1 /*
2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "gc_implementation/g1/concurrentMark.inline.hpp"
28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
32 #include "gc_implementation/g1/g1Log.hpp"
33 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
34 #include "gc_implementation/g1/g1RemSet.hpp"
35 #include "gc_implementation/g1/heapRegion.inline.hpp"
36 #include "gc_implementation/g1/heapRegionRemSet.hpp"
37 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
38 #include "gc_implementation/shared/vmGCOperations.hpp"
39 #include "gc_implementation/shared/gcTimer.hpp"
40 #include "gc_implementation/shared/gcTrace.hpp"
41 #include "gc_implementation/shared/gcTraceTime.hpp"
42 #include "memory/genOopClosures.inline.hpp"
43 #include "memory/referencePolicy.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "oops/oop.inline.hpp"
46 #include "runtime/handles.inline.hpp"
47 #include "runtime/java.hpp"
48 #include "runtime/atomic.inline.hpp"
49 #include "runtime/prefetch.inline.hpp"
50 #include "services/memTracker.hpp"
51
52 // Concurrent marking bit map wrapper
53
54 CMBitMapRO::CMBitMapRO(int shifter) :
55 _bm(),
56 _shifter(shifter) {
57 _bmStartWord = 0;
58 _bmWordSize = 0;
59 }
60
61 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr,
62 HeapWord* limit) const {
63 // First we must round addr *up* to a possible object boundary.
64 addr = (HeapWord*)align_size_up((intptr_t)addr,
65 HeapWordSize << _shifter);
66 size_t addrOffset = heapWordToOffset(addr);
67 if (limit == NULL) {
68 limit = _bmStartWord + _bmWordSize;
69 }
70 size_t limitOffset = heapWordToOffset(limit);
71 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
72 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
73 assert(nextAddr >= addr, "get_next_one postcondition");
74 assert(nextAddr == limit || isMarked(nextAddr),
75 "get_next_one postcondition");
76 return nextAddr;
77 }
78
79 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr,
80 HeapWord* limit) const {
81 size_t addrOffset = heapWordToOffset(addr);
82 if (limit == NULL) {
83 limit = _bmStartWord + _bmWordSize;
84 }
85 size_t limitOffset = heapWordToOffset(limit);
86 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
87 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
88 assert(nextAddr >= addr, "get_next_one postcondition");
89 assert(nextAddr == limit || !isMarked(nextAddr),
90 "get_next_one postcondition");
91 return nextAddr;
92 }
93
94 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
95 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
96 return (int) (diff >> _shifter);
97 }
98
99 #ifndef PRODUCT
100 bool CMBitMapRO::covers(ReservedSpace heap_rs) const {
101 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
102 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
103 "size inconsistency");
104 return _bmStartWord == (HeapWord*)(heap_rs.base()) &&
105 _bmWordSize == heap_rs.size()>>LogHeapWordSize;
106 }
107 #endif
108
109 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
110 _bm.print_on_error(st, prefix);
111 }
112
113 bool CMBitMap::allocate(ReservedSpace heap_rs) {
114 _bmStartWord = (HeapWord*)(heap_rs.base());
115 _bmWordSize = heap_rs.size()/HeapWordSize; // heap_rs.size() is in bytes
116 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
117 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
118 if (!brs.is_reserved()) {
119 warning("ConcurrentMark marking bit map allocation failure");
120 return false;
121 }
122 MemTracker::record_virtual_memory_type((address)brs.base(), mtGC);
123 // For now we'll just commit all of the bit map up front.
124 // Later on we'll try to be more parsimonious with swap.
125 if (!_virtual_space.initialize(brs, brs.size())) {
126 warning("ConcurrentMark marking bit map backing store failure");
127 return false;
128 }
129 assert(_virtual_space.committed_size() == brs.size(),
130 "didn't reserve backing store for all of concurrent marking bit map?");
131 _bm.set_map((uintptr_t*)_virtual_space.low());
132 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
133 _bmWordSize, "inconsistency in bit map sizing");
134 _bm.set_size(_bmWordSize >> _shifter);
135 return true;
136 }
137
138 void CMBitMap::clearAll() {
139 _bm.clear();
140 return;
141 }
142
143 void CMBitMap::markRange(MemRegion mr) {
144 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
145 assert(!mr.is_empty(), "unexpected empty region");
146 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
147 ((HeapWord *) mr.end())),
148 "markRange memory region end is not card aligned");
149 // convert address range into offset range
150 _bm.at_put_range(heapWordToOffset(mr.start()),
151 heapWordToOffset(mr.end()), true);
152 }
153
154 void CMBitMap::clearRange(MemRegion mr) {
155 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
156 assert(!mr.is_empty(), "unexpected empty region");
157 // convert address range into offset range
158 _bm.at_put_range(heapWordToOffset(mr.start()),
159 heapWordToOffset(mr.end()), false);
160 }
161
162 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
163 HeapWord* end_addr) {
164 HeapWord* start = getNextMarkedWordAddress(addr);
165 start = MIN2(start, end_addr);
166 HeapWord* end = getNextUnmarkedWordAddress(start);
167 end = MIN2(end, end_addr);
168 assert(start <= end, "Consistency check");
169 MemRegion mr(start, end);
170 if (!mr.is_empty()) {
171 clearRange(mr);
172 }
173 return mr;
174 }
175
176 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
177 _base(NULL), _cm(cm)
178 #ifdef ASSERT
179 , _drain_in_progress(false)
180 , _drain_in_progress_yields(false)
181 #endif
182 {}
183
184 bool CMMarkStack::allocate(size_t capacity) {
185 // allocate a stack of the requisite depth
186 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
187 if (!rs.is_reserved()) {
188 warning("ConcurrentMark MarkStack allocation failure");
189 return false;
190 }
191 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
192 if (!_virtual_space.initialize(rs, rs.size())) {
193 warning("ConcurrentMark MarkStack backing store failure");
194 // Release the virtual memory reserved for the marking stack
195 rs.release();
196 return false;
197 }
198 assert(_virtual_space.committed_size() == rs.size(),
199 "Didn't reserve backing store for all of ConcurrentMark stack?");
200 _base = (oop*) _virtual_space.low();
201 setEmpty();
202 _capacity = (jint) capacity;
203 _saved_index = -1;
204 _should_expand = false;
205 NOT_PRODUCT(_max_depth = 0);
206 return true;
207 }
208
209 void CMMarkStack::expand() {
210 // Called, during remark, if we've overflown the marking stack during marking.
211 assert(isEmpty(), "stack should been emptied while handling overflow");
212 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
213 // Clear expansion flag
214 _should_expand = false;
215 if (_capacity == (jint) MarkStackSizeMax) {
216 if (PrintGCDetails && Verbose) {
217 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
218 }
219 return;
220 }
221 // Double capacity if possible
222 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
223 // Do not give up existing stack until we have managed to
224 // get the double capacity that we desired.
225 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
226 sizeof(oop)));
227 if (rs.is_reserved()) {
228 // Release the backing store associated with old stack
229 _virtual_space.release();
230 // Reinitialize virtual space for new stack
231 if (!_virtual_space.initialize(rs, rs.size())) {
232 fatal("Not enough swap for expanded marking stack capacity");
233 }
234 _base = (oop*)(_virtual_space.low());
235 _index = 0;
236 _capacity = new_capacity;
237 } else {
238 if (PrintGCDetails && Verbose) {
239 // Failed to double capacity, continue;
240 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
241 SIZE_FORMAT"K to " SIZE_FORMAT"K",
242 _capacity / K, new_capacity / K);
243 }
244 }
245 }
246
247 void CMMarkStack::set_should_expand() {
248 // If we're resetting the marking state because of an
249 // marking stack overflow, record that we should, if
250 // possible, expand the stack.
251 _should_expand = _cm->has_overflown();
252 }
253
254 CMMarkStack::~CMMarkStack() {
255 if (_base != NULL) {
256 _base = NULL;
257 _virtual_space.release();
258 }
259 }
260
261 void CMMarkStack::par_push(oop ptr) {
262 while (true) {
263 if (isFull()) {
264 _overflow = true;
265 return;
266 }
267 // Otherwise...
268 jint index = _index;
269 jint next_index = index+1;
270 jint res = Atomic::cmpxchg(next_index, &_index, index);
271 if (res == index) {
272 _base[index] = ptr;
273 // Note that we don't maintain this atomically. We could, but it
274 // doesn't seem necessary.
275 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
276 return;
277 }
278 // Otherwise, we need to try again.
279 }
280 }
281
282 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
283 while (true) {
284 if (isFull()) {
285 _overflow = true;
286 return;
287 }
288 // Otherwise...
289 jint index = _index;
290 jint next_index = index + n;
291 if (next_index > _capacity) {
292 _overflow = true;
293 return;
294 }
295 jint res = Atomic::cmpxchg(next_index, &_index, index);
296 if (res == index) {
297 for (int i = 0; i < n; i++) {
298 int ind = index + i;
299 assert(ind < _capacity, "By overflow test above.");
300 _base[ind] = ptr_arr[i];
301 }
302 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
303 return;
304 }
305 // Otherwise, we need to try again.
306 }
307 }
308
309 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
310 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
311 jint start = _index;
312 jint next_index = start + n;
313 if (next_index > _capacity) {
314 _overflow = true;
315 return;
316 }
317 // Otherwise.
318 _index = next_index;
319 for (int i = 0; i < n; i++) {
320 int ind = start + i;
321 assert(ind < _capacity, "By overflow test above.");
322 _base[ind] = ptr_arr[i];
323 }
324 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
325 }
326
327 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
328 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
329 jint index = _index;
330 if (index == 0) {
331 *n = 0;
332 return false;
333 } else {
334 int k = MIN2(max, index);
335 jint new_ind = index - k;
336 for (int j = 0; j < k; j++) {
337 ptr_arr[j] = _base[new_ind + j];
338 }
339 _index = new_ind;
340 *n = k;
341 return true;
342 }
343 }
344
345 template<class OopClosureClass>
346 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
347 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
348 || SafepointSynchronize::is_at_safepoint(),
349 "Drain recursion must be yield-safe.");
350 bool res = true;
351 debug_only(_drain_in_progress = true);
352 debug_only(_drain_in_progress_yields = yield_after);
353 while (!isEmpty()) {
354 oop newOop = pop();
355 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
356 assert(newOop->is_oop(), "Expected an oop");
357 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
358 "only grey objects on this stack");
359 newOop->oop_iterate(cl);
360 if (yield_after && _cm->do_yield_check()) {
361 res = false;
362 break;
363 }
364 }
365 debug_only(_drain_in_progress = false);
366 return res;
367 }
368
369 void CMMarkStack::note_start_of_gc() {
370 assert(_saved_index == -1,
371 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
372 _saved_index = _index;
373 }
374
375 void CMMarkStack::note_end_of_gc() {
376 // This is intentionally a guarantee, instead of an assert. If we
377 // accidentally add something to the mark stack during GC, it
378 // will be a correctness issue so it's better if we crash. we'll
379 // only check this once per GC anyway, so it won't be a performance
380 // issue in any way.
381 guarantee(_saved_index == _index,
382 err_msg("saved index: %d index: %d", _saved_index, _index));
383 _saved_index = -1;
384 }
385
386 void CMMarkStack::oops_do(OopClosure* f) {
387 assert(_saved_index == _index,
388 err_msg("saved index: %d index: %d", _saved_index, _index));
389 for (int i = 0; i < _index; i += 1) {
390 f->do_oop(&_base[i]);
391 }
392 }
393
394 bool ConcurrentMark::not_yet_marked(oop obj) const {
395 return _g1h->is_obj_ill(obj);
396 }
397
398 CMRootRegions::CMRootRegions() :
399 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
400 _should_abort(false), _next_survivor(NULL) { }
401
402 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
403 _young_list = g1h->young_list();
404 _cm = cm;
405 }
406
407 void CMRootRegions::prepare_for_scan() {
408 assert(!scan_in_progress(), "pre-condition");
409
410 // Currently, only survivors can be root regions.
411 assert(_next_survivor == NULL, "pre-condition");
412 _next_survivor = _young_list->first_survivor_region();
413 _scan_in_progress = (_next_survivor != NULL);
414 _should_abort = false;
415 }
416
417 HeapRegion* CMRootRegions::claim_next() {
418 if (_should_abort) {
419 // If someone has set the should_abort flag, we return NULL to
420 // force the caller to bail out of their loop.
421 return NULL;
422 }
423
424 // Currently, only survivors can be root regions.
425 HeapRegion* res = _next_survivor;
426 if (res != NULL) {
427 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
428 // Read it again in case it changed while we were waiting for the lock.
429 res = _next_survivor;
430 if (res != NULL) {
431 if (res == _young_list->last_survivor_region()) {
432 // We just claimed the last survivor so store NULL to indicate
433 // that we're done.
434 _next_survivor = NULL;
435 } else {
436 _next_survivor = res->get_next_young_region();
437 }
438 } else {
439 // Someone else claimed the last survivor while we were trying
440 // to take the lock so nothing else to do.
441 }
442 }
443 assert(res == NULL || res->is_survivor(), "post-condition");
444
445 return res;
446 }
447
448 void CMRootRegions::scan_finished() {
449 assert(scan_in_progress(), "pre-condition");
450
451 // Currently, only survivors can be root regions.
452 if (!_should_abort) {
453 assert(_next_survivor == NULL, "we should have claimed all survivors");
454 }
455 _next_survivor = NULL;
456
457 {
458 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
459 _scan_in_progress = false;
460 RootRegionScan_lock->notify_all();
461 }
462 }
463
464 bool CMRootRegions::wait_until_scan_finished() {
465 if (!scan_in_progress()) return false;
466
467 {
468 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
469 while (scan_in_progress()) {
470 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
471 }
472 }
473 return true;
474 }
475
476 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
477 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
478 #endif // _MSC_VER
479
480 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
481 return MAX2((n_par_threads + 2) / 4, 1U);
482 }
483
484 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs) :
485 _g1h(g1h),
486 _markBitMap1(log2_intptr(MinObjAlignment)),
487 _markBitMap2(log2_intptr(MinObjAlignment)),
488 _parallel_marking_threads(0),
489 _max_parallel_marking_threads(0),
490 _sleep_factor(0.0),
491 _marking_task_overhead(1.0),
492 _cleanup_sleep_factor(0.0),
493 _cleanup_task_overhead(1.0),
494 _cleanup_list("Cleanup List"),
495 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
496 _card_bm((heap_rs.size() + CardTableModRefBS::card_size - 1) >>
497 CardTableModRefBS::card_shift,
498 false /* in_resource_area*/),
499
500 _prevMarkBitMap(&_markBitMap1),
501 _nextMarkBitMap(&_markBitMap2),
502
503 _markStack(this),
504 // _finger set in set_non_marking_state
505
506 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
507 // _active_tasks set in set_non_marking_state
508 // _tasks set inside the constructor
509 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
510 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
511
512 _has_overflown(false),
513 _concurrent(false),
514 _has_aborted(false),
515 _restart_for_overflow(false),
516 _concurrent_marking_in_progress(false),
517
518 // _verbose_level set below
519
520 _init_times(),
521 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
522 _cleanup_times(),
523 _total_counting_time(0.0),
524 _total_rs_scrub_time(0.0),
525
526 _parallel_workers(NULL),
527
528 _count_card_bitmaps(NULL),
529 _count_marked_bytes(NULL),
530 _completed_initialization(false) {
531 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
532 if (verbose_level < no_verbose) {
533 verbose_level = no_verbose;
534 }
535 if (verbose_level > high_verbose) {
536 verbose_level = high_verbose;
537 }
538 _verbose_level = verbose_level;
539
540 if (verbose_low()) {
541 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
542 "heap end = " PTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
543 }
544
545 if (!_markBitMap1.allocate(heap_rs)) {
546 warning("Failed to allocate first CM bit map");
547 return;
548 }
549 if (!_markBitMap2.allocate(heap_rs)) {
550 warning("Failed to allocate second CM bit map");
551 return;
552 }
553
554 // Create & start a ConcurrentMark thread.
555 _cmThread = new ConcurrentMarkThread(this);
556 assert(cmThread() != NULL, "CM Thread should have been created");
557 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
558 if (_cmThread->osthread() == NULL) {
559 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
560 }
561
562 assert(CGC_lock != NULL, "Where's the CGC_lock?");
563 assert(_markBitMap1.covers(heap_rs), "_markBitMap1 inconsistency");
564 assert(_markBitMap2.covers(heap_rs), "_markBitMap2 inconsistency");
565
566 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
567 satb_qs.set_buffer_size(G1SATBBufferSize);
568
569 _root_regions.init(_g1h, this);
570
571 if (ConcGCThreads > ParallelGCThreads) {
572 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
573 "than ParallelGCThreads (" UINTX_FORMAT ").",
574 ConcGCThreads, ParallelGCThreads);
575 return;
576 }
577 if (ParallelGCThreads == 0) {
578 // if we are not running with any parallel GC threads we will not
579 // spawn any marking threads either
580 _parallel_marking_threads = 0;
581 _max_parallel_marking_threads = 0;
582 _sleep_factor = 0.0;
583 _marking_task_overhead = 1.0;
584 } else {
585 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
586 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
587 // if both are set
588 _sleep_factor = 0.0;
589 _marking_task_overhead = 1.0;
590 } else if (G1MarkingOverheadPercent > 0) {
591 // We will calculate the number of parallel marking threads based
592 // on a target overhead with respect to the soft real-time goal
593 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
594 double overall_cm_overhead =
595 (double) MaxGCPauseMillis * marking_overhead /
596 (double) GCPauseIntervalMillis;
597 double cpu_ratio = 1.0 / (double) os::processor_count();
598 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
599 double marking_task_overhead =
600 overall_cm_overhead / marking_thread_num *
601 (double) os::processor_count();
602 double sleep_factor =
603 (1.0 - marking_task_overhead) / marking_task_overhead;
604
605 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
606 _sleep_factor = sleep_factor;
607 _marking_task_overhead = marking_task_overhead;
608 } else {
609 // Calculate the number of parallel marking threads by scaling
610 // the number of parallel GC threads.
611 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
612 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
613 _sleep_factor = 0.0;
614 _marking_task_overhead = 1.0;
615 }
616
617 assert(ConcGCThreads > 0, "Should have been set");
618 _parallel_marking_threads = (uint) ConcGCThreads;
619 _max_parallel_marking_threads = _parallel_marking_threads;
620
621 if (parallel_marking_threads() > 1) {
622 _cleanup_task_overhead = 1.0;
623 } else {
624 _cleanup_task_overhead = marking_task_overhead();
625 }
626 _cleanup_sleep_factor =
627 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
628
629 #if 0
630 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
631 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
632 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
633 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
634 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
635 #endif
636
637 guarantee(parallel_marking_threads() > 0, "peace of mind");
638 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
639 _max_parallel_marking_threads, false, true);
640 if (_parallel_workers == NULL) {
641 vm_exit_during_initialization("Failed necessary allocation.");
642 } else {
643 _parallel_workers->initialize_workers();
644 }
645 }
646
647 if (FLAG_IS_DEFAULT(MarkStackSize)) {
648 uintx mark_stack_size =
649 MIN2(MarkStackSizeMax,
650 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
651 // Verify that the calculated value for MarkStackSize is in range.
652 // It would be nice to use the private utility routine from Arguments.
653 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
654 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
655 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
656 mark_stack_size, (uintx) 1, MarkStackSizeMax);
657 return;
658 }
659 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
660 } else {
661 // Verify MarkStackSize is in range.
662 if (FLAG_IS_CMDLINE(MarkStackSize)) {
663 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
664 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
665 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
666 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
667 MarkStackSize, (uintx) 1, MarkStackSizeMax);
668 return;
669 }
670 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
671 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
672 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
673 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
674 MarkStackSize, MarkStackSizeMax);
675 return;
676 }
677 }
678 }
679 }
680
681 if (!_markStack.allocate(MarkStackSize)) {
682 warning("Failed to allocate CM marking stack");
683 return;
684 }
685
686 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
687 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
688
689 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
690 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
691
692 BitMap::idx_t card_bm_size = _card_bm.size();
693
694 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
695 _active_tasks = _max_worker_id;
696
697 size_t max_regions = (size_t) _g1h->max_regions();
698 for (uint i = 0; i < _max_worker_id; ++i) {
699 CMTaskQueue* task_queue = new CMTaskQueue();
700 task_queue->initialize();
701 _task_queues->register_queue(i, task_queue);
702
703 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
704 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
705
706 _tasks[i] = new CMTask(i, this,
707 _count_marked_bytes[i],
708 &_count_card_bitmaps[i],
709 task_queue, _task_queues);
710
711 _accum_task_vtime[i] = 0.0;
712 }
713
714 // Calculate the card number for the bottom of the heap. Used
715 // in biasing indexes into the accounting card bitmaps.
716 _heap_bottom_card_num =
717 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
718 CardTableModRefBS::card_shift);
719
720 // Clear all the liveness counting data
721 clear_all_count_data();
722
723 // so that the call below can read a sensible value
724 _heap_start = (HeapWord*) heap_rs.base();
725 set_non_marking_state();
726 _completed_initialization = true;
727 }
728
729 void ConcurrentMark::update_g1_committed(bool force) {
730 // If concurrent marking is not in progress, then we do not need to
731 // update _heap_end.
732 if (!concurrent_marking_in_progress() && !force) return;
733
734 MemRegion committed = _g1h->g1_committed();
735 assert(committed.start() == _heap_start, "start shouldn't change");
736 HeapWord* new_end = committed.end();
737 if (new_end > _heap_end) {
738 // The heap has been expanded.
739
740 _heap_end = new_end;
741 }
742 // Notice that the heap can also shrink. However, this only happens
743 // during a Full GC (at least currently) and the entire marking
744 // phase will bail out and the task will not be restarted. So, let's
745 // do nothing.
746 }
747
748 void ConcurrentMark::reset() {
749 // Starting values for these two. This should be called in a STW
750 // phase. CM will be notified of any future g1_committed expansions
751 // will be at the end of evacuation pauses, when tasks are
752 // inactive.
753 MemRegion committed = _g1h->g1_committed();
754 _heap_start = committed.start();
755 _heap_end = committed.end();
756
757 // Separated the asserts so that we know which one fires.
758 assert(_heap_start != NULL, "heap bounds should look ok");
759 assert(_heap_end != NULL, "heap bounds should look ok");
760 assert(_heap_start < _heap_end, "heap bounds should look ok");
761
762 // Reset all the marking data structures and any necessary flags
763 reset_marking_state();
764
765 if (verbose_low()) {
766 gclog_or_tty->print_cr("[global] resetting");
767 }
768
769 // We do reset all of them, since different phases will use
770 // different number of active threads. So, it's easiest to have all
771 // of them ready.
772 for (uint i = 0; i < _max_worker_id; ++i) {
773 _tasks[i]->reset(_nextMarkBitMap);
774 }
775
776 // we need this to make sure that the flag is on during the evac
777 // pause with initial mark piggy-backed
778 set_concurrent_marking_in_progress();
779 }
780
781
782 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
783 _markStack.set_should_expand();
784 _markStack.setEmpty(); // Also clears the _markStack overflow flag
785 if (clear_overflow) {
786 clear_has_overflown();
787 } else {
788 assert(has_overflown(), "pre-condition");
789 }
790 _finger = _heap_start;
791
792 for (uint i = 0; i < _max_worker_id; ++i) {
793 CMTaskQueue* queue = _task_queues->queue(i);
794 queue->set_empty();
795 }
796 }
797
798 void ConcurrentMark::set_concurrency(uint active_tasks) {
799 assert(active_tasks <= _max_worker_id, "we should not have more");
800
801 _active_tasks = active_tasks;
802 // Need to update the three data structures below according to the
803 // number of active threads for this phase.
804 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
805 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
806 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
807 }
808
809 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
810 set_concurrency(active_tasks);
811
812 _concurrent = concurrent;
813 // We propagate this to all tasks, not just the active ones.
814 for (uint i = 0; i < _max_worker_id; ++i)
815 _tasks[i]->set_concurrent(concurrent);
816
817 if (concurrent) {
818 set_concurrent_marking_in_progress();
819 } else {
820 // We currently assume that the concurrent flag has been set to
821 // false before we start remark. At this point we should also be
822 // in a STW phase.
823 assert(!concurrent_marking_in_progress(), "invariant");
824 assert(out_of_regions(),
825 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
826 p2i(_finger), p2i(_heap_end)));
827 update_g1_committed(true);
828 }
829 }
830
831 void ConcurrentMark::set_non_marking_state() {
832 // We set the global marking state to some default values when we're
833 // not doing marking.
834 reset_marking_state();
835 _active_tasks = 0;
836 clear_concurrent_marking_in_progress();
837 }
838
839 ConcurrentMark::~ConcurrentMark() {
840 // The ConcurrentMark instance is never freed.
841 ShouldNotReachHere();
842 }
843
844 void ConcurrentMark::clearNextBitmap() {
845 G1CollectedHeap* g1h = G1CollectedHeap::heap();
846 G1CollectorPolicy* g1p = g1h->g1_policy();
847
848 // Make sure that the concurrent mark thread looks to still be in
849 // the current cycle.
850 guarantee(cmThread()->during_cycle(), "invariant");
851
852 // We are finishing up the current cycle by clearing the next
853 // marking bitmap and getting it ready for the next cycle. During
854 // this time no other cycle can start. So, let's make sure that this
855 // is the case.
856 guarantee(!g1h->mark_in_progress(), "invariant");
857
858 // clear the mark bitmap (no grey objects to start with).
859 // We need to do this in chunks and offer to yield in between
860 // each chunk.
861 HeapWord* start = _nextMarkBitMap->startWord();
862 HeapWord* end = _nextMarkBitMap->endWord();
863 HeapWord* cur = start;
864 size_t chunkSize = M;
865 while (cur < end) {
866 HeapWord* next = cur + chunkSize;
867 if (next > end) {
868 next = end;
869 }
870 MemRegion mr(cur,next);
871 _nextMarkBitMap->clearRange(mr);
872 cur = next;
873 do_yield_check();
874
875 // Repeat the asserts from above. We'll do them as asserts here to
876 // minimize their overhead on the product. However, we'll have
877 // them as guarantees at the beginning / end of the bitmap
878 // clearing to get some checking in the product.
879 assert(cmThread()->during_cycle(), "invariant");
880 assert(!g1h->mark_in_progress(), "invariant");
881 }
882
883 // Clear the liveness counting data
884 clear_all_count_data();
885
886 // Repeat the asserts from above.
887 guarantee(cmThread()->during_cycle(), "invariant");
888 guarantee(!g1h->mark_in_progress(), "invariant");
889 }
890
891 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
892 public:
893 bool doHeapRegion(HeapRegion* r) {
894 if (!r->continuesHumongous()) {
895 r->note_start_of_marking();
896 }
897 return false;
898 }
899 };
900
901 void ConcurrentMark::checkpointRootsInitialPre() {
902 G1CollectedHeap* g1h = G1CollectedHeap::heap();
903 G1CollectorPolicy* g1p = g1h->g1_policy();
904
905 _has_aborted = false;
906
907 #ifndef PRODUCT
908 if (G1PrintReachableAtInitialMark) {
909 print_reachable("at-cycle-start",
910 VerifyOption_G1UsePrevMarking, true /* all */);
911 }
912 #endif
913
914 // Initialize marking structures. This has to be done in a STW phase.
915 reset();
916
917 // For each region note start of marking.
918 NoteStartOfMarkHRClosure startcl;
919 g1h->heap_region_iterate(&startcl);
920 }
921
922
923 void ConcurrentMark::checkpointRootsInitialPost() {
924 G1CollectedHeap* g1h = G1CollectedHeap::heap();
925
926 // If we force an overflow during remark, the remark operation will
927 // actually abort and we'll restart concurrent marking. If we always
928 // force an overflow during remark we'll never actually complete the
929 // marking phase. So, we initialize this here, at the start of the
930 // cycle, so that at the remaining overflow number will decrease at
931 // every remark and we'll eventually not need to cause one.
932 force_overflow_stw()->init();
933
934 // Start Concurrent Marking weak-reference discovery.
935 ReferenceProcessor* rp = g1h->ref_processor_cm();
936 // enable ("weak") refs discovery
937 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
938 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
939
940 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
941 // This is the start of the marking cycle, we're expected all
942 // threads to have SATB queues with active set to false.
943 satb_mq_set.set_active_all_threads(true, /* new active value */
944 false /* expected_active */);
945
946 _root_regions.prepare_for_scan();
947
948 // update_g1_committed() will be called at the end of an evac pause
949 // when marking is on. So, it's also called at the end of the
950 // initial-mark pause to update the heap end, if the heap expands
951 // during it. No need to call it here.
952 }
953
954 /*
955 * Notice that in the next two methods, we actually leave the STS
956 * during the barrier sync and join it immediately afterwards. If we
957 * do not do this, the following deadlock can occur: one thread could
958 * be in the barrier sync code, waiting for the other thread to also
959 * sync up, whereas another one could be trying to yield, while also
960 * waiting for the other threads to sync up too.
961 *
962 * Note, however, that this code is also used during remark and in
963 * this case we should not attempt to leave / enter the STS, otherwise
964 * we'll either hit an assert (debug / fastdebug) or deadlock
965 * (product). So we should only leave / enter the STS if we are
966 * operating concurrently.
967 *
968 * Because the thread that does the sync barrier has left the STS, it
969 * is possible to be suspended for a Full GC or an evacuation pause
970 * could occur. This is actually safe, since the entering the sync
971 * barrier is one of the last things do_marking_step() does, and it
972 * doesn't manipulate any data structures afterwards.
973 */
974
975 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
976 if (verbose_low()) {
977 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
978 }
979
980 if (concurrent()) {
981 SuspendibleThreadSet::leave();
982 }
983
984 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
985
986 if (concurrent()) {
987 SuspendibleThreadSet::join();
988 }
989 // at this point everyone should have synced up and not be doing any
990 // more work
991
992 if (verbose_low()) {
993 if (barrier_aborted) {
994 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
995 } else {
996 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
997 }
998 }
999
1000 if (barrier_aborted) {
1001 // If the barrier aborted we ignore the overflow condition and
1002 // just abort the whole marking phase as quickly as possible.
1003 return;
1004 }
1005
1006 // If we're executing the concurrent phase of marking, reset the marking
1007 // state; otherwise the marking state is reset after reference processing,
1008 // during the remark pause.
1009 // If we reset here as a result of an overflow during the remark we will
1010 // see assertion failures from any subsequent set_concurrency_and_phase()
1011 // calls.
1012 if (concurrent()) {
1013 // let the task associated with with worker 0 do this
1014 if (worker_id == 0) {
1015 // task 0 is responsible for clearing the global data structures
1016 // We should be here because of an overflow. During STW we should
1017 // not clear the overflow flag since we rely on it being true when
1018 // we exit this method to abort the pause and restart concurrent
1019 // marking.
1020 reset_marking_state(true /* clear_overflow */);
1021 force_overflow()->update();
1022
1023 if (G1Log::fine()) {
1024 gclog_or_tty->date_stamp(PrintGCDateStamps);
1025 gclog_or_tty->stamp(PrintGCTimeStamps);
1026 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1027 }
1028 }
1029 }
1030
1031 // after this, each task should reset its own data structures then
1032 // then go into the second barrier
1033 }
1034
1035 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1036 if (verbose_low()) {
1037 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1038 }
1039
1040 if (concurrent()) {
1041 SuspendibleThreadSet::leave();
1042 }
1043
1044 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1045
1046 if (concurrent()) {
1047 SuspendibleThreadSet::join();
1048 }
1049 // at this point everything should be re-initialized and ready to go
1050
1051 if (verbose_low()) {
1052 if (barrier_aborted) {
1053 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1054 } else {
1055 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1056 }
1057 }
1058 }
1059
1060 #ifndef PRODUCT
1061 void ForceOverflowSettings::init() {
1062 _num_remaining = G1ConcMarkForceOverflow;
1063 _force = false;
1064 update();
1065 }
1066
1067 void ForceOverflowSettings::update() {
1068 if (_num_remaining > 0) {
1069 _num_remaining -= 1;
1070 _force = true;
1071 } else {
1072 _force = false;
1073 }
1074 }
1075
1076 bool ForceOverflowSettings::should_force() {
1077 if (_force) {
1078 _force = false;
1079 return true;
1080 } else {
1081 return false;
1082 }
1083 }
1084 #endif // !PRODUCT
1085
1086 class CMConcurrentMarkingTask: public AbstractGangTask {
1087 private:
1088 ConcurrentMark* _cm;
1089 ConcurrentMarkThread* _cmt;
1090
1091 public:
1092 void work(uint worker_id) {
1093 assert(Thread::current()->is_ConcurrentGC_thread(),
1094 "this should only be done by a conc GC thread");
1095 ResourceMark rm;
1096
1097 double start_vtime = os::elapsedVTime();
1098
1099 SuspendibleThreadSet::join();
1100
1101 assert(worker_id < _cm->active_tasks(), "invariant");
1102 CMTask* the_task = _cm->task(worker_id);
1103 the_task->record_start_time();
1104 if (!_cm->has_aborted()) {
1105 do {
1106 double start_vtime_sec = os::elapsedVTime();
1107 double start_time_sec = os::elapsedTime();
1108 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1109
1110 the_task->do_marking_step(mark_step_duration_ms,
1111 true /* do_termination */,
1112 false /* is_serial*/);
1113
1114 double end_time_sec = os::elapsedTime();
1115 double end_vtime_sec = os::elapsedVTime();
1116 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1117 double elapsed_time_sec = end_time_sec - start_time_sec;
1118 _cm->clear_has_overflown();
1119
1120 bool ret = _cm->do_yield_check(worker_id);
1121
1122 jlong sleep_time_ms;
1123 if (!_cm->has_aborted() && the_task->has_aborted()) {
1124 sleep_time_ms =
1125 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1126 SuspendibleThreadSet::leave();
1127 os::sleep(Thread::current(), sleep_time_ms, false);
1128 SuspendibleThreadSet::join();
1129 }
1130 double end_time2_sec = os::elapsedTime();
1131 double elapsed_time2_sec = end_time2_sec - start_time_sec;
1132
1133 #if 0
1134 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, "
1135 "overhead %1.4lf",
1136 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms,
1137 the_task->conc_overhead(os::elapsedTime()) * 8.0);
1138 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms",
1139 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0);
1140 #endif
1141 } while (!_cm->has_aborted() && the_task->has_aborted());
1142 }
1143 the_task->record_end_time();
1144 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1145
1146 SuspendibleThreadSet::leave();
1147
1148 double end_vtime = os::elapsedVTime();
1149 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1150 }
1151
1152 CMConcurrentMarkingTask(ConcurrentMark* cm,
1153 ConcurrentMarkThread* cmt) :
1154 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1155
1156 ~CMConcurrentMarkingTask() { }
1157 };
1158
1159 // Calculates the number of active workers for a concurrent
1160 // phase.
1161 uint ConcurrentMark::calc_parallel_marking_threads() {
1162 if (G1CollectedHeap::use_parallel_gc_threads()) {
1163 uint n_conc_workers = 0;
1164 if (!UseDynamicNumberOfGCThreads ||
1165 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1166 !ForceDynamicNumberOfGCThreads)) {
1167 n_conc_workers = max_parallel_marking_threads();
1168 } else {
1169 n_conc_workers =
1170 AdaptiveSizePolicy::calc_default_active_workers(
1171 max_parallel_marking_threads(),
1172 1, /* Minimum workers */
1173 parallel_marking_threads(),
1174 Threads::number_of_non_daemon_threads());
1175 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1176 // that scaling has already gone into "_max_parallel_marking_threads".
1177 }
1178 assert(n_conc_workers > 0, "Always need at least 1");
1179 return n_conc_workers;
1180 }
1181 // If we are not running with any parallel GC threads we will not
1182 // have spawned any marking threads either. Hence the number of
1183 // concurrent workers should be 0.
1184 return 0;
1185 }
1186
1187 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1188 // Currently, only survivors can be root regions.
1189 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1190 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1191
1192 const uintx interval = PrefetchScanIntervalInBytes;
1193 HeapWord* curr = hr->bottom();
1194 const HeapWord* end = hr->top();
1195 while (curr < end) {
1196 Prefetch::read(curr, interval);
1197 oop obj = oop(curr);
1198 int size = obj->oop_iterate(&cl);
1199 assert(size == obj->size(), "sanity");
1200 curr += size;
1201 }
1202 }
1203
1204 class CMRootRegionScanTask : public AbstractGangTask {
1205 private:
1206 ConcurrentMark* _cm;
1207
1208 public:
1209 CMRootRegionScanTask(ConcurrentMark* cm) :
1210 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1211
1212 void work(uint worker_id) {
1213 assert(Thread::current()->is_ConcurrentGC_thread(),
1214 "this should only be done by a conc GC thread");
1215
1216 CMRootRegions* root_regions = _cm->root_regions();
1217 HeapRegion* hr = root_regions->claim_next();
1218 while (hr != NULL) {
1219 _cm->scanRootRegion(hr, worker_id);
1220 hr = root_regions->claim_next();
1221 }
1222 }
1223 };
1224
1225 void ConcurrentMark::scanRootRegions() {
1226 // scan_in_progress() will have been set to true only if there was
1227 // at least one root region to scan. So, if it's false, we
1228 // should not attempt to do any further work.
1229 if (root_regions()->scan_in_progress()) {
1230 _parallel_marking_threads = calc_parallel_marking_threads();
1231 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1232 "Maximum number of marking threads exceeded");
1233 uint active_workers = MAX2(1U, parallel_marking_threads());
1234
1235 CMRootRegionScanTask task(this);
1236 if (use_parallel_marking_threads()) {
1237 _parallel_workers->set_active_workers((int) active_workers);
1238 _parallel_workers->run_task(&task);
1239 } else {
1240 task.work(0);
1241 }
1242
1243 // It's possible that has_aborted() is true here without actually
1244 // aborting the survivor scan earlier. This is OK as it's
1245 // mainly used for sanity checking.
1246 root_regions()->scan_finished();
1247 }
1248 }
1249
1250 void ConcurrentMark::markFromRoots() {
1251 // we might be tempted to assert that:
1252 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1253 // "inconsistent argument?");
1254 // However that wouldn't be right, because it's possible that
1255 // a safepoint is indeed in progress as a younger generation
1256 // stop-the-world GC happens even as we mark in this generation.
1257
1258 _restart_for_overflow = false;
1259 force_overflow_conc()->init();
1260
1261 // _g1h has _n_par_threads
1262 _parallel_marking_threads = calc_parallel_marking_threads();
1263 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1264 "Maximum number of marking threads exceeded");
1265
1266 uint active_workers = MAX2(1U, parallel_marking_threads());
1267
1268 // Parallel task terminator is set in "set_concurrency_and_phase()"
1269 set_concurrency_and_phase(active_workers, true /* concurrent */);
1270
1271 CMConcurrentMarkingTask markingTask(this, cmThread());
1272 if (use_parallel_marking_threads()) {
1273 _parallel_workers->set_active_workers((int)active_workers);
1274 // Don't set _n_par_threads because it affects MT in process_strong_roots()
1275 // and the decisions on that MT processing is made elsewhere.
1276 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1277 _parallel_workers->run_task(&markingTask);
1278 } else {
1279 markingTask.work(0);
1280 }
1281 print_stats();
1282 }
1283
1284 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1285 // world is stopped at this checkpoint
1286 assert(SafepointSynchronize::is_at_safepoint(),
1287 "world should be stopped");
1288
1289 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1290
1291 // If a full collection has happened, we shouldn't do this.
1292 if (has_aborted()) {
1293 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1294 return;
1295 }
1296
1297 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1298
1299 if (VerifyDuringGC) {
1300 HandleMark hm; // handle scope
1301 Universe::heap()->prepare_for_verify();
1302 Universe::verify(VerifyOption_G1UsePrevMarking,
1303 " VerifyDuringGC:(before)");
1304 }
1305 g1h->check_bitmaps("Remark Start");
1306
1307 G1CollectorPolicy* g1p = g1h->g1_policy();
1308 g1p->record_concurrent_mark_remark_start();
1309
1310 double start = os::elapsedTime();
1311
1312 checkpointRootsFinalWork();
1313
1314 double mark_work_end = os::elapsedTime();
1315
1316 weakRefsWork(clear_all_soft_refs);
1317
1318 if (has_overflown()) {
1319 // Oops. We overflowed. Restart concurrent marking.
1320 _restart_for_overflow = true;
1321 if (G1TraceMarkStackOverflow) {
1322 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1323 }
1324
1325 // Verify the heap w.r.t. the previous marking bitmap.
1326 if (VerifyDuringGC) {
1327 HandleMark hm; // handle scope
1328 Universe::heap()->prepare_for_verify();
1329 Universe::verify(VerifyOption_G1UsePrevMarking,
1330 " VerifyDuringGC:(overflow)");
1331 }
1332
1333 // Clear the marking state because we will be restarting
1334 // marking due to overflowing the global mark stack.
1335 reset_marking_state();
1336 } else {
1337 // Aggregate the per-task counting data that we have accumulated
1338 // while marking.
1339 aggregate_count_data();
1340
1341 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1342 // We're done with marking.
1343 // This is the end of the marking cycle, we're expected all
1344 // threads to have SATB queues with active set to true.
1345 satb_mq_set.set_active_all_threads(false, /* new active value */
1346 true /* expected_active */);
1347
1348 if (VerifyDuringGC) {
1349 HandleMark hm; // handle scope
1350 Universe::heap()->prepare_for_verify();
1351 Universe::verify(VerifyOption_G1UseNextMarking,
1352 " VerifyDuringGC:(after)");
1353 }
1354 g1h->check_bitmaps("Remark End");
1355 assert(!restart_for_overflow(), "sanity");
1356 // Completely reset the marking state since marking completed
1357 set_non_marking_state();
1358 }
1359
1360 // Expand the marking stack, if we have to and if we can.
1361 if (_markStack.should_expand()) {
1362 _markStack.expand();
1363 }
1364
1365 // Statistics
1366 double now = os::elapsedTime();
1367 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1368 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1369 _remark_times.add((now - start) * 1000.0);
1370
1371 g1p->record_concurrent_mark_remark_end();
1372
1373 G1CMIsAliveClosure is_alive(g1h);
1374 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1375 }
1376
1377 // Base class of the closures that finalize and verify the
1378 // liveness counting data.
1379 class CMCountDataClosureBase: public HeapRegionClosure {
1380 protected:
1381 G1CollectedHeap* _g1h;
1382 ConcurrentMark* _cm;
1383 CardTableModRefBS* _ct_bs;
1384
1385 BitMap* _region_bm;
1386 BitMap* _card_bm;
1387
1388 // Takes a region that's not empty (i.e., it has at least one
1389 // live object in it and sets its corresponding bit on the region
1390 // bitmap to 1. If the region is "starts humongous" it will also set
1391 // to 1 the bits on the region bitmap that correspond to its
1392 // associated "continues humongous" regions.
1393 void set_bit_for_region(HeapRegion* hr) {
1394 assert(!hr->continuesHumongous(), "should have filtered those out");
1395
1396 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1397 if (!hr->startsHumongous()) {
1398 // Normal (non-humongous) case: just set the bit.
1399 _region_bm->par_at_put(index, true);
1400 } else {
1401 // Starts humongous case: calculate how many regions are part of
1402 // this humongous region and then set the bit range.
1403 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1404 _region_bm->par_at_put_range(index, end_index, true);
1405 }
1406 }
1407
1408 public:
1409 CMCountDataClosureBase(G1CollectedHeap* g1h,
1410 BitMap* region_bm, BitMap* card_bm):
1411 _g1h(g1h), _cm(g1h->concurrent_mark()),
1412 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1413 _region_bm(region_bm), _card_bm(card_bm) { }
1414 };
1415
1416 // Closure that calculates the # live objects per region. Used
1417 // for verification purposes during the cleanup pause.
1418 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1419 CMBitMapRO* _bm;
1420 size_t _region_marked_bytes;
1421
1422 public:
1423 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1424 BitMap* region_bm, BitMap* card_bm) :
1425 CMCountDataClosureBase(g1h, region_bm, card_bm),
1426 _bm(bm), _region_marked_bytes(0) { }
1427
1428 bool doHeapRegion(HeapRegion* hr) {
1429
1430 if (hr->continuesHumongous()) {
1431 // We will ignore these here and process them when their
1432 // associated "starts humongous" region is processed (see
1433 // set_bit_for_heap_region()). Note that we cannot rely on their
1434 // associated "starts humongous" region to have their bit set to
1435 // 1 since, due to the region chunking in the parallel region
1436 // iteration, a "continues humongous" region might be visited
1437 // before its associated "starts humongous".
1438 return false;
1439 }
1440
1441 HeapWord* ntams = hr->next_top_at_mark_start();
1442 HeapWord* start = hr->bottom();
1443
1444 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1445 err_msg("Preconditions not met - "
1446 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1447 p2i(start), p2i(ntams), p2i(hr->end())));
1448
1449 // Find the first marked object at or after "start".
1450 start = _bm->getNextMarkedWordAddress(start, ntams);
1451
1452 size_t marked_bytes = 0;
1453
1454 while (start < ntams) {
1455 oop obj = oop(start);
1456 int obj_sz = obj->size();
1457 HeapWord* obj_end = start + obj_sz;
1458
1459 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1460 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1461
1462 // Note: if we're looking at the last region in heap - obj_end
1463 // could be actually just beyond the end of the heap; end_idx
1464 // will then correspond to a (non-existent) card that is also
1465 // just beyond the heap.
1466 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1467 // end of object is not card aligned - increment to cover
1468 // all the cards spanned by the object
1469 end_idx += 1;
1470 }
1471
1472 // Set the bits in the card BM for the cards spanned by this object.
1473 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1474
1475 // Add the size of this object to the number of marked bytes.
1476 marked_bytes += (size_t)obj_sz * HeapWordSize;
1477
1478 // Find the next marked object after this one.
1479 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1480 }
1481
1482 // Mark the allocated-since-marking portion...
1483 HeapWord* top = hr->top();
1484 if (ntams < top) {
1485 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1486 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1487
1488 // Note: if we're looking at the last region in heap - top
1489 // could be actually just beyond the end of the heap; end_idx
1490 // will then correspond to a (non-existent) card that is also
1491 // just beyond the heap.
1492 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1493 // end of object is not card aligned - increment to cover
1494 // all the cards spanned by the object
1495 end_idx += 1;
1496 }
1497 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1498
1499 // This definitely means the region has live objects.
1500 set_bit_for_region(hr);
1501 }
1502
1503 // Update the live region bitmap.
1504 if (marked_bytes > 0) {
1505 set_bit_for_region(hr);
1506 }
1507
1508 // Set the marked bytes for the current region so that
1509 // it can be queried by a calling verification routine
1510 _region_marked_bytes = marked_bytes;
1511
1512 return false;
1513 }
1514
1515 size_t region_marked_bytes() const { return _region_marked_bytes; }
1516 };
1517
1518 // Heap region closure used for verifying the counting data
1519 // that was accumulated concurrently and aggregated during
1520 // the remark pause. This closure is applied to the heap
1521 // regions during the STW cleanup pause.
1522
1523 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1524 G1CollectedHeap* _g1h;
1525 ConcurrentMark* _cm;
1526 CalcLiveObjectsClosure _calc_cl;
1527 BitMap* _region_bm; // Region BM to be verified
1528 BitMap* _card_bm; // Card BM to be verified
1529 bool _verbose; // verbose output?
1530
1531 BitMap* _exp_region_bm; // Expected Region BM values
1532 BitMap* _exp_card_bm; // Expected card BM values
1533
1534 int _failures;
1535
1536 public:
1537 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1538 BitMap* region_bm,
1539 BitMap* card_bm,
1540 BitMap* exp_region_bm,
1541 BitMap* exp_card_bm,
1542 bool verbose) :
1543 _g1h(g1h), _cm(g1h->concurrent_mark()),
1544 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1545 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1546 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1547 _failures(0) { }
1548
1549 int failures() const { return _failures; }
1550
1551 bool doHeapRegion(HeapRegion* hr) {
1552 if (hr->continuesHumongous()) {
1553 // We will ignore these here and process them when their
1554 // associated "starts humongous" region is processed (see
1555 // set_bit_for_heap_region()). Note that we cannot rely on their
1556 // associated "starts humongous" region to have their bit set to
1557 // 1 since, due to the region chunking in the parallel region
1558 // iteration, a "continues humongous" region might be visited
1559 // before its associated "starts humongous".
1560 return false;
1561 }
1562
1563 int failures = 0;
1564
1565 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1566 // this region and set the corresponding bits in the expected region
1567 // and card bitmaps.
1568 bool res = _calc_cl.doHeapRegion(hr);
1569 assert(res == false, "should be continuing");
1570
1571 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1572 Mutex::_no_safepoint_check_flag);
1573
1574 // Verify the marked bytes for this region.
1575 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1576 size_t act_marked_bytes = hr->next_marked_bytes();
1577
1578 // We're not OK if expected marked bytes > actual marked bytes. It means
1579 // we have missed accounting some objects during the actual marking.
1580 if (exp_marked_bytes > act_marked_bytes) {
1581 if (_verbose) {
1582 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1583 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1584 hr->hrs_index(), exp_marked_bytes, act_marked_bytes);
1585 }
1586 failures += 1;
1587 }
1588
1589 // Verify the bit, for this region, in the actual and expected
1590 // (which was just calculated) region bit maps.
1591 // We're not OK if the bit in the calculated expected region
1592 // bitmap is set and the bit in the actual region bitmap is not.
1593 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1594
1595 bool expected = _exp_region_bm->at(index);
1596 bool actual = _region_bm->at(index);
1597 if (expected && !actual) {
1598 if (_verbose) {
1599 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1600 "expected: %s, actual: %s",
1601 hr->hrs_index(),
1602 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1603 }
1604 failures += 1;
1605 }
1606
1607 // Verify that the card bit maps for the cards spanned by the current
1608 // region match. We have an error if we have a set bit in the expected
1609 // bit map and the corresponding bit in the actual bitmap is not set.
1610
1611 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1612 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1613
1614 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1615 expected = _exp_card_bm->at(i);
1616 actual = _card_bm->at(i);
1617
1618 if (expected && !actual) {
1619 if (_verbose) {
1620 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1621 "expected: %s, actual: %s",
1622 hr->hrs_index(), i,
1623 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1624 }
1625 failures += 1;
1626 }
1627 }
1628
1629 if (failures > 0 && _verbose) {
1630 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1631 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1632 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1633 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1634 }
1635
1636 _failures += failures;
1637
1638 // We could stop iteration over the heap when we
1639 // find the first violating region by returning true.
1640 return false;
1641 }
1642 };
1643
1644 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1645 protected:
1646 G1CollectedHeap* _g1h;
1647 ConcurrentMark* _cm;
1648 BitMap* _actual_region_bm;
1649 BitMap* _actual_card_bm;
1650
1651 uint _n_workers;
1652
1653 BitMap* _expected_region_bm;
1654 BitMap* _expected_card_bm;
1655
1656 int _failures;
1657 bool _verbose;
1658
1659 public:
1660 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1661 BitMap* region_bm, BitMap* card_bm,
1662 BitMap* expected_region_bm, BitMap* expected_card_bm)
1663 : AbstractGangTask("G1 verify final counting"),
1664 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1665 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1666 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1667 _failures(0), _verbose(false),
1668 _n_workers(0) {
1669 assert(VerifyDuringGC, "don't call this otherwise");
1670
1671 // Use the value already set as the number of active threads
1672 // in the call to run_task().
1673 if (G1CollectedHeap::use_parallel_gc_threads()) {
1674 assert( _g1h->workers()->active_workers() > 0,
1675 "Should have been previously set");
1676 _n_workers = _g1h->workers()->active_workers();
1677 } else {
1678 _n_workers = 1;
1679 }
1680
1681 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1682 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1683
1684 _verbose = _cm->verbose_medium();
1685 }
1686
1687 void work(uint worker_id) {
1688 assert(worker_id < _n_workers, "invariant");
1689
1690 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1691 _actual_region_bm, _actual_card_bm,
1692 _expected_region_bm,
1693 _expected_card_bm,
1694 _verbose);
1695
1696 if (G1CollectedHeap::use_parallel_gc_threads()) {
1697 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1698 worker_id,
1699 _n_workers,
1700 HeapRegion::VerifyCountClaimValue);
1701 } else {
1702 _g1h->heap_region_iterate(&verify_cl);
1703 }
1704
1705 Atomic::add(verify_cl.failures(), &_failures);
1706 }
1707
1708 int failures() const { return _failures; }
1709 };
1710
1711 // Closure that finalizes the liveness counting data.
1712 // Used during the cleanup pause.
1713 // Sets the bits corresponding to the interval [NTAMS, top]
1714 // (which contains the implicitly live objects) in the
1715 // card liveness bitmap. Also sets the bit for each region,
1716 // containing live data, in the region liveness bitmap.
1717
1718 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1719 public:
1720 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1721 BitMap* region_bm,
1722 BitMap* card_bm) :
1723 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1724
1725 bool doHeapRegion(HeapRegion* hr) {
1726
1727 if (hr->continuesHumongous()) {
1728 // We will ignore these here and process them when their
1729 // associated "starts humongous" region is processed (see
1730 // set_bit_for_heap_region()). Note that we cannot rely on their
1731 // associated "starts humongous" region to have their bit set to
1732 // 1 since, due to the region chunking in the parallel region
1733 // iteration, a "continues humongous" region might be visited
1734 // before its associated "starts humongous".
1735 return false;
1736 }
1737
1738 HeapWord* ntams = hr->next_top_at_mark_start();
1739 HeapWord* top = hr->top();
1740
1741 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1742
1743 // Mark the allocated-since-marking portion...
1744 if (ntams < top) {
1745 // This definitely means the region has live objects.
1746 set_bit_for_region(hr);
1747
1748 // Now set the bits in the card bitmap for [ntams, top)
1749 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1750 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1751
1752 // Note: if we're looking at the last region in heap - top
1753 // could be actually just beyond the end of the heap; end_idx
1754 // will then correspond to a (non-existent) card that is also
1755 // just beyond the heap.
1756 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1757 // end of object is not card aligned - increment to cover
1758 // all the cards spanned by the object
1759 end_idx += 1;
1760 }
1761
1762 assert(end_idx <= _card_bm->size(),
1763 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1764 end_idx, _card_bm->size()));
1765 assert(start_idx < _card_bm->size(),
1766 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1767 start_idx, _card_bm->size()));
1768
1769 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1770 }
1771
1772 // Set the bit for the region if it contains live data
1773 if (hr->next_marked_bytes() > 0) {
1774 set_bit_for_region(hr);
1775 }
1776
1777 return false;
1778 }
1779 };
1780
1781 class G1ParFinalCountTask: public AbstractGangTask {
1782 protected:
1783 G1CollectedHeap* _g1h;
1784 ConcurrentMark* _cm;
1785 BitMap* _actual_region_bm;
1786 BitMap* _actual_card_bm;
1787
1788 uint _n_workers;
1789
1790 public:
1791 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1792 : AbstractGangTask("G1 final counting"),
1793 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1794 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1795 _n_workers(0) {
1796 // Use the value already set as the number of active threads
1797 // in the call to run_task().
1798 if (G1CollectedHeap::use_parallel_gc_threads()) {
1799 assert( _g1h->workers()->active_workers() > 0,
1800 "Should have been previously set");
1801 _n_workers = _g1h->workers()->active_workers();
1802 } else {
1803 _n_workers = 1;
1804 }
1805 }
1806
1807 void work(uint worker_id) {
1808 assert(worker_id < _n_workers, "invariant");
1809
1810 FinalCountDataUpdateClosure final_update_cl(_g1h,
1811 _actual_region_bm,
1812 _actual_card_bm);
1813
1814 if (G1CollectedHeap::use_parallel_gc_threads()) {
1815 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1816 worker_id,
1817 _n_workers,
1818 HeapRegion::FinalCountClaimValue);
1819 } else {
1820 _g1h->heap_region_iterate(&final_update_cl);
1821 }
1822 }
1823 };
1824
1825 class G1ParNoteEndTask;
1826
1827 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1828 G1CollectedHeap* _g1;
1829 size_t _max_live_bytes;
1830 uint _regions_claimed;
1831 size_t _freed_bytes;
1832 FreeRegionList* _local_cleanup_list;
1833 HeapRegionSetCount _old_regions_removed;
1834 HeapRegionSetCount _humongous_regions_removed;
1835 HRRSCleanupTask* _hrrs_cleanup_task;
1836 double _claimed_region_time;
1837 double _max_region_time;
1838
1839 public:
1840 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1841 FreeRegionList* local_cleanup_list,
1842 HRRSCleanupTask* hrrs_cleanup_task) :
1843 _g1(g1),
1844 _max_live_bytes(0), _regions_claimed(0),
1845 _freed_bytes(0),
1846 _claimed_region_time(0.0), _max_region_time(0.0),
1847 _local_cleanup_list(local_cleanup_list),
1848 _old_regions_removed(),
1849 _humongous_regions_removed(),
1850 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1851
1852 size_t freed_bytes() { return _freed_bytes; }
1853 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1854 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1855
1856 bool doHeapRegion(HeapRegion *hr) {
1857 if (hr->continuesHumongous()) {
1858 return false;
1859 }
1860 // We use a claim value of zero here because all regions
1861 // were claimed with value 1 in the FinalCount task.
1862 _g1->reset_gc_time_stamps(hr);
1863 double start = os::elapsedTime();
1864 _regions_claimed++;
1865 hr->note_end_of_marking();
1866 _max_live_bytes += hr->max_live_bytes();
1867
1868 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1869 _freed_bytes += hr->used();
1870 hr->set_containing_set(NULL);
1871 if (hr->isHumongous()) {
1872 assert(hr->startsHumongous(), "we should only see starts humongous");
1873 _humongous_regions_removed.increment(1u, hr->capacity());
1874 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1875 } else {
1876 _old_regions_removed.increment(1u, hr->capacity());
1877 _g1->free_region(hr, _local_cleanup_list, true);
1878 }
1879 } else {
1880 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1881 }
1882
1883 double region_time = (os::elapsedTime() - start);
1884 _claimed_region_time += region_time;
1885 if (region_time > _max_region_time) {
1886 _max_region_time = region_time;
1887 }
1888 return false;
1889 }
1890
1891 size_t max_live_bytes() { return _max_live_bytes; }
1892 uint regions_claimed() { return _regions_claimed; }
1893 double claimed_region_time_sec() { return _claimed_region_time; }
1894 double max_region_time_sec() { return _max_region_time; }
1895 };
1896
1897 class G1ParNoteEndTask: public AbstractGangTask {
1898 friend class G1NoteEndOfConcMarkClosure;
1899
1900 protected:
1901 G1CollectedHeap* _g1h;
1902 size_t _max_live_bytes;
1903 size_t _freed_bytes;
1904 FreeRegionList* _cleanup_list;
1905
1906 public:
1907 G1ParNoteEndTask(G1CollectedHeap* g1h,
1908 FreeRegionList* cleanup_list) :
1909 AbstractGangTask("G1 note end"), _g1h(g1h),
1910 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1911
1912 void work(uint worker_id) {
1913 double start = os::elapsedTime();
1914 FreeRegionList local_cleanup_list("Local Cleanup List");
1915 HRRSCleanupTask hrrs_cleanup_task;
1916 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1917 &hrrs_cleanup_task);
1918 if (G1CollectedHeap::use_parallel_gc_threads()) {
1919 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1920 _g1h->workers()->active_workers(),
1921 HeapRegion::NoteEndClaimValue);
1922 } else {
1923 _g1h->heap_region_iterate(&g1_note_end);
1924 }
1925 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1926
1927 // Now update the lists
1928 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1929 {
1930 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1931 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1932 _max_live_bytes += g1_note_end.max_live_bytes();
1933 _freed_bytes += g1_note_end.freed_bytes();
1934
1935 // If we iterate over the global cleanup list at the end of
1936 // cleanup to do this printing we will not guarantee to only
1937 // generate output for the newly-reclaimed regions (the list
1938 // might not be empty at the beginning of cleanup; we might
1939 // still be working on its previous contents). So we do the
1940 // printing here, before we append the new regions to the global
1941 // cleanup list.
1942
1943 G1HRPrinter* hr_printer = _g1h->hr_printer();
1944 if (hr_printer->is_active()) {
1945 FreeRegionListIterator iter(&local_cleanup_list);
1946 while (iter.more_available()) {
1947 HeapRegion* hr = iter.get_next();
1948 hr_printer->cleanup(hr);
1949 }
1950 }
1951
1952 _cleanup_list->add_ordered(&local_cleanup_list);
1953 assert(local_cleanup_list.is_empty(), "post-condition");
1954
1955 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1956 }
1957 }
1958 size_t max_live_bytes() { return _max_live_bytes; }
1959 size_t freed_bytes() { return _freed_bytes; }
1960 };
1961
1962 class G1ParScrubRemSetTask: public AbstractGangTask {
1963 protected:
1964 G1RemSet* _g1rs;
1965 BitMap* _region_bm;
1966 BitMap* _card_bm;
1967 public:
1968 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1969 BitMap* region_bm, BitMap* card_bm) :
1970 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1971 _region_bm(region_bm), _card_bm(card_bm) { }
1972
1973 void work(uint worker_id) {
1974 if (G1CollectedHeap::use_parallel_gc_threads()) {
1975 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1976 HeapRegion::ScrubRemSetClaimValue);
1977 } else {
1978 _g1rs->scrub(_region_bm, _card_bm);
1979 }
1980 }
1981
1982 };
1983
1984 void ConcurrentMark::cleanup() {
1985 // world is stopped at this checkpoint
1986 assert(SafepointSynchronize::is_at_safepoint(),
1987 "world should be stopped");
1988 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1989
1990 // If a full collection has happened, we shouldn't do this.
1991 if (has_aborted()) {
1992 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1993 return;
1994 }
1995
1996 g1h->verify_region_sets_optional();
1997
1998 if (VerifyDuringGC) {
1999 HandleMark hm; // handle scope
2000 Universe::heap()->prepare_for_verify();
2001 Universe::verify(VerifyOption_G1UsePrevMarking,
2002 " VerifyDuringGC:(before)");
2003 }
2004 g1h->check_bitmaps("Cleanup Start");
2005
2006 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
2007 g1p->record_concurrent_mark_cleanup_start();
2008
2009 double start = os::elapsedTime();
2010
2011 HeapRegionRemSet::reset_for_cleanup_tasks();
2012
2013 uint n_workers;
2014
2015 // Do counting once more with the world stopped for good measure.
2016 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2017
2018 if (G1CollectedHeap::use_parallel_gc_threads()) {
2019 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2020 "sanity check");
2021
2022 g1h->set_par_threads();
2023 n_workers = g1h->n_par_threads();
2024 assert(g1h->n_par_threads() == n_workers,
2025 "Should not have been reset");
2026 g1h->workers()->run_task(&g1_par_count_task);
2027 // Done with the parallel phase so reset to 0.
2028 g1h->set_par_threads(0);
2029
2030 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2031 "sanity check");
2032 } else {
2033 n_workers = 1;
2034 g1_par_count_task.work(0);
2035 }
2036
2037 if (VerifyDuringGC) {
2038 // Verify that the counting data accumulated during marking matches
2039 // that calculated by walking the marking bitmap.
2040
2041 // Bitmaps to hold expected values
2042 BitMap expected_region_bm(_region_bm.size(), true);
2043 BitMap expected_card_bm(_card_bm.size(), true);
2044
2045 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2046 &_region_bm,
2047 &_card_bm,
2048 &expected_region_bm,
2049 &expected_card_bm);
2050
2051 if (G1CollectedHeap::use_parallel_gc_threads()) {
2052 g1h->set_par_threads((int)n_workers);
2053 g1h->workers()->run_task(&g1_par_verify_task);
2054 // Done with the parallel phase so reset to 0.
2055 g1h->set_par_threads(0);
2056
2057 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2058 "sanity check");
2059 } else {
2060 g1_par_verify_task.work(0);
2061 }
2062
2063 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2064 }
2065
2066 size_t start_used_bytes = g1h->used();
2067 g1h->set_marking_complete();
2068
2069 double count_end = os::elapsedTime();
2070 double this_final_counting_time = (count_end - start);
2071 _total_counting_time += this_final_counting_time;
2072
2073 if (G1PrintRegionLivenessInfo) {
2074 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2075 _g1h->heap_region_iterate(&cl);
2076 }
2077
2078 // Install newly created mark bitMap as "prev".
2079 swapMarkBitMaps();
2080
2081 g1h->reset_gc_time_stamp();
2082
2083 // Note end of marking in all heap regions.
2084 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2085 if (G1CollectedHeap::use_parallel_gc_threads()) {
2086 g1h->set_par_threads((int)n_workers);
2087 g1h->workers()->run_task(&g1_par_note_end_task);
2088 g1h->set_par_threads(0);
2089
2090 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2091 "sanity check");
2092 } else {
2093 g1_par_note_end_task.work(0);
2094 }
2095 g1h->check_gc_time_stamps();
2096
2097 if (!cleanup_list_is_empty()) {
2098 // The cleanup list is not empty, so we'll have to process it
2099 // concurrently. Notify anyone else that might be wanting free
2100 // regions that there will be more free regions coming soon.
2101 g1h->set_free_regions_coming();
2102 }
2103
2104 // call below, since it affects the metric by which we sort the heap
2105 // regions.
2106 if (G1ScrubRemSets) {
2107 double rs_scrub_start = os::elapsedTime();
2108 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2109 if (G1CollectedHeap::use_parallel_gc_threads()) {
2110 g1h->set_par_threads((int)n_workers);
2111 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2112 g1h->set_par_threads(0);
2113
2114 assert(g1h->check_heap_region_claim_values(
2115 HeapRegion::ScrubRemSetClaimValue),
2116 "sanity check");
2117 } else {
2118 g1_par_scrub_rs_task.work(0);
2119 }
2120
2121 double rs_scrub_end = os::elapsedTime();
2122 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2123 _total_rs_scrub_time += this_rs_scrub_time;
2124 }
2125
2126 // this will also free any regions totally full of garbage objects,
2127 // and sort the regions.
2128 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2129
2130 // Statistics.
2131 double end = os::elapsedTime();
2132 _cleanup_times.add((end - start) * 1000.0);
2133
2134 if (G1Log::fine()) {
2135 g1h->print_size_transition(gclog_or_tty,
2136 start_used_bytes,
2137 g1h->used(),
2138 g1h->capacity());
2139 }
2140
2141 // Clean up will have freed any regions completely full of garbage.
2142 // Update the soft reference policy with the new heap occupancy.
2143 Universe::update_heap_info_at_gc();
2144
2145 // We need to make this be a "collection" so any collection pause that
2146 // races with it goes around and waits for completeCleanup to finish.
2147 g1h->increment_total_collections();
2148
2149 // We reclaimed old regions so we should calculate the sizes to make
2150 // sure we update the old gen/space data.
2151 g1h->g1mm()->update_sizes();
2152
2153 if (VerifyDuringGC) {
2154 HandleMark hm; // handle scope
2155 Universe::heap()->prepare_for_verify();
2156 Universe::verify(VerifyOption_G1UsePrevMarking,
2157 " VerifyDuringGC:(after)");
2158 }
2159 g1h->check_bitmaps("Cleanup End");
2160
2161 g1h->verify_region_sets_optional();
2162 g1h->trace_heap_after_concurrent_cycle();
2163 }
2164
2165 void ConcurrentMark::completeCleanup() {
2166 if (has_aborted()) return;
2167
2168 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2169
2170 _cleanup_list.verify_optional();
2171 FreeRegionList tmp_free_list("Tmp Free List");
2172
2173 if (G1ConcRegionFreeingVerbose) {
2174 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2175 "cleanup list has %u entries",
2176 _cleanup_list.length());
2177 }
2178
2179 // Noone else should be accessing the _cleanup_list at this point,
2180 // so it's not necessary to take any locks
2181 while (!_cleanup_list.is_empty()) {
2182 HeapRegion* hr = _cleanup_list.remove_head();
2183 assert(hr != NULL, "Got NULL from a non-empty list");
2184 hr->par_clear();
2185 tmp_free_list.add_ordered(hr);
2186
2187 // Instead of adding one region at a time to the secondary_free_list,
2188 // we accumulate them in the local list and move them a few at a
2189 // time. This also cuts down on the number of notify_all() calls
2190 // we do during this process. We'll also append the local list when
2191 // _cleanup_list is empty (which means we just removed the last
2192 // region from the _cleanup_list).
2193 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2194 _cleanup_list.is_empty()) {
2195 if (G1ConcRegionFreeingVerbose) {
2196 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2197 "appending %u entries to the secondary_free_list, "
2198 "cleanup list still has %u entries",
2199 tmp_free_list.length(),
2200 _cleanup_list.length());
2201 }
2202
2203 {
2204 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2205 g1h->secondary_free_list_add(&tmp_free_list);
2206 SecondaryFreeList_lock->notify_all();
2207 }
2208
2209 if (G1StressConcRegionFreeing) {
2210 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2211 os::sleep(Thread::current(), (jlong) 1, false);
2212 }
2213 }
2214 }
2215 }
2216 assert(tmp_free_list.is_empty(), "post-condition");
2217 }
2218
2219 // Supporting Object and Oop closures for reference discovery
2220 // and processing in during marking
2221
2222 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2223 HeapWord* addr = (HeapWord*)obj;
2224 return addr != NULL &&
2225 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2226 }
2227
2228 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2229 // Uses the CMTask associated with a worker thread (for serial reference
2230 // processing the CMTask for worker 0 is used) to preserve (mark) and
2231 // trace referent objects.
2232 //
2233 // Using the CMTask and embedded local queues avoids having the worker
2234 // threads operating on the global mark stack. This reduces the risk
2235 // of overflowing the stack - which we would rather avoid at this late
2236 // state. Also using the tasks' local queues removes the potential
2237 // of the workers interfering with each other that could occur if
2238 // operating on the global stack.
2239
2240 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2241 ConcurrentMark* _cm;
2242 CMTask* _task;
2243 int _ref_counter_limit;
2244 int _ref_counter;
2245 bool _is_serial;
2246 public:
2247 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2248 _cm(cm), _task(task), _is_serial(is_serial),
2249 _ref_counter_limit(G1RefProcDrainInterval) {
2250 assert(_ref_counter_limit > 0, "sanity");
2251 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2252 _ref_counter = _ref_counter_limit;
2253 }
2254
2255 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2256 virtual void do_oop( oop* p) { do_oop_work(p); }
2257
2258 template <class T> void do_oop_work(T* p) {
2259 if (!_cm->has_overflown()) {
2260 oop obj = oopDesc::load_decode_heap_oop(p);
2261 if (_cm->verbose_high()) {
2262 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2263 "*"PTR_FORMAT" = "PTR_FORMAT,
2264 _task->worker_id(), p2i(p), p2i((void*) obj));
2265 }
2266
2267 _task->deal_with_reference(obj);
2268 _ref_counter--;
2269
2270 if (_ref_counter == 0) {
2271 // We have dealt with _ref_counter_limit references, pushing them
2272 // and objects reachable from them on to the local stack (and
2273 // possibly the global stack). Call CMTask::do_marking_step() to
2274 // process these entries.
2275 //
2276 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2277 // there's nothing more to do (i.e. we're done with the entries that
2278 // were pushed as a result of the CMTask::deal_with_reference() calls
2279 // above) or we overflow.
2280 //
2281 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2282 // flag while there may still be some work to do. (See the comment at
2283 // the beginning of CMTask::do_marking_step() for those conditions -
2284 // one of which is reaching the specified time target.) It is only
2285 // when CMTask::do_marking_step() returns without setting the
2286 // has_aborted() flag that the marking step has completed.
2287 do {
2288 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2289 _task->do_marking_step(mark_step_duration_ms,
2290 false /* do_termination */,
2291 _is_serial);
2292 } while (_task->has_aborted() && !_cm->has_overflown());
2293 _ref_counter = _ref_counter_limit;
2294 }
2295 } else {
2296 if (_cm->verbose_high()) {
2297 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2298 }
2299 }
2300 }
2301 };
2302
2303 // 'Drain' oop closure used by both serial and parallel reference processing.
2304 // Uses the CMTask associated with a given worker thread (for serial
2305 // reference processing the CMtask for worker 0 is used). Calls the
2306 // do_marking_step routine, with an unbelievably large timeout value,
2307 // to drain the marking data structures of the remaining entries
2308 // added by the 'keep alive' oop closure above.
2309
2310 class G1CMDrainMarkingStackClosure: public VoidClosure {
2311 ConcurrentMark* _cm;
2312 CMTask* _task;
2313 bool _is_serial;
2314 public:
2315 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2316 _cm(cm), _task(task), _is_serial(is_serial) {
2317 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2318 }
2319
2320 void do_void() {
2321 do {
2322 if (_cm->verbose_high()) {
2323 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2324 _task->worker_id(), BOOL_TO_STR(_is_serial));
2325 }
2326
2327 // We call CMTask::do_marking_step() to completely drain the local
2328 // and global marking stacks of entries pushed by the 'keep alive'
2329 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2330 //
2331 // CMTask::do_marking_step() is called in a loop, which we'll exit
2332 // if there's nothing more to do (i.e. we've completely drained the
2333 // entries that were pushed as a a result of applying the 'keep alive'
2334 // closure to the entries on the discovered ref lists) or we overflow
2335 // the global marking stack.
2336 //
2337 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2338 // flag while there may still be some work to do. (See the comment at
2339 // the beginning of CMTask::do_marking_step() for those conditions -
2340 // one of which is reaching the specified time target.) It is only
2341 // when CMTask::do_marking_step() returns without setting the
2342 // has_aborted() flag that the marking step has completed.
2343
2344 _task->do_marking_step(1000000000.0 /* something very large */,
2345 true /* do_termination */,
2346 _is_serial);
2347 } while (_task->has_aborted() && !_cm->has_overflown());
2348 }
2349 };
2350
2351 // Implementation of AbstractRefProcTaskExecutor for parallel
2352 // reference processing at the end of G1 concurrent marking
2353
2354 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2355 private:
2356 G1CollectedHeap* _g1h;
2357 ConcurrentMark* _cm;
2358 WorkGang* _workers;
2359 int _active_workers;
2360
2361 public:
2362 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2363 ConcurrentMark* cm,
2364 WorkGang* workers,
2365 int n_workers) :
2366 _g1h(g1h), _cm(cm),
2367 _workers(workers), _active_workers(n_workers) { }
2368
2369 // Executes the given task using concurrent marking worker threads.
2370 virtual void execute(ProcessTask& task);
2371 virtual void execute(EnqueueTask& task);
2372 };
2373
2374 class G1CMRefProcTaskProxy: public AbstractGangTask {
2375 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2376 ProcessTask& _proc_task;
2377 G1CollectedHeap* _g1h;
2378 ConcurrentMark* _cm;
2379
2380 public:
2381 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2382 G1CollectedHeap* g1h,
2383 ConcurrentMark* cm) :
2384 AbstractGangTask("Process reference objects in parallel"),
2385 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2386 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2387 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2388 }
2389
2390 virtual void work(uint worker_id) {
2391 CMTask* task = _cm->task(worker_id);
2392 G1CMIsAliveClosure g1_is_alive(_g1h);
2393 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2394 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2395
2396 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2397 }
2398 };
2399
2400 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2401 assert(_workers != NULL, "Need parallel worker threads.");
2402 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2403
2404 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2405
2406 // We need to reset the concurrency level before each
2407 // proxy task execution, so that the termination protocol
2408 // and overflow handling in CMTask::do_marking_step() knows
2409 // how many workers to wait for.
2410 _cm->set_concurrency(_active_workers);
2411 _g1h->set_par_threads(_active_workers);
2412 _workers->run_task(&proc_task_proxy);
2413 _g1h->set_par_threads(0);
2414 }
2415
2416 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2417 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2418 EnqueueTask& _enq_task;
2419
2420 public:
2421 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2422 AbstractGangTask("Enqueue reference objects in parallel"),
2423 _enq_task(enq_task) { }
2424
2425 virtual void work(uint worker_id) {
2426 _enq_task.work(worker_id);
2427 }
2428 };
2429
2430 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2431 assert(_workers != NULL, "Need parallel worker threads.");
2432 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2433
2434 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2435
2436 // Not strictly necessary but...
2437 //
2438 // We need to reset the concurrency level before each
2439 // proxy task execution, so that the termination protocol
2440 // and overflow handling in CMTask::do_marking_step() knows
2441 // how many workers to wait for.
2442 _cm->set_concurrency(_active_workers);
2443 _g1h->set_par_threads(_active_workers);
2444 _workers->run_task(&enq_task_proxy);
2445 _g1h->set_par_threads(0);
2446 }
2447
2448 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2449 if (has_overflown()) {
2450 // Skip processing the discovered references if we have
2451 // overflown the global marking stack. Reference objects
2452 // only get discovered once so it is OK to not
2453 // de-populate the discovered reference lists. We could have,
2454 // but the only benefit would be that, when marking restarts,
2455 // less reference objects are discovered.
2456 return;
2457 }
2458
2459 ResourceMark rm;
2460 HandleMark hm;
2461
2462 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2463
2464 // Is alive closure.
2465 G1CMIsAliveClosure g1_is_alive(g1h);
2466
2467 // Inner scope to exclude the cleaning of the string and symbol
2468 // tables from the displayed time.
2469 {
2470 if (G1Log::finer()) {
2471 gclog_or_tty->put(' ');
2472 }
2473 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm());
2474
2475 ReferenceProcessor* rp = g1h->ref_processor_cm();
2476
2477 // See the comment in G1CollectedHeap::ref_processing_init()
2478 // about how reference processing currently works in G1.
2479
2480 // Set the soft reference policy
2481 rp->setup_policy(clear_all_soft_refs);
2482 assert(_markStack.isEmpty(), "mark stack should be empty");
2483
2484 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2485 // in serial reference processing. Note these closures are also
2486 // used for serially processing (by the the current thread) the
2487 // JNI references during parallel reference processing.
2488 //
2489 // These closures do not need to synchronize with the worker
2490 // threads involved in parallel reference processing as these
2491 // instances are executed serially by the current thread (e.g.
2492 // reference processing is not multi-threaded and is thus
2493 // performed by the current thread instead of a gang worker).
2494 //
2495 // The gang tasks involved in parallel reference processing create
2496 // their own instances of these closures, which do their own
2497 // synchronization among themselves.
2498 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2499 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2500
2501 // We need at least one active thread. If reference processing
2502 // is not multi-threaded we use the current (VMThread) thread,
2503 // otherwise we use the work gang from the G1CollectedHeap and
2504 // we utilize all the worker threads we can.
2505 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2506 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2507 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2508
2509 // Parallel processing task executor.
2510 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2511 g1h->workers(), active_workers);
2512 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2513
2514 // Set the concurrency level. The phase was already set prior to
2515 // executing the remark task.
2516 set_concurrency(active_workers);
2517
2518 // Set the degree of MT processing here. If the discovery was done MT,
2519 // the number of threads involved during discovery could differ from
2520 // the number of active workers. This is OK as long as the discovered
2521 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2522 rp->set_active_mt_degree(active_workers);
2523
2524 // Process the weak references.
2525 const ReferenceProcessorStats& stats =
2526 rp->process_discovered_references(&g1_is_alive,
2527 &g1_keep_alive,
2528 &g1_drain_mark_stack,
2529 executor,
2530 g1h->gc_timer_cm());
2531 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2532
2533 // The do_oop work routines of the keep_alive and drain_marking_stack
2534 // oop closures will set the has_overflown flag if we overflow the
2535 // global marking stack.
2536
2537 assert(_markStack.overflow() || _markStack.isEmpty(),
2538 "mark stack should be empty (unless it overflowed)");
2539
2540 if (_markStack.overflow()) {
2541 // This should have been done already when we tried to push an
2542 // entry on to the global mark stack. But let's do it again.
2543 set_has_overflown();
2544 }
2545
2546 assert(rp->num_q() == active_workers, "why not");
2547
2548 rp->enqueue_discovered_references(executor);
2549
2550 rp->verify_no_references_recorded();
2551 assert(!rp->discovery_enabled(), "Post condition");
2552 }
2553
2554 if (has_overflown()) {
2555 // We can not trust g1_is_alive if the marking stack overflowed
2556 return;
2557 }
2558
2559 g1h->unlink_string_and_symbol_table(&g1_is_alive,
2560 /* process_strings */ false, // currently strings are always roots
2561 /* process_symbols */ true);
2562 }
2563
2564 void ConcurrentMark::swapMarkBitMaps() {
2565 CMBitMapRO* temp = _prevMarkBitMap;
2566 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2567 _nextMarkBitMap = (CMBitMap*) temp;
2568 }
2569
2570 class CMRemarkTask: public AbstractGangTask {
2571 private:
2572 ConcurrentMark* _cm;
2573 bool _is_serial;
2574 public:
2575 void work(uint worker_id) {
2576 // Since all available tasks are actually started, we should
2577 // only proceed if we're supposed to be active.
2578 if (worker_id < _cm->active_tasks()) {
2579 CMTask* task = _cm->task(worker_id);
2580 task->record_start_time();
2581 do {
2582 task->do_marking_step(1000000000.0 /* something very large */,
2583 true /* do_termination */,
2584 _is_serial);
2585 } while (task->has_aborted() && !_cm->has_overflown());
2586 // If we overflow, then we do not want to restart. We instead
2587 // want to abort remark and do concurrent marking again.
2588 task->record_end_time();
2589 }
2590 }
2591
2592 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2593 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2594 _cm->terminator()->reset_for_reuse(active_workers);
2595 }
2596 };
2597
2598 void ConcurrentMark::checkpointRootsFinalWork() {
2599 ResourceMark rm;
2600 HandleMark hm;
2601 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2602
2603 g1h->ensure_parsability(false);
2604
2605 if (G1CollectedHeap::use_parallel_gc_threads()) {
2606 G1CollectedHeap::StrongRootsScope srs(g1h);
2607 // this is remark, so we'll use up all active threads
2608 uint active_workers = g1h->workers()->active_workers();
2609 if (active_workers == 0) {
2610 assert(active_workers > 0, "Should have been set earlier");
2611 active_workers = (uint) ParallelGCThreads;
2612 g1h->workers()->set_active_workers(active_workers);
2613 }
2614 set_concurrency_and_phase(active_workers, false /* concurrent */);
2615 // Leave _parallel_marking_threads at it's
2616 // value originally calculated in the ConcurrentMark
2617 // constructor and pass values of the active workers
2618 // through the gang in the task.
2619
2620 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2621 // We will start all available threads, even if we decide that the
2622 // active_workers will be fewer. The extra ones will just bail out
2623 // immediately.
2624 g1h->set_par_threads(active_workers);
2625 g1h->workers()->run_task(&remarkTask);
2626 g1h->set_par_threads(0);
2627 } else {
2628 G1CollectedHeap::StrongRootsScope srs(g1h);
2629 uint active_workers = 1;
2630 set_concurrency_and_phase(active_workers, false /* concurrent */);
2631
2632 // Note - if there's no work gang then the VMThread will be
2633 // the thread to execute the remark - serially. We have
2634 // to pass true for the is_serial parameter so that
2635 // CMTask::do_marking_step() doesn't enter the sync
2636 // barriers in the event of an overflow. Doing so will
2637 // cause an assert that the current thread is not a
2638 // concurrent GC thread.
2639 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2640 remarkTask.work(0);
2641 }
2642 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2643 guarantee(has_overflown() ||
2644 satb_mq_set.completed_buffers_num() == 0,
2645 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2646 BOOL_TO_STR(has_overflown()),
2647 satb_mq_set.completed_buffers_num()));
2648
2649 print_stats();
2650 }
2651
2652 #ifndef PRODUCT
2653
2654 class PrintReachableOopClosure: public OopClosure {
2655 private:
2656 G1CollectedHeap* _g1h;
2657 outputStream* _out;
2658 VerifyOption _vo;
2659 bool _all;
2660
2661 public:
2662 PrintReachableOopClosure(outputStream* out,
2663 VerifyOption vo,
2664 bool all) :
2665 _g1h(G1CollectedHeap::heap()),
2666 _out(out), _vo(vo), _all(all) { }
2667
2668 void do_oop(narrowOop* p) { do_oop_work(p); }
2669 void do_oop( oop* p) { do_oop_work(p); }
2670
2671 template <class T> void do_oop_work(T* p) {
2672 oop obj = oopDesc::load_decode_heap_oop(p);
2673 const char* str = NULL;
2674 const char* str2 = "";
2675
2676 if (obj == NULL) {
2677 str = "";
2678 } else if (!_g1h->is_in_g1_reserved(obj)) {
2679 str = " O";
2680 } else {
2681 HeapRegion* hr = _g1h->heap_region_containing(obj);
2682 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2683 bool marked = _g1h->is_marked(obj, _vo);
2684
2685 if (over_tams) {
2686 str = " >";
2687 if (marked) {
2688 str2 = " AND MARKED";
2689 }
2690 } else if (marked) {
2691 str = " M";
2692 } else {
2693 str = " NOT";
2694 }
2695 }
2696
2697 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2698 p2i(p), p2i((void*) obj), str, str2);
2699 }
2700 };
2701
2702 class PrintReachableObjectClosure : public ObjectClosure {
2703 private:
2704 G1CollectedHeap* _g1h;
2705 outputStream* _out;
2706 VerifyOption _vo;
2707 bool _all;
2708 HeapRegion* _hr;
2709
2710 public:
2711 PrintReachableObjectClosure(outputStream* out,
2712 VerifyOption vo,
2713 bool all,
2714 HeapRegion* hr) :
2715 _g1h(G1CollectedHeap::heap()),
2716 _out(out), _vo(vo), _all(all), _hr(hr) { }
2717
2718 void do_object(oop o) {
2719 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2720 bool marked = _g1h->is_marked(o, _vo);
2721 bool print_it = _all || over_tams || marked;
2722
2723 if (print_it) {
2724 _out->print_cr(" "PTR_FORMAT"%s",
2725 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2726 PrintReachableOopClosure oopCl(_out, _vo, _all);
2727 o->oop_iterate_no_header(&oopCl);
2728 }
2729 }
2730 };
2731
2732 class PrintReachableRegionClosure : public HeapRegionClosure {
2733 private:
2734 G1CollectedHeap* _g1h;
2735 outputStream* _out;
2736 VerifyOption _vo;
2737 bool _all;
2738
2739 public:
2740 bool doHeapRegion(HeapRegion* hr) {
2741 HeapWord* b = hr->bottom();
2742 HeapWord* e = hr->end();
2743 HeapWord* t = hr->top();
2744 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2745 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2746 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2747 _out->cr();
2748
2749 HeapWord* from = b;
2750 HeapWord* to = t;
2751
2752 if (to > from) {
2753 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2754 _out->cr();
2755 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2756 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2757 _out->cr();
2758 }
2759
2760 return false;
2761 }
2762
2763 PrintReachableRegionClosure(outputStream* out,
2764 VerifyOption vo,
2765 bool all) :
2766 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2767 };
2768
2769 void ConcurrentMark::print_reachable(const char* str,
2770 VerifyOption vo,
2771 bool all) {
2772 gclog_or_tty->cr();
2773 gclog_or_tty->print_cr("== Doing heap dump... ");
2774
2775 if (G1PrintReachableBaseFile == NULL) {
2776 gclog_or_tty->print_cr(" #### error: no base file defined");
2777 return;
2778 }
2779
2780 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2781 (JVM_MAXPATHLEN - 1)) {
2782 gclog_or_tty->print_cr(" #### error: file name too long");
2783 return;
2784 }
2785
2786 char file_name[JVM_MAXPATHLEN];
2787 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2788 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2789
2790 fileStream fout(file_name);
2791 if (!fout.is_open()) {
2792 gclog_or_tty->print_cr(" #### error: could not open file");
2793 return;
2794 }
2795
2796 outputStream* out = &fout;
2797 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2798 out->cr();
2799
2800 out->print_cr("--- ITERATING OVER REGIONS");
2801 out->cr();
2802 PrintReachableRegionClosure rcl(out, vo, all);
2803 _g1h->heap_region_iterate(&rcl);
2804 out->cr();
2805
2806 gclog_or_tty->print_cr(" done");
2807 gclog_or_tty->flush();
2808 }
2809
2810 #endif // PRODUCT
2811
2812 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2813 // Note we are overriding the read-only view of the prev map here, via
2814 // the cast.
2815 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2816 }
2817
2818 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2819 _nextMarkBitMap->clearRange(mr);
2820 }
2821
2822 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) {
2823 clearRangePrevBitmap(mr);
2824 clearRangeNextBitmap(mr);
2825 }
2826
2827 HeapRegion*
2828 ConcurrentMark::claim_region(uint worker_id) {
2829 // "checkpoint" the finger
2830 HeapWord* finger = _finger;
2831
2832 // _heap_end will not change underneath our feet; it only changes at
2833 // yield points.
2834 while (finger < _heap_end) {
2835 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2836
2837 // Note on how this code handles humongous regions. In the
2838 // normal case the finger will reach the start of a "starts
2839 // humongous" (SH) region. Its end will either be the end of the
2840 // last "continues humongous" (CH) region in the sequence, or the
2841 // standard end of the SH region (if the SH is the only region in
2842 // the sequence). That way claim_region() will skip over the CH
2843 // regions. However, there is a subtle race between a CM thread
2844 // executing this method and a mutator thread doing a humongous
2845 // object allocation. The two are not mutually exclusive as the CM
2846 // thread does not need to hold the Heap_lock when it gets
2847 // here. So there is a chance that claim_region() will come across
2848 // a free region that's in the progress of becoming a SH or a CH
2849 // region. In the former case, it will either
2850 // a) Miss the update to the region's end, in which case it will
2851 // visit every subsequent CH region, will find their bitmaps
2852 // empty, and do nothing, or
2853 // b) Will observe the update of the region's end (in which case
2854 // it will skip the subsequent CH regions).
2855 // If it comes across a region that suddenly becomes CH, the
2856 // scenario will be similar to b). So, the race between
2857 // claim_region() and a humongous object allocation might force us
2858 // to do a bit of unnecessary work (due to some unnecessary bitmap
2859 // iterations) but it should not introduce and correctness issues.
2860 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2861 HeapWord* bottom = curr_region->bottom();
2862 HeapWord* end = curr_region->end();
2863 HeapWord* limit = curr_region->next_top_at_mark_start();
2864
2865 if (verbose_low()) {
2866 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2867 "["PTR_FORMAT", "PTR_FORMAT"), "
2868 "limit = "PTR_FORMAT,
2869 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2870 }
2871
2872 // Is the gap between reading the finger and doing the CAS too long?
2873 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2874 if (res == finger) {
2875 // we succeeded
2876
2877 // notice that _finger == end cannot be guaranteed here since,
2878 // someone else might have moved the finger even further
2879 assert(_finger >= end, "the finger should have moved forward");
2880
2881 if (verbose_low()) {
2882 gclog_or_tty->print_cr("[%u] we were successful with region = "
2883 PTR_FORMAT, worker_id, p2i(curr_region));
2884 }
2885
2886 if (limit > bottom) {
2887 if (verbose_low()) {
2888 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2889 "returning it ", worker_id, p2i(curr_region));
2890 }
2891 return curr_region;
2892 } else {
2893 assert(limit == bottom,
2894 "the region limit should be at bottom");
2895 if (verbose_low()) {
2896 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
2897 "returning NULL", worker_id, p2i(curr_region));
2898 }
2899 // we return NULL and the caller should try calling
2900 // claim_region() again.
2901 return NULL;
2902 }
2903 } else {
2904 assert(_finger > finger, "the finger should have moved forward");
2905 if (verbose_low()) {
2906 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2907 "global finger = "PTR_FORMAT", "
2908 "our finger = "PTR_FORMAT,
2909 worker_id, p2i(_finger), p2i(finger));
2910 }
2911
2912 // read it again
2913 finger = _finger;
2914 }
2915 }
2916
2917 return NULL;
2918 }
2919
2920 #ifndef PRODUCT
2921 enum VerifyNoCSetOopsPhase {
2922 VerifyNoCSetOopsStack,
2923 VerifyNoCSetOopsQueues,
2924 VerifyNoCSetOopsSATBCompleted,
2925 VerifyNoCSetOopsSATBThread
2926 };
2927
2928 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
2929 private:
2930 G1CollectedHeap* _g1h;
2931 VerifyNoCSetOopsPhase _phase;
2932 int _info;
2933
2934 const char* phase_str() {
2935 switch (_phase) {
2936 case VerifyNoCSetOopsStack: return "Stack";
2937 case VerifyNoCSetOopsQueues: return "Queue";
2938 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2939 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
2940 default: ShouldNotReachHere();
2941 }
2942 return NULL;
2943 }
2944
2945 void do_object_work(oop obj) {
2946 guarantee(!_g1h->obj_in_cs(obj),
2947 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2948 p2i((void*) obj), phase_str(), _info));
2949 }
2950
2951 public:
2952 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2953
2954 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2955 _phase = phase;
2956 _info = info;
2957 }
2958
2959 virtual void do_oop(oop* p) {
2960 oop obj = oopDesc::load_decode_heap_oop(p);
2961 do_object_work(obj);
2962 }
2963
2964 virtual void do_oop(narrowOop* p) {
2965 // We should not come across narrow oops while scanning marking
2966 // stacks and SATB buffers.
2967 ShouldNotReachHere();
2968 }
2969
2970 virtual void do_object(oop obj) {
2971 do_object_work(obj);
2972 }
2973 };
2974
2975 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
2976 bool verify_enqueued_buffers,
2977 bool verify_thread_buffers,
2978 bool verify_fingers) {
2979 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2980 if (!G1CollectedHeap::heap()->mark_in_progress()) {
2981 return;
2982 }
2983
2984 VerifyNoCSetOopsClosure cl;
2985
2986 if (verify_stacks) {
2987 // Verify entries on the global mark stack
2988 cl.set_phase(VerifyNoCSetOopsStack);
2989 _markStack.oops_do(&cl);
2990
2991 // Verify entries on the task queues
2992 for (uint i = 0; i < _max_worker_id; i += 1) {
2993 cl.set_phase(VerifyNoCSetOopsQueues, i);
2994 CMTaskQueue* queue = _task_queues->queue(i);
2995 queue->oops_do(&cl);
2996 }
2997 }
2998
2999 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3000
3001 // Verify entries on the enqueued SATB buffers
3002 if (verify_enqueued_buffers) {
3003 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3004 satb_qs.iterate_completed_buffers_read_only(&cl);
3005 }
3006
3007 // Verify entries on the per-thread SATB buffers
3008 if (verify_thread_buffers) {
3009 cl.set_phase(VerifyNoCSetOopsSATBThread);
3010 satb_qs.iterate_thread_buffers_read_only(&cl);
3011 }
3012
3013 if (verify_fingers) {
3014 // Verify the global finger
3015 HeapWord* global_finger = finger();
3016 if (global_finger != NULL && global_finger < _heap_end) {
3017 // The global finger always points to a heap region boundary. We
3018 // use heap_region_containing_raw() to get the containing region
3019 // given that the global finger could be pointing to a free region
3020 // which subsequently becomes continues humongous. If that
3021 // happens, heap_region_containing() will return the bottom of the
3022 // corresponding starts humongous region and the check below will
3023 // not hold any more.
3024 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3025 guarantee(global_finger == global_hr->bottom(),
3026 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3027 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3028 }
3029
3030 // Verify the task fingers
3031 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3032 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3033 CMTask* task = _tasks[i];
3034 HeapWord* task_finger = task->finger();
3035 if (task_finger != NULL && task_finger < _heap_end) {
3036 // See above note on the global finger verification.
3037 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3038 guarantee(task_finger == task_hr->bottom() ||
3039 !task_hr->in_collection_set(),
3040 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3041 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3042 }
3043 }
3044 }
3045 }
3046 #endif // PRODUCT
3047
3048 // Aggregate the counting data that was constructed concurrently
3049 // with marking.
3050 class AggregateCountDataHRClosure: public HeapRegionClosure {
3051 G1CollectedHeap* _g1h;
3052 ConcurrentMark* _cm;
3053 CardTableModRefBS* _ct_bs;
3054 BitMap* _cm_card_bm;
3055 uint _max_worker_id;
3056
3057 public:
3058 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3059 BitMap* cm_card_bm,
3060 uint max_worker_id) :
3061 _g1h(g1h), _cm(g1h->concurrent_mark()),
3062 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3063 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3064
3065 bool doHeapRegion(HeapRegion* hr) {
3066 if (hr->continuesHumongous()) {
3067 // We will ignore these here and process them when their
3068 // associated "starts humongous" region is processed.
3069 // Note that we cannot rely on their associated
3070 // "starts humongous" region to have their bit set to 1
3071 // since, due to the region chunking in the parallel region
3072 // iteration, a "continues humongous" region might be visited
3073 // before its associated "starts humongous".
3074 return false;
3075 }
3076
3077 HeapWord* start = hr->bottom();
3078 HeapWord* limit = hr->next_top_at_mark_start();
3079 HeapWord* end = hr->end();
3080
3081 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3082 err_msg("Preconditions not met - "
3083 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3084 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3085 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3086
3087 assert(hr->next_marked_bytes() == 0, "Precondition");
3088
3089 if (start == limit) {
3090 // NTAMS of this region has not been set so nothing to do.
3091 return false;
3092 }
3093
3094 // 'start' should be in the heap.
3095 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3096 // 'end' *may* be just beyond the end of the heap (if hr is the last region)
3097 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3098
3099 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3100 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3101 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3102
3103 // If ntams is not card aligned then we bump card bitmap index
3104 // for limit so that we get the all the cards spanned by
3105 // the object ending at ntams.
3106 // Note: if this is the last region in the heap then ntams
3107 // could be actually just beyond the end of the the heap;
3108 // limit_idx will then correspond to a (non-existent) card
3109 // that is also outside the heap.
3110 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3111 limit_idx += 1;
3112 }
3113
3114 assert(limit_idx <= end_idx, "or else use atomics");
3115
3116 // Aggregate the "stripe" in the count data associated with hr.
3117 uint hrs_index = hr->hrs_index();
3118 size_t marked_bytes = 0;
3119
3120 for (uint i = 0; i < _max_worker_id; i += 1) {
3121 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3122 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3123
3124 // Fetch the marked_bytes in this region for task i and
3125 // add it to the running total for this region.
3126 marked_bytes += marked_bytes_array[hrs_index];
3127
3128 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3129 // into the global card bitmap.
3130 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3131
3132 while (scan_idx < limit_idx) {
3133 assert(task_card_bm->at(scan_idx) == true, "should be");
3134 _cm_card_bm->set_bit(scan_idx);
3135 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3136
3137 // BitMap::get_next_one_offset() can handle the case when
3138 // its left_offset parameter is greater than its right_offset
3139 // parameter. It does, however, have an early exit if
3140 // left_offset == right_offset. So let's limit the value
3141 // passed in for left offset here.
3142 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3143 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3144 }
3145 }
3146
3147 // Update the marked bytes for this region.
3148 hr->add_to_marked_bytes(marked_bytes);
3149
3150 // Next heap region
3151 return false;
3152 }
3153 };
3154
3155 class G1AggregateCountDataTask: public AbstractGangTask {
3156 protected:
3157 G1CollectedHeap* _g1h;
3158 ConcurrentMark* _cm;
3159 BitMap* _cm_card_bm;
3160 uint _max_worker_id;
3161 int _active_workers;
3162
3163 public:
3164 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3165 ConcurrentMark* cm,
3166 BitMap* cm_card_bm,
3167 uint max_worker_id,
3168 int n_workers) :
3169 AbstractGangTask("Count Aggregation"),
3170 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3171 _max_worker_id(max_worker_id),
3172 _active_workers(n_workers) { }
3173
3174 void work(uint worker_id) {
3175 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3176
3177 if (G1CollectedHeap::use_parallel_gc_threads()) {
3178 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3179 _active_workers,
3180 HeapRegion::AggregateCountClaimValue);
3181 } else {
3182 _g1h->heap_region_iterate(&cl);
3183 }
3184 }
3185 };
3186
3187
3188 void ConcurrentMark::aggregate_count_data() {
3189 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3190 _g1h->workers()->active_workers() :
3191 1);
3192
3193 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3194 _max_worker_id, n_workers);
3195
3196 if (G1CollectedHeap::use_parallel_gc_threads()) {
3197 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3198 "sanity check");
3199 _g1h->set_par_threads(n_workers);
3200 _g1h->workers()->run_task(&g1_par_agg_task);
3201 _g1h->set_par_threads(0);
3202
3203 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3204 "sanity check");
3205 _g1h->reset_heap_region_claim_values();
3206 } else {
3207 g1_par_agg_task.work(0);
3208 }
3209 }
3210
3211 // Clear the per-worker arrays used to store the per-region counting data
3212 void ConcurrentMark::clear_all_count_data() {
3213 // Clear the global card bitmap - it will be filled during
3214 // liveness count aggregation (during remark) and the
3215 // final counting task.
3216 _card_bm.clear();
3217
3218 // Clear the global region bitmap - it will be filled as part
3219 // of the final counting task.
3220 _region_bm.clear();
3221
3222 uint max_regions = _g1h->max_regions();
3223 assert(_max_worker_id > 0, "uninitialized");
3224
3225 for (uint i = 0; i < _max_worker_id; i += 1) {
3226 BitMap* task_card_bm = count_card_bitmap_for(i);
3227 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3228
3229 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3230 assert(marked_bytes_array != NULL, "uninitialized");
3231
3232 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3233 task_card_bm->clear();
3234 }
3235 }
3236
3237 void ConcurrentMark::print_stats() {
3238 if (verbose_stats()) {
3239 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3240 for (size_t i = 0; i < _active_tasks; ++i) {
3241 _tasks[i]->print_stats();
3242 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3243 }
3244 }
3245 }
3246
3247 // abandon current marking iteration due to a Full GC
3248 void ConcurrentMark::abort() {
3249 // Clear all marks to force marking thread to do nothing
3250 _nextMarkBitMap->clearAll();
3251
3252 // Note we cannot clear the previous marking bitmap here
3253 // since VerifyDuringGC verifies the objects marked during
3254 // a full GC against the previous bitmap.
3255
3256 // Clear the liveness counting data
3257 clear_all_count_data();
3258 // Empty mark stack
3259 reset_marking_state();
3260 for (uint i = 0; i < _max_worker_id; ++i) {
3261 _tasks[i]->clear_region_fields();
3262 }
3263 _first_overflow_barrier_sync.abort();
3264 _second_overflow_barrier_sync.abort();
3265 _has_aborted = true;
3266
3267 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3268 satb_mq_set.abandon_partial_marking();
3269 // This can be called either during or outside marking, we'll read
3270 // the expected_active value from the SATB queue set.
3271 satb_mq_set.set_active_all_threads(
3272 false, /* new active value */
3273 satb_mq_set.is_active() /* expected_active */);
3274
3275 _g1h->trace_heap_after_concurrent_cycle();
3276 _g1h->register_concurrent_cycle_end();
3277 }
3278
3279 static void print_ms_time_info(const char* prefix, const char* name,
3280 NumberSeq& ns) {
3281 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3282 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3283 if (ns.num() > 0) {
3284 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3285 prefix, ns.sd(), ns.maximum());
3286 }
3287 }
3288
3289 void ConcurrentMark::print_summary_info() {
3290 gclog_or_tty->print_cr(" Concurrent marking:");
3291 print_ms_time_info(" ", "init marks", _init_times);
3292 print_ms_time_info(" ", "remarks", _remark_times);
3293 {
3294 print_ms_time_info(" ", "final marks", _remark_mark_times);
3295 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3296
3297 }
3298 print_ms_time_info(" ", "cleanups", _cleanup_times);
3299 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3300 _total_counting_time,
3301 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3302 (double)_cleanup_times.num()
3303 : 0.0));
3304 if (G1ScrubRemSets) {
3305 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3306 _total_rs_scrub_time,
3307 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3308 (double)_cleanup_times.num()
3309 : 0.0));
3310 }
3311 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3312 (_init_times.sum() + _remark_times.sum() +
3313 _cleanup_times.sum())/1000.0);
3314 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3315 "(%8.2f s marking).",
3316 cmThread()->vtime_accum(),
3317 cmThread()->vtime_mark_accum());
3318 }
3319
3320 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3321 if (use_parallel_marking_threads()) {
3322 _parallel_workers->print_worker_threads_on(st);
3323 }
3324 }
3325
3326 void ConcurrentMark::print_on_error(outputStream* st) const {
3327 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3328 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3329 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3330 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3331 }
3332
3333 // We take a break if someone is trying to stop the world.
3334 bool ConcurrentMark::do_yield_check(uint worker_id) {
3335 if (SuspendibleThreadSet::should_yield()) {
3336 if (worker_id == 0) {
3337 _g1h->g1_policy()->record_concurrent_pause();
3338 }
3339 SuspendibleThreadSet::yield();
3340 return true;
3341 } else {
3342 return false;
3343 }
3344 }
3345
3346 bool ConcurrentMark::containing_card_is_marked(void* p) {
3347 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1);
3348 return _card_bm.at(offset >> CardTableModRefBS::card_shift);
3349 }
3350
3351 bool ConcurrentMark::containing_cards_are_marked(void* start,
3352 void* last) {
3353 return containing_card_is_marked(start) &&
3354 containing_card_is_marked(last);
3355 }
3356
3357 #ifndef PRODUCT
3358 // for debugging purposes
3359 void ConcurrentMark::print_finger() {
3360 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3361 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3362 for (uint i = 0; i < _max_worker_id; ++i) {
3363 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3364 }
3365 gclog_or_tty->cr();
3366 }
3367 #endif
3368
3369 void CMTask::scan_object(oop obj) {
3370 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3371
3372 if (_cm->verbose_high()) {
3373 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3374 _worker_id, p2i((void*) obj));
3375 }
3376
3377 size_t obj_size = obj->size();
3378 _words_scanned += obj_size;
3379
3380 obj->oop_iterate(_cm_oop_closure);
3381 statsOnly( ++_objs_scanned );
3382 check_limits();
3383 }
3384
3385 // Closure for iteration over bitmaps
3386 class CMBitMapClosure : public BitMapClosure {
3387 private:
3388 // the bitmap that is being iterated over
3389 CMBitMap* _nextMarkBitMap;
3390 ConcurrentMark* _cm;
3391 CMTask* _task;
3392
3393 public:
3394 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3395 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3396
3397 bool do_bit(size_t offset) {
3398 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3399 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3400 assert( addr < _cm->finger(), "invariant");
3401
3402 statsOnly( _task->increase_objs_found_on_bitmap() );
3403 assert(addr >= _task->finger(), "invariant");
3404
3405 // We move that task's local finger along.
3406 _task->move_finger_to(addr);
3407
3408 _task->scan_object(oop(addr));
3409 // we only partially drain the local queue and global stack
3410 _task->drain_local_queue(true);
3411 _task->drain_global_stack(true);
3412
3413 // if the has_aborted flag has been raised, we need to bail out of
3414 // the iteration
3415 return !_task->has_aborted();
3416 }
3417 };
3418
3419 // Closure for iterating over objects, currently only used for
3420 // processing SATB buffers.
3421 class CMObjectClosure : public ObjectClosure {
3422 private:
3423 CMTask* _task;
3424
3425 public:
3426 void do_object(oop obj) {
3427 _task->deal_with_reference(obj);
3428 }
3429
3430 CMObjectClosure(CMTask* task) : _task(task) { }
3431 };
3432
3433 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3434 ConcurrentMark* cm,
3435 CMTask* task)
3436 : _g1h(g1h), _cm(cm), _task(task) {
3437 assert(_ref_processor == NULL, "should be initialized to NULL");
3438
3439 if (G1UseConcMarkReferenceProcessing) {
3440 _ref_processor = g1h->ref_processor_cm();
3441 assert(_ref_processor != NULL, "should not be NULL");
3442 }
3443 }
3444
3445 void CMTask::setup_for_region(HeapRegion* hr) {
3446 assert(hr != NULL,
3447 "claim_region() should have filtered out NULL regions");
3448 assert(!hr->continuesHumongous(),
3449 "claim_region() should have filtered out continues humongous regions");
3450
3451 if (_cm->verbose_low()) {
3452 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3453 _worker_id, p2i(hr));
3454 }
3455
3456 _curr_region = hr;
3457 _finger = hr->bottom();
3458 update_region_limit();
3459 }
3460
3461 void CMTask::update_region_limit() {
3462 HeapRegion* hr = _curr_region;
3463 HeapWord* bottom = hr->bottom();
3464 HeapWord* limit = hr->next_top_at_mark_start();
3465
3466 if (limit == bottom) {
3467 if (_cm->verbose_low()) {
3468 gclog_or_tty->print_cr("[%u] found an empty region "
3469 "["PTR_FORMAT", "PTR_FORMAT")",
3470 _worker_id, p2i(bottom), p2i(limit));
3471 }
3472 // The region was collected underneath our feet.
3473 // We set the finger to bottom to ensure that the bitmap
3474 // iteration that will follow this will not do anything.
3475 // (this is not a condition that holds when we set the region up,
3476 // as the region is not supposed to be empty in the first place)
3477 _finger = bottom;
3478 } else if (limit >= _region_limit) {
3479 assert(limit >= _finger, "peace of mind");
3480 } else {
3481 assert(limit < _region_limit, "only way to get here");
3482 // This can happen under some pretty unusual circumstances. An
3483 // evacuation pause empties the region underneath our feet (NTAMS
3484 // at bottom). We then do some allocation in the region (NTAMS
3485 // stays at bottom), followed by the region being used as a GC
3486 // alloc region (NTAMS will move to top() and the objects
3487 // originally below it will be grayed). All objects now marked in
3488 // the region are explicitly grayed, if below the global finger,
3489 // and we do not need in fact to scan anything else. So, we simply
3490 // set _finger to be limit to ensure that the bitmap iteration
3491 // doesn't do anything.
3492 _finger = limit;
3493 }
3494
3495 _region_limit = limit;
3496 }
3497
3498 void CMTask::giveup_current_region() {
3499 assert(_curr_region != NULL, "invariant");
3500 if (_cm->verbose_low()) {
3501 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3502 _worker_id, p2i(_curr_region));
3503 }
3504 clear_region_fields();
3505 }
3506
3507 void CMTask::clear_region_fields() {
3508 // Values for these three fields that indicate that we're not
3509 // holding on to a region.
3510 _curr_region = NULL;
3511 _finger = NULL;
3512 _region_limit = NULL;
3513 }
3514
3515 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3516 if (cm_oop_closure == NULL) {
3517 assert(_cm_oop_closure != NULL, "invariant");
3518 } else {
3519 assert(_cm_oop_closure == NULL, "invariant");
3520 }
3521 _cm_oop_closure = cm_oop_closure;
3522 }
3523
3524 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3525 guarantee(nextMarkBitMap != NULL, "invariant");
3526
3527 if (_cm->verbose_low()) {
3528 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3529 }
3530
3531 _nextMarkBitMap = nextMarkBitMap;
3532 clear_region_fields();
3533
3534 _calls = 0;
3535 _elapsed_time_ms = 0.0;
3536 _termination_time_ms = 0.0;
3537 _termination_start_time_ms = 0.0;
3538
3539 #if _MARKING_STATS_
3540 _local_pushes = 0;
3541 _local_pops = 0;
3542 _local_max_size = 0;
3543 _objs_scanned = 0;
3544 _global_pushes = 0;
3545 _global_pops = 0;
3546 _global_max_size = 0;
3547 _global_transfers_to = 0;
3548 _global_transfers_from = 0;
3549 _regions_claimed = 0;
3550 _objs_found_on_bitmap = 0;
3551 _satb_buffers_processed = 0;
3552 _steal_attempts = 0;
3553 _steals = 0;
3554 _aborted = 0;
3555 _aborted_overflow = 0;
3556 _aborted_cm_aborted = 0;
3557 _aborted_yield = 0;
3558 _aborted_timed_out = 0;
3559 _aborted_satb = 0;
3560 _aborted_termination = 0;
3561 #endif // _MARKING_STATS_
3562 }
3563
3564 bool CMTask::should_exit_termination() {
3565 regular_clock_call();
3566 // This is called when we are in the termination protocol. We should
3567 // quit if, for some reason, this task wants to abort or the global
3568 // stack is not empty (this means that we can get work from it).
3569 return !_cm->mark_stack_empty() || has_aborted();
3570 }
3571
3572 void CMTask::reached_limit() {
3573 assert(_words_scanned >= _words_scanned_limit ||
3574 _refs_reached >= _refs_reached_limit ,
3575 "shouldn't have been called otherwise");
3576 regular_clock_call();
3577 }
3578
3579 void CMTask::regular_clock_call() {
3580 if (has_aborted()) return;
3581
3582 // First, we need to recalculate the words scanned and refs reached
3583 // limits for the next clock call.
3584 recalculate_limits();
3585
3586 // During the regular clock call we do the following
3587
3588 // (1) If an overflow has been flagged, then we abort.
3589 if (_cm->has_overflown()) {
3590 set_has_aborted();
3591 return;
3592 }
3593
3594 // If we are not concurrent (i.e. we're doing remark) we don't need
3595 // to check anything else. The other steps are only needed during
3596 // the concurrent marking phase.
3597 if (!concurrent()) return;
3598
3599 // (2) If marking has been aborted for Full GC, then we also abort.
3600 if (_cm->has_aborted()) {
3601 set_has_aborted();
3602 statsOnly( ++_aborted_cm_aborted );
3603 return;
3604 }
3605
3606 double curr_time_ms = os::elapsedVTime() * 1000.0;
3607
3608 // (3) If marking stats are enabled, then we update the step history.
3609 #if _MARKING_STATS_
3610 if (_words_scanned >= _words_scanned_limit) {
3611 ++_clock_due_to_scanning;
3612 }
3613 if (_refs_reached >= _refs_reached_limit) {
3614 ++_clock_due_to_marking;
3615 }
3616
3617 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3618 _interval_start_time_ms = curr_time_ms;
3619 _all_clock_intervals_ms.add(last_interval_ms);
3620
3621 if (_cm->verbose_medium()) {
3622 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3623 "scanned = %d%s, refs reached = %d%s",
3624 _worker_id, last_interval_ms,
3625 _words_scanned,
3626 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3627 _refs_reached,
3628 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3629 }
3630 #endif // _MARKING_STATS_
3631
3632 // (4) We check whether we should yield. If we have to, then we abort.
3633 if (SuspendibleThreadSet::should_yield()) {
3634 // We should yield. To do this we abort the task. The caller is
3635 // responsible for yielding.
3636 set_has_aborted();
3637 statsOnly( ++_aborted_yield );
3638 return;
3639 }
3640
3641 // (5) We check whether we've reached our time quota. If we have,
3642 // then we abort.
3643 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3644 if (elapsed_time_ms > _time_target_ms) {
3645 set_has_aborted();
3646 _has_timed_out = true;
3647 statsOnly( ++_aborted_timed_out );
3648 return;
3649 }
3650
3651 // (6) Finally, we check whether there are enough completed STAB
3652 // buffers available for processing. If there are, we abort.
3653 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3654 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3655 if (_cm->verbose_low()) {
3656 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3657 _worker_id);
3658 }
3659 // we do need to process SATB buffers, we'll abort and restart
3660 // the marking task to do so
3661 set_has_aborted();
3662 statsOnly( ++_aborted_satb );
3663 return;
3664 }
3665 }
3666
3667 void CMTask::recalculate_limits() {
3668 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3669 _words_scanned_limit = _real_words_scanned_limit;
3670
3671 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3672 _refs_reached_limit = _real_refs_reached_limit;
3673 }
3674
3675 void CMTask::decrease_limits() {
3676 // This is called when we believe that we're going to do an infrequent
3677 // operation which will increase the per byte scanned cost (i.e. move
3678 // entries to/from the global stack). It basically tries to decrease the
3679 // scanning limit so that the clock is called earlier.
3680
3681 if (_cm->verbose_medium()) {
3682 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3683 }
3684
3685 _words_scanned_limit = _real_words_scanned_limit -
3686 3 * words_scanned_period / 4;
3687 _refs_reached_limit = _real_refs_reached_limit -
3688 3 * refs_reached_period / 4;
3689 }
3690
3691 void CMTask::move_entries_to_global_stack() {
3692 // local array where we'll store the entries that will be popped
3693 // from the local queue
3694 oop buffer[global_stack_transfer_size];
3695
3696 int n = 0;
3697 oop obj;
3698 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3699 buffer[n] = obj;
3700 ++n;
3701 }
3702
3703 if (n > 0) {
3704 // we popped at least one entry from the local queue
3705
3706 statsOnly( ++_global_transfers_to; _local_pops += n );
3707
3708 if (!_cm->mark_stack_push(buffer, n)) {
3709 if (_cm->verbose_low()) {
3710 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3711 _worker_id);
3712 }
3713 set_has_aborted();
3714 } else {
3715 // the transfer was successful
3716
3717 if (_cm->verbose_medium()) {
3718 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3719 _worker_id, n);
3720 }
3721 statsOnly( int tmp_size = _cm->mark_stack_size();
3722 if (tmp_size > _global_max_size) {
3723 _global_max_size = tmp_size;
3724 }
3725 _global_pushes += n );
3726 }
3727 }
3728
3729 // this operation was quite expensive, so decrease the limits
3730 decrease_limits();
3731 }
3732
3733 void CMTask::get_entries_from_global_stack() {
3734 // local array where we'll store the entries that will be popped
3735 // from the global stack.
3736 oop buffer[global_stack_transfer_size];
3737 int n;
3738 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3739 assert(n <= global_stack_transfer_size,
3740 "we should not pop more than the given limit");
3741 if (n > 0) {
3742 // yes, we did actually pop at least one entry
3743
3744 statsOnly( ++_global_transfers_from; _global_pops += n );
3745 if (_cm->verbose_medium()) {
3746 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3747 _worker_id, n);
3748 }
3749 for (int i = 0; i < n; ++i) {
3750 bool success = _task_queue->push(buffer[i]);
3751 // We only call this when the local queue is empty or under a
3752 // given target limit. So, we do not expect this push to fail.
3753 assert(success, "invariant");
3754 }
3755
3756 statsOnly( int tmp_size = _task_queue->size();
3757 if (tmp_size > _local_max_size) {
3758 _local_max_size = tmp_size;
3759 }
3760 _local_pushes += n );
3761 }
3762
3763 // this operation was quite expensive, so decrease the limits
3764 decrease_limits();
3765 }
3766
3767 void CMTask::drain_local_queue(bool partially) {
3768 if (has_aborted()) return;
3769
3770 // Decide what the target size is, depending whether we're going to
3771 // drain it partially (so that other tasks can steal if they run out
3772 // of things to do) or totally (at the very end).
3773 size_t target_size;
3774 if (partially) {
3775 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3776 } else {
3777 target_size = 0;
3778 }
3779
3780 if (_task_queue->size() > target_size) {
3781 if (_cm->verbose_high()) {
3782 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3783 _worker_id, target_size);
3784 }
3785
3786 oop obj;
3787 bool ret = _task_queue->pop_local(obj);
3788 while (ret) {
3789 statsOnly( ++_local_pops );
3790
3791 if (_cm->verbose_high()) {
3792 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3793 p2i((void*) obj));
3794 }
3795
3796 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3797 assert(!_g1h->is_on_master_free_list(
3798 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3799
3800 scan_object(obj);
3801
3802 if (_task_queue->size() <= target_size || has_aborted()) {
3803 ret = false;
3804 } else {
3805 ret = _task_queue->pop_local(obj);
3806 }
3807 }
3808
3809 if (_cm->verbose_high()) {
3810 gclog_or_tty->print_cr("[%u] drained local queue, size = %u",
3811 _worker_id, _task_queue->size());
3812 }
3813 }
3814 }
3815
3816 void CMTask::drain_global_stack(bool partially) {
3817 if (has_aborted()) return;
3818
3819 // We have a policy to drain the local queue before we attempt to
3820 // drain the global stack.
3821 assert(partially || _task_queue->size() == 0, "invariant");
3822
3823 // Decide what the target size is, depending whether we're going to
3824 // drain it partially (so that other tasks can steal if they run out
3825 // of things to do) or totally (at the very end). Notice that,
3826 // because we move entries from the global stack in chunks or
3827 // because another task might be doing the same, we might in fact
3828 // drop below the target. But, this is not a problem.
3829 size_t target_size;
3830 if (partially) {
3831 target_size = _cm->partial_mark_stack_size_target();
3832 } else {
3833 target_size = 0;
3834 }
3835
3836 if (_cm->mark_stack_size() > target_size) {
3837 if (_cm->verbose_low()) {
3838 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3839 _worker_id, target_size);
3840 }
3841
3842 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3843 get_entries_from_global_stack();
3844 drain_local_queue(partially);
3845 }
3846
3847 if (_cm->verbose_low()) {
3848 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3849 _worker_id, _cm->mark_stack_size());
3850 }
3851 }
3852 }
3853
3854 // SATB Queue has several assumptions on whether to call the par or
3855 // non-par versions of the methods. this is why some of the code is
3856 // replicated. We should really get rid of the single-threaded version
3857 // of the code to simplify things.
3858 void CMTask::drain_satb_buffers() {
3859 if (has_aborted()) return;
3860
3861 // We set this so that the regular clock knows that we're in the
3862 // middle of draining buffers and doesn't set the abort flag when it
3863 // notices that SATB buffers are available for draining. It'd be
3864 // very counter productive if it did that. :-)
3865 _draining_satb_buffers = true;
3866
3867 CMObjectClosure oc(this);
3868 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3869 if (G1CollectedHeap::use_parallel_gc_threads()) {
3870 satb_mq_set.set_par_closure(_worker_id, &oc);
3871 } else {
3872 satb_mq_set.set_closure(&oc);
3873 }
3874
3875 // This keeps claiming and applying the closure to completed buffers
3876 // until we run out of buffers or we need to abort.
3877 if (G1CollectedHeap::use_parallel_gc_threads()) {
3878 while (!has_aborted() &&
3879 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
3880 if (_cm->verbose_medium()) {
3881 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3882 }
3883 statsOnly( ++_satb_buffers_processed );
3884 regular_clock_call();
3885 }
3886 } else {
3887 while (!has_aborted() &&
3888 satb_mq_set.apply_closure_to_completed_buffer()) {
3889 if (_cm->verbose_medium()) {
3890 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3891 }
3892 statsOnly( ++_satb_buffers_processed );
3893 regular_clock_call();
3894 }
3895 }
3896
3897 if (!concurrent() && !has_aborted()) {
3898 // We should only do this during remark.
3899 if (G1CollectedHeap::use_parallel_gc_threads()) {
3900 satb_mq_set.par_iterate_closure_all_threads(_worker_id);
3901 } else {
3902 satb_mq_set.iterate_closure_all_threads();
3903 }
3904 }
3905
3906 _draining_satb_buffers = false;
3907
3908 assert(has_aborted() ||
3909 concurrent() ||
3910 satb_mq_set.completed_buffers_num() == 0, "invariant");
3911
3912 if (G1CollectedHeap::use_parallel_gc_threads()) {
3913 satb_mq_set.set_par_closure(_worker_id, NULL);
3914 } else {
3915 satb_mq_set.set_closure(NULL);
3916 }
3917
3918 // again, this was a potentially expensive operation, decrease the
3919 // limits to get the regular clock call early
3920 decrease_limits();
3921 }
3922
3923 void CMTask::print_stats() {
3924 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3925 _worker_id, _calls);
3926 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3927 _elapsed_time_ms, _termination_time_ms);
3928 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3929 _step_times_ms.num(), _step_times_ms.avg(),
3930 _step_times_ms.sd());
3931 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3932 _step_times_ms.maximum(), _step_times_ms.sum());
3933
3934 #if _MARKING_STATS_
3935 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3936 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3937 _all_clock_intervals_ms.sd());
3938 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
3939 _all_clock_intervals_ms.maximum(),
3940 _all_clock_intervals_ms.sum());
3941 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
3942 _clock_due_to_scanning, _clock_due_to_marking);
3943 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
3944 _objs_scanned, _objs_found_on_bitmap);
3945 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
3946 _local_pushes, _local_pops, _local_max_size);
3947 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
3948 _global_pushes, _global_pops, _global_max_size);
3949 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
3950 _global_transfers_to,_global_transfers_from);
3951 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
3952 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
3953 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
3954 _steal_attempts, _steals);
3955 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
3956 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
3957 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3958 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
3959 _aborted_timed_out, _aborted_satb, _aborted_termination);
3960 #endif // _MARKING_STATS_
3961 }
3962
3963 /*****************************************************************************
3964
3965 The do_marking_step(time_target_ms, ...) method is the building
3966 block of the parallel marking framework. It can be called in parallel
3967 with other invocations of do_marking_step() on different tasks
3968 (but only one per task, obviously) and concurrently with the
3969 mutator threads, or during remark, hence it eliminates the need
3970 for two versions of the code. When called during remark, it will
3971 pick up from where the task left off during the concurrent marking
3972 phase. Interestingly, tasks are also claimable during evacuation
3973 pauses too, since do_marking_step() ensures that it aborts before
3974 it needs to yield.
3975
3976 The data structures that it uses to do marking work are the
3977 following:
3978
3979 (1) Marking Bitmap. If there are gray objects that appear only
3980 on the bitmap (this happens either when dealing with an overflow
3981 or when the initial marking phase has simply marked the roots
3982 and didn't push them on the stack), then tasks claim heap
3983 regions whose bitmap they then scan to find gray objects. A
3984 global finger indicates where the end of the last claimed region
3985 is. A local finger indicates how far into the region a task has
3986 scanned. The two fingers are used to determine how to gray an
3987 object (i.e. whether simply marking it is OK, as it will be
3988 visited by a task in the future, or whether it needs to be also
3989 pushed on a stack).
3990
3991 (2) Local Queue. The local queue of the task which is accessed
3992 reasonably efficiently by the task. Other tasks can steal from
3993 it when they run out of work. Throughout the marking phase, a
3994 task attempts to keep its local queue short but not totally
3995 empty, so that entries are available for stealing by other
3996 tasks. Only when there is no more work, a task will totally
3997 drain its local queue.
3998
3999 (3) Global Mark Stack. This handles local queue overflow. During
4000 marking only sets of entries are moved between it and the local
4001 queues, as access to it requires a mutex and more fine-grain
4002 interaction with it which might cause contention. If it
4003 overflows, then the marking phase should restart and iterate
4004 over the bitmap to identify gray objects. Throughout the marking
4005 phase, tasks attempt to keep the global mark stack at a small
4006 length but not totally empty, so that entries are available for
4007 popping by other tasks. Only when there is no more work, tasks
4008 will totally drain the global mark stack.
4009
4010 (4) SATB Buffer Queue. This is where completed SATB buffers are
4011 made available. Buffers are regularly removed from this queue
4012 and scanned for roots, so that the queue doesn't get too
4013 long. During remark, all completed buffers are processed, as
4014 well as the filled in parts of any uncompleted buffers.
4015
4016 The do_marking_step() method tries to abort when the time target
4017 has been reached. There are a few other cases when the
4018 do_marking_step() method also aborts:
4019
4020 (1) When the marking phase has been aborted (after a Full GC).
4021
4022 (2) When a global overflow (on the global stack) has been
4023 triggered. Before the task aborts, it will actually sync up with
4024 the other tasks to ensure that all the marking data structures
4025 (local queues, stacks, fingers etc.) are re-initialized so that
4026 when do_marking_step() completes, the marking phase can
4027 immediately restart.
4028
4029 (3) When enough completed SATB buffers are available. The
4030 do_marking_step() method only tries to drain SATB buffers right
4031 at the beginning. So, if enough buffers are available, the
4032 marking step aborts and the SATB buffers are processed at
4033 the beginning of the next invocation.
4034
4035 (4) To yield. when we have to yield then we abort and yield
4036 right at the end of do_marking_step(). This saves us from a lot
4037 of hassle as, by yielding we might allow a Full GC. If this
4038 happens then objects will be compacted underneath our feet, the
4039 heap might shrink, etc. We save checking for this by just
4040 aborting and doing the yield right at the end.
4041
4042 From the above it follows that the do_marking_step() method should
4043 be called in a loop (or, otherwise, regularly) until it completes.
4044
4045 If a marking step completes without its has_aborted() flag being
4046 true, it means it has completed the current marking phase (and
4047 also all other marking tasks have done so and have all synced up).
4048
4049 A method called regular_clock_call() is invoked "regularly" (in
4050 sub ms intervals) throughout marking. It is this clock method that
4051 checks all the abort conditions which were mentioned above and
4052 decides when the task should abort. A work-based scheme is used to
4053 trigger this clock method: when the number of object words the
4054 marking phase has scanned or the number of references the marking
4055 phase has visited reach a given limit. Additional invocations to
4056 the method clock have been planted in a few other strategic places
4057 too. The initial reason for the clock method was to avoid calling
4058 vtime too regularly, as it is quite expensive. So, once it was in
4059 place, it was natural to piggy-back all the other conditions on it
4060 too and not constantly check them throughout the code.
4061
4062 If do_termination is true then do_marking_step will enter its
4063 termination protocol.
4064
4065 The value of is_serial must be true when do_marking_step is being
4066 called serially (i.e. by the VMThread) and do_marking_step should
4067 skip any synchronization in the termination and overflow code.
4068 Examples include the serial remark code and the serial reference
4069 processing closures.
4070
4071 The value of is_serial must be false when do_marking_step is
4072 being called by any of the worker threads in a work gang.
4073 Examples include the concurrent marking code (CMMarkingTask),
4074 the MT remark code, and the MT reference processing closures.
4075
4076 *****************************************************************************/
4077
4078 void CMTask::do_marking_step(double time_target_ms,
4079 bool do_termination,
4080 bool is_serial) {
4081 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4082 assert(concurrent() == _cm->concurrent(), "they should be the same");
4083
4084 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4085 assert(_task_queues != NULL, "invariant");
4086 assert(_task_queue != NULL, "invariant");
4087 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4088
4089 assert(!_claimed,
4090 "only one thread should claim this task at any one time");
4091
4092 // OK, this doesn't safeguard again all possible scenarios, as it is
4093 // possible for two threads to set the _claimed flag at the same
4094 // time. But it is only for debugging purposes anyway and it will
4095 // catch most problems.
4096 _claimed = true;
4097
4098 _start_time_ms = os::elapsedVTime() * 1000.0;
4099 statsOnly( _interval_start_time_ms = _start_time_ms );
4100
4101 // If do_stealing is true then do_marking_step will attempt to
4102 // steal work from the other CMTasks. It only makes sense to
4103 // enable stealing when the termination protocol is enabled
4104 // and do_marking_step() is not being called serially.
4105 bool do_stealing = do_termination && !is_serial;
4106
4107 double diff_prediction_ms =
4108 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4109 _time_target_ms = time_target_ms - diff_prediction_ms;
4110
4111 // set up the variables that are used in the work-based scheme to
4112 // call the regular clock method
4113 _words_scanned = 0;
4114 _refs_reached = 0;
4115 recalculate_limits();
4116
4117 // clear all flags
4118 clear_has_aborted();
4119 _has_timed_out = false;
4120 _draining_satb_buffers = false;
4121
4122 ++_calls;
4123
4124 if (_cm->verbose_low()) {
4125 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4126 "target = %1.2lfms >>>>>>>>>>",
4127 _worker_id, _calls, _time_target_ms);
4128 }
4129
4130 // Set up the bitmap and oop closures. Anything that uses them is
4131 // eventually called from this method, so it is OK to allocate these
4132 // statically.
4133 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4134 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4135 set_cm_oop_closure(&cm_oop_closure);
4136
4137 if (_cm->has_overflown()) {
4138 // This can happen if the mark stack overflows during a GC pause
4139 // and this task, after a yield point, restarts. We have to abort
4140 // as we need to get into the overflow protocol which happens
4141 // right at the end of this task.
4142 set_has_aborted();
4143 }
4144
4145 // First drain any available SATB buffers. After this, we will not
4146 // look at SATB buffers before the next invocation of this method.
4147 // If enough completed SATB buffers are queued up, the regular clock
4148 // will abort this task so that it restarts.
4149 drain_satb_buffers();
4150 // ...then partially drain the local queue and the global stack
4151 drain_local_queue(true);
4152 drain_global_stack(true);
4153
4154 do {
4155 if (!has_aborted() && _curr_region != NULL) {
4156 // This means that we're already holding on to a region.
4157 assert(_finger != NULL, "if region is not NULL, then the finger "
4158 "should not be NULL either");
4159
4160 // We might have restarted this task after an evacuation pause
4161 // which might have evacuated the region we're holding on to
4162 // underneath our feet. Let's read its limit again to make sure
4163 // that we do not iterate over a region of the heap that
4164 // contains garbage (update_region_limit() will also move
4165 // _finger to the start of the region if it is found empty).
4166 update_region_limit();
4167 // We will start from _finger not from the start of the region,
4168 // as we might be restarting this task after aborting half-way
4169 // through scanning this region. In this case, _finger points to
4170 // the address where we last found a marked object. If this is a
4171 // fresh region, _finger points to start().
4172 MemRegion mr = MemRegion(_finger, _region_limit);
4173
4174 if (_cm->verbose_low()) {
4175 gclog_or_tty->print_cr("[%u] we're scanning part "
4176 "["PTR_FORMAT", "PTR_FORMAT") "
4177 "of region "HR_FORMAT,
4178 _worker_id, p2i(_finger), p2i(_region_limit),
4179 HR_FORMAT_PARAMS(_curr_region));
4180 }
4181
4182 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4183 "humongous regions should go around loop once only");
4184
4185 // Some special cases:
4186 // If the memory region is empty, we can just give up the region.
4187 // If the current region is humongous then we only need to check
4188 // the bitmap for the bit associated with the start of the object,
4189 // scan the object if it's live, and give up the region.
4190 // Otherwise, let's iterate over the bitmap of the part of the region
4191 // that is left.
4192 // If the iteration is successful, give up the region.
4193 if (mr.is_empty()) {
4194 giveup_current_region();
4195 regular_clock_call();
4196 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4197 if (_nextMarkBitMap->isMarked(mr.start())) {
4198 // The object is marked - apply the closure
4199 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4200 bitmap_closure.do_bit(offset);
4201 }
4202 // Even if this task aborted while scanning the humongous object
4203 // we can (and should) give up the current region.
4204 giveup_current_region();
4205 regular_clock_call();
4206 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4207 giveup_current_region();
4208 regular_clock_call();
4209 } else {
4210 assert(has_aborted(), "currently the only way to do so");
4211 // The only way to abort the bitmap iteration is to return
4212 // false from the do_bit() method. However, inside the
4213 // do_bit() method we move the _finger to point to the
4214 // object currently being looked at. So, if we bail out, we
4215 // have definitely set _finger to something non-null.
4216 assert(_finger != NULL, "invariant");
4217
4218 // Region iteration was actually aborted. So now _finger
4219 // points to the address of the object we last scanned. If we
4220 // leave it there, when we restart this task, we will rescan
4221 // the object. It is easy to avoid this. We move the finger by
4222 // enough to point to the next possible object header (the
4223 // bitmap knows by how much we need to move it as it knows its
4224 // granularity).
4225 assert(_finger < _region_limit, "invariant");
4226 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4227 // Check if bitmap iteration was aborted while scanning the last object
4228 if (new_finger >= _region_limit) {
4229 giveup_current_region();
4230 } else {
4231 move_finger_to(new_finger);
4232 }
4233 }
4234 }
4235 // At this point we have either completed iterating over the
4236 // region we were holding on to, or we have aborted.
4237
4238 // We then partially drain the local queue and the global stack.
4239 // (Do we really need this?)
4240 drain_local_queue(true);
4241 drain_global_stack(true);
4242
4243 // Read the note on the claim_region() method on why it might
4244 // return NULL with potentially more regions available for
4245 // claiming and why we have to check out_of_regions() to determine
4246 // whether we're done or not.
4247 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4248 // We are going to try to claim a new region. We should have
4249 // given up on the previous one.
4250 // Separated the asserts so that we know which one fires.
4251 assert(_curr_region == NULL, "invariant");
4252 assert(_finger == NULL, "invariant");
4253 assert(_region_limit == NULL, "invariant");
4254 if (_cm->verbose_low()) {
4255 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4256 }
4257 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4258 if (claimed_region != NULL) {
4259 // Yes, we managed to claim one
4260 statsOnly( ++_regions_claimed );
4261
4262 if (_cm->verbose_low()) {
4263 gclog_or_tty->print_cr("[%u] we successfully claimed "
4264 "region "PTR_FORMAT,
4265 _worker_id, p2i(claimed_region));
4266 }
4267
4268 setup_for_region(claimed_region);
4269 assert(_curr_region == claimed_region, "invariant");
4270 }
4271 // It is important to call the regular clock here. It might take
4272 // a while to claim a region if, for example, we hit a large
4273 // block of empty regions. So we need to call the regular clock
4274 // method once round the loop to make sure it's called
4275 // frequently enough.
4276 regular_clock_call();
4277 }
4278
4279 if (!has_aborted() && _curr_region == NULL) {
4280 assert(_cm->out_of_regions(),
4281 "at this point we should be out of regions");
4282 }
4283 } while ( _curr_region != NULL && !has_aborted());
4284
4285 if (!has_aborted()) {
4286 // We cannot check whether the global stack is empty, since other
4287 // tasks might be pushing objects to it concurrently.
4288 assert(_cm->out_of_regions(),
4289 "at this point we should be out of regions");
4290
4291 if (_cm->verbose_low()) {
4292 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4293 }
4294
4295 // Try to reduce the number of available SATB buffers so that
4296 // remark has less work to do.
4297 drain_satb_buffers();
4298 }
4299
4300 // Since we've done everything else, we can now totally drain the
4301 // local queue and global stack.
4302 drain_local_queue(false);
4303 drain_global_stack(false);
4304
4305 // Attempt at work stealing from other task's queues.
4306 if (do_stealing && !has_aborted()) {
4307 // We have not aborted. This means that we have finished all that
4308 // we could. Let's try to do some stealing...
4309
4310 // We cannot check whether the global stack is empty, since other
4311 // tasks might be pushing objects to it concurrently.
4312 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4313 "only way to reach here");
4314
4315 if (_cm->verbose_low()) {
4316 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4317 }
4318
4319 while (!has_aborted()) {
4320 oop obj;
4321 statsOnly( ++_steal_attempts );
4322
4323 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4324 if (_cm->verbose_medium()) {
4325 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4326 _worker_id, p2i((void*) obj));
4327 }
4328
4329 statsOnly( ++_steals );
4330
4331 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4332 "any stolen object should be marked");
4333 scan_object(obj);
4334
4335 // And since we're towards the end, let's totally drain the
4336 // local queue and global stack.
4337 drain_local_queue(false);
4338 drain_global_stack(false);
4339 } else {
4340 break;
4341 }
4342 }
4343 }
4344
4345 // If we are about to wrap up and go into termination, check if we
4346 // should raise the overflow flag.
4347 if (do_termination && !has_aborted()) {
4348 if (_cm->force_overflow()->should_force()) {
4349 _cm->set_has_overflown();
4350 regular_clock_call();
4351 }
4352 }
4353
4354 // We still haven't aborted. Now, let's try to get into the
4355 // termination protocol.
4356 if (do_termination && !has_aborted()) {
4357 // We cannot check whether the global stack is empty, since other
4358 // tasks might be concurrently pushing objects on it.
4359 // Separated the asserts so that we know which one fires.
4360 assert(_cm->out_of_regions(), "only way to reach here");
4361 assert(_task_queue->size() == 0, "only way to reach here");
4362
4363 if (_cm->verbose_low()) {
4364 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4365 }
4366
4367 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4368
4369 // The CMTask class also extends the TerminatorTerminator class,
4370 // hence its should_exit_termination() method will also decide
4371 // whether to exit the termination protocol or not.
4372 bool finished = (is_serial ||
4373 _cm->terminator()->offer_termination(this));
4374 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4375 _termination_time_ms +=
4376 termination_end_time_ms - _termination_start_time_ms;
4377
4378 if (finished) {
4379 // We're all done.
4380
4381 if (_worker_id == 0) {
4382 // let's allow task 0 to do this
4383 if (concurrent()) {
4384 assert(_cm->concurrent_marking_in_progress(), "invariant");
4385 // we need to set this to false before the next
4386 // safepoint. This way we ensure that the marking phase
4387 // doesn't observe any more heap expansions.
4388 _cm->clear_concurrent_marking_in_progress();
4389 }
4390 }
4391
4392 // We can now guarantee that the global stack is empty, since
4393 // all other tasks have finished. We separated the guarantees so
4394 // that, if a condition is false, we can immediately find out
4395 // which one.
4396 guarantee(_cm->out_of_regions(), "only way to reach here");
4397 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4398 guarantee(_task_queue->size() == 0, "only way to reach here");
4399 guarantee(!_cm->has_overflown(), "only way to reach here");
4400 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4401
4402 if (_cm->verbose_low()) {
4403 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4404 }
4405 } else {
4406 // Apparently there's more work to do. Let's abort this task. It
4407 // will restart it and we can hopefully find more things to do.
4408
4409 if (_cm->verbose_low()) {
4410 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4411 _worker_id);
4412 }
4413
4414 set_has_aborted();
4415 statsOnly( ++_aborted_termination );
4416 }
4417 }
4418
4419 // Mainly for debugging purposes to make sure that a pointer to the
4420 // closure which was statically allocated in this frame doesn't
4421 // escape it by accident.
4422 set_cm_oop_closure(NULL);
4423 double end_time_ms = os::elapsedVTime() * 1000.0;
4424 double elapsed_time_ms = end_time_ms - _start_time_ms;
4425 // Update the step history.
4426 _step_times_ms.add(elapsed_time_ms);
4427
4428 if (has_aborted()) {
4429 // The task was aborted for some reason.
4430
4431 statsOnly( ++_aborted );
4432
4433 if (_has_timed_out) {
4434 double diff_ms = elapsed_time_ms - _time_target_ms;
4435 // Keep statistics of how well we did with respect to hitting
4436 // our target only if we actually timed out (if we aborted for
4437 // other reasons, then the results might get skewed).
4438 _marking_step_diffs_ms.add(diff_ms);
4439 }
4440
4441 if (_cm->has_overflown()) {
4442 // This is the interesting one. We aborted because a global
4443 // overflow was raised. This means we have to restart the
4444 // marking phase and start iterating over regions. However, in
4445 // order to do this we have to make sure that all tasks stop
4446 // what they are doing and re-initialize in a safe manner. We
4447 // will achieve this with the use of two barrier sync points.
4448
4449 if (_cm->verbose_low()) {
4450 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4451 }
4452
4453 if (!is_serial) {
4454 // We only need to enter the sync barrier if being called
4455 // from a parallel context
4456 _cm->enter_first_sync_barrier(_worker_id);
4457
4458 // When we exit this sync barrier we know that all tasks have
4459 // stopped doing marking work. So, it's now safe to
4460 // re-initialize our data structures. At the end of this method,
4461 // task 0 will clear the global data structures.
4462 }
4463
4464 statsOnly( ++_aborted_overflow );
4465
4466 // We clear the local state of this task...
4467 clear_region_fields();
4468
4469 if (!is_serial) {
4470 // ...and enter the second barrier.
4471 _cm->enter_second_sync_barrier(_worker_id);
4472 }
4473 // At this point, if we're during the concurrent phase of
4474 // marking, everything has been re-initialized and we're
4475 // ready to restart.
4476 }
4477
4478 if (_cm->verbose_low()) {
4479 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4480 "elapsed = %1.2lfms <<<<<<<<<<",
4481 _worker_id, _time_target_ms, elapsed_time_ms);
4482 if (_cm->has_aborted()) {
4483 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4484 _worker_id);
4485 }
4486 }
4487 } else {
4488 if (_cm->verbose_low()) {
4489 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4490 "elapsed = %1.2lfms <<<<<<<<<<",
4491 _worker_id, _time_target_ms, elapsed_time_ms);
4492 }
4493 }
4494
4495 _claimed = false;
4496 }
4497
4498 CMTask::CMTask(uint worker_id,
4499 ConcurrentMark* cm,
4500 size_t* marked_bytes,
4501 BitMap* card_bm,
4502 CMTaskQueue* task_queue,
4503 CMTaskQueueSet* task_queues)
4504 : _g1h(G1CollectedHeap::heap()),
4505 _worker_id(worker_id), _cm(cm),
4506 _claimed(false),
4507 _nextMarkBitMap(NULL), _hash_seed(17),
4508 _task_queue(task_queue),
4509 _task_queues(task_queues),
4510 _cm_oop_closure(NULL),
4511 _marked_bytes_array(marked_bytes),
4512 _card_bm(card_bm) {
4513 guarantee(task_queue != NULL, "invariant");
4514 guarantee(task_queues != NULL, "invariant");
4515
4516 statsOnly( _clock_due_to_scanning = 0;
4517 _clock_due_to_marking = 0 );
4518
4519 _marking_step_diffs_ms.add(0.5);
4520 }
4521
4522 // These are formatting macros that are used below to ensure
4523 // consistent formatting. The *_H_* versions are used to format the
4524 // header for a particular value and they should be kept consistent
4525 // with the corresponding macro. Also note that most of the macros add
4526 // the necessary white space (as a prefix) which makes them a bit
4527 // easier to compose.
4528
4529 // All the output lines are prefixed with this string to be able to
4530 // identify them easily in a large log file.
4531 #define G1PPRL_LINE_PREFIX "###"
4532
4533 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4534 #ifdef _LP64
4535 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4536 #else // _LP64
4537 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4538 #endif // _LP64
4539
4540 // For per-region info
4541 #define G1PPRL_TYPE_FORMAT " %-4s"
4542 #define G1PPRL_TYPE_H_FORMAT " %4s"
4543 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4544 #define G1PPRL_BYTE_H_FORMAT " %9s"
4545 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4546 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4547
4548 // For summary info
4549 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4550 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4551 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4552 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4553
4554 G1PrintRegionLivenessInfoClosure::
4555 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4556 : _out(out),
4557 _total_used_bytes(0), _total_capacity_bytes(0),
4558 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4559 _hum_used_bytes(0), _hum_capacity_bytes(0),
4560 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4561 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4562 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4563 MemRegion g1_committed = g1h->g1_committed();
4564 MemRegion g1_reserved = g1h->g1_reserved();
4565 double now = os::elapsedTime();
4566
4567 // Print the header of the output.
4568 _out->cr();
4569 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4570 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4571 G1PPRL_SUM_ADDR_FORMAT("committed")
4572 G1PPRL_SUM_ADDR_FORMAT("reserved")
4573 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4574 p2i(g1_committed.start()), p2i(g1_committed.end()),
4575 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4576 HeapRegion::GrainBytes);
4577 _out->print_cr(G1PPRL_LINE_PREFIX);
4578 _out->print_cr(G1PPRL_LINE_PREFIX
4579 G1PPRL_TYPE_H_FORMAT
4580 G1PPRL_ADDR_BASE_H_FORMAT
4581 G1PPRL_BYTE_H_FORMAT
4582 G1PPRL_BYTE_H_FORMAT
4583 G1PPRL_BYTE_H_FORMAT
4584 G1PPRL_DOUBLE_H_FORMAT
4585 G1PPRL_BYTE_H_FORMAT
4586 G1PPRL_BYTE_H_FORMAT,
4587 "type", "address-range",
4588 "used", "prev-live", "next-live", "gc-eff",
4589 "remset", "code-roots");
4590 _out->print_cr(G1PPRL_LINE_PREFIX
4591 G1PPRL_TYPE_H_FORMAT
4592 G1PPRL_ADDR_BASE_H_FORMAT
4593 G1PPRL_BYTE_H_FORMAT
4594 G1PPRL_BYTE_H_FORMAT
4595 G1PPRL_BYTE_H_FORMAT
4596 G1PPRL_DOUBLE_H_FORMAT
4597 G1PPRL_BYTE_H_FORMAT
4598 G1PPRL_BYTE_H_FORMAT,
4599 "", "",
4600 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4601 "(bytes)", "(bytes)");
4602 }
4603
4604 // It takes as a parameter a reference to one of the _hum_* fields, it
4605 // deduces the corresponding value for a region in a humongous region
4606 // series (either the region size, or what's left if the _hum_* field
4607 // is < the region size), and updates the _hum_* field accordingly.
4608 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4609 size_t bytes = 0;
4610 // The > 0 check is to deal with the prev and next live bytes which
4611 // could be 0.
4612 if (*hum_bytes > 0) {
4613 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4614 *hum_bytes -= bytes;
4615 }
4616 return bytes;
4617 }
4618
4619 // It deduces the values for a region in a humongous region series
4620 // from the _hum_* fields and updates those accordingly. It assumes
4621 // that that _hum_* fields have already been set up from the "starts
4622 // humongous" region and we visit the regions in address order.
4623 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4624 size_t* capacity_bytes,
4625 size_t* prev_live_bytes,
4626 size_t* next_live_bytes) {
4627 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4628 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4629 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4630 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4631 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4632 }
4633
4634 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4635 const char* type = "";
4636 HeapWord* bottom = r->bottom();
4637 HeapWord* end = r->end();
4638 size_t capacity_bytes = r->capacity();
4639 size_t used_bytes = r->used();
4640 size_t prev_live_bytes = r->live_bytes();
4641 size_t next_live_bytes = r->next_live_bytes();
4642 double gc_eff = r->gc_efficiency();
4643 size_t remset_bytes = r->rem_set()->mem_size();
4644 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4645
4646 if (r->used() == 0) {
4647 type = "FREE";
4648 } else if (r->is_survivor()) {
4649 type = "SURV";
4650 } else if (r->is_young()) {
4651 type = "EDEN";
4652 } else if (r->startsHumongous()) {
4653 type = "HUMS";
4654
4655 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4656 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4657 "they should have been zeroed after the last time we used them");
4658 // Set up the _hum_* fields.
4659 _hum_capacity_bytes = capacity_bytes;
4660 _hum_used_bytes = used_bytes;
4661 _hum_prev_live_bytes = prev_live_bytes;
4662 _hum_next_live_bytes = next_live_bytes;
4663 get_hum_bytes(&used_bytes, &capacity_bytes,
4664 &prev_live_bytes, &next_live_bytes);
4665 end = bottom + HeapRegion::GrainWords;
4666 } else if (r->continuesHumongous()) {
4667 type = "HUMC";
4668 get_hum_bytes(&used_bytes, &capacity_bytes,
4669 &prev_live_bytes, &next_live_bytes);
4670 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4671 } else {
4672 type = "OLD";
4673 }
4674
4675 _total_used_bytes += used_bytes;
4676 _total_capacity_bytes += capacity_bytes;
4677 _total_prev_live_bytes += prev_live_bytes;
4678 _total_next_live_bytes += next_live_bytes;
4679 _total_remset_bytes += remset_bytes;
4680 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4681
4682 // Print a line for this particular region.
4683 _out->print_cr(G1PPRL_LINE_PREFIX
4684 G1PPRL_TYPE_FORMAT
4685 G1PPRL_ADDR_BASE_FORMAT
4686 G1PPRL_BYTE_FORMAT
4687 G1PPRL_BYTE_FORMAT
4688 G1PPRL_BYTE_FORMAT
4689 G1PPRL_DOUBLE_FORMAT
4690 G1PPRL_BYTE_FORMAT
4691 G1PPRL_BYTE_FORMAT,
4692 type, p2i(bottom), p2i(end),
4693 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4694 remset_bytes, strong_code_roots_bytes);
4695
4696 return false;
4697 }
4698
4699 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4700 // add static memory usages to remembered set sizes
4701 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4702 // Print the footer of the output.
4703 _out->print_cr(G1PPRL_LINE_PREFIX);
4704 _out->print_cr(G1PPRL_LINE_PREFIX
4705 " SUMMARY"
4706 G1PPRL_SUM_MB_FORMAT("capacity")
4707 G1PPRL_SUM_MB_PERC_FORMAT("used")
4708 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4709 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4710 G1PPRL_SUM_MB_FORMAT("remset")
4711 G1PPRL_SUM_MB_FORMAT("code-roots"),
4712 bytes_to_mb(_total_capacity_bytes),
4713 bytes_to_mb(_total_used_bytes),
4714 perc(_total_used_bytes, _total_capacity_bytes),
4715 bytes_to_mb(_total_prev_live_bytes),
4716 perc(_total_prev_live_bytes, _total_capacity_bytes),
4717 bytes_to_mb(_total_next_live_bytes),
4718 perc(_total_next_live_bytes, _total_capacity_bytes),
4719 bytes_to_mb(_total_remset_bytes),
4720 bytes_to_mb(_total_strong_code_roots_bytes));
4721 _out->cr();
4722 }