Tue, 19 Aug 2014 10:50:27 +0200
8054818: Refactor HeapRegionSeq to manage heap region and auxiliary data
Summary: Let HeapRegionSeq manage the heap region and auxiliary data to decrease the amount of responsibilities of G1CollectedHeap, and encapsulate this work from other code.
Reviewed-by: jwilhelm, jmasa, mgerdin, brutisso
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_heap_boundaries(MemRegion bounds, 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 assert(bounds.start() == _heap_start, "start shouldn't change");
737 HeapWord* new_end = bounds.end();
738 if (new_end > _heap_end) {
739 // The heap has been expanded.
741 _heap_end = new_end;
742 }
743 // Notice that the heap can also shrink. However, this only happens
744 // during a Full GC (at least currently) and the entire marking
745 // phase will bail out and the task will not be restarted. So, let's
746 // do nothing.
747 }
749 void ConcurrentMark::reset() {
750 // Starting values for these two. This should be called in a STW
751 // phase. CM will be notified of any future g1_committed expansions
752 // will be at the end of evacuation pauses, when tasks are
753 // inactive.
754 MemRegion committed = _g1h->g1_committed();
755 _heap_start = committed.start();
756 _heap_end = committed.end();
758 // Separated the asserts so that we know which one fires.
759 assert(_heap_start != NULL, "heap bounds should look ok");
760 assert(_heap_end != NULL, "heap bounds should look ok");
761 assert(_heap_start < _heap_end, "heap bounds should look ok");
763 // Reset all the marking data structures and any necessary flags
764 reset_marking_state();
766 if (verbose_low()) {
767 gclog_or_tty->print_cr("[global] resetting");
768 }
770 // We do reset all of them, since different phases will use
771 // different number of active threads. So, it's easiest to have all
772 // of them ready.
773 for (uint i = 0; i < _max_worker_id; ++i) {
774 _tasks[i]->reset(_nextMarkBitMap);
775 }
777 // we need this to make sure that the flag is on during the evac
778 // pause with initial mark piggy-backed
779 set_concurrent_marking_in_progress();
780 }
783 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
784 _markStack.set_should_expand();
785 _markStack.setEmpty(); // Also clears the _markStack overflow flag
786 if (clear_overflow) {
787 clear_has_overflown();
788 } else {
789 assert(has_overflown(), "pre-condition");
790 }
791 _finger = _heap_start;
793 for (uint i = 0; i < _max_worker_id; ++i) {
794 CMTaskQueue* queue = _task_queues->queue(i);
795 queue->set_empty();
796 }
797 }
799 void ConcurrentMark::set_concurrency(uint active_tasks) {
800 assert(active_tasks <= _max_worker_id, "we should not have more");
802 _active_tasks = active_tasks;
803 // Need to update the three data structures below according to the
804 // number of active threads for this phase.
805 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
806 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
807 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
808 }
810 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
811 set_concurrency(active_tasks);
813 _concurrent = concurrent;
814 // We propagate this to all tasks, not just the active ones.
815 for (uint i = 0; i < _max_worker_id; ++i)
816 _tasks[i]->set_concurrent(concurrent);
818 if (concurrent) {
819 set_concurrent_marking_in_progress();
820 } else {
821 // We currently assume that the concurrent flag has been set to
822 // false before we start remark. At this point we should also be
823 // in a STW phase.
824 assert(!concurrent_marking_in_progress(), "invariant");
825 assert(out_of_regions(),
826 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
827 p2i(_finger), p2i(_heap_end)));
828 update_heap_boundaries(_g1h->g1_committed(), true);
829 }
830 }
832 void ConcurrentMark::set_non_marking_state() {
833 // We set the global marking state to some default values when we're
834 // not doing marking.
835 reset_marking_state();
836 _active_tasks = 0;
837 clear_concurrent_marking_in_progress();
838 }
840 ConcurrentMark::~ConcurrentMark() {
841 // The ConcurrentMark instance is never freed.
842 ShouldNotReachHere();
843 }
845 void ConcurrentMark::clearNextBitmap() {
846 G1CollectedHeap* g1h = G1CollectedHeap::heap();
847 G1CollectorPolicy* g1p = g1h->g1_policy();
849 // Make sure that the concurrent mark thread looks to still be in
850 // the current cycle.
851 guarantee(cmThread()->during_cycle(), "invariant");
853 // We are finishing up the current cycle by clearing the next
854 // marking bitmap and getting it ready for the next cycle. During
855 // this time no other cycle can start. So, let's make sure that this
856 // is the case.
857 guarantee(!g1h->mark_in_progress(), "invariant");
859 // clear the mark bitmap (no grey objects to start with).
860 // We need to do this in chunks and offer to yield in between
861 // each chunk.
862 HeapWord* start = _nextMarkBitMap->startWord();
863 HeapWord* end = _nextMarkBitMap->endWord();
864 HeapWord* cur = start;
865 size_t chunkSize = M;
866 while (cur < end) {
867 HeapWord* next = cur + chunkSize;
868 if (next > end) {
869 next = end;
870 }
871 MemRegion mr(cur,next);
872 _nextMarkBitMap->clearRange(mr);
873 cur = next;
874 do_yield_check();
876 // Repeat the asserts from above. We'll do them as asserts here to
877 // minimize their overhead on the product. However, we'll have
878 // them as guarantees at the beginning / end of the bitmap
879 // clearing to get some checking in the product.
880 assert(cmThread()->during_cycle(), "invariant");
881 assert(!g1h->mark_in_progress(), "invariant");
882 }
884 // Clear the liveness counting data
885 clear_all_count_data();
887 // Repeat the asserts from above.
888 guarantee(cmThread()->during_cycle(), "invariant");
889 guarantee(!g1h->mark_in_progress(), "invariant");
890 }
892 bool ConcurrentMark::nextMarkBitmapIsClear() {
893 return _nextMarkBitMap->getNextMarkedWordAddress(_heap_start, _heap_end) == _heap_end;
894 }
896 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
897 public:
898 bool doHeapRegion(HeapRegion* r) {
899 if (!r->continuesHumongous()) {
900 r->note_start_of_marking();
901 }
902 return false;
903 }
904 };
906 void ConcurrentMark::checkpointRootsInitialPre() {
907 G1CollectedHeap* g1h = G1CollectedHeap::heap();
908 G1CollectorPolicy* g1p = g1h->g1_policy();
910 _has_aborted = false;
912 #ifndef PRODUCT
913 if (G1PrintReachableAtInitialMark) {
914 print_reachable("at-cycle-start",
915 VerifyOption_G1UsePrevMarking, true /* all */);
916 }
917 #endif
919 // Initialise marking structures. This has to be done in a STW phase.
920 reset();
922 // For each region note start of marking.
923 NoteStartOfMarkHRClosure startcl;
924 g1h->heap_region_iterate(&startcl);
925 }
928 void ConcurrentMark::checkpointRootsInitialPost() {
929 G1CollectedHeap* g1h = G1CollectedHeap::heap();
931 // If we force an overflow during remark, the remark operation will
932 // actually abort and we'll restart concurrent marking. If we always
933 // force an oveflow during remark we'll never actually complete the
934 // marking phase. So, we initilize this here, at the start of the
935 // cycle, so that at the remaining overflow number will decrease at
936 // every remark and we'll eventually not need to cause one.
937 force_overflow_stw()->init();
939 // Start Concurrent Marking weak-reference discovery.
940 ReferenceProcessor* rp = g1h->ref_processor_cm();
941 // enable ("weak") refs discovery
942 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
943 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
945 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
946 // This is the start of the marking cycle, we're expected all
947 // threads to have SATB queues with active set to false.
948 satb_mq_set.set_active_all_threads(true, /* new active value */
949 false /* expected_active */);
951 _root_regions.prepare_for_scan();
953 // update_g1_committed() will be called at the end of an evac pause
954 // when marking is on. So, it's also called at the end of the
955 // initial-mark pause to update the heap end, if the heap expands
956 // during it. No need to call it here.
957 }
959 /*
960 * Notice that in the next two methods, we actually leave the STS
961 * during the barrier sync and join it immediately afterwards. If we
962 * do not do this, the following deadlock can occur: one thread could
963 * be in the barrier sync code, waiting for the other thread to also
964 * sync up, whereas another one could be trying to yield, while also
965 * waiting for the other threads to sync up too.
966 *
967 * Note, however, that this code is also used during remark and in
968 * this case we should not attempt to leave / enter the STS, otherwise
969 * we'll either hit an asseert (debug / fastdebug) or deadlock
970 * (product). So we should only leave / enter the STS if we are
971 * operating concurrently.
972 *
973 * Because the thread that does the sync barrier has left the STS, it
974 * is possible to be suspended for a Full GC or an evacuation pause
975 * could occur. This is actually safe, since the entering the sync
976 * barrier is one of the last things do_marking_step() does, and it
977 * doesn't manipulate any data structures afterwards.
978 */
980 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
981 if (verbose_low()) {
982 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
983 }
985 if (concurrent()) {
986 SuspendibleThreadSet::leave();
987 }
989 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
991 if (concurrent()) {
992 SuspendibleThreadSet::join();
993 }
994 // at this point everyone should have synced up and not be doing any
995 // more work
997 if (verbose_low()) {
998 if (barrier_aborted) {
999 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1000 } else {
1001 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1002 }
1003 }
1005 if (barrier_aborted) {
1006 // If the barrier aborted we ignore the overflow condition and
1007 // just abort the whole marking phase as quickly as possible.
1008 return;
1009 }
1011 // If we're executing the concurrent phase of marking, reset the marking
1012 // state; otherwise the marking state is reset after reference processing,
1013 // during the remark pause.
1014 // If we reset here as a result of an overflow during the remark we will
1015 // see assertion failures from any subsequent set_concurrency_and_phase()
1016 // calls.
1017 if (concurrent()) {
1018 // let the task associated with with worker 0 do this
1019 if (worker_id == 0) {
1020 // task 0 is responsible for clearing the global data structures
1021 // We should be here because of an overflow. During STW we should
1022 // not clear the overflow flag since we rely on it being true when
1023 // we exit this method to abort the pause and restart concurent
1024 // marking.
1025 reset_marking_state(true /* clear_overflow */);
1026 force_overflow()->update();
1028 if (G1Log::fine()) {
1029 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1030 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1031 }
1032 }
1033 }
1035 // after this, each task should reset its own data structures then
1036 // then go into the second barrier
1037 }
1039 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1040 if (verbose_low()) {
1041 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1042 }
1044 if (concurrent()) {
1045 SuspendibleThreadSet::leave();
1046 }
1048 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1050 if (concurrent()) {
1051 SuspendibleThreadSet::join();
1052 }
1053 // at this point everything should be re-initialized and ready to go
1055 if (verbose_low()) {
1056 if (barrier_aborted) {
1057 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1058 } else {
1059 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1060 }
1061 }
1062 }
1064 #ifndef PRODUCT
1065 void ForceOverflowSettings::init() {
1066 _num_remaining = G1ConcMarkForceOverflow;
1067 _force = false;
1068 update();
1069 }
1071 void ForceOverflowSettings::update() {
1072 if (_num_remaining > 0) {
1073 _num_remaining -= 1;
1074 _force = true;
1075 } else {
1076 _force = false;
1077 }
1078 }
1080 bool ForceOverflowSettings::should_force() {
1081 if (_force) {
1082 _force = false;
1083 return true;
1084 } else {
1085 return false;
1086 }
1087 }
1088 #endif // !PRODUCT
1090 class CMConcurrentMarkingTask: public AbstractGangTask {
1091 private:
1092 ConcurrentMark* _cm;
1093 ConcurrentMarkThread* _cmt;
1095 public:
1096 void work(uint worker_id) {
1097 assert(Thread::current()->is_ConcurrentGC_thread(),
1098 "this should only be done by a conc GC thread");
1099 ResourceMark rm;
1101 double start_vtime = os::elapsedVTime();
1103 SuspendibleThreadSet::join();
1105 assert(worker_id < _cm->active_tasks(), "invariant");
1106 CMTask* the_task = _cm->task(worker_id);
1107 the_task->record_start_time();
1108 if (!_cm->has_aborted()) {
1109 do {
1110 double start_vtime_sec = os::elapsedVTime();
1111 double start_time_sec = os::elapsedTime();
1112 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1114 the_task->do_marking_step(mark_step_duration_ms,
1115 true /* do_termination */,
1116 false /* is_serial*/);
1118 double end_time_sec = os::elapsedTime();
1119 double end_vtime_sec = os::elapsedVTime();
1120 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1121 double elapsed_time_sec = end_time_sec - start_time_sec;
1122 _cm->clear_has_overflown();
1124 bool ret = _cm->do_yield_check(worker_id);
1126 jlong sleep_time_ms;
1127 if (!_cm->has_aborted() && the_task->has_aborted()) {
1128 sleep_time_ms =
1129 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1130 SuspendibleThreadSet::leave();
1131 os::sleep(Thread::current(), sleep_time_ms, false);
1132 SuspendibleThreadSet::join();
1133 }
1134 double end_time2_sec = os::elapsedTime();
1135 double elapsed_time2_sec = end_time2_sec - start_time_sec;
1137 #if 0
1138 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, "
1139 "overhead %1.4lf",
1140 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms,
1141 the_task->conc_overhead(os::elapsedTime()) * 8.0);
1142 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms",
1143 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0);
1144 #endif
1145 } while (!_cm->has_aborted() && the_task->has_aborted());
1146 }
1147 the_task->record_end_time();
1148 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1150 SuspendibleThreadSet::leave();
1152 double end_vtime = os::elapsedVTime();
1153 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1154 }
1156 CMConcurrentMarkingTask(ConcurrentMark* cm,
1157 ConcurrentMarkThread* cmt) :
1158 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1160 ~CMConcurrentMarkingTask() { }
1161 };
1163 // Calculates the number of active workers for a concurrent
1164 // phase.
1165 uint ConcurrentMark::calc_parallel_marking_threads() {
1166 if (G1CollectedHeap::use_parallel_gc_threads()) {
1167 uint n_conc_workers = 0;
1168 if (!UseDynamicNumberOfGCThreads ||
1169 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1170 !ForceDynamicNumberOfGCThreads)) {
1171 n_conc_workers = max_parallel_marking_threads();
1172 } else {
1173 n_conc_workers =
1174 AdaptiveSizePolicy::calc_default_active_workers(
1175 max_parallel_marking_threads(),
1176 1, /* Minimum workers */
1177 parallel_marking_threads(),
1178 Threads::number_of_non_daemon_threads());
1179 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1180 // that scaling has already gone into "_max_parallel_marking_threads".
1181 }
1182 assert(n_conc_workers > 0, "Always need at least 1");
1183 return n_conc_workers;
1184 }
1185 // If we are not running with any parallel GC threads we will not
1186 // have spawned any marking threads either. Hence the number of
1187 // concurrent workers should be 0.
1188 return 0;
1189 }
1191 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1192 // Currently, only survivors can be root regions.
1193 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1194 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1196 const uintx interval = PrefetchScanIntervalInBytes;
1197 HeapWord* curr = hr->bottom();
1198 const HeapWord* end = hr->top();
1199 while (curr < end) {
1200 Prefetch::read(curr, interval);
1201 oop obj = oop(curr);
1202 int size = obj->oop_iterate(&cl);
1203 assert(size == obj->size(), "sanity");
1204 curr += size;
1205 }
1206 }
1208 class CMRootRegionScanTask : public AbstractGangTask {
1209 private:
1210 ConcurrentMark* _cm;
1212 public:
1213 CMRootRegionScanTask(ConcurrentMark* cm) :
1214 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1216 void work(uint worker_id) {
1217 assert(Thread::current()->is_ConcurrentGC_thread(),
1218 "this should only be done by a conc GC thread");
1220 CMRootRegions* root_regions = _cm->root_regions();
1221 HeapRegion* hr = root_regions->claim_next();
1222 while (hr != NULL) {
1223 _cm->scanRootRegion(hr, worker_id);
1224 hr = root_regions->claim_next();
1225 }
1226 }
1227 };
1229 void ConcurrentMark::scanRootRegions() {
1230 // Start of concurrent marking.
1231 ClassLoaderDataGraph::clear_claimed_marks();
1233 // scan_in_progress() will have been set to true only if there was
1234 // at least one root region to scan. So, if it's false, we
1235 // should not attempt to do any further work.
1236 if (root_regions()->scan_in_progress()) {
1237 _parallel_marking_threads = calc_parallel_marking_threads();
1238 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1239 "Maximum number of marking threads exceeded");
1240 uint active_workers = MAX2(1U, parallel_marking_threads());
1242 CMRootRegionScanTask task(this);
1243 if (use_parallel_marking_threads()) {
1244 _parallel_workers->set_active_workers((int) active_workers);
1245 _parallel_workers->run_task(&task);
1246 } else {
1247 task.work(0);
1248 }
1250 // It's possible that has_aborted() is true here without actually
1251 // aborting the survivor scan earlier. This is OK as it's
1252 // mainly used for sanity checking.
1253 root_regions()->scan_finished();
1254 }
1255 }
1257 void ConcurrentMark::markFromRoots() {
1258 // we might be tempted to assert that:
1259 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1260 // "inconsistent argument?");
1261 // However that wouldn't be right, because it's possible that
1262 // a safepoint is indeed in progress as a younger generation
1263 // stop-the-world GC happens even as we mark in this generation.
1265 _restart_for_overflow = false;
1266 force_overflow_conc()->init();
1268 // _g1h has _n_par_threads
1269 _parallel_marking_threads = calc_parallel_marking_threads();
1270 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1271 "Maximum number of marking threads exceeded");
1273 uint active_workers = MAX2(1U, parallel_marking_threads());
1275 // Parallel task terminator is set in "set_concurrency_and_phase()"
1276 set_concurrency_and_phase(active_workers, true /* concurrent */);
1278 CMConcurrentMarkingTask markingTask(this, cmThread());
1279 if (use_parallel_marking_threads()) {
1280 _parallel_workers->set_active_workers((int)active_workers);
1281 // Don't set _n_par_threads because it affects MT in process_roots()
1282 // and the decisions on that MT processing is made elsewhere.
1283 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1284 _parallel_workers->run_task(&markingTask);
1285 } else {
1286 markingTask.work(0);
1287 }
1288 print_stats();
1289 }
1291 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1292 // world is stopped at this checkpoint
1293 assert(SafepointSynchronize::is_at_safepoint(),
1294 "world should be stopped");
1296 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1298 // If a full collection has happened, we shouldn't do this.
1299 if (has_aborted()) {
1300 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1301 return;
1302 }
1304 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1306 if (VerifyDuringGC) {
1307 HandleMark hm; // handle scope
1308 Universe::heap()->prepare_for_verify();
1309 Universe::verify(VerifyOption_G1UsePrevMarking,
1310 " VerifyDuringGC:(before)");
1311 }
1312 g1h->check_bitmaps("Remark Start");
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 g1h->check_bitmaps("Remark End");
1362 assert(!restart_for_overflow(), "sanity");
1363 // Completely reset the marking state since marking completed
1364 set_non_marking_state();
1365 }
1367 // Expand the marking stack, if we have to and if we can.
1368 if (_markStack.should_expand()) {
1369 _markStack.expand();
1370 }
1372 // Statistics
1373 double now = os::elapsedTime();
1374 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1375 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1376 _remark_times.add((now - start) * 1000.0);
1378 g1p->record_concurrent_mark_remark_end();
1380 G1CMIsAliveClosure is_alive(g1h);
1381 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1382 }
1384 // Base class of the closures that finalize and verify the
1385 // liveness counting data.
1386 class CMCountDataClosureBase: public HeapRegionClosure {
1387 protected:
1388 G1CollectedHeap* _g1h;
1389 ConcurrentMark* _cm;
1390 CardTableModRefBS* _ct_bs;
1392 BitMap* _region_bm;
1393 BitMap* _card_bm;
1395 // Takes a region that's not empty (i.e., it has at least one
1396 // live object in it and sets its corresponding bit on the region
1397 // bitmap to 1. If the region is "starts humongous" it will also set
1398 // to 1 the bits on the region bitmap that correspond to its
1399 // associated "continues humongous" regions.
1400 void set_bit_for_region(HeapRegion* hr) {
1401 assert(!hr->continuesHumongous(), "should have filtered those out");
1403 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1404 if (!hr->startsHumongous()) {
1405 // Normal (non-humongous) case: just set the bit.
1406 _region_bm->par_at_put(index, true);
1407 } else {
1408 // Starts humongous case: calculate how many regions are part of
1409 // this humongous region and then set the bit range.
1410 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1411 _region_bm->par_at_put_range(index, end_index, true);
1412 }
1413 }
1415 public:
1416 CMCountDataClosureBase(G1CollectedHeap* g1h,
1417 BitMap* region_bm, BitMap* card_bm):
1418 _g1h(g1h), _cm(g1h->concurrent_mark()),
1419 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1420 _region_bm(region_bm), _card_bm(card_bm) { }
1421 };
1423 // Closure that calculates the # live objects per region. Used
1424 // for verification purposes during the cleanup pause.
1425 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1426 CMBitMapRO* _bm;
1427 size_t _region_marked_bytes;
1429 public:
1430 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1431 BitMap* region_bm, BitMap* card_bm) :
1432 CMCountDataClosureBase(g1h, region_bm, card_bm),
1433 _bm(bm), _region_marked_bytes(0) { }
1435 bool doHeapRegion(HeapRegion* hr) {
1437 if (hr->continuesHumongous()) {
1438 // We will ignore these here and process them when their
1439 // associated "starts humongous" region is processed (see
1440 // set_bit_for_heap_region()). Note that we cannot rely on their
1441 // associated "starts humongous" region to have their bit set to
1442 // 1 since, due to the region chunking in the parallel region
1443 // iteration, a "continues humongous" region might be visited
1444 // before its associated "starts humongous".
1445 return false;
1446 }
1448 HeapWord* ntams = hr->next_top_at_mark_start();
1449 HeapWord* start = hr->bottom();
1451 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1452 err_msg("Preconditions not met - "
1453 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1454 p2i(start), p2i(ntams), p2i(hr->end())));
1456 // Find the first marked object at or after "start".
1457 start = _bm->getNextMarkedWordAddress(start, ntams);
1459 size_t marked_bytes = 0;
1461 while (start < ntams) {
1462 oop obj = oop(start);
1463 int obj_sz = obj->size();
1464 HeapWord* obj_end = start + obj_sz;
1466 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1467 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1469 // Note: if we're looking at the last region in heap - obj_end
1470 // could be actually just beyond the end of the heap; end_idx
1471 // will then correspond to a (non-existent) card that is also
1472 // just beyond the heap.
1473 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1474 // end of object is not card aligned - increment to cover
1475 // all the cards spanned by the object
1476 end_idx += 1;
1477 }
1479 // Set the bits in the card BM for the cards spanned by this object.
1480 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1482 // Add the size of this object to the number of marked bytes.
1483 marked_bytes += (size_t)obj_sz * HeapWordSize;
1485 // Find the next marked object after this one.
1486 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1487 }
1489 // Mark the allocated-since-marking portion...
1490 HeapWord* top = hr->top();
1491 if (ntams < top) {
1492 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1493 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1495 // Note: if we're looking at the last region in heap - top
1496 // could be actually just beyond the end of the heap; end_idx
1497 // will then correspond to a (non-existent) card that is also
1498 // just beyond the heap.
1499 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1500 // end of object is not card aligned - increment to cover
1501 // all the cards spanned by the object
1502 end_idx += 1;
1503 }
1504 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1506 // This definitely means the region has live objects.
1507 set_bit_for_region(hr);
1508 }
1510 // Update the live region bitmap.
1511 if (marked_bytes > 0) {
1512 set_bit_for_region(hr);
1513 }
1515 // Set the marked bytes for the current region so that
1516 // it can be queried by a calling verificiation routine
1517 _region_marked_bytes = marked_bytes;
1519 return false;
1520 }
1522 size_t region_marked_bytes() const { return _region_marked_bytes; }
1523 };
1525 // Heap region closure used for verifying the counting data
1526 // that was accumulated concurrently and aggregated during
1527 // the remark pause. This closure is applied to the heap
1528 // regions during the STW cleanup pause.
1530 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1531 G1CollectedHeap* _g1h;
1532 ConcurrentMark* _cm;
1533 CalcLiveObjectsClosure _calc_cl;
1534 BitMap* _region_bm; // Region BM to be verified
1535 BitMap* _card_bm; // Card BM to be verified
1536 bool _verbose; // verbose output?
1538 BitMap* _exp_region_bm; // Expected Region BM values
1539 BitMap* _exp_card_bm; // Expected card BM values
1541 int _failures;
1543 public:
1544 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1545 BitMap* region_bm,
1546 BitMap* card_bm,
1547 BitMap* exp_region_bm,
1548 BitMap* exp_card_bm,
1549 bool verbose) :
1550 _g1h(g1h), _cm(g1h->concurrent_mark()),
1551 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1552 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1553 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1554 _failures(0) { }
1556 int failures() const { return _failures; }
1558 bool doHeapRegion(HeapRegion* hr) {
1559 if (hr->continuesHumongous()) {
1560 // We will ignore these here and process them when their
1561 // associated "starts humongous" region is processed (see
1562 // set_bit_for_heap_region()). Note that we cannot rely on their
1563 // associated "starts humongous" region to have their bit set to
1564 // 1 since, due to the region chunking in the parallel region
1565 // iteration, a "continues humongous" region might be visited
1566 // before its associated "starts humongous".
1567 return false;
1568 }
1570 int failures = 0;
1572 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1573 // this region and set the corresponding bits in the expected region
1574 // and card bitmaps.
1575 bool res = _calc_cl.doHeapRegion(hr);
1576 assert(res == false, "should be continuing");
1578 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1579 Mutex::_no_safepoint_check_flag);
1581 // Verify the marked bytes for this region.
1582 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1583 size_t act_marked_bytes = hr->next_marked_bytes();
1585 // We're not OK if expected marked bytes > actual marked bytes. It means
1586 // we have missed accounting some objects during the actual marking.
1587 if (exp_marked_bytes > act_marked_bytes) {
1588 if (_verbose) {
1589 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1590 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1591 hr->hrs_index(), exp_marked_bytes, act_marked_bytes);
1592 }
1593 failures += 1;
1594 }
1596 // Verify the bit, for this region, in the actual and expected
1597 // (which was just calculated) region bit maps.
1598 // We're not OK if the bit in the calculated expected region
1599 // bitmap is set and the bit in the actual region bitmap is not.
1600 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1602 bool expected = _exp_region_bm->at(index);
1603 bool actual = _region_bm->at(index);
1604 if (expected && !actual) {
1605 if (_verbose) {
1606 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1607 "expected: %s, actual: %s",
1608 hr->hrs_index(),
1609 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1610 }
1611 failures += 1;
1612 }
1614 // Verify that the card bit maps for the cards spanned by the current
1615 // region match. We have an error if we have a set bit in the expected
1616 // bit map and the corresponding bit in the actual bitmap is not set.
1618 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1619 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1621 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1622 expected = _exp_card_bm->at(i);
1623 actual = _card_bm->at(i);
1625 if (expected && !actual) {
1626 if (_verbose) {
1627 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1628 "expected: %s, actual: %s",
1629 hr->hrs_index(), i,
1630 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1631 }
1632 failures += 1;
1633 }
1634 }
1636 if (failures > 0 && _verbose) {
1637 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1638 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1639 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1640 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1641 }
1643 _failures += failures;
1645 // We could stop iteration over the heap when we
1646 // find the first violating region by returning true.
1647 return false;
1648 }
1649 };
1651 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1652 protected:
1653 G1CollectedHeap* _g1h;
1654 ConcurrentMark* _cm;
1655 BitMap* _actual_region_bm;
1656 BitMap* _actual_card_bm;
1658 uint _n_workers;
1660 BitMap* _expected_region_bm;
1661 BitMap* _expected_card_bm;
1663 int _failures;
1664 bool _verbose;
1666 public:
1667 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1668 BitMap* region_bm, BitMap* card_bm,
1669 BitMap* expected_region_bm, BitMap* expected_card_bm)
1670 : AbstractGangTask("G1 verify final counting"),
1671 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1672 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1673 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1674 _failures(0), _verbose(false),
1675 _n_workers(0) {
1676 assert(VerifyDuringGC, "don't call this otherwise");
1678 // Use the value already set as the number of active threads
1679 // in the call to run_task().
1680 if (G1CollectedHeap::use_parallel_gc_threads()) {
1681 assert( _g1h->workers()->active_workers() > 0,
1682 "Should have been previously set");
1683 _n_workers = _g1h->workers()->active_workers();
1684 } else {
1685 _n_workers = 1;
1686 }
1688 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1689 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1691 _verbose = _cm->verbose_medium();
1692 }
1694 void work(uint worker_id) {
1695 assert(worker_id < _n_workers, "invariant");
1697 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1698 _actual_region_bm, _actual_card_bm,
1699 _expected_region_bm,
1700 _expected_card_bm,
1701 _verbose);
1703 if (G1CollectedHeap::use_parallel_gc_threads()) {
1704 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1705 worker_id,
1706 _n_workers,
1707 HeapRegion::VerifyCountClaimValue);
1708 } else {
1709 _g1h->heap_region_iterate(&verify_cl);
1710 }
1712 Atomic::add(verify_cl.failures(), &_failures);
1713 }
1715 int failures() const { return _failures; }
1716 };
1718 // Closure that finalizes the liveness counting data.
1719 // Used during the cleanup pause.
1720 // Sets the bits corresponding to the interval [NTAMS, top]
1721 // (which contains the implicitly live objects) in the
1722 // card liveness bitmap. Also sets the bit for each region,
1723 // containing live data, in the region liveness bitmap.
1725 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1726 public:
1727 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1728 BitMap* region_bm,
1729 BitMap* card_bm) :
1730 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1732 bool doHeapRegion(HeapRegion* hr) {
1734 if (hr->continuesHumongous()) {
1735 // We will ignore these here and process them when their
1736 // associated "starts humongous" region is processed (see
1737 // set_bit_for_heap_region()). Note that we cannot rely on their
1738 // associated "starts humongous" region to have their bit set to
1739 // 1 since, due to the region chunking in the parallel region
1740 // iteration, a "continues humongous" region might be visited
1741 // before its associated "starts humongous".
1742 return false;
1743 }
1745 HeapWord* ntams = hr->next_top_at_mark_start();
1746 HeapWord* top = hr->top();
1748 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1750 // Mark the allocated-since-marking portion...
1751 if (ntams < top) {
1752 // This definitely means the region has live objects.
1753 set_bit_for_region(hr);
1755 // Now set the bits in the card bitmap for [ntams, top)
1756 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1757 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1759 // Note: if we're looking at the last region in heap - top
1760 // could be actually just beyond the end of the heap; end_idx
1761 // will then correspond to a (non-existent) card that is also
1762 // just beyond the heap.
1763 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1764 // end of object is not card aligned - increment to cover
1765 // all the cards spanned by the object
1766 end_idx += 1;
1767 }
1769 assert(end_idx <= _card_bm->size(),
1770 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1771 end_idx, _card_bm->size()));
1772 assert(start_idx < _card_bm->size(),
1773 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1774 start_idx, _card_bm->size()));
1776 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1777 }
1779 // Set the bit for the region if it contains live data
1780 if (hr->next_marked_bytes() > 0) {
1781 set_bit_for_region(hr);
1782 }
1784 return false;
1785 }
1786 };
1788 class G1ParFinalCountTask: public AbstractGangTask {
1789 protected:
1790 G1CollectedHeap* _g1h;
1791 ConcurrentMark* _cm;
1792 BitMap* _actual_region_bm;
1793 BitMap* _actual_card_bm;
1795 uint _n_workers;
1797 public:
1798 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1799 : AbstractGangTask("G1 final counting"),
1800 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1801 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1802 _n_workers(0) {
1803 // Use the value already set as the number of active threads
1804 // in the call to run_task().
1805 if (G1CollectedHeap::use_parallel_gc_threads()) {
1806 assert( _g1h->workers()->active_workers() > 0,
1807 "Should have been previously set");
1808 _n_workers = _g1h->workers()->active_workers();
1809 } else {
1810 _n_workers = 1;
1811 }
1812 }
1814 void work(uint worker_id) {
1815 assert(worker_id < _n_workers, "invariant");
1817 FinalCountDataUpdateClosure final_update_cl(_g1h,
1818 _actual_region_bm,
1819 _actual_card_bm);
1821 if (G1CollectedHeap::use_parallel_gc_threads()) {
1822 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1823 worker_id,
1824 _n_workers,
1825 HeapRegion::FinalCountClaimValue);
1826 } else {
1827 _g1h->heap_region_iterate(&final_update_cl);
1828 }
1829 }
1830 };
1832 class G1ParNoteEndTask;
1834 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1835 G1CollectedHeap* _g1;
1836 size_t _max_live_bytes;
1837 uint _regions_claimed;
1838 size_t _freed_bytes;
1839 FreeRegionList* _local_cleanup_list;
1840 HeapRegionSetCount _old_regions_removed;
1841 HeapRegionSetCount _humongous_regions_removed;
1842 HRRSCleanupTask* _hrrs_cleanup_task;
1843 double _claimed_region_time;
1844 double _max_region_time;
1846 public:
1847 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1848 FreeRegionList* local_cleanup_list,
1849 HRRSCleanupTask* hrrs_cleanup_task) :
1850 _g1(g1),
1851 _max_live_bytes(0), _regions_claimed(0),
1852 _freed_bytes(0),
1853 _claimed_region_time(0.0), _max_region_time(0.0),
1854 _local_cleanup_list(local_cleanup_list),
1855 _old_regions_removed(),
1856 _humongous_regions_removed(),
1857 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1859 size_t freed_bytes() { return _freed_bytes; }
1860 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1861 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1863 bool doHeapRegion(HeapRegion *hr) {
1864 if (hr->continuesHumongous()) {
1865 return false;
1866 }
1867 // We use a claim value of zero here because all regions
1868 // were claimed with value 1 in the FinalCount task.
1869 _g1->reset_gc_time_stamps(hr);
1870 double start = os::elapsedTime();
1871 _regions_claimed++;
1872 hr->note_end_of_marking();
1873 _max_live_bytes += hr->max_live_bytes();
1875 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1876 _freed_bytes += hr->used();
1877 hr->set_containing_set(NULL);
1878 if (hr->isHumongous()) {
1879 assert(hr->startsHumongous(), "we should only see starts humongous");
1880 _humongous_regions_removed.increment(1u, hr->capacity());
1881 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1882 } else {
1883 _old_regions_removed.increment(1u, hr->capacity());
1884 _g1->free_region(hr, _local_cleanup_list, true);
1885 }
1886 } else {
1887 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1888 }
1890 double region_time = (os::elapsedTime() - start);
1891 _claimed_region_time += region_time;
1892 if (region_time > _max_region_time) {
1893 _max_region_time = region_time;
1894 }
1895 return false;
1896 }
1898 size_t max_live_bytes() { return _max_live_bytes; }
1899 uint regions_claimed() { return _regions_claimed; }
1900 double claimed_region_time_sec() { return _claimed_region_time; }
1901 double max_region_time_sec() { return _max_region_time; }
1902 };
1904 class G1ParNoteEndTask: public AbstractGangTask {
1905 friend class G1NoteEndOfConcMarkClosure;
1907 protected:
1908 G1CollectedHeap* _g1h;
1909 size_t _max_live_bytes;
1910 size_t _freed_bytes;
1911 FreeRegionList* _cleanup_list;
1913 public:
1914 G1ParNoteEndTask(G1CollectedHeap* g1h,
1915 FreeRegionList* cleanup_list) :
1916 AbstractGangTask("G1 note end"), _g1h(g1h),
1917 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1919 void work(uint worker_id) {
1920 double start = os::elapsedTime();
1921 FreeRegionList local_cleanup_list("Local Cleanup List");
1922 HRRSCleanupTask hrrs_cleanup_task;
1923 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1924 &hrrs_cleanup_task);
1925 if (G1CollectedHeap::use_parallel_gc_threads()) {
1926 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1927 _g1h->workers()->active_workers(),
1928 HeapRegion::NoteEndClaimValue);
1929 } else {
1930 _g1h->heap_region_iterate(&g1_note_end);
1931 }
1932 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1934 // Now update the lists
1935 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1936 {
1937 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1938 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1939 _max_live_bytes += g1_note_end.max_live_bytes();
1940 _freed_bytes += g1_note_end.freed_bytes();
1942 // If we iterate over the global cleanup list at the end of
1943 // cleanup to do this printing we will not guarantee to only
1944 // generate output for the newly-reclaimed regions (the list
1945 // might not be empty at the beginning of cleanup; we might
1946 // still be working on its previous contents). So we do the
1947 // printing here, before we append the new regions to the global
1948 // cleanup list.
1950 G1HRPrinter* hr_printer = _g1h->hr_printer();
1951 if (hr_printer->is_active()) {
1952 FreeRegionListIterator iter(&local_cleanup_list);
1953 while (iter.more_available()) {
1954 HeapRegion* hr = iter.get_next();
1955 hr_printer->cleanup(hr);
1956 }
1957 }
1959 _cleanup_list->add_ordered(&local_cleanup_list);
1960 assert(local_cleanup_list.is_empty(), "post-condition");
1962 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1963 }
1964 }
1965 size_t max_live_bytes() { return _max_live_bytes; }
1966 size_t freed_bytes() { return _freed_bytes; }
1967 };
1969 class G1ParScrubRemSetTask: public AbstractGangTask {
1970 protected:
1971 G1RemSet* _g1rs;
1972 BitMap* _region_bm;
1973 BitMap* _card_bm;
1974 public:
1975 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1976 BitMap* region_bm, BitMap* card_bm) :
1977 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1978 _region_bm(region_bm), _card_bm(card_bm) { }
1980 void work(uint worker_id) {
1981 if (G1CollectedHeap::use_parallel_gc_threads()) {
1982 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1983 HeapRegion::ScrubRemSetClaimValue);
1984 } else {
1985 _g1rs->scrub(_region_bm, _card_bm);
1986 }
1987 }
1989 };
1991 void ConcurrentMark::cleanup() {
1992 // world is stopped at this checkpoint
1993 assert(SafepointSynchronize::is_at_safepoint(),
1994 "world should be stopped");
1995 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1997 // If a full collection has happened, we shouldn't do this.
1998 if (has_aborted()) {
1999 g1h->set_marking_complete(); // So bitmap clearing isn't confused
2000 return;
2001 }
2003 g1h->verify_region_sets_optional();
2005 if (VerifyDuringGC) {
2006 HandleMark hm; // handle scope
2007 Universe::heap()->prepare_for_verify();
2008 Universe::verify(VerifyOption_G1UsePrevMarking,
2009 " VerifyDuringGC:(before)");
2010 }
2011 g1h->check_bitmaps("Cleanup Start");
2013 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
2014 g1p->record_concurrent_mark_cleanup_start();
2016 double start = os::elapsedTime();
2018 HeapRegionRemSet::reset_for_cleanup_tasks();
2020 uint n_workers;
2022 // Do counting once more with the world stopped for good measure.
2023 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2025 if (G1CollectedHeap::use_parallel_gc_threads()) {
2026 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2027 "sanity check");
2029 g1h->set_par_threads();
2030 n_workers = g1h->n_par_threads();
2031 assert(g1h->n_par_threads() == n_workers,
2032 "Should not have been reset");
2033 g1h->workers()->run_task(&g1_par_count_task);
2034 // Done with the parallel phase so reset to 0.
2035 g1h->set_par_threads(0);
2037 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2038 "sanity check");
2039 } else {
2040 n_workers = 1;
2041 g1_par_count_task.work(0);
2042 }
2044 if (VerifyDuringGC) {
2045 // Verify that the counting data accumulated during marking matches
2046 // that calculated by walking the marking bitmap.
2048 // Bitmaps to hold expected values
2049 BitMap expected_region_bm(_region_bm.size(), true);
2050 BitMap expected_card_bm(_card_bm.size(), true);
2052 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2053 &_region_bm,
2054 &_card_bm,
2055 &expected_region_bm,
2056 &expected_card_bm);
2058 if (G1CollectedHeap::use_parallel_gc_threads()) {
2059 g1h->set_par_threads((int)n_workers);
2060 g1h->workers()->run_task(&g1_par_verify_task);
2061 // Done with the parallel phase so reset to 0.
2062 g1h->set_par_threads(0);
2064 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2065 "sanity check");
2066 } else {
2067 g1_par_verify_task.work(0);
2068 }
2070 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2071 }
2073 size_t start_used_bytes = g1h->used();
2074 g1h->set_marking_complete();
2076 double count_end = os::elapsedTime();
2077 double this_final_counting_time = (count_end - start);
2078 _total_counting_time += this_final_counting_time;
2080 if (G1PrintRegionLivenessInfo) {
2081 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2082 _g1h->heap_region_iterate(&cl);
2083 }
2085 // Install newly created mark bitMap as "prev".
2086 swapMarkBitMaps();
2088 g1h->reset_gc_time_stamp();
2090 // Note end of marking in all heap regions.
2091 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2092 if (G1CollectedHeap::use_parallel_gc_threads()) {
2093 g1h->set_par_threads((int)n_workers);
2094 g1h->workers()->run_task(&g1_par_note_end_task);
2095 g1h->set_par_threads(0);
2097 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2098 "sanity check");
2099 } else {
2100 g1_par_note_end_task.work(0);
2101 }
2102 g1h->check_gc_time_stamps();
2104 if (!cleanup_list_is_empty()) {
2105 // The cleanup list is not empty, so we'll have to process it
2106 // concurrently. Notify anyone else that might be wanting free
2107 // regions that there will be more free regions coming soon.
2108 g1h->set_free_regions_coming();
2109 }
2111 // call below, since it affects the metric by which we sort the heap
2112 // regions.
2113 if (G1ScrubRemSets) {
2114 double rs_scrub_start = os::elapsedTime();
2115 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2116 if (G1CollectedHeap::use_parallel_gc_threads()) {
2117 g1h->set_par_threads((int)n_workers);
2118 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2119 g1h->set_par_threads(0);
2121 assert(g1h->check_heap_region_claim_values(
2122 HeapRegion::ScrubRemSetClaimValue),
2123 "sanity check");
2124 } else {
2125 g1_par_scrub_rs_task.work(0);
2126 }
2128 double rs_scrub_end = os::elapsedTime();
2129 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2130 _total_rs_scrub_time += this_rs_scrub_time;
2131 }
2133 // this will also free any regions totally full of garbage objects,
2134 // and sort the regions.
2135 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2137 // Statistics.
2138 double end = os::elapsedTime();
2139 _cleanup_times.add((end - start) * 1000.0);
2141 if (G1Log::fine()) {
2142 g1h->print_size_transition(gclog_or_tty,
2143 start_used_bytes,
2144 g1h->used(),
2145 g1h->capacity());
2146 }
2148 // Clean up will have freed any regions completely full of garbage.
2149 // Update the soft reference policy with the new heap occupancy.
2150 Universe::update_heap_info_at_gc();
2152 if (VerifyDuringGC) {
2153 HandleMark hm; // handle scope
2154 Universe::heap()->prepare_for_verify();
2155 Universe::verify(VerifyOption_G1UsePrevMarking,
2156 " VerifyDuringGC:(after)");
2157 }
2158 g1h->check_bitmaps("Cleanup End");
2160 g1h->verify_region_sets_optional();
2162 // We need to make this be a "collection" so any collection pause that
2163 // races with it goes around and waits for completeCleanup to finish.
2164 g1h->increment_total_collections();
2166 // Clean out dead classes and update Metaspace sizes.
2167 if (ClassUnloadingWithConcurrentMark) {
2168 ClassLoaderDataGraph::purge();
2169 }
2170 MetaspaceGC::compute_new_size();
2172 // We reclaimed old regions so we should calculate the sizes to make
2173 // sure we update the old gen/space data.
2174 g1h->g1mm()->update_sizes();
2176 g1h->trace_heap_after_concurrent_cycle();
2177 }
2179 void ConcurrentMark::completeCleanup() {
2180 if (has_aborted()) return;
2182 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2184 _cleanup_list.verify_optional();
2185 FreeRegionList tmp_free_list("Tmp Free List");
2187 if (G1ConcRegionFreeingVerbose) {
2188 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2189 "cleanup list has %u entries",
2190 _cleanup_list.length());
2191 }
2193 // Noone else should be accessing the _cleanup_list at this point,
2194 // so it's not necessary to take any locks
2195 while (!_cleanup_list.is_empty()) {
2196 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2197 assert(hr != NULL, "Got NULL from a non-empty list");
2198 hr->par_clear();
2199 tmp_free_list.add_ordered(hr);
2201 // Instead of adding one region at a time to the secondary_free_list,
2202 // we accumulate them in the local list and move them a few at a
2203 // time. This also cuts down on the number of notify_all() calls
2204 // we do during this process. We'll also append the local list when
2205 // _cleanup_list is empty (which means we just removed the last
2206 // region from the _cleanup_list).
2207 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2208 _cleanup_list.is_empty()) {
2209 if (G1ConcRegionFreeingVerbose) {
2210 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2211 "appending %u entries to the secondary_free_list, "
2212 "cleanup list still has %u entries",
2213 tmp_free_list.length(),
2214 _cleanup_list.length());
2215 }
2217 {
2218 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2219 g1h->secondary_free_list_add(&tmp_free_list);
2220 SecondaryFreeList_lock->notify_all();
2221 }
2223 if (G1StressConcRegionFreeing) {
2224 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2225 os::sleep(Thread::current(), (jlong) 1, false);
2226 }
2227 }
2228 }
2229 }
2230 assert(tmp_free_list.is_empty(), "post-condition");
2231 }
2233 // Supporting Object and Oop closures for reference discovery
2234 // and processing in during marking
2236 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2237 HeapWord* addr = (HeapWord*)obj;
2238 return addr != NULL &&
2239 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2240 }
2242 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2243 // Uses the CMTask associated with a worker thread (for serial reference
2244 // processing the CMTask for worker 0 is used) to preserve (mark) and
2245 // trace referent objects.
2246 //
2247 // Using the CMTask and embedded local queues avoids having the worker
2248 // threads operating on the global mark stack. This reduces the risk
2249 // of overflowing the stack - which we would rather avoid at this late
2250 // state. Also using the tasks' local queues removes the potential
2251 // of the workers interfering with each other that could occur if
2252 // operating on the global stack.
2254 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2255 ConcurrentMark* _cm;
2256 CMTask* _task;
2257 int _ref_counter_limit;
2258 int _ref_counter;
2259 bool _is_serial;
2260 public:
2261 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2262 _cm(cm), _task(task), _is_serial(is_serial),
2263 _ref_counter_limit(G1RefProcDrainInterval) {
2264 assert(_ref_counter_limit > 0, "sanity");
2265 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2266 _ref_counter = _ref_counter_limit;
2267 }
2269 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2270 virtual void do_oop( oop* p) { do_oop_work(p); }
2272 template <class T> void do_oop_work(T* p) {
2273 if (!_cm->has_overflown()) {
2274 oop obj = oopDesc::load_decode_heap_oop(p);
2275 if (_cm->verbose_high()) {
2276 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2277 "*"PTR_FORMAT" = "PTR_FORMAT,
2278 _task->worker_id(), p2i(p), p2i((void*) obj));
2279 }
2281 _task->deal_with_reference(obj);
2282 _ref_counter--;
2284 if (_ref_counter == 0) {
2285 // We have dealt with _ref_counter_limit references, pushing them
2286 // and objects reachable from them on to the local stack (and
2287 // possibly the global stack). Call CMTask::do_marking_step() to
2288 // process these entries.
2289 //
2290 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2291 // there's nothing more to do (i.e. we're done with the entries that
2292 // were pushed as a result of the CMTask::deal_with_reference() calls
2293 // above) or we overflow.
2294 //
2295 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2296 // flag while there may still be some work to do. (See the comment at
2297 // the beginning of CMTask::do_marking_step() for those conditions -
2298 // one of which is reaching the specified time target.) It is only
2299 // when CMTask::do_marking_step() returns without setting the
2300 // has_aborted() flag that the marking step has completed.
2301 do {
2302 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2303 _task->do_marking_step(mark_step_duration_ms,
2304 false /* do_termination */,
2305 _is_serial);
2306 } while (_task->has_aborted() && !_cm->has_overflown());
2307 _ref_counter = _ref_counter_limit;
2308 }
2309 } else {
2310 if (_cm->verbose_high()) {
2311 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2312 }
2313 }
2314 }
2315 };
2317 // 'Drain' oop closure used by both serial and parallel reference processing.
2318 // Uses the CMTask associated with a given worker thread (for serial
2319 // reference processing the CMtask for worker 0 is used). Calls the
2320 // do_marking_step routine, with an unbelievably large timeout value,
2321 // to drain the marking data structures of the remaining entries
2322 // added by the 'keep alive' oop closure above.
2324 class G1CMDrainMarkingStackClosure: public VoidClosure {
2325 ConcurrentMark* _cm;
2326 CMTask* _task;
2327 bool _is_serial;
2328 public:
2329 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2330 _cm(cm), _task(task), _is_serial(is_serial) {
2331 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2332 }
2334 void do_void() {
2335 do {
2336 if (_cm->verbose_high()) {
2337 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2338 _task->worker_id(), BOOL_TO_STR(_is_serial));
2339 }
2341 // We call CMTask::do_marking_step() to completely drain the local
2342 // and global marking stacks of entries pushed by the 'keep alive'
2343 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2344 //
2345 // CMTask::do_marking_step() is called in a loop, which we'll exit
2346 // if there's nothing more to do (i.e. we'completely drained the
2347 // entries that were pushed as a a result of applying the 'keep alive'
2348 // closure to the entries on the discovered ref lists) or we overflow
2349 // the global marking stack.
2350 //
2351 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2352 // flag while there may still be some work to do. (See the comment at
2353 // the beginning of CMTask::do_marking_step() for those conditions -
2354 // one of which is reaching the specified time target.) It is only
2355 // when CMTask::do_marking_step() returns without setting the
2356 // has_aborted() flag that the marking step has completed.
2358 _task->do_marking_step(1000000000.0 /* something very large */,
2359 true /* do_termination */,
2360 _is_serial);
2361 } while (_task->has_aborted() && !_cm->has_overflown());
2362 }
2363 };
2365 // Implementation of AbstractRefProcTaskExecutor for parallel
2366 // reference processing at the end of G1 concurrent marking
2368 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2369 private:
2370 G1CollectedHeap* _g1h;
2371 ConcurrentMark* _cm;
2372 WorkGang* _workers;
2373 int _active_workers;
2375 public:
2376 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2377 ConcurrentMark* cm,
2378 WorkGang* workers,
2379 int n_workers) :
2380 _g1h(g1h), _cm(cm),
2381 _workers(workers), _active_workers(n_workers) { }
2383 // Executes the given task using concurrent marking worker threads.
2384 virtual void execute(ProcessTask& task);
2385 virtual void execute(EnqueueTask& task);
2386 };
2388 class G1CMRefProcTaskProxy: public AbstractGangTask {
2389 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2390 ProcessTask& _proc_task;
2391 G1CollectedHeap* _g1h;
2392 ConcurrentMark* _cm;
2394 public:
2395 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2396 G1CollectedHeap* g1h,
2397 ConcurrentMark* cm) :
2398 AbstractGangTask("Process reference objects in parallel"),
2399 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2400 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2401 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2402 }
2404 virtual void work(uint worker_id) {
2405 ResourceMark rm;
2406 HandleMark hm;
2407 CMTask* task = _cm->task(worker_id);
2408 G1CMIsAliveClosure g1_is_alive(_g1h);
2409 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2410 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2412 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2413 }
2414 };
2416 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2417 assert(_workers != NULL, "Need parallel worker threads.");
2418 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2420 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2422 // We need to reset the concurrency level before each
2423 // proxy task execution, so that the termination protocol
2424 // and overflow handling in CMTask::do_marking_step() knows
2425 // how many workers to wait for.
2426 _cm->set_concurrency(_active_workers);
2427 _g1h->set_par_threads(_active_workers);
2428 _workers->run_task(&proc_task_proxy);
2429 _g1h->set_par_threads(0);
2430 }
2432 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2433 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2434 EnqueueTask& _enq_task;
2436 public:
2437 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2438 AbstractGangTask("Enqueue reference objects in parallel"),
2439 _enq_task(enq_task) { }
2441 virtual void work(uint worker_id) {
2442 _enq_task.work(worker_id);
2443 }
2444 };
2446 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2447 assert(_workers != NULL, "Need parallel worker threads.");
2448 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2450 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2452 // Not strictly necessary but...
2453 //
2454 // We need to reset the concurrency level before each
2455 // proxy task execution, so that the termination protocol
2456 // and overflow handling in CMTask::do_marking_step() knows
2457 // how many workers to wait for.
2458 _cm->set_concurrency(_active_workers);
2459 _g1h->set_par_threads(_active_workers);
2460 _workers->run_task(&enq_task_proxy);
2461 _g1h->set_par_threads(0);
2462 }
2464 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2465 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2466 }
2468 // Helper class to get rid of some boilerplate code.
2469 class G1RemarkGCTraceTime : public GCTraceTime {
2470 static bool doit_and_prepend(bool doit) {
2471 if (doit) {
2472 gclog_or_tty->put(' ');
2473 }
2474 return doit;
2475 }
2477 public:
2478 G1RemarkGCTraceTime(const char* title, bool doit)
2479 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
2480 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
2481 }
2482 };
2484 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2485 if (has_overflown()) {
2486 // Skip processing the discovered references if we have
2487 // overflown the global marking stack. Reference objects
2488 // only get discovered once so it is OK to not
2489 // de-populate the discovered reference lists. We could have,
2490 // but the only benefit would be that, when marking restarts,
2491 // less reference objects are discovered.
2492 return;
2493 }
2495 ResourceMark rm;
2496 HandleMark hm;
2498 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2500 // Is alive closure.
2501 G1CMIsAliveClosure g1_is_alive(g1h);
2503 // Inner scope to exclude the cleaning of the string and symbol
2504 // tables from the displayed time.
2505 {
2506 if (G1Log::finer()) {
2507 gclog_or_tty->put(' ');
2508 }
2509 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm(), concurrent_gc_id());
2511 ReferenceProcessor* rp = g1h->ref_processor_cm();
2513 // See the comment in G1CollectedHeap::ref_processing_init()
2514 // about how reference processing currently works in G1.
2516 // Set the soft reference policy
2517 rp->setup_policy(clear_all_soft_refs);
2518 assert(_markStack.isEmpty(), "mark stack should be empty");
2520 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2521 // in serial reference processing. Note these closures are also
2522 // used for serially processing (by the the current thread) the
2523 // JNI references during parallel reference processing.
2524 //
2525 // These closures do not need to synchronize with the worker
2526 // threads involved in parallel reference processing as these
2527 // instances are executed serially by the current thread (e.g.
2528 // reference processing is not multi-threaded and is thus
2529 // performed by the current thread instead of a gang worker).
2530 //
2531 // The gang tasks involved in parallel reference procssing create
2532 // their own instances of these closures, which do their own
2533 // synchronization among themselves.
2534 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2535 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2537 // We need at least one active thread. If reference processing
2538 // is not multi-threaded we use the current (VMThread) thread,
2539 // otherwise we use the work gang from the G1CollectedHeap and
2540 // we utilize all the worker threads we can.
2541 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2542 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2543 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2545 // Parallel processing task executor.
2546 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2547 g1h->workers(), active_workers);
2548 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2550 // Set the concurrency level. The phase was already set prior to
2551 // executing the remark task.
2552 set_concurrency(active_workers);
2554 // Set the degree of MT processing here. If the discovery was done MT,
2555 // the number of threads involved during discovery could differ from
2556 // the number of active workers. This is OK as long as the discovered
2557 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2558 rp->set_active_mt_degree(active_workers);
2560 // Process the weak references.
2561 const ReferenceProcessorStats& stats =
2562 rp->process_discovered_references(&g1_is_alive,
2563 &g1_keep_alive,
2564 &g1_drain_mark_stack,
2565 executor,
2566 g1h->gc_timer_cm(),
2567 concurrent_gc_id());
2568 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2570 // The do_oop work routines of the keep_alive and drain_marking_stack
2571 // oop closures will set the has_overflown flag if we overflow the
2572 // global marking stack.
2574 assert(_markStack.overflow() || _markStack.isEmpty(),
2575 "mark stack should be empty (unless it overflowed)");
2577 if (_markStack.overflow()) {
2578 // This should have been done already when we tried to push an
2579 // entry on to the global mark stack. But let's do it again.
2580 set_has_overflown();
2581 }
2583 assert(rp->num_q() == active_workers, "why not");
2585 rp->enqueue_discovered_references(executor);
2587 rp->verify_no_references_recorded();
2588 assert(!rp->discovery_enabled(), "Post condition");
2589 }
2591 if (has_overflown()) {
2592 // We can not trust g1_is_alive if the marking stack overflowed
2593 return;
2594 }
2596 assert(_markStack.isEmpty(), "Marking should have completed");
2598 // Unload Klasses, String, Symbols, Code Cache, etc.
2599 {
2600 G1RemarkGCTraceTime trace("Unloading", G1Log::finer());
2602 if (ClassUnloadingWithConcurrentMark) {
2603 bool purged_classes;
2605 {
2606 G1RemarkGCTraceTime trace("System Dictionary Unloading", G1Log::finest());
2607 purged_classes = SystemDictionary::do_unloading(&g1_is_alive);
2608 }
2610 {
2611 G1RemarkGCTraceTime trace("Parallel Unloading", G1Log::finest());
2612 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2613 }
2614 }
2616 if (G1StringDedup::is_enabled()) {
2617 G1RemarkGCTraceTime trace("String Deduplication Unlink", G1Log::finest());
2618 G1StringDedup::unlink(&g1_is_alive);
2619 }
2620 }
2621 }
2623 void ConcurrentMark::swapMarkBitMaps() {
2624 CMBitMapRO* temp = _prevMarkBitMap;
2625 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2626 _nextMarkBitMap = (CMBitMap*) temp;
2627 }
2629 class CMObjectClosure;
2631 // Closure for iterating over objects, currently only used for
2632 // processing SATB buffers.
2633 class CMObjectClosure : public ObjectClosure {
2634 private:
2635 CMTask* _task;
2637 public:
2638 void do_object(oop obj) {
2639 _task->deal_with_reference(obj);
2640 }
2642 CMObjectClosure(CMTask* task) : _task(task) { }
2643 };
2645 class G1RemarkThreadsClosure : public ThreadClosure {
2646 CMObjectClosure _cm_obj;
2647 G1CMOopClosure _cm_cl;
2648 MarkingCodeBlobClosure _code_cl;
2649 int _thread_parity;
2650 bool _is_par;
2652 public:
2653 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) :
2654 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2655 _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {}
2657 void do_thread(Thread* thread) {
2658 if (thread->is_Java_thread()) {
2659 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2660 JavaThread* jt = (JavaThread*)thread;
2662 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2663 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2664 // * Alive if on the stack of an executing method
2665 // * Weakly reachable otherwise
2666 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2667 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2668 jt->nmethods_do(&_code_cl);
2670 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2671 }
2672 } else if (thread->is_VM_thread()) {
2673 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2674 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2675 }
2676 }
2677 }
2678 };
2680 class CMRemarkTask: public AbstractGangTask {
2681 private:
2682 ConcurrentMark* _cm;
2683 bool _is_serial;
2684 public:
2685 void work(uint worker_id) {
2686 // Since all available tasks are actually started, we should
2687 // only proceed if we're supposed to be actived.
2688 if (worker_id < _cm->active_tasks()) {
2689 CMTask* task = _cm->task(worker_id);
2690 task->record_start_time();
2691 {
2692 ResourceMark rm;
2693 HandleMark hm;
2695 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial);
2696 Threads::threads_do(&threads_f);
2697 }
2699 do {
2700 task->do_marking_step(1000000000.0 /* something very large */,
2701 true /* do_termination */,
2702 _is_serial);
2703 } while (task->has_aborted() && !_cm->has_overflown());
2704 // If we overflow, then we do not want to restart. We instead
2705 // want to abort remark and do concurrent marking again.
2706 task->record_end_time();
2707 }
2708 }
2710 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2711 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2712 _cm->terminator()->reset_for_reuse(active_workers);
2713 }
2714 };
2716 void ConcurrentMark::checkpointRootsFinalWork() {
2717 ResourceMark rm;
2718 HandleMark hm;
2719 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2721 G1RemarkGCTraceTime trace("Finalize Marking", G1Log::finer());
2723 g1h->ensure_parsability(false);
2725 if (G1CollectedHeap::use_parallel_gc_threads()) {
2726 G1CollectedHeap::StrongRootsScope srs(g1h);
2727 // this is remark, so we'll use up all active threads
2728 uint active_workers = g1h->workers()->active_workers();
2729 if (active_workers == 0) {
2730 assert(active_workers > 0, "Should have been set earlier");
2731 active_workers = (uint) ParallelGCThreads;
2732 g1h->workers()->set_active_workers(active_workers);
2733 }
2734 set_concurrency_and_phase(active_workers, false /* concurrent */);
2735 // Leave _parallel_marking_threads at it's
2736 // value originally calculated in the ConcurrentMark
2737 // constructor and pass values of the active workers
2738 // through the gang in the task.
2740 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2741 // We will start all available threads, even if we decide that the
2742 // active_workers will be fewer. The extra ones will just bail out
2743 // immediately.
2744 g1h->set_par_threads(active_workers);
2745 g1h->workers()->run_task(&remarkTask);
2746 g1h->set_par_threads(0);
2747 } else {
2748 G1CollectedHeap::StrongRootsScope srs(g1h);
2749 uint active_workers = 1;
2750 set_concurrency_and_phase(active_workers, false /* concurrent */);
2752 // Note - if there's no work gang then the VMThread will be
2753 // the thread to execute the remark - serially. We have
2754 // to pass true for the is_serial parameter so that
2755 // CMTask::do_marking_step() doesn't enter the sync
2756 // barriers in the event of an overflow. Doing so will
2757 // cause an assert that the current thread is not a
2758 // concurrent GC thread.
2759 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2760 remarkTask.work(0);
2761 }
2762 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2763 guarantee(has_overflown() ||
2764 satb_mq_set.completed_buffers_num() == 0,
2765 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2766 BOOL_TO_STR(has_overflown()),
2767 satb_mq_set.completed_buffers_num()));
2769 print_stats();
2770 }
2772 #ifndef PRODUCT
2774 class PrintReachableOopClosure: public OopClosure {
2775 private:
2776 G1CollectedHeap* _g1h;
2777 outputStream* _out;
2778 VerifyOption _vo;
2779 bool _all;
2781 public:
2782 PrintReachableOopClosure(outputStream* out,
2783 VerifyOption vo,
2784 bool all) :
2785 _g1h(G1CollectedHeap::heap()),
2786 _out(out), _vo(vo), _all(all) { }
2788 void do_oop(narrowOop* p) { do_oop_work(p); }
2789 void do_oop( oop* p) { do_oop_work(p); }
2791 template <class T> void do_oop_work(T* p) {
2792 oop obj = oopDesc::load_decode_heap_oop(p);
2793 const char* str = NULL;
2794 const char* str2 = "";
2796 if (obj == NULL) {
2797 str = "";
2798 } else if (!_g1h->is_in_g1_reserved(obj)) {
2799 str = " O";
2800 } else {
2801 HeapRegion* hr = _g1h->heap_region_containing(obj);
2802 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2803 bool marked = _g1h->is_marked(obj, _vo);
2805 if (over_tams) {
2806 str = " >";
2807 if (marked) {
2808 str2 = " AND MARKED";
2809 }
2810 } else if (marked) {
2811 str = " M";
2812 } else {
2813 str = " NOT";
2814 }
2815 }
2817 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2818 p2i(p), p2i((void*) obj), str, str2);
2819 }
2820 };
2822 class PrintReachableObjectClosure : public ObjectClosure {
2823 private:
2824 G1CollectedHeap* _g1h;
2825 outputStream* _out;
2826 VerifyOption _vo;
2827 bool _all;
2828 HeapRegion* _hr;
2830 public:
2831 PrintReachableObjectClosure(outputStream* out,
2832 VerifyOption vo,
2833 bool all,
2834 HeapRegion* hr) :
2835 _g1h(G1CollectedHeap::heap()),
2836 _out(out), _vo(vo), _all(all), _hr(hr) { }
2838 void do_object(oop o) {
2839 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2840 bool marked = _g1h->is_marked(o, _vo);
2841 bool print_it = _all || over_tams || marked;
2843 if (print_it) {
2844 _out->print_cr(" "PTR_FORMAT"%s",
2845 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2846 PrintReachableOopClosure oopCl(_out, _vo, _all);
2847 o->oop_iterate_no_header(&oopCl);
2848 }
2849 }
2850 };
2852 class PrintReachableRegionClosure : public HeapRegionClosure {
2853 private:
2854 G1CollectedHeap* _g1h;
2855 outputStream* _out;
2856 VerifyOption _vo;
2857 bool _all;
2859 public:
2860 bool doHeapRegion(HeapRegion* hr) {
2861 HeapWord* b = hr->bottom();
2862 HeapWord* e = hr->end();
2863 HeapWord* t = hr->top();
2864 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2865 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2866 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2867 _out->cr();
2869 HeapWord* from = b;
2870 HeapWord* to = t;
2872 if (to > from) {
2873 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2874 _out->cr();
2875 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2876 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2877 _out->cr();
2878 }
2880 return false;
2881 }
2883 PrintReachableRegionClosure(outputStream* out,
2884 VerifyOption vo,
2885 bool all) :
2886 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2887 };
2889 void ConcurrentMark::print_reachable(const char* str,
2890 VerifyOption vo,
2891 bool all) {
2892 gclog_or_tty->cr();
2893 gclog_or_tty->print_cr("== Doing heap dump... ");
2895 if (G1PrintReachableBaseFile == NULL) {
2896 gclog_or_tty->print_cr(" #### error: no base file defined");
2897 return;
2898 }
2900 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2901 (JVM_MAXPATHLEN - 1)) {
2902 gclog_or_tty->print_cr(" #### error: file name too long");
2903 return;
2904 }
2906 char file_name[JVM_MAXPATHLEN];
2907 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2908 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2910 fileStream fout(file_name);
2911 if (!fout.is_open()) {
2912 gclog_or_tty->print_cr(" #### error: could not open file");
2913 return;
2914 }
2916 outputStream* out = &fout;
2917 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2918 out->cr();
2920 out->print_cr("--- ITERATING OVER REGIONS");
2921 out->cr();
2922 PrintReachableRegionClosure rcl(out, vo, all);
2923 _g1h->heap_region_iterate(&rcl);
2924 out->cr();
2926 gclog_or_tty->print_cr(" done");
2927 gclog_or_tty->flush();
2928 }
2930 #endif // PRODUCT
2932 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2933 // Note we are overriding the read-only view of the prev map here, via
2934 // the cast.
2935 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2936 }
2938 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2939 _nextMarkBitMap->clearRange(mr);
2940 }
2942 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) {
2943 clearRangePrevBitmap(mr);
2944 clearRangeNextBitmap(mr);
2945 }
2947 HeapRegion*
2948 ConcurrentMark::claim_region(uint worker_id) {
2949 // "checkpoint" the finger
2950 HeapWord* finger = _finger;
2952 // _heap_end will not change underneath our feet; it only changes at
2953 // yield points.
2954 while (finger < _heap_end) {
2955 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2957 // Note on how this code handles humongous regions. In the
2958 // normal case the finger will reach the start of a "starts
2959 // humongous" (SH) region. Its end will either be the end of the
2960 // last "continues humongous" (CH) region in the sequence, or the
2961 // standard end of the SH region (if the SH is the only region in
2962 // the sequence). That way claim_region() will skip over the CH
2963 // regions. However, there is a subtle race between a CM thread
2964 // executing this method and a mutator thread doing a humongous
2965 // object allocation. The two are not mutually exclusive as the CM
2966 // thread does not need to hold the Heap_lock when it gets
2967 // here. So there is a chance that claim_region() will come across
2968 // a free region that's in the progress of becoming a SH or a CH
2969 // region. In the former case, it will either
2970 // a) Miss the update to the region's end, in which case it will
2971 // visit every subsequent CH region, will find their bitmaps
2972 // empty, and do nothing, or
2973 // b) Will observe the update of the region's end (in which case
2974 // it will skip the subsequent CH regions).
2975 // If it comes across a region that suddenly becomes CH, the
2976 // scenario will be similar to b). So, the race between
2977 // claim_region() and a humongous object allocation might force us
2978 // to do a bit of unnecessary work (due to some unnecessary bitmap
2979 // iterations) but it should not introduce and correctness issues.
2980 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2981 HeapWord* bottom = curr_region->bottom();
2982 HeapWord* end = curr_region->end();
2983 HeapWord* limit = curr_region->next_top_at_mark_start();
2985 if (verbose_low()) {
2986 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2987 "["PTR_FORMAT", "PTR_FORMAT"), "
2988 "limit = "PTR_FORMAT,
2989 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2990 }
2992 // Is the gap between reading the finger and doing the CAS too long?
2993 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2994 if (res == finger) {
2995 // we succeeded
2997 // notice that _finger == end cannot be guaranteed here since,
2998 // someone else might have moved the finger even further
2999 assert(_finger >= end, "the finger should have moved forward");
3001 if (verbose_low()) {
3002 gclog_or_tty->print_cr("[%u] we were successful with region = "
3003 PTR_FORMAT, worker_id, p2i(curr_region));
3004 }
3006 if (limit > bottom) {
3007 if (verbose_low()) {
3008 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
3009 "returning it ", worker_id, p2i(curr_region));
3010 }
3011 return curr_region;
3012 } else {
3013 assert(limit == bottom,
3014 "the region limit should be at bottom");
3015 if (verbose_low()) {
3016 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
3017 "returning NULL", worker_id, p2i(curr_region));
3018 }
3019 // we return NULL and the caller should try calling
3020 // claim_region() again.
3021 return NULL;
3022 }
3023 } else {
3024 assert(_finger > finger, "the finger should have moved forward");
3025 if (verbose_low()) {
3026 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
3027 "global finger = "PTR_FORMAT", "
3028 "our finger = "PTR_FORMAT,
3029 worker_id, p2i(_finger), p2i(finger));
3030 }
3032 // read it again
3033 finger = _finger;
3034 }
3035 }
3037 return NULL;
3038 }
3040 #ifndef PRODUCT
3041 enum VerifyNoCSetOopsPhase {
3042 VerifyNoCSetOopsStack,
3043 VerifyNoCSetOopsQueues,
3044 VerifyNoCSetOopsSATBCompleted,
3045 VerifyNoCSetOopsSATBThread
3046 };
3048 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
3049 private:
3050 G1CollectedHeap* _g1h;
3051 VerifyNoCSetOopsPhase _phase;
3052 int _info;
3054 const char* phase_str() {
3055 switch (_phase) {
3056 case VerifyNoCSetOopsStack: return "Stack";
3057 case VerifyNoCSetOopsQueues: return "Queue";
3058 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
3059 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
3060 default: ShouldNotReachHere();
3061 }
3062 return NULL;
3063 }
3065 void do_object_work(oop obj) {
3066 guarantee(!_g1h->obj_in_cs(obj),
3067 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
3068 p2i((void*) obj), phase_str(), _info));
3069 }
3071 public:
3072 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3074 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3075 _phase = phase;
3076 _info = info;
3077 }
3079 virtual void do_oop(oop* p) {
3080 oop obj = oopDesc::load_decode_heap_oop(p);
3081 do_object_work(obj);
3082 }
3084 virtual void do_oop(narrowOop* p) {
3085 // We should not come across narrow oops while scanning marking
3086 // stacks and SATB buffers.
3087 ShouldNotReachHere();
3088 }
3090 virtual void do_object(oop obj) {
3091 do_object_work(obj);
3092 }
3093 };
3095 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3096 bool verify_enqueued_buffers,
3097 bool verify_thread_buffers,
3098 bool verify_fingers) {
3099 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3100 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3101 return;
3102 }
3104 VerifyNoCSetOopsClosure cl;
3106 if (verify_stacks) {
3107 // Verify entries on the global mark stack
3108 cl.set_phase(VerifyNoCSetOopsStack);
3109 _markStack.oops_do(&cl);
3111 // Verify entries on the task queues
3112 for (uint i = 0; i < _max_worker_id; i += 1) {
3113 cl.set_phase(VerifyNoCSetOopsQueues, i);
3114 CMTaskQueue* queue = _task_queues->queue(i);
3115 queue->oops_do(&cl);
3116 }
3117 }
3119 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3121 // Verify entries on the enqueued SATB buffers
3122 if (verify_enqueued_buffers) {
3123 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3124 satb_qs.iterate_completed_buffers_read_only(&cl);
3125 }
3127 // Verify entries on the per-thread SATB buffers
3128 if (verify_thread_buffers) {
3129 cl.set_phase(VerifyNoCSetOopsSATBThread);
3130 satb_qs.iterate_thread_buffers_read_only(&cl);
3131 }
3133 if (verify_fingers) {
3134 // Verify the global finger
3135 HeapWord* global_finger = finger();
3136 if (global_finger != NULL && global_finger < _heap_end) {
3137 // The global finger always points to a heap region boundary. We
3138 // use heap_region_containing_raw() to get the containing region
3139 // given that the global finger could be pointing to a free region
3140 // which subsequently becomes continues humongous. If that
3141 // happens, heap_region_containing() will return the bottom of the
3142 // corresponding starts humongous region and the check below will
3143 // not hold any more.
3144 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3145 guarantee(global_finger == global_hr->bottom(),
3146 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3147 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3148 }
3150 // Verify the task fingers
3151 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3152 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3153 CMTask* task = _tasks[i];
3154 HeapWord* task_finger = task->finger();
3155 if (task_finger != NULL && task_finger < _heap_end) {
3156 // See above note on the global finger verification.
3157 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3158 guarantee(task_finger == task_hr->bottom() ||
3159 !task_hr->in_collection_set(),
3160 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3161 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3162 }
3163 }
3164 }
3165 }
3166 #endif // PRODUCT
3168 // Aggregate the counting data that was constructed concurrently
3169 // with marking.
3170 class AggregateCountDataHRClosure: public HeapRegionClosure {
3171 G1CollectedHeap* _g1h;
3172 ConcurrentMark* _cm;
3173 CardTableModRefBS* _ct_bs;
3174 BitMap* _cm_card_bm;
3175 uint _max_worker_id;
3177 public:
3178 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3179 BitMap* cm_card_bm,
3180 uint max_worker_id) :
3181 _g1h(g1h), _cm(g1h->concurrent_mark()),
3182 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3183 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3185 bool doHeapRegion(HeapRegion* hr) {
3186 if (hr->continuesHumongous()) {
3187 // We will ignore these here and process them when their
3188 // associated "starts humongous" region is processed.
3189 // Note that we cannot rely on their associated
3190 // "starts humongous" region to have their bit set to 1
3191 // since, due to the region chunking in the parallel region
3192 // iteration, a "continues humongous" region might be visited
3193 // before its associated "starts humongous".
3194 return false;
3195 }
3197 HeapWord* start = hr->bottom();
3198 HeapWord* limit = hr->next_top_at_mark_start();
3199 HeapWord* end = hr->end();
3201 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3202 err_msg("Preconditions not met - "
3203 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3204 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3205 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3207 assert(hr->next_marked_bytes() == 0, "Precondition");
3209 if (start == limit) {
3210 // NTAMS of this region has not been set so nothing to do.
3211 return false;
3212 }
3214 // 'start' should be in the heap.
3215 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3216 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3217 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3219 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3220 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3221 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3223 // If ntams is not card aligned then we bump card bitmap index
3224 // for limit so that we get the all the cards spanned by
3225 // the object ending at ntams.
3226 // Note: if this is the last region in the heap then ntams
3227 // could be actually just beyond the end of the the heap;
3228 // limit_idx will then correspond to a (non-existent) card
3229 // that is also outside the heap.
3230 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3231 limit_idx += 1;
3232 }
3234 assert(limit_idx <= end_idx, "or else use atomics");
3236 // Aggregate the "stripe" in the count data associated with hr.
3237 uint hrs_index = hr->hrs_index();
3238 size_t marked_bytes = 0;
3240 for (uint i = 0; i < _max_worker_id; i += 1) {
3241 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3242 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3244 // Fetch the marked_bytes in this region for task i and
3245 // add it to the running total for this region.
3246 marked_bytes += marked_bytes_array[hrs_index];
3248 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3249 // into the global card bitmap.
3250 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3252 while (scan_idx < limit_idx) {
3253 assert(task_card_bm->at(scan_idx) == true, "should be");
3254 _cm_card_bm->set_bit(scan_idx);
3255 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3257 // BitMap::get_next_one_offset() can handle the case when
3258 // its left_offset parameter is greater than its right_offset
3259 // parameter. It does, however, have an early exit if
3260 // left_offset == right_offset. So let's limit the value
3261 // passed in for left offset here.
3262 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3263 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3264 }
3265 }
3267 // Update the marked bytes for this region.
3268 hr->add_to_marked_bytes(marked_bytes);
3270 // Next heap region
3271 return false;
3272 }
3273 };
3275 class G1AggregateCountDataTask: public AbstractGangTask {
3276 protected:
3277 G1CollectedHeap* _g1h;
3278 ConcurrentMark* _cm;
3279 BitMap* _cm_card_bm;
3280 uint _max_worker_id;
3281 int _active_workers;
3283 public:
3284 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3285 ConcurrentMark* cm,
3286 BitMap* cm_card_bm,
3287 uint max_worker_id,
3288 int n_workers) :
3289 AbstractGangTask("Count Aggregation"),
3290 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3291 _max_worker_id(max_worker_id),
3292 _active_workers(n_workers) { }
3294 void work(uint worker_id) {
3295 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3297 if (G1CollectedHeap::use_parallel_gc_threads()) {
3298 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3299 _active_workers,
3300 HeapRegion::AggregateCountClaimValue);
3301 } else {
3302 _g1h->heap_region_iterate(&cl);
3303 }
3304 }
3305 };
3308 void ConcurrentMark::aggregate_count_data() {
3309 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3310 _g1h->workers()->active_workers() :
3311 1);
3313 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3314 _max_worker_id, n_workers);
3316 if (G1CollectedHeap::use_parallel_gc_threads()) {
3317 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3318 "sanity check");
3319 _g1h->set_par_threads(n_workers);
3320 _g1h->workers()->run_task(&g1_par_agg_task);
3321 _g1h->set_par_threads(0);
3323 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3324 "sanity check");
3325 _g1h->reset_heap_region_claim_values();
3326 } else {
3327 g1_par_agg_task.work(0);
3328 }
3329 }
3331 // Clear the per-worker arrays used to store the per-region counting data
3332 void ConcurrentMark::clear_all_count_data() {
3333 // Clear the global card bitmap - it will be filled during
3334 // liveness count aggregation (during remark) and the
3335 // final counting task.
3336 _card_bm.clear();
3338 // Clear the global region bitmap - it will be filled as part
3339 // of the final counting task.
3340 _region_bm.clear();
3342 uint max_regions = _g1h->max_regions();
3343 assert(_max_worker_id > 0, "uninitialized");
3345 for (uint i = 0; i < _max_worker_id; i += 1) {
3346 BitMap* task_card_bm = count_card_bitmap_for(i);
3347 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3349 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3350 assert(marked_bytes_array != NULL, "uninitialized");
3352 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3353 task_card_bm->clear();
3354 }
3355 }
3357 void ConcurrentMark::print_stats() {
3358 if (verbose_stats()) {
3359 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3360 for (size_t i = 0; i < _active_tasks; ++i) {
3361 _tasks[i]->print_stats();
3362 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3363 }
3364 }
3365 }
3367 // abandon current marking iteration due to a Full GC
3368 void ConcurrentMark::abort() {
3369 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3370 // concurrent bitmap clearing.
3371 _nextMarkBitMap->clearAll();
3373 // Note we cannot clear the previous marking bitmap here
3374 // since VerifyDuringGC verifies the objects marked during
3375 // a full GC against the previous bitmap.
3377 // Clear the liveness counting data
3378 clear_all_count_data();
3379 // Empty mark stack
3380 reset_marking_state();
3381 for (uint i = 0; i < _max_worker_id; ++i) {
3382 _tasks[i]->clear_region_fields();
3383 }
3384 _first_overflow_barrier_sync.abort();
3385 _second_overflow_barrier_sync.abort();
3386 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3387 if (!gc_id.is_undefined()) {
3388 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3389 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3390 _aborted_gc_id = gc_id;
3391 }
3392 _has_aborted = true;
3394 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3395 satb_mq_set.abandon_partial_marking();
3396 // This can be called either during or outside marking, we'll read
3397 // the expected_active value from the SATB queue set.
3398 satb_mq_set.set_active_all_threads(
3399 false, /* new active value */
3400 satb_mq_set.is_active() /* expected_active */);
3402 _g1h->trace_heap_after_concurrent_cycle();
3403 _g1h->register_concurrent_cycle_end();
3404 }
3406 const GCId& ConcurrentMark::concurrent_gc_id() {
3407 if (has_aborted()) {
3408 return _aborted_gc_id;
3409 }
3410 return _g1h->gc_tracer_cm()->gc_id();
3411 }
3413 static void print_ms_time_info(const char* prefix, const char* name,
3414 NumberSeq& ns) {
3415 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3416 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3417 if (ns.num() > 0) {
3418 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3419 prefix, ns.sd(), ns.maximum());
3420 }
3421 }
3423 void ConcurrentMark::print_summary_info() {
3424 gclog_or_tty->print_cr(" Concurrent marking:");
3425 print_ms_time_info(" ", "init marks", _init_times);
3426 print_ms_time_info(" ", "remarks", _remark_times);
3427 {
3428 print_ms_time_info(" ", "final marks", _remark_mark_times);
3429 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3431 }
3432 print_ms_time_info(" ", "cleanups", _cleanup_times);
3433 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3434 _total_counting_time,
3435 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3436 (double)_cleanup_times.num()
3437 : 0.0));
3438 if (G1ScrubRemSets) {
3439 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3440 _total_rs_scrub_time,
3441 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3442 (double)_cleanup_times.num()
3443 : 0.0));
3444 }
3445 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3446 (_init_times.sum() + _remark_times.sum() +
3447 _cleanup_times.sum())/1000.0);
3448 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3449 "(%8.2f s marking).",
3450 cmThread()->vtime_accum(),
3451 cmThread()->vtime_mark_accum());
3452 }
3454 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3455 if (use_parallel_marking_threads()) {
3456 _parallel_workers->print_worker_threads_on(st);
3457 }
3458 }
3460 void ConcurrentMark::print_on_error(outputStream* st) const {
3461 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3462 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3463 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3464 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3465 }
3467 // We take a break if someone is trying to stop the world.
3468 bool ConcurrentMark::do_yield_check(uint worker_id) {
3469 if (SuspendibleThreadSet::should_yield()) {
3470 if (worker_id == 0) {
3471 _g1h->g1_policy()->record_concurrent_pause();
3472 }
3473 SuspendibleThreadSet::yield();
3474 return true;
3475 } else {
3476 return false;
3477 }
3478 }
3480 bool ConcurrentMark::containing_card_is_marked(void* p) {
3481 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1);
3482 return _card_bm.at(offset >> CardTableModRefBS::card_shift);
3483 }
3485 bool ConcurrentMark::containing_cards_are_marked(void* start,
3486 void* last) {
3487 return containing_card_is_marked(start) &&
3488 containing_card_is_marked(last);
3489 }
3491 #ifndef PRODUCT
3492 // for debugging purposes
3493 void ConcurrentMark::print_finger() {
3494 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3495 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3496 for (uint i = 0; i < _max_worker_id; ++i) {
3497 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3498 }
3499 gclog_or_tty->cr();
3500 }
3501 #endif
3503 void CMTask::scan_object(oop obj) {
3504 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3506 if (_cm->verbose_high()) {
3507 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3508 _worker_id, p2i((void*) obj));
3509 }
3511 size_t obj_size = obj->size();
3512 _words_scanned += obj_size;
3514 obj->oop_iterate(_cm_oop_closure);
3515 statsOnly( ++_objs_scanned );
3516 check_limits();
3517 }
3519 // Closure for iteration over bitmaps
3520 class CMBitMapClosure : public BitMapClosure {
3521 private:
3522 // the bitmap that is being iterated over
3523 CMBitMap* _nextMarkBitMap;
3524 ConcurrentMark* _cm;
3525 CMTask* _task;
3527 public:
3528 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3529 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3531 bool do_bit(size_t offset) {
3532 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3533 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3534 assert( addr < _cm->finger(), "invariant");
3536 statsOnly( _task->increase_objs_found_on_bitmap() );
3537 assert(addr >= _task->finger(), "invariant");
3539 // We move that task's local finger along.
3540 _task->move_finger_to(addr);
3542 _task->scan_object(oop(addr));
3543 // we only partially drain the local queue and global stack
3544 _task->drain_local_queue(true);
3545 _task->drain_global_stack(true);
3547 // if the has_aborted flag has been raised, we need to bail out of
3548 // the iteration
3549 return !_task->has_aborted();
3550 }
3551 };
3553 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3554 ConcurrentMark* cm,
3555 CMTask* task)
3556 : _g1h(g1h), _cm(cm), _task(task) {
3557 assert(_ref_processor == NULL, "should be initialized to NULL");
3559 if (G1UseConcMarkReferenceProcessing) {
3560 _ref_processor = g1h->ref_processor_cm();
3561 assert(_ref_processor != NULL, "should not be NULL");
3562 }
3563 }
3565 void CMTask::setup_for_region(HeapRegion* hr) {
3566 assert(hr != NULL,
3567 "claim_region() should have filtered out NULL regions");
3568 assert(!hr->continuesHumongous(),
3569 "claim_region() should have filtered out continues humongous regions");
3571 if (_cm->verbose_low()) {
3572 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3573 _worker_id, p2i(hr));
3574 }
3576 _curr_region = hr;
3577 _finger = hr->bottom();
3578 update_region_limit();
3579 }
3581 void CMTask::update_region_limit() {
3582 HeapRegion* hr = _curr_region;
3583 HeapWord* bottom = hr->bottom();
3584 HeapWord* limit = hr->next_top_at_mark_start();
3586 if (limit == bottom) {
3587 if (_cm->verbose_low()) {
3588 gclog_or_tty->print_cr("[%u] found an empty region "
3589 "["PTR_FORMAT", "PTR_FORMAT")",
3590 _worker_id, p2i(bottom), p2i(limit));
3591 }
3592 // The region was collected underneath our feet.
3593 // We set the finger to bottom to ensure that the bitmap
3594 // iteration that will follow this will not do anything.
3595 // (this is not a condition that holds when we set the region up,
3596 // as the region is not supposed to be empty in the first place)
3597 _finger = bottom;
3598 } else if (limit >= _region_limit) {
3599 assert(limit >= _finger, "peace of mind");
3600 } else {
3601 assert(limit < _region_limit, "only way to get here");
3602 // This can happen under some pretty unusual circumstances. An
3603 // evacuation pause empties the region underneath our feet (NTAMS
3604 // at bottom). We then do some allocation in the region (NTAMS
3605 // stays at bottom), followed by the region being used as a GC
3606 // alloc region (NTAMS will move to top() and the objects
3607 // originally below it will be grayed). All objects now marked in
3608 // the region are explicitly grayed, if below the global finger,
3609 // and we do not need in fact to scan anything else. So, we simply
3610 // set _finger to be limit to ensure that the bitmap iteration
3611 // doesn't do anything.
3612 _finger = limit;
3613 }
3615 _region_limit = limit;
3616 }
3618 void CMTask::giveup_current_region() {
3619 assert(_curr_region != NULL, "invariant");
3620 if (_cm->verbose_low()) {
3621 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3622 _worker_id, p2i(_curr_region));
3623 }
3624 clear_region_fields();
3625 }
3627 void CMTask::clear_region_fields() {
3628 // Values for these three fields that indicate that we're not
3629 // holding on to a region.
3630 _curr_region = NULL;
3631 _finger = NULL;
3632 _region_limit = NULL;
3633 }
3635 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3636 if (cm_oop_closure == NULL) {
3637 assert(_cm_oop_closure != NULL, "invariant");
3638 } else {
3639 assert(_cm_oop_closure == NULL, "invariant");
3640 }
3641 _cm_oop_closure = cm_oop_closure;
3642 }
3644 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3645 guarantee(nextMarkBitMap != NULL, "invariant");
3647 if (_cm->verbose_low()) {
3648 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3649 }
3651 _nextMarkBitMap = nextMarkBitMap;
3652 clear_region_fields();
3654 _calls = 0;
3655 _elapsed_time_ms = 0.0;
3656 _termination_time_ms = 0.0;
3657 _termination_start_time_ms = 0.0;
3659 #if _MARKING_STATS_
3660 _local_pushes = 0;
3661 _local_pops = 0;
3662 _local_max_size = 0;
3663 _objs_scanned = 0;
3664 _global_pushes = 0;
3665 _global_pops = 0;
3666 _global_max_size = 0;
3667 _global_transfers_to = 0;
3668 _global_transfers_from = 0;
3669 _regions_claimed = 0;
3670 _objs_found_on_bitmap = 0;
3671 _satb_buffers_processed = 0;
3672 _steal_attempts = 0;
3673 _steals = 0;
3674 _aborted = 0;
3675 _aborted_overflow = 0;
3676 _aborted_cm_aborted = 0;
3677 _aborted_yield = 0;
3678 _aborted_timed_out = 0;
3679 _aborted_satb = 0;
3680 _aborted_termination = 0;
3681 #endif // _MARKING_STATS_
3682 }
3684 bool CMTask::should_exit_termination() {
3685 regular_clock_call();
3686 // This is called when we are in the termination protocol. We should
3687 // quit if, for some reason, this task wants to abort or the global
3688 // stack is not empty (this means that we can get work from it).
3689 return !_cm->mark_stack_empty() || has_aborted();
3690 }
3692 void CMTask::reached_limit() {
3693 assert(_words_scanned >= _words_scanned_limit ||
3694 _refs_reached >= _refs_reached_limit ,
3695 "shouldn't have been called otherwise");
3696 regular_clock_call();
3697 }
3699 void CMTask::regular_clock_call() {
3700 if (has_aborted()) return;
3702 // First, we need to recalculate the words scanned and refs reached
3703 // limits for the next clock call.
3704 recalculate_limits();
3706 // During the regular clock call we do the following
3708 // (1) If an overflow has been flagged, then we abort.
3709 if (_cm->has_overflown()) {
3710 set_has_aborted();
3711 return;
3712 }
3714 // If we are not concurrent (i.e. we're doing remark) we don't need
3715 // to check anything else. The other steps are only needed during
3716 // the concurrent marking phase.
3717 if (!concurrent()) return;
3719 // (2) If marking has been aborted for Full GC, then we also abort.
3720 if (_cm->has_aborted()) {
3721 set_has_aborted();
3722 statsOnly( ++_aborted_cm_aborted );
3723 return;
3724 }
3726 double curr_time_ms = os::elapsedVTime() * 1000.0;
3728 // (3) If marking stats are enabled, then we update the step history.
3729 #if _MARKING_STATS_
3730 if (_words_scanned >= _words_scanned_limit) {
3731 ++_clock_due_to_scanning;
3732 }
3733 if (_refs_reached >= _refs_reached_limit) {
3734 ++_clock_due_to_marking;
3735 }
3737 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3738 _interval_start_time_ms = curr_time_ms;
3739 _all_clock_intervals_ms.add(last_interval_ms);
3741 if (_cm->verbose_medium()) {
3742 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3743 "scanned = %d%s, refs reached = %d%s",
3744 _worker_id, last_interval_ms,
3745 _words_scanned,
3746 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3747 _refs_reached,
3748 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3749 }
3750 #endif // _MARKING_STATS_
3752 // (4) We check whether we should yield. If we have to, then we abort.
3753 if (SuspendibleThreadSet::should_yield()) {
3754 // We should yield. To do this we abort the task. The caller is
3755 // responsible for yielding.
3756 set_has_aborted();
3757 statsOnly( ++_aborted_yield );
3758 return;
3759 }
3761 // (5) We check whether we've reached our time quota. If we have,
3762 // then we abort.
3763 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3764 if (elapsed_time_ms > _time_target_ms) {
3765 set_has_aborted();
3766 _has_timed_out = true;
3767 statsOnly( ++_aborted_timed_out );
3768 return;
3769 }
3771 // (6) Finally, we check whether there are enough completed STAB
3772 // buffers available for processing. If there are, we abort.
3773 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3774 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3775 if (_cm->verbose_low()) {
3776 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3777 _worker_id);
3778 }
3779 // we do need to process SATB buffers, we'll abort and restart
3780 // the marking task to do so
3781 set_has_aborted();
3782 statsOnly( ++_aborted_satb );
3783 return;
3784 }
3785 }
3787 void CMTask::recalculate_limits() {
3788 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3789 _words_scanned_limit = _real_words_scanned_limit;
3791 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3792 _refs_reached_limit = _real_refs_reached_limit;
3793 }
3795 void CMTask::decrease_limits() {
3796 // This is called when we believe that we're going to do an infrequent
3797 // operation which will increase the per byte scanned cost (i.e. move
3798 // entries to/from the global stack). It basically tries to decrease the
3799 // scanning limit so that the clock is called earlier.
3801 if (_cm->verbose_medium()) {
3802 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3803 }
3805 _words_scanned_limit = _real_words_scanned_limit -
3806 3 * words_scanned_period / 4;
3807 _refs_reached_limit = _real_refs_reached_limit -
3808 3 * refs_reached_period / 4;
3809 }
3811 void CMTask::move_entries_to_global_stack() {
3812 // local array where we'll store the entries that will be popped
3813 // from the local queue
3814 oop buffer[global_stack_transfer_size];
3816 int n = 0;
3817 oop obj;
3818 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3819 buffer[n] = obj;
3820 ++n;
3821 }
3823 if (n > 0) {
3824 // we popped at least one entry from the local queue
3826 statsOnly( ++_global_transfers_to; _local_pops += n );
3828 if (!_cm->mark_stack_push(buffer, n)) {
3829 if (_cm->verbose_low()) {
3830 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3831 _worker_id);
3832 }
3833 set_has_aborted();
3834 } else {
3835 // the transfer was successful
3837 if (_cm->verbose_medium()) {
3838 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3839 _worker_id, n);
3840 }
3841 statsOnly( int tmp_size = _cm->mark_stack_size();
3842 if (tmp_size > _global_max_size) {
3843 _global_max_size = tmp_size;
3844 }
3845 _global_pushes += n );
3846 }
3847 }
3849 // this operation was quite expensive, so decrease the limits
3850 decrease_limits();
3851 }
3853 void CMTask::get_entries_from_global_stack() {
3854 // local array where we'll store the entries that will be popped
3855 // from the global stack.
3856 oop buffer[global_stack_transfer_size];
3857 int n;
3858 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3859 assert(n <= global_stack_transfer_size,
3860 "we should not pop more than the given limit");
3861 if (n > 0) {
3862 // yes, we did actually pop at least one entry
3864 statsOnly( ++_global_transfers_from; _global_pops += n );
3865 if (_cm->verbose_medium()) {
3866 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3867 _worker_id, n);
3868 }
3869 for (int i = 0; i < n; ++i) {
3870 bool success = _task_queue->push(buffer[i]);
3871 // We only call this when the local queue is empty or under a
3872 // given target limit. So, we do not expect this push to fail.
3873 assert(success, "invariant");
3874 }
3876 statsOnly( int tmp_size = _task_queue->size();
3877 if (tmp_size > _local_max_size) {
3878 _local_max_size = tmp_size;
3879 }
3880 _local_pushes += n );
3881 }
3883 // this operation was quite expensive, so decrease the limits
3884 decrease_limits();
3885 }
3887 void CMTask::drain_local_queue(bool partially) {
3888 if (has_aborted()) return;
3890 // Decide what the target size is, depending whether we're going to
3891 // drain it partially (so that other tasks can steal if they run out
3892 // of things to do) or totally (at the very end).
3893 size_t target_size;
3894 if (partially) {
3895 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3896 } else {
3897 target_size = 0;
3898 }
3900 if (_task_queue->size() > target_size) {
3901 if (_cm->verbose_high()) {
3902 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3903 _worker_id, target_size);
3904 }
3906 oop obj;
3907 bool ret = _task_queue->pop_local(obj);
3908 while (ret) {
3909 statsOnly( ++_local_pops );
3911 if (_cm->verbose_high()) {
3912 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3913 p2i((void*) obj));
3914 }
3916 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3917 assert(!_g1h->is_on_master_free_list(
3918 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3920 scan_object(obj);
3922 if (_task_queue->size() <= target_size || has_aborted()) {
3923 ret = false;
3924 } else {
3925 ret = _task_queue->pop_local(obj);
3926 }
3927 }
3929 if (_cm->verbose_high()) {
3930 gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3931 _worker_id, _task_queue->size());
3932 }
3933 }
3934 }
3936 void CMTask::drain_global_stack(bool partially) {
3937 if (has_aborted()) return;
3939 // We have a policy to drain the local queue before we attempt to
3940 // drain the global stack.
3941 assert(partially || _task_queue->size() == 0, "invariant");
3943 // Decide what the target size is, depending whether we're going to
3944 // drain it partially (so that other tasks can steal if they run out
3945 // of things to do) or totally (at the very end). Notice that,
3946 // because we move entries from the global stack in chunks or
3947 // because another task might be doing the same, we might in fact
3948 // drop below the target. But, this is not a problem.
3949 size_t target_size;
3950 if (partially) {
3951 target_size = _cm->partial_mark_stack_size_target();
3952 } else {
3953 target_size = 0;
3954 }
3956 if (_cm->mark_stack_size() > target_size) {
3957 if (_cm->verbose_low()) {
3958 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3959 _worker_id, target_size);
3960 }
3962 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3963 get_entries_from_global_stack();
3964 drain_local_queue(partially);
3965 }
3967 if (_cm->verbose_low()) {
3968 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3969 _worker_id, _cm->mark_stack_size());
3970 }
3971 }
3972 }
3974 // SATB Queue has several assumptions on whether to call the par or
3975 // non-par versions of the methods. this is why some of the code is
3976 // replicated. We should really get rid of the single-threaded version
3977 // of the code to simplify things.
3978 void CMTask::drain_satb_buffers() {
3979 if (has_aborted()) return;
3981 // We set this so that the regular clock knows that we're in the
3982 // middle of draining buffers and doesn't set the abort flag when it
3983 // notices that SATB buffers are available for draining. It'd be
3984 // very counter productive if it did that. :-)
3985 _draining_satb_buffers = true;
3987 CMObjectClosure oc(this);
3988 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3989 if (G1CollectedHeap::use_parallel_gc_threads()) {
3990 satb_mq_set.set_par_closure(_worker_id, &oc);
3991 } else {
3992 satb_mq_set.set_closure(&oc);
3993 }
3995 // This keeps claiming and applying the closure to completed buffers
3996 // until we run out of buffers or we need to abort.
3997 if (G1CollectedHeap::use_parallel_gc_threads()) {
3998 while (!has_aborted() &&
3999 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
4000 if (_cm->verbose_medium()) {
4001 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4002 }
4003 statsOnly( ++_satb_buffers_processed );
4004 regular_clock_call();
4005 }
4006 } else {
4007 while (!has_aborted() &&
4008 satb_mq_set.apply_closure_to_completed_buffer()) {
4009 if (_cm->verbose_medium()) {
4010 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4011 }
4012 statsOnly( ++_satb_buffers_processed );
4013 regular_clock_call();
4014 }
4015 }
4017 _draining_satb_buffers = false;
4019 assert(has_aborted() ||
4020 concurrent() ||
4021 satb_mq_set.completed_buffers_num() == 0, "invariant");
4023 if (G1CollectedHeap::use_parallel_gc_threads()) {
4024 satb_mq_set.set_par_closure(_worker_id, NULL);
4025 } else {
4026 satb_mq_set.set_closure(NULL);
4027 }
4029 // again, this was a potentially expensive operation, decrease the
4030 // limits to get the regular clock call early
4031 decrease_limits();
4032 }
4034 void CMTask::print_stats() {
4035 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
4036 _worker_id, _calls);
4037 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
4038 _elapsed_time_ms, _termination_time_ms);
4039 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4040 _step_times_ms.num(), _step_times_ms.avg(),
4041 _step_times_ms.sd());
4042 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4043 _step_times_ms.maximum(), _step_times_ms.sum());
4045 #if _MARKING_STATS_
4046 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4047 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
4048 _all_clock_intervals_ms.sd());
4049 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4050 _all_clock_intervals_ms.maximum(),
4051 _all_clock_intervals_ms.sum());
4052 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
4053 _clock_due_to_scanning, _clock_due_to_marking);
4054 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
4055 _objs_scanned, _objs_found_on_bitmap);
4056 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
4057 _local_pushes, _local_pops, _local_max_size);
4058 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
4059 _global_pushes, _global_pops, _global_max_size);
4060 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
4061 _global_transfers_to,_global_transfers_from);
4062 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
4063 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
4064 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
4065 _steal_attempts, _steals);
4066 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
4067 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
4068 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
4069 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
4070 _aborted_timed_out, _aborted_satb, _aborted_termination);
4071 #endif // _MARKING_STATS_
4072 }
4074 /*****************************************************************************
4076 The do_marking_step(time_target_ms, ...) method is the building
4077 block of the parallel marking framework. It can be called in parallel
4078 with other invocations of do_marking_step() on different tasks
4079 (but only one per task, obviously) and concurrently with the
4080 mutator threads, or during remark, hence it eliminates the need
4081 for two versions of the code. When called during remark, it will
4082 pick up from where the task left off during the concurrent marking
4083 phase. Interestingly, tasks are also claimable during evacuation
4084 pauses too, since do_marking_step() ensures that it aborts before
4085 it needs to yield.
4087 The data structures that it uses to do marking work are the
4088 following:
4090 (1) Marking Bitmap. If there are gray objects that appear only
4091 on the bitmap (this happens either when dealing with an overflow
4092 or when the initial marking phase has simply marked the roots
4093 and didn't push them on the stack), then tasks claim heap
4094 regions whose bitmap they then scan to find gray objects. A
4095 global finger indicates where the end of the last claimed region
4096 is. A local finger indicates how far into the region a task has
4097 scanned. The two fingers are used to determine how to gray an
4098 object (i.e. whether simply marking it is OK, as it will be
4099 visited by a task in the future, or whether it needs to be also
4100 pushed on a stack).
4102 (2) Local Queue. The local queue of the task which is accessed
4103 reasonably efficiently by the task. Other tasks can steal from
4104 it when they run out of work. Throughout the marking phase, a
4105 task attempts to keep its local queue short but not totally
4106 empty, so that entries are available for stealing by other
4107 tasks. Only when there is no more work, a task will totally
4108 drain its local queue.
4110 (3) Global Mark Stack. This handles local queue overflow. During
4111 marking only sets of entries are moved between it and the local
4112 queues, as access to it requires a mutex and more fine-grain
4113 interaction with it which might cause contention. If it
4114 overflows, then the marking phase should restart and iterate
4115 over the bitmap to identify gray objects. Throughout the marking
4116 phase, tasks attempt to keep the global mark stack at a small
4117 length but not totally empty, so that entries are available for
4118 popping by other tasks. Only when there is no more work, tasks
4119 will totally drain the global mark stack.
4121 (4) SATB Buffer Queue. This is where completed SATB buffers are
4122 made available. Buffers are regularly removed from this queue
4123 and scanned for roots, so that the queue doesn't get too
4124 long. During remark, all completed buffers are processed, as
4125 well as the filled in parts of any uncompleted buffers.
4127 The do_marking_step() method tries to abort when the time target
4128 has been reached. There are a few other cases when the
4129 do_marking_step() method also aborts:
4131 (1) When the marking phase has been aborted (after a Full GC).
4133 (2) When a global overflow (on the global stack) has been
4134 triggered. Before the task aborts, it will actually sync up with
4135 the other tasks to ensure that all the marking data structures
4136 (local queues, stacks, fingers etc.) are re-initialized so that
4137 when do_marking_step() completes, the marking phase can
4138 immediately restart.
4140 (3) When enough completed SATB buffers are available. The
4141 do_marking_step() method only tries to drain SATB buffers right
4142 at the beginning. So, if enough buffers are available, the
4143 marking step aborts and the SATB buffers are processed at
4144 the beginning of the next invocation.
4146 (4) To yield. when we have to yield then we abort and yield
4147 right at the end of do_marking_step(). This saves us from a lot
4148 of hassle as, by yielding we might allow a Full GC. If this
4149 happens then objects will be compacted underneath our feet, the
4150 heap might shrink, etc. We save checking for this by just
4151 aborting and doing the yield right at the end.
4153 From the above it follows that the do_marking_step() method should
4154 be called in a loop (or, otherwise, regularly) until it completes.
4156 If a marking step completes without its has_aborted() flag being
4157 true, it means it has completed the current marking phase (and
4158 also all other marking tasks have done so and have all synced up).
4160 A method called regular_clock_call() is invoked "regularly" (in
4161 sub ms intervals) throughout marking. It is this clock method that
4162 checks all the abort conditions which were mentioned above and
4163 decides when the task should abort. A work-based scheme is used to
4164 trigger this clock method: when the number of object words the
4165 marking phase has scanned or the number of references the marking
4166 phase has visited reach a given limit. Additional invocations to
4167 the method clock have been planted in a few other strategic places
4168 too. The initial reason for the clock method was to avoid calling
4169 vtime too regularly, as it is quite expensive. So, once it was in
4170 place, it was natural to piggy-back all the other conditions on it
4171 too and not constantly check them throughout the code.
4173 If do_termination is true then do_marking_step will enter its
4174 termination protocol.
4176 The value of is_serial must be true when do_marking_step is being
4177 called serially (i.e. by the VMThread) and do_marking_step should
4178 skip any synchronization in the termination and overflow code.
4179 Examples include the serial remark code and the serial reference
4180 processing closures.
4182 The value of is_serial must be false when do_marking_step is
4183 being called by any of the worker threads in a work gang.
4184 Examples include the concurrent marking code (CMMarkingTask),
4185 the MT remark code, and the MT reference processing closures.
4187 *****************************************************************************/
4189 void CMTask::do_marking_step(double time_target_ms,
4190 bool do_termination,
4191 bool is_serial) {
4192 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4193 assert(concurrent() == _cm->concurrent(), "they should be the same");
4195 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4196 assert(_task_queues != NULL, "invariant");
4197 assert(_task_queue != NULL, "invariant");
4198 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4200 assert(!_claimed,
4201 "only one thread should claim this task at any one time");
4203 // OK, this doesn't safeguard again all possible scenarios, as it is
4204 // possible for two threads to set the _claimed flag at the same
4205 // time. But it is only for debugging purposes anyway and it will
4206 // catch most problems.
4207 _claimed = true;
4209 _start_time_ms = os::elapsedVTime() * 1000.0;
4210 statsOnly( _interval_start_time_ms = _start_time_ms );
4212 // If do_stealing is true then do_marking_step will attempt to
4213 // steal work from the other CMTasks. It only makes sense to
4214 // enable stealing when the termination protocol is enabled
4215 // and do_marking_step() is not being called serially.
4216 bool do_stealing = do_termination && !is_serial;
4218 double diff_prediction_ms =
4219 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4220 _time_target_ms = time_target_ms - diff_prediction_ms;
4222 // set up the variables that are used in the work-based scheme to
4223 // call the regular clock method
4224 _words_scanned = 0;
4225 _refs_reached = 0;
4226 recalculate_limits();
4228 // clear all flags
4229 clear_has_aborted();
4230 _has_timed_out = false;
4231 _draining_satb_buffers = false;
4233 ++_calls;
4235 if (_cm->verbose_low()) {
4236 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4237 "target = %1.2lfms >>>>>>>>>>",
4238 _worker_id, _calls, _time_target_ms);
4239 }
4241 // Set up the bitmap and oop closures. Anything that uses them is
4242 // eventually called from this method, so it is OK to allocate these
4243 // statically.
4244 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4245 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4246 set_cm_oop_closure(&cm_oop_closure);
4248 if (_cm->has_overflown()) {
4249 // This can happen if the mark stack overflows during a GC pause
4250 // and this task, after a yield point, restarts. We have to abort
4251 // as we need to get into the overflow protocol which happens
4252 // right at the end of this task.
4253 set_has_aborted();
4254 }
4256 // First drain any available SATB buffers. After this, we will not
4257 // look at SATB buffers before the next invocation of this method.
4258 // If enough completed SATB buffers are queued up, the regular clock
4259 // will abort this task so that it restarts.
4260 drain_satb_buffers();
4261 // ...then partially drain the local queue and the global stack
4262 drain_local_queue(true);
4263 drain_global_stack(true);
4265 do {
4266 if (!has_aborted() && _curr_region != NULL) {
4267 // This means that we're already holding on to a region.
4268 assert(_finger != NULL, "if region is not NULL, then the finger "
4269 "should not be NULL either");
4271 // We might have restarted this task after an evacuation pause
4272 // which might have evacuated the region we're holding on to
4273 // underneath our feet. Let's read its limit again to make sure
4274 // that we do not iterate over a region of the heap that
4275 // contains garbage (update_region_limit() will also move
4276 // _finger to the start of the region if it is found empty).
4277 update_region_limit();
4278 // We will start from _finger not from the start of the region,
4279 // as we might be restarting this task after aborting half-way
4280 // through scanning this region. In this case, _finger points to
4281 // the address where we last found a marked object. If this is a
4282 // fresh region, _finger points to start().
4283 MemRegion mr = MemRegion(_finger, _region_limit);
4285 if (_cm->verbose_low()) {
4286 gclog_or_tty->print_cr("[%u] we're scanning part "
4287 "["PTR_FORMAT", "PTR_FORMAT") "
4288 "of region "HR_FORMAT,
4289 _worker_id, p2i(_finger), p2i(_region_limit),
4290 HR_FORMAT_PARAMS(_curr_region));
4291 }
4293 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4294 "humongous regions should go around loop once only");
4296 // Some special cases:
4297 // If the memory region is empty, we can just give up the region.
4298 // If the current region is humongous then we only need to check
4299 // the bitmap for the bit associated with the start of the object,
4300 // scan the object if it's live, and give up the region.
4301 // Otherwise, let's iterate over the bitmap of the part of the region
4302 // that is left.
4303 // If the iteration is successful, give up the region.
4304 if (mr.is_empty()) {
4305 giveup_current_region();
4306 regular_clock_call();
4307 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4308 if (_nextMarkBitMap->isMarked(mr.start())) {
4309 // The object is marked - apply the closure
4310 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4311 bitmap_closure.do_bit(offset);
4312 }
4313 // Even if this task aborted while scanning the humongous object
4314 // we can (and should) give up the current region.
4315 giveup_current_region();
4316 regular_clock_call();
4317 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4318 giveup_current_region();
4319 regular_clock_call();
4320 } else {
4321 assert(has_aborted(), "currently the only way to do so");
4322 // The only way to abort the bitmap iteration is to return
4323 // false from the do_bit() method. However, inside the
4324 // do_bit() method we move the _finger to point to the
4325 // object currently being looked at. So, if we bail out, we
4326 // have definitely set _finger to something non-null.
4327 assert(_finger != NULL, "invariant");
4329 // Region iteration was actually aborted. So now _finger
4330 // points to the address of the object we last scanned. If we
4331 // leave it there, when we restart this task, we will rescan
4332 // the object. It is easy to avoid this. We move the finger by
4333 // enough to point to the next possible object header (the
4334 // bitmap knows by how much we need to move it as it knows its
4335 // granularity).
4336 assert(_finger < _region_limit, "invariant");
4337 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4338 // Check if bitmap iteration was aborted while scanning the last object
4339 if (new_finger >= _region_limit) {
4340 giveup_current_region();
4341 } else {
4342 move_finger_to(new_finger);
4343 }
4344 }
4345 }
4346 // At this point we have either completed iterating over the
4347 // region we were holding on to, or we have aborted.
4349 // We then partially drain the local queue and the global stack.
4350 // (Do we really need this?)
4351 drain_local_queue(true);
4352 drain_global_stack(true);
4354 // Read the note on the claim_region() method on why it might
4355 // return NULL with potentially more regions available for
4356 // claiming and why we have to check out_of_regions() to determine
4357 // whether we're done or not.
4358 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4359 // We are going to try to claim a new region. We should have
4360 // given up on the previous one.
4361 // Separated the asserts so that we know which one fires.
4362 assert(_curr_region == NULL, "invariant");
4363 assert(_finger == NULL, "invariant");
4364 assert(_region_limit == NULL, "invariant");
4365 if (_cm->verbose_low()) {
4366 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4367 }
4368 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4369 if (claimed_region != NULL) {
4370 // Yes, we managed to claim one
4371 statsOnly( ++_regions_claimed );
4373 if (_cm->verbose_low()) {
4374 gclog_or_tty->print_cr("[%u] we successfully claimed "
4375 "region "PTR_FORMAT,
4376 _worker_id, p2i(claimed_region));
4377 }
4379 setup_for_region(claimed_region);
4380 assert(_curr_region == claimed_region, "invariant");
4381 }
4382 // It is important to call the regular clock here. It might take
4383 // a while to claim a region if, for example, we hit a large
4384 // block of empty regions. So we need to call the regular clock
4385 // method once round the loop to make sure it's called
4386 // frequently enough.
4387 regular_clock_call();
4388 }
4390 if (!has_aborted() && _curr_region == NULL) {
4391 assert(_cm->out_of_regions(),
4392 "at this point we should be out of regions");
4393 }
4394 } while ( _curr_region != NULL && !has_aborted());
4396 if (!has_aborted()) {
4397 // We cannot check whether the global stack is empty, since other
4398 // tasks might be pushing objects to it concurrently.
4399 assert(_cm->out_of_regions(),
4400 "at this point we should be out of regions");
4402 if (_cm->verbose_low()) {
4403 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4404 }
4406 // Try to reduce the number of available SATB buffers so that
4407 // remark has less work to do.
4408 drain_satb_buffers();
4409 }
4411 // Since we've done everything else, we can now totally drain the
4412 // local queue and global stack.
4413 drain_local_queue(false);
4414 drain_global_stack(false);
4416 // Attempt at work stealing from other task's queues.
4417 if (do_stealing && !has_aborted()) {
4418 // We have not aborted. This means that we have finished all that
4419 // we could. Let's try to do some stealing...
4421 // We cannot check whether the global stack is empty, since other
4422 // tasks might be pushing objects to it concurrently.
4423 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4424 "only way to reach here");
4426 if (_cm->verbose_low()) {
4427 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4428 }
4430 while (!has_aborted()) {
4431 oop obj;
4432 statsOnly( ++_steal_attempts );
4434 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4435 if (_cm->verbose_medium()) {
4436 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4437 _worker_id, p2i((void*) obj));
4438 }
4440 statsOnly( ++_steals );
4442 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4443 "any stolen object should be marked");
4444 scan_object(obj);
4446 // And since we're towards the end, let's totally drain the
4447 // local queue and global stack.
4448 drain_local_queue(false);
4449 drain_global_stack(false);
4450 } else {
4451 break;
4452 }
4453 }
4454 }
4456 // If we are about to wrap up and go into termination, check if we
4457 // should raise the overflow flag.
4458 if (do_termination && !has_aborted()) {
4459 if (_cm->force_overflow()->should_force()) {
4460 _cm->set_has_overflown();
4461 regular_clock_call();
4462 }
4463 }
4465 // We still haven't aborted. Now, let's try to get into the
4466 // termination protocol.
4467 if (do_termination && !has_aborted()) {
4468 // We cannot check whether the global stack is empty, since other
4469 // tasks might be concurrently pushing objects on it.
4470 // Separated the asserts so that we know which one fires.
4471 assert(_cm->out_of_regions(), "only way to reach here");
4472 assert(_task_queue->size() == 0, "only way to reach here");
4474 if (_cm->verbose_low()) {
4475 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4476 }
4478 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4480 // The CMTask class also extends the TerminatorTerminator class,
4481 // hence its should_exit_termination() method will also decide
4482 // whether to exit the termination protocol or not.
4483 bool finished = (is_serial ||
4484 _cm->terminator()->offer_termination(this));
4485 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4486 _termination_time_ms +=
4487 termination_end_time_ms - _termination_start_time_ms;
4489 if (finished) {
4490 // We're all done.
4492 if (_worker_id == 0) {
4493 // let's allow task 0 to do this
4494 if (concurrent()) {
4495 assert(_cm->concurrent_marking_in_progress(), "invariant");
4496 // we need to set this to false before the next
4497 // safepoint. This way we ensure that the marking phase
4498 // doesn't observe any more heap expansions.
4499 _cm->clear_concurrent_marking_in_progress();
4500 }
4501 }
4503 // We can now guarantee that the global stack is empty, since
4504 // all other tasks have finished. We separated the guarantees so
4505 // that, if a condition is false, we can immediately find out
4506 // which one.
4507 guarantee(_cm->out_of_regions(), "only way to reach here");
4508 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4509 guarantee(_task_queue->size() == 0, "only way to reach here");
4510 guarantee(!_cm->has_overflown(), "only way to reach here");
4511 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4513 if (_cm->verbose_low()) {
4514 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4515 }
4516 } else {
4517 // Apparently there's more work to do. Let's abort this task. It
4518 // will restart it and we can hopefully find more things to do.
4520 if (_cm->verbose_low()) {
4521 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4522 _worker_id);
4523 }
4525 set_has_aborted();
4526 statsOnly( ++_aborted_termination );
4527 }
4528 }
4530 // Mainly for debugging purposes to make sure that a pointer to the
4531 // closure which was statically allocated in this frame doesn't
4532 // escape it by accident.
4533 set_cm_oop_closure(NULL);
4534 double end_time_ms = os::elapsedVTime() * 1000.0;
4535 double elapsed_time_ms = end_time_ms - _start_time_ms;
4536 // Update the step history.
4537 _step_times_ms.add(elapsed_time_ms);
4539 if (has_aborted()) {
4540 // The task was aborted for some reason.
4542 statsOnly( ++_aborted );
4544 if (_has_timed_out) {
4545 double diff_ms = elapsed_time_ms - _time_target_ms;
4546 // Keep statistics of how well we did with respect to hitting
4547 // our target only if we actually timed out (if we aborted for
4548 // other reasons, then the results might get skewed).
4549 _marking_step_diffs_ms.add(diff_ms);
4550 }
4552 if (_cm->has_overflown()) {
4553 // This is the interesting one. We aborted because a global
4554 // overflow was raised. This means we have to restart the
4555 // marking phase and start iterating over regions. However, in
4556 // order to do this we have to make sure that all tasks stop
4557 // what they are doing and re-initialise in a safe manner. We
4558 // will achieve this with the use of two barrier sync points.
4560 if (_cm->verbose_low()) {
4561 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4562 }
4564 if (!is_serial) {
4565 // We only need to enter the sync barrier if being called
4566 // from a parallel context
4567 _cm->enter_first_sync_barrier(_worker_id);
4569 // When we exit this sync barrier we know that all tasks have
4570 // stopped doing marking work. So, it's now safe to
4571 // re-initialise our data structures. At the end of this method,
4572 // task 0 will clear the global data structures.
4573 }
4575 statsOnly( ++_aborted_overflow );
4577 // We clear the local state of this task...
4578 clear_region_fields();
4580 if (!is_serial) {
4581 // ...and enter the second barrier.
4582 _cm->enter_second_sync_barrier(_worker_id);
4583 }
4584 // At this point, if we're during the concurrent phase of
4585 // marking, everything has been re-initialized and we're
4586 // ready to restart.
4587 }
4589 if (_cm->verbose_low()) {
4590 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4591 "elapsed = %1.2lfms <<<<<<<<<<",
4592 _worker_id, _time_target_ms, elapsed_time_ms);
4593 if (_cm->has_aborted()) {
4594 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4595 _worker_id);
4596 }
4597 }
4598 } else {
4599 if (_cm->verbose_low()) {
4600 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4601 "elapsed = %1.2lfms <<<<<<<<<<",
4602 _worker_id, _time_target_ms, elapsed_time_ms);
4603 }
4604 }
4606 _claimed = false;
4607 }
4609 CMTask::CMTask(uint worker_id,
4610 ConcurrentMark* cm,
4611 size_t* marked_bytes,
4612 BitMap* card_bm,
4613 CMTaskQueue* task_queue,
4614 CMTaskQueueSet* task_queues)
4615 : _g1h(G1CollectedHeap::heap()),
4616 _worker_id(worker_id), _cm(cm),
4617 _claimed(false),
4618 _nextMarkBitMap(NULL), _hash_seed(17),
4619 _task_queue(task_queue),
4620 _task_queues(task_queues),
4621 _cm_oop_closure(NULL),
4622 _marked_bytes_array(marked_bytes),
4623 _card_bm(card_bm) {
4624 guarantee(task_queue != NULL, "invariant");
4625 guarantee(task_queues != NULL, "invariant");
4627 statsOnly( _clock_due_to_scanning = 0;
4628 _clock_due_to_marking = 0 );
4630 _marking_step_diffs_ms.add(0.5);
4631 }
4633 // These are formatting macros that are used below to ensure
4634 // consistent formatting. The *_H_* versions are used to format the
4635 // header for a particular value and they should be kept consistent
4636 // with the corresponding macro. Also note that most of the macros add
4637 // the necessary white space (as a prefix) which makes them a bit
4638 // easier to compose.
4640 // All the output lines are prefixed with this string to be able to
4641 // identify them easily in a large log file.
4642 #define G1PPRL_LINE_PREFIX "###"
4644 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4645 #ifdef _LP64
4646 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4647 #else // _LP64
4648 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4649 #endif // _LP64
4651 // For per-region info
4652 #define G1PPRL_TYPE_FORMAT " %-4s"
4653 #define G1PPRL_TYPE_H_FORMAT " %4s"
4654 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4655 #define G1PPRL_BYTE_H_FORMAT " %9s"
4656 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4657 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4659 // For summary info
4660 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4661 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4662 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4663 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4665 G1PrintRegionLivenessInfoClosure::
4666 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4667 : _out(out),
4668 _total_used_bytes(0), _total_capacity_bytes(0),
4669 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4670 _hum_used_bytes(0), _hum_capacity_bytes(0),
4671 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4672 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4673 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4674 MemRegion g1_committed = g1h->g1_committed();
4675 MemRegion g1_reserved = g1h->g1_reserved();
4676 double now = os::elapsedTime();
4678 // Print the header of the output.
4679 _out->cr();
4680 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4681 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4682 G1PPRL_SUM_ADDR_FORMAT("committed")
4683 G1PPRL_SUM_ADDR_FORMAT("reserved")
4684 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4685 p2i(g1_committed.start()), p2i(g1_committed.end()),
4686 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4687 HeapRegion::GrainBytes);
4688 _out->print_cr(G1PPRL_LINE_PREFIX);
4689 _out->print_cr(G1PPRL_LINE_PREFIX
4690 G1PPRL_TYPE_H_FORMAT
4691 G1PPRL_ADDR_BASE_H_FORMAT
4692 G1PPRL_BYTE_H_FORMAT
4693 G1PPRL_BYTE_H_FORMAT
4694 G1PPRL_BYTE_H_FORMAT
4695 G1PPRL_DOUBLE_H_FORMAT
4696 G1PPRL_BYTE_H_FORMAT
4697 G1PPRL_BYTE_H_FORMAT,
4698 "type", "address-range",
4699 "used", "prev-live", "next-live", "gc-eff",
4700 "remset", "code-roots");
4701 _out->print_cr(G1PPRL_LINE_PREFIX
4702 G1PPRL_TYPE_H_FORMAT
4703 G1PPRL_ADDR_BASE_H_FORMAT
4704 G1PPRL_BYTE_H_FORMAT
4705 G1PPRL_BYTE_H_FORMAT
4706 G1PPRL_BYTE_H_FORMAT
4707 G1PPRL_DOUBLE_H_FORMAT
4708 G1PPRL_BYTE_H_FORMAT
4709 G1PPRL_BYTE_H_FORMAT,
4710 "", "",
4711 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4712 "(bytes)", "(bytes)");
4713 }
4715 // It takes as a parameter a reference to one of the _hum_* fields, it
4716 // deduces the corresponding value for a region in a humongous region
4717 // series (either the region size, or what's left if the _hum_* field
4718 // is < the region size), and updates the _hum_* field accordingly.
4719 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4720 size_t bytes = 0;
4721 // The > 0 check is to deal with the prev and next live bytes which
4722 // could be 0.
4723 if (*hum_bytes > 0) {
4724 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4725 *hum_bytes -= bytes;
4726 }
4727 return bytes;
4728 }
4730 // It deduces the values for a region in a humongous region series
4731 // from the _hum_* fields and updates those accordingly. It assumes
4732 // that that _hum_* fields have already been set up from the "starts
4733 // humongous" region and we visit the regions in address order.
4734 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4735 size_t* capacity_bytes,
4736 size_t* prev_live_bytes,
4737 size_t* next_live_bytes) {
4738 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4739 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4740 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4741 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4742 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4743 }
4745 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4746 const char* type = "";
4747 HeapWord* bottom = r->bottom();
4748 HeapWord* end = r->end();
4749 size_t capacity_bytes = r->capacity();
4750 size_t used_bytes = r->used();
4751 size_t prev_live_bytes = r->live_bytes();
4752 size_t next_live_bytes = r->next_live_bytes();
4753 double gc_eff = r->gc_efficiency();
4754 size_t remset_bytes = r->rem_set()->mem_size();
4755 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4757 if (r->used() == 0) {
4758 type = "FREE";
4759 } else if (r->is_survivor()) {
4760 type = "SURV";
4761 } else if (r->is_young()) {
4762 type = "EDEN";
4763 } else if (r->startsHumongous()) {
4764 type = "HUMS";
4766 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4767 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4768 "they should have been zeroed after the last time we used them");
4769 // Set up the _hum_* fields.
4770 _hum_capacity_bytes = capacity_bytes;
4771 _hum_used_bytes = used_bytes;
4772 _hum_prev_live_bytes = prev_live_bytes;
4773 _hum_next_live_bytes = next_live_bytes;
4774 get_hum_bytes(&used_bytes, &capacity_bytes,
4775 &prev_live_bytes, &next_live_bytes);
4776 end = bottom + HeapRegion::GrainWords;
4777 } else if (r->continuesHumongous()) {
4778 type = "HUMC";
4779 get_hum_bytes(&used_bytes, &capacity_bytes,
4780 &prev_live_bytes, &next_live_bytes);
4781 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4782 } else {
4783 type = "OLD";
4784 }
4786 _total_used_bytes += used_bytes;
4787 _total_capacity_bytes += capacity_bytes;
4788 _total_prev_live_bytes += prev_live_bytes;
4789 _total_next_live_bytes += next_live_bytes;
4790 _total_remset_bytes += remset_bytes;
4791 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4793 // Print a line for this particular region.
4794 _out->print_cr(G1PPRL_LINE_PREFIX
4795 G1PPRL_TYPE_FORMAT
4796 G1PPRL_ADDR_BASE_FORMAT
4797 G1PPRL_BYTE_FORMAT
4798 G1PPRL_BYTE_FORMAT
4799 G1PPRL_BYTE_FORMAT
4800 G1PPRL_DOUBLE_FORMAT
4801 G1PPRL_BYTE_FORMAT
4802 G1PPRL_BYTE_FORMAT,
4803 type, p2i(bottom), p2i(end),
4804 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4805 remset_bytes, strong_code_roots_bytes);
4807 return false;
4808 }
4810 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4811 // add static memory usages to remembered set sizes
4812 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4813 // Print the footer of the output.
4814 _out->print_cr(G1PPRL_LINE_PREFIX);
4815 _out->print_cr(G1PPRL_LINE_PREFIX
4816 " SUMMARY"
4817 G1PPRL_SUM_MB_FORMAT("capacity")
4818 G1PPRL_SUM_MB_PERC_FORMAT("used")
4819 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4820 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4821 G1PPRL_SUM_MB_FORMAT("remset")
4822 G1PPRL_SUM_MB_FORMAT("code-roots"),
4823 bytes_to_mb(_total_capacity_bytes),
4824 bytes_to_mb(_total_used_bytes),
4825 perc(_total_used_bytes, _total_capacity_bytes),
4826 bytes_to_mb(_total_prev_live_bytes),
4827 perc(_total_prev_live_bytes, _total_capacity_bytes),
4828 bytes_to_mb(_total_next_live_bytes),
4829 perc(_total_next_live_bytes, _total_capacity_bytes),
4830 bytes_to_mb(_total_remset_bytes),
4831 bytes_to_mb(_total_strong_code_roots_bytes));
4832 _out->cr();
4833 }