par_at_put(index, true);
1376 } else {
1377 // Starts humongous case: calculate how many regions are part of
1378 // this humongous region and then set the bit range.
1379 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1380 _region_bm->par_at_put_range(index, end_index, true);
1381 }
1382 }
1383
1384 public:
1385 CMCountDataClosureBase(G1CollectedHeap* g1h,
1386 BitMap* region_bm, BitMap* card_bm):
1387 _g1h(g1h), _cm(g1h->concurrent_mark()),
1388 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1389 _region_bm(region_bm), _card_bm(card_bm) { }
1390 };
1391
1392 // Closure that calculates the # live objects per region. Used
1393 // for verification purposes during the cleanup pause.
1394 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1395 CMBitMapRO* _bm;
1396 size_t _region_marked_bytes;
1397
1398 public:
1399 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1400 BitMap* region_bm, BitMap* card_bm) :
1401 CMCountDataClosureBase(g1h, region_bm, card_bm),
1402 _bm(bm), _region_marked_bytes(0) { }
1403
1404 bool doHeapRegion(HeapRegion* hr) {
1405
1406 if (hr->continuesHumongous()) {
1407 // We will ignore these here and process them when their
1408 // associated "starts humongous" region is processed (see
1409 // set_bit_for_heap_region()). Note that we cannot rely on their
1410 // associated "starts humongous" region to have their bit set to
1411 // 1 since, due to the region chunking in the parallel region
1412 // iteration, a "continues humongous" region might be visited
1413 // before its associated "starts humongous".
1414 return false;
1415 }
1416
1417 HeapWord* ntams = hr->next_top_at_mark_start();
1418 HeapWord* start = hr->bottom();
1419
1420 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1421 err_msg("Preconditions not met - "
1422 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1423 start, ntams, hr->end()));
1424
1425 // Find the first marked object at or after "start".
1426 start = _bm->getNextMarkedWordAddress(start, ntams);
1427
1428 size_t marked_bytes = 0;
1429
1430 while (start < ntams) {
1431 oop obj = oop(start);
1432 int obj_sz = obj->size();
1433 HeapWord* obj_end = start + obj_sz;
1434
1435 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1436 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1437
1438 // Note: if we're looking at the last region in heap - obj_end
1439 // could be actually just beyond the end of the heap; end_idx
1440 // will then correspond to a (non-existent) card that is also
1441 // just beyond the heap.
1442 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1443 // end of object is not card aligned - increment to cover
1444 // all the cards spanned by the object
1445 end_idx += 1;
1446 }
1447
1448 // Set the bits in the card BM for the cards spanned by this object.
1449 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1450
1451 // Add the size of this object to the number of marked bytes.
1452 marked_bytes += (size_t)obj_sz * HeapWordSize;
1453
1454 // Find the next marked object after this one.
1455 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1456 }
1457
1458 // Mark the allocated-since-marking portion...
1459 HeapWord* top = hr->top();
1460 if (ntams < top) {
1461 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1462 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1463
1464 // Note: if we're looking at the last region in heap - top
1465 // could be actually just beyond the end of the heap; end_idx
1466 // will then correspond to a (non-existent) card that is also
1467 // just beyond the heap.
1468 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1469 // end of object is not card aligned - increment to cover
1470 // all the cards spanned by the object
1471 end_idx += 1;
1472 }
1473 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1474
1475 // This definitely means the region has live objects.
1476 set_bit_for_region(hr);
1477 }
1478
1479 // Update the live region bitmap.
1480 if (marked_bytes > 0) {
1481 set_bit_for_region(hr);
1482 }
1483
1484 // Set the marked bytes for the current region so that
1485 // it can be queried by a calling verificiation routine
1486 _region_marked_bytes = marked_bytes;
1487
1488 return false;
1489 }
1490
1491 size_t region_marked_bytes() const { return _region_marked_bytes; }
1492 };
1493
1494 // Heap region closure used for verifying the counting data
1495 // that was accumulated concurrently and aggregated during
1496 // the remark pause. This closure is applied to the heap
1497 // regions during the STW cleanup pause.
1498
1499 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1500 G1CollectedHeap* _g1h;
1501 ConcurrentMark* _cm;
1502 CalcLiveObjectsClosure _calc_cl;
1503 BitMap* _region_bm; // Region BM to be verified
1504 BitMap* _card_bm; // Card BM to be verified
1505 bool _verbose; // verbose output?
1506
1507 BitMap* _exp_region_bm; // Expected Region BM values
1508 BitMap* _exp_card_bm; // Expected card BM values
1509
1510 int _failures;
1511
1512 public:
1513 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1514 BitMap* region_bm,
1515 BitMap* card_bm,
1516 BitMap* exp_region_bm,
1517 BitMap* exp_card_bm,
1518 bool verbose) :
1519 _g1h(g1h), _cm(g1h->concurrent_mark()),
1520 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1521 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1522 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1523 _failures(0) { }
1524
1525 int failures() const { return _failures; }
1526
1527 bool doHeapRegion(HeapRegion* hr) {
1528 if (hr->continuesHumongous()) {
1529 // We will ignore these here and process them when their
1530 // associated "starts humongous" region is processed (see
1531 // set_bit_for_heap_region()). Note that we cannot rely on their
1532 // associated "starts humongous" region to have their bit set to
1533 // 1 since, due to the region chunking in the parallel region
1534 // iteration, a "continues humongous" region might be visited
1535 // before its associated "starts humongous".
1536 return false;
1537 }
1538
1539 int failures = 0;
1540
1541 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1542 // this region and set the corresponding bits in the expected region
1543 // and card bitmaps.
1544 bool res = _calc_cl.doHeapRegion(hr);
1545 assert(res == false, "should be continuing");
1546
1547 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1548 Mutex::_no_safepoint_check_flag);
1549
1550 // Verify the marked bytes for this region.
1551 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1552 size_t act_marked_bytes = hr->next_marked_bytes();
1553
1554 // We're not OK if expected marked bytes > actual marked bytes. It means
1555 // we have missed accounting some objects during the actual marking.
1556 if (exp_marked_bytes > act_marked_bytes) {
1557 if (_verbose) {
1558 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1559 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1560 hr->hrs_index(), exp_marked_bytes, act_marked_bytes);
1561 }
1562 failures += 1;
1563 }
1564
1565 // Verify the bit, for this region, in the actual and expected
1566 // (which was just calculated) region bit maps.
1567 // We're not OK if the bit in the calculated expected region
1568 // bitmap is set and the bit in the actual region bitmap is not.
1569 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1570
1571 bool expected = _exp_region_bm->at(index);
1572 bool actual = _region_bm->at(index);
1573 if (expected && !actual) {
1574 if (_verbose) {
1575 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1576 "expected: %s, actual: %s",
1577 hr->hrs_index(),
1578 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1579 }
1580 failures += 1;
1581 }
1582
1583 // Verify that the card bit maps for the cards spanned by the current
1584 // region match. We have an error if we have a set bit in the expected
1585 // bit map and the corresponding bit in the actual bitmap is not set.
1586
1587 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1588 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1589
1590 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1591 expected = _exp_card_bm->at(i);
1592 actual = _card_bm->at(i);
1593
1594 if (expected && !actual) {
1595 if (_verbose) {
1596 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1597 "expected: %s, actual: %s",
1598 hr->hrs_index(), i,
1599 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1600 }
1601 failures += 1;
1602 }
1603 }
1604
1605 if (failures > 0 && _verbose) {
1606 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1607 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1608 HR_FORMAT_PARAMS(hr), hr->next_top_at_mark_start(),
1609 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1610 }
1611
1612 _failures += failures;
1613
1614 // We could stop iteration over the heap when we
1615 // find the first violating region by returning true.
1616 return false;
1617 }
1618 };
1619
1620
1621 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1622 protected:
1623 G1CollectedHeap* _g1h;
1624 ConcurrentMark* _cm;
1625 BitMap* _actual_region_bm;
1626 BitMap* _actual_card_bm;
1627
1628 uint _n_workers;
1629
1630 BitMap* _expected_region_bm;
1631 BitMap* _expected_card_bm;
1632
1633 int _failures;
1634 bool _verbose;
1635
1636 public:
1637 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1638 BitMap* region_bm, BitMap* card_bm,
1639 BitMap* expected_region_bm, BitMap* expected_card_bm)
1640 : AbstractGangTask("G1 verify final counting"),
1641 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1642 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1643 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1644 _failures(0), _verbose(false),
1645 _n_workers(0) {
1646 assert(VerifyDuringGC, "don't call this otherwise");
1647
1648 // Use the value already set as the number of active threads
1649 // in the call to run_task().
1650 if (G1CollectedHeap::use_parallel_gc_threads()) {
1651 assert( _g1h->workers()->active_workers() > 0,
1652 "Should have been previously set");
1653 _n_workers = _g1h->workers()->active_workers();
1654 } else {
1655 _n_workers = 1;
1656 }
1657
1658 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1659 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1660
1661 _verbose = _cm->verbose_medium();
1662 }
1663
1664 void work(uint worker_id) {
1665 assert(worker_id < _n_workers, "invariant");
1666
1667 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1668 _actual_region_bm, _actual_card_bm,
1669 _expected_region_bm,
1670 _expected_card_bm,
1671 _verbose);
1672
1673 if (G1CollectedHeap::use_parallel_gc_threads()) {
1674 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1675 worker_id,
1676 _n_workers,
1677 HeapRegion::VerifyCountClaimValue);
1678 } else {
1679 _g1h->heap_region_iterate(&verify_cl);
1680 }
1681
1682 Atomic::add(verify_cl.failures(), &_failures);
1683 }
1684
1685 int failures() const { return _failures; }
1686 };
1687
1688 // Closure that finalizes the liveness counting data.
1689 // Used during the cleanup pause.
1690 // Sets the bits corresponding to the interval [NTAMS, top]
1691 // (which contains the implicitly live objects) in the
1692 // card liveness bitmap. Also sets the bit for each region,
1693 // containing live data, in the region liveness bitmap.
1694
1695 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1696 public:
1697 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1698 BitMap* region_bm,
1699 BitMap* card_bm) :
1700 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1701
1702 bool doHeapRegion(HeapRegion* hr) {
1703
1704 if (hr->continuesHumongous()) {
1705 // We will ignore these here and process them when their
1706 // associated "starts humongous" region is processed (see
1707 // set_bit_for_heap_region()). Note that we cannot rely on their
1708 // associated "starts humongous" region to have their bit set to
1709 // 1 since, due to the region chunking in the parallel region
1710 // iteration, a "continues humongous" region might be visited
1711 // before its associated "starts humongous".
1712 return false;
1713 }
1714
1715 HeapWord* ntams = hr->next_top_at_mark_start();
1716 HeapWord* top = hr->top();
1717
1718 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1719
1720 // Mark the allocated-since-marking portion...
1721 if (ntams < top) {
1722 // This definitely means the region has live objects.
1723 set_bit_for_region(hr);
1724
1725 // Now set the bits in the card bitmap for [ntams, top)
1726 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1727 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1728
1729 // Note: if we're looking at the last region in heap - top
1730 // could be actually just beyond the end of the heap; end_idx
1731 // will then correspond to a (non-existent) card that is also
1732 // just beyond the heap.
1733 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1734 // end of object is not card aligned - increment to cover
1735 // all the cards spanned by the object
1736 end_idx += 1;
1737 }
1738
1739 assert(end_idx <= _card_bm->size(),
1740 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1741 end_idx, _card_bm->size()));
1742 assert(start_idx < _card_bm->size(),
1743 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1744 start_idx, _card_bm->size()));
1745
1746 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1747 }
1748
1749 // Set the bit for the region if it contains live data
1750 if (hr->next_marked_bytes() > 0) {
1751 set_bit_for_region(hr);
1752 }
1753
1754 return false;
1755 }
1756 };
1757
1758 class G1ParFinalCountTask: public AbstractGangTask {
1759 protected:
1760 G1CollectedHeap* _g1h;
1761 ConcurrentMark* _cm;
1762 BitMap* _actual_region_bm;
1763 BitMap* _actual_card_bm;
1764
1765 uint _n_workers;
1766
1767 public:
1768 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1769 : AbstractGangTask("G1 final counting"),
1770 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1771 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1772 _n_workers(0) {
1773 // Use the value already set as the number of active threads
1774 // in the call to run_task().
1775 if (G1CollectedHeap::use_parallel_gc_threads()) {
1776 assert( _g1h->workers()->active_workers() > 0,
1777 "Should have been previously set");
1778 _n_workers = _g1h->workers()->active_workers();
1779 } else {
1780 _n_workers = 1;
1781 }
1782 }
1783
1784 void work(uint worker_id) {
1785 assert(worker_id < _n_workers, "invariant");
1786
1787 FinalCountDataUpdateClosure final_update_cl(_g1h,
1788 _actual_region_bm,
1789 _actual_card_bm);
1790
1791 if (G1CollectedHeap::use_parallel_gc_threads()) {
1792 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1793 worker_id,
1794 _n_workers,
1795 HeapRegion::FinalCountClaimValue);
1796 } else {
1797 _g1h->heap_region_iterate(&final_update_cl);
1798 }
1799 }
1800 };
1801
1802 class G1ParNoteEndTask;
1803
1804 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1805 G1CollectedHeap* _g1;
1806 int _worker_num;
1807 size_t _max_live_bytes;
1808 uint _regions_claimed;
1809 size_t _freed_bytes;
1810 FreeRegionList* _local_cleanup_list;
1811 OldRegionSet* _old_proxy_set;
1812 HumongousRegionSet* _humongous_proxy_set;
1813 HRRSCleanupTask* _hrrs_cleanup_task;
1814 double _claimed_region_time;
1815 double _max_region_time;
1816
1817 public:
1818 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1819 int worker_num,
1820 FreeRegionList* local_cleanup_list,
1821 OldRegionSet* old_proxy_set,
1822 HumongousRegionSet* humongous_proxy_set,
1823 HRRSCleanupTask* hrrs_cleanup_task) :
1824 _g1(g1), _worker_num(worker_num),
1825 _max_live_bytes(0), _regions_claimed(0),
1826 _freed_bytes(0),
1827 _claimed_region_time(0.0), _max_region_time(0.0),
1828 _local_cleanup_list(local_cleanup_list),
1829 _old_proxy_set(old_proxy_set),
1830 _humongous_proxy_set(humongous_proxy_set),
1831 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1832
1833 size_t freed_bytes() { return _freed_bytes; }
1834
1835 bool doHeapRegion(HeapRegion *hr) {
1836 if (hr->continuesHumongous()) {
1837 return false;
1838 }
1839 // We use a claim value of zero here because all regions
1840 // were claimed with value 1 in the FinalCount task.
1841 _g1->reset_gc_time_stamps(hr);
1842 double start = os::elapsedTime();
1843 _regions_claimed++;
1844 hr->note_end_of_marking();
1845 _max_live_bytes += hr->max_live_bytes();
1846 _g1->free_region_if_empty(hr,
1847 &_freed_bytes,
1848 _local_cleanup_list,
1849 _old_proxy_set,
1850 _humongous_proxy_set,
1851 _hrrs_cleanup_task,
1852 true /* par */);
1853 double region_time = (os::elapsedTime() - start);
1854 _claimed_region_time += region_time;
1855 if (region_time > _max_region_time) {
1856 _max_region_time = region_time;
1857 }
1858 return false;
1859 }
1860
1861 size_t max_live_bytes() { return _max_live_bytes; }
1862 uint regions_claimed() { return _regions_claimed; }
1863 double claimed_region_time_sec() { return _claimed_region_time; }
1864 double max_region_time_sec() { return _max_region_time; }
1865 };
1866
1867 class G1ParNoteEndTask: public AbstractGangTask {
1868 friend class G1NoteEndOfConcMarkClosure;
1869
1870 protected:
1871 G1CollectedHeap* _g1h;
1872 size_t _max_live_bytes;
1873 size_t _freed_bytes;
1874 FreeRegionList* _cleanup_list;
1875
1876 public:
1877 G1ParNoteEndTask(G1CollectedHeap* g1h,
1878 FreeRegionList* cleanup_list) :
1879 AbstractGangTask("G1 note end"), _g1h(g1h),
1880 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1881
1882 void work(uint worker_id) {
1883 double start = os::elapsedTime();
1884 FreeRegionList local_cleanup_list("Local Cleanup List");
1885 OldRegionSet old_proxy_set("Local Cleanup Old Proxy Set");
1886 HumongousRegionSet humongous_proxy_set("Local Cleanup Humongous Proxy Set");
1887 HRRSCleanupTask hrrs_cleanup_task;
1888 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, worker_id, &local_cleanup_list,
1889 &old_proxy_set,
1890 &humongous_proxy_set,
1891 &hrrs_cleanup_task);
1892 if (G1CollectedHeap::use_parallel_gc_threads()) {
1893 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1894 _g1h->workers()->active_workers(),
1895 HeapRegion::NoteEndClaimValue);
1896 } else {
1897 _g1h->heap_region_iterate(&g1_note_end);
1898 }
1899 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1900
1901 // Now update the lists
1902 _g1h->update_sets_after_freeing_regions(g1_note_end.freed_bytes(),
1903 NULL /* free_list */,
1904 &old_proxy_set,
1905 &humongous_proxy_set,
1906 true /* par */);
1907 {
1908 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1909 _max_live_bytes += g1_note_end.max_live_bytes();
1910 _freed_bytes += g1_note_end.freed_bytes();
1911
1912 // If we iterate over the global cleanup list at the end of
1913 // cleanup to do this printing we will not guarantee to only
1914 // generate output for the newly-reclaimed regions (the list
1915 // might not be empty at the beginning of cleanup; we might
1916 // still be working on its previous contents). So we do the
1917 // printing here, before we append the new regions to the global
1918 // cleanup list.
1919
1920 G1HRPrinter* hr_printer = _g1h->hr_printer();
1921 if (hr_printer->is_active()) {
1922 HeapRegionLinkedListIterator iter(&local_cleanup_list);
1923 while (iter.more_available()) {
1924 HeapRegion* hr = iter.get_next();
1925 hr_printer->cleanup(hr);
1926 }
1927 }
1928
1929 _cleanup_list->add_as_tail(&local_cleanup_list);
1930 assert(local_cleanup_list.is_empty(), "post-condition");
1931
1932 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1933 }
1934 }
1935 size_t max_live_bytes() { return _max_live_bytes; }
1936 size_t freed_bytes() { return _freed_bytes; }
1937 };
1938
1939 class G1ParScrubRemSetTask: public AbstractGangTask {
1940 protected:
1941 G1RemSet* _g1rs;
1942 BitMap* _region_bm;
1943 BitMap* _card_bm;
1944 public:
1945 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1946 BitMap* region_bm, BitMap* card_bm) :
1947 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1948 _region_bm(region_bm), _card_bm(card_bm) { }
1949
1950 void work(uint worker_id) {
1951 if (G1CollectedHeap::use_parallel_gc_threads()) {
1952 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1953 HeapRegion::ScrubRemSetClaimValue);
1954 } else {
1955 _g1rs->scrub(_region_bm, _card_bm);
1956 }
1957 }
1958
1959 };
1960
1961 void ConcurrentMark::cleanup() {
1962 // world is stopped at this checkpoint
1963 assert(SafepointSynchronize::is_at_safepoint(),
1964 "world should be stopped");
1965 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1966
1967 // If a full collection has happened, we shouldn't do this.
1968 if (has_aborted()) {
1969 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1970 return;
1971 }
1972
1973 HRSPhaseSetter x(HRSPhaseCleanup);
1974 g1h->verify_region_sets_optional();
1975
1976 if (VerifyDuringGC) {
1977 HandleMark hm; // handle scope
1978 Universe::heap()->prepare_for_verify();
1979 Universe::verify(VerifyOption_G1UsePrevMarking,
1980 " VerifyDuringGC:(before)");
1981 }
1982
1983 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
1984 g1p->record_concurrent_mark_cleanup_start();
1985
1986 double start = os::elapsedTime();
1987
1988 HeapRegionRemSet::reset_for_cleanup_tasks();
1989
1990 uint n_workers;
1991
1992 // Do counting once more with the world stopped for good measure.
1993 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1994
1995 if (G1CollectedHeap::use_parallel_gc_threads()) {
1996 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
1997 "sanity check");
1998
1999 g1h->set_par_threads();
2000 n_workers = g1h->n_par_threads();
2001 assert(g1h->n_par_threads() == n_workers,
2002 "Should not have been reset");
2003 g1h->workers()->run_task(&g1_par_count_task);
2004 // Done with the parallel phase so reset to 0.
2005 g1h->set_par_threads(0);
2006
2007 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2008 "sanity check");
2009 } else {
2010 n_workers = 1;
2011 g1_par_count_task.work(0);
2012 }
2013
2014 if (VerifyDuringGC) {
2015 // Verify that the counting data accumulated during marking matches
2016 // that calculated by walking the marking bitmap.
2017
2018 // Bitmaps to hold expected values
2019 BitMap expected_region_bm(_region_bm.size(), false);
2020 BitMap expected_card_bm(_card_bm.size(), false);
2021
2022 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2023 &_region_bm,
2024 &_card_bm,
2025 &expected_region_bm,
2026 &expected_card_bm);
2027
2028 if (G1CollectedHeap::use_parallel_gc_threads()) {
2029 g1h->set_par_threads((int)n_workers);
2030 g1h->workers()->run_task(&g1_par_verify_task);
2031 // Done with the parallel phase so reset to 0.
2032 g1h->set_par_threads(0);
2033
2034 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2035 "sanity check");
2036 } else {
2037 g1_par_verify_task.work(0);
2038 }
2039
2040 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2041 }
2042
2043 size_t start_used_bytes = g1h->used();
2044 g1h->set_marking_complete();
2045
2046 double count_end = os::elapsedTime();
2047 double this_final_counting_time = (count_end - start);
2048 _total_counting_time += this_final_counting_time;
2049
2050 if (G1PrintRegionLivenessInfo) {
2051 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2052 _g1h->heap_region_iterate(&cl);
2053 }
2054
2055 // Install newly created mark bitMap as "prev".
2056 swapMarkBitMaps();
2057
2058 g1h->reset_gc_time_stamp();
2059
2060 // Note end of marking in all heap regions.
2061 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2062 if (G1CollectedHeap::use_parallel_gc_threads()) {
2063 g1h->set_par_threads((int)n_workers);
2064 g1h->workers()->run_task(&g1_par_note_end_task);
2065 g1h->set_par_threads(0);
2066
2067 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2068 "sanity check");
2069 } else {
2070 g1_par_note_end_task.work(0);
2071 }
2072 g1h->check_gc_time_stamps();
2073
2074 if (!cleanup_list_is_empty()) {
2075 // The cleanup list is not empty, so we'll have to process it
2076 // concurrently. Notify anyone else that might be wanting free
2077 // regions that there will be more free regions coming soon.
2078 g1h->set_free_regions_coming();
2079 }
2080
2081 // call below, since it affects the metric by which we sort the heap
2082 // regions.
2083 if (G1ScrubRemSets) {
2084 double rs_scrub_start = os::elapsedTime();
2085 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2086 if (G1CollectedHeap::use_parallel_gc_threads()) {
2087 g1h->set_par_threads((int)n_workers);
2088 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2089 g1h->set_par_threads(0);
2090
2091 assert(g1h->check_heap_region_claim_values(
2092 HeapRegion::ScrubRemSetClaimValue),
2093 "sanity check");
2094 } else {
2095 g1_par_scrub_rs_task.work(0);
2096 }
2097
2098 double rs_scrub_end = os::elapsedTime();
2099 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2100 _total_rs_scrub_time += this_rs_scrub_time;
2101 }
2102
2103 // this will also free any regions totally full of garbage objects,
2104 // and sort the regions.
2105 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2106
2107 // Statistics.
2108 double end = os::elapsedTime();
2109 _cleanup_times.add((end - start) * 1000.0);
2110
2111 if (G1Log::fine()) {
2112 g1h->print_size_transition(gclog_or_tty,
2113 start_used_bytes,
2114 g1h->used(),
2115 g1h->capacity());
2116 }
2117
2118 // Clean up will have freed any regions completely full of garbage.
2119 // Update the soft reference policy with the new heap occupancy.
2120 Universe::update_heap_info_at_gc();
2121
2122 // We need to make this be a "collection" so any collection pause that
2123 // races with it goes around and waits for completeCleanup to finish.
2124 g1h->increment_total_collections();
2125
2126 // We reclaimed old regions so we should calculate the sizes to make
2127 // sure we update the old gen/space data.
2128 g1h->g1mm()->update_sizes();
2129
2130 if (VerifyDuringGC) {
2131 HandleMark hm; // handle scope
2132 Universe::heap()->prepare_for_verify();
2133 Universe::verify(VerifyOption_G1UsePrevMarking,
2134 " VerifyDuringGC:(after)");
2135 }
2136
2137 g1h->verify_region_sets_optional();
2138 g1h->trace_heap_after_concurrent_cycle();
2139 }
2140
2141 void ConcurrentMark::completeCleanup() {
2142 if (has_aborted()) return;
2143
2144 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2145
2146 _cleanup_list.verify_optional();
2147 FreeRegionList tmp_free_list("Tmp Free List");
2148
2149 if (G1ConcRegionFreeingVerbose) {
2150 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2151 "cleanup list has %u entries",
2152 _cleanup_list.length());
2153 }
2154
2155 // Noone else should be accessing the _cleanup_list at this point,
2156 // so it's not necessary to take any locks
2157 while (!_cleanup_list.is_empty()) {
2158 HeapRegion* hr = _cleanup_list.remove_head();
2159 assert(hr != NULL, "the list was not empty");
2160 hr->par_clear();
2161 tmp_free_list.add_as_tail(hr);
2162
2163 // Instead of adding one region at a time to the secondary_free_list,
2164 // we accumulate them in the local list and move them a few at a
2165 // time. This also cuts down on the number of notify_all() calls
2166 // we do during this process. We'll also append the local list when
2167 // _cleanup_list is empty (which means we just removed the last
2168 // region from the _cleanup_list).
2169 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2170 _cleanup_list.is_empty()) {
2171 if (G1ConcRegionFreeingVerbose) {
2172 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2173 "appending %u entries to the secondary_free_list, "
2174 "cleanup list still has %u entries",
2175 tmp_free_list.length(),
2176 _cleanup_list.length());
2177 }
2178
2179 {
2180 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2181 g1h->secondary_free_list_add_as_tail(&tmp_free_list);
2182 SecondaryFreeList_lock->notify_all();
2183 }
2184
2185 if (G1StressConcRegionFreeing) {
2186 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2187 os::sleep(Thread::current(), (jlong) 1, false);
2188 }
2189 }
2190 }
2191 }
2192 assert(tmp_free_list.is_empty(), "post-condition");
2193 }
2194
2195 // Supporting Object and Oop closures for reference discovery
2196 // and processing in during marking
2197
2198 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2199 HeapWord* addr = (HeapWord*)obj;
2200 return addr != NULL &&
2201 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2202 }
2203
2204 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2205 // Uses the CMTask associated with a worker thread (for serial reference
2206 // processing the CMTask for worker 0 is used) to preserve (mark) and
2207 // trace referent objects.
2208 //
2209 // Using the CMTask and embedded local queues avoids having the worker
2210 // threads operating on the global mark stack. This reduces the risk
2211 // of overflowing the stack - which we would rather avoid at this late
2212 // state. Also using the tasks' local queues removes the potential
2213 // of the workers interfering with each other that could occur if
2214 // operating on the global stack.
2215
2216 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2217 ConcurrentMark* _cm;
2218 CMTask* _task;
2219 int _ref_counter_limit;
2220 int _ref_counter;
2221 bool _is_serial;
2222 public:
2223 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2224 _cm(cm), _task(task), _is_serial(is_serial),
2225 _ref_counter_limit(G1RefProcDrainInterval) {
2226 assert(_ref_counter_limit > 0, "sanity");
2227 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2228 _ref_counter = _ref_counter_limit;
2229 }
2230
2231 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2232 virtual void do_oop( oop* p) { do_oop_work(p); }
2233
2234 template <class T> void do_oop_work(T* p) {
2235 if (!_cm->has_overflown()) {
2236 oop obj = oopDesc::load_decode_heap_oop(p);
2237 if (_cm->verbose_high()) {
2238 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2239 "*"PTR_FORMAT" = "PTR_FORMAT,
2240 _task->worker_id(), p, (void*) obj);
2241 }
2242
2243 _task->deal_with_reference(obj);
2244 _ref_counter--;
2245
2246 if (_ref_counter == 0) {
2247 // We have dealt with _ref_counter_limit references, pushing them
2248 // and objects reachable from them on to the local stack (and
2249 // possibly the global stack). Call CMTask::do_marking_step() to
2250 // process these entries.
2251 //
2252 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2253 // there's nothing more to do (i.e. we're done with the entries that
2254 // were pushed as a result of the CMTask::deal_with_reference() calls
2255 // above) or we overflow.
2256 //
2257 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2258 // flag while there may still be some work to do. (See the comment at
2259 // the beginning of CMTask::do_marking_step() for those conditions -
2260 // one of which is reaching the specified time target.) It is only
2261 // when CMTask::do_marking_step() returns without setting the
2262 // has_aborted() flag that the marking step has completed.
2263 do {
2264 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2265 _task->do_marking_step(mark_step_duration_ms,
2266 false /* do_termination */,
2267 _is_serial);
2268 } while (_task->has_aborted() && !_cm->has_overflown());
2269 _ref_counter = _ref_counter_limit;
2270 }
2271 } else {
2272 if (_cm->verbose_high()) {
2273 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2274 }
2275 }
2276 }
2277 };
2278
2279 // 'Drain' oop closure used by both serial and parallel reference processing.
2280 // Uses the CMTask associated with a given worker thread (for serial
2281 // reference processing the CMtask for worker 0 is used). Calls the
2282 // do_marking_step routine, with an unbelievably large timeout value,
2283 // to drain the marking data structures of the remaining entries
2284 // added by the 'keep alive' oop closure above.
2285
2286 class G1CMDrainMarkingStackClosure: public VoidClosure {
2287 ConcurrentMark* _cm;
2288 CMTask* _task;
2289 bool _is_serial;
2290 public:
2291 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2292 _cm(cm), _task(task), _is_serial(is_serial) {
2293 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2294 }
2295
2296 void do_void() {
2297 do {
2298 if (_cm->verbose_high()) {
2299 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2300 _task->worker_id(), BOOL_TO_STR(_is_serial));
2301 }
2302
2303 // We call CMTask::do_marking_step() to completely drain the local
2304 // and global marking stacks of entries pushed by the 'keep alive'
2305 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2306 //
2307 // CMTask::do_marking_step() is called in a loop, which we'll exit
2308 // if there's nothing more to do (i.e. we'completely drained the
2309 // entries that were pushed as a a result of applying the 'keep alive'
2310 // closure to the entries on the discovered ref lists) or we overflow
2311 // the global marking stack.
2312 //
2313 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2314 // flag while there may still be some work to do. (See the comment at
2315 // the beginning of CMTask::do_marking_step() for those conditions -
2316 // one of which is reaching the specified time target.) It is only
2317 // when CMTask::do_marking_step() returns without setting the
2318 // has_aborted() flag that the marking step has completed.
2319
2320 _task->do_marking_step(1000000000.0 /* something very large */,
2321 true /* do_termination */,
2322 _is_serial);
2323 } while (_task->has_aborted() && !_cm->has_overflown());
2324 }
2325 };
2326
2327 // Implementation of AbstractRefProcTaskExecutor for parallel
2328 // reference processing at the end of G1 concurrent marking
2329
2330 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2331 private:
2332 G1CollectedHeap* _g1h;
2333 ConcurrentMark* _cm;
2334 WorkGang* _workers;
2335 int _active_workers;
2336
2337 public:
2338 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2339 ConcurrentMark* cm,
2340 WorkGang* workers,
2341 int n_workers) :
2342 _g1h(g1h), _cm(cm),
2343 _workers(workers), _active_workers(n_workers) { }
2344
2345 // Executes the given task using concurrent marking worker threads.
2346 virtual void execute(ProcessTask& task);
2347 virtual void execute(EnqueueTask& task);
2348 };
2349
2350 class G1CMRefProcTaskProxy: public AbstractGangTask {
2351 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2352 ProcessTask& _proc_task;
2353 G1CollectedHeap* _g1h;
2354 ConcurrentMark* _cm;
2355
2356 public:
2357 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2358 G1CollectedHeap* g1h,
2359 ConcurrentMark* cm) :
2360 AbstractGangTask("Process reference objects in parallel"),
2361 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2362 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2363 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2364 }
2365
2366 virtual void work(uint worker_id) {
2367 CMTask* task = _cm->task(worker_id);
2368 G1CMIsAliveClosure g1_is_alive(_g1h);
2369 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2370 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2371
2372 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2373 }
2374 };
2375
2376 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2377 assert(_workers != NULL, "Need parallel worker threads.");
2378 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2379
2380 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2381
2382 // We need to reset the concurrency level before each
2383 // proxy task execution, so that the termination protocol
2384 // and overflow handling in CMTask::do_marking_step() knows
2385 // how many workers to wait for.
2386 _cm->set_concurrency(_active_workers);
2387 _g1h->set_par_threads(_active_workers);
2388 _workers->run_task(&proc_task_proxy);
2389 _g1h->set_par_threads(0);
2390 }
2391
2392 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2393 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2394 EnqueueTask& _enq_task;
2395
2396 public:
2397 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2398 AbstractGangTask("Enqueue reference objects in parallel"),
2399 _enq_task(enq_task) { }
2400
2401 virtual void work(uint worker_id) {
2402 _enq_task.work(worker_id);
2403 }
2404 };
2405
2406 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2407 assert(_workers != NULL, "Need parallel worker threads.");
2408 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2409
2410 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2411
2412 // Not strictly necessary but...
2413 //
2414 // We need to reset the concurrency level before each
2415 // proxy task execution, so that the termination protocol
2416 // and overflow handling in CMTask::do_marking_step() knows
2417 // how many workers to wait for.
2418 _cm->set_concurrency(_active_workers);
2419 _g1h->set_par_threads(_active_workers);
2420 _workers->run_task(&enq_task_proxy);
2421 _g1h->set_par_threads(0);
2422 }
2423
2424 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2425 if (has_overflown()) {
2426 // Skip processing the discovered references if we have
2427 // overflown the global marking stack. Reference objects
2428 // only get discovered once so it is OK to not
2429 // de-populate the discovered reference lists. We could have,
2430 // but the only benefit would be that, when marking restarts,
2431 // less reference objects are discovered.
2432 return;
2433 }
2434
2435 ResourceMark rm;
2436 HandleMark hm;
2437
2438 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2439
2440 // Is alive closure.
2441 G1CMIsAliveClosure g1_is_alive(g1h);
2442
2443 // Inner scope to exclude the cleaning of the string and symbol
2444 // tables from the displayed time.
2445 {
2446 if (G1Log::finer()) {
2447 gclog_or_tty->put(' ');
2448 }
2449 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm());
2450
2451 ReferenceProcessor* rp = g1h->ref_processor_cm();
2452
2453 // See the comment in G1CollectedHeap::ref_processing_init()
2454 // about how reference processing currently works in G1.
2455
2456 // Set the soft reference policy
2457 rp->setup_policy(clear_all_soft_refs);
2458 assert(_markStack.isEmpty(), "mark stack should be empty");
2459
2460 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2461 // in serial reference processing. Note these closures are also
2462 // used for serially processing (by the the current thread) the
2463 // JNI references during parallel reference processing.
2464 //
2465 // These closures do not need to synchronize with the worker
2466 // threads involved in parallel reference processing as these
2467 // instances are executed serially by the current thread (e.g.
2468 // reference processing is not multi-threaded and is thus
2469 // performed by the current thread instead of a gang worker).
2470 //
2471 // The gang tasks involved in parallel reference procssing create
2472 // their own instances of these closures, which do their own
2473 // synchronization among themselves.
2474 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2475 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2476
2477 // We need at least one active thread. If reference processing
2478 // is not multi-threaded we use the current (VMThread) thread,
2479 // otherwise we use the work gang from the G1CollectedHeap and
2480 // we utilize all the worker threads we can.
2481 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2482 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2483 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2484
2485 // Parallel processing task executor.
2486 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2487 g1h->workers(), active_workers);
2488 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2489
2490 // Set the concurrency level. The phase was already set prior to
2491 // executing the remark task.
2492 set_concurrency(active_workers);
2493
2494 // Set the degree of MT processing here. If the discovery was done MT,
2495 // the number of threads involved during discovery could differ from
2496 // the number of active workers. This is OK as long as the discovered
2497 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2498 rp->set_active_mt_degree(active_workers);
2499
2500 // Process the weak references.
2501 const ReferenceProcessorStats& stats =
2502 rp->process_discovered_references(&g1_is_alive,
2503 &g1_keep_alive,
2504 &g1_drain_mark_stack,
2505 executor,
2506 g1h->gc_timer_cm());
2507 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2508
2509 // The do_oop work routines of the keep_alive and drain_marking_stack
2510 // oop closures will set the has_overflown flag if we overflow the
2511 // global marking stack.
2512
2513 assert(_markStack.overflow() || _markStack.isEmpty(),
2514 "mark stack should be empty (unless it overflowed)");
2515
2516 if (_markStack.overflow()) {
2517 // This should have been done already when we tried to push an
2518 // entry on to the global mark stack. But let's do it again.
2519 set_has_overflown();
2520 }
2521
2522 assert(rp->num_q() == active_workers, "why not");
2523
2524 rp->enqueue_discovered_references(executor);
2525
2526 rp->verify_no_references_recorded();
2527 assert(!rp->discovery_enabled(), "Post condition");
2528 }
2529
2530 // Now clean up stale oops in StringTable
2531 StringTable::unlink(&g1_is_alive);
2532 // Clean up unreferenced symbols in symbol table.
2533 SymbolTable::unlink();
2534 }
2535
2536 void ConcurrentMark::swapMarkBitMaps() {
2537 CMBitMapRO* temp = _prevMarkBitMap;
2538 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2539 _nextMarkBitMap = (CMBitMap*) temp;
2540 }
2541
2542 class CMRemarkTask: public AbstractGangTask {
2543 private:
2544 ConcurrentMark* _cm;
2545 bool _is_serial;
2546 public:
2547 void work(uint worker_id) {
2548 // Since all available tasks are actually started, we should
2549 // only proceed if we're supposed to be actived.
2550 if (worker_id < _cm->active_tasks()) {
2551 CMTask* task = _cm->task(worker_id);
2552 task->record_start_time();
2553 do {
2554 task->do_marking_step(1000000000.0 /* something very large */,
2555 true /* do_termination */,
2556 _is_serial);
2557 } while (task->has_aborted() && !_cm->has_overflown());
2558 // If we overflow, then we do not want to restart. We instead
2559 // want to abort remark and do concurrent marking again.
2560 task->record_end_time();
2561 }
2562 }
2563
2564 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2565 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2566 _cm->terminator()->reset_for_reuse(active_workers);
2567 }
2568 };
2569
2570 void ConcurrentMark::checkpointRootsFinalWork() {
2571 ResourceMark rm;
2572 HandleMark hm;
2573 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2574
2575 g1h->ensure_parsability(false);
2576
2577 if (G1CollectedHeap::use_parallel_gc_threads()) {
2578 G1CollectedHeap::StrongRootsScope srs(g1h);
2579 // this is remark, so we'll use up all active threads
2580 uint active_workers = g1h->workers()->active_workers();
2581 if (active_workers == 0) {
2582 assert(active_workers > 0, "Should have been set earlier");
2583 active_workers = (uint) ParallelGCThreads;
2584 g1h->workers()->set_active_workers(active_workers);
2585 }
2586 set_concurrency_and_phase(active_workers, false /* concurrent */);
2587 // Leave _parallel_marking_threads at it's
2588 // value originally calculated in the ConcurrentMark
2589 // constructor and pass values of the active workers
2590 // through the gang in the task.
2591
2592 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2593 // We will start all available threads, even if we decide that the
2594 // active_workers will be fewer. The extra ones will just bail out
2595 // immediately.
2596 g1h->set_par_threads(active_workers);
2597 g1h->workers()->run_task(&remarkTask);
2598 g1h->set_par_threads(0);
2599 } else {
2600 G1CollectedHeap::StrongRootsScope srs(g1h);
2601 uint active_workers = 1;
2602 set_concurrency_and_phase(active_workers, false /* concurrent */);
2603
2604 // Note - if there's no work gang then the VMThread will be
2605 // the thread to execute the remark - serially. We have
2606 // to pass true for the is_serial parameter so that
2607 // CMTask::do_marking_step() doesn't enter the sync
2608 // barriers in the event of an overflow. Doing so will
2609 // cause an assert that the current thread is not a
2610 // concurrent GC thread.
2611 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2612 remarkTask.work(0);
2613 }
2614 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2615 guarantee(has_overflown() ||
2616 satb_mq_set.completed_buffers_num() == 0,
2617 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2618 BOOL_TO_STR(has_overflown()),
2619 satb_mq_set.completed_buffers_num()));
2620
2621 print_stats();
2622 }
2623
2624 #ifndef PRODUCT
2625
2626 class PrintReachableOopClosure: public OopClosure {
2627 private:
2628 G1CollectedHeap* _g1h;
2629 outputStream* _out;
2630 VerifyOption _vo;
2631 bool _all;
2632
2633 public:
2634 PrintReachableOopClosure(outputStream* out,
2635 VerifyOption vo,
2636 bool all) :
2637 _g1h(G1CollectedHeap::heap()),
2638 _out(out), _vo(vo), _all(all) { }
2639
2640 void do_oop(narrowOop* p) { do_oop_work(p); }
2641 void do_oop( oop* p) { do_oop_work(p); }
2642
2643 template <class T> void do_oop_work(T* p) {
2644 oop obj = oopDesc::load_decode_heap_oop(p);
2645 const char* str = NULL;
2646 const char* str2 = "";
2647
2648 if (obj == NULL) {
2649 str = "";
2650 } else if (!_g1h->is_in_g1_reserved(obj)) {
2651 str = " O";
2652 } else {
2653 HeapRegion* hr = _g1h->heap_region_containing(obj);
2654 guarantee(hr != NULL, "invariant");
2655 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2656 bool marked = _g1h->is_marked(obj, _vo);
2657
2658 if (over_tams) {
2659 str = " >";
2660 if (marked) {
2661 str2 = " AND MARKED";
2662 }
2663 } else if (marked) {
2664 str = " M";
2665 } else {
2666 str = " NOT";
2667 }
2668 }
2669
2670 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2671 p, (void*) obj, str, str2);
2672 }
2673 };
2674
2675 class PrintReachableObjectClosure : public ObjectClosure {
2676 private:
2677 G1CollectedHeap* _g1h;
2678 outputStream* _out;
2679 VerifyOption _vo;
2680 bool _all;
2681 HeapRegion* _hr;
2682
2683 public:
2684 PrintReachableObjectClosure(outputStream* out,
2685 VerifyOption vo,
2686 bool all,
2687 HeapRegion* hr) :
2688 _g1h(G1CollectedHeap::heap()),
2689 _out(out), _vo(vo), _all(all), _hr(hr) { }
2690
2691 void do_object(oop o) {
2692 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2693 bool marked = _g1h->is_marked(o, _vo);
2694 bool print_it = _all || over_tams || marked;
2695
2696 if (print_it) {
2697 _out->print_cr(" "PTR_FORMAT"%s",
2698 o, (over_tams) ? " >" : (marked) ? " M" : "");
2699 PrintReachableOopClosure oopCl(_out, _vo, _all);
2700 o->oop_iterate_no_header(&oopCl);
2701 }
2702 }
2703 };
2704
2705 class PrintReachableRegionClosure : public HeapRegionClosure {
2706 private:
2707 G1CollectedHeap* _g1h;
2708 outputStream* _out;
2709 VerifyOption _vo;
2710 bool _all;
2711
2712 public:
2713 bool doHeapRegion(HeapRegion* hr) {
2714 HeapWord* b = hr->bottom();
2715 HeapWord* e = hr->end();
2716 HeapWord* t = hr->top();
2717 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2718 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2719 "TAMS: "PTR_FORMAT, b, e, t, p);
2720 _out->cr();
2721
2722 HeapWord* from = b;
2723 HeapWord* to = t;
2724
2725 if (to > from) {
2726 _out->print_cr("Objects in ["PTR_FORMAT", "PTR_FORMAT"]", from, to);
2727 _out->cr();
2728 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2729 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2730 _out->cr();
2731 }
2732
2733 return false;
2734 }
2735
2736 PrintReachableRegionClosure(outputStream* out,
2737 VerifyOption vo,
2738 bool all) :
2739 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2740 };
2741
2742 void ConcurrentMark::print_reachable(const char* str,
2743 VerifyOption vo,
2744 bool all) {
2745 gclog_or_tty->cr();
2746 gclog_or_tty->print_cr("== Doing heap dump... ");
2747
2748 if (G1PrintReachableBaseFile == NULL) {
2749 gclog_or_tty->print_cr(" #### error: no base file defined");
2750 return;
2751 }
2752
2753 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2754 (JVM_MAXPATHLEN - 1)) {
2755 gclog_or_tty->print_cr(" #### error: file name too long");
2756 return;
2757 }
2758
2759 char file_name[JVM_MAXPATHLEN];
2760 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2761 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2762
2763 fileStream fout(file_name);
2764 if (!fout.is_open()) {
2765 gclog_or_tty->print_cr(" #### error: could not open file");
2766 return;
2767 }
2768
2769 outputStream* out = &fout;
2770 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2771 out->cr();
2772
2773 out->print_cr("--- ITERATING OVER REGIONS");
2774 out->cr();
2775 PrintReachableRegionClosure rcl(out, vo, all);
2776 _g1h->heap_region_iterate(&rcl);
2777 out->cr();
2778
2779 gclog_or_tty->print_cr(" done");
2780 gclog_or_tty->flush();
2781 }
2782
2783 #endif // PRODUCT
2784
2785 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2786 // Note we are overriding the read-only view of the prev map here, via
2787 // the cast.
2788 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2789 }
2790
2791 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2792 _nextMarkBitMap->clearRange(mr);
2793 }
2794
2795 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) {
2796 clearRangePrevBitmap(mr);
2797 clearRangeNextBitmap(mr);
2798 }
2799
2800 HeapRegion*
2801 ConcurrentMark::claim_region(uint worker_id) {
2802 // "checkpoint" the finger
2803 HeapWord* finger = _finger;
2804
2805 // _heap_end will not change underneath our feet; it only changes at
2806 // yield points.
2807 while (finger < _heap_end) {
2808 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2809
2810 // Note on how this code handles humongous regions. In the
2811 // normal case the finger will reach the start of a "starts
2812 // humongous" (SH) region. Its end will either be the end of the
2813 // last "continues humongous" (CH) region in the sequence, or the
2814 // standard end of the SH region (if the SH is the only region in
2815 // the sequence). That way claim_region() will skip over the CH
2816 // regions. However, there is a subtle race between a CM thread
2817 // executing this method and a mutator thread doing a humongous
2818 // object allocation. The two are not mutually exclusive as the CM
2819 // thread does not need to hold the Heap_lock when it gets
2820 // here. So there is a chance that claim_region() will come across
2821 // a free region that's in the progress of becoming a SH or a CH
2822 // region. In the former case, it will either
2823 // a) Miss the update to the region's end, in which case it will
2824 // visit every subsequent CH region, will find their bitmaps
2825 // empty, and do nothing, or
2826 // b) Will observe the update of the region's end (in which case
2827 // it will skip the subsequent CH regions).
2828 // If it comes across a region that suddenly becomes CH, the
2829 // scenario will be similar to b). So, the race between
2830 // claim_region() and a humongous object allocation might force us
2831 // to do a bit of unnecessary work (due to some unnecessary bitmap
2832 // iterations) but it should not introduce and correctness issues.
2833 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2834 HeapWord* bottom = curr_region->bottom();
2835 HeapWord* end = curr_region->end();
2836 HeapWord* limit = curr_region->next_top_at_mark_start();
2837
2838 if (verbose_low()) {
2839 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2840 "["PTR_FORMAT", "PTR_FORMAT"), "
2841 "limit = "PTR_FORMAT,
2842 worker_id, curr_region, bottom, end, limit);
2843 }
2844
2845 // Is the gap between reading the finger and doing the CAS too long?
2846 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2847 if (res == finger) {
2848 // we succeeded
2849
2850 // notice that _finger == end cannot be guaranteed here since,
2851 // someone else might have moved the finger even further
2852 assert(_finger >= end, "the finger should have moved forward");
2853
2854 if (verbose_low()) {
2855 gclog_or_tty->print_cr("[%u] we were successful with region = "
2856 PTR_FORMAT, worker_id, curr_region);
2857 }
2858
2859 if (limit > bottom) {
2860 if (verbose_low()) {
2861 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2862 "returning it ", worker_id, curr_region);
2863 }
2864 return curr_region;
2865 } else {
2866 assert(limit == bottom,
2867 "the region limit should be at bottom");
2868 if (verbose_low()) {
2869 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
2870 "returning NULL", worker_id, curr_region);
2871 }
2872 // we return NULL and the caller should try calling
2873 // claim_region() again.
2874 return NULL;
2875 }
2876 } else {
2877 assert(_finger > finger, "the finger should have moved forward");
2878 if (verbose_low()) {
2879 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2880 "global finger = "PTR_FORMAT", "
2881 "our finger = "PTR_FORMAT,
2882 worker_id, _finger, finger);
2883 }
2884
2885 // read it again
2886 finger = _finger;
2887 }
2888 }
2889
2890 return NULL;
2891 }
2892
2893 #ifndef PRODUCT
2894 enum VerifyNoCSetOopsPhase {
2895 VerifyNoCSetOopsStack,
2896 VerifyNoCSetOopsQueues,
2897 VerifyNoCSetOopsSATBCompleted,
2898 VerifyNoCSetOopsSATBThread
2899 };
2900
2901 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
2902 private:
2903 G1CollectedHeap* _g1h;
2904 VerifyNoCSetOopsPhase _phase;
2905 int _info;
2906
2907 const char* phase_str() {
2908 switch (_phase) {
2909 case VerifyNoCSetOopsStack: return "Stack";
2910 case VerifyNoCSetOopsQueues: return "Queue";
2911 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2912 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
2913 default: ShouldNotReachHere();
2914 }
2915 return NULL;
2916 }
2917
2918 void do_object_work(oop obj) {
2919 guarantee(!_g1h->obj_in_cs(obj),
2920 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2921 (void*) obj, phase_str(), _info));
2922 }
2923
2924 public:
2925 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2926
2927 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2928 _phase = phase;
2929 _info = info;
2930 }
2931
2932 virtual void do_oop(oop* p) {
2933 oop obj = oopDesc::load_decode_heap_oop(p);
2934 do_object_work(obj);
2935 }
2936
2937 virtual void do_oop(narrowOop* p) {
2938 // We should not come across narrow oops while scanning marking
2939 // stacks and SATB buffers.
2940 ShouldNotReachHere();
2941 }
2942
2943 virtual void do_object(oop obj) {
2944 do_object_work(obj);
2945 }
2946 };
2947
2948 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
2949 bool verify_enqueued_buffers,
2950 bool verify_thread_buffers,
2951 bool verify_fingers) {
2952 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2953 if (!G1CollectedHeap::heap()->mark_in_progress()) {
2954 return;
2955 }
2956
2957 VerifyNoCSetOopsClosure cl;
2958
2959 if (verify_stacks) {
2960 // Verify entries on the global mark stack
2961 cl.set_phase(VerifyNoCSetOopsStack);
2962 _markStack.oops_do(&cl);
2963
2964 // Verify entries on the task queues
2965 for (uint i = 0; i < _max_worker_id; i += 1) {
2966 cl.set_phase(VerifyNoCSetOopsQueues, i);
2967 CMTaskQueue* queue = _task_queues->queue(i);
2968 queue->oops_do(&cl);
2969 }
2970 }
2971
2972 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
2973
2974 // Verify entries on the enqueued SATB buffers
2975 if (verify_enqueued_buffers) {
2976 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
2977 satb_qs.iterate_completed_buffers_read_only(&cl);
2978 }
2979
2980 // Verify entries on the per-thread SATB buffers
2981 if (verify_thread_buffers) {
2982 cl.set_phase(VerifyNoCSetOopsSATBThread);
2983 satb_qs.iterate_thread_buffers_read_only(&cl);
2984 }
2985
2986 if (verify_fingers) {
2987 // Verify the global finger
2988 HeapWord* global_finger = finger();
2989 if (global_finger != NULL && global_finger < _heap_end) {
2990 // The global finger always points to a heap region boundary. We
2991 // use heap_region_containing_raw() to get the containing region
2992 // given that the global finger could be pointing to a free region
2993 // which subsequently becomes continues humongous. If that
2994 // happens, heap_region_containing() will return the bottom of the
2995 // corresponding starts humongous region and the check below will
2996 // not hold any more.
2997 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
2998 guarantee(global_finger == global_hr->bottom(),
2999 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3000 global_finger, HR_FORMAT_PARAMS(global_hr)));
3001 }
3002
3003 // Verify the task fingers
3004 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3005 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3006 CMTask* task = _tasks[i];
3007 HeapWord* task_finger = task->finger();
3008 if (task_finger != NULL && task_finger < _heap_end) {
3009 // See above note on the global finger verification.
3010 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3011 guarantee(task_finger == task_hr->bottom() ||
3012 !task_hr->in_collection_set(),
3013 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3014 task_finger, HR_FORMAT_PARAMS(task_hr)));
3015 }
3016 }
3017 }
3018 }
3019 #endif // PRODUCT
3020
3021 // Aggregate the counting data that was constructed concurrently
3022 // with marking.
3023 class AggregateCountDataHRClosure: public HeapRegionClosure {
3024 G1CollectedHeap* _g1h;
3025 ConcurrentMark* _cm;
3026 CardTableModRefBS* _ct_bs;
3027 BitMap* _cm_card_bm;
3028 uint _max_worker_id;
3029
3030 public:
3031 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3032 BitMap* cm_card_bm,
3033 uint max_worker_id) :
3034 _g1h(g1h), _cm(g1h->concurrent_mark()),
3035 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3036 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3037
3038 bool doHeapRegion(HeapRegion* hr) {
3039 if (hr->continuesHumongous()) {
3040 // We will ignore these here and process them when their
3041 // associated "starts humongous" region is processed.
3042 // Note that we cannot rely on their associated
3043 // "starts humongous" region to have their bit set to 1
3044 // since, due to the region chunking in the parallel region
3045 // iteration, a "continues humongous" region might be visited
3046 // before its associated "starts humongous".
3047 return false;
3048 }
3049
3050 HeapWord* start = hr->bottom();
3051 HeapWord* limit = hr->next_top_at_mark_start();
3052 HeapWord* end = hr->end();
3053
3054 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3055 err_msg("Preconditions not met - "
3056 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3057 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3058 start, limit, hr->top(), hr->end()));
3059
3060 assert(hr->next_marked_bytes() == 0, "Precondition");
3061
3062 if (start == limit) {
3063 // NTAMS of this region has not been set so nothing to do.
3064 return false;
3065 }
3066
3067 // 'start' should be in the heap.
3068 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3069 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3070 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3071
3072 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3073 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3074 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3075
3076 // If ntams is not card aligned then we bump card bitmap index
3077 // for limit so that we get the all the cards spanned by
3078 // the object ending at ntams.
3079 // Note: if this is the last region in the heap then ntams
3080 // could be actually just beyond the end of the the heap;
3081 // limit_idx will then correspond to a (non-existent) card
3082 // that is also outside the heap.
3083 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3084 limit_idx += 1;
3085 }
3086
3087 assert(limit_idx <= end_idx, "or else use atomics");
3088
3089 // Aggregate the "stripe" in the count data associated with hr.
3090 uint hrs_index = hr->hrs_index();
3091 size_t marked_bytes = 0;
3092
3093 for (uint i = 0; i < _max_worker_id; i += 1) {
3094 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3095 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3096
3097 // Fetch the marked_bytes in this region for task i and
3098 // add it to the running total for this region.
3099 marked_bytes += marked_bytes_array[hrs_index];
3100
3101 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3102 // into the global card bitmap.
3103 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3104
3105 while (scan_idx < limit_idx) {
3106 assert(task_card_bm->at(scan_idx) == true, "should be");
3107 _cm_card_bm->set_bit(scan_idx);
3108 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3109
3110 // BitMap::get_next_one_offset() can handle the case when
3111 // its left_offset parameter is greater than its right_offset
3112 // parameter. It does, however, have an early exit if
3113 // left_offset == right_offset. So let's limit the value
3114 // passed in for left offset here.
3115 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3116 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3117 }
3118 }
3119
3120 // Update the marked bytes for this region.
3121 hr->add_to_marked_bytes(marked_bytes);
3122
3123 // Next heap region
3124 return false;
3125 }
3126 };
3127
3128 class G1AggregateCountDataTask: public AbstractGangTask {
3129 protected:
3130 G1CollectedHeap* _g1h;
3131 ConcurrentMark* _cm;
3132 BitMap* _cm_card_bm;
3133 uint _max_worker_id;
3134 int _active_workers;
3135
3136 public:
3137 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3138 ConcurrentMark* cm,
3139 BitMap* cm_card_bm,
3140 uint max_worker_id,
3141 int n_workers) :
3142 AbstractGangTask("Count Aggregation"),
3143 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3144 _max_worker_id(max_worker_id),
3145 _active_workers(n_workers) { }
3146
3147 void work(uint worker_id) {
3148 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3149
3150 if (G1CollectedHeap::use_parallel_gc_threads()) {
3151 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3152 _active_workers,
3153 HeapRegion::AggregateCountClaimValue);
3154 } else {
3155 _g1h->heap_region_iterate(&cl);
3156 }
3157 }
3158 };
3159
3160
3161 void ConcurrentMark::aggregate_count_data() {
3162 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3163 _g1h->workers()->active_workers() :
3164 1);
3165
3166 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3167 _max_worker_id, n_workers);
3168
3169 if (G1CollectedHeap::use_parallel_gc_threads()) {
3170 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3171 "sanity check");
3172 _g1h->set_par_threads(n_workers);
3173 _g1h->workers()->run_task(&g1_par_agg_task);
3174 _g1h->set_par_threads(0);
3175
3176 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3177 "sanity check");
3178 _g1h->reset_heap_region_claim_values();
3179 } else {
3180 g1_par_agg_task.work(0);
3181 }
3182 }
3183
3184 // Clear the per-worker arrays used to store the per-region counting data
3185 void ConcurrentMark::clear_all_count_data() {
3186 // Clear the global card bitmap - it will be filled during
3187 // liveness count aggregation (during remark) and the
3188 // final counting task.
3189 _card_bm.clear();
3190
3191 // Clear the global region bitmap - it will be filled as part
3192 // of the final counting task.
3193 _region_bm.clear();
3194
3195 uint max_regions = _g1h->max_regions();
3196 assert(_max_worker_id > 0, "uninitialized");
3197
3198 for (uint i = 0; i < _max_worker_id; i += 1) {
3199 BitMap* task_card_bm = count_card_bitmap_for(i);
3200 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3201
3202 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3203 assert(marked_bytes_array != NULL, "uninitialized");
3204
3205 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3206 task_card_bm->clear();
3207 }
3208 }
3209
3210 void ConcurrentMark::print_stats() {
3211 if (verbose_stats()) {
3212 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3213 for (size_t i = 0; i < _active_tasks; ++i) {
3214 _tasks[i]->print_stats();
3215 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3216 }
3217 }
3218 }
3219
3220 // abandon current marking iteration due to a Full GC
3221 void ConcurrentMark::abort() {
3222 // Clear all marks to force marking thread to do nothing
3223 _nextMarkBitMap->clearAll();
3224 // Clear the liveness counting data
3225 clear_all_count_data();
3226 // Empty mark stack
3227 reset_marking_state();
3228 for (uint i = 0; i < _max_worker_id; ++i) {
3229 _tasks[i]->clear_region_fields();
3230 }
3231 _has_aborted = true;
3232
3233 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3234 satb_mq_set.abandon_partial_marking();
3235 // This can be called either during or outside marking, we'll read
3236 // the expected_active value from the SATB queue set.
3237 satb_mq_set.set_active_all_threads(
3238 false, /* new active value */
3239 satb_mq_set.is_active() /* expected_active */);
3240
3241 _g1h->trace_heap_after_concurrent_cycle();
3242 _g1h->register_concurrent_cycle_end();
3243 }
3244
3245 static void print_ms_time_info(const char* prefix, const char* name,
3246 NumberSeq& ns) {
3247 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3248 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3249 if (ns.num() > 0) {
3250 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3251 prefix, ns.sd(), ns.maximum());
3252 }
3253 }
3254
3255 void ConcurrentMark::print_summary_info() {
3256 gclog_or_tty->print_cr(" Concurrent marking:");
3257 print_ms_time_info(" ", "init marks", _init_times);
3258 print_ms_time_info(" ", "remarks", _remark_times);
3259 {
3260 print_ms_time_info(" ", "final marks", _remark_mark_times);
3261 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3262
3263 }
3264 print_ms_time_info(" ", "cleanups", _cleanup_times);
3265 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3266 _total_counting_time,
3267 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3268 (double)_cleanup_times.num()
3269 : 0.0));
3270 if (G1ScrubRemSets) {
3271 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3272 _total_rs_scrub_time,
3273 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3274 (double)_cleanup_times.num()
3275 : 0.0));
3276 }
3277 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3278 (_init_times.sum() + _remark_times.sum() +
3279 _cleanup_times.sum())/1000.0);
3280 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3281 "(%8.2f s marking).",
3282 cmThread()->vtime_accum(),
3283 cmThread()->vtime_mark_accum());
3284 }
3285
3286 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3287 if (use_parallel_marking_threads()) {
3288 _parallel_workers->print_worker_threads_on(st);
3289 }
3290 }
3291
3292 void ConcurrentMark::print_on_error(outputStream* st) const {
3293 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3294 _prevMarkBitMap, _nextMarkBitMap);
3295 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3296 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3297 }
3298
3299 // We take a break if someone is trying to stop the world.
3300 bool ConcurrentMark::do_yield_check(uint worker_id) {
3301 if (should_yield()) {
3302 if (worker_id == 0) {
3303 _g1h->g1_policy()->record_concurrent_pause();
3304 }
3305 cmThread()->yield();
3306 return true;
3307 } else {
3308 return false;
3309 }
3310 }
3311
3312 bool ConcurrentMark::should_yield() {
3313 return cmThread()->should_yield();
3314 }
3315
3316 bool ConcurrentMark::containing_card_is_marked(void* p) {
3317 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1);
3318 return _card_bm.at(offset >> CardTableModRefBS::card_shift);
3319 }
3320
3321 bool ConcurrentMark::containing_cards_are_marked(void* start,
3322 void* last) {
3323 return containing_card_is_marked(start) &&
3324 containing_card_is_marked(last);
3325 }
3326
3327 #ifndef PRODUCT
3328 // for debugging purposes
3329 void ConcurrentMark::print_finger() {
3330 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3331 _heap_start, _heap_end, _finger);
3332 for (uint i = 0; i < _max_worker_id; ++i) {
3333 gclog_or_tty->print(" %u: "PTR_FORMAT, i, _tasks[i]->finger());
3334 }
3335 gclog_or_tty->print_cr("");
3336 }
3337 #endif
3338
3339 void CMTask::scan_object(oop obj) {
3340 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3341
3342 if (_cm->verbose_high()) {
3343 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3344 _worker_id, (void*) obj);
3345 }
3346
3347 size_t obj_size = obj->size();
3348 _words_scanned += obj_size;
3349
3350 obj->oop_iterate(_cm_oop_closure);
3351 statsOnly( ++_objs_scanned );
3352 check_limits();
3353 }
3354
3355 // Closure for iteration over bitmaps
3356 class CMBitMapClosure : public BitMapClosure {
3357 private:
3358 // the bitmap that is being iterated over
3359 CMBitMap* _nextMarkBitMap;
3360 ConcurrentMark* _cm;
3361 CMTask* _task;
3362
3363 public:
3364 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3365 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3366
3367 bool do_bit(size_t offset) {
3368 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3369 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3370 assert( addr < _cm->finger(), "invariant");
3371
3372 statsOnly( _task->increase_objs_found_on_bitmap() );
3373 assert(addr >= _task->finger(), "invariant");
3374
3375 // We move that task's local finger along.
3376 _task->move_finger_to(addr);
3377
3378 _task->scan_object(oop(addr));
3379 // we only partially drain the local queue and global stack
3380 _task->drain_local_queue(true);
3381 _task->drain_global_stack(true);
3382
3383 // if the has_aborted flag has been raised, we need to bail out of
3384 // the iteration
3385 return !_task->has_aborted();
3386 }
3387 };
3388
3389 // Closure for iterating over objects, currently only used for
3390 // processing SATB buffers.
3391 class CMObjectClosure : public ObjectClosure {
3392 private:
3393 CMTask* _task;
3394
3395 public:
3396 void do_object(oop obj) {
3397 _task->deal_with_reference(obj);
3398 }
3399
3400 CMObjectClosure(CMTask* task) : _task(task) { }
3401 };
3402
3403 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3404 ConcurrentMark* cm,
3405 CMTask* task)
3406 : _g1h(g1h), _cm(cm), _task(task) {
3407 assert(_ref_processor == NULL, "should be initialized to NULL");
3408
3409 if (G1UseConcMarkReferenceProcessing) {
3410 _ref_processor = g1h->ref_processor_cm();
3411 assert(_ref_processor != NULL, "should not be NULL");
3412 }
3413 }
3414
3415 void CMTask::setup_for_region(HeapRegion* hr) {
3416 // Separated the asserts so that we know which one fires.
3417 assert(hr != NULL,
3418 "claim_region() should have filtered out continues humongous regions");
3419 assert(!hr->continuesHumongous(),
3420 "claim_region() should have filtered out continues humongous regions");
3421
3422 if (_cm->verbose_low()) {
3423 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3424 _worker_id, hr);
3425 }
3426
3427 _curr_region = hr;
3428 _finger = hr->bottom();
3429 update_region_limit();
3430 }
3431
3432 void CMTask::update_region_limit() {
3433 HeapRegion* hr = _curr_region;
3434 HeapWord* bottom = hr->bottom();
3435 HeapWord* limit = hr->next_top_at_mark_start();
3436
3437 if (limit == bottom) {
3438 if (_cm->verbose_low()) {
3439 gclog_or_tty->print_cr("[%u] found an empty region "
3440 "["PTR_FORMAT", "PTR_FORMAT")",
3441 _worker_id, bottom, limit);
3442 }
3443 // The region was collected underneath our feet.
3444 // We set the finger to bottom to ensure that the bitmap
3445 // iteration that will follow this will not do anything.
3446 // (this is not a condition that holds when we set the region up,
3447 // as the region is not supposed to be empty in the first place)
3448 _finger = bottom;
3449 } else if (limit >= _region_limit) {
3450 assert(limit >= _finger, "peace of mind");
3451 } else {
3452 assert(limit < _region_limit, "only way to get here");
3453 // This can happen under some pretty unusual circumstances. An
3454 // evacuation pause empties the region underneath our feet (NTAMS
3455 // at bottom). We then do some allocation in the region (NTAMS
3456 // stays at bottom), followed by the region being used as a GC
3457 // alloc region (NTAMS will move to top() and the objects
3458 // originally below it will be grayed). All objects now marked in
3459 // the region are explicitly grayed, if below the global finger,
3460 // and we do not need in fact to scan anything else. So, we simply
3461 // set _finger to be limit to ensure that the bitmap iteration
3462 // doesn't do anything.
3463 _finger = limit;
3464 }
3465
3466 _region_limit = limit;
3467 }
3468
3469 void CMTask::giveup_current_region() {
3470 assert(_curr_region != NULL, "invariant");
3471 if (_cm->verbose_low()) {
3472 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3473 _worker_id, _curr_region);
3474 }
3475 clear_region_fields();
3476 }
3477
3478 void CMTask::clear_region_fields() {
3479 // Values for these three fields that indicate that we're not
3480 // holding on to a region.
3481 _curr_region = NULL;
3482 _finger = NULL;
3483 _region_limit = NULL;
3484 }
3485
3486 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3487 if (cm_oop_closure == NULL) {
3488 assert(_cm_oop_closure != NULL, "invariant");
3489 } else {
3490 assert(_cm_oop_closure == NULL, "invariant");
3491 }
3492 _cm_oop_closure = cm_oop_closure;
3493 }
3494
3495 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3496 guarantee(nextMarkBitMap != NULL, "invariant");
3497
3498 if (_cm->verbose_low()) {
3499 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3500 }
3501
3502 _nextMarkBitMap = nextMarkBitMap;
3503 clear_region_fields();
3504
3505 _calls = 0;
3506 _elapsed_time_ms = 0.0;
3507 _termination_time_ms = 0.0;
3508 _termination_start_time_ms = 0.0;
3509
3510 #if _MARKING_STATS_
3511 _local_pushes = 0;
3512 _local_pops = 0;
3513 _local_max_size = 0;
3514 _objs_scanned = 0;
3515 _global_pushes = 0;
3516 _global_pops = 0;
3517 _global_max_size = 0;
3518 _global_transfers_to = 0;
3519 _global_transfers_from = 0;
3520 _regions_claimed = 0;
3521 _objs_found_on_bitmap = 0;
3522 _satb_buffers_processed = 0;
3523 _steal_attempts = 0;
3524 _steals = 0;
3525 _aborted = 0;
3526 _aborted_overflow = 0;
3527 _aborted_cm_aborted = 0;
3528 _aborted_yield = 0;
3529 _aborted_timed_out = 0;
3530 _aborted_satb = 0;
3531 _aborted_termination = 0;
3532 #endif // _MARKING_STATS_
3533 }
3534
3535 bool CMTask::should_exit_termination() {
3536 regular_clock_call();
3537 // This is called when we are in the termination protocol. We should
3538 // quit if, for some reason, this task wants to abort or the global
3539 // stack is not empty (this means that we can get work from it).
3540 return !_cm->mark_stack_empty() || has_aborted();
3541 }
3542
3543 void CMTask::reached_limit() {
3544 assert(_words_scanned >= _words_scanned_limit ||
3545 _refs_reached >= _refs_reached_limit ,
3546 "shouldn't have been called otherwise");
3547 regular_clock_call();
3548 }
3549
3550 void CMTask::regular_clock_call() {
3551 if (has_aborted()) return;
3552
3553 // First, we need to recalculate the words scanned and refs reached
3554 // limits for the next clock call.
3555 recalculate_limits();
3556
3557 // During the regular clock call we do the following
3558
3559 // (1) If an overflow has been flagged, then we abort.
3560 if (_cm->has_overflown()) {
3561 set_has_aborted();
3562 return;
3563 }
3564
3565 // If we are not concurrent (i.e. we're doing remark) we don't need
3566 // to check anything else. The other steps are only needed during
3567 // the concurrent marking phase.
3568 if (!concurrent()) return;
3569
3570 // (2) If marking has been aborted for Full GC, then we also abort.
3571 if (_cm->has_aborted()) {
3572 set_has_aborted();
3573 statsOnly( ++_aborted_cm_aborted );
3574 return;
3575 }
3576
3577 double curr_time_ms = os::elapsedVTime() * 1000.0;
3578
3579 // (3) If marking stats are enabled, then we update the step history.
3580 #if _MARKING_STATS_
3581 if (_words_scanned >= _words_scanned_limit) {
3582 ++_clock_due_to_scanning;
3583 }
3584 if (_refs_reached >= _refs_reached_limit) {
3585 ++_clock_due_to_marking;
3586 }
3587
3588 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3589 _interval_start_time_ms = curr_time_ms;
3590 _all_clock_intervals_ms.add(last_interval_ms);
3591
3592 if (_cm->verbose_medium()) {
3593 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3594 "scanned = %d%s, refs reached = %d%s",
3595 _worker_id, last_interval_ms,
3596 _words_scanned,
3597 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3598 _refs_reached,
3599 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3600 }
3601 #endif // _MARKING_STATS_
3602
3603 // (4) We check whether we should yield. If we have to, then we abort.
3604 if (_cm->should_yield()) {
3605 // We should yield. To do this we abort the task. The caller is
3606 // responsible for yielding.
3607 set_has_aborted();
3608 statsOnly( ++_aborted_yield );
3609 return;
3610 }
3611
3612 // (5) We check whether we've reached our time quota. If we have,
3613 // then we abort.
3614 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3615 if (elapsed_time_ms > _time_target_ms) {
3616 set_has_aborted();
3617 _has_timed_out = true;
3618 statsOnly( ++_aborted_timed_out );
3619 return;
3620 }
3621
3622 // (6) Finally, we check whether there are enough completed STAB
3623 // buffers available for processing. If there are, we abort.
3624 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3625 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3626 if (_cm->verbose_low()) {
3627 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3628 _worker_id);
3629 }
3630 // we do need to process SATB buffers, we'll abort and restart
3631 // the marking task to do so
3632 set_has_aborted();
3633 statsOnly( ++_aborted_satb );
3634 return;
3635 }
3636 }
3637
3638 void CMTask::recalculate_limits() {
3639 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3640 _words_scanned_limit = _real_words_scanned_limit;
3641
3642 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3643 _refs_reached_limit = _real_refs_reached_limit;
3644 }
3645
3646 void CMTask::decrease_limits() {
3647 // This is called when we believe that we're going to do an infrequent
3648 // operation which will increase the per byte scanned cost (i.e. move
3649 // entries to/from the global stack). It basically tries to decrease the
3650 // scanning limit so that the clock is called earlier.
3651
3652 if (_cm->verbose_medium()) {
3653 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3654 }
3655
3656 _words_scanned_limit = _real_words_scanned_limit -
3657 3 * words_scanned_period / 4;
3658 _refs_reached_limit = _real_refs_reached_limit -
3659 3 * refs_reached_period / 4;
3660 }
3661
3662 void CMTask::move_entries_to_global_stack() {
3663 // local array where we'll store the entries that will be popped
3664 // from the local queue
3665 oop buffer[global_stack_transfer_size];
3666
3667 int n = 0;
3668 oop obj;
3669 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3670 buffer[n] = obj;
3671 ++n;
3672 }
3673
3674 if (n > 0) {
3675 // we popped at least one entry from the local queue
3676
3677 statsOnly( ++_global_transfers_to; _local_pops += n );
3678
3679 if (!_cm->mark_stack_push(buffer, n)) {
3680 if (_cm->verbose_low()) {
3681 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3682 _worker_id);
3683 }
3684 set_has_aborted();
3685 } else {
3686 // the transfer was successful
3687
3688 if (_cm->verbose_medium()) {
3689 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3690 _worker_id, n);
3691 }
3692 statsOnly( int tmp_size = _cm->mark_stack_size();
3693 if (tmp_size > _global_max_size) {
3694 _global_max_size = tmp_size;
3695 }
3696 _global_pushes += n );
3697 }
3698 }
3699
3700 // this operation was quite expensive, so decrease the limits
3701 decrease_limits();
3702 }
3703
3704 void CMTask::get_entries_from_global_stack() {
3705 // local array where we'll store the entries that will be popped
3706 // from the global stack.
3707 oop buffer[global_stack_transfer_size];
3708 int n;
3709 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3710 assert(n <= global_stack_transfer_size,
3711 "we should not pop more than the given limit");
3712 if (n > 0) {
3713 // yes, we did actually pop at least one entry
3714
3715 statsOnly( ++_global_transfers_from; _global_pops += n );
3716 if (_cm->verbose_medium()) {
3717 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3718 _worker_id, n);
3719 }
3720 for (int i = 0; i < n; ++i) {
3721 bool success = _task_queue->push(buffer[i]);
3722 // We only call this when the local queue is empty or under a
3723 // given target limit. So, we do not expect this push to fail.
3724 assert(success, "invariant");
3725 }
3726
3727 statsOnly( int tmp_size = _task_queue->size();
3728 if (tmp_size > _local_max_size) {
3729 _local_max_size = tmp_size;
3730 }
3731 _local_pushes += n );
3732 }
3733
3734 // this operation was quite expensive, so decrease the limits
3735 decrease_limits();
3736 }
3737
3738 void CMTask::drain_local_queue(bool partially) {
3739 if (has_aborted()) return;
3740
3741 // Decide what the target size is, depending whether we're going to
3742 // drain it partially (so that other tasks can steal if they run out
3743 // of things to do) or totally (at the very end).
3744 size_t target_size;
3745 if (partially) {
3746 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3747 } else {
3748 target_size = 0;
3749 }
3750
3751 if (_task_queue->size() > target_size) {
3752 if (_cm->verbose_high()) {
3753 gclog_or_tty->print_cr("[%u] draining local queue, target size = %d",
3754 _worker_id, target_size);
3755 }
3756
3757 oop obj;
3758 bool ret = _task_queue->pop_local(obj);
3759 while (ret) {
3760 statsOnly( ++_local_pops );
3761
3762 if (_cm->verbose_high()) {
3763 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3764 (void*) obj);
3765 }
3766
3767 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3768 assert(!_g1h->is_on_master_free_list(
3769 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3770
3771 scan_object(obj);
3772
3773 if (_task_queue->size() <= target_size || has_aborted()) {
3774 ret = false;
3775 } else {
3776 ret = _task_queue->pop_local(obj);
3777 }
3778 }
3779
3780 if (_cm->verbose_high()) {
3781 gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3782 _worker_id, _task_queue->size());
3783 }
3784 }
3785 }
3786
3787 void CMTask::drain_global_stack(bool partially) {
3788 if (has_aborted()) return;
3789
3790 // We have a policy to drain the local queue before we attempt to
3791 // drain the global stack.
3792 assert(partially || _task_queue->size() == 0, "invariant");
3793
3794 // Decide what the target size is, depending whether we're going to
3795 // drain it partially (so that other tasks can steal if they run out
3796 // of things to do) or totally (at the very end). Notice that,
3797 // because we move entries from the global stack in chunks or
3798 // because another task might be doing the same, we might in fact
3799 // drop below the target. But, this is not a problem.
3800 size_t target_size;
3801 if (partially) {
3802 target_size = _cm->partial_mark_stack_size_target();
3803 } else {
3804 target_size = 0;
3805 }
3806
3807 if (_cm->mark_stack_size() > target_size) {
3808 if (_cm->verbose_low()) {
3809 gclog_or_tty->print_cr("[%u] draining global_stack, target size %d",
3810 _worker_id, target_size);
3811 }
3812
3813 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3814 get_entries_from_global_stack();
3815 drain_local_queue(partially);
3816 }
3817
3818 if (_cm->verbose_low()) {
3819 gclog_or_tty->print_cr("[%u] drained global stack, size = %d",
3820 _worker_id, _cm->mark_stack_size());
3821 }
3822 }
3823 }
3824
3825 // SATB Queue has several assumptions on whether to call the par or
3826 // non-par versions of the methods. this is why some of the code is
3827 // replicated. We should really get rid of the single-threaded version
3828 // of the code to simplify things.
3829 void CMTask::drain_satb_buffers() {
3830 if (has_aborted()) return;
3831
3832 // We set this so that the regular clock knows that we're in the
3833 // middle of draining buffers and doesn't set the abort flag when it
3834 // notices that SATB buffers are available for draining. It'd be
3835 // very counter productive if it did that. :-)
3836 _draining_satb_buffers = true;
3837
3838 CMObjectClosure oc(this);
3839 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3840 if (G1CollectedHeap::use_parallel_gc_threads()) {
3841 satb_mq_set.set_par_closure(_worker_id, &oc);
3842 } else {
3843 satb_mq_set.set_closure(&oc);
3844 }
3845
3846 // This keeps claiming and applying the closure to completed buffers
3847 // until we run out of buffers or we need to abort.
3848 if (G1CollectedHeap::use_parallel_gc_threads()) {
3849 while (!has_aborted() &&
3850 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
3851 if (_cm->verbose_medium()) {
3852 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3853 }
3854 statsOnly( ++_satb_buffers_processed );
3855 regular_clock_call();
3856 }
3857 } else {
3858 while (!has_aborted() &&
3859 satb_mq_set.apply_closure_to_completed_buffer()) {
3860 if (_cm->verbose_medium()) {
3861 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3862 }
3863 statsOnly( ++_satb_buffers_processed );
3864 regular_clock_call();
3865 }
3866 }
3867
3868 if (!concurrent() && !has_aborted()) {
3869 // We should only do this during remark.
3870 if (G1CollectedHeap::use_parallel_gc_threads()) {
3871 satb_mq_set.par_iterate_closure_all_threads(_worker_id);
3872 } else {
3873 satb_mq_set.iterate_closure_all_threads();
3874 }
3875 }
3876
3877 _draining_satb_buffers = false;
3878
3879 assert(has_aborted() ||
3880 concurrent() ||
3881 satb_mq_set.completed_buffers_num() == 0, "invariant");
3882
3883 if (G1CollectedHeap::use_parallel_gc_threads()) {
3884 satb_mq_set.set_par_closure(_worker_id, NULL);
3885 } else {
3886 satb_mq_set.set_closure(NULL);
3887 }
3888
3889 // again, this was a potentially expensive operation, decrease the
3890 // limits to get the regular clock call early
3891 decrease_limits();
3892 }
3893
3894 void CMTask::print_stats() {
3895 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3896 _worker_id, _calls);
3897 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3898 _elapsed_time_ms, _termination_time_ms);
3899 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3900 _step_times_ms.num(), _step_times_ms.avg(),
3901 _step_times_ms.sd());
3902 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3903 _step_times_ms.maximum(), _step_times_ms.sum());
3904
3905 #if _MARKING_STATS_
3906 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3907 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3908 _all_clock_intervals_ms.sd());
3909 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3910 _all_clock_intervals_ms.maximum(),
3911 _all_clock_intervals_ms.sum());
3912 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
3913 _clock_due_to_scanning, _clock_due_to_marking);
3914 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
3915 _objs_scanned, _objs_found_on_bitmap);
3916 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
3917 _local_pushes, _local_pops, _local_max_size);
3918 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
3919 _global_pushes, _global_pops, _global_max_size);
3920 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
3921 _global_transfers_to,_global_transfers_from);
3922 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
3923 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
3924 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
3925 _steal_attempts, _steals);
3926 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
3927 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
3928 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3929 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
3930 _aborted_timed_out, _aborted_satb, _aborted_termination);
3931 #endif // _MARKING_STATS_
3932 }
3933
3934 /*****************************************************************************
3935
3936 The do_marking_step(time_target_ms, ...) method is the building
3937 block of the parallel marking framework. It can be called in parallel
3938 with other invocations of do_marking_step() on different tasks
3939 (but only one per task, obviously) and concurrently with the
3940 mutator threads, or during remark, hence it eliminates the need
3941 for two versions of the code. When called during remark, it will
3942 pick up from where the task left off during the concurrent marking
3943 phase. Interestingly, tasks are also claimable during evacuation
3944 pauses too, since do_marking_step() ensures that it aborts before
3945 it needs to yield.
3946
3947 The data structures that it uses to do marking work are the
3948 following:
3949
3950 (1) Marking Bitmap. If there are gray objects that appear only
3951 on the bitmap (this happens either when dealing with an overflow
3952 or when the initial marking phase has simply marked the roots
3953 and didn't push them on the stack), then tasks claim heap
3954 regions whose bitmap they then scan to find gray objects. A
3955 global finger indicates where the end of the last claimed region
3956 is. A local finger indicates how far into the region a task has
3957 scanned. The two fingers are used to determine how to gray an
3958 object (i.e. whether simply marking it is OK, as it will be
3959 visited by a task in the future, or whether it needs to be also
3960 pushed on a stack).
3961
3962 (2) Local Queue. The local queue of the task which is accessed
3963 reasonably efficiently by the task. Other tasks can steal from
3964 it when they run out of work. Throughout the marking phase, a
3965 task attempts to keep its local queue short but not totally
3966 empty, so that entries are available for stealing by other
3967 tasks. Only when there is no more work, a task will totally
3968 drain its local queue.
3969
3970 (3) Global Mark Stack. This handles local queue overflow. During
3971 marking only sets of entries are moved between it and the local
3972 queues, as access to it requires a mutex and more fine-grain
3973 interaction with it which might cause contention. If it
3974 overflows, then the marking phase should restart and iterate
3975 over the bitmap to identify gray objects. Throughout the marking
3976 phase, tasks attempt to keep the global mark stack at a small
3977 length but not totally empty, so that entries are available for
3978 popping by other tasks. Only when there is no more work, tasks
3979 will totally drain the global mark stack.
3980
3981 (4) SATB Buffer Queue. This is where completed SATB buffers are
3982 made available. Buffers are regularly removed from this queue
3983 and scanned for roots, so that the queue doesn't get too
3984 long. During remark, all completed buffers are processed, as
3985 well as the filled in parts of any uncompleted buffers.
3986
3987 The do_marking_step() method tries to abort when the time target
3988 has been reached. There are a few other cases when the
3989 do_marking_step() method also aborts:
3990
3991 (1) When the marking phase has been aborted (after a Full GC).
3992
3993 (2) When a global overflow (on the global stack) has been
3994 triggered. Before the task aborts, it will actually sync up with
3995 the other tasks to ensure that all the marking data structures
3996 (local queues, stacks, fingers etc.) are re-initialized so that
3997 when do_marking_step() completes, the marking phase can
3998 immediately restart.
3999
4000 (3) When enough completed SATB buffers are available. The
4001 do_marking_step() method only tries to drain SATB buffers right
4002 at the beginning. So, if enough buffers are available, the
4003 marking step aborts and the SATB buffers are processed at
4004 the beginning of the next invocation.
4005
4006 (4) To yield. when we have to yield then we abort and yield
4007 right at the end of do_marking_step(). This saves us from a lot
4008 of hassle as, by yielding we might allow a Full GC. If this
4009 happens then objects will be compacted underneath our feet, the
4010 heap might shrink, etc. We save checking for this by just
4011 aborting and doing the yield right at the end.
4012
4013 From the above it follows that the do_marking_step() method should
4014 be called in a loop (or, otherwise, regularly) until it completes.
4015
4016 If a marking step completes without its has_aborted() flag being
4017 true, it means it has completed the current marking phase (and
4018 also all other marking tasks have done so and have all synced up).
4019
4020 A method called regular_clock_call() is invoked "regularly" (in
4021 sub ms intervals) throughout marking. It is this clock method that
4022 checks all the abort conditions which were mentioned above and
4023 decides when the task should abort. A work-based scheme is used to
4024 trigger this clock method: when the number of object words the
4025 marking phase has scanned or the number of references the marking
4026 phase has visited reach a given limit. Additional invocations to
4027 the method clock have been planted in a few other strategic places
4028 too. The initial reason for the clock method was to avoid calling
4029 vtime too regularly, as it is quite expensive. So, once it was in
4030 place, it was natural to piggy-back all the other conditions on it
4031 too and not constantly check them throughout the code.
4032
4033 If do_termination is true then do_marking_step will enter its
4034 termination protocol.
4035
4036 The value of is_serial must be true when do_marking_step is being
4037 called serially (i.e. by the VMThread) and do_marking_step should
4038 skip any synchronization in the termination and overflow code.
4039 Examples include the serial remark code and the serial reference
4040 processing closures.
4041
4042 The value of is_serial must be false when do_marking_step is
4043 being called by any of the worker threads in a work gang.
4044 Examples include the concurrent marking code (CMMarkingTask),
4045 the MT remark code, and the MT reference processing closures.
4046
4047 *****************************************************************************/
4048
4049 void CMTask::do_marking_step(double time_target_ms,
4050 bool do_termination,
4051 bool is_serial) {
4052 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4053 assert(concurrent() == _cm->concurrent(), "they should be the same");
4054
4055 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4056 assert(_task_queues != NULL, "invariant");
4057 assert(_task_queue != NULL, "invariant");
4058 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4059
4060 assert(!_claimed,
4061 "only one thread should claim this task at any one time");
4062
4063 // OK, this doesn't safeguard again all possible scenarios, as it is
4064 // possible for two threads to set the _claimed flag at the same
4065 // time. But it is only for debugging purposes anyway and it will
4066 // catch most problems.
4067 _claimed = true;
4068
4069 _start_time_ms = os::elapsedVTime() * 1000.0;
4070 statsOnly( _interval_start_time_ms = _start_time_ms );
4071
4072 // If do_stealing is true then do_marking_step will attempt to
4073 // steal work from the other CMTasks. It only makes sense to
4074 // enable stealing when the termination protocol is enabled
4075 // and do_marking_step() is not being called serially.
4076 bool do_stealing = do_termination && !is_serial;
4077
4078 double diff_prediction_ms =
4079 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4080 _time_target_ms = time_target_ms - diff_prediction_ms;
4081
4082 // set up the variables that are used in the work-based scheme to
4083 // call the regular clock method
4084 _words_scanned = 0;
4085 _refs_reached = 0;
4086 recalculate_limits();
4087
4088 // clear all flags
4089 clear_has_aborted();
4090 _has_timed_out = false;
4091 _draining_satb_buffers = false;
4092
4093 ++_calls;
4094
4095 if (_cm->verbose_low()) {
4096 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4097 "target = %1.2lfms >>>>>>>>>>",
4098 _worker_id, _calls, _time_target_ms);
4099 }
4100
4101 // Set up the bitmap and oop closures. Anything that uses them is
4102 // eventually called from this method, so it is OK to allocate these
4103 // statically.
4104 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4105 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4106 set_cm_oop_closure(&cm_oop_closure);
4107
4108 if (_cm->has_overflown()) {
4109 // This can happen if the mark stack overflows during a GC pause
4110 // and this task, after a yield point, restarts. We have to abort
4111 // as we need to get into the overflow protocol which happens
4112 // right at the end of this task.
4113 set_has_aborted();
4114 }
4115
4116 // First drain any available SATB buffers. After this, we will not
4117 // look at SATB buffers before the next invocation of this method.
4118 // If enough completed SATB buffers are queued up, the regular clock
4119 // will abort this task so that it restarts.
4120 drain_satb_buffers();
4121 // ...then partially drain the local queue and the global stack
4122 drain_local_queue(true);
4123 drain_global_stack(true);
4124
4125 do {
4126 if (!has_aborted() && _curr_region != NULL) {
4127 // This means that we're already holding on to a region.
4128 assert(_finger != NULL, "if region is not NULL, then the finger "
4129 "should not be NULL either");
4130
4131 // We might have restarted this task after an evacuation pause
4132 // which might have evacuated the region we're holding on to
4133 // underneath our feet. Let's read its limit again to make sure
4134 // that we do not iterate over a region of the heap that
4135 // contains garbage (update_region_limit() will also move
4136 // _finger to the start of the region if it is found empty).
4137 update_region_limit();
4138 // We will start from _finger not from the start of the region,
4139 // as we might be restarting this task after aborting half-way
4140 // through scanning this region. In this case, _finger points to
4141 // the address where we last found a marked object. If this is a
4142 // fresh region, _finger points to start().
4143 MemRegion mr = MemRegion(_finger, _region_limit);
4144
4145 if (_cm->verbose_low()) {
4146 gclog_or_tty->print_cr("[%u] we're scanning part "
4147 "["PTR_FORMAT", "PTR_FORMAT") "
4148 "of region "HR_FORMAT,
4149 _worker_id, _finger, _region_limit,
4150 HR_FORMAT_PARAMS(_curr_region));
4151 }
4152
4153 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4154 "humongous regions should go around loop once only");
4155
4156 // Some special cases:
4157 // If the memory region is empty, we can just give up the region.
4158 // If the current region is humongous then we only need to check
4159 // the bitmap for the bit associated with the start of the object,
4160 // scan the object if it's live, and give up the region.
4161 // Otherwise, let's iterate over the bitmap of the part of the region
4162 // that is left.
4163 // If the iteration is successful, give up the region.
4164 if (mr.is_empty()) {
4165 giveup_current_region();
4166 regular_clock_call();
4167 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4168 if (_nextMarkBitMap->isMarked(mr.start())) {
4169 // The object is marked - apply the closure
4170 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4171 bitmap_closure.do_bit(offset);
4172 }
4173 // Even if this task aborted while scanning the humongous object
4174 // we can (and should) give up the current region.
4175 giveup_current_region();
4176 regular_clock_call();
4177 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4178 giveup_current_region();
4179 regular_clock_call();
4180 } else {
4181 assert(has_aborted(), "currently the only way to do so");
4182 // The only way to abort the bitmap iteration is to return
4183 // false from the do_bit() method. However, inside the
4184 // do_bit() method we move the _finger to point to the
4185 // object currently being looked at. So, if we bail out, we
4186 // have definitely set _finger to something non-null.
4187 assert(_finger != NULL, "invariant");
4188
4189 // Region iteration was actually aborted. So now _finger
4190 // points to the address of the object we last scanned. If we
4191 // leave it there, when we restart this task, we will rescan
4192 // the object. It is easy to avoid this. We move the finger by
4193 // enough to point to the next possible object header (the
4194 // bitmap knows by how much we need to move it as it knows its
4195 // granularity).
4196 assert(_finger < _region_limit, "invariant");
4197 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4198 // Check if bitmap iteration was aborted while scanning the last object
4199 if (new_finger >= _region_limit) {
4200 giveup_current_region();
4201 } else {
4202 move_finger_to(new_finger);
4203 }
4204 }
4205 }
4206 // At this point we have either completed iterating over the
4207 // region we were holding on to, or we have aborted.
4208
4209 // We then partially drain the local queue and the global stack.
4210 // (Do we really need this?)
4211 drain_local_queue(true);
4212 drain_global_stack(true);
4213
4214 // Read the note on the claim_region() method on why it might
4215 // return NULL with potentially more regions available for
4216 // claiming and why we have to check out_of_regions() to determine
4217 // whether we're done or not.
4218 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4219 // We are going to try to claim a new region. We should have
4220 // given up on the previous one.
4221 // Separated the asserts so that we know which one fires.
4222 assert(_curr_region == NULL, "invariant");
4223 assert(_finger == NULL, "invariant");
4224 assert(_region_limit == NULL, "invariant");
4225 if (_cm->verbose_low()) {
4226 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4227 }
4228 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4229 if (claimed_region != NULL) {
4230 // Yes, we managed to claim one
4231 statsOnly( ++_regions_claimed );
4232
4233 if (_cm->verbose_low()) {
4234 gclog_or_tty->print_cr("[%u] we successfully claimed "
4235 "region "PTR_FORMAT,
4236 _worker_id, claimed_region);
4237 }
4238
4239 setup_for_region(claimed_region);
4240 assert(_curr_region == claimed_region, "invariant");
4241 }
4242 // It is important to call the regular clock here. It might take
4243 // a while to claim a region if, for example, we hit a large
4244 // block of empty regions. So we need to call the regular clock
4245 // method once round the loop to make sure it's called
4246 // frequently enough.
4247 regular_clock_call();
4248 }
4249
4250 if (!has_aborted() && _curr_region == NULL) {
4251 assert(_cm->out_of_regions(),
4252 "at this point we should be out of regions");
4253 }
4254 } while ( _curr_region != NULL && !has_aborted());
4255
4256 if (!has_aborted()) {
4257 // We cannot check whether the global stack is empty, since other
4258 // tasks might be pushing objects to it concurrently.
4259 assert(_cm->out_of_regions(),
4260 "at this point we should be out of regions");
4261
4262 if (_cm->verbose_low()) {
4263 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4264 }
4265
4266 // Try to reduce the number of available SATB buffers so that
4267 // remark has less work to do.
4268 drain_satb_buffers();
4269 }
4270
4271 // Since we've done everything else, we can now totally drain the
4272 // local queue and global stack.
4273 drain_local_queue(false);
4274 drain_global_stack(false);
4275
4276 // Attempt at work stealing from other task's queues.
4277 if (do_stealing && !has_aborted()) {
4278 // We have not aborted. This means that we have finished all that
4279 // we could. Let's try to do some stealing...
4280
4281 // We cannot check whether the global stack is empty, since other
4282 // tasks might be pushing objects to it concurrently.
4283 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4284 "only way to reach here");
4285
4286 if (_cm->verbose_low()) {
4287 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4288 }
4289
4290 while (!has_aborted()) {
4291 oop obj;
4292 statsOnly( ++_steal_attempts );
4293
4294 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4295 if (_cm->verbose_medium()) {
4296 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4297 _worker_id, (void*) obj);
4298 }
4299
4300 statsOnly( ++_steals );
4301
4302 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4303 "any stolen object should be marked");
4304 scan_object(obj);
4305
4306 // And since we're towards the end, let's totally drain the
4307 // local queue and global stack.
4308 drain_local_queue(false);
4309 drain_global_stack(false);
4310 } else {
4311 break;
4312 }
4313 }
4314 }
4315
4316 // If we are about to wrap up and go into termination, check if we
4317 // should raise the overflow flag.
4318 if (do_termination && !has_aborted()) {
4319 if (_cm->force_overflow()->should_force()) {
4320 _cm->set_has_overflown();
4321 regular_clock_call();
4322 }
4323 }
4324
4325 // We still haven't aborted. Now, let's try to get into the
4326 // termination protocol.
4327 if (do_termination && !has_aborted()) {
4328 // We cannot check whether the global stack is empty, since other
4329 // tasks might be concurrently pushing objects on it.
4330 // Separated the asserts so that we know which one fires.
4331 assert(_cm->out_of_regions(), "only way to reach here");
4332 assert(_task_queue->size() == 0, "only way to reach here");
4333
4334 if (_cm->verbose_low()) {
4335 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4336 }
4337
4338 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4339
4340 // The CMTask class also extends the TerminatorTerminator class,
4341 // hence its should_exit_termination() method will also decide
4342 // whether to exit the termination protocol or not.
4343 bool finished = (is_serial ||
4344 _cm->terminator()->offer_termination(this));
4345 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4346 _termination_time_ms +=
4347 termination_end_time_ms - _termination_start_time_ms;
4348
4349 if (finished) {
4350 // We're all done.
4351
4352 if (_worker_id == 0) {
4353 // let's allow task 0 to do this
4354 if (concurrent()) {
4355 assert(_cm->concurrent_marking_in_progress(), "invariant");
4356 // we need to set this to false before the next
4357 // safepoint. This way we ensure that the marking phase
4358 // doesn't observe any more heap expansions.
4359 _cm->clear_concurrent_marking_in_progress();
4360 }
4361 }
4362
4363 // We can now guarantee that the global stack is empty, since
4364 // all other tasks have finished. We separated the guarantees so
4365 // that, if a condition is false, we can immediately find out
4366 // which one.
4367 guarantee(_cm->out_of_regions(), "only way to reach here");
4368 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4369 guarantee(_task_queue->size() == 0, "only way to reach here");
4370 guarantee(!_cm->has_overflown(), "only way to reach here");
4371 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4372
4373 if (_cm->verbose_low()) {
4374 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4375 }
4376 } else {
4377 // Apparently there's more work to do. Let's abort this task. It
4378 // will restart it and we can hopefully find more things to do.
4379
4380 if (_cm->verbose_low()) {
4381 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4382 _worker_id);
4383 }
4384
4385 set_has_aborted();
4386 statsOnly( ++_aborted_termination );
4387 }
4388 }
4389
4390 // Mainly for debugging purposes to make sure that a pointer to the
4391 // closure which was statically allocated in this frame doesn't
4392 // escape it by accident.
4393 set_cm_oop_closure(NULL);
4394 double end_time_ms = os::elapsedVTime() * 1000.0;
4395 double elapsed_time_ms = end_time_ms - _start_time_ms;
4396 // Update the step history.
4397 _step_times_ms.add(elapsed_time_ms);
4398
4399 if (has_aborted()) {
4400 // The task was aborted for some reason.
4401
4402 statsOnly( ++_aborted );
4403
4404 if (_has_timed_out) {
4405 double diff_ms = elapsed_time_ms - _time_target_ms;
4406 // Keep statistics of how well we did with respect to hitting
4407 // our target only if we actually timed out (if we aborted for
4408 // other reasons, then the results might get skewed).
4409 _marking_step_diffs_ms.add(diff_ms);
4410 }
4411
4412 if (_cm->has_overflown()) {
4413 // This is the interesting one. We aborted because a global
4414 // overflow was raised. This means we have to restart the
4415 // marking phase and start iterating over regions. However, in
4416 // order to do this we have to make sure that all tasks stop
4417 // what they are doing and re-initialise in a safe manner. We
4418 // will achieve this with the use of two barrier sync points.
4419
4420 if (_cm->verbose_low()) {
4421 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4422 }
4423
4424 if (!is_serial) {
4425 // We only need to enter the sync barrier if being called
4426 // from a parallel context
4427 _cm->enter_first_sync_barrier(_worker_id);
4428
4429 // When we exit this sync barrier we know that all tasks have
4430 // stopped doing marking work. So, it's now safe to
4431 // re-initialise our data structures. At the end of this method,
4432 // task 0 will clear the global data structures.
4433 }
4434
4435 statsOnly( ++_aborted_overflow );
4436
4437 // We clear the local state of this task...
4438 clear_region_fields();
4439
4440 if (!is_serial) {
4441 // ...and enter the second barrier.
4442 _cm->enter_second_sync_barrier(_worker_id);
4443 }
4444 // At this point, if we're during the concurrent phase of
4445 // marking, everything has been re-initialized and we're
4446 // ready to restart.
4447 }
4448
4449 if (_cm->verbose_low()) {
4450 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4451 "elapsed = %1.2lfms <<<<<<<<<<",
4452 _worker_id, _time_target_ms, elapsed_time_ms);
4453 if (_cm->has_aborted()) {
4454 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4455 _worker_id);
4456 }
4457 }
4458 } else {
4459 if (_cm->verbose_low()) {
4460 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4461 "elapsed = %1.2lfms <<<<<<<<<<",
4462 _worker_id, _time_target_ms, elapsed_time_ms);
4463 }
4464 }
4465
4466 _claimed = false;
4467 }
4468
4469 CMTask::CMTask(uint worker_id,
4470 ConcurrentMark* cm,
4471 size_t* marked_bytes,
4472 BitMap* card_bm,
4473 CMTaskQueue* task_queue,
4474 CMTaskQueueSet* task_queues)
4475 : _g1h(G1CollectedHeap::heap()),
4476 _worker_id(worker_id), _cm(cm),
4477 _claimed(false),
4478 _nextMarkBitMap(NULL), _hash_seed(17),
4479 _task_queue(task_queue),
4480 _task_queues(task_queues),
4481 _cm_oop_closure(NULL),
4482 _marked_bytes_array(marked_bytes),
4483 _card_bm(card_bm) {
4484 guarantee(task_queue != NULL, "invariant");
4485 guarantee(task_queues != NULL, "invariant");
4486
4487 statsOnly( _clock_due_to_scanning = 0;
4488 _clock_due_to_marking = 0 );
4489
4490 _marking_step_diffs_ms.add(0.5);
4491 }
4492
4493 // These are formatting macros that are used below to ensure
4494 // consistent formatting. The *_H_* versions are used to format the
4495 // header for a particular value and they should be kept consistent
4496 // with the corresponding macro. Also note that most of the macros add
4497 // the necessary white space (as a prefix) which makes them a bit
4498 // easier to compose.
4499
4500 // All the output lines are prefixed with this string to be able to
4501 // identify them easily in a large log file.
4502 #define G1PPRL_LINE_PREFIX "###"
4503
4504 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4505 #ifdef _LP64
4506 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4507 #else // _LP64
4508 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4509 #endif // _LP64
4510
4511 // For per-region info
4512 #define G1PPRL_TYPE_FORMAT " %-4s"
4513 #define G1PPRL_TYPE_H_FORMAT " %4s"
4514 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4515 #define G1PPRL_BYTE_H_FORMAT " %9s"
4516 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4517 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4518
4519 // For summary info
4520 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4521 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4522 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4523 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4524
4525 G1PrintRegionLivenessInfoClosure::
4526 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4527 : _out(out),
4528 _total_used_bytes(0), _total_capacity_bytes(0),
4529 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4530 _hum_used_bytes(0), _hum_capacity_bytes(0),
4531 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4532 _total_remset_bytes(0) {
4533 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4534 MemRegion g1_committed = g1h->g1_committed();
4535 MemRegion g1_reserved = g1h->g1_reserved();
4536 double now = os::elapsedTime();
4537
4538 // Print the header of the output.
4539 _out->cr();
4540 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4541 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4542 G1PPRL_SUM_ADDR_FORMAT("committed")
4543 G1PPRL_SUM_ADDR_FORMAT("reserved")
4544 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4545 g1_committed.start(), g1_committed.end(),
4546 g1_reserved.start(), g1_reserved.end(),
4547 HeapRegion::GrainBytes);
4548 _out->print_cr(G1PPRL_LINE_PREFIX);
4549 _out->print_cr(G1PPRL_LINE_PREFIX
4550 G1PPRL_TYPE_H_FORMAT
4551 G1PPRL_ADDR_BASE_H_FORMAT
4552 G1PPRL_BYTE_H_FORMAT
4553 G1PPRL_BYTE_H_FORMAT
4554 G1PPRL_BYTE_H_FORMAT
4555 G1PPRL_DOUBLE_H_FORMAT
4556 G1PPRL_BYTE_H_FORMAT,
4557 "type", "address-range",
4558 "used", "prev-live", "next-live", "gc-eff", "remset");
4559 _out->print_cr(G1PPRL_LINE_PREFIX
4560 G1PPRL_TYPE_H_FORMAT
4561 G1PPRL_ADDR_BASE_H_FORMAT
4562 G1PPRL_BYTE_H_FORMAT
4563 G1PPRL_BYTE_H_FORMAT
4564 G1PPRL_BYTE_H_FORMAT
4565 G1PPRL_DOUBLE_H_FORMAT
4566 G1PPRL_BYTE_H_FORMAT,
4567 "", "",
4568 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)", "(bytes)");
4569 }
4570
4571 // It takes as a parameter a reference to one of the _hum_* fields, it
4572 // deduces the corresponding value for a region in a humongous region
4573 // series (either the region size, or what's left if the _hum_* field
4574 // is < the region size), and updates the _hum_* field accordingly.
4575 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4576 size_t bytes = 0;
4577 // The > 0 check is to deal with the prev and next live bytes which
4578 // could be 0.
4579 if (*hum_bytes > 0) {
4580 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4581 *hum_bytes -= bytes;
4582 }
4583 return bytes;
4584 }
4585
4586 // It deduces the values for a region in a humongous region series
4587 // from the _hum_* fields and updates those accordingly. It assumes
4588 // that that _hum_* fields have already been set up from the "starts
4589 // humongous" region and we visit the regions in address order.
4590 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4591 size_t* capacity_bytes,
4592 size_t* prev_live_bytes,
4593 size_t* next_live_bytes) {
4594 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4595 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4596 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4597 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4598 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4599 }
4600
4601 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4602 const char* type = "";
4603 HeapWord* bottom = r->bottom();
4604 HeapWord* end = r->end();
4605 size_t capacity_bytes = r->capacity();
4606 size_t used_bytes = r->used();
4607 size_t prev_live_bytes = r->live_bytes();
4608 size_t next_live_bytes = r->next_live_bytes();
4609 double gc_eff = r->gc_efficiency();
4610 size_t remset_bytes = r->rem_set()->mem_size();
4611 if (r->used() == 0) {
4612 type = "FREE";
4613 } else if (r->is_survivor()) {
4614 type = "SURV";
4615 } else if (r->is_young()) {
4616 type = "EDEN";
4617 } else if (r->startsHumongous()) {
4618 type = "HUMS";
4619
4620 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4621 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4622 "they should have been zeroed after the last time we used them");
4623 // Set up the _hum_* fields.
4624 _hum_capacity_bytes = capacity_bytes;
4625 _hum_used_bytes = used_bytes;
4626 _hum_prev_live_bytes = prev_live_bytes;
4627 _hum_next_live_bytes = next_live_bytes;
4628 get_hum_bytes(&used_bytes, &capacity_bytes,
4629 &prev_live_bytes, &next_live_bytes);
4630 end = bottom + HeapRegion::GrainWords;
4631 } else if (r->continuesHumongous()) {
4632 type = "HUMC";
4633 get_hum_bytes(&used_bytes, &capacity_bytes,
4634 &prev_live_bytes, &next_live_bytes);
4635 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4636 } else {
4637 type = "OLD";
4638 }
4639
4640 _total_used_bytes += used_bytes;
4641 _total_capacity_bytes += capacity_bytes;
4642 _total_prev_live_bytes += prev_live_bytes;
4643 _total_next_live_bytes += next_live_bytes;
4644 _total_remset_bytes += remset_bytes;
4645
4646 // Print a line for this particular region.
4647 _out->print_cr(G1PPRL_LINE_PREFIX
4648 G1PPRL_TYPE_FORMAT
4649 G1PPRL_ADDR_BASE_FORMAT
4650 G1PPRL_BYTE_FORMAT
4651 G1PPRL_BYTE_FORMAT
4652 G1PPRL_BYTE_FORMAT
4653 G1PPRL_DOUBLE_FORMAT
4654 G1PPRL_BYTE_FORMAT,
4655 type, bottom, end,
4656 used_bytes, prev_live_bytes, next_live_bytes, gc_eff , remset_bytes);
4657
4658 return false;
4659 }
4660
4661 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4662 // add static memory usages to remembered set sizes
4663 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4664 // Print the footer of the output.
4665 _out->print_cr(G1PPRL_LINE_PREFIX);
4666 _out->print_cr(G1PPRL_LINE_PREFIX
4667 " SUMMARY"
4668 G1PPRL_SUM_MB_FORMAT("capacity")
4669 G1PPRL_SUM_MB_PERC_FORMAT("used")
4670 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4671 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4672 G1PPRL_SUM_MB_FORMAT("remset"),
4673 bytes_to_mb(_total_capacity_bytes),
4674 bytes_to_mb(_total_used_bytes),
4675 perc(_total_used_bytes, _total_capacity_bytes),
4676 bytes_to_mb(_total_prev_live_bytes),
4677 perc(_total_prev_live_bytes, _total_capacity_bytes),
4678 bytes_to_mb(_total_next_live_bytes),
4679 perc(_total_next_live_bytes, _total_capacity_bytes),
4680 bytes_to_mb(_total_remset_bytes));
4681 _out->cr();
4682 }