Tue, 29 Jul 2014 10:26:09 +0200
8050973: CMS/G1 GC: add missing Resource and Handle mark
Summary: Add Resource/HandleMark in the work() method of some AbstractGangTask to reclaim these resources earlier.
Reviewed-by: tschatzl, goetz
1 /*
2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "code/codeCache.hpp"
28 #include "gc_implementation/g1/concurrentMark.inline.hpp"
29 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
30 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
31 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
32 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
33 #include "gc_implementation/g1/g1Log.hpp"
34 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
35 #include "gc_implementation/g1/g1RemSet.hpp"
36 #include "gc_implementation/g1/heapRegion.inline.hpp"
37 #include "gc_implementation/g1/heapRegionRemSet.hpp"
38 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
39 #include "gc_implementation/shared/vmGCOperations.hpp"
40 #include "gc_implementation/shared/gcTimer.hpp"
41 #include "gc_implementation/shared/gcTrace.hpp"
42 #include "gc_implementation/shared/gcTraceTime.hpp"
43 #include "memory/allocation.hpp"
44 #include "memory/genOopClosures.inline.hpp"
45 #include "memory/referencePolicy.hpp"
46 #include "memory/resourceArea.hpp"
47 #include "oops/oop.inline.hpp"
48 #include "runtime/handles.inline.hpp"
49 #include "runtime/java.hpp"
50 #include "runtime/prefetch.inline.hpp"
51 #include "services/memTracker.hpp"
53 // Concurrent marking bit map wrapper
55 CMBitMapRO::CMBitMapRO(int shifter) :
56 _bm(),
57 _shifter(shifter) {
58 _bmStartWord = 0;
59 _bmWordSize = 0;
60 }
62 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
63 const HeapWord* limit) const {
64 // First we must round addr *up* to a possible object boundary.
65 addr = (HeapWord*)align_size_up((intptr_t)addr,
66 HeapWordSize << _shifter);
67 size_t addrOffset = heapWordToOffset(addr);
68 if (limit == NULL) {
69 limit = _bmStartWord + _bmWordSize;
70 }
71 size_t limitOffset = heapWordToOffset(limit);
72 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
73 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
74 assert(nextAddr >= addr, "get_next_one postcondition");
75 assert(nextAddr == limit || isMarked(nextAddr),
76 "get_next_one postcondition");
77 return nextAddr;
78 }
80 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
81 const HeapWord* limit) const {
82 size_t addrOffset = heapWordToOffset(addr);
83 if (limit == NULL) {
84 limit = _bmStartWord + _bmWordSize;
85 }
86 size_t limitOffset = heapWordToOffset(limit);
87 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
88 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
89 assert(nextAddr >= addr, "get_next_one postcondition");
90 assert(nextAddr == limit || !isMarked(nextAddr),
91 "get_next_one postcondition");
92 return nextAddr;
93 }
95 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
96 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
97 return (int) (diff >> _shifter);
98 }
100 #ifndef PRODUCT
101 bool CMBitMapRO::covers(ReservedSpace heap_rs) const {
102 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
103 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
104 "size inconsistency");
105 return _bmStartWord == (HeapWord*)(heap_rs.base()) &&
106 _bmWordSize == heap_rs.size()>>LogHeapWordSize;
107 }
108 #endif
110 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
111 _bm.print_on_error(st, prefix);
112 }
114 bool CMBitMap::allocate(ReservedSpace heap_rs) {
115 _bmStartWord = (HeapWord*)(heap_rs.base());
116 _bmWordSize = heap_rs.size()/HeapWordSize; // heap_rs.size() is in bytes
117 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
118 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
119 if (!brs.is_reserved()) {
120 warning("ConcurrentMark marking bit map allocation failure");
121 return false;
122 }
123 MemTracker::record_virtual_memory_type((address)brs.base(), mtGC);
124 // For now we'll just commit all of the bit map up front.
125 // Later on we'll try to be more parsimonious with swap.
126 if (!_virtual_space.initialize(brs, brs.size())) {
127 warning("ConcurrentMark marking bit map backing store failure");
128 return false;
129 }
130 assert(_virtual_space.committed_size() == brs.size(),
131 "didn't reserve backing store for all of concurrent marking bit map?");
132 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
133 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
134 _bmWordSize, "inconsistency in bit map sizing");
135 _bm.set_size(_bmWordSize >> _shifter);
136 return true;
137 }
139 void CMBitMap::clearAll() {
140 _bm.clear();
141 return;
142 }
144 void CMBitMap::markRange(MemRegion mr) {
145 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
146 assert(!mr.is_empty(), "unexpected empty region");
147 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
148 ((HeapWord *) mr.end())),
149 "markRange memory region end is not card aligned");
150 // convert address range into offset range
151 _bm.at_put_range(heapWordToOffset(mr.start()),
152 heapWordToOffset(mr.end()), true);
153 }
155 void CMBitMap::clearRange(MemRegion mr) {
156 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
157 assert(!mr.is_empty(), "unexpected empty region");
158 // convert address range into offset range
159 _bm.at_put_range(heapWordToOffset(mr.start()),
160 heapWordToOffset(mr.end()), false);
161 }
163 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
164 HeapWord* end_addr) {
165 HeapWord* start = getNextMarkedWordAddress(addr);
166 start = MIN2(start, end_addr);
167 HeapWord* end = getNextUnmarkedWordAddress(start);
168 end = MIN2(end, end_addr);
169 assert(start <= end, "Consistency check");
170 MemRegion mr(start, end);
171 if (!mr.is_empty()) {
172 clearRange(mr);
173 }
174 return mr;
175 }
177 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
178 _base(NULL), _cm(cm)
179 #ifdef ASSERT
180 , _drain_in_progress(false)
181 , _drain_in_progress_yields(false)
182 #endif
183 {}
185 bool CMMarkStack::allocate(size_t capacity) {
186 // allocate a stack of the requisite depth
187 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
188 if (!rs.is_reserved()) {
189 warning("ConcurrentMark MarkStack allocation failure");
190 return false;
191 }
192 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
193 if (!_virtual_space.initialize(rs, rs.size())) {
194 warning("ConcurrentMark MarkStack backing store failure");
195 // Release the virtual memory reserved for the marking stack
196 rs.release();
197 return false;
198 }
199 assert(_virtual_space.committed_size() == rs.size(),
200 "Didn't reserve backing store for all of ConcurrentMark stack?");
201 _base = (oop*) _virtual_space.low();
202 setEmpty();
203 _capacity = (jint) capacity;
204 _saved_index = -1;
205 _should_expand = false;
206 NOT_PRODUCT(_max_depth = 0);
207 return true;
208 }
210 void CMMarkStack::expand() {
211 // Called, during remark, if we've overflown the marking stack during marking.
212 assert(isEmpty(), "stack should been emptied while handling overflow");
213 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
214 // Clear expansion flag
215 _should_expand = false;
216 if (_capacity == (jint) MarkStackSizeMax) {
217 if (PrintGCDetails && Verbose) {
218 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
219 }
220 return;
221 }
222 // Double capacity if possible
223 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
224 // Do not give up existing stack until we have managed to
225 // get the double capacity that we desired.
226 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
227 sizeof(oop)));
228 if (rs.is_reserved()) {
229 // Release the backing store associated with old stack
230 _virtual_space.release();
231 // Reinitialize virtual space for new stack
232 if (!_virtual_space.initialize(rs, rs.size())) {
233 fatal("Not enough swap for expanded marking stack capacity");
234 }
235 _base = (oop*)(_virtual_space.low());
236 _index = 0;
237 _capacity = new_capacity;
238 } else {
239 if (PrintGCDetails && Verbose) {
240 // Failed to double capacity, continue;
241 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
242 SIZE_FORMAT"K to " SIZE_FORMAT"K",
243 _capacity / K, new_capacity / K);
244 }
245 }
246 }
248 void CMMarkStack::set_should_expand() {
249 // If we're resetting the marking state because of an
250 // marking stack overflow, record that we should, if
251 // possible, expand the stack.
252 _should_expand = _cm->has_overflown();
253 }
255 CMMarkStack::~CMMarkStack() {
256 if (_base != NULL) {
257 _base = NULL;
258 _virtual_space.release();
259 }
260 }
262 void CMMarkStack::par_push(oop ptr) {
263 while (true) {
264 if (isFull()) {
265 _overflow = true;
266 return;
267 }
268 // Otherwise...
269 jint index = _index;
270 jint next_index = index+1;
271 jint res = Atomic::cmpxchg(next_index, &_index, index);
272 if (res == index) {
273 _base[index] = ptr;
274 // Note that we don't maintain this atomically. We could, but it
275 // doesn't seem necessary.
276 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
277 return;
278 }
279 // Otherwise, we need to try again.
280 }
281 }
283 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
284 while (true) {
285 if (isFull()) {
286 _overflow = true;
287 return;
288 }
289 // Otherwise...
290 jint index = _index;
291 jint next_index = index + n;
292 if (next_index > _capacity) {
293 _overflow = true;
294 return;
295 }
296 jint res = Atomic::cmpxchg(next_index, &_index, index);
297 if (res == index) {
298 for (int i = 0; i < n; i++) {
299 int ind = index + i;
300 assert(ind < _capacity, "By overflow test above.");
301 _base[ind] = ptr_arr[i];
302 }
303 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
304 return;
305 }
306 // Otherwise, we need to try again.
307 }
308 }
310 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
311 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
312 jint start = _index;
313 jint next_index = start + n;
314 if (next_index > _capacity) {
315 _overflow = true;
316 return;
317 }
318 // Otherwise.
319 _index = next_index;
320 for (int i = 0; i < n; i++) {
321 int ind = start + i;
322 assert(ind < _capacity, "By overflow test above.");
323 _base[ind] = ptr_arr[i];
324 }
325 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
326 }
328 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
329 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
330 jint index = _index;
331 if (index == 0) {
332 *n = 0;
333 return false;
334 } else {
335 int k = MIN2(max, index);
336 jint new_ind = index - k;
337 for (int j = 0; j < k; j++) {
338 ptr_arr[j] = _base[new_ind + j];
339 }
340 _index = new_ind;
341 *n = k;
342 return true;
343 }
344 }
346 template<class OopClosureClass>
347 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
348 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
349 || SafepointSynchronize::is_at_safepoint(),
350 "Drain recursion must be yield-safe.");
351 bool res = true;
352 debug_only(_drain_in_progress = true);
353 debug_only(_drain_in_progress_yields = yield_after);
354 while (!isEmpty()) {
355 oop newOop = pop();
356 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
357 assert(newOop->is_oop(), "Expected an oop");
358 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
359 "only grey objects on this stack");
360 newOop->oop_iterate(cl);
361 if (yield_after && _cm->do_yield_check()) {
362 res = false;
363 break;
364 }
365 }
366 debug_only(_drain_in_progress = false);
367 return res;
368 }
370 void CMMarkStack::note_start_of_gc() {
371 assert(_saved_index == -1,
372 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
373 _saved_index = _index;
374 }
376 void CMMarkStack::note_end_of_gc() {
377 // This is intentionally a guarantee, instead of an assert. If we
378 // accidentally add something to the mark stack during GC, it
379 // will be a correctness issue so it's better if we crash. we'll
380 // only check this once per GC anyway, so it won't be a performance
381 // issue in any way.
382 guarantee(_saved_index == _index,
383 err_msg("saved index: %d index: %d", _saved_index, _index));
384 _saved_index = -1;
385 }
387 void CMMarkStack::oops_do(OopClosure* f) {
388 assert(_saved_index == _index,
389 err_msg("saved index: %d index: %d", _saved_index, _index));
390 for (int i = 0; i < _index; i += 1) {
391 f->do_oop(&_base[i]);
392 }
393 }
395 bool ConcurrentMark::not_yet_marked(oop obj) const {
396 return _g1h->is_obj_ill(obj);
397 }
399 CMRootRegions::CMRootRegions() :
400 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
401 _should_abort(false), _next_survivor(NULL) { }
403 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
404 _young_list = g1h->young_list();
405 _cm = cm;
406 }
408 void CMRootRegions::prepare_for_scan() {
409 assert(!scan_in_progress(), "pre-condition");
411 // Currently, only survivors can be root regions.
412 assert(_next_survivor == NULL, "pre-condition");
413 _next_survivor = _young_list->first_survivor_region();
414 _scan_in_progress = (_next_survivor != NULL);
415 _should_abort = false;
416 }
418 HeapRegion* CMRootRegions::claim_next() {
419 if (_should_abort) {
420 // If someone has set the should_abort flag, we return NULL to
421 // force the caller to bail out of their loop.
422 return NULL;
423 }
425 // Currently, only survivors can be root regions.
426 HeapRegion* res = _next_survivor;
427 if (res != NULL) {
428 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
429 // Read it again in case it changed while we were waiting for the lock.
430 res = _next_survivor;
431 if (res != NULL) {
432 if (res == _young_list->last_survivor_region()) {
433 // We just claimed the last survivor so store NULL to indicate
434 // that we're done.
435 _next_survivor = NULL;
436 } else {
437 _next_survivor = res->get_next_young_region();
438 }
439 } else {
440 // Someone else claimed the last survivor while we were trying
441 // to take the lock so nothing else to do.
442 }
443 }
444 assert(res == NULL || res->is_survivor(), "post-condition");
446 return res;
447 }
449 void CMRootRegions::scan_finished() {
450 assert(scan_in_progress(), "pre-condition");
452 // Currently, only survivors can be root regions.
453 if (!_should_abort) {
454 assert(_next_survivor == NULL, "we should have claimed all survivors");
455 }
456 _next_survivor = NULL;
458 {
459 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
460 _scan_in_progress = false;
461 RootRegionScan_lock->notify_all();
462 }
463 }
465 bool CMRootRegions::wait_until_scan_finished() {
466 if (!scan_in_progress()) return false;
468 {
469 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
470 while (scan_in_progress()) {
471 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
472 }
473 }
474 return true;
475 }
477 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
478 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
479 #endif // _MSC_VER
481 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
482 return MAX2((n_par_threads + 2) / 4, 1U);
483 }
485 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs) :
486 _g1h(g1h),
487 _markBitMap1(log2_intptr(MinObjAlignment)),
488 _markBitMap2(log2_intptr(MinObjAlignment)),
489 _parallel_marking_threads(0),
490 _max_parallel_marking_threads(0),
491 _sleep_factor(0.0),
492 _marking_task_overhead(1.0),
493 _cleanup_sleep_factor(0.0),
494 _cleanup_task_overhead(1.0),
495 _cleanup_list("Cleanup List"),
496 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
497 _card_bm((heap_rs.size() + CardTableModRefBS::card_size - 1) >>
498 CardTableModRefBS::card_shift,
499 false /* in_resource_area*/),
501 _prevMarkBitMap(&_markBitMap1),
502 _nextMarkBitMap(&_markBitMap2),
504 _markStack(this),
505 // _finger set in set_non_marking_state
507 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
508 // _active_tasks set in set_non_marking_state
509 // _tasks set inside the constructor
510 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
511 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
513 _has_overflown(false),
514 _concurrent(false),
515 _has_aborted(false),
516 _aborted_gc_id(GCId::undefined()),
517 _restart_for_overflow(false),
518 _concurrent_marking_in_progress(false),
520 // _verbose_level set below
522 _init_times(),
523 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
524 _cleanup_times(),
525 _total_counting_time(0.0),
526 _total_rs_scrub_time(0.0),
528 _parallel_workers(NULL),
530 _count_card_bitmaps(NULL),
531 _count_marked_bytes(NULL),
532 _completed_initialization(false) {
533 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
534 if (verbose_level < no_verbose) {
535 verbose_level = no_verbose;
536 }
537 if (verbose_level > high_verbose) {
538 verbose_level = high_verbose;
539 }
540 _verbose_level = verbose_level;
542 if (verbose_low()) {
543 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
544 "heap end = " INTPTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
545 }
547 if (!_markBitMap1.allocate(heap_rs)) {
548 warning("Failed to allocate first CM bit map");
549 return;
550 }
551 if (!_markBitMap2.allocate(heap_rs)) {
552 warning("Failed to allocate second CM bit map");
553 return;
554 }
556 // Create & start a ConcurrentMark thread.
557 _cmThread = new ConcurrentMarkThread(this);
558 assert(cmThread() != NULL, "CM Thread should have been created");
559 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
560 if (_cmThread->osthread() == NULL) {
561 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
562 }
564 assert(CGC_lock != NULL, "Where's the CGC_lock?");
565 assert(_markBitMap1.covers(heap_rs), "_markBitMap1 inconsistency");
566 assert(_markBitMap2.covers(heap_rs), "_markBitMap2 inconsistency");
568 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
569 satb_qs.set_buffer_size(G1SATBBufferSize);
571 _root_regions.init(_g1h, this);
573 if (ConcGCThreads > ParallelGCThreads) {
574 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
575 "than ParallelGCThreads (" UINTX_FORMAT ").",
576 ConcGCThreads, ParallelGCThreads);
577 return;
578 }
579 if (ParallelGCThreads == 0) {
580 // if we are not running with any parallel GC threads we will not
581 // spawn any marking threads either
582 _parallel_marking_threads = 0;
583 _max_parallel_marking_threads = 0;
584 _sleep_factor = 0.0;
585 _marking_task_overhead = 1.0;
586 } else {
587 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
588 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
589 // if both are set
590 _sleep_factor = 0.0;
591 _marking_task_overhead = 1.0;
592 } else if (G1MarkingOverheadPercent > 0) {
593 // We will calculate the number of parallel marking threads based
594 // on a target overhead with respect to the soft real-time goal
595 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
596 double overall_cm_overhead =
597 (double) MaxGCPauseMillis * marking_overhead /
598 (double) GCPauseIntervalMillis;
599 double cpu_ratio = 1.0 / (double) os::processor_count();
600 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
601 double marking_task_overhead =
602 overall_cm_overhead / marking_thread_num *
603 (double) os::processor_count();
604 double sleep_factor =
605 (1.0 - marking_task_overhead) / marking_task_overhead;
607 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
608 _sleep_factor = sleep_factor;
609 _marking_task_overhead = marking_task_overhead;
610 } else {
611 // Calculate the number of parallel marking threads by scaling
612 // the number of parallel GC threads.
613 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
614 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
615 _sleep_factor = 0.0;
616 _marking_task_overhead = 1.0;
617 }
619 assert(ConcGCThreads > 0, "Should have been set");
620 _parallel_marking_threads = (uint) ConcGCThreads;
621 _max_parallel_marking_threads = _parallel_marking_threads;
623 if (parallel_marking_threads() > 1) {
624 _cleanup_task_overhead = 1.0;
625 } else {
626 _cleanup_task_overhead = marking_task_overhead();
627 }
628 _cleanup_sleep_factor =
629 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
631 #if 0
632 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
633 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
634 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
635 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
636 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
637 #endif
639 guarantee(parallel_marking_threads() > 0, "peace of mind");
640 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
641 _max_parallel_marking_threads, false, true);
642 if (_parallel_workers == NULL) {
643 vm_exit_during_initialization("Failed necessary allocation.");
644 } else {
645 _parallel_workers->initialize_workers();
646 }
647 }
649 if (FLAG_IS_DEFAULT(MarkStackSize)) {
650 uintx mark_stack_size =
651 MIN2(MarkStackSizeMax,
652 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
653 // Verify that the calculated value for MarkStackSize is in range.
654 // It would be nice to use the private utility routine from Arguments.
655 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
656 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
657 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
658 mark_stack_size, (uintx) 1, MarkStackSizeMax);
659 return;
660 }
661 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
662 } else {
663 // Verify MarkStackSize is in range.
664 if (FLAG_IS_CMDLINE(MarkStackSize)) {
665 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
666 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
667 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
668 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
669 MarkStackSize, (uintx) 1, MarkStackSizeMax);
670 return;
671 }
672 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
673 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
674 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
675 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
676 MarkStackSize, MarkStackSizeMax);
677 return;
678 }
679 }
680 }
681 }
683 if (!_markStack.allocate(MarkStackSize)) {
684 warning("Failed to allocate CM marking stack");
685 return;
686 }
688 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
689 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
691 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
692 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
694 BitMap::idx_t card_bm_size = _card_bm.size();
696 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
697 _active_tasks = _max_worker_id;
699 size_t max_regions = (size_t) _g1h->max_regions();
700 for (uint i = 0; i < _max_worker_id; ++i) {
701 CMTaskQueue* task_queue = new CMTaskQueue();
702 task_queue->initialize();
703 _task_queues->register_queue(i, task_queue);
705 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
706 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
708 _tasks[i] = new CMTask(i, this,
709 _count_marked_bytes[i],
710 &_count_card_bitmaps[i],
711 task_queue, _task_queues);
713 _accum_task_vtime[i] = 0.0;
714 }
716 // Calculate the card number for the bottom of the heap. Used
717 // in biasing indexes into the accounting card bitmaps.
718 _heap_bottom_card_num =
719 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
720 CardTableModRefBS::card_shift);
722 // Clear all the liveness counting data
723 clear_all_count_data();
725 // so that the call below can read a sensible value
726 _heap_start = (HeapWord*) heap_rs.base();
727 set_non_marking_state();
728 _completed_initialization = true;
729 }
731 void ConcurrentMark::update_g1_committed(bool force) {
732 // If concurrent marking is not in progress, then we do not need to
733 // update _heap_end.
734 if (!concurrent_marking_in_progress() && !force) return;
736 MemRegion committed = _g1h->g1_committed();
737 assert(committed.start() == _heap_start, "start shouldn't change");
738 HeapWord* new_end = committed.end();
739 if (new_end > _heap_end) {
740 // The heap has been expanded.
742 _heap_end = new_end;
743 }
744 // Notice that the heap can also shrink. However, this only happens
745 // during a Full GC (at least currently) and the entire marking
746 // phase will bail out and the task will not be restarted. So, let's
747 // do nothing.
748 }
750 void ConcurrentMark::reset() {
751 // Starting values for these two. This should be called in a STW
752 // phase. CM will be notified of any future g1_committed expansions
753 // will be at the end of evacuation pauses, when tasks are
754 // inactive.
755 MemRegion committed = _g1h->g1_committed();
756 _heap_start = committed.start();
757 _heap_end = committed.end();
759 // Separated the asserts so that we know which one fires.
760 assert(_heap_start != NULL, "heap bounds should look ok");
761 assert(_heap_end != NULL, "heap bounds should look ok");
762 assert(_heap_start < _heap_end, "heap bounds should look ok");
764 // Reset all the marking data structures and any necessary flags
765 reset_marking_state();
767 if (verbose_low()) {
768 gclog_or_tty->print_cr("[global] resetting");
769 }
771 // We do reset all of them, since different phases will use
772 // different number of active threads. So, it's easiest to have all
773 // of them ready.
774 for (uint i = 0; i < _max_worker_id; ++i) {
775 _tasks[i]->reset(_nextMarkBitMap);
776 }
778 // we need this to make sure that the flag is on during the evac
779 // pause with initial mark piggy-backed
780 set_concurrent_marking_in_progress();
781 }
784 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
785 _markStack.set_should_expand();
786 _markStack.setEmpty(); // Also clears the _markStack overflow flag
787 if (clear_overflow) {
788 clear_has_overflown();
789 } else {
790 assert(has_overflown(), "pre-condition");
791 }
792 _finger = _heap_start;
794 for (uint i = 0; i < _max_worker_id; ++i) {
795 CMTaskQueue* queue = _task_queues->queue(i);
796 queue->set_empty();
797 }
798 }
800 void ConcurrentMark::set_concurrency(uint active_tasks) {
801 assert(active_tasks <= _max_worker_id, "we should not have more");
803 _active_tasks = active_tasks;
804 // Need to update the three data structures below according to the
805 // number of active threads for this phase.
806 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
807 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
808 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
809 }
811 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
812 set_concurrency(active_tasks);
814 _concurrent = concurrent;
815 // We propagate this to all tasks, not just the active ones.
816 for (uint i = 0; i < _max_worker_id; ++i)
817 _tasks[i]->set_concurrent(concurrent);
819 if (concurrent) {
820 set_concurrent_marking_in_progress();
821 } else {
822 // We currently assume that the concurrent flag has been set to
823 // false before we start remark. At this point we should also be
824 // in a STW phase.
825 assert(!concurrent_marking_in_progress(), "invariant");
826 assert(out_of_regions(),
827 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
828 p2i(_finger), p2i(_heap_end)));
829 update_g1_committed(true);
830 }
831 }
833 void ConcurrentMark::set_non_marking_state() {
834 // We set the global marking state to some default values when we're
835 // not doing marking.
836 reset_marking_state();
837 _active_tasks = 0;
838 clear_concurrent_marking_in_progress();
839 }
841 ConcurrentMark::~ConcurrentMark() {
842 // The ConcurrentMark instance is never freed.
843 ShouldNotReachHere();
844 }
846 void ConcurrentMark::clearNextBitmap() {
847 G1CollectedHeap* g1h = G1CollectedHeap::heap();
848 G1CollectorPolicy* g1p = g1h->g1_policy();
850 // Make sure that the concurrent mark thread looks to still be in
851 // the current cycle.
852 guarantee(cmThread()->during_cycle(), "invariant");
854 // We are finishing up the current cycle by clearing the next
855 // marking bitmap and getting it ready for the next cycle. During
856 // this time no other cycle can start. So, let's make sure that this
857 // is the case.
858 guarantee(!g1h->mark_in_progress(), "invariant");
860 // clear the mark bitmap (no grey objects to start with).
861 // We need to do this in chunks and offer to yield in between
862 // each chunk.
863 HeapWord* start = _nextMarkBitMap->startWord();
864 HeapWord* end = _nextMarkBitMap->endWord();
865 HeapWord* cur = start;
866 size_t chunkSize = M;
867 while (cur < end) {
868 HeapWord* next = cur + chunkSize;
869 if (next > end) {
870 next = end;
871 }
872 MemRegion mr(cur,next);
873 _nextMarkBitMap->clearRange(mr);
874 cur = next;
875 do_yield_check();
877 // Repeat the asserts from above. We'll do them as asserts here to
878 // minimize their overhead on the product. However, we'll have
879 // them as guarantees at the beginning / end of the bitmap
880 // clearing to get some checking in the product.
881 assert(cmThread()->during_cycle(), "invariant");
882 assert(!g1h->mark_in_progress(), "invariant");
883 }
885 // Clear the liveness counting data
886 clear_all_count_data();
888 // Repeat the asserts from above.
889 guarantee(cmThread()->during_cycle(), "invariant");
890 guarantee(!g1h->mark_in_progress(), "invariant");
891 }
893 bool ConcurrentMark::nextMarkBitmapIsClear() {
894 return _nextMarkBitMap->getNextMarkedWordAddress(_heap_start, _heap_end) == _heap_end;
895 }
897 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
898 public:
899 bool doHeapRegion(HeapRegion* r) {
900 if (!r->continuesHumongous()) {
901 r->note_start_of_marking();
902 }
903 return false;
904 }
905 };
907 void ConcurrentMark::checkpointRootsInitialPre() {
908 G1CollectedHeap* g1h = G1CollectedHeap::heap();
909 G1CollectorPolicy* g1p = g1h->g1_policy();
911 _has_aborted = false;
913 #ifndef PRODUCT
914 if (G1PrintReachableAtInitialMark) {
915 print_reachable("at-cycle-start",
916 VerifyOption_G1UsePrevMarking, true /* all */);
917 }
918 #endif
920 // Initialise marking structures. This has to be done in a STW phase.
921 reset();
923 // For each region note start of marking.
924 NoteStartOfMarkHRClosure startcl;
925 g1h->heap_region_iterate(&startcl);
926 }
929 void ConcurrentMark::checkpointRootsInitialPost() {
930 G1CollectedHeap* g1h = G1CollectedHeap::heap();
932 // If we force an overflow during remark, the remark operation will
933 // actually abort and we'll restart concurrent marking. If we always
934 // force an oveflow during remark we'll never actually complete the
935 // marking phase. So, we initilize this here, at the start of the
936 // cycle, so that at the remaining overflow number will decrease at
937 // every remark and we'll eventually not need to cause one.
938 force_overflow_stw()->init();
940 // Start Concurrent Marking weak-reference discovery.
941 ReferenceProcessor* rp = g1h->ref_processor_cm();
942 // enable ("weak") refs discovery
943 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
944 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
946 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
947 // This is the start of the marking cycle, we're expected all
948 // threads to have SATB queues with active set to false.
949 satb_mq_set.set_active_all_threads(true, /* new active value */
950 false /* expected_active */);
952 _root_regions.prepare_for_scan();
954 // update_g1_committed() will be called at the end of an evac pause
955 // when marking is on. So, it's also called at the end of the
956 // initial-mark pause to update the heap end, if the heap expands
957 // during it. No need to call it here.
958 }
960 /*
961 * Notice that in the next two methods, we actually leave the STS
962 * during the barrier sync and join it immediately afterwards. If we
963 * do not do this, the following deadlock can occur: one thread could
964 * be in the barrier sync code, waiting for the other thread to also
965 * sync up, whereas another one could be trying to yield, while also
966 * waiting for the other threads to sync up too.
967 *
968 * Note, however, that this code is also used during remark and in
969 * this case we should not attempt to leave / enter the STS, otherwise
970 * we'll either hit an asseert (debug / fastdebug) or deadlock
971 * (product). So we should only leave / enter the STS if we are
972 * operating concurrently.
973 *
974 * Because the thread that does the sync barrier has left the STS, it
975 * is possible to be suspended for a Full GC or an evacuation pause
976 * could occur. This is actually safe, since the entering the sync
977 * barrier is one of the last things do_marking_step() does, and it
978 * doesn't manipulate any data structures afterwards.
979 */
981 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
982 if (verbose_low()) {
983 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
984 }
986 if (concurrent()) {
987 SuspendibleThreadSet::leave();
988 }
990 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
992 if (concurrent()) {
993 SuspendibleThreadSet::join();
994 }
995 // at this point everyone should have synced up and not be doing any
996 // more work
998 if (verbose_low()) {
999 if (barrier_aborted) {
1000 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1001 } else {
1002 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1003 }
1004 }
1006 if (barrier_aborted) {
1007 // If the barrier aborted we ignore the overflow condition and
1008 // just abort the whole marking phase as quickly as possible.
1009 return;
1010 }
1012 // If we're executing the concurrent phase of marking, reset the marking
1013 // state; otherwise the marking state is reset after reference processing,
1014 // during the remark pause.
1015 // If we reset here as a result of an overflow during the remark we will
1016 // see assertion failures from any subsequent set_concurrency_and_phase()
1017 // calls.
1018 if (concurrent()) {
1019 // let the task associated with with worker 0 do this
1020 if (worker_id == 0) {
1021 // task 0 is responsible for clearing the global data structures
1022 // We should be here because of an overflow. During STW we should
1023 // not clear the overflow flag since we rely on it being true when
1024 // we exit this method to abort the pause and restart concurent
1025 // marking.
1026 reset_marking_state(true /* clear_overflow */);
1027 force_overflow()->update();
1029 if (G1Log::fine()) {
1030 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1031 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1032 }
1033 }
1034 }
1036 // after this, each task should reset its own data structures then
1037 // then go into the second barrier
1038 }
1040 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1041 if (verbose_low()) {
1042 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1043 }
1045 if (concurrent()) {
1046 SuspendibleThreadSet::leave();
1047 }
1049 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1051 if (concurrent()) {
1052 SuspendibleThreadSet::join();
1053 }
1054 // at this point everything should be re-initialized and ready to go
1056 if (verbose_low()) {
1057 if (barrier_aborted) {
1058 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1059 } else {
1060 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1061 }
1062 }
1063 }
1065 #ifndef PRODUCT
1066 void ForceOverflowSettings::init() {
1067 _num_remaining = G1ConcMarkForceOverflow;
1068 _force = false;
1069 update();
1070 }
1072 void ForceOverflowSettings::update() {
1073 if (_num_remaining > 0) {
1074 _num_remaining -= 1;
1075 _force = true;
1076 } else {
1077 _force = false;
1078 }
1079 }
1081 bool ForceOverflowSettings::should_force() {
1082 if (_force) {
1083 _force = false;
1084 return true;
1085 } else {
1086 return false;
1087 }
1088 }
1089 #endif // !PRODUCT
1091 class CMConcurrentMarkingTask: public AbstractGangTask {
1092 private:
1093 ConcurrentMark* _cm;
1094 ConcurrentMarkThread* _cmt;
1096 public:
1097 void work(uint worker_id) {
1098 assert(Thread::current()->is_ConcurrentGC_thread(),
1099 "this should only be done by a conc GC thread");
1100 ResourceMark rm;
1102 double start_vtime = os::elapsedVTime();
1104 SuspendibleThreadSet::join();
1106 assert(worker_id < _cm->active_tasks(), "invariant");
1107 CMTask* the_task = _cm->task(worker_id);
1108 the_task->record_start_time();
1109 if (!_cm->has_aborted()) {
1110 do {
1111 double start_vtime_sec = os::elapsedVTime();
1112 double start_time_sec = os::elapsedTime();
1113 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1115 the_task->do_marking_step(mark_step_duration_ms,
1116 true /* do_termination */,
1117 false /* is_serial*/);
1119 double end_time_sec = os::elapsedTime();
1120 double end_vtime_sec = os::elapsedVTime();
1121 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1122 double elapsed_time_sec = end_time_sec - start_time_sec;
1123 _cm->clear_has_overflown();
1125 bool ret = _cm->do_yield_check(worker_id);
1127 jlong sleep_time_ms;
1128 if (!_cm->has_aborted() && the_task->has_aborted()) {
1129 sleep_time_ms =
1130 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1131 SuspendibleThreadSet::leave();
1132 os::sleep(Thread::current(), sleep_time_ms, false);
1133 SuspendibleThreadSet::join();
1134 }
1135 double end_time2_sec = os::elapsedTime();
1136 double elapsed_time2_sec = end_time2_sec - start_time_sec;
1138 #if 0
1139 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, "
1140 "overhead %1.4lf",
1141 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms,
1142 the_task->conc_overhead(os::elapsedTime()) * 8.0);
1143 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms",
1144 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0);
1145 #endif
1146 } while (!_cm->has_aborted() && the_task->has_aborted());
1147 }
1148 the_task->record_end_time();
1149 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1151 SuspendibleThreadSet::leave();
1153 double end_vtime = os::elapsedVTime();
1154 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1155 }
1157 CMConcurrentMarkingTask(ConcurrentMark* cm,
1158 ConcurrentMarkThread* cmt) :
1159 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1161 ~CMConcurrentMarkingTask() { }
1162 };
1164 // Calculates the number of active workers for a concurrent
1165 // phase.
1166 uint ConcurrentMark::calc_parallel_marking_threads() {
1167 if (G1CollectedHeap::use_parallel_gc_threads()) {
1168 uint n_conc_workers = 0;
1169 if (!UseDynamicNumberOfGCThreads ||
1170 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1171 !ForceDynamicNumberOfGCThreads)) {
1172 n_conc_workers = max_parallel_marking_threads();
1173 } else {
1174 n_conc_workers =
1175 AdaptiveSizePolicy::calc_default_active_workers(
1176 max_parallel_marking_threads(),
1177 1, /* Minimum workers */
1178 parallel_marking_threads(),
1179 Threads::number_of_non_daemon_threads());
1180 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1181 // that scaling has already gone into "_max_parallel_marking_threads".
1182 }
1183 assert(n_conc_workers > 0, "Always need at least 1");
1184 return n_conc_workers;
1185 }
1186 // If we are not running with any parallel GC threads we will not
1187 // have spawned any marking threads either. Hence the number of
1188 // concurrent workers should be 0.
1189 return 0;
1190 }
1192 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1193 // Currently, only survivors can be root regions.
1194 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1195 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1197 const uintx interval = PrefetchScanIntervalInBytes;
1198 HeapWord* curr = hr->bottom();
1199 const HeapWord* end = hr->top();
1200 while (curr < end) {
1201 Prefetch::read(curr, interval);
1202 oop obj = oop(curr);
1203 int size = obj->oop_iterate(&cl);
1204 assert(size == obj->size(), "sanity");
1205 curr += size;
1206 }
1207 }
1209 class CMRootRegionScanTask : public AbstractGangTask {
1210 private:
1211 ConcurrentMark* _cm;
1213 public:
1214 CMRootRegionScanTask(ConcurrentMark* cm) :
1215 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1217 void work(uint worker_id) {
1218 assert(Thread::current()->is_ConcurrentGC_thread(),
1219 "this should only be done by a conc GC thread");
1221 CMRootRegions* root_regions = _cm->root_regions();
1222 HeapRegion* hr = root_regions->claim_next();
1223 while (hr != NULL) {
1224 _cm->scanRootRegion(hr, worker_id);
1225 hr = root_regions->claim_next();
1226 }
1227 }
1228 };
1230 void ConcurrentMark::scanRootRegions() {
1231 // Start of concurrent marking.
1232 ClassLoaderDataGraph::clear_claimed_marks();
1234 // scan_in_progress() will have been set to true only if there was
1235 // at least one root region to scan. So, if it's false, we
1236 // should not attempt to do any further work.
1237 if (root_regions()->scan_in_progress()) {
1238 _parallel_marking_threads = calc_parallel_marking_threads();
1239 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1240 "Maximum number of marking threads exceeded");
1241 uint active_workers = MAX2(1U, parallel_marking_threads());
1243 CMRootRegionScanTask task(this);
1244 if (use_parallel_marking_threads()) {
1245 _parallel_workers->set_active_workers((int) active_workers);
1246 _parallel_workers->run_task(&task);
1247 } else {
1248 task.work(0);
1249 }
1251 // It's possible that has_aborted() is true here without actually
1252 // aborting the survivor scan earlier. This is OK as it's
1253 // mainly used for sanity checking.
1254 root_regions()->scan_finished();
1255 }
1256 }
1258 void ConcurrentMark::markFromRoots() {
1259 // we might be tempted to assert that:
1260 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1261 // "inconsistent argument?");
1262 // However that wouldn't be right, because it's possible that
1263 // a safepoint is indeed in progress as a younger generation
1264 // stop-the-world GC happens even as we mark in this generation.
1266 _restart_for_overflow = false;
1267 force_overflow_conc()->init();
1269 // _g1h has _n_par_threads
1270 _parallel_marking_threads = calc_parallel_marking_threads();
1271 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1272 "Maximum number of marking threads exceeded");
1274 uint active_workers = MAX2(1U, parallel_marking_threads());
1276 // Parallel task terminator is set in "set_concurrency_and_phase()"
1277 set_concurrency_and_phase(active_workers, true /* concurrent */);
1279 CMConcurrentMarkingTask markingTask(this, cmThread());
1280 if (use_parallel_marking_threads()) {
1281 _parallel_workers->set_active_workers((int)active_workers);
1282 // Don't set _n_par_threads because it affects MT in process_roots()
1283 // and the decisions on that MT processing is made elsewhere.
1284 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1285 _parallel_workers->run_task(&markingTask);
1286 } else {
1287 markingTask.work(0);
1288 }
1289 print_stats();
1290 }
1292 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1293 // world is stopped at this checkpoint
1294 assert(SafepointSynchronize::is_at_safepoint(),
1295 "world should be stopped");
1297 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1299 // If a full collection has happened, we shouldn't do this.
1300 if (has_aborted()) {
1301 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1302 return;
1303 }
1305 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1307 if (VerifyDuringGC) {
1308 HandleMark hm; // handle scope
1309 Universe::heap()->prepare_for_verify();
1310 Universe::verify(VerifyOption_G1UsePrevMarking,
1311 " VerifyDuringGC:(before)");
1312 }
1314 G1CollectorPolicy* g1p = g1h->g1_policy();
1315 g1p->record_concurrent_mark_remark_start();
1317 double start = os::elapsedTime();
1319 checkpointRootsFinalWork();
1321 double mark_work_end = os::elapsedTime();
1323 weakRefsWork(clear_all_soft_refs);
1325 if (has_overflown()) {
1326 // Oops. We overflowed. Restart concurrent marking.
1327 _restart_for_overflow = true;
1328 if (G1TraceMarkStackOverflow) {
1329 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1330 }
1332 // Verify the heap w.r.t. the previous marking bitmap.
1333 if (VerifyDuringGC) {
1334 HandleMark hm; // handle scope
1335 Universe::heap()->prepare_for_verify();
1336 Universe::verify(VerifyOption_G1UsePrevMarking,
1337 " VerifyDuringGC:(overflow)");
1338 }
1340 // Clear the marking state because we will be restarting
1341 // marking due to overflowing the global mark stack.
1342 reset_marking_state();
1343 } else {
1344 // Aggregate the per-task counting data that we have accumulated
1345 // while marking.
1346 aggregate_count_data();
1348 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1349 // We're done with marking.
1350 // This is the end of the marking cycle, we're expected all
1351 // threads to have SATB queues with active set to true.
1352 satb_mq_set.set_active_all_threads(false, /* new active value */
1353 true /* expected_active */);
1355 if (VerifyDuringGC) {
1356 HandleMark hm; // handle scope
1357 Universe::heap()->prepare_for_verify();
1358 Universe::verify(VerifyOption_G1UseNextMarking,
1359 " VerifyDuringGC:(after)");
1360 }
1361 assert(!restart_for_overflow(), "sanity");
1362 // Completely reset the marking state since marking completed
1363 set_non_marking_state();
1364 }
1366 // Expand the marking stack, if we have to and if we can.
1367 if (_markStack.should_expand()) {
1368 _markStack.expand();
1369 }
1371 // Statistics
1372 double now = os::elapsedTime();
1373 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1374 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1375 _remark_times.add((now - start) * 1000.0);
1377 g1p->record_concurrent_mark_remark_end();
1379 G1CMIsAliveClosure is_alive(g1h);
1380 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1381 }
1383 // Base class of the closures that finalize and verify the
1384 // liveness counting data.
1385 class CMCountDataClosureBase: public HeapRegionClosure {
1386 protected:
1387 G1CollectedHeap* _g1h;
1388 ConcurrentMark* _cm;
1389 CardTableModRefBS* _ct_bs;
1391 BitMap* _region_bm;
1392 BitMap* _card_bm;
1394 // Takes a region that's not empty (i.e., it has at least one
1395 // live object in it and sets its corresponding bit on the region
1396 // bitmap to 1. If the region is "starts humongous" it will also set
1397 // to 1 the bits on the region bitmap that correspond to its
1398 // associated "continues humongous" regions.
1399 void set_bit_for_region(HeapRegion* hr) {
1400 assert(!hr->continuesHumongous(), "should have filtered those out");
1402 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1403 if (!hr->startsHumongous()) {
1404 // Normal (non-humongous) case: just set the bit.
1405 _region_bm->par_at_put(index, true);
1406 } else {
1407 // Starts humongous case: calculate how many regions are part of
1408 // this humongous region and then set the bit range.
1409 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1410 _region_bm->par_at_put_range(index, end_index, true);
1411 }
1412 }
1414 public:
1415 CMCountDataClosureBase(G1CollectedHeap* g1h,
1416 BitMap* region_bm, BitMap* card_bm):
1417 _g1h(g1h), _cm(g1h->concurrent_mark()),
1418 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1419 _region_bm(region_bm), _card_bm(card_bm) { }
1420 };
1422 // Closure that calculates the # live objects per region. Used
1423 // for verification purposes during the cleanup pause.
1424 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1425 CMBitMapRO* _bm;
1426 size_t _region_marked_bytes;
1428 public:
1429 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1430 BitMap* region_bm, BitMap* card_bm) :
1431 CMCountDataClosureBase(g1h, region_bm, card_bm),
1432 _bm(bm), _region_marked_bytes(0) { }
1434 bool doHeapRegion(HeapRegion* hr) {
1436 if (hr->continuesHumongous()) {
1437 // We will ignore these here and process them when their
1438 // associated "starts humongous" region is processed (see
1439 // set_bit_for_heap_region()). Note that we cannot rely on their
1440 // associated "starts humongous" region to have their bit set to
1441 // 1 since, due to the region chunking in the parallel region
1442 // iteration, a "continues humongous" region might be visited
1443 // before its associated "starts humongous".
1444 return false;
1445 }
1447 HeapWord* ntams = hr->next_top_at_mark_start();
1448 HeapWord* start = hr->bottom();
1450 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1451 err_msg("Preconditions not met - "
1452 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1453 p2i(start), p2i(ntams), p2i(hr->end())));
1455 // Find the first marked object at or after "start".
1456 start = _bm->getNextMarkedWordAddress(start, ntams);
1458 size_t marked_bytes = 0;
1460 while (start < ntams) {
1461 oop obj = oop(start);
1462 int obj_sz = obj->size();
1463 HeapWord* obj_end = start + obj_sz;
1465 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1466 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1468 // Note: if we're looking at the last region in heap - obj_end
1469 // could be actually just beyond the end of the heap; end_idx
1470 // will then correspond to a (non-existent) card that is also
1471 // just beyond the heap.
1472 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1473 // end of object is not card aligned - increment to cover
1474 // all the cards spanned by the object
1475 end_idx += 1;
1476 }
1478 // Set the bits in the card BM for the cards spanned by this object.
1479 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1481 // Add the size of this object to the number of marked bytes.
1482 marked_bytes += (size_t)obj_sz * HeapWordSize;
1484 // Find the next marked object after this one.
1485 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1486 }
1488 // Mark the allocated-since-marking portion...
1489 HeapWord* top = hr->top();
1490 if (ntams < top) {
1491 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1492 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1494 // Note: if we're looking at the last region in heap - top
1495 // could be actually just beyond the end of the heap; end_idx
1496 // will then correspond to a (non-existent) card that is also
1497 // just beyond the heap.
1498 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1499 // end of object is not card aligned - increment to cover
1500 // all the cards spanned by the object
1501 end_idx += 1;
1502 }
1503 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1505 // This definitely means the region has live objects.
1506 set_bit_for_region(hr);
1507 }
1509 // Update the live region bitmap.
1510 if (marked_bytes > 0) {
1511 set_bit_for_region(hr);
1512 }
1514 // Set the marked bytes for the current region so that
1515 // it can be queried by a calling verificiation routine
1516 _region_marked_bytes = marked_bytes;
1518 return false;
1519 }
1521 size_t region_marked_bytes() const { return _region_marked_bytes; }
1522 };
1524 // Heap region closure used for verifying the counting data
1525 // that was accumulated concurrently and aggregated during
1526 // the remark pause. This closure is applied to the heap
1527 // regions during the STW cleanup pause.
1529 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1530 G1CollectedHeap* _g1h;
1531 ConcurrentMark* _cm;
1532 CalcLiveObjectsClosure _calc_cl;
1533 BitMap* _region_bm; // Region BM to be verified
1534 BitMap* _card_bm; // Card BM to be verified
1535 bool _verbose; // verbose output?
1537 BitMap* _exp_region_bm; // Expected Region BM values
1538 BitMap* _exp_card_bm; // Expected card BM values
1540 int _failures;
1542 public:
1543 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1544 BitMap* region_bm,
1545 BitMap* card_bm,
1546 BitMap* exp_region_bm,
1547 BitMap* exp_card_bm,
1548 bool verbose) :
1549 _g1h(g1h), _cm(g1h->concurrent_mark()),
1550 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1551 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1552 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1553 _failures(0) { }
1555 int failures() const { return _failures; }
1557 bool doHeapRegion(HeapRegion* hr) {
1558 if (hr->continuesHumongous()) {
1559 // We will ignore these here and process them when their
1560 // associated "starts humongous" region is processed (see
1561 // set_bit_for_heap_region()). Note that we cannot rely on their
1562 // associated "starts humongous" region to have their bit set to
1563 // 1 since, due to the region chunking in the parallel region
1564 // iteration, a "continues humongous" region might be visited
1565 // before its associated "starts humongous".
1566 return false;
1567 }
1569 int failures = 0;
1571 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1572 // this region and set the corresponding bits in the expected region
1573 // and card bitmaps.
1574 bool res = _calc_cl.doHeapRegion(hr);
1575 assert(res == false, "should be continuing");
1577 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1578 Mutex::_no_safepoint_check_flag);
1580 // Verify the marked bytes for this region.
1581 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1582 size_t act_marked_bytes = hr->next_marked_bytes();
1584 // We're not OK if expected marked bytes > actual marked bytes. It means
1585 // we have missed accounting some objects during the actual marking.
1586 if (exp_marked_bytes > act_marked_bytes) {
1587 if (_verbose) {
1588 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1589 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1590 hr->hrs_index(), exp_marked_bytes, act_marked_bytes);
1591 }
1592 failures += 1;
1593 }
1595 // Verify the bit, for this region, in the actual and expected
1596 // (which was just calculated) region bit maps.
1597 // We're not OK if the bit in the calculated expected region
1598 // bitmap is set and the bit in the actual region bitmap is not.
1599 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1601 bool expected = _exp_region_bm->at(index);
1602 bool actual = _region_bm->at(index);
1603 if (expected && !actual) {
1604 if (_verbose) {
1605 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1606 "expected: %s, actual: %s",
1607 hr->hrs_index(),
1608 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1609 }
1610 failures += 1;
1611 }
1613 // Verify that the card bit maps for the cards spanned by the current
1614 // region match. We have an error if we have a set bit in the expected
1615 // bit map and the corresponding bit in the actual bitmap is not set.
1617 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1618 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1620 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1621 expected = _exp_card_bm->at(i);
1622 actual = _card_bm->at(i);
1624 if (expected && !actual) {
1625 if (_verbose) {
1626 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1627 "expected: %s, actual: %s",
1628 hr->hrs_index(), i,
1629 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1630 }
1631 failures += 1;
1632 }
1633 }
1635 if (failures > 0 && _verbose) {
1636 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1637 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1638 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1639 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1640 }
1642 _failures += failures;
1644 // We could stop iteration over the heap when we
1645 // find the first violating region by returning true.
1646 return false;
1647 }
1648 };
1650 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1651 protected:
1652 G1CollectedHeap* _g1h;
1653 ConcurrentMark* _cm;
1654 BitMap* _actual_region_bm;
1655 BitMap* _actual_card_bm;
1657 uint _n_workers;
1659 BitMap* _expected_region_bm;
1660 BitMap* _expected_card_bm;
1662 int _failures;
1663 bool _verbose;
1665 public:
1666 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1667 BitMap* region_bm, BitMap* card_bm,
1668 BitMap* expected_region_bm, BitMap* expected_card_bm)
1669 : AbstractGangTask("G1 verify final counting"),
1670 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1671 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1672 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1673 _failures(0), _verbose(false),
1674 _n_workers(0) {
1675 assert(VerifyDuringGC, "don't call this otherwise");
1677 // Use the value already set as the number of active threads
1678 // in the call to run_task().
1679 if (G1CollectedHeap::use_parallel_gc_threads()) {
1680 assert( _g1h->workers()->active_workers() > 0,
1681 "Should have been previously set");
1682 _n_workers = _g1h->workers()->active_workers();
1683 } else {
1684 _n_workers = 1;
1685 }
1687 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1688 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1690 _verbose = _cm->verbose_medium();
1691 }
1693 void work(uint worker_id) {
1694 assert(worker_id < _n_workers, "invariant");
1696 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1697 _actual_region_bm, _actual_card_bm,
1698 _expected_region_bm,
1699 _expected_card_bm,
1700 _verbose);
1702 if (G1CollectedHeap::use_parallel_gc_threads()) {
1703 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1704 worker_id,
1705 _n_workers,
1706 HeapRegion::VerifyCountClaimValue);
1707 } else {
1708 _g1h->heap_region_iterate(&verify_cl);
1709 }
1711 Atomic::add(verify_cl.failures(), &_failures);
1712 }
1714 int failures() const { return _failures; }
1715 };
1717 // Closure that finalizes the liveness counting data.
1718 // Used during the cleanup pause.
1719 // Sets the bits corresponding to the interval [NTAMS, top]
1720 // (which contains the implicitly live objects) in the
1721 // card liveness bitmap. Also sets the bit for each region,
1722 // containing live data, in the region liveness bitmap.
1724 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1725 public:
1726 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1727 BitMap* region_bm,
1728 BitMap* card_bm) :
1729 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1731 bool doHeapRegion(HeapRegion* hr) {
1733 if (hr->continuesHumongous()) {
1734 // We will ignore these here and process them when their
1735 // associated "starts humongous" region is processed (see
1736 // set_bit_for_heap_region()). Note that we cannot rely on their
1737 // associated "starts humongous" region to have their bit set to
1738 // 1 since, due to the region chunking in the parallel region
1739 // iteration, a "continues humongous" region might be visited
1740 // before its associated "starts humongous".
1741 return false;
1742 }
1744 HeapWord* ntams = hr->next_top_at_mark_start();
1745 HeapWord* top = hr->top();
1747 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1749 // Mark the allocated-since-marking portion...
1750 if (ntams < top) {
1751 // This definitely means the region has live objects.
1752 set_bit_for_region(hr);
1754 // Now set the bits in the card bitmap for [ntams, top)
1755 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1756 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1758 // Note: if we're looking at the last region in heap - top
1759 // could be actually just beyond the end of the heap; end_idx
1760 // will then correspond to a (non-existent) card that is also
1761 // just beyond the heap.
1762 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1763 // end of object is not card aligned - increment to cover
1764 // all the cards spanned by the object
1765 end_idx += 1;
1766 }
1768 assert(end_idx <= _card_bm->size(),
1769 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1770 end_idx, _card_bm->size()));
1771 assert(start_idx < _card_bm->size(),
1772 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1773 start_idx, _card_bm->size()));
1775 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1776 }
1778 // Set the bit for the region if it contains live data
1779 if (hr->next_marked_bytes() > 0) {
1780 set_bit_for_region(hr);
1781 }
1783 return false;
1784 }
1785 };
1787 class G1ParFinalCountTask: public AbstractGangTask {
1788 protected:
1789 G1CollectedHeap* _g1h;
1790 ConcurrentMark* _cm;
1791 BitMap* _actual_region_bm;
1792 BitMap* _actual_card_bm;
1794 uint _n_workers;
1796 public:
1797 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1798 : AbstractGangTask("G1 final counting"),
1799 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1800 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1801 _n_workers(0) {
1802 // Use the value already set as the number of active threads
1803 // in the call to run_task().
1804 if (G1CollectedHeap::use_parallel_gc_threads()) {
1805 assert( _g1h->workers()->active_workers() > 0,
1806 "Should have been previously set");
1807 _n_workers = _g1h->workers()->active_workers();
1808 } else {
1809 _n_workers = 1;
1810 }
1811 }
1813 void work(uint worker_id) {
1814 assert(worker_id < _n_workers, "invariant");
1816 FinalCountDataUpdateClosure final_update_cl(_g1h,
1817 _actual_region_bm,
1818 _actual_card_bm);
1820 if (G1CollectedHeap::use_parallel_gc_threads()) {
1821 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1822 worker_id,
1823 _n_workers,
1824 HeapRegion::FinalCountClaimValue);
1825 } else {
1826 _g1h->heap_region_iterate(&final_update_cl);
1827 }
1828 }
1829 };
1831 class G1ParNoteEndTask;
1833 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1834 G1CollectedHeap* _g1;
1835 size_t _max_live_bytes;
1836 uint _regions_claimed;
1837 size_t _freed_bytes;
1838 FreeRegionList* _local_cleanup_list;
1839 HeapRegionSetCount _old_regions_removed;
1840 HeapRegionSetCount _humongous_regions_removed;
1841 HRRSCleanupTask* _hrrs_cleanup_task;
1842 double _claimed_region_time;
1843 double _max_region_time;
1845 public:
1846 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1847 FreeRegionList* local_cleanup_list,
1848 HRRSCleanupTask* hrrs_cleanup_task) :
1849 _g1(g1),
1850 _max_live_bytes(0), _regions_claimed(0),
1851 _freed_bytes(0),
1852 _claimed_region_time(0.0), _max_region_time(0.0),
1853 _local_cleanup_list(local_cleanup_list),
1854 _old_regions_removed(),
1855 _humongous_regions_removed(),
1856 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1858 size_t freed_bytes() { return _freed_bytes; }
1859 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1860 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1862 bool doHeapRegion(HeapRegion *hr) {
1863 if (hr->continuesHumongous()) {
1864 return false;
1865 }
1866 // We use a claim value of zero here because all regions
1867 // were claimed with value 1 in the FinalCount task.
1868 _g1->reset_gc_time_stamps(hr);
1869 double start = os::elapsedTime();
1870 _regions_claimed++;
1871 hr->note_end_of_marking();
1872 _max_live_bytes += hr->max_live_bytes();
1874 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1875 _freed_bytes += hr->used();
1876 hr->set_containing_set(NULL);
1877 if (hr->isHumongous()) {
1878 assert(hr->startsHumongous(), "we should only see starts humongous");
1879 _humongous_regions_removed.increment(1u, hr->capacity());
1880 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1881 } else {
1882 _old_regions_removed.increment(1u, hr->capacity());
1883 _g1->free_region(hr, _local_cleanup_list, true);
1884 }
1885 } else {
1886 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1887 }
1889 double region_time = (os::elapsedTime() - start);
1890 _claimed_region_time += region_time;
1891 if (region_time > _max_region_time) {
1892 _max_region_time = region_time;
1893 }
1894 return false;
1895 }
1897 size_t max_live_bytes() { return _max_live_bytes; }
1898 uint regions_claimed() { return _regions_claimed; }
1899 double claimed_region_time_sec() { return _claimed_region_time; }
1900 double max_region_time_sec() { return _max_region_time; }
1901 };
1903 class G1ParNoteEndTask: public AbstractGangTask {
1904 friend class G1NoteEndOfConcMarkClosure;
1906 protected:
1907 G1CollectedHeap* _g1h;
1908 size_t _max_live_bytes;
1909 size_t _freed_bytes;
1910 FreeRegionList* _cleanup_list;
1912 public:
1913 G1ParNoteEndTask(G1CollectedHeap* g1h,
1914 FreeRegionList* cleanup_list) :
1915 AbstractGangTask("G1 note end"), _g1h(g1h),
1916 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1918 void work(uint worker_id) {
1919 double start = os::elapsedTime();
1920 FreeRegionList local_cleanup_list("Local Cleanup List");
1921 HRRSCleanupTask hrrs_cleanup_task;
1922 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1923 &hrrs_cleanup_task);
1924 if (G1CollectedHeap::use_parallel_gc_threads()) {
1925 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1926 _g1h->workers()->active_workers(),
1927 HeapRegion::NoteEndClaimValue);
1928 } else {
1929 _g1h->heap_region_iterate(&g1_note_end);
1930 }
1931 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1933 // Now update the lists
1934 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1935 {
1936 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1937 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1938 _max_live_bytes += g1_note_end.max_live_bytes();
1939 _freed_bytes += g1_note_end.freed_bytes();
1941 // If we iterate over the global cleanup list at the end of
1942 // cleanup to do this printing we will not guarantee to only
1943 // generate output for the newly-reclaimed regions (the list
1944 // might not be empty at the beginning of cleanup; we might
1945 // still be working on its previous contents). So we do the
1946 // printing here, before we append the new regions to the global
1947 // cleanup list.
1949 G1HRPrinter* hr_printer = _g1h->hr_printer();
1950 if (hr_printer->is_active()) {
1951 FreeRegionListIterator iter(&local_cleanup_list);
1952 while (iter.more_available()) {
1953 HeapRegion* hr = iter.get_next();
1954 hr_printer->cleanup(hr);
1955 }
1956 }
1958 _cleanup_list->add_ordered(&local_cleanup_list);
1959 assert(local_cleanup_list.is_empty(), "post-condition");
1961 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1962 }
1963 }
1964 size_t max_live_bytes() { return _max_live_bytes; }
1965 size_t freed_bytes() { return _freed_bytes; }
1966 };
1968 class G1ParScrubRemSetTask: public AbstractGangTask {
1969 protected:
1970 G1RemSet* _g1rs;
1971 BitMap* _region_bm;
1972 BitMap* _card_bm;
1973 public:
1974 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1975 BitMap* region_bm, BitMap* card_bm) :
1976 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1977 _region_bm(region_bm), _card_bm(card_bm) { }
1979 void work(uint worker_id) {
1980 if (G1CollectedHeap::use_parallel_gc_threads()) {
1981 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1982 HeapRegion::ScrubRemSetClaimValue);
1983 } else {
1984 _g1rs->scrub(_region_bm, _card_bm);
1985 }
1986 }
1988 };
1990 void ConcurrentMark::cleanup() {
1991 // world is stopped at this checkpoint
1992 assert(SafepointSynchronize::is_at_safepoint(),
1993 "world should be stopped");
1994 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1996 // If a full collection has happened, we shouldn't do this.
1997 if (has_aborted()) {
1998 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1999 return;
2000 }
2002 g1h->verify_region_sets_optional();
2004 if (VerifyDuringGC) {
2005 HandleMark hm; // handle scope
2006 Universe::heap()->prepare_for_verify();
2007 Universe::verify(VerifyOption_G1UsePrevMarking,
2008 " VerifyDuringGC:(before)");
2009 }
2011 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
2012 g1p->record_concurrent_mark_cleanup_start();
2014 double start = os::elapsedTime();
2016 HeapRegionRemSet::reset_for_cleanup_tasks();
2018 uint n_workers;
2020 // Do counting once more with the world stopped for good measure.
2021 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2023 if (G1CollectedHeap::use_parallel_gc_threads()) {
2024 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2025 "sanity check");
2027 g1h->set_par_threads();
2028 n_workers = g1h->n_par_threads();
2029 assert(g1h->n_par_threads() == n_workers,
2030 "Should not have been reset");
2031 g1h->workers()->run_task(&g1_par_count_task);
2032 // Done with the parallel phase so reset to 0.
2033 g1h->set_par_threads(0);
2035 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2036 "sanity check");
2037 } else {
2038 n_workers = 1;
2039 g1_par_count_task.work(0);
2040 }
2042 if (VerifyDuringGC) {
2043 // Verify that the counting data accumulated during marking matches
2044 // that calculated by walking the marking bitmap.
2046 // Bitmaps to hold expected values
2047 BitMap expected_region_bm(_region_bm.size(), true);
2048 BitMap expected_card_bm(_card_bm.size(), true);
2050 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2051 &_region_bm,
2052 &_card_bm,
2053 &expected_region_bm,
2054 &expected_card_bm);
2056 if (G1CollectedHeap::use_parallel_gc_threads()) {
2057 g1h->set_par_threads((int)n_workers);
2058 g1h->workers()->run_task(&g1_par_verify_task);
2059 // Done with the parallel phase so reset to 0.
2060 g1h->set_par_threads(0);
2062 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2063 "sanity check");
2064 } else {
2065 g1_par_verify_task.work(0);
2066 }
2068 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2069 }
2071 size_t start_used_bytes = g1h->used();
2072 g1h->set_marking_complete();
2074 double count_end = os::elapsedTime();
2075 double this_final_counting_time = (count_end - start);
2076 _total_counting_time += this_final_counting_time;
2078 if (G1PrintRegionLivenessInfo) {
2079 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2080 _g1h->heap_region_iterate(&cl);
2081 }
2083 // Install newly created mark bitMap as "prev".
2084 swapMarkBitMaps();
2086 g1h->reset_gc_time_stamp();
2088 // Note end of marking in all heap regions.
2089 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2090 if (G1CollectedHeap::use_parallel_gc_threads()) {
2091 g1h->set_par_threads((int)n_workers);
2092 g1h->workers()->run_task(&g1_par_note_end_task);
2093 g1h->set_par_threads(0);
2095 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2096 "sanity check");
2097 } else {
2098 g1_par_note_end_task.work(0);
2099 }
2100 g1h->check_gc_time_stamps();
2102 if (!cleanup_list_is_empty()) {
2103 // The cleanup list is not empty, so we'll have to process it
2104 // concurrently. Notify anyone else that might be wanting free
2105 // regions that there will be more free regions coming soon.
2106 g1h->set_free_regions_coming();
2107 }
2109 // call below, since it affects the metric by which we sort the heap
2110 // regions.
2111 if (G1ScrubRemSets) {
2112 double rs_scrub_start = os::elapsedTime();
2113 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2114 if (G1CollectedHeap::use_parallel_gc_threads()) {
2115 g1h->set_par_threads((int)n_workers);
2116 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2117 g1h->set_par_threads(0);
2119 assert(g1h->check_heap_region_claim_values(
2120 HeapRegion::ScrubRemSetClaimValue),
2121 "sanity check");
2122 } else {
2123 g1_par_scrub_rs_task.work(0);
2124 }
2126 double rs_scrub_end = os::elapsedTime();
2127 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2128 _total_rs_scrub_time += this_rs_scrub_time;
2129 }
2131 // this will also free any regions totally full of garbage objects,
2132 // and sort the regions.
2133 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2135 // Statistics.
2136 double end = os::elapsedTime();
2137 _cleanup_times.add((end - start) * 1000.0);
2139 if (G1Log::fine()) {
2140 g1h->print_size_transition(gclog_or_tty,
2141 start_used_bytes,
2142 g1h->used(),
2143 g1h->capacity());
2144 }
2146 // Clean up will have freed any regions completely full of garbage.
2147 // Update the soft reference policy with the new heap occupancy.
2148 Universe::update_heap_info_at_gc();
2150 if (VerifyDuringGC) {
2151 HandleMark hm; // handle scope
2152 Universe::heap()->prepare_for_verify();
2153 Universe::verify(VerifyOption_G1UsePrevMarking,
2154 " VerifyDuringGC:(after)");
2155 }
2157 g1h->verify_region_sets_optional();
2159 // We need to make this be a "collection" so any collection pause that
2160 // races with it goes around and waits for completeCleanup to finish.
2161 g1h->increment_total_collections();
2163 // Clean out dead classes and update Metaspace sizes.
2164 if (ClassUnloadingWithConcurrentMark) {
2165 ClassLoaderDataGraph::purge();
2166 }
2167 MetaspaceGC::compute_new_size();
2169 // We reclaimed old regions so we should calculate the sizes to make
2170 // sure we update the old gen/space data.
2171 g1h->g1mm()->update_sizes();
2173 g1h->trace_heap_after_concurrent_cycle();
2174 }
2176 void ConcurrentMark::completeCleanup() {
2177 if (has_aborted()) return;
2179 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2181 _cleanup_list.verify_optional();
2182 FreeRegionList tmp_free_list("Tmp Free List");
2184 if (G1ConcRegionFreeingVerbose) {
2185 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2186 "cleanup list has %u entries",
2187 _cleanup_list.length());
2188 }
2190 // Noone else should be accessing the _cleanup_list at this point,
2191 // so it's not necessary to take any locks
2192 while (!_cleanup_list.is_empty()) {
2193 HeapRegion* hr = _cleanup_list.remove_head();
2194 assert(hr != NULL, "Got NULL from a non-empty list");
2195 hr->par_clear();
2196 tmp_free_list.add_ordered(hr);
2198 // Instead of adding one region at a time to the secondary_free_list,
2199 // we accumulate them in the local list and move them a few at a
2200 // time. This also cuts down on the number of notify_all() calls
2201 // we do during this process. We'll also append the local list when
2202 // _cleanup_list is empty (which means we just removed the last
2203 // region from the _cleanup_list).
2204 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2205 _cleanup_list.is_empty()) {
2206 if (G1ConcRegionFreeingVerbose) {
2207 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2208 "appending %u entries to the secondary_free_list, "
2209 "cleanup list still has %u entries",
2210 tmp_free_list.length(),
2211 _cleanup_list.length());
2212 }
2214 {
2215 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2216 g1h->secondary_free_list_add(&tmp_free_list);
2217 SecondaryFreeList_lock->notify_all();
2218 }
2220 if (G1StressConcRegionFreeing) {
2221 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2222 os::sleep(Thread::current(), (jlong) 1, false);
2223 }
2224 }
2225 }
2226 }
2227 assert(tmp_free_list.is_empty(), "post-condition");
2228 }
2230 // Supporting Object and Oop closures for reference discovery
2231 // and processing in during marking
2233 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2234 HeapWord* addr = (HeapWord*)obj;
2235 return addr != NULL &&
2236 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2237 }
2239 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2240 // Uses the CMTask associated with a worker thread (for serial reference
2241 // processing the CMTask for worker 0 is used) to preserve (mark) and
2242 // trace referent objects.
2243 //
2244 // Using the CMTask and embedded local queues avoids having the worker
2245 // threads operating on the global mark stack. This reduces the risk
2246 // of overflowing the stack - which we would rather avoid at this late
2247 // state. Also using the tasks' local queues removes the potential
2248 // of the workers interfering with each other that could occur if
2249 // operating on the global stack.
2251 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2252 ConcurrentMark* _cm;
2253 CMTask* _task;
2254 int _ref_counter_limit;
2255 int _ref_counter;
2256 bool _is_serial;
2257 public:
2258 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2259 _cm(cm), _task(task), _is_serial(is_serial),
2260 _ref_counter_limit(G1RefProcDrainInterval) {
2261 assert(_ref_counter_limit > 0, "sanity");
2262 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2263 _ref_counter = _ref_counter_limit;
2264 }
2266 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2267 virtual void do_oop( oop* p) { do_oop_work(p); }
2269 template <class T> void do_oop_work(T* p) {
2270 if (!_cm->has_overflown()) {
2271 oop obj = oopDesc::load_decode_heap_oop(p);
2272 if (_cm->verbose_high()) {
2273 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2274 "*"PTR_FORMAT" = "PTR_FORMAT,
2275 _task->worker_id(), p2i(p), p2i((void*) obj));
2276 }
2278 _task->deal_with_reference(obj);
2279 _ref_counter--;
2281 if (_ref_counter == 0) {
2282 // We have dealt with _ref_counter_limit references, pushing them
2283 // and objects reachable from them on to the local stack (and
2284 // possibly the global stack). Call CMTask::do_marking_step() to
2285 // process these entries.
2286 //
2287 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2288 // there's nothing more to do (i.e. we're done with the entries that
2289 // were pushed as a result of the CMTask::deal_with_reference() calls
2290 // above) or we overflow.
2291 //
2292 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2293 // flag while there may still be some work to do. (See the comment at
2294 // the beginning of CMTask::do_marking_step() for those conditions -
2295 // one of which is reaching the specified time target.) It is only
2296 // when CMTask::do_marking_step() returns without setting the
2297 // has_aborted() flag that the marking step has completed.
2298 do {
2299 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2300 _task->do_marking_step(mark_step_duration_ms,
2301 false /* do_termination */,
2302 _is_serial);
2303 } while (_task->has_aborted() && !_cm->has_overflown());
2304 _ref_counter = _ref_counter_limit;
2305 }
2306 } else {
2307 if (_cm->verbose_high()) {
2308 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2309 }
2310 }
2311 }
2312 };
2314 // 'Drain' oop closure used by both serial and parallel reference processing.
2315 // Uses the CMTask associated with a given worker thread (for serial
2316 // reference processing the CMtask for worker 0 is used). Calls the
2317 // do_marking_step routine, with an unbelievably large timeout value,
2318 // to drain the marking data structures of the remaining entries
2319 // added by the 'keep alive' oop closure above.
2321 class G1CMDrainMarkingStackClosure: public VoidClosure {
2322 ConcurrentMark* _cm;
2323 CMTask* _task;
2324 bool _is_serial;
2325 public:
2326 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2327 _cm(cm), _task(task), _is_serial(is_serial) {
2328 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2329 }
2331 void do_void() {
2332 do {
2333 if (_cm->verbose_high()) {
2334 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2335 _task->worker_id(), BOOL_TO_STR(_is_serial));
2336 }
2338 // We call CMTask::do_marking_step() to completely drain the local
2339 // and global marking stacks of entries pushed by the 'keep alive'
2340 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2341 //
2342 // CMTask::do_marking_step() is called in a loop, which we'll exit
2343 // if there's nothing more to do (i.e. we'completely drained the
2344 // entries that were pushed as a a result of applying the 'keep alive'
2345 // closure to the entries on the discovered ref lists) or we overflow
2346 // the global marking stack.
2347 //
2348 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2349 // flag while there may still be some work to do. (See the comment at
2350 // the beginning of CMTask::do_marking_step() for those conditions -
2351 // one of which is reaching the specified time target.) It is only
2352 // when CMTask::do_marking_step() returns without setting the
2353 // has_aborted() flag that the marking step has completed.
2355 _task->do_marking_step(1000000000.0 /* something very large */,
2356 true /* do_termination */,
2357 _is_serial);
2358 } while (_task->has_aborted() && !_cm->has_overflown());
2359 }
2360 };
2362 // Implementation of AbstractRefProcTaskExecutor for parallel
2363 // reference processing at the end of G1 concurrent marking
2365 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2366 private:
2367 G1CollectedHeap* _g1h;
2368 ConcurrentMark* _cm;
2369 WorkGang* _workers;
2370 int _active_workers;
2372 public:
2373 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2374 ConcurrentMark* cm,
2375 WorkGang* workers,
2376 int n_workers) :
2377 _g1h(g1h), _cm(cm),
2378 _workers(workers), _active_workers(n_workers) { }
2380 // Executes the given task using concurrent marking worker threads.
2381 virtual void execute(ProcessTask& task);
2382 virtual void execute(EnqueueTask& task);
2383 };
2385 class G1CMRefProcTaskProxy: public AbstractGangTask {
2386 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2387 ProcessTask& _proc_task;
2388 G1CollectedHeap* _g1h;
2389 ConcurrentMark* _cm;
2391 public:
2392 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2393 G1CollectedHeap* g1h,
2394 ConcurrentMark* cm) :
2395 AbstractGangTask("Process reference objects in parallel"),
2396 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2397 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2398 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2399 }
2401 virtual void work(uint worker_id) {
2402 ResourceMark rm;
2403 HandleMark hm;
2404 CMTask* task = _cm->task(worker_id);
2405 G1CMIsAliveClosure g1_is_alive(_g1h);
2406 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2407 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2409 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2410 }
2411 };
2413 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2414 assert(_workers != NULL, "Need parallel worker threads.");
2415 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2417 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2419 // We need to reset the concurrency level before each
2420 // proxy task execution, so that the termination protocol
2421 // and overflow handling in CMTask::do_marking_step() knows
2422 // how many workers to wait for.
2423 _cm->set_concurrency(_active_workers);
2424 _g1h->set_par_threads(_active_workers);
2425 _workers->run_task(&proc_task_proxy);
2426 _g1h->set_par_threads(0);
2427 }
2429 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2430 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2431 EnqueueTask& _enq_task;
2433 public:
2434 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2435 AbstractGangTask("Enqueue reference objects in parallel"),
2436 _enq_task(enq_task) { }
2438 virtual void work(uint worker_id) {
2439 _enq_task.work(worker_id);
2440 }
2441 };
2443 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2444 assert(_workers != NULL, "Need parallel worker threads.");
2445 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2447 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2449 // Not strictly necessary but...
2450 //
2451 // We need to reset the concurrency level before each
2452 // proxy task execution, so that the termination protocol
2453 // and overflow handling in CMTask::do_marking_step() knows
2454 // how many workers to wait for.
2455 _cm->set_concurrency(_active_workers);
2456 _g1h->set_par_threads(_active_workers);
2457 _workers->run_task(&enq_task_proxy);
2458 _g1h->set_par_threads(0);
2459 }
2461 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2462 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2463 }
2465 // Helper class to get rid of some boilerplate code.
2466 class G1RemarkGCTraceTime : public GCTraceTime {
2467 static bool doit_and_prepend(bool doit) {
2468 if (doit) {
2469 gclog_or_tty->put(' ');
2470 }
2471 return doit;
2472 }
2474 public:
2475 G1RemarkGCTraceTime(const char* title, bool doit)
2476 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
2477 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
2478 }
2479 };
2481 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2482 if (has_overflown()) {
2483 // Skip processing the discovered references if we have
2484 // overflown the global marking stack. Reference objects
2485 // only get discovered once so it is OK to not
2486 // de-populate the discovered reference lists. We could have,
2487 // but the only benefit would be that, when marking restarts,
2488 // less reference objects are discovered.
2489 return;
2490 }
2492 ResourceMark rm;
2493 HandleMark hm;
2495 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2497 // Is alive closure.
2498 G1CMIsAliveClosure g1_is_alive(g1h);
2500 // Inner scope to exclude the cleaning of the string and symbol
2501 // tables from the displayed time.
2502 {
2503 if (G1Log::finer()) {
2504 gclog_or_tty->put(' ');
2505 }
2506 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm(), concurrent_gc_id());
2508 ReferenceProcessor* rp = g1h->ref_processor_cm();
2510 // See the comment in G1CollectedHeap::ref_processing_init()
2511 // about how reference processing currently works in G1.
2513 // Set the soft reference policy
2514 rp->setup_policy(clear_all_soft_refs);
2515 assert(_markStack.isEmpty(), "mark stack should be empty");
2517 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2518 // in serial reference processing. Note these closures are also
2519 // used for serially processing (by the the current thread) the
2520 // JNI references during parallel reference processing.
2521 //
2522 // These closures do not need to synchronize with the worker
2523 // threads involved in parallel reference processing as these
2524 // instances are executed serially by the current thread (e.g.
2525 // reference processing is not multi-threaded and is thus
2526 // performed by the current thread instead of a gang worker).
2527 //
2528 // The gang tasks involved in parallel reference procssing create
2529 // their own instances of these closures, which do their own
2530 // synchronization among themselves.
2531 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2532 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2534 // We need at least one active thread. If reference processing
2535 // is not multi-threaded we use the current (VMThread) thread,
2536 // otherwise we use the work gang from the G1CollectedHeap and
2537 // we utilize all the worker threads we can.
2538 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2539 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2540 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2542 // Parallel processing task executor.
2543 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2544 g1h->workers(), active_workers);
2545 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2547 // Set the concurrency level. The phase was already set prior to
2548 // executing the remark task.
2549 set_concurrency(active_workers);
2551 // Set the degree of MT processing here. If the discovery was done MT,
2552 // the number of threads involved during discovery could differ from
2553 // the number of active workers. This is OK as long as the discovered
2554 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2555 rp->set_active_mt_degree(active_workers);
2557 // Process the weak references.
2558 const ReferenceProcessorStats& stats =
2559 rp->process_discovered_references(&g1_is_alive,
2560 &g1_keep_alive,
2561 &g1_drain_mark_stack,
2562 executor,
2563 g1h->gc_timer_cm(),
2564 concurrent_gc_id());
2565 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2567 // The do_oop work routines of the keep_alive and drain_marking_stack
2568 // oop closures will set the has_overflown flag if we overflow the
2569 // global marking stack.
2571 assert(_markStack.overflow() || _markStack.isEmpty(),
2572 "mark stack should be empty (unless it overflowed)");
2574 if (_markStack.overflow()) {
2575 // This should have been done already when we tried to push an
2576 // entry on to the global mark stack. But let's do it again.
2577 set_has_overflown();
2578 }
2580 assert(rp->num_q() == active_workers, "why not");
2582 rp->enqueue_discovered_references(executor);
2584 rp->verify_no_references_recorded();
2585 assert(!rp->discovery_enabled(), "Post condition");
2586 }
2588 if (has_overflown()) {
2589 // We can not trust g1_is_alive if the marking stack overflowed
2590 return;
2591 }
2593 assert(_markStack.isEmpty(), "Marking should have completed");
2595 // Unload Klasses, String, Symbols, Code Cache, etc.
2596 {
2597 G1RemarkGCTraceTime trace("Unloading", G1Log::finer());
2599 if (ClassUnloadingWithConcurrentMark) {
2600 bool purged_classes;
2602 {
2603 G1RemarkGCTraceTime trace("System Dictionary Unloading", G1Log::finest());
2604 purged_classes = SystemDictionary::do_unloading(&g1_is_alive);
2605 }
2607 {
2608 G1RemarkGCTraceTime trace("Parallel Unloading", G1Log::finest());
2609 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2610 }
2611 }
2613 if (G1StringDedup::is_enabled()) {
2614 G1RemarkGCTraceTime trace("String Deduplication Unlink", G1Log::finest());
2615 G1StringDedup::unlink(&g1_is_alive);
2616 }
2617 }
2618 }
2620 void ConcurrentMark::swapMarkBitMaps() {
2621 CMBitMapRO* temp = _prevMarkBitMap;
2622 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2623 _nextMarkBitMap = (CMBitMap*) temp;
2624 }
2626 class CMObjectClosure;
2628 // Closure for iterating over objects, currently only used for
2629 // processing SATB buffers.
2630 class CMObjectClosure : public ObjectClosure {
2631 private:
2632 CMTask* _task;
2634 public:
2635 void do_object(oop obj) {
2636 _task->deal_with_reference(obj);
2637 }
2639 CMObjectClosure(CMTask* task) : _task(task) { }
2640 };
2642 class G1RemarkThreadsClosure : public ThreadClosure {
2643 CMObjectClosure _cm_obj;
2644 G1CMOopClosure _cm_cl;
2645 MarkingCodeBlobClosure _code_cl;
2646 int _thread_parity;
2647 bool _is_par;
2649 public:
2650 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) :
2651 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2652 _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {}
2654 void do_thread(Thread* thread) {
2655 if (thread->is_Java_thread()) {
2656 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2657 JavaThread* jt = (JavaThread*)thread;
2659 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2660 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2661 // * Alive if on the stack of an executing method
2662 // * Weakly reachable otherwise
2663 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2664 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2665 jt->nmethods_do(&_code_cl);
2667 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2668 }
2669 } else if (thread->is_VM_thread()) {
2670 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2671 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2672 }
2673 }
2674 }
2675 };
2677 class CMRemarkTask: public AbstractGangTask {
2678 private:
2679 ConcurrentMark* _cm;
2680 bool _is_serial;
2681 public:
2682 void work(uint worker_id) {
2683 // Since all available tasks are actually started, we should
2684 // only proceed if we're supposed to be actived.
2685 if (worker_id < _cm->active_tasks()) {
2686 CMTask* task = _cm->task(worker_id);
2687 task->record_start_time();
2688 {
2689 ResourceMark rm;
2690 HandleMark hm;
2692 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial);
2693 Threads::threads_do(&threads_f);
2694 }
2696 do {
2697 task->do_marking_step(1000000000.0 /* something very large */,
2698 true /* do_termination */,
2699 _is_serial);
2700 } while (task->has_aborted() && !_cm->has_overflown());
2701 // If we overflow, then we do not want to restart. We instead
2702 // want to abort remark and do concurrent marking again.
2703 task->record_end_time();
2704 }
2705 }
2707 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2708 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2709 _cm->terminator()->reset_for_reuse(active_workers);
2710 }
2711 };
2713 void ConcurrentMark::checkpointRootsFinalWork() {
2714 ResourceMark rm;
2715 HandleMark hm;
2716 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2718 G1RemarkGCTraceTime trace("Finalize Marking", G1Log::finer());
2720 g1h->ensure_parsability(false);
2722 if (G1CollectedHeap::use_parallel_gc_threads()) {
2723 G1CollectedHeap::StrongRootsScope srs(g1h);
2724 // this is remark, so we'll use up all active threads
2725 uint active_workers = g1h->workers()->active_workers();
2726 if (active_workers == 0) {
2727 assert(active_workers > 0, "Should have been set earlier");
2728 active_workers = (uint) ParallelGCThreads;
2729 g1h->workers()->set_active_workers(active_workers);
2730 }
2731 set_concurrency_and_phase(active_workers, false /* concurrent */);
2732 // Leave _parallel_marking_threads at it's
2733 // value originally calculated in the ConcurrentMark
2734 // constructor and pass values of the active workers
2735 // through the gang in the task.
2737 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2738 // We will start all available threads, even if we decide that the
2739 // active_workers will be fewer. The extra ones will just bail out
2740 // immediately.
2741 g1h->set_par_threads(active_workers);
2742 g1h->workers()->run_task(&remarkTask);
2743 g1h->set_par_threads(0);
2744 } else {
2745 G1CollectedHeap::StrongRootsScope srs(g1h);
2746 uint active_workers = 1;
2747 set_concurrency_and_phase(active_workers, false /* concurrent */);
2749 // Note - if there's no work gang then the VMThread will be
2750 // the thread to execute the remark - serially. We have
2751 // to pass true for the is_serial parameter so that
2752 // CMTask::do_marking_step() doesn't enter the sync
2753 // barriers in the event of an overflow. Doing so will
2754 // cause an assert that the current thread is not a
2755 // concurrent GC thread.
2756 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2757 remarkTask.work(0);
2758 }
2759 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2760 guarantee(has_overflown() ||
2761 satb_mq_set.completed_buffers_num() == 0,
2762 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2763 BOOL_TO_STR(has_overflown()),
2764 satb_mq_set.completed_buffers_num()));
2766 print_stats();
2767 }
2769 #ifndef PRODUCT
2771 class PrintReachableOopClosure: public OopClosure {
2772 private:
2773 G1CollectedHeap* _g1h;
2774 outputStream* _out;
2775 VerifyOption _vo;
2776 bool _all;
2778 public:
2779 PrintReachableOopClosure(outputStream* out,
2780 VerifyOption vo,
2781 bool all) :
2782 _g1h(G1CollectedHeap::heap()),
2783 _out(out), _vo(vo), _all(all) { }
2785 void do_oop(narrowOop* p) { do_oop_work(p); }
2786 void do_oop( oop* p) { do_oop_work(p); }
2788 template <class T> void do_oop_work(T* p) {
2789 oop obj = oopDesc::load_decode_heap_oop(p);
2790 const char* str = NULL;
2791 const char* str2 = "";
2793 if (obj == NULL) {
2794 str = "";
2795 } else if (!_g1h->is_in_g1_reserved(obj)) {
2796 str = " O";
2797 } else {
2798 HeapRegion* hr = _g1h->heap_region_containing(obj);
2799 guarantee(hr != NULL, "invariant");
2800 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2801 bool marked = _g1h->is_marked(obj, _vo);
2803 if (over_tams) {
2804 str = " >";
2805 if (marked) {
2806 str2 = " AND MARKED";
2807 }
2808 } else if (marked) {
2809 str = " M";
2810 } else {
2811 str = " NOT";
2812 }
2813 }
2815 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2816 p2i(p), p2i((void*) obj), str, str2);
2817 }
2818 };
2820 class PrintReachableObjectClosure : public ObjectClosure {
2821 private:
2822 G1CollectedHeap* _g1h;
2823 outputStream* _out;
2824 VerifyOption _vo;
2825 bool _all;
2826 HeapRegion* _hr;
2828 public:
2829 PrintReachableObjectClosure(outputStream* out,
2830 VerifyOption vo,
2831 bool all,
2832 HeapRegion* hr) :
2833 _g1h(G1CollectedHeap::heap()),
2834 _out(out), _vo(vo), _all(all), _hr(hr) { }
2836 void do_object(oop o) {
2837 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2838 bool marked = _g1h->is_marked(o, _vo);
2839 bool print_it = _all || over_tams || marked;
2841 if (print_it) {
2842 _out->print_cr(" "PTR_FORMAT"%s",
2843 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2844 PrintReachableOopClosure oopCl(_out, _vo, _all);
2845 o->oop_iterate_no_header(&oopCl);
2846 }
2847 }
2848 };
2850 class PrintReachableRegionClosure : public HeapRegionClosure {
2851 private:
2852 G1CollectedHeap* _g1h;
2853 outputStream* _out;
2854 VerifyOption _vo;
2855 bool _all;
2857 public:
2858 bool doHeapRegion(HeapRegion* hr) {
2859 HeapWord* b = hr->bottom();
2860 HeapWord* e = hr->end();
2861 HeapWord* t = hr->top();
2862 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2863 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2864 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2865 _out->cr();
2867 HeapWord* from = b;
2868 HeapWord* to = t;
2870 if (to > from) {
2871 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2872 _out->cr();
2873 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2874 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2875 _out->cr();
2876 }
2878 return false;
2879 }
2881 PrintReachableRegionClosure(outputStream* out,
2882 VerifyOption vo,
2883 bool all) :
2884 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2885 };
2887 void ConcurrentMark::print_reachable(const char* str,
2888 VerifyOption vo,
2889 bool all) {
2890 gclog_or_tty->cr();
2891 gclog_or_tty->print_cr("== Doing heap dump... ");
2893 if (G1PrintReachableBaseFile == NULL) {
2894 gclog_or_tty->print_cr(" #### error: no base file defined");
2895 return;
2896 }
2898 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2899 (JVM_MAXPATHLEN - 1)) {
2900 gclog_or_tty->print_cr(" #### error: file name too long");
2901 return;
2902 }
2904 char file_name[JVM_MAXPATHLEN];
2905 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2906 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2908 fileStream fout(file_name);
2909 if (!fout.is_open()) {
2910 gclog_or_tty->print_cr(" #### error: could not open file");
2911 return;
2912 }
2914 outputStream* out = &fout;
2915 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2916 out->cr();
2918 out->print_cr("--- ITERATING OVER REGIONS");
2919 out->cr();
2920 PrintReachableRegionClosure rcl(out, vo, all);
2921 _g1h->heap_region_iterate(&rcl);
2922 out->cr();
2924 gclog_or_tty->print_cr(" done");
2925 gclog_or_tty->flush();
2926 }
2928 #endif // PRODUCT
2930 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2931 // Note we are overriding the read-only view of the prev map here, via
2932 // the cast.
2933 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2934 }
2936 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2937 _nextMarkBitMap->clearRange(mr);
2938 }
2940 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) {
2941 clearRangePrevBitmap(mr);
2942 clearRangeNextBitmap(mr);
2943 }
2945 HeapRegion*
2946 ConcurrentMark::claim_region(uint worker_id) {
2947 // "checkpoint" the finger
2948 HeapWord* finger = _finger;
2950 // _heap_end will not change underneath our feet; it only changes at
2951 // yield points.
2952 while (finger < _heap_end) {
2953 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2955 // Note on how this code handles humongous regions. In the
2956 // normal case the finger will reach the start of a "starts
2957 // humongous" (SH) region. Its end will either be the end of the
2958 // last "continues humongous" (CH) region in the sequence, or the
2959 // standard end of the SH region (if the SH is the only region in
2960 // the sequence). That way claim_region() will skip over the CH
2961 // regions. However, there is a subtle race between a CM thread
2962 // executing this method and a mutator thread doing a humongous
2963 // object allocation. The two are not mutually exclusive as the CM
2964 // thread does not need to hold the Heap_lock when it gets
2965 // here. So there is a chance that claim_region() will come across
2966 // a free region that's in the progress of becoming a SH or a CH
2967 // region. In the former case, it will either
2968 // a) Miss the update to the region's end, in which case it will
2969 // visit every subsequent CH region, will find their bitmaps
2970 // empty, and do nothing, or
2971 // b) Will observe the update of the region's end (in which case
2972 // it will skip the subsequent CH regions).
2973 // If it comes across a region that suddenly becomes CH, the
2974 // scenario will be similar to b). So, the race between
2975 // claim_region() and a humongous object allocation might force us
2976 // to do a bit of unnecessary work (due to some unnecessary bitmap
2977 // iterations) but it should not introduce and correctness issues.
2978 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2979 HeapWord* bottom = curr_region->bottom();
2980 HeapWord* end = curr_region->end();
2981 HeapWord* limit = curr_region->next_top_at_mark_start();
2983 if (verbose_low()) {
2984 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2985 "["PTR_FORMAT", "PTR_FORMAT"), "
2986 "limit = "PTR_FORMAT,
2987 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2988 }
2990 // Is the gap between reading the finger and doing the CAS too long?
2991 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2992 if (res == finger) {
2993 // we succeeded
2995 // notice that _finger == end cannot be guaranteed here since,
2996 // someone else might have moved the finger even further
2997 assert(_finger >= end, "the finger should have moved forward");
2999 if (verbose_low()) {
3000 gclog_or_tty->print_cr("[%u] we were successful with region = "
3001 PTR_FORMAT, worker_id, p2i(curr_region));
3002 }
3004 if (limit > bottom) {
3005 if (verbose_low()) {
3006 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
3007 "returning it ", worker_id, p2i(curr_region));
3008 }
3009 return curr_region;
3010 } else {
3011 assert(limit == bottom,
3012 "the region limit should be at bottom");
3013 if (verbose_low()) {
3014 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
3015 "returning NULL", worker_id, p2i(curr_region));
3016 }
3017 // we return NULL and the caller should try calling
3018 // claim_region() again.
3019 return NULL;
3020 }
3021 } else {
3022 assert(_finger > finger, "the finger should have moved forward");
3023 if (verbose_low()) {
3024 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
3025 "global finger = "PTR_FORMAT", "
3026 "our finger = "PTR_FORMAT,
3027 worker_id, p2i(_finger), p2i(finger));
3028 }
3030 // read it again
3031 finger = _finger;
3032 }
3033 }
3035 return NULL;
3036 }
3038 #ifndef PRODUCT
3039 enum VerifyNoCSetOopsPhase {
3040 VerifyNoCSetOopsStack,
3041 VerifyNoCSetOopsQueues,
3042 VerifyNoCSetOopsSATBCompleted,
3043 VerifyNoCSetOopsSATBThread
3044 };
3046 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
3047 private:
3048 G1CollectedHeap* _g1h;
3049 VerifyNoCSetOopsPhase _phase;
3050 int _info;
3052 const char* phase_str() {
3053 switch (_phase) {
3054 case VerifyNoCSetOopsStack: return "Stack";
3055 case VerifyNoCSetOopsQueues: return "Queue";
3056 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
3057 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
3058 default: ShouldNotReachHere();
3059 }
3060 return NULL;
3061 }
3063 void do_object_work(oop obj) {
3064 guarantee(!_g1h->obj_in_cs(obj),
3065 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
3066 p2i((void*) obj), phase_str(), _info));
3067 }
3069 public:
3070 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3072 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3073 _phase = phase;
3074 _info = info;
3075 }
3077 virtual void do_oop(oop* p) {
3078 oop obj = oopDesc::load_decode_heap_oop(p);
3079 do_object_work(obj);
3080 }
3082 virtual void do_oop(narrowOop* p) {
3083 // We should not come across narrow oops while scanning marking
3084 // stacks and SATB buffers.
3085 ShouldNotReachHere();
3086 }
3088 virtual void do_object(oop obj) {
3089 do_object_work(obj);
3090 }
3091 };
3093 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3094 bool verify_enqueued_buffers,
3095 bool verify_thread_buffers,
3096 bool verify_fingers) {
3097 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3098 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3099 return;
3100 }
3102 VerifyNoCSetOopsClosure cl;
3104 if (verify_stacks) {
3105 // Verify entries on the global mark stack
3106 cl.set_phase(VerifyNoCSetOopsStack);
3107 _markStack.oops_do(&cl);
3109 // Verify entries on the task queues
3110 for (uint i = 0; i < _max_worker_id; i += 1) {
3111 cl.set_phase(VerifyNoCSetOopsQueues, i);
3112 CMTaskQueue* queue = _task_queues->queue(i);
3113 queue->oops_do(&cl);
3114 }
3115 }
3117 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3119 // Verify entries on the enqueued SATB buffers
3120 if (verify_enqueued_buffers) {
3121 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3122 satb_qs.iterate_completed_buffers_read_only(&cl);
3123 }
3125 // Verify entries on the per-thread SATB buffers
3126 if (verify_thread_buffers) {
3127 cl.set_phase(VerifyNoCSetOopsSATBThread);
3128 satb_qs.iterate_thread_buffers_read_only(&cl);
3129 }
3131 if (verify_fingers) {
3132 // Verify the global finger
3133 HeapWord* global_finger = finger();
3134 if (global_finger != NULL && global_finger < _heap_end) {
3135 // The global finger always points to a heap region boundary. We
3136 // use heap_region_containing_raw() to get the containing region
3137 // given that the global finger could be pointing to a free region
3138 // which subsequently becomes continues humongous. If that
3139 // happens, heap_region_containing() will return the bottom of the
3140 // corresponding starts humongous region and the check below will
3141 // not hold any more.
3142 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3143 guarantee(global_finger == global_hr->bottom(),
3144 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3145 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3146 }
3148 // Verify the task fingers
3149 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3150 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3151 CMTask* task = _tasks[i];
3152 HeapWord* task_finger = task->finger();
3153 if (task_finger != NULL && task_finger < _heap_end) {
3154 // See above note on the global finger verification.
3155 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3156 guarantee(task_finger == task_hr->bottom() ||
3157 !task_hr->in_collection_set(),
3158 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3159 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3160 }
3161 }
3162 }
3163 }
3164 #endif // PRODUCT
3166 // Aggregate the counting data that was constructed concurrently
3167 // with marking.
3168 class AggregateCountDataHRClosure: public HeapRegionClosure {
3169 G1CollectedHeap* _g1h;
3170 ConcurrentMark* _cm;
3171 CardTableModRefBS* _ct_bs;
3172 BitMap* _cm_card_bm;
3173 uint _max_worker_id;
3175 public:
3176 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3177 BitMap* cm_card_bm,
3178 uint max_worker_id) :
3179 _g1h(g1h), _cm(g1h->concurrent_mark()),
3180 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3181 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3183 bool doHeapRegion(HeapRegion* hr) {
3184 if (hr->continuesHumongous()) {
3185 // We will ignore these here and process them when their
3186 // associated "starts humongous" region is processed.
3187 // Note that we cannot rely on their associated
3188 // "starts humongous" region to have their bit set to 1
3189 // since, due to the region chunking in the parallel region
3190 // iteration, a "continues humongous" region might be visited
3191 // before its associated "starts humongous".
3192 return false;
3193 }
3195 HeapWord* start = hr->bottom();
3196 HeapWord* limit = hr->next_top_at_mark_start();
3197 HeapWord* end = hr->end();
3199 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3200 err_msg("Preconditions not met - "
3201 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3202 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3203 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3205 assert(hr->next_marked_bytes() == 0, "Precondition");
3207 if (start == limit) {
3208 // NTAMS of this region has not been set so nothing to do.
3209 return false;
3210 }
3212 // 'start' should be in the heap.
3213 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3214 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3215 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3217 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3218 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3219 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3221 // If ntams is not card aligned then we bump card bitmap index
3222 // for limit so that we get the all the cards spanned by
3223 // the object ending at ntams.
3224 // Note: if this is the last region in the heap then ntams
3225 // could be actually just beyond the end of the the heap;
3226 // limit_idx will then correspond to a (non-existent) card
3227 // that is also outside the heap.
3228 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3229 limit_idx += 1;
3230 }
3232 assert(limit_idx <= end_idx, "or else use atomics");
3234 // Aggregate the "stripe" in the count data associated with hr.
3235 uint hrs_index = hr->hrs_index();
3236 size_t marked_bytes = 0;
3238 for (uint i = 0; i < _max_worker_id; i += 1) {
3239 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3240 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3242 // Fetch the marked_bytes in this region for task i and
3243 // add it to the running total for this region.
3244 marked_bytes += marked_bytes_array[hrs_index];
3246 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3247 // into the global card bitmap.
3248 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3250 while (scan_idx < limit_idx) {
3251 assert(task_card_bm->at(scan_idx) == true, "should be");
3252 _cm_card_bm->set_bit(scan_idx);
3253 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3255 // BitMap::get_next_one_offset() can handle the case when
3256 // its left_offset parameter is greater than its right_offset
3257 // parameter. It does, however, have an early exit if
3258 // left_offset == right_offset. So let's limit the value
3259 // passed in for left offset here.
3260 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3261 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3262 }
3263 }
3265 // Update the marked bytes for this region.
3266 hr->add_to_marked_bytes(marked_bytes);
3268 // Next heap region
3269 return false;
3270 }
3271 };
3273 class G1AggregateCountDataTask: public AbstractGangTask {
3274 protected:
3275 G1CollectedHeap* _g1h;
3276 ConcurrentMark* _cm;
3277 BitMap* _cm_card_bm;
3278 uint _max_worker_id;
3279 int _active_workers;
3281 public:
3282 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3283 ConcurrentMark* cm,
3284 BitMap* cm_card_bm,
3285 uint max_worker_id,
3286 int n_workers) :
3287 AbstractGangTask("Count Aggregation"),
3288 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3289 _max_worker_id(max_worker_id),
3290 _active_workers(n_workers) { }
3292 void work(uint worker_id) {
3293 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3295 if (G1CollectedHeap::use_parallel_gc_threads()) {
3296 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3297 _active_workers,
3298 HeapRegion::AggregateCountClaimValue);
3299 } else {
3300 _g1h->heap_region_iterate(&cl);
3301 }
3302 }
3303 };
3306 void ConcurrentMark::aggregate_count_data() {
3307 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3308 _g1h->workers()->active_workers() :
3309 1);
3311 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3312 _max_worker_id, n_workers);
3314 if (G1CollectedHeap::use_parallel_gc_threads()) {
3315 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3316 "sanity check");
3317 _g1h->set_par_threads(n_workers);
3318 _g1h->workers()->run_task(&g1_par_agg_task);
3319 _g1h->set_par_threads(0);
3321 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3322 "sanity check");
3323 _g1h->reset_heap_region_claim_values();
3324 } else {
3325 g1_par_agg_task.work(0);
3326 }
3327 }
3329 // Clear the per-worker arrays used to store the per-region counting data
3330 void ConcurrentMark::clear_all_count_data() {
3331 // Clear the global card bitmap - it will be filled during
3332 // liveness count aggregation (during remark) and the
3333 // final counting task.
3334 _card_bm.clear();
3336 // Clear the global region bitmap - it will be filled as part
3337 // of the final counting task.
3338 _region_bm.clear();
3340 uint max_regions = _g1h->max_regions();
3341 assert(_max_worker_id > 0, "uninitialized");
3343 for (uint i = 0; i < _max_worker_id; i += 1) {
3344 BitMap* task_card_bm = count_card_bitmap_for(i);
3345 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3347 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3348 assert(marked_bytes_array != NULL, "uninitialized");
3350 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3351 task_card_bm->clear();
3352 }
3353 }
3355 void ConcurrentMark::print_stats() {
3356 if (verbose_stats()) {
3357 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3358 for (size_t i = 0; i < _active_tasks; ++i) {
3359 _tasks[i]->print_stats();
3360 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3361 }
3362 }
3363 }
3365 // abandon current marking iteration due to a Full GC
3366 void ConcurrentMark::abort() {
3367 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3368 // concurrent bitmap clearing.
3369 _nextMarkBitMap->clearAll();
3370 // Clear the liveness counting data
3371 clear_all_count_data();
3372 // Empty mark stack
3373 reset_marking_state();
3374 for (uint i = 0; i < _max_worker_id; ++i) {
3375 _tasks[i]->clear_region_fields();
3376 }
3377 _first_overflow_barrier_sync.abort();
3378 _second_overflow_barrier_sync.abort();
3379 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3380 if (!gc_id.is_undefined()) {
3381 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3382 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3383 _aborted_gc_id = gc_id;
3384 }
3385 _has_aborted = true;
3387 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3388 satb_mq_set.abandon_partial_marking();
3389 // This can be called either during or outside marking, we'll read
3390 // the expected_active value from the SATB queue set.
3391 satb_mq_set.set_active_all_threads(
3392 false, /* new active value */
3393 satb_mq_set.is_active() /* expected_active */);
3395 _g1h->trace_heap_after_concurrent_cycle();
3396 _g1h->register_concurrent_cycle_end();
3397 }
3399 const GCId& ConcurrentMark::concurrent_gc_id() {
3400 if (has_aborted()) {
3401 return _aborted_gc_id;
3402 }
3403 return _g1h->gc_tracer_cm()->gc_id();
3404 }
3406 static void print_ms_time_info(const char* prefix, const char* name,
3407 NumberSeq& ns) {
3408 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3409 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3410 if (ns.num() > 0) {
3411 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3412 prefix, ns.sd(), ns.maximum());
3413 }
3414 }
3416 void ConcurrentMark::print_summary_info() {
3417 gclog_or_tty->print_cr(" Concurrent marking:");
3418 print_ms_time_info(" ", "init marks", _init_times);
3419 print_ms_time_info(" ", "remarks", _remark_times);
3420 {
3421 print_ms_time_info(" ", "final marks", _remark_mark_times);
3422 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3424 }
3425 print_ms_time_info(" ", "cleanups", _cleanup_times);
3426 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3427 _total_counting_time,
3428 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3429 (double)_cleanup_times.num()
3430 : 0.0));
3431 if (G1ScrubRemSets) {
3432 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3433 _total_rs_scrub_time,
3434 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3435 (double)_cleanup_times.num()
3436 : 0.0));
3437 }
3438 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3439 (_init_times.sum() + _remark_times.sum() +
3440 _cleanup_times.sum())/1000.0);
3441 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3442 "(%8.2f s marking).",
3443 cmThread()->vtime_accum(),
3444 cmThread()->vtime_mark_accum());
3445 }
3447 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3448 if (use_parallel_marking_threads()) {
3449 _parallel_workers->print_worker_threads_on(st);
3450 }
3451 }
3453 void ConcurrentMark::print_on_error(outputStream* st) const {
3454 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3455 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3456 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3457 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3458 }
3460 // We take a break if someone is trying to stop the world.
3461 bool ConcurrentMark::do_yield_check(uint worker_id) {
3462 if (SuspendibleThreadSet::should_yield()) {
3463 if (worker_id == 0) {
3464 _g1h->g1_policy()->record_concurrent_pause();
3465 }
3466 SuspendibleThreadSet::yield();
3467 return true;
3468 } else {
3469 return false;
3470 }
3471 }
3473 bool ConcurrentMark::containing_card_is_marked(void* p) {
3474 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1);
3475 return _card_bm.at(offset >> CardTableModRefBS::card_shift);
3476 }
3478 bool ConcurrentMark::containing_cards_are_marked(void* start,
3479 void* last) {
3480 return containing_card_is_marked(start) &&
3481 containing_card_is_marked(last);
3482 }
3484 #ifndef PRODUCT
3485 // for debugging purposes
3486 void ConcurrentMark::print_finger() {
3487 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3488 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3489 for (uint i = 0; i < _max_worker_id; ++i) {
3490 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3491 }
3492 gclog_or_tty->cr();
3493 }
3494 #endif
3496 void CMTask::scan_object(oop obj) {
3497 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3499 if (_cm->verbose_high()) {
3500 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3501 _worker_id, p2i((void*) obj));
3502 }
3504 size_t obj_size = obj->size();
3505 _words_scanned += obj_size;
3507 obj->oop_iterate(_cm_oop_closure);
3508 statsOnly( ++_objs_scanned );
3509 check_limits();
3510 }
3512 // Closure for iteration over bitmaps
3513 class CMBitMapClosure : public BitMapClosure {
3514 private:
3515 // the bitmap that is being iterated over
3516 CMBitMap* _nextMarkBitMap;
3517 ConcurrentMark* _cm;
3518 CMTask* _task;
3520 public:
3521 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3522 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3524 bool do_bit(size_t offset) {
3525 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3526 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3527 assert( addr < _cm->finger(), "invariant");
3529 statsOnly( _task->increase_objs_found_on_bitmap() );
3530 assert(addr >= _task->finger(), "invariant");
3532 // We move that task's local finger along.
3533 _task->move_finger_to(addr);
3535 _task->scan_object(oop(addr));
3536 // we only partially drain the local queue and global stack
3537 _task->drain_local_queue(true);
3538 _task->drain_global_stack(true);
3540 // if the has_aborted flag has been raised, we need to bail out of
3541 // the iteration
3542 return !_task->has_aborted();
3543 }
3544 };
3546 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3547 ConcurrentMark* cm,
3548 CMTask* task)
3549 : _g1h(g1h), _cm(cm), _task(task) {
3550 assert(_ref_processor == NULL, "should be initialized to NULL");
3552 if (G1UseConcMarkReferenceProcessing) {
3553 _ref_processor = g1h->ref_processor_cm();
3554 assert(_ref_processor != NULL, "should not be NULL");
3555 }
3556 }
3558 void CMTask::setup_for_region(HeapRegion* hr) {
3559 // Separated the asserts so that we know which one fires.
3560 assert(hr != NULL,
3561 "claim_region() should have filtered out continues humongous regions");
3562 assert(!hr->continuesHumongous(),
3563 "claim_region() should have filtered out continues humongous regions");
3565 if (_cm->verbose_low()) {
3566 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3567 _worker_id, p2i(hr));
3568 }
3570 _curr_region = hr;
3571 _finger = hr->bottom();
3572 update_region_limit();
3573 }
3575 void CMTask::update_region_limit() {
3576 HeapRegion* hr = _curr_region;
3577 HeapWord* bottom = hr->bottom();
3578 HeapWord* limit = hr->next_top_at_mark_start();
3580 if (limit == bottom) {
3581 if (_cm->verbose_low()) {
3582 gclog_or_tty->print_cr("[%u] found an empty region "
3583 "["PTR_FORMAT", "PTR_FORMAT")",
3584 _worker_id, p2i(bottom), p2i(limit));
3585 }
3586 // The region was collected underneath our feet.
3587 // We set the finger to bottom to ensure that the bitmap
3588 // iteration that will follow this will not do anything.
3589 // (this is not a condition that holds when we set the region up,
3590 // as the region is not supposed to be empty in the first place)
3591 _finger = bottom;
3592 } else if (limit >= _region_limit) {
3593 assert(limit >= _finger, "peace of mind");
3594 } else {
3595 assert(limit < _region_limit, "only way to get here");
3596 // This can happen under some pretty unusual circumstances. An
3597 // evacuation pause empties the region underneath our feet (NTAMS
3598 // at bottom). We then do some allocation in the region (NTAMS
3599 // stays at bottom), followed by the region being used as a GC
3600 // alloc region (NTAMS will move to top() and the objects
3601 // originally below it will be grayed). All objects now marked in
3602 // the region are explicitly grayed, if below the global finger,
3603 // and we do not need in fact to scan anything else. So, we simply
3604 // set _finger to be limit to ensure that the bitmap iteration
3605 // doesn't do anything.
3606 _finger = limit;
3607 }
3609 _region_limit = limit;
3610 }
3612 void CMTask::giveup_current_region() {
3613 assert(_curr_region != NULL, "invariant");
3614 if (_cm->verbose_low()) {
3615 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3616 _worker_id, p2i(_curr_region));
3617 }
3618 clear_region_fields();
3619 }
3621 void CMTask::clear_region_fields() {
3622 // Values for these three fields that indicate that we're not
3623 // holding on to a region.
3624 _curr_region = NULL;
3625 _finger = NULL;
3626 _region_limit = NULL;
3627 }
3629 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3630 if (cm_oop_closure == NULL) {
3631 assert(_cm_oop_closure != NULL, "invariant");
3632 } else {
3633 assert(_cm_oop_closure == NULL, "invariant");
3634 }
3635 _cm_oop_closure = cm_oop_closure;
3636 }
3638 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3639 guarantee(nextMarkBitMap != NULL, "invariant");
3641 if (_cm->verbose_low()) {
3642 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3643 }
3645 _nextMarkBitMap = nextMarkBitMap;
3646 clear_region_fields();
3648 _calls = 0;
3649 _elapsed_time_ms = 0.0;
3650 _termination_time_ms = 0.0;
3651 _termination_start_time_ms = 0.0;
3653 #if _MARKING_STATS_
3654 _local_pushes = 0;
3655 _local_pops = 0;
3656 _local_max_size = 0;
3657 _objs_scanned = 0;
3658 _global_pushes = 0;
3659 _global_pops = 0;
3660 _global_max_size = 0;
3661 _global_transfers_to = 0;
3662 _global_transfers_from = 0;
3663 _regions_claimed = 0;
3664 _objs_found_on_bitmap = 0;
3665 _satb_buffers_processed = 0;
3666 _steal_attempts = 0;
3667 _steals = 0;
3668 _aborted = 0;
3669 _aborted_overflow = 0;
3670 _aborted_cm_aborted = 0;
3671 _aborted_yield = 0;
3672 _aborted_timed_out = 0;
3673 _aborted_satb = 0;
3674 _aborted_termination = 0;
3675 #endif // _MARKING_STATS_
3676 }
3678 bool CMTask::should_exit_termination() {
3679 regular_clock_call();
3680 // This is called when we are in the termination protocol. We should
3681 // quit if, for some reason, this task wants to abort or the global
3682 // stack is not empty (this means that we can get work from it).
3683 return !_cm->mark_stack_empty() || has_aborted();
3684 }
3686 void CMTask::reached_limit() {
3687 assert(_words_scanned >= _words_scanned_limit ||
3688 _refs_reached >= _refs_reached_limit ,
3689 "shouldn't have been called otherwise");
3690 regular_clock_call();
3691 }
3693 void CMTask::regular_clock_call() {
3694 if (has_aborted()) return;
3696 // First, we need to recalculate the words scanned and refs reached
3697 // limits for the next clock call.
3698 recalculate_limits();
3700 // During the regular clock call we do the following
3702 // (1) If an overflow has been flagged, then we abort.
3703 if (_cm->has_overflown()) {
3704 set_has_aborted();
3705 return;
3706 }
3708 // If we are not concurrent (i.e. we're doing remark) we don't need
3709 // to check anything else. The other steps are only needed during
3710 // the concurrent marking phase.
3711 if (!concurrent()) return;
3713 // (2) If marking has been aborted for Full GC, then we also abort.
3714 if (_cm->has_aborted()) {
3715 set_has_aborted();
3716 statsOnly( ++_aborted_cm_aborted );
3717 return;
3718 }
3720 double curr_time_ms = os::elapsedVTime() * 1000.0;
3722 // (3) If marking stats are enabled, then we update the step history.
3723 #if _MARKING_STATS_
3724 if (_words_scanned >= _words_scanned_limit) {
3725 ++_clock_due_to_scanning;
3726 }
3727 if (_refs_reached >= _refs_reached_limit) {
3728 ++_clock_due_to_marking;
3729 }
3731 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3732 _interval_start_time_ms = curr_time_ms;
3733 _all_clock_intervals_ms.add(last_interval_ms);
3735 if (_cm->verbose_medium()) {
3736 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3737 "scanned = %d%s, refs reached = %d%s",
3738 _worker_id, last_interval_ms,
3739 _words_scanned,
3740 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3741 _refs_reached,
3742 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3743 }
3744 #endif // _MARKING_STATS_
3746 // (4) We check whether we should yield. If we have to, then we abort.
3747 if (SuspendibleThreadSet::should_yield()) {
3748 // We should yield. To do this we abort the task. The caller is
3749 // responsible for yielding.
3750 set_has_aborted();
3751 statsOnly( ++_aborted_yield );
3752 return;
3753 }
3755 // (5) We check whether we've reached our time quota. If we have,
3756 // then we abort.
3757 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3758 if (elapsed_time_ms > _time_target_ms) {
3759 set_has_aborted();
3760 _has_timed_out = true;
3761 statsOnly( ++_aborted_timed_out );
3762 return;
3763 }
3765 // (6) Finally, we check whether there are enough completed STAB
3766 // buffers available for processing. If there are, we abort.
3767 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3768 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3769 if (_cm->verbose_low()) {
3770 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3771 _worker_id);
3772 }
3773 // we do need to process SATB buffers, we'll abort and restart
3774 // the marking task to do so
3775 set_has_aborted();
3776 statsOnly( ++_aborted_satb );
3777 return;
3778 }
3779 }
3781 void CMTask::recalculate_limits() {
3782 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3783 _words_scanned_limit = _real_words_scanned_limit;
3785 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3786 _refs_reached_limit = _real_refs_reached_limit;
3787 }
3789 void CMTask::decrease_limits() {
3790 // This is called when we believe that we're going to do an infrequent
3791 // operation which will increase the per byte scanned cost (i.e. move
3792 // entries to/from the global stack). It basically tries to decrease the
3793 // scanning limit so that the clock is called earlier.
3795 if (_cm->verbose_medium()) {
3796 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3797 }
3799 _words_scanned_limit = _real_words_scanned_limit -
3800 3 * words_scanned_period / 4;
3801 _refs_reached_limit = _real_refs_reached_limit -
3802 3 * refs_reached_period / 4;
3803 }
3805 void CMTask::move_entries_to_global_stack() {
3806 // local array where we'll store the entries that will be popped
3807 // from the local queue
3808 oop buffer[global_stack_transfer_size];
3810 int n = 0;
3811 oop obj;
3812 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3813 buffer[n] = obj;
3814 ++n;
3815 }
3817 if (n > 0) {
3818 // we popped at least one entry from the local queue
3820 statsOnly( ++_global_transfers_to; _local_pops += n );
3822 if (!_cm->mark_stack_push(buffer, n)) {
3823 if (_cm->verbose_low()) {
3824 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3825 _worker_id);
3826 }
3827 set_has_aborted();
3828 } else {
3829 // the transfer was successful
3831 if (_cm->verbose_medium()) {
3832 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3833 _worker_id, n);
3834 }
3835 statsOnly( int tmp_size = _cm->mark_stack_size();
3836 if (tmp_size > _global_max_size) {
3837 _global_max_size = tmp_size;
3838 }
3839 _global_pushes += n );
3840 }
3841 }
3843 // this operation was quite expensive, so decrease the limits
3844 decrease_limits();
3845 }
3847 void CMTask::get_entries_from_global_stack() {
3848 // local array where we'll store the entries that will be popped
3849 // from the global stack.
3850 oop buffer[global_stack_transfer_size];
3851 int n;
3852 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3853 assert(n <= global_stack_transfer_size,
3854 "we should not pop more than the given limit");
3855 if (n > 0) {
3856 // yes, we did actually pop at least one entry
3858 statsOnly( ++_global_transfers_from; _global_pops += n );
3859 if (_cm->verbose_medium()) {
3860 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3861 _worker_id, n);
3862 }
3863 for (int i = 0; i < n; ++i) {
3864 bool success = _task_queue->push(buffer[i]);
3865 // We only call this when the local queue is empty or under a
3866 // given target limit. So, we do not expect this push to fail.
3867 assert(success, "invariant");
3868 }
3870 statsOnly( int tmp_size = _task_queue->size();
3871 if (tmp_size > _local_max_size) {
3872 _local_max_size = tmp_size;
3873 }
3874 _local_pushes += n );
3875 }
3877 // this operation was quite expensive, so decrease the limits
3878 decrease_limits();
3879 }
3881 void CMTask::drain_local_queue(bool partially) {
3882 if (has_aborted()) return;
3884 // Decide what the target size is, depending whether we're going to
3885 // drain it partially (so that other tasks can steal if they run out
3886 // of things to do) or totally (at the very end).
3887 size_t target_size;
3888 if (partially) {
3889 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3890 } else {
3891 target_size = 0;
3892 }
3894 if (_task_queue->size() > target_size) {
3895 if (_cm->verbose_high()) {
3896 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3897 _worker_id, target_size);
3898 }
3900 oop obj;
3901 bool ret = _task_queue->pop_local(obj);
3902 while (ret) {
3903 statsOnly( ++_local_pops );
3905 if (_cm->verbose_high()) {
3906 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3907 p2i((void*) obj));
3908 }
3910 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3911 assert(!_g1h->is_on_master_free_list(
3912 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3914 scan_object(obj);
3916 if (_task_queue->size() <= target_size || has_aborted()) {
3917 ret = false;
3918 } else {
3919 ret = _task_queue->pop_local(obj);
3920 }
3921 }
3923 if (_cm->verbose_high()) {
3924 gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3925 _worker_id, _task_queue->size());
3926 }
3927 }
3928 }
3930 void CMTask::drain_global_stack(bool partially) {
3931 if (has_aborted()) return;
3933 // We have a policy to drain the local queue before we attempt to
3934 // drain the global stack.
3935 assert(partially || _task_queue->size() == 0, "invariant");
3937 // Decide what the target size is, depending whether we're going to
3938 // drain it partially (so that other tasks can steal if they run out
3939 // of things to do) or totally (at the very end). Notice that,
3940 // because we move entries from the global stack in chunks or
3941 // because another task might be doing the same, we might in fact
3942 // drop below the target. But, this is not a problem.
3943 size_t target_size;
3944 if (partially) {
3945 target_size = _cm->partial_mark_stack_size_target();
3946 } else {
3947 target_size = 0;
3948 }
3950 if (_cm->mark_stack_size() > target_size) {
3951 if (_cm->verbose_low()) {
3952 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3953 _worker_id, target_size);
3954 }
3956 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3957 get_entries_from_global_stack();
3958 drain_local_queue(partially);
3959 }
3961 if (_cm->verbose_low()) {
3962 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3963 _worker_id, _cm->mark_stack_size());
3964 }
3965 }
3966 }
3968 // SATB Queue has several assumptions on whether to call the par or
3969 // non-par versions of the methods. this is why some of the code is
3970 // replicated. We should really get rid of the single-threaded version
3971 // of the code to simplify things.
3972 void CMTask::drain_satb_buffers() {
3973 if (has_aborted()) return;
3975 // We set this so that the regular clock knows that we're in the
3976 // middle of draining buffers and doesn't set the abort flag when it
3977 // notices that SATB buffers are available for draining. It'd be
3978 // very counter productive if it did that. :-)
3979 _draining_satb_buffers = true;
3981 CMObjectClosure oc(this);
3982 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3983 if (G1CollectedHeap::use_parallel_gc_threads()) {
3984 satb_mq_set.set_par_closure(_worker_id, &oc);
3985 } else {
3986 satb_mq_set.set_closure(&oc);
3987 }
3989 // This keeps claiming and applying the closure to completed buffers
3990 // until we run out of buffers or we need to abort.
3991 if (G1CollectedHeap::use_parallel_gc_threads()) {
3992 while (!has_aborted() &&
3993 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
3994 if (_cm->verbose_medium()) {
3995 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3996 }
3997 statsOnly( ++_satb_buffers_processed );
3998 regular_clock_call();
3999 }
4000 } else {
4001 while (!has_aborted() &&
4002 satb_mq_set.apply_closure_to_completed_buffer()) {
4003 if (_cm->verbose_medium()) {
4004 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4005 }
4006 statsOnly( ++_satb_buffers_processed );
4007 regular_clock_call();
4008 }
4009 }
4011 _draining_satb_buffers = false;
4013 assert(has_aborted() ||
4014 concurrent() ||
4015 satb_mq_set.completed_buffers_num() == 0, "invariant");
4017 if (G1CollectedHeap::use_parallel_gc_threads()) {
4018 satb_mq_set.set_par_closure(_worker_id, NULL);
4019 } else {
4020 satb_mq_set.set_closure(NULL);
4021 }
4023 // again, this was a potentially expensive operation, decrease the
4024 // limits to get the regular clock call early
4025 decrease_limits();
4026 }
4028 void CMTask::print_stats() {
4029 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
4030 _worker_id, _calls);
4031 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
4032 _elapsed_time_ms, _termination_time_ms);
4033 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4034 _step_times_ms.num(), _step_times_ms.avg(),
4035 _step_times_ms.sd());
4036 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4037 _step_times_ms.maximum(), _step_times_ms.sum());
4039 #if _MARKING_STATS_
4040 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4041 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
4042 _all_clock_intervals_ms.sd());
4043 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4044 _all_clock_intervals_ms.maximum(),
4045 _all_clock_intervals_ms.sum());
4046 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
4047 _clock_due_to_scanning, _clock_due_to_marking);
4048 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
4049 _objs_scanned, _objs_found_on_bitmap);
4050 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
4051 _local_pushes, _local_pops, _local_max_size);
4052 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
4053 _global_pushes, _global_pops, _global_max_size);
4054 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
4055 _global_transfers_to,_global_transfers_from);
4056 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
4057 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
4058 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
4059 _steal_attempts, _steals);
4060 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
4061 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
4062 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
4063 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
4064 _aborted_timed_out, _aborted_satb, _aborted_termination);
4065 #endif // _MARKING_STATS_
4066 }
4068 /*****************************************************************************
4070 The do_marking_step(time_target_ms, ...) method is the building
4071 block of the parallel marking framework. It can be called in parallel
4072 with other invocations of do_marking_step() on different tasks
4073 (but only one per task, obviously) and concurrently with the
4074 mutator threads, or during remark, hence it eliminates the need
4075 for two versions of the code. When called during remark, it will
4076 pick up from where the task left off during the concurrent marking
4077 phase. Interestingly, tasks are also claimable during evacuation
4078 pauses too, since do_marking_step() ensures that it aborts before
4079 it needs to yield.
4081 The data structures that it uses to do marking work are the
4082 following:
4084 (1) Marking Bitmap. If there are gray objects that appear only
4085 on the bitmap (this happens either when dealing with an overflow
4086 or when the initial marking phase has simply marked the roots
4087 and didn't push them on the stack), then tasks claim heap
4088 regions whose bitmap they then scan to find gray objects. A
4089 global finger indicates where the end of the last claimed region
4090 is. A local finger indicates how far into the region a task has
4091 scanned. The two fingers are used to determine how to gray an
4092 object (i.e. whether simply marking it is OK, as it will be
4093 visited by a task in the future, or whether it needs to be also
4094 pushed on a stack).
4096 (2) Local Queue. The local queue of the task which is accessed
4097 reasonably efficiently by the task. Other tasks can steal from
4098 it when they run out of work. Throughout the marking phase, a
4099 task attempts to keep its local queue short but not totally
4100 empty, so that entries are available for stealing by other
4101 tasks. Only when there is no more work, a task will totally
4102 drain its local queue.
4104 (3) Global Mark Stack. This handles local queue overflow. During
4105 marking only sets of entries are moved between it and the local
4106 queues, as access to it requires a mutex and more fine-grain
4107 interaction with it which might cause contention. If it
4108 overflows, then the marking phase should restart and iterate
4109 over the bitmap to identify gray objects. Throughout the marking
4110 phase, tasks attempt to keep the global mark stack at a small
4111 length but not totally empty, so that entries are available for
4112 popping by other tasks. Only when there is no more work, tasks
4113 will totally drain the global mark stack.
4115 (4) SATB Buffer Queue. This is where completed SATB buffers are
4116 made available. Buffers are regularly removed from this queue
4117 and scanned for roots, so that the queue doesn't get too
4118 long. During remark, all completed buffers are processed, as
4119 well as the filled in parts of any uncompleted buffers.
4121 The do_marking_step() method tries to abort when the time target
4122 has been reached. There are a few other cases when the
4123 do_marking_step() method also aborts:
4125 (1) When the marking phase has been aborted (after a Full GC).
4127 (2) When a global overflow (on the global stack) has been
4128 triggered. Before the task aborts, it will actually sync up with
4129 the other tasks to ensure that all the marking data structures
4130 (local queues, stacks, fingers etc.) are re-initialized so that
4131 when do_marking_step() completes, the marking phase can
4132 immediately restart.
4134 (3) When enough completed SATB buffers are available. The
4135 do_marking_step() method only tries to drain SATB buffers right
4136 at the beginning. So, if enough buffers are available, the
4137 marking step aborts and the SATB buffers are processed at
4138 the beginning of the next invocation.
4140 (4) To yield. when we have to yield then we abort and yield
4141 right at the end of do_marking_step(). This saves us from a lot
4142 of hassle as, by yielding we might allow a Full GC. If this
4143 happens then objects will be compacted underneath our feet, the
4144 heap might shrink, etc. We save checking for this by just
4145 aborting and doing the yield right at the end.
4147 From the above it follows that the do_marking_step() method should
4148 be called in a loop (or, otherwise, regularly) until it completes.
4150 If a marking step completes without its has_aborted() flag being
4151 true, it means it has completed the current marking phase (and
4152 also all other marking tasks have done so and have all synced up).
4154 A method called regular_clock_call() is invoked "regularly" (in
4155 sub ms intervals) throughout marking. It is this clock method that
4156 checks all the abort conditions which were mentioned above and
4157 decides when the task should abort. A work-based scheme is used to
4158 trigger this clock method: when the number of object words the
4159 marking phase has scanned or the number of references the marking
4160 phase has visited reach a given limit. Additional invocations to
4161 the method clock have been planted in a few other strategic places
4162 too. The initial reason for the clock method was to avoid calling
4163 vtime too regularly, as it is quite expensive. So, once it was in
4164 place, it was natural to piggy-back all the other conditions on it
4165 too and not constantly check them throughout the code.
4167 If do_termination is true then do_marking_step will enter its
4168 termination protocol.
4170 The value of is_serial must be true when do_marking_step is being
4171 called serially (i.e. by the VMThread) and do_marking_step should
4172 skip any synchronization in the termination and overflow code.
4173 Examples include the serial remark code and the serial reference
4174 processing closures.
4176 The value of is_serial must be false when do_marking_step is
4177 being called by any of the worker threads in a work gang.
4178 Examples include the concurrent marking code (CMMarkingTask),
4179 the MT remark code, and the MT reference processing closures.
4181 *****************************************************************************/
4183 void CMTask::do_marking_step(double time_target_ms,
4184 bool do_termination,
4185 bool is_serial) {
4186 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4187 assert(concurrent() == _cm->concurrent(), "they should be the same");
4189 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4190 assert(_task_queues != NULL, "invariant");
4191 assert(_task_queue != NULL, "invariant");
4192 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4194 assert(!_claimed,
4195 "only one thread should claim this task at any one time");
4197 // OK, this doesn't safeguard again all possible scenarios, as it is
4198 // possible for two threads to set the _claimed flag at the same
4199 // time. But it is only for debugging purposes anyway and it will
4200 // catch most problems.
4201 _claimed = true;
4203 _start_time_ms = os::elapsedVTime() * 1000.0;
4204 statsOnly( _interval_start_time_ms = _start_time_ms );
4206 // If do_stealing is true then do_marking_step will attempt to
4207 // steal work from the other CMTasks. It only makes sense to
4208 // enable stealing when the termination protocol is enabled
4209 // and do_marking_step() is not being called serially.
4210 bool do_stealing = do_termination && !is_serial;
4212 double diff_prediction_ms =
4213 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4214 _time_target_ms = time_target_ms - diff_prediction_ms;
4216 // set up the variables that are used in the work-based scheme to
4217 // call the regular clock method
4218 _words_scanned = 0;
4219 _refs_reached = 0;
4220 recalculate_limits();
4222 // clear all flags
4223 clear_has_aborted();
4224 _has_timed_out = false;
4225 _draining_satb_buffers = false;
4227 ++_calls;
4229 if (_cm->verbose_low()) {
4230 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4231 "target = %1.2lfms >>>>>>>>>>",
4232 _worker_id, _calls, _time_target_ms);
4233 }
4235 // Set up the bitmap and oop closures. Anything that uses them is
4236 // eventually called from this method, so it is OK to allocate these
4237 // statically.
4238 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4239 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4240 set_cm_oop_closure(&cm_oop_closure);
4242 if (_cm->has_overflown()) {
4243 // This can happen if the mark stack overflows during a GC pause
4244 // and this task, after a yield point, restarts. We have to abort
4245 // as we need to get into the overflow protocol which happens
4246 // right at the end of this task.
4247 set_has_aborted();
4248 }
4250 // First drain any available SATB buffers. After this, we will not
4251 // look at SATB buffers before the next invocation of this method.
4252 // If enough completed SATB buffers are queued up, the regular clock
4253 // will abort this task so that it restarts.
4254 drain_satb_buffers();
4255 // ...then partially drain the local queue and the global stack
4256 drain_local_queue(true);
4257 drain_global_stack(true);
4259 do {
4260 if (!has_aborted() && _curr_region != NULL) {
4261 // This means that we're already holding on to a region.
4262 assert(_finger != NULL, "if region is not NULL, then the finger "
4263 "should not be NULL either");
4265 // We might have restarted this task after an evacuation pause
4266 // which might have evacuated the region we're holding on to
4267 // underneath our feet. Let's read its limit again to make sure
4268 // that we do not iterate over a region of the heap that
4269 // contains garbage (update_region_limit() will also move
4270 // _finger to the start of the region if it is found empty).
4271 update_region_limit();
4272 // We will start from _finger not from the start of the region,
4273 // as we might be restarting this task after aborting half-way
4274 // through scanning this region. In this case, _finger points to
4275 // the address where we last found a marked object. If this is a
4276 // fresh region, _finger points to start().
4277 MemRegion mr = MemRegion(_finger, _region_limit);
4279 if (_cm->verbose_low()) {
4280 gclog_or_tty->print_cr("[%u] we're scanning part "
4281 "["PTR_FORMAT", "PTR_FORMAT") "
4282 "of region "HR_FORMAT,
4283 _worker_id, p2i(_finger), p2i(_region_limit),
4284 HR_FORMAT_PARAMS(_curr_region));
4285 }
4287 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4288 "humongous regions should go around loop once only");
4290 // Some special cases:
4291 // If the memory region is empty, we can just give up the region.
4292 // If the current region is humongous then we only need to check
4293 // the bitmap for the bit associated with the start of the object,
4294 // scan the object if it's live, and give up the region.
4295 // Otherwise, let's iterate over the bitmap of the part of the region
4296 // that is left.
4297 // If the iteration is successful, give up the region.
4298 if (mr.is_empty()) {
4299 giveup_current_region();
4300 regular_clock_call();
4301 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4302 if (_nextMarkBitMap->isMarked(mr.start())) {
4303 // The object is marked - apply the closure
4304 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4305 bitmap_closure.do_bit(offset);
4306 }
4307 // Even if this task aborted while scanning the humongous object
4308 // we can (and should) give up the current region.
4309 giveup_current_region();
4310 regular_clock_call();
4311 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4312 giveup_current_region();
4313 regular_clock_call();
4314 } else {
4315 assert(has_aborted(), "currently the only way to do so");
4316 // The only way to abort the bitmap iteration is to return
4317 // false from the do_bit() method. However, inside the
4318 // do_bit() method we move the _finger to point to the
4319 // object currently being looked at. So, if we bail out, we
4320 // have definitely set _finger to something non-null.
4321 assert(_finger != NULL, "invariant");
4323 // Region iteration was actually aborted. So now _finger
4324 // points to the address of the object we last scanned. If we
4325 // leave it there, when we restart this task, we will rescan
4326 // the object. It is easy to avoid this. We move the finger by
4327 // enough to point to the next possible object header (the
4328 // bitmap knows by how much we need to move it as it knows its
4329 // granularity).
4330 assert(_finger < _region_limit, "invariant");
4331 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4332 // Check if bitmap iteration was aborted while scanning the last object
4333 if (new_finger >= _region_limit) {
4334 giveup_current_region();
4335 } else {
4336 move_finger_to(new_finger);
4337 }
4338 }
4339 }
4340 // At this point we have either completed iterating over the
4341 // region we were holding on to, or we have aborted.
4343 // We then partially drain the local queue and the global stack.
4344 // (Do we really need this?)
4345 drain_local_queue(true);
4346 drain_global_stack(true);
4348 // Read the note on the claim_region() method on why it might
4349 // return NULL with potentially more regions available for
4350 // claiming and why we have to check out_of_regions() to determine
4351 // whether we're done or not.
4352 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4353 // We are going to try to claim a new region. We should have
4354 // given up on the previous one.
4355 // Separated the asserts so that we know which one fires.
4356 assert(_curr_region == NULL, "invariant");
4357 assert(_finger == NULL, "invariant");
4358 assert(_region_limit == NULL, "invariant");
4359 if (_cm->verbose_low()) {
4360 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4361 }
4362 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4363 if (claimed_region != NULL) {
4364 // Yes, we managed to claim one
4365 statsOnly( ++_regions_claimed );
4367 if (_cm->verbose_low()) {
4368 gclog_or_tty->print_cr("[%u] we successfully claimed "
4369 "region "PTR_FORMAT,
4370 _worker_id, p2i(claimed_region));
4371 }
4373 setup_for_region(claimed_region);
4374 assert(_curr_region == claimed_region, "invariant");
4375 }
4376 // It is important to call the regular clock here. It might take
4377 // a while to claim a region if, for example, we hit a large
4378 // block of empty regions. So we need to call the regular clock
4379 // method once round the loop to make sure it's called
4380 // frequently enough.
4381 regular_clock_call();
4382 }
4384 if (!has_aborted() && _curr_region == NULL) {
4385 assert(_cm->out_of_regions(),
4386 "at this point we should be out of regions");
4387 }
4388 } while ( _curr_region != NULL && !has_aborted());
4390 if (!has_aborted()) {
4391 // We cannot check whether the global stack is empty, since other
4392 // tasks might be pushing objects to it concurrently.
4393 assert(_cm->out_of_regions(),
4394 "at this point we should be out of regions");
4396 if (_cm->verbose_low()) {
4397 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4398 }
4400 // Try to reduce the number of available SATB buffers so that
4401 // remark has less work to do.
4402 drain_satb_buffers();
4403 }
4405 // Since we've done everything else, we can now totally drain the
4406 // local queue and global stack.
4407 drain_local_queue(false);
4408 drain_global_stack(false);
4410 // Attempt at work stealing from other task's queues.
4411 if (do_stealing && !has_aborted()) {
4412 // We have not aborted. This means that we have finished all that
4413 // we could. Let's try to do some stealing...
4415 // We cannot check whether the global stack is empty, since other
4416 // tasks might be pushing objects to it concurrently.
4417 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4418 "only way to reach here");
4420 if (_cm->verbose_low()) {
4421 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4422 }
4424 while (!has_aborted()) {
4425 oop obj;
4426 statsOnly( ++_steal_attempts );
4428 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4429 if (_cm->verbose_medium()) {
4430 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4431 _worker_id, p2i((void*) obj));
4432 }
4434 statsOnly( ++_steals );
4436 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4437 "any stolen object should be marked");
4438 scan_object(obj);
4440 // And since we're towards the end, let's totally drain the
4441 // local queue and global stack.
4442 drain_local_queue(false);
4443 drain_global_stack(false);
4444 } else {
4445 break;
4446 }
4447 }
4448 }
4450 // If we are about to wrap up and go into termination, check if we
4451 // should raise the overflow flag.
4452 if (do_termination && !has_aborted()) {
4453 if (_cm->force_overflow()->should_force()) {
4454 _cm->set_has_overflown();
4455 regular_clock_call();
4456 }
4457 }
4459 // We still haven't aborted. Now, let's try to get into the
4460 // termination protocol.
4461 if (do_termination && !has_aborted()) {
4462 // We cannot check whether the global stack is empty, since other
4463 // tasks might be concurrently pushing objects on it.
4464 // Separated the asserts so that we know which one fires.
4465 assert(_cm->out_of_regions(), "only way to reach here");
4466 assert(_task_queue->size() == 0, "only way to reach here");
4468 if (_cm->verbose_low()) {
4469 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4470 }
4472 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4474 // The CMTask class also extends the TerminatorTerminator class,
4475 // hence its should_exit_termination() method will also decide
4476 // whether to exit the termination protocol or not.
4477 bool finished = (is_serial ||
4478 _cm->terminator()->offer_termination(this));
4479 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4480 _termination_time_ms +=
4481 termination_end_time_ms - _termination_start_time_ms;
4483 if (finished) {
4484 // We're all done.
4486 if (_worker_id == 0) {
4487 // let's allow task 0 to do this
4488 if (concurrent()) {
4489 assert(_cm->concurrent_marking_in_progress(), "invariant");
4490 // we need to set this to false before the next
4491 // safepoint. This way we ensure that the marking phase
4492 // doesn't observe any more heap expansions.
4493 _cm->clear_concurrent_marking_in_progress();
4494 }
4495 }
4497 // We can now guarantee that the global stack is empty, since
4498 // all other tasks have finished. We separated the guarantees so
4499 // that, if a condition is false, we can immediately find out
4500 // which one.
4501 guarantee(_cm->out_of_regions(), "only way to reach here");
4502 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4503 guarantee(_task_queue->size() == 0, "only way to reach here");
4504 guarantee(!_cm->has_overflown(), "only way to reach here");
4505 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4507 if (_cm->verbose_low()) {
4508 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4509 }
4510 } else {
4511 // Apparently there's more work to do. Let's abort this task. It
4512 // will restart it and we can hopefully find more things to do.
4514 if (_cm->verbose_low()) {
4515 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4516 _worker_id);
4517 }
4519 set_has_aborted();
4520 statsOnly( ++_aborted_termination );
4521 }
4522 }
4524 // Mainly for debugging purposes to make sure that a pointer to the
4525 // closure which was statically allocated in this frame doesn't
4526 // escape it by accident.
4527 set_cm_oop_closure(NULL);
4528 double end_time_ms = os::elapsedVTime() * 1000.0;
4529 double elapsed_time_ms = end_time_ms - _start_time_ms;
4530 // Update the step history.
4531 _step_times_ms.add(elapsed_time_ms);
4533 if (has_aborted()) {
4534 // The task was aborted for some reason.
4536 statsOnly( ++_aborted );
4538 if (_has_timed_out) {
4539 double diff_ms = elapsed_time_ms - _time_target_ms;
4540 // Keep statistics of how well we did with respect to hitting
4541 // our target only if we actually timed out (if we aborted for
4542 // other reasons, then the results might get skewed).
4543 _marking_step_diffs_ms.add(diff_ms);
4544 }
4546 if (_cm->has_overflown()) {
4547 // This is the interesting one. We aborted because a global
4548 // overflow was raised. This means we have to restart the
4549 // marking phase and start iterating over regions. However, in
4550 // order to do this we have to make sure that all tasks stop
4551 // what they are doing and re-initialise in a safe manner. We
4552 // will achieve this with the use of two barrier sync points.
4554 if (_cm->verbose_low()) {
4555 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4556 }
4558 if (!is_serial) {
4559 // We only need to enter the sync barrier if being called
4560 // from a parallel context
4561 _cm->enter_first_sync_barrier(_worker_id);
4563 // When we exit this sync barrier we know that all tasks have
4564 // stopped doing marking work. So, it's now safe to
4565 // re-initialise our data structures. At the end of this method,
4566 // task 0 will clear the global data structures.
4567 }
4569 statsOnly( ++_aborted_overflow );
4571 // We clear the local state of this task...
4572 clear_region_fields();
4574 if (!is_serial) {
4575 // ...and enter the second barrier.
4576 _cm->enter_second_sync_barrier(_worker_id);
4577 }
4578 // At this point, if we're during the concurrent phase of
4579 // marking, everything has been re-initialized and we're
4580 // ready to restart.
4581 }
4583 if (_cm->verbose_low()) {
4584 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4585 "elapsed = %1.2lfms <<<<<<<<<<",
4586 _worker_id, _time_target_ms, elapsed_time_ms);
4587 if (_cm->has_aborted()) {
4588 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4589 _worker_id);
4590 }
4591 }
4592 } else {
4593 if (_cm->verbose_low()) {
4594 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4595 "elapsed = %1.2lfms <<<<<<<<<<",
4596 _worker_id, _time_target_ms, elapsed_time_ms);
4597 }
4598 }
4600 _claimed = false;
4601 }
4603 CMTask::CMTask(uint worker_id,
4604 ConcurrentMark* cm,
4605 size_t* marked_bytes,
4606 BitMap* card_bm,
4607 CMTaskQueue* task_queue,
4608 CMTaskQueueSet* task_queues)
4609 : _g1h(G1CollectedHeap::heap()),
4610 _worker_id(worker_id), _cm(cm),
4611 _claimed(false),
4612 _nextMarkBitMap(NULL), _hash_seed(17),
4613 _task_queue(task_queue),
4614 _task_queues(task_queues),
4615 _cm_oop_closure(NULL),
4616 _marked_bytes_array(marked_bytes),
4617 _card_bm(card_bm) {
4618 guarantee(task_queue != NULL, "invariant");
4619 guarantee(task_queues != NULL, "invariant");
4621 statsOnly( _clock_due_to_scanning = 0;
4622 _clock_due_to_marking = 0 );
4624 _marking_step_diffs_ms.add(0.5);
4625 }
4627 // These are formatting macros that are used below to ensure
4628 // consistent formatting. The *_H_* versions are used to format the
4629 // header for a particular value and they should be kept consistent
4630 // with the corresponding macro. Also note that most of the macros add
4631 // the necessary white space (as a prefix) which makes them a bit
4632 // easier to compose.
4634 // All the output lines are prefixed with this string to be able to
4635 // identify them easily in a large log file.
4636 #define G1PPRL_LINE_PREFIX "###"
4638 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4639 #ifdef _LP64
4640 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4641 #else // _LP64
4642 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4643 #endif // _LP64
4645 // For per-region info
4646 #define G1PPRL_TYPE_FORMAT " %-4s"
4647 #define G1PPRL_TYPE_H_FORMAT " %4s"
4648 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4649 #define G1PPRL_BYTE_H_FORMAT " %9s"
4650 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4651 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4653 // For summary info
4654 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4655 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4656 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4657 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4659 G1PrintRegionLivenessInfoClosure::
4660 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4661 : _out(out),
4662 _total_used_bytes(0), _total_capacity_bytes(0),
4663 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4664 _hum_used_bytes(0), _hum_capacity_bytes(0),
4665 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4666 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4667 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4668 MemRegion g1_committed = g1h->g1_committed();
4669 MemRegion g1_reserved = g1h->g1_reserved();
4670 double now = os::elapsedTime();
4672 // Print the header of the output.
4673 _out->cr();
4674 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4675 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4676 G1PPRL_SUM_ADDR_FORMAT("committed")
4677 G1PPRL_SUM_ADDR_FORMAT("reserved")
4678 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4679 p2i(g1_committed.start()), p2i(g1_committed.end()),
4680 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4681 HeapRegion::GrainBytes);
4682 _out->print_cr(G1PPRL_LINE_PREFIX);
4683 _out->print_cr(G1PPRL_LINE_PREFIX
4684 G1PPRL_TYPE_H_FORMAT
4685 G1PPRL_ADDR_BASE_H_FORMAT
4686 G1PPRL_BYTE_H_FORMAT
4687 G1PPRL_BYTE_H_FORMAT
4688 G1PPRL_BYTE_H_FORMAT
4689 G1PPRL_DOUBLE_H_FORMAT
4690 G1PPRL_BYTE_H_FORMAT
4691 G1PPRL_BYTE_H_FORMAT,
4692 "type", "address-range",
4693 "used", "prev-live", "next-live", "gc-eff",
4694 "remset", "code-roots");
4695 _out->print_cr(G1PPRL_LINE_PREFIX
4696 G1PPRL_TYPE_H_FORMAT
4697 G1PPRL_ADDR_BASE_H_FORMAT
4698 G1PPRL_BYTE_H_FORMAT
4699 G1PPRL_BYTE_H_FORMAT
4700 G1PPRL_BYTE_H_FORMAT
4701 G1PPRL_DOUBLE_H_FORMAT
4702 G1PPRL_BYTE_H_FORMAT
4703 G1PPRL_BYTE_H_FORMAT,
4704 "", "",
4705 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4706 "(bytes)", "(bytes)");
4707 }
4709 // It takes as a parameter a reference to one of the _hum_* fields, it
4710 // deduces the corresponding value for a region in a humongous region
4711 // series (either the region size, or what's left if the _hum_* field
4712 // is < the region size), and updates the _hum_* field accordingly.
4713 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4714 size_t bytes = 0;
4715 // The > 0 check is to deal with the prev and next live bytes which
4716 // could be 0.
4717 if (*hum_bytes > 0) {
4718 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4719 *hum_bytes -= bytes;
4720 }
4721 return bytes;
4722 }
4724 // It deduces the values for a region in a humongous region series
4725 // from the _hum_* fields and updates those accordingly. It assumes
4726 // that that _hum_* fields have already been set up from the "starts
4727 // humongous" region and we visit the regions in address order.
4728 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4729 size_t* capacity_bytes,
4730 size_t* prev_live_bytes,
4731 size_t* next_live_bytes) {
4732 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4733 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4734 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4735 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4736 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4737 }
4739 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4740 const char* type = "";
4741 HeapWord* bottom = r->bottom();
4742 HeapWord* end = r->end();
4743 size_t capacity_bytes = r->capacity();
4744 size_t used_bytes = r->used();
4745 size_t prev_live_bytes = r->live_bytes();
4746 size_t next_live_bytes = r->next_live_bytes();
4747 double gc_eff = r->gc_efficiency();
4748 size_t remset_bytes = r->rem_set()->mem_size();
4749 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4751 if (r->used() == 0) {
4752 type = "FREE";
4753 } else if (r->is_survivor()) {
4754 type = "SURV";
4755 } else if (r->is_young()) {
4756 type = "EDEN";
4757 } else if (r->startsHumongous()) {
4758 type = "HUMS";
4760 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4761 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4762 "they should have been zeroed after the last time we used them");
4763 // Set up the _hum_* fields.
4764 _hum_capacity_bytes = capacity_bytes;
4765 _hum_used_bytes = used_bytes;
4766 _hum_prev_live_bytes = prev_live_bytes;
4767 _hum_next_live_bytes = next_live_bytes;
4768 get_hum_bytes(&used_bytes, &capacity_bytes,
4769 &prev_live_bytes, &next_live_bytes);
4770 end = bottom + HeapRegion::GrainWords;
4771 } else if (r->continuesHumongous()) {
4772 type = "HUMC";
4773 get_hum_bytes(&used_bytes, &capacity_bytes,
4774 &prev_live_bytes, &next_live_bytes);
4775 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4776 } else {
4777 type = "OLD";
4778 }
4780 _total_used_bytes += used_bytes;
4781 _total_capacity_bytes += capacity_bytes;
4782 _total_prev_live_bytes += prev_live_bytes;
4783 _total_next_live_bytes += next_live_bytes;
4784 _total_remset_bytes += remset_bytes;
4785 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4787 // Print a line for this particular region.
4788 _out->print_cr(G1PPRL_LINE_PREFIX
4789 G1PPRL_TYPE_FORMAT
4790 G1PPRL_ADDR_BASE_FORMAT
4791 G1PPRL_BYTE_FORMAT
4792 G1PPRL_BYTE_FORMAT
4793 G1PPRL_BYTE_FORMAT
4794 G1PPRL_DOUBLE_FORMAT
4795 G1PPRL_BYTE_FORMAT
4796 G1PPRL_BYTE_FORMAT,
4797 type, p2i(bottom), p2i(end),
4798 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4799 remset_bytes, strong_code_roots_bytes);
4801 return false;
4802 }
4804 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4805 // add static memory usages to remembered set sizes
4806 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4807 // Print the footer of the output.
4808 _out->print_cr(G1PPRL_LINE_PREFIX);
4809 _out->print_cr(G1PPRL_LINE_PREFIX
4810 " SUMMARY"
4811 G1PPRL_SUM_MB_FORMAT("capacity")
4812 G1PPRL_SUM_MB_PERC_FORMAT("used")
4813 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4814 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4815 G1PPRL_SUM_MB_FORMAT("remset")
4816 G1PPRL_SUM_MB_FORMAT("code-roots"),
4817 bytes_to_mb(_total_capacity_bytes),
4818 bytes_to_mb(_total_used_bytes),
4819 perc(_total_used_bytes, _total_capacity_bytes),
4820 bytes_to_mb(_total_prev_live_bytes),
4821 perc(_total_prev_live_bytes, _total_capacity_bytes),
4822 bytes_to_mb(_total_next_live_bytes),
4823 perc(_total_next_live_bytes, _total_capacity_bytes),
4824 bytes_to_mb(_total_remset_bytes),
4825 bytes_to_mb(_total_strong_code_roots_bytes));
4826 _out->cr();
4827 }