1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/symbolTable.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/codeCache.hpp"
31 #include "gc/cms/cmsCollectorPolicy.hpp"
32 #include "gc/cms/cmsGCStats.hpp"
33 #include "gc/cms/cmsHeap.hpp"
34 #include "gc/cms/cmsOopClosures.inline.hpp"
35 #include "gc/cms/compactibleFreeListSpace.hpp"
36 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
37 #include "gc/cms/concurrentMarkSweepThread.hpp"
38 #include "gc/cms/parNewGeneration.hpp"
39 #include "gc/cms/promotionInfo.inline.hpp"
40 #include "gc/cms/vmCMSOperations.hpp"
41 #include "gc/serial/genMarkSweep.hpp"
42 #include "gc/serial/tenuredGeneration.hpp"
43 #include "gc/shared/adaptiveSizePolicy.hpp"
44 #include "gc/shared/cardGeneration.inline.hpp"
45 #include "gc/shared/cardTableRS.hpp"
46 #include "gc/shared/collectedHeap.inline.hpp"
47 #include "gc/shared/collectorCounters.hpp"
48 #include "gc/shared/collectorPolicy.hpp"
49 #include "gc/shared/gcLocker.hpp"
50 #include "gc/shared/gcPolicyCounters.hpp"
51 #include "gc/shared/gcTimer.hpp"
52 #include "gc/shared/gcTrace.hpp"
53 #include "gc/shared/gcTraceTime.inline.hpp"
54 #include "gc/shared/genCollectedHeap.hpp"
55 #include "gc/shared/genOopClosures.inline.hpp"
56 #include "gc/shared/isGCActiveMark.hpp"
57 #include "gc/shared/oopStorageParState.hpp"
58 #include "gc/shared/referencePolicy.hpp"
59 #include "gc/shared/space.inline.hpp"
60 #include "gc/shared/strongRootsScope.hpp"
61 #include "gc/shared/taskqueue.inline.hpp"
62 #include "gc/shared/weakProcessor.hpp"
63 #include "logging/log.hpp"
64 #include "logging/logStream.hpp"
65 #include "memory/allocation.hpp"
66 #include "memory/binaryTreeDictionary.inline.hpp"
67 #include "memory/iterator.inline.hpp"
68 #include "memory/padded.hpp"
69 #include "memory/resourceArea.hpp"
70 #include "oops/access.inline.hpp"
71 #include "oops/oop.inline.hpp"
72 #include "prims/jvmtiExport.hpp"
73 #include "runtime/atomic.hpp"
74 #include "runtime/flags/flagSetting.hpp"
75 #include "runtime/globals_extension.hpp"
76 #include "runtime/handles.inline.hpp"
77 #include "runtime/java.hpp"
78 #include "runtime/orderAccess.hpp"
79 #include "runtime/timer.hpp"
80 #include "runtime/vmThread.hpp"
81 #include "services/memoryService.hpp"
82 #include "services/runtimeService.hpp"
83 #include "utilities/align.hpp"
84 #include "utilities/stack.inline.hpp"
85
86 // statics
87 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
88 bool CMSCollector::_full_gc_requested = false;
89 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
90
91 //////////////////////////////////////////////////////////////////
92 // In support of CMS/VM thread synchronization
93 //////////////////////////////////////////////////////////////////
94 // We split use of the CGC_lock into 2 "levels".
95 // The low-level locking is of the usual CGC_lock monitor. We introduce
96 // a higher level "token" (hereafter "CMS token") built on top of the
97 // low level monitor (hereafter "CGC lock").
98 // The token-passing protocol gives priority to the VM thread. The
99 // CMS-lock doesn't provide any fairness guarantees, but clients
100 // should ensure that it is only held for very short, bounded
101 // durations.
102 //
103 // When either of the CMS thread or the VM thread is involved in
104 // collection operations during which it does not want the other
105 // thread to interfere, it obtains the CMS token.
106 //
107 // If either thread tries to get the token while the other has
108 // it, that thread waits. However, if the VM thread and CMS thread
109 // both want the token, then the VM thread gets priority while the
110 // CMS thread waits. This ensures, for instance, that the "concurrent"
111 // phases of the CMS thread's work do not block out the VM thread
112 // for long periods of time as the CMS thread continues to hog
113 // the token. (See bug 4616232).
114 //
115 // The baton-passing functions are, however, controlled by the
116 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
117 // and here the low-level CMS lock, not the high level token,
118 // ensures mutual exclusion.
119 //
120 // Two important conditions that we have to satisfy:
121 // 1. if a thread does a low-level wait on the CMS lock, then it
122 // relinquishes the CMS token if it were holding that token
123 // when it acquired the low-level CMS lock.
124 // 2. any low-level notifications on the low-level lock
125 // should only be sent when a thread has relinquished the token.
126 //
127 // In the absence of either property, we'd have potential deadlock.
128 //
129 // We protect each of the CMS (concurrent and sequential) phases
130 // with the CMS _token_, not the CMS _lock_.
131 //
132 // The only code protected by CMS lock is the token acquisition code
133 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
134 // baton-passing code.
135 //
136 // Unfortunately, i couldn't come up with a good abstraction to factor and
137 // hide the naked CGC_lock manipulation in the baton-passing code
138 // further below. That's something we should try to do. Also, the proof
139 // of correctness of this 2-level locking scheme is far from obvious,
140 // and potentially quite slippery. We have an uneasy suspicion, for instance,
141 // that there may be a theoretical possibility of delay/starvation in the
142 // low-level lock/wait/notify scheme used for the baton-passing because of
143 // potential interference with the priority scheme embodied in the
144 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
145 // invocation further below and marked with "XXX 20011219YSR".
146 // Indeed, as we note elsewhere, this may become yet more slippery
147 // in the presence of multiple CMS and/or multiple VM threads. XXX
148
149 class CMSTokenSync: public StackObj {
150 private:
151 bool _is_cms_thread;
152 public:
153 CMSTokenSync(bool is_cms_thread):
154 _is_cms_thread(is_cms_thread) {
155 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
156 "Incorrect argument to constructor");
157 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
158 }
159
160 ~CMSTokenSync() {
161 assert(_is_cms_thread ?
162 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
163 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
164 "Incorrect state");
165 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
166 }
167 };
168
169 // Convenience class that does a CMSTokenSync, and then acquires
170 // upto three locks.
171 class CMSTokenSyncWithLocks: public CMSTokenSync {
172 private:
173 // Note: locks are acquired in textual declaration order
174 // and released in the opposite order
175 MutexLockerEx _locker1, _locker2, _locker3;
176 public:
177 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
178 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
179 CMSTokenSync(is_cms_thread),
180 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
181 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
182 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
183 { }
184 };
185
186
187 //////////////////////////////////////////////////////////////////
188 // Concurrent Mark-Sweep Generation /////////////////////////////
189 //////////////////////////////////////////////////////////////////
190
191 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
192
193 // This struct contains per-thread things necessary to support parallel
194 // young-gen collection.
195 class CMSParGCThreadState: public CHeapObj<mtGC> {
196 public:
197 CompactibleFreeListSpaceLAB lab;
198 PromotionInfo promo;
199
200 // Constructor.
201 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
202 promo.setSpace(cfls);
203 }
204 };
205
206 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
207 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
208 CardGeneration(rs, initial_byte_size, ct),
209 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
210 _did_compact(false)
211 {
212 HeapWord* bottom = (HeapWord*) _virtual_space.low();
213 HeapWord* end = (HeapWord*) _virtual_space.high();
214
215 _direct_allocated_words = 0;
216 NOT_PRODUCT(
217 _numObjectsPromoted = 0;
218 _numWordsPromoted = 0;
219 _numObjectsAllocated = 0;
220 _numWordsAllocated = 0;
221 )
222
223 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
224 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
225 _cmsSpace->_old_gen = this;
226
227 _gc_stats = new CMSGCStats();
228
229 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
230 // offsets match. The ability to tell free chunks from objects
231 // depends on this property.
232 debug_only(
233 FreeChunk* junk = NULL;
234 assert(UseCompressedClassPointers ||
235 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
236 "Offset of FreeChunk::_prev within FreeChunk must match"
237 " that of OopDesc::_klass within OopDesc");
238 )
239
240 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
241 for (uint i = 0; i < ParallelGCThreads; i++) {
242 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
243 }
244
245 _incremental_collection_failed = false;
246 // The "dilatation_factor" is the expansion that can occur on
247 // account of the fact that the minimum object size in the CMS
248 // generation may be larger than that in, say, a contiguous young
249 // generation.
250 // Ideally, in the calculation below, we'd compute the dilatation
251 // factor as: MinChunkSize/(promoting_gen's min object size)
252 // Since we do not have such a general query interface for the
253 // promoting generation, we'll instead just use the minimum
254 // object size (which today is a header's worth of space);
255 // note that all arithmetic is in units of HeapWords.
256 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
257 assert(_dilatation_factor >= 1.0, "from previous assert");
258 }
259
260
261 // The field "_initiating_occupancy" represents the occupancy percentage
262 // at which we trigger a new collection cycle. Unless explicitly specified
263 // via CMSInitiatingOccupancyFraction (argument "io" below), it
264 // is calculated by:
265 //
266 // Let "f" be MinHeapFreeRatio in
267 //
268 // _initiating_occupancy = 100-f +
269 // f * (CMSTriggerRatio/100)
270 // where CMSTriggerRatio is the argument "tr" below.
271 //
272 // That is, if we assume the heap is at its desired maximum occupancy at the
273 // end of a collection, we let CMSTriggerRatio of the (purported) free
274 // space be allocated before initiating a new collection cycle.
275 //
276 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
277 assert(io <= 100 && tr <= 100, "Check the arguments");
278 if (io >= 0) {
279 _initiating_occupancy = (double)io / 100.0;
280 } else {
281 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
282 (double)(tr * MinHeapFreeRatio) / 100.0)
283 / 100.0;
284 }
285 }
286
287 void ConcurrentMarkSweepGeneration::ref_processor_init() {
288 assert(collector() != NULL, "no collector");
289 collector()->ref_processor_init();
290 }
291
292 void CMSCollector::ref_processor_init() {
293 if (_ref_processor == NULL) {
294 // Allocate and initialize a reference processor
295 _ref_processor =
296 new ReferenceProcessor(&_span_based_discoverer,
297 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
298 ParallelGCThreads, // mt processing degree
299 _cmsGen->refs_discovery_is_mt(), // mt discovery
300 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
301 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
302 &_is_alive_closure); // closure for liveness info
303 // Initialize the _ref_processor field of CMSGen
304 _cmsGen->set_ref_processor(_ref_processor);
305
306 }
307 }
308
309 AdaptiveSizePolicy* CMSCollector::size_policy() {
310 return CMSHeap::heap()->size_policy();
311 }
312
313 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
314
315 const char* gen_name = "old";
316 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy();
317 // Generation Counters - generation 1, 1 subspace
318 _gen_counters = new GenerationCounters(gen_name, 1, 1,
319 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
320
321 _space_counters = new GSpaceCounters(gen_name, 0,
322 _virtual_space.reserved_size(),
323 this, _gen_counters);
324 }
325
326 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
327 _cms_gen(cms_gen)
328 {
329 assert(alpha <= 100, "bad value");
330 _saved_alpha = alpha;
331
332 // Initialize the alphas to the bootstrap value of 100.
333 _gc0_alpha = _cms_alpha = 100;
334
335 _cms_begin_time.update();
336 _cms_end_time.update();
337
338 _gc0_duration = 0.0;
339 _gc0_period = 0.0;
340 _gc0_promoted = 0;
341
342 _cms_duration = 0.0;
343 _cms_period = 0.0;
344 _cms_allocated = 0;
345
346 _cms_used_at_gc0_begin = 0;
347 _cms_used_at_gc0_end = 0;
348 _allow_duty_cycle_reduction = false;
349 _valid_bits = 0;
350 }
351
352 double CMSStats::cms_free_adjustment_factor(size_t free) const {
353 // TBD: CR 6909490
354 return 1.0;
355 }
356
357 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
358 }
359
360 // If promotion failure handling is on use
361 // the padded average size of the promotion for each
362 // young generation collection.
363 double CMSStats::time_until_cms_gen_full() const {
364 size_t cms_free = _cms_gen->cmsSpace()->free();
365 CMSHeap* heap = CMSHeap::heap();
366 size_t expected_promotion = MIN2(heap->young_gen()->capacity(),
367 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
368 if (cms_free > expected_promotion) {
369 // Start a cms collection if there isn't enough space to promote
370 // for the next young collection. Use the padded average as
371 // a safety factor.
372 cms_free -= expected_promotion;
373
374 // Adjust by the safety factor.
375 double cms_free_dbl = (double)cms_free;
376 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
377 // Apply a further correction factor which tries to adjust
378 // for recent occurance of concurrent mode failures.
379 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
380 cms_free_dbl = cms_free_dbl * cms_adjustment;
381
382 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
383 cms_free, expected_promotion);
384 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
385 // Add 1 in case the consumption rate goes to zero.
386 return cms_free_dbl / (cms_consumption_rate() + 1.0);
387 }
388 return 0.0;
389 }
390
391 // Compare the duration of the cms collection to the
392 // time remaining before the cms generation is empty.
393 // Note that the time from the start of the cms collection
394 // to the start of the cms sweep (less than the total
395 // duration of the cms collection) can be used. This
396 // has been tried and some applications experienced
397 // promotion failures early in execution. This was
398 // possibly because the averages were not accurate
399 // enough at the beginning.
400 double CMSStats::time_until_cms_start() const {
401 // We add "gc0_period" to the "work" calculation
402 // below because this query is done (mostly) at the
403 // end of a scavenge, so we need to conservatively
404 // account for that much possible delay
405 // in the query so as to avoid concurrent mode failures
406 // due to starting the collection just a wee bit too
407 // late.
408 double work = cms_duration() + gc0_period();
409 double deadline = time_until_cms_gen_full();
410 // If a concurrent mode failure occurred recently, we want to be
411 // more conservative and halve our expected time_until_cms_gen_full()
412 if (work > deadline) {
413 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
414 cms_duration(), gc0_period(), time_until_cms_gen_full());
415 return 0.0;
416 }
417 return work - deadline;
418 }
419
420 #ifndef PRODUCT
421 void CMSStats::print_on(outputStream *st) const {
422 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
423 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
424 gc0_duration(), gc0_period(), gc0_promoted());
425 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
426 cms_duration(), cms_period(), cms_allocated());
427 st->print(",cms_since_beg=%g,cms_since_end=%g",
428 cms_time_since_begin(), cms_time_since_end());
429 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
430 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
431
432 if (valid()) {
433 st->print(",promo_rate=%g,cms_alloc_rate=%g",
434 promotion_rate(), cms_allocation_rate());
435 st->print(",cms_consumption_rate=%g,time_until_full=%g",
436 cms_consumption_rate(), time_until_cms_gen_full());
437 }
438 st->cr();
439 }
440 #endif // #ifndef PRODUCT
441
442 CMSCollector::CollectorState CMSCollector::_collectorState =
443 CMSCollector::Idling;
444 bool CMSCollector::_foregroundGCIsActive = false;
445 bool CMSCollector::_foregroundGCShouldWait = false;
446
447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
448 CardTableRS* ct,
449 ConcurrentMarkSweepPolicy* cp):
450 _cmsGen(cmsGen),
451 // Adjust span to cover old (cms) gen
452 _span(cmsGen->reserved()),
453 _ct(ct),
454 _span_based_discoverer(_span),
455 _ref_processor(NULL), // will be set later
456 _conc_workers(NULL), // may be set later
457 _abort_preclean(false),
458 _start_sampling(false),
459 _between_prologue_and_epilogue(false),
460 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
461 _modUnionTable((CardTable::card_shift - LogHeapWordSize),
462 -1 /* lock-free */, "No_lock" /* dummy */),
463 _modUnionClosurePar(&_modUnionTable),
464 // Construct the is_alive_closure with _span & markBitMap
465 _is_alive_closure(_span, &_markBitMap),
466 _restart_addr(NULL),
467 _overflow_list(NULL),
468 _stats(cmsGen),
469 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
470 //verify that this lock should be acquired with safepoint check.
471 Monitor::_safepoint_check_sometimes)),
472 _eden_chunk_array(NULL), // may be set in ctor body
473 _eden_chunk_capacity(0), // -- ditto --
474 _eden_chunk_index(0), // -- ditto --
475 _survivor_plab_array(NULL), // -- ditto --
476 _survivor_chunk_array(NULL), // -- ditto --
477 _survivor_chunk_capacity(0), // -- ditto --
478 _survivor_chunk_index(0), // -- ditto --
479 _ser_pmc_preclean_ovflw(0),
480 _ser_kac_preclean_ovflw(0),
481 _ser_pmc_remark_ovflw(0),
482 _par_pmc_remark_ovflw(0),
483 _ser_kac_ovflw(0),
484 _par_kac_ovflw(0),
485 #ifndef PRODUCT
486 _num_par_pushes(0),
487 #endif
488 _collection_count_start(0),
489 _verifying(false),
490 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
491 _completed_initialization(false),
492 _collector_policy(cp),
493 _should_unload_classes(CMSClassUnloadingEnabled),
494 _concurrent_cycles_since_last_unload(0),
495 _roots_scanning_options(GenCollectedHeap::SO_None),
496 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
497 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
498 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
499 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
500 _cms_start_registered(false)
501 {
502 // Now expand the span and allocate the collection support structures
503 // (MUT, marking bit map etc.) to cover both generations subject to
504 // collection.
505
506 // For use by dirty card to oop closures.
507 _cmsGen->cmsSpace()->set_collector(this);
508
509 // Allocate MUT and marking bit map
510 {
511 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
512 if (!_markBitMap.allocate(_span)) {
513 log_warning(gc)("Failed to allocate CMS Bit Map");
514 return;
515 }
516 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
517 }
518 {
519 _modUnionTable.allocate(_span);
520 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
521 }
522
523 if (!_markStack.allocate(MarkStackSize)) {
524 log_warning(gc)("Failed to allocate CMS Marking Stack");
525 return;
526 }
527
528 // Support for multi-threaded concurrent phases
529 if (CMSConcurrentMTEnabled) {
530 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
531 // just for now
532 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
533 }
534 if (ConcGCThreads > 1) {
535 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
536 ConcGCThreads, true);
537 if (_conc_workers == NULL) {
538 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
539 CMSConcurrentMTEnabled = false;
540 } else {
541 _conc_workers->initialize_workers();
542 }
543 } else {
544 CMSConcurrentMTEnabled = false;
545 }
546 }
547 if (!CMSConcurrentMTEnabled) {
548 ConcGCThreads = 0;
549 } else {
550 // Turn off CMSCleanOnEnter optimization temporarily for
551 // the MT case where it's not fixed yet; see 6178663.
552 CMSCleanOnEnter = false;
553 }
554 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
555 "Inconsistency");
556 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
557 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
558
559 // Parallel task queues; these are shared for the
560 // concurrent and stop-world phases of CMS, but
561 // are not shared with parallel scavenge (ParNew).
562 {
563 uint i;
564 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
565
566 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
567 || ParallelRefProcEnabled)
568 && num_queues > 0) {
569 _task_queues = new OopTaskQueueSet(num_queues);
570 if (_task_queues == NULL) {
571 log_warning(gc)("task_queues allocation failure.");
572 return;
573 }
574 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
575 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
576 for (i = 0; i < num_queues; i++) {
577 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
578 if (q == NULL) {
579 log_warning(gc)("work_queue allocation failure.");
580 return;
581 }
582 _task_queues->register_queue(i, q);
583 }
584 for (i = 0; i < num_queues; i++) {
585 _task_queues->queue(i)->initialize();
586 _hash_seed[i] = 17; // copied from ParNew
587 }
588 }
589 }
590
591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
592
593 // Clip CMSBootstrapOccupancy between 0 and 100.
594 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
595
596 // Now tell CMS generations the identity of their collector
597 ConcurrentMarkSweepGeneration::set_collector(this);
598
599 // Create & start a CMS thread for this CMS collector
600 _cmsThread = ConcurrentMarkSweepThread::start(this);
601 assert(cmsThread() != NULL, "CMS Thread should have been created");
602 assert(cmsThread()->collector() == this,
603 "CMS Thread should refer to this gen");
604 assert(CGC_lock != NULL, "Where's the CGC_lock?");
605
606 // Support for parallelizing young gen rescan
607 CMSHeap* heap = CMSHeap::heap();
608 _young_gen = heap->young_gen();
609 if (heap->supports_inline_contig_alloc()) {
610 _top_addr = heap->top_addr();
611 _end_addr = heap->end_addr();
612 assert(_young_gen != NULL, "no _young_gen");
613 _eden_chunk_index = 0;
614 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
615 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
616 }
617
618 // Support for parallelizing survivor space rescan
619 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
620 const size_t max_plab_samples =
621 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
622
623 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
624 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
625 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
626 _survivor_chunk_capacity = max_plab_samples;
627 for (uint i = 0; i < ParallelGCThreads; i++) {
628 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
629 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
630 assert(cur->end() == 0, "Should be 0");
631 assert(cur->array() == vec, "Should be vec");
632 assert(cur->capacity() == max_plab_samples, "Error");
633 }
634 }
635
636 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
637 _gc_counters = new CollectorCounters("CMS", 1);
638 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2);
639 _completed_initialization = true;
640 _inter_sweep_timer.start(); // start of time
641 }
642
643 const char* ConcurrentMarkSweepGeneration::name() const {
644 return "concurrent mark-sweep generation";
645 }
646 void ConcurrentMarkSweepGeneration::update_counters() {
647 if (UsePerfData) {
648 _space_counters->update_all();
649 _gen_counters->update_all();
650 }
651 }
652
653 // this is an optimized version of update_counters(). it takes the
654 // used value as a parameter rather than computing it.
655 //
656 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
657 if (UsePerfData) {
658 _space_counters->update_used(used);
659 _space_counters->update_capacity();
660 _gen_counters->update_all();
661 }
662 }
663
664 void ConcurrentMarkSweepGeneration::print() const {
665 Generation::print();
666 cmsSpace()->print();
667 }
668
669 #ifndef PRODUCT
670 void ConcurrentMarkSweepGeneration::print_statistics() {
671 cmsSpace()->printFLCensus(0);
672 }
673 #endif
674
675 size_t
676 ConcurrentMarkSweepGeneration::contiguous_available() const {
677 // dld proposes an improvement in precision here. If the committed
678 // part of the space ends in a free block we should add that to
679 // uncommitted size in the calculation below. Will make this
680 // change later, staying with the approximation below for the
681 // time being. -- ysr.
682 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
683 }
684
685 size_t
686 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
687 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
688 }
689
690 size_t ConcurrentMarkSweepGeneration::max_available() const {
691 return free() + _virtual_space.uncommitted_size();
692 }
693
694 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
695 size_t available = max_available();
696 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
697 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
698 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
699 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
700 return res;
701 }
702
703 // At a promotion failure dump information on block layout in heap
704 // (cms old generation).
705 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
706 Log(gc, promotion) log;
707 if (log.is_trace()) {
708 LogStream ls(log.trace());
709 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
710 }
711 }
712
713 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
714 // Clear the promotion information. These pointers can be adjusted
715 // along with all the other pointers into the heap but
716 // compaction is expected to be a rare event with
717 // a heap using cms so don't do it without seeing the need.
718 for (uint i = 0; i < ParallelGCThreads; i++) {
719 _par_gc_thread_states[i]->promo.reset();
720 }
721 }
722
723 void ConcurrentMarkSweepGeneration::compute_new_size() {
724 assert_locked_or_safepoint(Heap_lock);
725
726 // If incremental collection failed, we just want to expand
727 // to the limit.
728 if (incremental_collection_failed()) {
729 clear_incremental_collection_failed();
730 grow_to_reserved();
731 return;
732 }
733
734 // The heap has been compacted but not reset yet.
735 // Any metric such as free() or used() will be incorrect.
736
737 CardGeneration::compute_new_size();
738
739 // Reset again after a possible resizing
740 if (did_compact()) {
741 cmsSpace()->reset_after_compaction();
742 }
743 }
744
745 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
746 assert_locked_or_safepoint(Heap_lock);
747
748 // If incremental collection failed, we just want to expand
749 // to the limit.
750 if (incremental_collection_failed()) {
751 clear_incremental_collection_failed();
752 grow_to_reserved();
753 return;
754 }
755
756 double free_percentage = ((double) free()) / capacity();
757 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
758 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
759
760 // compute expansion delta needed for reaching desired free percentage
761 if (free_percentage < desired_free_percentage) {
762 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
763 assert(desired_capacity >= capacity(), "invalid expansion size");
764 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
765 Log(gc) log;
766 if (log.is_trace()) {
767 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
768 log.trace("From compute_new_size: ");
769 log.trace(" Free fraction %f", free_percentage);
770 log.trace(" Desired free fraction %f", desired_free_percentage);
771 log.trace(" Maximum free fraction %f", maximum_free_percentage);
772 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000);
773 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
774 CMSHeap* heap = CMSHeap::heap();
775 size_t young_size = heap->young_gen()->capacity();
776 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000);
777 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
778 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
779 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
780 }
781 // safe if expansion fails
782 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
783 log.trace(" Expanded free fraction %f", ((double) free()) / capacity());
784 } else {
785 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
786 assert(desired_capacity <= capacity(), "invalid expansion size");
787 size_t shrink_bytes = capacity() - desired_capacity;
788 // Don't shrink unless the delta is greater than the minimum shrink we want
789 if (shrink_bytes >= MinHeapDeltaBytes) {
790 shrink_free_list_by(shrink_bytes);
791 }
792 }
793 }
794
795 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
796 return cmsSpace()->freelistLock();
797 }
798
799 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
800 CMSSynchronousYieldRequest yr;
801 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
802 return have_lock_and_allocate(size, tlab);
803 }
804
805 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
806 bool tlab /* ignored */) {
807 assert_lock_strong(freelistLock());
808 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
809 HeapWord* res = cmsSpace()->allocate(adjustedSize);
810 // Allocate the object live (grey) if the background collector has
811 // started marking. This is necessary because the marker may
812 // have passed this address and consequently this object will
813 // not otherwise be greyed and would be incorrectly swept up.
814 // Note that if this object contains references, the writing
815 // of those references will dirty the card containing this object
816 // allowing the object to be blackened (and its references scanned)
817 // either during a preclean phase or at the final checkpoint.
818 if (res != NULL) {
819 // We may block here with an uninitialized object with
820 // its mark-bit or P-bits not yet set. Such objects need
821 // to be safely navigable by block_start().
822 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
823 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
824 collector()->direct_allocated(res, adjustedSize);
825 _direct_allocated_words += adjustedSize;
826 // allocation counters
827 NOT_PRODUCT(
828 _numObjectsAllocated++;
829 _numWordsAllocated += (int)adjustedSize;
830 )
831 }
832 return res;
833 }
834
835 // In the case of direct allocation by mutators in a generation that
836 // is being concurrently collected, the object must be allocated
837 // live (grey) if the background collector has started marking.
838 // This is necessary because the marker may
839 // have passed this address and consequently this object will
840 // not otherwise be greyed and would be incorrectly swept up.
841 // Note that if this object contains references, the writing
842 // of those references will dirty the card containing this object
843 // allowing the object to be blackened (and its references scanned)
844 // either during a preclean phase or at the final checkpoint.
845 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
846 assert(_markBitMap.covers(start, size), "Out of bounds");
847 if (_collectorState >= Marking) {
848 MutexLockerEx y(_markBitMap.lock(),
849 Mutex::_no_safepoint_check_flag);
850 // [see comments preceding SweepClosure::do_blk() below for details]
851 //
852 // Can the P-bits be deleted now? JJJ
853 //
854 // 1. need to mark the object as live so it isn't collected
855 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
856 // 3. need to mark the end of the object so marking, precleaning or sweeping
857 // can skip over uninitialized or unparsable objects. An allocated
858 // object is considered uninitialized for our purposes as long as
859 // its klass word is NULL. All old gen objects are parsable
860 // as soon as they are initialized.)
861 _markBitMap.mark(start); // object is live
862 _markBitMap.mark(start + 1); // object is potentially uninitialized?
863 _markBitMap.mark(start + size - 1);
864 // mark end of object
865 }
866 // check that oop looks uninitialized
867 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
868 }
869
870 void CMSCollector::promoted(bool par, HeapWord* start,
871 bool is_obj_array, size_t obj_size) {
872 assert(_markBitMap.covers(start), "Out of bounds");
873 // See comment in direct_allocated() about when objects should
874 // be allocated live.
875 if (_collectorState >= Marking) {
876 // we already hold the marking bit map lock, taken in
877 // the prologue
878 if (par) {
879 _markBitMap.par_mark(start);
880 } else {
881 _markBitMap.mark(start);
882 }
883 // We don't need to mark the object as uninitialized (as
884 // in direct_allocated above) because this is being done with the
885 // world stopped and the object will be initialized by the
886 // time the marking, precleaning or sweeping get to look at it.
887 // But see the code for copying objects into the CMS generation,
888 // where we need to ensure that concurrent readers of the
889 // block offset table are able to safely navigate a block that
890 // is in flux from being free to being allocated (and in
891 // transition while being copied into) and subsequently
892 // becoming a bona-fide object when the copy/promotion is complete.
893 assert(SafepointSynchronize::is_at_safepoint(),
894 "expect promotion only at safepoints");
895
896 if (_collectorState < Sweeping) {
897 // Mark the appropriate cards in the modUnionTable, so that
898 // this object gets scanned before the sweep. If this is
899 // not done, CMS generation references in the object might
900 // not get marked.
901 // For the case of arrays, which are otherwise precisely
902 // marked, we need to dirty the entire array, not just its head.
903 if (is_obj_array) {
904 // The [par_]mark_range() method expects mr.end() below to
905 // be aligned to the granularity of a bit's representation
906 // in the heap. In the case of the MUT below, that's a
907 // card size.
908 MemRegion mr(start,
909 align_up(start + obj_size,
910 CardTable::card_size /* bytes */));
911 if (par) {
912 _modUnionTable.par_mark_range(mr);
913 } else {
914 _modUnionTable.mark_range(mr);
915 }
916 } else { // not an obj array; we can just mark the head
917 if (par) {
918 _modUnionTable.par_mark(start);
919 } else {
920 _modUnionTable.mark(start);
921 }
922 }
923 }
924 }
925 }
926
927 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
928 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
929 // allocate, copy and if necessary update promoinfo --
930 // delegate to underlying space.
931 assert_lock_strong(freelistLock());
932
933 #ifndef PRODUCT
934 if (CMSHeap::heap()->promotion_should_fail()) {
935 return NULL;
936 }
937 #endif // #ifndef PRODUCT
938
939 oop res = _cmsSpace->promote(obj, obj_size);
940 if (res == NULL) {
941 // expand and retry
942 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
943 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
944 // Since this is the old generation, we don't try to promote
945 // into a more senior generation.
946 res = _cmsSpace->promote(obj, obj_size);
947 }
948 if (res != NULL) {
949 // See comment in allocate() about when objects should
950 // be allocated live.
951 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below");
952 collector()->promoted(false, // Not parallel
953 (HeapWord*)res, obj->is_objArray(), obj_size);
954 // promotion counters
955 NOT_PRODUCT(
956 _numObjectsPromoted++;
957 _numWordsPromoted +=
958 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
959 )
960 }
961 return res;
962 }
963
964
965 // IMPORTANT: Notes on object size recognition in CMS.
966 // ---------------------------------------------------
967 // A block of storage in the CMS generation is always in
968 // one of three states. A free block (FREE), an allocated
969 // object (OBJECT) whose size() method reports the correct size,
970 // and an intermediate state (TRANSIENT) in which its size cannot
971 // be accurately determined.
972 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
973 // -----------------------------------------------------
974 // FREE: klass_word & 1 == 1; mark_word holds block size
975 //
976 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
977 // obj->size() computes correct size
978 //
979 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
980 //
981 // STATE IDENTIFICATION: (64 bit+COOPS)
982 // ------------------------------------
983 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
984 //
985 // OBJECT: klass_word installed; klass_word != 0;
986 // obj->size() computes correct size
987 //
988 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
989 //
990 //
991 // STATE TRANSITION DIAGRAM
992 //
993 // mut / parnew mut / parnew
994 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
995 // ^ |
996 // |------------------------ DEAD <------------------------------------|
997 // sweep mut
998 //
999 // While a block is in TRANSIENT state its size cannot be determined
1000 // so readers will either need to come back later or stall until
1001 // the size can be determined. Note that for the case of direct
1002 // allocation, P-bits, when available, may be used to determine the
1003 // size of an object that may not yet have been initialized.
1004
1005 // Things to support parallel young-gen collection.
1006 oop
1007 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1008 oop old, markOop m,
1009 size_t word_sz) {
1010 #ifndef PRODUCT
1011 if (CMSHeap::heap()->promotion_should_fail()) {
1012 return NULL;
1013 }
1014 #endif // #ifndef PRODUCT
1015
1016 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1017 PromotionInfo* promoInfo = &ps->promo;
1018 // if we are tracking promotions, then first ensure space for
1019 // promotion (including spooling space for saving header if necessary).
1020 // then allocate and copy, then track promoted info if needed.
1021 // When tracking (see PromotionInfo::track()), the mark word may
1022 // be displaced and in this case restoration of the mark word
1023 // occurs in the (oop_since_save_marks_)iterate phase.
1024 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1025 // Out of space for allocating spooling buffers;
1026 // try expanding and allocating spooling buffers.
1027 if (!expand_and_ensure_spooling_space(promoInfo)) {
1028 return NULL;
1029 }
1030 }
1031 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1032 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1033 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1034 if (obj_ptr == NULL) {
1035 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1036 if (obj_ptr == NULL) {
1037 return NULL;
1038 }
1039 }
1040 oop obj = oop(obj_ptr);
1041 OrderAccess::storestore();
1042 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1043 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1044 // IMPORTANT: See note on object initialization for CMS above.
1045 // Otherwise, copy the object. Here we must be careful to insert the
1046 // klass pointer last, since this marks the block as an allocated object.
1047 // Except with compressed oops it's the mark word.
1048 HeapWord* old_ptr = (HeapWord*)old;
1049 // Restore the mark word copied above.
1050 obj->set_mark_raw(m);
1051 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1052 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1053 OrderAccess::storestore();
1054
1055 if (UseCompressedClassPointers) {
1056 // Copy gap missed by (aligned) header size calculation below
1057 obj->set_klass_gap(old->klass_gap());
1058 }
1059 if (word_sz > (size_t)oopDesc::header_size()) {
1060 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1061 obj_ptr + oopDesc::header_size(),
1062 word_sz - oopDesc::header_size());
1063 }
1064
1065 // Now we can track the promoted object, if necessary. We take care
1066 // to delay the transition from uninitialized to full object
1067 // (i.e., insertion of klass pointer) until after, so that it
1068 // atomically becomes a promoted object.
1069 if (promoInfo->tracking()) {
1070 promoInfo->track((PromotedObject*)obj, old->klass());
1071 }
1072 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1073 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1074 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below");
1075
1076 // Finally, install the klass pointer (this should be volatile).
1077 OrderAccess::storestore();
1078 obj->set_klass(old->klass());
1079 // We should now be able to calculate the right size for this object
1080 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1081
1082 collector()->promoted(true, // parallel
1083 obj_ptr, old->is_objArray(), word_sz);
1084
1085 NOT_PRODUCT(
1086 Atomic::inc(&_numObjectsPromoted);
1087 Atomic::add(alloc_sz, &_numWordsPromoted);
1088 )
1089
1090 return obj;
1091 }
1092
1093 void
1094 ConcurrentMarkSweepGeneration::
1095 par_promote_alloc_done(int thread_num) {
1096 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1097 ps->lab.retire(thread_num);
1098 }
1099
1100 void
1101 ConcurrentMarkSweepGeneration::
1102 par_oop_since_save_marks_iterate_done(int thread_num) {
1103 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1104 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1105 ps->promo.promoted_oops_iterate(dummy_cl);
1106
1107 // Because card-scanning has been completed, subsequent phases
1108 // (e.g., reference processing) will not need to recognize which
1109 // objects have been promoted during this GC. So, we can now disable
1110 // promotion tracking.
1111 ps->promo.stopTrackingPromotions();
1112 }
1113
1114 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1115 size_t size,
1116 bool tlab)
1117 {
1118 // We allow a STW collection only if a full
1119 // collection was requested.
1120 return full || should_allocate(size, tlab); // FIX ME !!!
1121 // This and promotion failure handling are connected at the
1122 // hip and should be fixed by untying them.
1123 }
1124
1125 bool CMSCollector::shouldConcurrentCollect() {
1126 LogTarget(Trace, gc) log;
1127
1128 if (_full_gc_requested) {
1129 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)");
1130 return true;
1131 }
1132
1133 FreelistLocker x(this);
1134 // ------------------------------------------------------------------
1135 // Print out lots of information which affects the initiation of
1136 // a collection.
1137 if (log.is_enabled() && stats().valid()) {
1138 log.print("CMSCollector shouldConcurrentCollect: ");
1139
1140 LogStream out(log);
1141 stats().print_on(&out);
1142
1143 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1144 log.print("free=" SIZE_FORMAT, _cmsGen->free());
1145 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1146 log.print("promotion_rate=%g", stats().promotion_rate());
1147 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1148 log.print("occupancy=%3.7f", _cmsGen->occupancy());
1149 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1150 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1151 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1152 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1153 }
1154 // ------------------------------------------------------------------
1155
1156 // If the estimated time to complete a cms collection (cms_duration())
1157 // is less than the estimated time remaining until the cms generation
1158 // is full, start a collection.
1159 if (!UseCMSInitiatingOccupancyOnly) {
1160 if (stats().valid()) {
1161 if (stats().time_until_cms_start() == 0.0) {
1162 return true;
1163 }
1164 } else {
1165 // We want to conservatively collect somewhat early in order
1166 // to try and "bootstrap" our CMS/promotion statistics;
1167 // this branch will not fire after the first successful CMS
1168 // collection because the stats should then be valid.
1169 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1170 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1171 _cmsGen->occupancy(), _bootstrap_occupancy);
1172 return true;
1173 }
1174 }
1175 }
1176
1177 // Otherwise, we start a collection cycle if
1178 // old gen want a collection cycle started. Each may use
1179 // an appropriate criterion for making this decision.
1180 // XXX We need to make sure that the gen expansion
1181 // criterion dovetails well with this. XXX NEED TO FIX THIS
1182 if (_cmsGen->should_concurrent_collect()) {
1183 log.print("CMS old gen initiated");
1184 return true;
1185 }
1186
1187 // We start a collection if we believe an incremental collection may fail;
1188 // this is not likely to be productive in practice because it's probably too
1189 // late anyway.
1190 CMSHeap* heap = CMSHeap::heap();
1191 if (heap->incremental_collection_will_fail(true /* consult_young */)) {
1192 log.print("CMSCollector: collect because incremental collection will fail ");
1193 return true;
1194 }
1195
1196 if (MetaspaceGC::should_concurrent_collect()) {
1197 log.print("CMSCollector: collect for metadata allocation ");
1198 return true;
1199 }
1200
1201 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1202 if (CMSTriggerInterval >= 0) {
1203 if (CMSTriggerInterval == 0) {
1204 // Trigger always
1205 return true;
1206 }
1207
1208 // Check the CMS time since begin (we do not check the stats validity
1209 // as we want to be able to trigger the first CMS cycle as well)
1210 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1211 if (stats().valid()) {
1212 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1213 stats().cms_time_since_begin());
1214 } else {
1215 log.print("CMSCollector: collect because of trigger interval (first collection)");
1216 }
1217 return true;
1218 }
1219 }
1220
1221 return false;
1222 }
1223
1224 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1225
1226 // Clear _expansion_cause fields of constituent generations
1227 void CMSCollector::clear_expansion_cause() {
1228 _cmsGen->clear_expansion_cause();
1229 }
1230
1231 // We should be conservative in starting a collection cycle. To
1232 // start too eagerly runs the risk of collecting too often in the
1233 // extreme. To collect too rarely falls back on full collections,
1234 // which works, even if not optimum in terms of concurrent work.
1235 // As a work around for too eagerly collecting, use the flag
1236 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1237 // giving the user an easily understandable way of controlling the
1238 // collections.
1239 // We want to start a new collection cycle if any of the following
1240 // conditions hold:
1241 // . our current occupancy exceeds the configured initiating occupancy
1242 // for this generation, or
1243 // . we recently needed to expand this space and have not, since that
1244 // expansion, done a collection of this generation, or
1245 // . the underlying space believes that it may be a good idea to initiate
1246 // a concurrent collection (this may be based on criteria such as the
1247 // following: the space uses linear allocation and linear allocation is
1248 // going to fail, or there is believed to be excessive fragmentation in
1249 // the generation, etc... or ...
1250 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1251 // the case of the old generation; see CR 6543076):
1252 // we may be approaching a point at which allocation requests may fail because
1253 // we will be out of sufficient free space given allocation rate estimates.]
1254 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1255
1256 assert_lock_strong(freelistLock());
1257 if (occupancy() > initiating_occupancy()) {
1258 log_trace(gc)(" %s: collect because of occupancy %f / %f ",
1259 short_name(), occupancy(), initiating_occupancy());
1260 return true;
1261 }
1262 if (UseCMSInitiatingOccupancyOnly) {
1263 return false;
1264 }
1265 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1266 log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1267 return true;
1268 }
1269 return false;
1270 }
1271
1272 void ConcurrentMarkSweepGeneration::collect(bool full,
1273 bool clear_all_soft_refs,
1274 size_t size,
1275 bool tlab)
1276 {
1277 collector()->collect(full, clear_all_soft_refs, size, tlab);
1278 }
1279
1280 void CMSCollector::collect(bool full,
1281 bool clear_all_soft_refs,
1282 size_t size,
1283 bool tlab)
1284 {
1285 // The following "if" branch is present for defensive reasons.
1286 // In the current uses of this interface, it can be replaced with:
1287 // assert(!GCLocker.is_active(), "Can't be called otherwise");
1288 // But I am not placing that assert here to allow future
1289 // generality in invoking this interface.
1290 if (GCLocker::is_active()) {
1291 // A consistency test for GCLocker
1292 assert(GCLocker::needs_gc(), "Should have been set already");
1293 // Skip this foreground collection, instead
1294 // expanding the heap if necessary.
1295 // Need the free list locks for the call to free() in compute_new_size()
1296 compute_new_size();
1297 return;
1298 }
1299 acquire_control_and_collect(full, clear_all_soft_refs);
1300 }
1301
1302 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1303 CMSHeap* heap = CMSHeap::heap();
1304 unsigned int gc_count = heap->total_full_collections();
1305 if (gc_count == full_gc_count) {
1306 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1307 _full_gc_requested = true;
1308 _full_gc_cause = cause;
1309 CGC_lock->notify(); // nudge CMS thread
1310 } else {
1311 assert(gc_count > full_gc_count, "Error: causal loop");
1312 }
1313 }
1314
1315 bool CMSCollector::is_external_interruption() {
1316 GCCause::Cause cause = CMSHeap::heap()->gc_cause();
1317 return GCCause::is_user_requested_gc(cause) ||
1318 GCCause::is_serviceability_requested_gc(cause);
1319 }
1320
1321 void CMSCollector::report_concurrent_mode_interruption() {
1322 if (is_external_interruption()) {
1323 log_debug(gc)("Concurrent mode interrupted");
1324 } else {
1325 log_debug(gc)("Concurrent mode failure");
1326 _gc_tracer_cm->report_concurrent_mode_failure();
1327 }
1328 }
1329
1330
1331 // The foreground and background collectors need to coordinate in order
1332 // to make sure that they do not mutually interfere with CMS collections.
1333 // When a background collection is active,
1334 // the foreground collector may need to take over (preempt) and
1335 // synchronously complete an ongoing collection. Depending on the
1336 // frequency of the background collections and the heap usage
1337 // of the application, this preemption can be seldom or frequent.
1338 // There are only certain
1339 // points in the background collection that the "collection-baton"
1340 // can be passed to the foreground collector.
1341 //
1342 // The foreground collector will wait for the baton before
1343 // starting any part of the collection. The foreground collector
1344 // will only wait at one location.
1345 //
1346 // The background collector will yield the baton before starting a new
1347 // phase of the collection (e.g., before initial marking, marking from roots,
1348 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1349 // of the loop which switches the phases. The background collector does some
1350 // of the phases (initial mark, final re-mark) with the world stopped.
1351 // Because of locking involved in stopping the world,
1352 // the foreground collector should not block waiting for the background
1353 // collector when it is doing a stop-the-world phase. The background
1354 // collector will yield the baton at an additional point just before
1355 // it enters a stop-the-world phase. Once the world is stopped, the
1356 // background collector checks the phase of the collection. If the
1357 // phase has not changed, it proceeds with the collection. If the
1358 // phase has changed, it skips that phase of the collection. See
1359 // the comments on the use of the Heap_lock in collect_in_background().
1360 //
1361 // Variable used in baton passing.
1362 // _foregroundGCIsActive - Set to true by the foreground collector when
1363 // it wants the baton. The foreground clears it when it has finished
1364 // the collection.
1365 // _foregroundGCShouldWait - Set to true by the background collector
1366 // when it is running. The foreground collector waits while
1367 // _foregroundGCShouldWait is true.
1368 // CGC_lock - monitor used to protect access to the above variables
1369 // and to notify the foreground and background collectors.
1370 // _collectorState - current state of the CMS collection.
1371 //
1372 // The foreground collector
1373 // acquires the CGC_lock
1374 // sets _foregroundGCIsActive
1375 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1376 // various locks acquired in preparation for the collection
1377 // are released so as not to block the background collector
1378 // that is in the midst of a collection
1379 // proceeds with the collection
1380 // clears _foregroundGCIsActive
1381 // returns
1382 //
1383 // The background collector in a loop iterating on the phases of the
1384 // collection
1385 // acquires the CGC_lock
1386 // sets _foregroundGCShouldWait
1387 // if _foregroundGCIsActive is set
1388 // clears _foregroundGCShouldWait, notifies _CGC_lock
1389 // waits on _CGC_lock for _foregroundGCIsActive to become false
1390 // and exits the loop.
1391 // otherwise
1392 // proceed with that phase of the collection
1393 // if the phase is a stop-the-world phase,
1394 // yield the baton once more just before enqueueing
1395 // the stop-world CMS operation (executed by the VM thread).
1396 // returns after all phases of the collection are done
1397 //
1398
1399 void CMSCollector::acquire_control_and_collect(bool full,
1400 bool clear_all_soft_refs) {
1401 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1402 assert(!Thread::current()->is_ConcurrentGC_thread(),
1403 "shouldn't try to acquire control from self!");
1404
1405 // Start the protocol for acquiring control of the
1406 // collection from the background collector (aka CMS thread).
1407 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1408 "VM thread should have CMS token");
1409 // Remember the possibly interrupted state of an ongoing
1410 // concurrent collection
1411 CollectorState first_state = _collectorState;
1412
1413 // Signal to a possibly ongoing concurrent collection that
1414 // we want to do a foreground collection.
1415 _foregroundGCIsActive = true;
1416
1417 // release locks and wait for a notify from the background collector
1418 // releasing the locks in only necessary for phases which
1419 // do yields to improve the granularity of the collection.
1420 assert_lock_strong(bitMapLock());
1421 // We need to lock the Free list lock for the space that we are
1422 // currently collecting.
1423 assert(haveFreelistLocks(), "Must be holding free list locks");
1424 bitMapLock()->unlock();
1425 releaseFreelistLocks();
1426 {
1427 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1428 if (_foregroundGCShouldWait) {
1429 // We are going to be waiting for action for the CMS thread;
1430 // it had better not be gone (for instance at shutdown)!
1431 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1432 "CMS thread must be running");
1433 // Wait here until the background collector gives us the go-ahead
1434 ConcurrentMarkSweepThread::clear_CMS_flag(
1435 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1436 // Get a possibly blocked CMS thread going:
1437 // Note that we set _foregroundGCIsActive true above,
1438 // without protection of the CGC_lock.
1439 CGC_lock->notify();
1440 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1441 "Possible deadlock");
1442 while (_foregroundGCShouldWait) {
1443 // wait for notification
1444 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1445 // Possibility of delay/starvation here, since CMS token does
1446 // not know to give priority to VM thread? Actually, i think
1447 // there wouldn't be any delay/starvation, but the proof of
1448 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1449 }
1450 ConcurrentMarkSweepThread::set_CMS_flag(
1451 ConcurrentMarkSweepThread::CMS_vm_has_token);
1452 }
1453 }
1454 // The CMS_token is already held. Get back the other locks.
1455 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1456 "VM thread should have CMS token");
1457 getFreelistLocks();
1458 bitMapLock()->lock_without_safepoint_check();
1459 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1460 p2i(Thread::current()), first_state);
1461 log_debug(gc, state)(" gets control with state %d", _collectorState);
1462
1463 // Inform cms gen if this was due to partial collection failing.
1464 // The CMS gen may use this fact to determine its expansion policy.
1465 CMSHeap* heap = CMSHeap::heap();
1466 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) {
1467 assert(!_cmsGen->incremental_collection_failed(),
1468 "Should have been noticed, reacted to and cleared");
1469 _cmsGen->set_incremental_collection_failed();
1470 }
1471
1472 if (first_state > Idling) {
1473 report_concurrent_mode_interruption();
1474 }
1475
1476 set_did_compact(true);
1477
1478 // If the collection is being acquired from the background
1479 // collector, there may be references on the discovered
1480 // references lists. Abandon those references, since some
1481 // of them may have become unreachable after concurrent
1482 // discovery; the STW compacting collector will redo discovery
1483 // more precisely, without being subject to floating garbage.
1484 // Leaving otherwise unreachable references in the discovered
1485 // lists would require special handling.
1486 ref_processor()->disable_discovery();
1487 ref_processor()->abandon_partial_discovery();
1488 ref_processor()->verify_no_references_recorded();
1489
1490 if (first_state > Idling) {
1491 save_heap_summary();
1492 }
1493
1494 do_compaction_work(clear_all_soft_refs);
1495
1496 // Has the GC time limit been exceeded?
1497 size_t max_eden_size = _young_gen->max_eden_size();
1498 GCCause::Cause gc_cause = heap->gc_cause();
1499 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1500 _young_gen->eden()->used(),
1501 _cmsGen->max_capacity(),
1502 max_eden_size,
1503 full,
1504 gc_cause,
1505 heap->soft_ref_policy());
1506
1507 // Reset the expansion cause, now that we just completed
1508 // a collection cycle.
1509 clear_expansion_cause();
1510 _foregroundGCIsActive = false;
1511 return;
1512 }
1513
1514 // Resize the tenured generation
1515 // after obtaining the free list locks for the
1516 // two generations.
1517 void CMSCollector::compute_new_size() {
1518 assert_locked_or_safepoint(Heap_lock);
1519 FreelistLocker z(this);
1520 MetaspaceGC::compute_new_size();
1521 _cmsGen->compute_new_size_free_list();
1522 }
1523
1524 // A work method used by the foreground collector to do
1525 // a mark-sweep-compact.
1526 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1527 CMSHeap* heap = CMSHeap::heap();
1528
1529 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1530 gc_timer->register_gc_start();
1531
1532 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1533 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start());
1534
1535 heap->pre_full_gc_dump(gc_timer);
1536
1537 GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1538
1539 // Temporarily widen the span of the weak reference processing to
1540 // the entire heap.
1541 MemRegion new_span(CMSHeap::heap()->reserved_region());
1542 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1543 // Temporarily, clear the "is_alive_non_header" field of the
1544 // reference processor.
1545 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1546 // Temporarily make reference _processing_ single threaded (non-MT).
1547 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1548 // Temporarily make refs discovery atomic
1549 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1550 // Temporarily make reference _discovery_ single threaded (non-MT)
1551 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1552
1553 ref_processor()->set_enqueuing_is_done(false);
1554 ref_processor()->enable_discovery();
1555 ref_processor()->setup_policy(clear_all_soft_refs);
1556 // If an asynchronous collection finishes, the _modUnionTable is
1557 // all clear. If we are assuming the collection from an asynchronous
1558 // collection, clear the _modUnionTable.
1559 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1560 "_modUnionTable should be clear if the baton was not passed");
1561 _modUnionTable.clear_all();
1562 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(),
1563 "mod union for klasses should be clear if the baton was passed");
1564 _ct->cld_rem_set()->clear_mod_union();
1565
1566
1567 // We must adjust the allocation statistics being maintained
1568 // in the free list space. We do so by reading and clearing
1569 // the sweep timer and updating the block flux rate estimates below.
1570 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1571 if (_inter_sweep_timer.is_active()) {
1572 _inter_sweep_timer.stop();
1573 // Note that we do not use this sample to update the _inter_sweep_estimate.
1574 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1575 _inter_sweep_estimate.padded_average(),
1576 _intra_sweep_estimate.padded_average());
1577 }
1578
1579 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1580 #ifdef ASSERT
1581 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1582 size_t free_size = cms_space->free();
1583 assert(free_size ==
1584 pointer_delta(cms_space->end(), cms_space->compaction_top())
1585 * HeapWordSize,
1586 "All the free space should be compacted into one chunk at top");
1587 assert(cms_space->dictionary()->total_chunk_size(
1588 debug_only(cms_space->freelistLock())) == 0 ||
1589 cms_space->totalSizeInIndexedFreeLists() == 0,
1590 "All the free space should be in a single chunk");
1591 size_t num = cms_space->totalCount();
1592 assert((free_size == 0 && num == 0) ||
1593 (free_size > 0 && (num == 1 || num == 2)),
1594 "There should be at most 2 free chunks after compaction");
1595 #endif // ASSERT
1596 _collectorState = Resetting;
1597 assert(_restart_addr == NULL,
1598 "Should have been NULL'd before baton was passed");
1599 reset_stw();
1600 _cmsGen->reset_after_compaction();
1601 _concurrent_cycles_since_last_unload = 0;
1602
1603 // Clear any data recorded in the PLAB chunk arrays.
1604 if (_survivor_plab_array != NULL) {
1605 reset_survivor_plab_arrays();
1606 }
1607
1608 // Adjust the per-size allocation stats for the next epoch.
1609 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1610 // Restart the "inter sweep timer" for the next epoch.
1611 _inter_sweep_timer.reset();
1612 _inter_sweep_timer.start();
1613
1614 // No longer a need to do a concurrent collection for Metaspace.
1615 MetaspaceGC::set_should_concurrent_collect(false);
1616
1617 heap->post_full_gc_dump(gc_timer);
1618
1619 gc_timer->register_gc_end();
1620
1621 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1622
1623 // For a mark-sweep-compact, compute_new_size() will be called
1624 // in the heap's do_collection() method.
1625 }
1626
1627 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1628 Log(gc, heap) log;
1629 if (!log.is_trace()) {
1630 return;
1631 }
1632
1633 ContiguousSpace* eden_space = _young_gen->eden();
1634 ContiguousSpace* from_space = _young_gen->from();
1635 ContiguousSpace* to_space = _young_gen->to();
1636 // Eden
1637 if (_eden_chunk_array != NULL) {
1638 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1639 p2i(eden_space->bottom()), p2i(eden_space->top()),
1640 p2i(eden_space->end()), eden_space->capacity());
1641 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1642 _eden_chunk_index, _eden_chunk_capacity);
1643 for (size_t i = 0; i < _eden_chunk_index; i++) {
1644 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1645 }
1646 }
1647 // Survivor
1648 if (_survivor_chunk_array != NULL) {
1649 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1650 p2i(from_space->bottom()), p2i(from_space->top()),
1651 p2i(from_space->end()), from_space->capacity());
1652 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1653 _survivor_chunk_index, _survivor_chunk_capacity);
1654 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1655 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1656 }
1657 }
1658 }
1659
1660 void CMSCollector::getFreelistLocks() const {
1661 // Get locks for all free lists in all generations that this
1662 // collector is responsible for
1663 _cmsGen->freelistLock()->lock_without_safepoint_check();
1664 }
1665
1666 void CMSCollector::releaseFreelistLocks() const {
1667 // Release locks for all free lists in all generations that this
1668 // collector is responsible for
1669 _cmsGen->freelistLock()->unlock();
1670 }
1671
1672 bool CMSCollector::haveFreelistLocks() const {
1673 // Check locks for all free lists in all generations that this
1674 // collector is responsible for
1675 assert_lock_strong(_cmsGen->freelistLock());
1676 PRODUCT_ONLY(ShouldNotReachHere());
1677 return true;
1678 }
1679
1680 // A utility class that is used by the CMS collector to
1681 // temporarily "release" the foreground collector from its
1682 // usual obligation to wait for the background collector to
1683 // complete an ongoing phase before proceeding.
1684 class ReleaseForegroundGC: public StackObj {
1685 private:
1686 CMSCollector* _c;
1687 public:
1688 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1689 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1690 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1691 // allow a potentially blocked foreground collector to proceed
1692 _c->_foregroundGCShouldWait = false;
1693 if (_c->_foregroundGCIsActive) {
1694 CGC_lock->notify();
1695 }
1696 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1697 "Possible deadlock");
1698 }
1699
1700 ~ReleaseForegroundGC() {
1701 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1702 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1703 _c->_foregroundGCShouldWait = true;
1704 }
1705 };
1706
1707 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1708 assert(Thread::current()->is_ConcurrentGC_thread(),
1709 "A CMS asynchronous collection is only allowed on a CMS thread.");
1710
1711 CMSHeap* heap = CMSHeap::heap();
1712 {
1713 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1714 MutexLockerEx hl(Heap_lock, safepoint_check);
1715 FreelistLocker fll(this);
1716 MutexLockerEx x(CGC_lock, safepoint_check);
1717 if (_foregroundGCIsActive) {
1718 // The foreground collector is. Skip this
1719 // background collection.
1720 assert(!_foregroundGCShouldWait, "Should be clear");
1721 return;
1722 } else {
1723 assert(_collectorState == Idling, "Should be idling before start.");
1724 _collectorState = InitialMarking;
1725 register_gc_start(cause);
1726 // Reset the expansion cause, now that we are about to begin
1727 // a new cycle.
1728 clear_expansion_cause();
1729
1730 // Clear the MetaspaceGC flag since a concurrent collection
1731 // is starting but also clear it after the collection.
1732 MetaspaceGC::set_should_concurrent_collect(false);
1733 }
1734 // Decide if we want to enable class unloading as part of the
1735 // ensuing concurrent GC cycle.
1736 update_should_unload_classes();
1737 _full_gc_requested = false; // acks all outstanding full gc requests
1738 _full_gc_cause = GCCause::_no_gc;
1739 // Signal that we are about to start a collection
1740 heap->increment_total_full_collections(); // ... starting a collection cycle
1741 _collection_count_start = heap->total_full_collections();
1742 }
1743
1744 size_t prev_used = _cmsGen->used();
1745
1746 // The change of the collection state is normally done at this level;
1747 // the exceptions are phases that are executed while the world is
1748 // stopped. For those phases the change of state is done while the
1749 // world is stopped. For baton passing purposes this allows the
1750 // background collector to finish the phase and change state atomically.
1751 // The foreground collector cannot wait on a phase that is done
1752 // while the world is stopped because the foreground collector already
1753 // has the world stopped and would deadlock.
1754 while (_collectorState != Idling) {
1755 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1756 p2i(Thread::current()), _collectorState);
1757 // The foreground collector
1758 // holds the Heap_lock throughout its collection.
1759 // holds the CMS token (but not the lock)
1760 // except while it is waiting for the background collector to yield.
1761 //
1762 // The foreground collector should be blocked (not for long)
1763 // if the background collector is about to start a phase
1764 // executed with world stopped. If the background
1765 // collector has already started such a phase, the
1766 // foreground collector is blocked waiting for the
1767 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1768 // are executed in the VM thread.
1769 //
1770 // The locking order is
1771 // PendingListLock (PLL) -- if applicable (FinalMarking)
1772 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1773 // CMS token (claimed in
1774 // stop_world_and_do() -->
1775 // safepoint_synchronize() -->
1776 // CMSThread::synchronize())
1777
1778 {
1779 // Check if the FG collector wants us to yield.
1780 CMSTokenSync x(true); // is cms thread
1781 if (waitForForegroundGC()) {
1782 // We yielded to a foreground GC, nothing more to be
1783 // done this round.
1784 assert(_foregroundGCShouldWait == false, "We set it to false in "
1785 "waitForForegroundGC()");
1786 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1787 p2i(Thread::current()), _collectorState);
1788 return;
1789 } else {
1790 // The background collector can run but check to see if the
1791 // foreground collector has done a collection while the
1792 // background collector was waiting to get the CGC_lock
1793 // above. If yes, break so that _foregroundGCShouldWait
1794 // is cleared before returning.
1795 if (_collectorState == Idling) {
1796 break;
1797 }
1798 }
1799 }
1800
1801 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1802 "should be waiting");
1803
1804 switch (_collectorState) {
1805 case InitialMarking:
1806 {
1807 ReleaseForegroundGC x(this);
1808 stats().record_cms_begin();
1809 VM_CMS_Initial_Mark initial_mark_op(this);
1810 VMThread::execute(&initial_mark_op);
1811 }
1812 // The collector state may be any legal state at this point
1813 // since the background collector may have yielded to the
1814 // foreground collector.
1815 break;
1816 case Marking:
1817 // initial marking in checkpointRootsInitialWork has been completed
1818 if (markFromRoots()) { // we were successful
1819 assert(_collectorState == Precleaning, "Collector state should "
1820 "have changed");
1821 } else {
1822 assert(_foregroundGCIsActive, "Internal state inconsistency");
1823 }
1824 break;
1825 case Precleaning:
1826 // marking from roots in markFromRoots has been completed
1827 preclean();
1828 assert(_collectorState == AbortablePreclean ||
1829 _collectorState == FinalMarking,
1830 "Collector state should have changed");
1831 break;
1832 case AbortablePreclean:
1833 abortable_preclean();
1834 assert(_collectorState == FinalMarking, "Collector state should "
1835 "have changed");
1836 break;
1837 case FinalMarking:
1838 {
1839 ReleaseForegroundGC x(this);
1840
1841 VM_CMS_Final_Remark final_remark_op(this);
1842 VMThread::execute(&final_remark_op);
1843 }
1844 assert(_foregroundGCShouldWait, "block post-condition");
1845 break;
1846 case Sweeping:
1847 // final marking in checkpointRootsFinal has been completed
1848 sweep();
1849 assert(_collectorState == Resizing, "Collector state change "
1850 "to Resizing must be done under the free_list_lock");
1851
1852 case Resizing: {
1853 // Sweeping has been completed...
1854 // At this point the background collection has completed.
1855 // Don't move the call to compute_new_size() down
1856 // into code that might be executed if the background
1857 // collection was preempted.
1858 {
1859 ReleaseForegroundGC x(this); // unblock FG collection
1860 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1861 CMSTokenSync z(true); // not strictly needed.
1862 if (_collectorState == Resizing) {
1863 compute_new_size();
1864 save_heap_summary();
1865 _collectorState = Resetting;
1866 } else {
1867 assert(_collectorState == Idling, "The state should only change"
1868 " because the foreground collector has finished the collection");
1869 }
1870 }
1871 break;
1872 }
1873 case Resetting:
1874 // CMS heap resizing has been completed
1875 reset_concurrent();
1876 assert(_collectorState == Idling, "Collector state should "
1877 "have changed");
1878
1879 MetaspaceGC::set_should_concurrent_collect(false);
1880
1881 stats().record_cms_end();
1882 // Don't move the concurrent_phases_end() and compute_new_size()
1883 // calls to here because a preempted background collection
1884 // has it's state set to "Resetting".
1885 break;
1886 case Idling:
1887 default:
1888 ShouldNotReachHere();
1889 break;
1890 }
1891 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1892 p2i(Thread::current()), _collectorState);
1893 assert(_foregroundGCShouldWait, "block post-condition");
1894 }
1895
1896 // Should this be in gc_epilogue?
1897 heap->counters()->update_counters();
1898
1899 {
1900 // Clear _foregroundGCShouldWait and, in the event that the
1901 // foreground collector is waiting, notify it, before
1902 // returning.
1903 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1904 _foregroundGCShouldWait = false;
1905 if (_foregroundGCIsActive) {
1906 CGC_lock->notify();
1907 }
1908 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1909 "Possible deadlock");
1910 }
1911 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1912 p2i(Thread::current()), _collectorState);
1913 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)",
1914 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1915 }
1916
1917 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1918 _cms_start_registered = true;
1919 _gc_timer_cm->register_gc_start();
1920 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1921 }
1922
1923 void CMSCollector::register_gc_end() {
1924 if (_cms_start_registered) {
1925 report_heap_summary(GCWhen::AfterGC);
1926
1927 _gc_timer_cm->register_gc_end();
1928 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1929 _cms_start_registered = false;
1930 }
1931 }
1932
1933 void CMSCollector::save_heap_summary() {
1934 CMSHeap* heap = CMSHeap::heap();
1935 _last_heap_summary = heap->create_heap_summary();
1936 _last_metaspace_summary = heap->create_metaspace_summary();
1937 }
1938
1939 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1940 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1941 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1942 }
1943
1944 bool CMSCollector::waitForForegroundGC() {
1945 bool res = false;
1946 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1947 "CMS thread should have CMS token");
1948 // Block the foreground collector until the
1949 // background collectors decides whether to
1950 // yield.
1951 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1952 _foregroundGCShouldWait = true;
1953 if (_foregroundGCIsActive) {
1954 // The background collector yields to the
1955 // foreground collector and returns a value
1956 // indicating that it has yielded. The foreground
1957 // collector can proceed.
1958 res = true;
1959 _foregroundGCShouldWait = false;
1960 ConcurrentMarkSweepThread::clear_CMS_flag(
1961 ConcurrentMarkSweepThread::CMS_cms_has_token);
1962 ConcurrentMarkSweepThread::set_CMS_flag(
1963 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1964 // Get a possibly blocked foreground thread going
1965 CGC_lock->notify();
1966 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1967 p2i(Thread::current()), _collectorState);
1968 while (_foregroundGCIsActive) {
1969 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1970 }
1971 ConcurrentMarkSweepThread::set_CMS_flag(
1972 ConcurrentMarkSweepThread::CMS_cms_has_token);
1973 ConcurrentMarkSweepThread::clear_CMS_flag(
1974 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1975 }
1976 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1977 p2i(Thread::current()), _collectorState);
1978 return res;
1979 }
1980
1981 // Because of the need to lock the free lists and other structures in
1982 // the collector, common to all the generations that the collector is
1983 // collecting, we need the gc_prologues of individual CMS generations
1984 // delegate to their collector. It may have been simpler had the
1985 // current infrastructure allowed one to call a prologue on a
1986 // collector. In the absence of that we have the generation's
1987 // prologue delegate to the collector, which delegates back
1988 // some "local" work to a worker method in the individual generations
1989 // that it's responsible for collecting, while itself doing any
1990 // work common to all generations it's responsible for. A similar
1991 // comment applies to the gc_epilogue()'s.
1992 // The role of the variable _between_prologue_and_epilogue is to
1993 // enforce the invocation protocol.
1994 void CMSCollector::gc_prologue(bool full) {
1995 // Call gc_prologue_work() for the CMSGen
1996 // we are responsible for.
1997
1998 // The following locking discipline assumes that we are only called
1999 // when the world is stopped.
2000 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2001
2002 // The CMSCollector prologue must call the gc_prologues for the
2003 // "generations" that it's responsible
2004 // for.
2005
2006 assert( Thread::current()->is_VM_thread()
2007 || ( CMSScavengeBeforeRemark
2008 && Thread::current()->is_ConcurrentGC_thread()),
2009 "Incorrect thread type for prologue execution");
2010
2011 if (_between_prologue_and_epilogue) {
2012 // We have already been invoked; this is a gc_prologue delegation
2013 // from yet another CMS generation that we are responsible for, just
2014 // ignore it since all relevant work has already been done.
2015 return;
2016 }
2017
2018 // set a bit saying prologue has been called; cleared in epilogue
2019 _between_prologue_and_epilogue = true;
2020 // Claim locks for common data structures, then call gc_prologue_work()
2021 // for each CMSGen.
2022
2023 getFreelistLocks(); // gets free list locks on constituent spaces
2024 bitMapLock()->lock_without_safepoint_check();
2025
2026 // Should call gc_prologue_work() for all cms gens we are responsible for
2027 bool duringMarking = _collectorState >= Marking
2028 && _collectorState < Sweeping;
2029
2030 // The young collections clear the modified oops state, which tells if
2031 // there are any modified oops in the class. The remark phase also needs
2032 // that information. Tell the young collection to save the union of all
2033 // modified klasses.
2034 if (duringMarking) {
2035 _ct->cld_rem_set()->set_accumulate_modified_oops(true);
2036 }
2037
2038 bool registerClosure = duringMarking;
2039
2040 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2041
2042 if (!full) {
2043 stats().record_gc0_begin();
2044 }
2045 }
2046
2047 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2048
2049 _capacity_at_prologue = capacity();
2050 _used_at_prologue = used();
2051
2052 // We enable promotion tracking so that card-scanning can recognize
2053 // which objects have been promoted during this GC and skip them.
2054 for (uint i = 0; i < ParallelGCThreads; i++) {
2055 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2056 }
2057
2058 // Delegate to CMScollector which knows how to coordinate between
2059 // this and any other CMS generations that it is responsible for
2060 // collecting.
2061 collector()->gc_prologue(full);
2062 }
2063
2064 // This is a "private" interface for use by this generation's CMSCollector.
2065 // Not to be called directly by any other entity (for instance,
2066 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2067 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2068 bool registerClosure, ModUnionClosure* modUnionClosure) {
2069 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2070 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2071 "Should be NULL");
2072 if (registerClosure) {
2073 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2074 }
2075 cmsSpace()->gc_prologue();
2076 // Clear stat counters
2077 NOT_PRODUCT(
2078 assert(_numObjectsPromoted == 0, "check");
2079 assert(_numWordsPromoted == 0, "check");
2080 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2081 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2082 _numObjectsAllocated = 0;
2083 _numWordsAllocated = 0;
2084 )
2085 }
2086
2087 void CMSCollector::gc_epilogue(bool full) {
2088 // The following locking discipline assumes that we are only called
2089 // when the world is stopped.
2090 assert(SafepointSynchronize::is_at_safepoint(),
2091 "world is stopped assumption");
2092
2093 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2094 // if linear allocation blocks need to be appropriately marked to allow the
2095 // the blocks to be parsable. We also check here whether we need to nudge the
2096 // CMS collector thread to start a new cycle (if it's not already active).
2097 assert( Thread::current()->is_VM_thread()
2098 || ( CMSScavengeBeforeRemark
2099 && Thread::current()->is_ConcurrentGC_thread()),
2100 "Incorrect thread type for epilogue execution");
2101
2102 if (!_between_prologue_and_epilogue) {
2103 // We have already been invoked; this is a gc_epilogue delegation
2104 // from yet another CMS generation that we are responsible for, just
2105 // ignore it since all relevant work has already been done.
2106 return;
2107 }
2108 assert(haveFreelistLocks(), "must have freelist locks");
2109 assert_lock_strong(bitMapLock());
2110
2111 _ct->cld_rem_set()->set_accumulate_modified_oops(false);
2112
2113 _cmsGen->gc_epilogue_work(full);
2114
2115 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2116 // in case sampling was not already enabled, enable it
2117 _start_sampling = true;
2118 }
2119 // reset _eden_chunk_array so sampling starts afresh
2120 _eden_chunk_index = 0;
2121
2122 size_t cms_used = _cmsGen->cmsSpace()->used();
2123
2124 // update performance counters - this uses a special version of
2125 // update_counters() that allows the utilization to be passed as a
2126 // parameter, avoiding multiple calls to used().
2127 //
2128 _cmsGen->update_counters(cms_used);
2129
2130 bitMapLock()->unlock();
2131 releaseFreelistLocks();
2132
2133 if (!CleanChunkPoolAsync) {
2134 Chunk::clean_chunk_pool();
2135 }
2136
2137 set_did_compact(false);
2138 _between_prologue_and_epilogue = false; // ready for next cycle
2139 }
2140
2141 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2142 collector()->gc_epilogue(full);
2143
2144 // When using ParNew, promotion tracking should have already been
2145 // disabled. However, the prologue (which enables promotion
2146 // tracking) and epilogue are called irrespective of the type of
2147 // GC. So they will also be called before and after Full GCs, during
2148 // which promotion tracking will not be explicitly disabled. So,
2149 // it's safer to also disable it here too (to be symmetric with
2150 // enabling it in the prologue).
2151 for (uint i = 0; i < ParallelGCThreads; i++) {
2152 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2153 }
2154 }
2155
2156 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2157 assert(!incremental_collection_failed(), "Should have been cleared");
2158 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2159 cmsSpace()->gc_epilogue();
2160 // Print stat counters
2161 NOT_PRODUCT(
2162 assert(_numObjectsAllocated == 0, "check");
2163 assert(_numWordsAllocated == 0, "check");
2164 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2165 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2166 _numObjectsPromoted = 0;
2167 _numWordsPromoted = 0;
2168 )
2169
2170 // Call down the chain in contiguous_available needs the freelistLock
2171 // so print this out before releasing the freeListLock.
2172 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2173 }
2174
2175 #ifndef PRODUCT
2176 bool CMSCollector::have_cms_token() {
2177 Thread* thr = Thread::current();
2178 if (thr->is_VM_thread()) {
2179 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2180 } else if (thr->is_ConcurrentGC_thread()) {
2181 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2182 } else if (thr->is_GC_task_thread()) {
2183 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2184 ParGCRareEvent_lock->owned_by_self();
2185 }
2186 return false;
2187 }
2188
2189 // Check reachability of the given heap address in CMS generation,
2190 // treating all other generations as roots.
2191 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2192 // We could "guarantee" below, rather than assert, but I'll
2193 // leave these as "asserts" so that an adventurous debugger
2194 // could try this in the product build provided some subset of
2195 // the conditions were met, provided they were interested in the
2196 // results and knew that the computation below wouldn't interfere
2197 // with other concurrent computations mutating the structures
2198 // being read or written.
2199 assert(SafepointSynchronize::is_at_safepoint(),
2200 "Else mutations in object graph will make answer suspect");
2201 assert(have_cms_token(), "Should hold cms token");
2202 assert(haveFreelistLocks(), "must hold free list locks");
2203 assert_lock_strong(bitMapLock());
2204
2205 // Clear the marking bit map array before starting, but, just
2206 // for kicks, first report if the given address is already marked
2207 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2208 _markBitMap.isMarked(addr) ? "" : " not");
2209
2210 if (verify_after_remark()) {
2211 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2212 bool result = verification_mark_bm()->isMarked(addr);
2213 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2214 result ? "IS" : "is NOT");
2215 return result;
2216 } else {
2217 tty->print_cr("Could not compute result");
2218 return false;
2219 }
2220 }
2221 #endif
2222
2223 void
2224 CMSCollector::print_on_error(outputStream* st) {
2225 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2226 if (collector != NULL) {
2227 CMSBitMap* bitmap = &collector->_markBitMap;
2228 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2229 bitmap->print_on_error(st, " Bits: ");
2230
2231 st->cr();
2232
2233 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2234 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2235 mut_bitmap->print_on_error(st, " Bits: ");
2236 }
2237 }
2238
2239 ////////////////////////////////////////////////////////
2240 // CMS Verification Support
2241 ////////////////////////////////////////////////////////
2242 // Following the remark phase, the following invariant
2243 // should hold -- each object in the CMS heap which is
2244 // marked in markBitMap() should be marked in the verification_mark_bm().
2245
2246 class VerifyMarkedClosure: public BitMapClosure {
2247 CMSBitMap* _marks;
2248 bool _failed;
2249
2250 public:
2251 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2252
2253 bool do_bit(size_t offset) {
2254 HeapWord* addr = _marks->offsetToHeapWord(offset);
2255 if (!_marks->isMarked(addr)) {
2256 Log(gc, verify) log;
2257 ResourceMark rm;
2258 LogStream ls(log.error());
2259 oop(addr)->print_on(&ls);
2260 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2261 _failed = true;
2262 }
2263 return true;
2264 }
2265
2266 bool failed() { return _failed; }
2267 };
2268
2269 bool CMSCollector::verify_after_remark() {
2270 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2271 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2272 static bool init = false;
2273
2274 assert(SafepointSynchronize::is_at_safepoint(),
2275 "Else mutations in object graph will make answer suspect");
2276 assert(have_cms_token(),
2277 "Else there may be mutual interference in use of "
2278 " verification data structures");
2279 assert(_collectorState > Marking && _collectorState <= Sweeping,
2280 "Else marking info checked here may be obsolete");
2281 assert(haveFreelistLocks(), "must hold free list locks");
2282 assert_lock_strong(bitMapLock());
2283
2284
2285 // Allocate marking bit map if not already allocated
2286 if (!init) { // first time
2287 if (!verification_mark_bm()->allocate(_span)) {
2288 return false;
2289 }
2290 init = true;
2291 }
2292
2293 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2294
2295 // Turn off refs discovery -- so we will be tracing through refs.
2296 // This is as intended, because by this time
2297 // GC must already have cleared any refs that need to be cleared,
2298 // and traced those that need to be marked; moreover,
2299 // the marking done here is not going to interfere in any
2300 // way with the marking information used by GC.
2301 NoRefDiscovery no_discovery(ref_processor());
2302
2303 #if COMPILER2_OR_JVMCI
2304 DerivedPointerTableDeactivate dpt_deact;
2305 #endif
2306
2307 // Clear any marks from a previous round
2308 verification_mark_bm()->clear_all();
2309 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2310 verify_work_stacks_empty();
2311
2312 CMSHeap* heap = CMSHeap::heap();
2313 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2314 // Update the saved marks which may affect the root scans.
2315 heap->save_marks();
2316
2317 if (CMSRemarkVerifyVariant == 1) {
2318 // In this first variant of verification, we complete
2319 // all marking, then check if the new marks-vector is
2320 // a subset of the CMS marks-vector.
2321 verify_after_remark_work_1();
2322 } else {
2323 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2324 // In this second variant of verification, we flag an error
2325 // (i.e. an object reachable in the new marks-vector not reachable
2326 // in the CMS marks-vector) immediately, also indicating the
2327 // identify of an object (A) that references the unmarked object (B) --
2328 // presumably, a mutation to A failed to be picked up by preclean/remark?
2329 verify_after_remark_work_2();
2330 }
2331
2332 return true;
2333 }
2334
2335 void CMSCollector::verify_after_remark_work_1() {
2336 ResourceMark rm;
2337 HandleMark hm;
2338 CMSHeap* heap = CMSHeap::heap();
2339
2340 // Get a clear set of claim bits for the roots processing to work with.
2341 ClassLoaderDataGraph::clear_claimed_marks();
2342
2343 // Mark from roots one level into CMS
2344 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2345 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2346
2347 {
2348 StrongRootsScope srs(1);
2349
2350 heap->cms_process_roots(&srs,
2351 true, // young gen as roots
2352 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2353 should_unload_classes(),
2354 ¬Older,
2355 NULL);
2356 }
2357
2358 // Now mark from the roots
2359 MarkFromRootsClosure markFromRootsClosure(this, _span,
2360 verification_mark_bm(), verification_mark_stack(),
2361 false /* don't yield */, true /* verifying */);
2362 assert(_restart_addr == NULL, "Expected pre-condition");
2363 verification_mark_bm()->iterate(&markFromRootsClosure);
2364 while (_restart_addr != NULL) {
2365 // Deal with stack overflow: by restarting at the indicated
2366 // address.
2367 HeapWord* ra = _restart_addr;
2368 markFromRootsClosure.reset(ra);
2369 _restart_addr = NULL;
2370 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2371 }
2372 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2373 verify_work_stacks_empty();
2374
2375 // Marking completed -- now verify that each bit marked in
2376 // verification_mark_bm() is also marked in markBitMap(); flag all
2377 // errors by printing corresponding objects.
2378 VerifyMarkedClosure vcl(markBitMap());
2379 verification_mark_bm()->iterate(&vcl);
2380 if (vcl.failed()) {
2381 Log(gc, verify) log;
2382 log.error("Failed marking verification after remark");
2383 ResourceMark rm;
2384 LogStream ls(log.error());
2385 heap->print_on(&ls);
2386 fatal("CMS: failed marking verification after remark");
2387 }
2388 }
2389
2390 class VerifyCLDOopsCLDClosure : public CLDClosure {
2391 class VerifyCLDOopsClosure : public OopClosure {
2392 CMSBitMap* _bitmap;
2393 public:
2394 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2395 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2396 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2397 } _oop_closure;
2398 public:
2399 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2400 void do_cld(ClassLoaderData* cld) {
2401 cld->oops_do(&_oop_closure, false, false);
2402 }
2403 };
2404
2405 void CMSCollector::verify_after_remark_work_2() {
2406 ResourceMark rm;
2407 HandleMark hm;
2408 CMSHeap* heap = CMSHeap::heap();
2409
2410 // Get a clear set of claim bits for the roots processing to work with.
2411 ClassLoaderDataGraph::clear_claimed_marks();
2412
2413 // Mark from roots one level into CMS
2414 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2415 markBitMap());
2416 CLDToOopClosure cld_closure(¬Older, true);
2417
2418 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2419
2420 {
2421 StrongRootsScope srs(1);
2422
2423 heap->cms_process_roots(&srs,
2424 true, // young gen as roots
2425 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2426 should_unload_classes(),
2427 ¬Older,
2428 &cld_closure);
2429 }
2430
2431 // Now mark from the roots
2432 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2433 verification_mark_bm(), markBitMap(), verification_mark_stack());
2434 assert(_restart_addr == NULL, "Expected pre-condition");
2435 verification_mark_bm()->iterate(&markFromRootsClosure);
2436 while (_restart_addr != NULL) {
2437 // Deal with stack overflow: by restarting at the indicated
2438 // address.
2439 HeapWord* ra = _restart_addr;
2440 markFromRootsClosure.reset(ra);
2441 _restart_addr = NULL;
2442 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2443 }
2444 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2445 verify_work_stacks_empty();
2446
2447 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm());
2448 ClassLoaderDataGraph::cld_do(&verify_cld_oops);
2449
2450 // Marking completed -- now verify that each bit marked in
2451 // verification_mark_bm() is also marked in markBitMap(); flag all
2452 // errors by printing corresponding objects.
2453 VerifyMarkedClosure vcl(markBitMap());
2454 verification_mark_bm()->iterate(&vcl);
2455 assert(!vcl.failed(), "Else verification above should not have succeeded");
2456 }
2457
2458 void ConcurrentMarkSweepGeneration::save_marks() {
2459 // delegate to CMS space
2460 cmsSpace()->save_marks();
2461 }
2462
2463 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2464 return cmsSpace()->no_allocs_since_save_marks();
2465 }
2466
2467 void
2468 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2469 if (freelistLock()->owned_by_self()) {
2470 Generation::oop_iterate(cl);
2471 } else {
2472 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2473 Generation::oop_iterate(cl);
2474 }
2475 }
2476
2477 void
2478 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2479 if (freelistLock()->owned_by_self()) {
2480 Generation::object_iterate(cl);
2481 } else {
2482 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2483 Generation::object_iterate(cl);
2484 }
2485 }
2486
2487 void
2488 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2489 if (freelistLock()->owned_by_self()) {
2490 Generation::safe_object_iterate(cl);
2491 } else {
2492 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2493 Generation::safe_object_iterate(cl);
2494 }
2495 }
2496
2497 void
2498 ConcurrentMarkSweepGeneration::post_compact() {
2499 }
2500
2501 void
2502 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2503 // Fix the linear allocation blocks to look like free blocks.
2504
2505 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2506 // are not called when the heap is verified during universe initialization and
2507 // at vm shutdown.
2508 if (freelistLock()->owned_by_self()) {
2509 cmsSpace()->prepare_for_verify();
2510 } else {
2511 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2512 cmsSpace()->prepare_for_verify();
2513 }
2514 }
2515
2516 void
2517 ConcurrentMarkSweepGeneration::verify() {
2518 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2519 // are not called when the heap is verified during universe initialization and
2520 // at vm shutdown.
2521 if (freelistLock()->owned_by_self()) {
2522 cmsSpace()->verify();
2523 } else {
2524 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2525 cmsSpace()->verify();
2526 }
2527 }
2528
2529 void CMSCollector::verify() {
2530 _cmsGen->verify();
2531 }
2532
2533 #ifndef PRODUCT
2534 bool CMSCollector::overflow_list_is_empty() const {
2535 assert(_num_par_pushes >= 0, "Inconsistency");
2536 if (_overflow_list == NULL) {
2537 assert(_num_par_pushes == 0, "Inconsistency");
2538 }
2539 return _overflow_list == NULL;
2540 }
2541
2542 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2543 // merely consolidate assertion checks that appear to occur together frequently.
2544 void CMSCollector::verify_work_stacks_empty() const {
2545 assert(_markStack.isEmpty(), "Marking stack should be empty");
2546 assert(overflow_list_is_empty(), "Overflow list should be empty");
2547 }
2548
2549 void CMSCollector::verify_overflow_empty() const {
2550 assert(overflow_list_is_empty(), "Overflow list should be empty");
2551 assert(no_preserved_marks(), "No preserved marks");
2552 }
2553 #endif // PRODUCT
2554
2555 // Decide if we want to enable class unloading as part of the
2556 // ensuing concurrent GC cycle. We will collect and
2557 // unload classes if it's the case that:
2558 // (a) class unloading is enabled at the command line, and
2559 // (b) old gen is getting really full
2560 // NOTE: Provided there is no change in the state of the heap between
2561 // calls to this method, it should have idempotent results. Moreover,
2562 // its results should be monotonically increasing (i.e. going from 0 to 1,
2563 // but not 1 to 0) between successive calls between which the heap was
2564 // not collected. For the implementation below, it must thus rely on
2565 // the property that concurrent_cycles_since_last_unload()
2566 // will not decrease unless a collection cycle happened and that
2567 // _cmsGen->is_too_full() are
2568 // themselves also monotonic in that sense. See check_monotonicity()
2569 // below.
2570 void CMSCollector::update_should_unload_classes() {
2571 _should_unload_classes = false;
2572 if (CMSClassUnloadingEnabled) {
2573 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2574 CMSClassUnloadingMaxInterval)
2575 || _cmsGen->is_too_full();
2576 }
2577 }
2578
2579 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2580 bool res = should_concurrent_collect();
2581 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2582 return res;
2583 }
2584
2585 void CMSCollector::setup_cms_unloading_and_verification_state() {
2586 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2587 || VerifyBeforeExit;
2588 const int rso = GenCollectedHeap::SO_AllCodeCache;
2589
2590 // We set the proper root for this CMS cycle here.
2591 if (should_unload_classes()) { // Should unload classes this cycle
2592 remove_root_scanning_option(rso); // Shrink the root set appropriately
2593 set_verifying(should_verify); // Set verification state for this cycle
2594 return; // Nothing else needs to be done at this time
2595 }
2596
2597 // Not unloading classes this cycle
2598 assert(!should_unload_classes(), "Inconsistency!");
2599
2600 // If we are not unloading classes then add SO_AllCodeCache to root
2601 // scanning options.
2602 add_root_scanning_option(rso);
2603
2604 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2605 set_verifying(true);
2606 } else if (verifying() && !should_verify) {
2607 // We were verifying, but some verification flags got disabled.
2608 set_verifying(false);
2609 // Exclude symbols, strings and code cache elements from root scanning to
2610 // reduce IM and RM pauses.
2611 remove_root_scanning_option(rso);
2612 }
2613 }
2614
2615
2616 #ifndef PRODUCT
2617 HeapWord* CMSCollector::block_start(const void* p) const {
2618 const HeapWord* addr = (HeapWord*)p;
2619 if (_span.contains(p)) {
2620 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2621 return _cmsGen->cmsSpace()->block_start(p);
2622 }
2623 }
2624 return NULL;
2625 }
2626 #endif
2627
2628 HeapWord*
2629 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2630 bool tlab,
2631 bool parallel) {
2632 CMSSynchronousYieldRequest yr;
2633 assert(!tlab, "Can't deal with TLAB allocation");
2634 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2635 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2636 if (GCExpandToAllocateDelayMillis > 0) {
2637 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2638 }
2639 return have_lock_and_allocate(word_size, tlab);
2640 }
2641
2642 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2643 size_t bytes,
2644 size_t expand_bytes,
2645 CMSExpansionCause::Cause cause)
2646 {
2647
2648 bool success = expand(bytes, expand_bytes);
2649
2650 // remember why we expanded; this information is used
2651 // by shouldConcurrentCollect() when making decisions on whether to start
2652 // a new CMS cycle.
2653 if (success) {
2654 set_expansion_cause(cause);
2655 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause));
2656 }
2657 }
2658
2659 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2660 HeapWord* res = NULL;
2661 MutexLocker x(ParGCRareEvent_lock);
2662 while (true) {
2663 // Expansion by some other thread might make alloc OK now:
2664 res = ps->lab.alloc(word_sz);
2665 if (res != NULL) return res;
2666 // If there's not enough expansion space available, give up.
2667 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2668 return NULL;
2669 }
2670 // Otherwise, we try expansion.
2671 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2672 // Now go around the loop and try alloc again;
2673 // A competing par_promote might beat us to the expansion space,
2674 // so we may go around the loop again if promotion fails again.
2675 if (GCExpandToAllocateDelayMillis > 0) {
2676 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2677 }
2678 }
2679 }
2680
2681
2682 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2683 PromotionInfo* promo) {
2684 MutexLocker x(ParGCRareEvent_lock);
2685 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2686 while (true) {
2687 // Expansion by some other thread might make alloc OK now:
2688 if (promo->ensure_spooling_space()) {
2689 assert(promo->has_spooling_space(),
2690 "Post-condition of successful ensure_spooling_space()");
2691 return true;
2692 }
2693 // If there's not enough expansion space available, give up.
2694 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2695 return false;
2696 }
2697 // Otherwise, we try expansion.
2698 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2699 // Now go around the loop and try alloc again;
2700 // A competing allocation might beat us to the expansion space,
2701 // so we may go around the loop again if allocation fails again.
2702 if (GCExpandToAllocateDelayMillis > 0) {
2703 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2704 }
2705 }
2706 }
2707
2708 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2709 // Only shrink if a compaction was done so that all the free space
2710 // in the generation is in a contiguous block at the end.
2711 if (did_compact()) {
2712 CardGeneration::shrink(bytes);
2713 }
2714 }
2715
2716 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2717 assert_locked_or_safepoint(Heap_lock);
2718 }
2719
2720 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2721 assert_locked_or_safepoint(Heap_lock);
2722 assert_lock_strong(freelistLock());
2723 log_trace(gc)("Shrinking of CMS not yet implemented");
2724 return;
2725 }
2726
2727
2728 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2729 // phases.
2730 class CMSPhaseAccounting: public StackObj {
2731 public:
2732 CMSPhaseAccounting(CMSCollector *collector,
2733 const char *title);
2734 ~CMSPhaseAccounting();
2735
2736 private:
2737 CMSCollector *_collector;
2738 const char *_title;
2739 GCTraceConcTime(Info, gc) _trace_time;
2740
2741 public:
2742 // Not MT-safe; so do not pass around these StackObj's
2743 // where they may be accessed by other threads.
2744 double wallclock_millis() {
2745 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2746 }
2747 };
2748
2749 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2750 const char *title) :
2751 _collector(collector), _title(title), _trace_time(title) {
2752
2753 _collector->resetYields();
2754 _collector->resetTimer();
2755 _collector->startTimer();
2756 _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2757 }
2758
2759 CMSPhaseAccounting::~CMSPhaseAccounting() {
2760 _collector->gc_timer_cm()->register_gc_concurrent_end();
2761 _collector->stopTimer();
2762 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2763 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2764 }
2765
2766 // CMS work
2767
2768 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2769 class CMSParMarkTask : public AbstractGangTask {
2770 protected:
2771 CMSCollector* _collector;
2772 uint _n_workers;
2773 OopStorage::ParState<false, false> _par_state_string;
2774 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2775 AbstractGangTask(name),
2776 _collector(collector),
2777 _n_workers(n_workers),
2778 _par_state_string(StringTable::weak_storage()) {
2779 StringTable::reset_dead_counter();
2780 }
2781 ~CMSParMarkTask() {
2782 StringTable::finish_dead_counter();
2783 }
2784 // Work method in support of parallel rescan ... of young gen spaces
2785 void do_young_space_rescan(OopsInGenClosure* cl,
2786 ContiguousSpace* space,
2787 HeapWord** chunk_array, size_t chunk_top);
2788 void work_on_young_gen_roots(OopsInGenClosure* cl);
2789 };
2790
2791 // Parallel initial mark task
2792 class CMSParInitialMarkTask: public CMSParMarkTask {
2793 StrongRootsScope* _strong_roots_scope;
2794 public:
2795 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2796 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2797 _strong_roots_scope(strong_roots_scope) {}
2798 void work(uint worker_id);
2799 };
2800
2801 // Checkpoint the roots into this generation from outside
2802 // this generation. [Note this initial checkpoint need only
2803 // be approximate -- we'll do a catch up phase subsequently.]
2804 void CMSCollector::checkpointRootsInitial() {
2805 assert(_collectorState == InitialMarking, "Wrong collector state");
2806 check_correct_thread_executing();
2807 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
2808
2809 save_heap_summary();
2810 report_heap_summary(GCWhen::BeforeGC);
2811
2812 ReferenceProcessor* rp = ref_processor();
2813 assert(_restart_addr == NULL, "Control point invariant");
2814 {
2815 // acquire locks for subsequent manipulations
2816 MutexLockerEx x(bitMapLock(),
2817 Mutex::_no_safepoint_check_flag);
2818 checkpointRootsInitialWork();
2819 // enable ("weak") refs discovery
2820 rp->enable_discovery();
2821 _collectorState = Marking;
2822 }
2823 }
2824
2825 void CMSCollector::checkpointRootsInitialWork() {
2826 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2827 assert(_collectorState == InitialMarking, "just checking");
2828
2829 // Already have locks.
2830 assert_lock_strong(bitMapLock());
2831 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2832
2833 // Setup the verification and class unloading state for this
2834 // CMS collection cycle.
2835 setup_cms_unloading_and_verification_state();
2836
2837 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2838
2839 // Reset all the PLAB chunk arrays if necessary.
2840 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2841 reset_survivor_plab_arrays();
2842 }
2843
2844 ResourceMark rm;
2845 HandleMark hm;
2846
2847 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2848 CMSHeap* heap = CMSHeap::heap();
2849
2850 verify_work_stacks_empty();
2851 verify_overflow_empty();
2852
2853 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2854 // Update the saved marks which may affect the root scans.
2855 heap->save_marks();
2856
2857 // weak reference processing has not started yet.
2858 ref_processor()->set_enqueuing_is_done(false);
2859
2860 // Need to remember all newly created CLDs,
2861 // so that we can guarantee that the remark finds them.
2862 ClassLoaderDataGraph::remember_new_clds(true);
2863
2864 // Whenever a CLD is found, it will be claimed before proceeding to mark
2865 // the klasses. The claimed marks need to be cleared before marking starts.
2866 ClassLoaderDataGraph::clear_claimed_marks();
2867
2868 print_eden_and_survivor_chunk_arrays();
2869
2870 {
2871 #if COMPILER2_OR_JVMCI
2872 DerivedPointerTableDeactivate dpt_deact;
2873 #endif
2874 if (CMSParallelInitialMarkEnabled) {
2875 // The parallel version.
2876 WorkGang* workers = heap->workers();
2877 assert(workers != NULL, "Need parallel worker threads.");
2878 uint n_workers = workers->active_workers();
2879
2880 StrongRootsScope srs(n_workers);
2881
2882 CMSParInitialMarkTask tsk(this, &srs, n_workers);
2883 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2884 // If the total workers is greater than 1, then multiple workers
2885 // may be used at some time and the initialization has been set
2886 // such that the single threaded path cannot be used.
2887 if (workers->total_workers() > 1) {
2888 workers->run_task(&tsk);
2889 } else {
2890 tsk.work(0);
2891 }
2892 } else {
2893 // The serial version.
2894 CLDToOopClosure cld_closure(¬Older, true);
2895 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2896
2897 StrongRootsScope srs(1);
2898
2899 heap->cms_process_roots(&srs,
2900 true, // young gen as roots
2901 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2902 should_unload_classes(),
2903 ¬Older,
2904 &cld_closure);
2905 }
2906 }
2907
2908 // Clear mod-union table; it will be dirtied in the prologue of
2909 // CMS generation per each young generation collection.
2910
2911 assert(_modUnionTable.isAllClear(),
2912 "Was cleared in most recent final checkpoint phase"
2913 " or no bits are set in the gc_prologue before the start of the next "
2914 "subsequent marking phase.");
2915
2916 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be");
2917
2918 // Save the end of the used_region of the constituent generations
2919 // to be used to limit the extent of sweep in each generation.
2920 save_sweep_limits();
2921 verify_overflow_empty();
2922 }
2923
2924 bool CMSCollector::markFromRoots() {
2925 // we might be tempted to assert that:
2926 // assert(!SafepointSynchronize::is_at_safepoint(),
2927 // "inconsistent argument?");
2928 // However that wouldn't be right, because it's possible that
2929 // a safepoint is indeed in progress as a young generation
2930 // stop-the-world GC happens even as we mark in this generation.
2931 assert(_collectorState == Marking, "inconsistent state?");
2932 check_correct_thread_executing();
2933 verify_overflow_empty();
2934
2935 // Weak ref discovery note: We may be discovering weak
2936 // refs in this generation concurrent (but interleaved) with
2937 // weak ref discovery by the young generation collector.
2938
2939 CMSTokenSyncWithLocks ts(true, bitMapLock());
2940 GCTraceCPUTime tcpu;
2941 CMSPhaseAccounting pa(this, "Concurrent Mark");
2942 bool res = markFromRootsWork();
2943 if (res) {
2944 _collectorState = Precleaning;
2945 } else { // We failed and a foreground collection wants to take over
2946 assert(_foregroundGCIsActive, "internal state inconsistency");
2947 assert(_restart_addr == NULL, "foreground will restart from scratch");
2948 log_debug(gc)("bailing out to foreground collection");
2949 }
2950 verify_overflow_empty();
2951 return res;
2952 }
2953
2954 bool CMSCollector::markFromRootsWork() {
2955 // iterate over marked bits in bit map, doing a full scan and mark
2956 // from these roots using the following algorithm:
2957 // . if oop is to the right of the current scan pointer,
2958 // mark corresponding bit (we'll process it later)
2959 // . else (oop is to left of current scan pointer)
2960 // push oop on marking stack
2961 // . drain the marking stack
2962
2963 // Note that when we do a marking step we need to hold the
2964 // bit map lock -- recall that direct allocation (by mutators)
2965 // and promotion (by the young generation collector) is also
2966 // marking the bit map. [the so-called allocate live policy.]
2967 // Because the implementation of bit map marking is not
2968 // robust wrt simultaneous marking of bits in the same word,
2969 // we need to make sure that there is no such interference
2970 // between concurrent such updates.
2971
2972 // already have locks
2973 assert_lock_strong(bitMapLock());
2974
2975 verify_work_stacks_empty();
2976 verify_overflow_empty();
2977 bool result = false;
2978 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2979 result = do_marking_mt();
2980 } else {
2981 result = do_marking_st();
2982 }
2983 return result;
2984 }
2985
2986 // Forward decl
2987 class CMSConcMarkingTask;
2988
2989 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2990 CMSCollector* _collector;
2991 CMSConcMarkingTask* _task;
2992 public:
2993 virtual void yield();
2994
2995 // "n_threads" is the number of threads to be terminated.
2996 // "queue_set" is a set of work queues of other threads.
2997 // "collector" is the CMS collector associated with this task terminator.
2998 // "yield" indicates whether we need the gang as a whole to yield.
2999 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3000 ParallelTaskTerminator(n_threads, queue_set),
3001 _collector(collector) { }
3002
3003 void set_task(CMSConcMarkingTask* task) {
3004 _task = task;
3005 }
3006 };
3007
3008 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3009 CMSConcMarkingTask* _task;
3010 public:
3011 bool should_exit_termination();
3012 void set_task(CMSConcMarkingTask* task) {
3013 _task = task;
3014 }
3015 };
3016
3017 // MT Concurrent Marking Task
3018 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3019 CMSCollector* _collector;
3020 uint _n_workers; // requested/desired # workers
3021 bool _result;
3022 CompactibleFreeListSpace* _cms_space;
3023 char _pad_front[64]; // padding to ...
3024 HeapWord* volatile _global_finger; // ... avoid sharing cache line
3025 char _pad_back[64];
3026 HeapWord* _restart_addr;
3027
3028 // Exposed here for yielding support
3029 Mutex* const _bit_map_lock;
3030
3031 // The per thread work queues, available here for stealing
3032 OopTaskQueueSet* _task_queues;
3033
3034 // Termination (and yielding) support
3035 CMSConcMarkingTerminator _term;
3036 CMSConcMarkingTerminatorTerminator _term_term;
3037
3038 public:
3039 CMSConcMarkingTask(CMSCollector* collector,
3040 CompactibleFreeListSpace* cms_space,
3041 YieldingFlexibleWorkGang* workers,
3042 OopTaskQueueSet* task_queues):
3043 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3044 _collector(collector),
3045 _cms_space(cms_space),
3046 _n_workers(0), _result(true),
3047 _task_queues(task_queues),
3048 _term(_n_workers, task_queues, _collector),
3049 _bit_map_lock(collector->bitMapLock())
3050 {
3051 _requested_size = _n_workers;
3052 _term.set_task(this);
3053 _term_term.set_task(this);
3054 _restart_addr = _global_finger = _cms_space->bottom();
3055 }
3056
3057
3058 OopTaskQueueSet* task_queues() { return _task_queues; }
3059
3060 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3061
3062 HeapWord* volatile* global_finger_addr() { return &_global_finger; }
3063
3064 CMSConcMarkingTerminator* terminator() { return &_term; }
3065
3066 virtual void set_for_termination(uint active_workers) {
3067 terminator()->reset_for_reuse(active_workers);
3068 }
3069
3070 void work(uint worker_id);
3071 bool should_yield() {
3072 return ConcurrentMarkSweepThread::should_yield()
3073 && !_collector->foregroundGCIsActive();
3074 }
3075
3076 virtual void coordinator_yield(); // stuff done by coordinator
3077 bool result() { return _result; }
3078
3079 void reset(HeapWord* ra) {
3080 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3081 _restart_addr = _global_finger = ra;
3082 _term.reset_for_reuse();
3083 }
3084
3085 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3086 OopTaskQueue* work_q);
3087
3088 private:
3089 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3090 void do_work_steal(int i);
3091 void bump_global_finger(HeapWord* f);
3092 };
3093
3094 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3095 assert(_task != NULL, "Error");
3096 return _task->yielding();
3097 // Note that we do not need the disjunct || _task->should_yield() above
3098 // because we want terminating threads to yield only if the task
3099 // is already in the midst of yielding, which happens only after at least one
3100 // thread has yielded.
3101 }
3102
3103 void CMSConcMarkingTerminator::yield() {
3104 if (_task->should_yield()) {
3105 _task->yield();
3106 } else {
3107 ParallelTaskTerminator::yield();
3108 }
3109 }
3110
3111 ////////////////////////////////////////////////////////////////
3112 // Concurrent Marking Algorithm Sketch
3113 ////////////////////////////////////////////////////////////////
3114 // Until all tasks exhausted (both spaces):
3115 // -- claim next available chunk
3116 // -- bump global finger via CAS
3117 // -- find first object that starts in this chunk
3118 // and start scanning bitmap from that position
3119 // -- scan marked objects for oops
3120 // -- CAS-mark target, and if successful:
3121 // . if target oop is above global finger (volatile read)
3122 // nothing to do
3123 // . if target oop is in chunk and above local finger
3124 // then nothing to do
3125 // . else push on work-queue
3126 // -- Deal with possible overflow issues:
3127 // . local work-queue overflow causes stuff to be pushed on
3128 // global (common) overflow queue
3129 // . always first empty local work queue
3130 // . then get a batch of oops from global work queue if any
3131 // . then do work stealing
3132 // -- When all tasks claimed (both spaces)
3133 // and local work queue empty,
3134 // then in a loop do:
3135 // . check global overflow stack; steal a batch of oops and trace
3136 // . try to steal from other threads oif GOS is empty
3137 // . if neither is available, offer termination
3138 // -- Terminate and return result
3139 //
3140 void CMSConcMarkingTask::work(uint worker_id) {
3141 elapsedTimer _timer;
3142 ResourceMark rm;
3143 HandleMark hm;
3144
3145 DEBUG_ONLY(_collector->verify_overflow_empty();)
3146
3147 // Before we begin work, our work queue should be empty
3148 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3149 // Scan the bitmap covering _cms_space, tracing through grey objects.
3150 _timer.start();
3151 do_scan_and_mark(worker_id, _cms_space);
3152 _timer.stop();
3153 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3154
3155 // ... do work stealing
3156 _timer.reset();
3157 _timer.start();
3158 do_work_steal(worker_id);
3159 _timer.stop();
3160 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3161 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3162 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3163 // Note that under the current task protocol, the
3164 // following assertion is true even of the spaces
3165 // expanded since the completion of the concurrent
3166 // marking. XXX This will likely change under a strict
3167 // ABORT semantics.
3168 // After perm removal the comparison was changed to
3169 // greater than or equal to from strictly greater than.
3170 // Before perm removal the highest address sweep would
3171 // have been at the end of perm gen but now is at the
3172 // end of the tenured gen.
3173 assert(_global_finger >= _cms_space->end(),
3174 "All tasks have been completed");
3175 DEBUG_ONLY(_collector->verify_overflow_empty();)
3176 }
3177
3178 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3179 HeapWord* read = _global_finger;
3180 HeapWord* cur = read;
3181 while (f > read) {
3182 cur = read;
3183 read = Atomic::cmpxchg(f, &_global_finger, cur);
3184 if (cur == read) {
3185 // our cas succeeded
3186 assert(_global_finger >= f, "protocol consistency");
3187 break;
3188 }
3189 }
3190 }
3191
3192 // This is really inefficient, and should be redone by
3193 // using (not yet available) block-read and -write interfaces to the
3194 // stack and the work_queue. XXX FIX ME !!!
3195 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3196 OopTaskQueue* work_q) {
3197 // Fast lock-free check
3198 if (ovflw_stk->length() == 0) {
3199 return false;
3200 }
3201 assert(work_q->size() == 0, "Shouldn't steal");
3202 MutexLockerEx ml(ovflw_stk->par_lock(),
3203 Mutex::_no_safepoint_check_flag);
3204 // Grab up to 1/4 the size of the work queue
3205 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3206 (size_t)ParGCDesiredObjsFromOverflowList);
3207 num = MIN2(num, ovflw_stk->length());
3208 for (int i = (int) num; i > 0; i--) {
3209 oop cur = ovflw_stk->pop();
3210 assert(cur != NULL, "Counted wrong?");
3211 work_q->push(cur);
3212 }
3213 return num > 0;
3214 }
3215
3216 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3217 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3218 int n_tasks = pst->n_tasks();
3219 // We allow that there may be no tasks to do here because
3220 // we are restarting after a stack overflow.
3221 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3222 uint nth_task = 0;
3223
3224 HeapWord* aligned_start = sp->bottom();
3225 if (sp->used_region().contains(_restart_addr)) {
3226 // Align down to a card boundary for the start of 0th task
3227 // for this space.
3228 aligned_start = align_down(_restart_addr, CardTable::card_size);
3229 }
3230
3231 size_t chunk_size = sp->marking_task_size();
3232 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3233 // Having claimed the nth task in this space,
3234 // compute the chunk that it corresponds to:
3235 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3236 aligned_start + (nth_task+1)*chunk_size);
3237 // Try and bump the global finger via a CAS;
3238 // note that we need to do the global finger bump
3239 // _before_ taking the intersection below, because
3240 // the task corresponding to that region will be
3241 // deemed done even if the used_region() expands
3242 // because of allocation -- as it almost certainly will
3243 // during start-up while the threads yield in the
3244 // closure below.
3245 HeapWord* finger = span.end();
3246 bump_global_finger(finger); // atomically
3247 // There are null tasks here corresponding to chunks
3248 // beyond the "top" address of the space.
3249 span = span.intersection(sp->used_region());
3250 if (!span.is_empty()) { // Non-null task
3251 HeapWord* prev_obj;
3252 assert(!span.contains(_restart_addr) || nth_task == 0,
3253 "Inconsistency");
3254 if (nth_task == 0) {
3255 // For the 0th task, we'll not need to compute a block_start.
3256 if (span.contains(_restart_addr)) {
3257 // In the case of a restart because of stack overflow,
3258 // we might additionally skip a chunk prefix.
3259 prev_obj = _restart_addr;
3260 } else {
3261 prev_obj = span.start();
3262 }
3263 } else {
3264 // We want to skip the first object because
3265 // the protocol is to scan any object in its entirety
3266 // that _starts_ in this span; a fortiori, any
3267 // object starting in an earlier span is scanned
3268 // as part of an earlier claimed task.
3269 // Below we use the "careful" version of block_start
3270 // so we do not try to navigate uninitialized objects.
3271 prev_obj = sp->block_start_careful(span.start());
3272 // Below we use a variant of block_size that uses the
3273 // Printezis bits to avoid waiting for allocated
3274 // objects to become initialized/parsable.
3275 while (prev_obj < span.start()) {
3276 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3277 if (sz > 0) {
3278 prev_obj += sz;
3279 } else {
3280 // In this case we may end up doing a bit of redundant
3281 // scanning, but that appears unavoidable, short of
3282 // locking the free list locks; see bug 6324141.
3283 break;
3284 }
3285 }
3286 }
3287 if (prev_obj < span.end()) {
3288 MemRegion my_span = MemRegion(prev_obj, span.end());
3289 // Do the marking work within a non-empty span --
3290 // the last argument to the constructor indicates whether the
3291 // iteration should be incremental with periodic yields.
3292 ParMarkFromRootsClosure cl(this, _collector, my_span,
3293 &_collector->_markBitMap,
3294 work_queue(i),
3295 &_collector->_markStack);
3296 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3297 } // else nothing to do for this task
3298 } // else nothing to do for this task
3299 }
3300 // We'd be tempted to assert here that since there are no
3301 // more tasks left to claim in this space, the global_finger
3302 // must exceed space->top() and a fortiori space->end(). However,
3303 // that would not quite be correct because the bumping of
3304 // global_finger occurs strictly after the claiming of a task,
3305 // so by the time we reach here the global finger may not yet
3306 // have been bumped up by the thread that claimed the last
3307 // task.
3308 pst->all_tasks_completed();
3309 }
3310
3311 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3312 private:
3313 CMSCollector* _collector;
3314 CMSConcMarkingTask* _task;
3315 MemRegion _span;
3316 CMSBitMap* _bit_map;
3317 CMSMarkStack* _overflow_stack;
3318 OopTaskQueue* _work_queue;
3319 protected:
3320 DO_OOP_WORK_DEFN
3321 public:
3322 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3323 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3324 MetadataAwareOopClosure(collector->ref_processor()),
3325 _collector(collector),
3326 _task(task),
3327 _span(collector->_span),
3328 _work_queue(work_queue),
3329 _bit_map(bit_map),
3330 _overflow_stack(overflow_stack)
3331 { }
3332 virtual void do_oop(oop* p);
3333 virtual void do_oop(narrowOop* p);
3334
3335 void trim_queue(size_t max);
3336 void handle_stack_overflow(HeapWord* lost);
3337 void do_yield_check() {
3338 if (_task->should_yield()) {
3339 _task->yield();
3340 }
3341 }
3342 };
3343
3344 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3345
3346 // Grey object scanning during work stealing phase --
3347 // the salient assumption here is that any references
3348 // that are in these stolen objects being scanned must
3349 // already have been initialized (else they would not have
3350 // been published), so we do not need to check for
3351 // uninitialized objects before pushing here.
3352 void ParConcMarkingClosure::do_oop(oop obj) {
3353 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3354 HeapWord* addr = (HeapWord*)obj;
3355 // Check if oop points into the CMS generation
3356 // and is not marked
3357 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3358 // a white object ...
3359 // If we manage to "claim" the object, by being the
3360 // first thread to mark it, then we push it on our
3361 // marking stack
3362 if (_bit_map->par_mark(addr)) { // ... now grey
3363 // push on work queue (grey set)
3364 bool simulate_overflow = false;
3365 NOT_PRODUCT(
3366 if (CMSMarkStackOverflowALot &&
3367 _collector->simulate_overflow()) {
3368 // simulate a stack overflow
3369 simulate_overflow = true;
3370 }
3371 )
3372 if (simulate_overflow ||
3373 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3374 // stack overflow
3375 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3376 // We cannot assert that the overflow stack is full because
3377 // it may have been emptied since.
3378 assert(simulate_overflow ||
3379 _work_queue->size() == _work_queue->max_elems(),
3380 "Else push should have succeeded");
3381 handle_stack_overflow(addr);
3382 }
3383 } // Else, some other thread got there first
3384 do_yield_check();
3385 }
3386 }
3387
3388 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); }
3389 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3390
3391 void ParConcMarkingClosure::trim_queue(size_t max) {
3392 while (_work_queue->size() > max) {
3393 oop new_oop;
3394 if (_work_queue->pop_local(new_oop)) {
3395 assert(oopDesc::is_oop(new_oop), "Should be an oop");
3396 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3397 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3398 new_oop->oop_iterate(this); // do_oop() above
3399 do_yield_check();
3400 }
3401 }
3402 }
3403
3404 // Upon stack overflow, we discard (part of) the stack,
3405 // remembering the least address amongst those discarded
3406 // in CMSCollector's _restart_address.
3407 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3408 // We need to do this under a mutex to prevent other
3409 // workers from interfering with the work done below.
3410 MutexLockerEx ml(_overflow_stack->par_lock(),
3411 Mutex::_no_safepoint_check_flag);
3412 // Remember the least grey address discarded
3413 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3414 _collector->lower_restart_addr(ra);
3415 _overflow_stack->reset(); // discard stack contents
3416 _overflow_stack->expand(); // expand the stack if possible
3417 }
3418
3419
3420 void CMSConcMarkingTask::do_work_steal(int i) {
3421 OopTaskQueue* work_q = work_queue(i);
3422 oop obj_to_scan;
3423 CMSBitMap* bm = &(_collector->_markBitMap);
3424 CMSMarkStack* ovflw = &(_collector->_markStack);
3425 int* seed = _collector->hash_seed(i);
3426 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3427 while (true) {
3428 cl.trim_queue(0);
3429 assert(work_q->size() == 0, "Should have been emptied above");
3430 if (get_work_from_overflow_stack(ovflw, work_q)) {
3431 // Can't assert below because the work obtained from the
3432 // overflow stack may already have been stolen from us.
3433 // assert(work_q->size() > 0, "Work from overflow stack");
3434 continue;
3435 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3436 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3437 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3438 obj_to_scan->oop_iterate(&cl);
3439 } else if (terminator()->offer_termination(&_term_term)) {
3440 assert(work_q->size() == 0, "Impossible!");
3441 break;
3442 } else if (yielding() || should_yield()) {
3443 yield();
3444 }
3445 }
3446 }
3447
3448 // This is run by the CMS (coordinator) thread.
3449 void CMSConcMarkingTask::coordinator_yield() {
3450 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3451 "CMS thread should hold CMS token");
3452 // First give up the locks, then yield, then re-lock
3453 // We should probably use a constructor/destructor idiom to
3454 // do this unlock/lock or modify the MutexUnlocker class to
3455 // serve our purpose. XXX
3456 assert_lock_strong(_bit_map_lock);
3457 _bit_map_lock->unlock();
3458 ConcurrentMarkSweepThread::desynchronize(true);
3459 _collector->stopTimer();
3460 _collector->incrementYields();
3461
3462 // It is possible for whichever thread initiated the yield request
3463 // not to get a chance to wake up and take the bitmap lock between
3464 // this thread releasing it and reacquiring it. So, while the
3465 // should_yield() flag is on, let's sleep for a bit to give the
3466 // other thread a chance to wake up. The limit imposed on the number
3467 // of iterations is defensive, to avoid any unforseen circumstances
3468 // putting us into an infinite loop. Since it's always been this
3469 // (coordinator_yield()) method that was observed to cause the
3470 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3471 // which is by default non-zero. For the other seven methods that
3472 // also perform the yield operation, as are using a different
3473 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3474 // can enable the sleeping for those methods too, if necessary.
3475 // See 6442774.
3476 //
3477 // We really need to reconsider the synchronization between the GC
3478 // thread and the yield-requesting threads in the future and we
3479 // should really use wait/notify, which is the recommended
3480 // way of doing this type of interaction. Additionally, we should
3481 // consolidate the eight methods that do the yield operation and they
3482 // are almost identical into one for better maintainability and
3483 // readability. See 6445193.
3484 //
3485 // Tony 2006.06.29
3486 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3487 ConcurrentMarkSweepThread::should_yield() &&
3488 !CMSCollector::foregroundGCIsActive(); ++i) {
3489 os::sleep(Thread::current(), 1, false);
3490 }
3491
3492 ConcurrentMarkSweepThread::synchronize(true);
3493 _bit_map_lock->lock_without_safepoint_check();
3494 _collector->startTimer();
3495 }
3496
3497 bool CMSCollector::do_marking_mt() {
3498 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3499 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3500 conc_workers()->active_workers(),
3501 Threads::number_of_non_daemon_threads());
3502 num_workers = conc_workers()->update_active_workers(num_workers);
3503 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3504
3505 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3506
3507 CMSConcMarkingTask tsk(this,
3508 cms_space,
3509 conc_workers(),
3510 task_queues());
3511
3512 // Since the actual number of workers we get may be different
3513 // from the number we requested above, do we need to do anything different
3514 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3515 // class?? XXX
3516 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3517
3518 // Refs discovery is already non-atomic.
3519 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3520 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3521 conc_workers()->start_task(&tsk);
3522 while (tsk.yielded()) {
3523 tsk.coordinator_yield();
3524 conc_workers()->continue_task(&tsk);
3525 }
3526 // If the task was aborted, _restart_addr will be non-NULL
3527 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3528 while (_restart_addr != NULL) {
3529 // XXX For now we do not make use of ABORTED state and have not
3530 // yet implemented the right abort semantics (even in the original
3531 // single-threaded CMS case). That needs some more investigation
3532 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3533 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3534 // If _restart_addr is non-NULL, a marking stack overflow
3535 // occurred; we need to do a fresh marking iteration from the
3536 // indicated restart address.
3537 if (_foregroundGCIsActive) {
3538 // We may be running into repeated stack overflows, having
3539 // reached the limit of the stack size, while making very
3540 // slow forward progress. It may be best to bail out and
3541 // let the foreground collector do its job.
3542 // Clear _restart_addr, so that foreground GC
3543 // works from scratch. This avoids the headache of
3544 // a "rescan" which would otherwise be needed because
3545 // of the dirty mod union table & card table.
3546 _restart_addr = NULL;
3547 return false;
3548 }
3549 // Adjust the task to restart from _restart_addr
3550 tsk.reset(_restart_addr);
3551 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3552 _restart_addr);
3553 _restart_addr = NULL;
3554 // Get the workers going again
3555 conc_workers()->start_task(&tsk);
3556 while (tsk.yielded()) {
3557 tsk.coordinator_yield();
3558 conc_workers()->continue_task(&tsk);
3559 }
3560 }
3561 assert(tsk.completed(), "Inconsistency");
3562 assert(tsk.result() == true, "Inconsistency");
3563 return true;
3564 }
3565
3566 bool CMSCollector::do_marking_st() {
3567 ResourceMark rm;
3568 HandleMark hm;
3569
3570 // Temporarily make refs discovery single threaded (non-MT)
3571 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3572 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3573 &_markStack, CMSYield);
3574 // the last argument to iterate indicates whether the iteration
3575 // should be incremental with periodic yields.
3576 _markBitMap.iterate(&markFromRootsClosure);
3577 // If _restart_addr is non-NULL, a marking stack overflow
3578 // occurred; we need to do a fresh iteration from the
3579 // indicated restart address.
3580 while (_restart_addr != NULL) {
3581 if (_foregroundGCIsActive) {
3582 // We may be running into repeated stack overflows, having
3583 // reached the limit of the stack size, while making very
3584 // slow forward progress. It may be best to bail out and
3585 // let the foreground collector do its job.
3586 // Clear _restart_addr, so that foreground GC
3587 // works from scratch. This avoids the headache of
3588 // a "rescan" which would otherwise be needed because
3589 // of the dirty mod union table & card table.
3590 _restart_addr = NULL;
3591 return false; // indicating failure to complete marking
3592 }
3593 // Deal with stack overflow:
3594 // we restart marking from _restart_addr
3595 HeapWord* ra = _restart_addr;
3596 markFromRootsClosure.reset(ra);
3597 _restart_addr = NULL;
3598 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3599 }
3600 return true;
3601 }
3602
3603 void CMSCollector::preclean() {
3604 check_correct_thread_executing();
3605 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3606 verify_work_stacks_empty();
3607 verify_overflow_empty();
3608 _abort_preclean = false;
3609 if (CMSPrecleaningEnabled) {
3610 if (!CMSEdenChunksRecordAlways) {
3611 _eden_chunk_index = 0;
3612 }
3613 size_t used = get_eden_used();
3614 size_t capacity = get_eden_capacity();
3615 // Don't start sampling unless we will get sufficiently
3616 // many samples.
3617 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3618 * CMSScheduleRemarkEdenPenetration)) {
3619 _start_sampling = true;
3620 } else {
3621 _start_sampling = false;
3622 }
3623 GCTraceCPUTime tcpu;
3624 CMSPhaseAccounting pa(this, "Concurrent Preclean");
3625 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3626 }
3627 CMSTokenSync x(true); // is cms thread
3628 if (CMSPrecleaningEnabled) {
3629 sample_eden();
3630 _collectorState = AbortablePreclean;
3631 } else {
3632 _collectorState = FinalMarking;
3633 }
3634 verify_work_stacks_empty();
3635 verify_overflow_empty();
3636 }
3637
3638 // Try and schedule the remark such that young gen
3639 // occupancy is CMSScheduleRemarkEdenPenetration %.
3640 void CMSCollector::abortable_preclean() {
3641 check_correct_thread_executing();
3642 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3643 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3644
3645 // If Eden's current occupancy is below this threshold,
3646 // immediately schedule the remark; else preclean
3647 // past the next scavenge in an effort to
3648 // schedule the pause as described above. By choosing
3649 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3650 // we will never do an actual abortable preclean cycle.
3651 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3652 GCTraceCPUTime tcpu;
3653 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3654 // We need more smarts in the abortable preclean
3655 // loop below to deal with cases where allocation
3656 // in young gen is very very slow, and our precleaning
3657 // is running a losing race against a horde of
3658 // mutators intent on flooding us with CMS updates
3659 // (dirty cards).
3660 // One, admittedly dumb, strategy is to give up
3661 // after a certain number of abortable precleaning loops
3662 // or after a certain maximum time. We want to make
3663 // this smarter in the next iteration.
3664 // XXX FIX ME!!! YSR
3665 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3666 while (!(should_abort_preclean() ||
3667 ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3668 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3669 cumworkdone += workdone;
3670 loops++;
3671 // Voluntarily terminate abortable preclean phase if we have
3672 // been at it for too long.
3673 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3674 loops >= CMSMaxAbortablePrecleanLoops) {
3675 log_debug(gc)(" CMS: abort preclean due to loops ");
3676 break;
3677 }
3678 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3679 log_debug(gc)(" CMS: abort preclean due to time ");
3680 break;
3681 }
3682 // If we are doing little work each iteration, we should
3683 // take a short break.
3684 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3685 // Sleep for some time, waiting for work to accumulate
3686 stopTimer();
3687 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3688 startTimer();
3689 waited++;
3690 }
3691 }
3692 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3693 loops, waited, cumworkdone);
3694 }
3695 CMSTokenSync x(true); // is cms thread
3696 if (_collectorState != Idling) {
3697 assert(_collectorState == AbortablePreclean,
3698 "Spontaneous state transition?");
3699 _collectorState = FinalMarking;
3700 } // Else, a foreground collection completed this CMS cycle.
3701 return;
3702 }
3703
3704 // Respond to an Eden sampling opportunity
3705 void CMSCollector::sample_eden() {
3706 // Make sure a young gc cannot sneak in between our
3707 // reading and recording of a sample.
3708 assert(Thread::current()->is_ConcurrentGC_thread(),
3709 "Only the cms thread may collect Eden samples");
3710 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3711 "Should collect samples while holding CMS token");
3712 if (!_start_sampling) {
3713 return;
3714 }
3715 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3716 // is populated by the young generation.
3717 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3718 if (_eden_chunk_index < _eden_chunk_capacity) {
3719 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3720 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3721 "Unexpected state of Eden");
3722 // We'd like to check that what we just sampled is an oop-start address;
3723 // however, we cannot do that here since the object may not yet have been
3724 // initialized. So we'll instead do the check when we _use_ this sample
3725 // later.
3726 if (_eden_chunk_index == 0 ||
3727 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3728 _eden_chunk_array[_eden_chunk_index-1])
3729 >= CMSSamplingGrain)) {
3730 _eden_chunk_index++; // commit sample
3731 }
3732 }
3733 }
3734 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3735 size_t used = get_eden_used();
3736 size_t capacity = get_eden_capacity();
3737 assert(used <= capacity, "Unexpected state of Eden");
3738 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3739 _abort_preclean = true;
3740 }
3741 }
3742 }
3743
3744 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3745 assert(_collectorState == Precleaning ||
3746 _collectorState == AbortablePreclean, "incorrect state");
3747 ResourceMark rm;
3748 HandleMark hm;
3749
3750 // Precleaning is currently not MT but the reference processor
3751 // may be set for MT. Disable it temporarily here.
3752 ReferenceProcessor* rp = ref_processor();
3753 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3754
3755 // Do one pass of scrubbing the discovered reference lists
3756 // to remove any reference objects with strongly-reachable
3757 // referents.
3758 if (clean_refs) {
3759 CMSPrecleanRefsYieldClosure yield_cl(this);
3760 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal");
3761 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3762 &_markStack, true /* preclean */);
3763 CMSDrainMarkingStackClosure complete_trace(this,
3764 _span, &_markBitMap, &_markStack,
3765 &keep_alive, true /* preclean */);
3766
3767 // We don't want this step to interfere with a young
3768 // collection because we don't want to take CPU
3769 // or memory bandwidth away from the young GC threads
3770 // (which may be as many as there are CPUs).
3771 // Note that we don't need to protect ourselves from
3772 // interference with mutators because they can't
3773 // manipulate the discovered reference lists nor affect
3774 // the computed reachability of the referents, the
3775 // only properties manipulated by the precleaning
3776 // of these reference lists.
3777 stopTimer();
3778 CMSTokenSyncWithLocks x(true /* is cms thread */,
3779 bitMapLock());
3780 startTimer();
3781 sample_eden();
3782
3783 // The following will yield to allow foreground
3784 // collection to proceed promptly. XXX YSR:
3785 // The code in this method may need further
3786 // tweaking for better performance and some restructuring
3787 // for cleaner interfaces.
3788 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3789 rp->preclean_discovered_references(
3790 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3791 gc_timer);
3792 }
3793
3794 if (clean_survivor) { // preclean the active survivor space(s)
3795 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3796 &_markBitMap, &_modUnionTable,
3797 &_markStack, true /* precleaning phase */);
3798 stopTimer();
3799 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3800 bitMapLock());
3801 startTimer();
3802 unsigned int before_count =
3803 CMSHeap::heap()->total_collections();
3804 SurvivorSpacePrecleanClosure
3805 sss_cl(this, _span, &_markBitMap, &_markStack,
3806 &pam_cl, before_count, CMSYield);
3807 _young_gen->from()->object_iterate_careful(&sss_cl);
3808 _young_gen->to()->object_iterate_careful(&sss_cl);
3809 }
3810 MarkRefsIntoAndScanClosure
3811 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3812 &_markStack, this, CMSYield,
3813 true /* precleaning phase */);
3814 // CAUTION: The following closure has persistent state that may need to
3815 // be reset upon a decrease in the sequence of addresses it
3816 // processes.
3817 ScanMarkedObjectsAgainCarefullyClosure
3818 smoac_cl(this, _span,
3819 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3820
3821 // Preclean dirty cards in ModUnionTable and CardTable using
3822 // appropriate convergence criterion;
3823 // repeat CMSPrecleanIter times unless we find that
3824 // we are losing.
3825 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3826 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3827 "Bad convergence multiplier");
3828 assert(CMSPrecleanThreshold >= 100,
3829 "Unreasonably low CMSPrecleanThreshold");
3830
3831 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3832 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3833 numIter < CMSPrecleanIter;
3834 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3835 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3836 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3837 // Either there are very few dirty cards, so re-mark
3838 // pause will be small anyway, or our pre-cleaning isn't
3839 // that much faster than the rate at which cards are being
3840 // dirtied, so we might as well stop and re-mark since
3841 // precleaning won't improve our re-mark time by much.
3842 if (curNumCards <= CMSPrecleanThreshold ||
3843 (numIter > 0 &&
3844 (curNumCards * CMSPrecleanDenominator >
3845 lastNumCards * CMSPrecleanNumerator))) {
3846 numIter++;
3847 cumNumCards += curNumCards;
3848 break;
3849 }
3850 }
3851
3852 preclean_cld(&mrias_cl, _cmsGen->freelistLock());
3853
3854 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3855 cumNumCards += curNumCards;
3856 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3857 curNumCards, cumNumCards, numIter);
3858 return cumNumCards; // as a measure of useful work done
3859 }
3860
3861 // PRECLEANING NOTES:
3862 // Precleaning involves:
3863 // . reading the bits of the modUnionTable and clearing the set bits.
3864 // . For the cards corresponding to the set bits, we scan the
3865 // objects on those cards. This means we need the free_list_lock
3866 // so that we can safely iterate over the CMS space when scanning
3867 // for oops.
3868 // . When we scan the objects, we'll be both reading and setting
3869 // marks in the marking bit map, so we'll need the marking bit map.
3870 // . For protecting _collector_state transitions, we take the CGC_lock.
3871 // Note that any races in the reading of of card table entries by the
3872 // CMS thread on the one hand and the clearing of those entries by the
3873 // VM thread or the setting of those entries by the mutator threads on the
3874 // other are quite benign. However, for efficiency it makes sense to keep
3875 // the VM thread from racing with the CMS thread while the latter is
3876 // dirty card info to the modUnionTable. We therefore also use the
3877 // CGC_lock to protect the reading of the card table and the mod union
3878 // table by the CM thread.
3879 // . We run concurrently with mutator updates, so scanning
3880 // needs to be done carefully -- we should not try to scan
3881 // potentially uninitialized objects.
3882 //
3883 // Locking strategy: While holding the CGC_lock, we scan over and
3884 // reset a maximal dirty range of the mod union / card tables, then lock
3885 // the free_list_lock and bitmap lock to do a full marking, then
3886 // release these locks; and repeat the cycle. This allows for a
3887 // certain amount of fairness in the sharing of these locks between
3888 // the CMS collector on the one hand, and the VM thread and the
3889 // mutators on the other.
3890
3891 // NOTE: preclean_mod_union_table() and preclean_card_table()
3892 // further below are largely identical; if you need to modify
3893 // one of these methods, please check the other method too.
3894
3895 size_t CMSCollector::preclean_mod_union_table(
3896 ConcurrentMarkSweepGeneration* old_gen,
3897 ScanMarkedObjectsAgainCarefullyClosure* cl) {
3898 verify_work_stacks_empty();
3899 verify_overflow_empty();
3900
3901 // strategy: starting with the first card, accumulate contiguous
3902 // ranges of dirty cards; clear these cards, then scan the region
3903 // covered by these cards.
3904
3905 // Since all of the MUT is committed ahead, we can just use
3906 // that, in case the generations expand while we are precleaning.
3907 // It might also be fine to just use the committed part of the
3908 // generation, but we might potentially miss cards when the
3909 // generation is rapidly expanding while we are in the midst
3910 // of precleaning.
3911 HeapWord* startAddr = old_gen->reserved().start();
3912 HeapWord* endAddr = old_gen->reserved().end();
3913
3914 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
3915
3916 size_t numDirtyCards, cumNumDirtyCards;
3917 HeapWord *nextAddr, *lastAddr;
3918 for (cumNumDirtyCards = numDirtyCards = 0,
3919 nextAddr = lastAddr = startAddr;
3920 nextAddr < endAddr;
3921 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3922
3923 ResourceMark rm;
3924 HandleMark hm;
3925
3926 MemRegion dirtyRegion;
3927 {
3928 stopTimer();
3929 // Potential yield point
3930 CMSTokenSync ts(true);
3931 startTimer();
3932 sample_eden();
3933 // Get dirty region starting at nextOffset (inclusive),
3934 // simultaneously clearing it.
3935 dirtyRegion =
3936 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3937 assert(dirtyRegion.start() >= nextAddr,
3938 "returned region inconsistent?");
3939 }
3940 // Remember where the next search should begin.
3941 // The returned region (if non-empty) is a right open interval,
3942 // so lastOffset is obtained from the right end of that
3943 // interval.
3944 lastAddr = dirtyRegion.end();
3945 // Should do something more transparent and less hacky XXX
3946 numDirtyCards =
3947 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3948
3949 // We'll scan the cards in the dirty region (with periodic
3950 // yields for foreground GC as needed).
3951 if (!dirtyRegion.is_empty()) {
3952 assert(numDirtyCards > 0, "consistency check");
3953 HeapWord* stop_point = NULL;
3954 stopTimer();
3955 // Potential yield point
3956 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3957 bitMapLock());
3958 startTimer();
3959 {
3960 verify_work_stacks_empty();
3961 verify_overflow_empty();
3962 sample_eden();
3963 stop_point =
3964 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3965 }
3966 if (stop_point != NULL) {
3967 // The careful iteration stopped early either because it found an
3968 // uninitialized object, or because we were in the midst of an
3969 // "abortable preclean", which should now be aborted. Redirty
3970 // the bits corresponding to the partially-scanned or unscanned
3971 // cards. We'll either restart at the next block boundary or
3972 // abort the preclean.
3973 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3974 "Should only be AbortablePreclean.");
3975 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3976 if (should_abort_preclean()) {
3977 break; // out of preclean loop
3978 } else {
3979 // Compute the next address at which preclean should pick up;
3980 // might need bitMapLock in order to read P-bits.
3981 lastAddr = next_card_start_after_block(stop_point);
3982 }
3983 }
3984 } else {
3985 assert(lastAddr == endAddr, "consistency check");
3986 assert(numDirtyCards == 0, "consistency check");
3987 break;
3988 }
3989 }
3990 verify_work_stacks_empty();
3991 verify_overflow_empty();
3992 return cumNumDirtyCards;
3993 }
3994
3995 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3996 // below are largely identical; if you need to modify
3997 // one of these methods, please check the other method too.
3998
3999 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4000 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4001 // strategy: it's similar to precleamModUnionTable above, in that
4002 // we accumulate contiguous ranges of dirty cards, mark these cards
4003 // precleaned, then scan the region covered by these cards.
4004 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4005 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4006
4007 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4008
4009 size_t numDirtyCards, cumNumDirtyCards;
4010 HeapWord *lastAddr, *nextAddr;
4011
4012 for (cumNumDirtyCards = numDirtyCards = 0,
4013 nextAddr = lastAddr = startAddr;
4014 nextAddr < endAddr;
4015 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4016
4017 ResourceMark rm;
4018 HandleMark hm;
4019
4020 MemRegion dirtyRegion;
4021 {
4022 // See comments in "Precleaning notes" above on why we
4023 // do this locking. XXX Could the locking overheads be
4024 // too high when dirty cards are sparse? [I don't think so.]
4025 stopTimer();
4026 CMSTokenSync x(true); // is cms thread
4027 startTimer();
4028 sample_eden();
4029 // Get and clear dirty region from card table
4030 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr),
4031 true,
4032 CardTable::precleaned_card_val());
4033
4034 assert(dirtyRegion.start() >= nextAddr,
4035 "returned region inconsistent?");
4036 }
4037 lastAddr = dirtyRegion.end();
4038 numDirtyCards =
4039 dirtyRegion.word_size()/CardTable::card_size_in_words;
4040
4041 if (!dirtyRegion.is_empty()) {
4042 stopTimer();
4043 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4044 startTimer();
4045 sample_eden();
4046 verify_work_stacks_empty();
4047 verify_overflow_empty();
4048 HeapWord* stop_point =
4049 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4050 if (stop_point != NULL) {
4051 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4052 "Should only be AbortablePreclean.");
4053 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4054 if (should_abort_preclean()) {
4055 break; // out of preclean loop
4056 } else {
4057 // Compute the next address at which preclean should pick up.
4058 lastAddr = next_card_start_after_block(stop_point);
4059 }
4060 }
4061 } else {
4062 break;
4063 }
4064 }
4065 verify_work_stacks_empty();
4066 verify_overflow_empty();
4067 return cumNumDirtyCards;
4068 }
4069
4070 class PrecleanCLDClosure : public CLDClosure {
4071 MetadataAwareOopsInGenClosure* _cm_closure;
4072 public:
4073 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {}
4074 void do_cld(ClassLoaderData* cld) {
4075 if (cld->has_accumulated_modified_oops()) {
4076 cld->clear_accumulated_modified_oops();
4077
4078 _cm_closure->do_cld(cld);
4079 }
4080 }
4081 };
4082
4083 // The freelist lock is needed to prevent asserts, is it really needed?
4084 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4085
4086 cl->set_freelistLock(freelistLock);
4087
4088 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4089
4090 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4091 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4092 PrecleanCLDClosure preclean_closure(cl);
4093 ClassLoaderDataGraph::cld_do(&preclean_closure);
4094
4095 verify_work_stacks_empty();
4096 verify_overflow_empty();
4097 }
4098
4099 void CMSCollector::checkpointRootsFinal() {
4100 assert(_collectorState == FinalMarking, "incorrect state transition?");
4101 check_correct_thread_executing();
4102 // world is stopped at this checkpoint
4103 assert(SafepointSynchronize::is_at_safepoint(),
4104 "world should be stopped");
4105 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
4106
4107 verify_work_stacks_empty();
4108 verify_overflow_empty();
4109
4110 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4111 _young_gen->used() / K, _young_gen->capacity() / K);
4112 {
4113 if (CMSScavengeBeforeRemark) {
4114 CMSHeap* heap = CMSHeap::heap();
4115 // Temporarily set flag to false, GCH->do_collection will
4116 // expect it to be false and set to true
4117 FlagSetting fl(heap->_is_gc_active, false);
4118
4119 heap->do_collection(true, // full (i.e. force, see below)
4120 false, // !clear_all_soft_refs
4121 0, // size
4122 false, // is_tlab
4123 GenCollectedHeap::YoungGen // type
4124 );
4125 }
4126 FreelistLocker x(this);
4127 MutexLockerEx y(bitMapLock(),
4128 Mutex::_no_safepoint_check_flag);
4129 checkpointRootsFinalWork();
4130 }
4131 verify_work_stacks_empty();
4132 verify_overflow_empty();
4133 }
4134
4135 void CMSCollector::checkpointRootsFinalWork() {
4136 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4137
4138 assert(haveFreelistLocks(), "must have free list locks");
4139 assert_lock_strong(bitMapLock());
4140
4141 ResourceMark rm;
4142 HandleMark hm;
4143
4144 CMSHeap* heap = CMSHeap::heap();
4145
4146 if (should_unload_classes()) {
4147 CodeCache::gc_prologue();
4148 }
4149 assert(haveFreelistLocks(), "must have free list locks");
4150 assert_lock_strong(bitMapLock());
4151
4152 // We might assume that we need not fill TLAB's when
4153 // CMSScavengeBeforeRemark is set, because we may have just done
4154 // a scavenge which would have filled all TLAB's -- and besides
4155 // Eden would be empty. This however may not always be the case --
4156 // for instance although we asked for a scavenge, it may not have
4157 // happened because of a JNI critical section. We probably need
4158 // a policy for deciding whether we can in that case wait until
4159 // the critical section releases and then do the remark following
4160 // the scavenge, and skip it here. In the absence of that policy,
4161 // or of an indication of whether the scavenge did indeed occur,
4162 // we cannot rely on TLAB's having been filled and must do
4163 // so here just in case a scavenge did not happen.
4164 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them
4165 // Update the saved marks which may affect the root scans.
4166 heap->save_marks();
4167
4168 print_eden_and_survivor_chunk_arrays();
4169
4170 {
4171 #if COMPILER2_OR_JVMCI
4172 DerivedPointerTableDeactivate dpt_deact;
4173 #endif
4174
4175 // Note on the role of the mod union table:
4176 // Since the marker in "markFromRoots" marks concurrently with
4177 // mutators, it is possible for some reachable objects not to have been
4178 // scanned. For instance, an only reference to an object A was
4179 // placed in object B after the marker scanned B. Unless B is rescanned,
4180 // A would be collected. Such updates to references in marked objects
4181 // are detected via the mod union table which is the set of all cards
4182 // dirtied since the first checkpoint in this GC cycle and prior to
4183 // the most recent young generation GC, minus those cleaned up by the
4184 // concurrent precleaning.
4185 if (CMSParallelRemarkEnabled) {
4186 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4187 do_remark_parallel();
4188 } else {
4189 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4190 do_remark_non_parallel();
4191 }
4192 }
4193 verify_work_stacks_empty();
4194 verify_overflow_empty();
4195
4196 {
4197 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4198 refProcessingWork();
4199 }
4200 verify_work_stacks_empty();
4201 verify_overflow_empty();
4202
4203 if (should_unload_classes()) {
4204 CodeCache::gc_epilogue();
4205 }
4206 JvmtiExport::gc_epilogue();
4207
4208 // If we encountered any (marking stack / work queue) overflow
4209 // events during the current CMS cycle, take appropriate
4210 // remedial measures, where possible, so as to try and avoid
4211 // recurrence of that condition.
4212 assert(_markStack.isEmpty(), "No grey objects");
4213 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4214 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4215 if (ser_ovflw > 0) {
4216 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4217 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4218 _markStack.expand();
4219 _ser_pmc_remark_ovflw = 0;
4220 _ser_pmc_preclean_ovflw = 0;
4221 _ser_kac_preclean_ovflw = 0;
4222 _ser_kac_ovflw = 0;
4223 }
4224 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4225 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4226 _par_pmc_remark_ovflw, _par_kac_ovflw);
4227 _par_pmc_remark_ovflw = 0;
4228 _par_kac_ovflw = 0;
4229 }
4230 if (_markStack._hit_limit > 0) {
4231 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4232 _markStack._hit_limit);
4233 }
4234 if (_markStack._failed_double > 0) {
4235 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4236 _markStack._failed_double, _markStack.capacity());
4237 }
4238 _markStack._hit_limit = 0;
4239 _markStack._failed_double = 0;
4240
4241 if ((VerifyAfterGC || VerifyDuringGC) &&
4242 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4243 verify_after_remark();
4244 }
4245
4246 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4247
4248 // Change under the freelistLocks.
4249 _collectorState = Sweeping;
4250 // Call isAllClear() under bitMapLock
4251 assert(_modUnionTable.isAllClear(),
4252 "Should be clear by end of the final marking");
4253 assert(_ct->cld_rem_set()->mod_union_is_clear(),
4254 "Should be clear by end of the final marking");
4255 }
4256
4257 void CMSParInitialMarkTask::work(uint worker_id) {
4258 elapsedTimer _timer;
4259 ResourceMark rm;
4260 HandleMark hm;
4261
4262 // ---------- scan from roots --------------
4263 _timer.start();
4264 CMSHeap* heap = CMSHeap::heap();
4265 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4266
4267 // ---------- young gen roots --------------
4268 {
4269 work_on_young_gen_roots(&par_mri_cl);
4270 _timer.stop();
4271 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4272 }
4273
4274 // ---------- remaining roots --------------
4275 _timer.reset();
4276 _timer.start();
4277
4278 CLDToOopClosure cld_closure(&par_mri_cl, true);
4279
4280 heap->cms_process_roots(_strong_roots_scope,
4281 false, // yg was scanned above
4282 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4283 _collector->should_unload_classes(),
4284 &par_mri_cl,
4285 &cld_closure,
4286 &_par_state_string);
4287
4288 assert(_collector->should_unload_classes()
4289 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4290 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4291 _timer.stop();
4292 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4293 }
4294
4295 // Parallel remark task
4296 class CMSParRemarkTask: public CMSParMarkTask {
4297 CompactibleFreeListSpace* _cms_space;
4298
4299 // The per-thread work queues, available here for stealing.
4300 OopTaskQueueSet* _task_queues;
4301 ParallelTaskTerminator _term;
4302 StrongRootsScope* _strong_roots_scope;
4303
4304 public:
4305 // A value of 0 passed to n_workers will cause the number of
4306 // workers to be taken from the active workers in the work gang.
4307 CMSParRemarkTask(CMSCollector* collector,
4308 CompactibleFreeListSpace* cms_space,
4309 uint n_workers, WorkGang* workers,
4310 OopTaskQueueSet* task_queues,
4311 StrongRootsScope* strong_roots_scope):
4312 CMSParMarkTask("Rescan roots and grey objects in parallel",
4313 collector, n_workers),
4314 _cms_space(cms_space),
4315 _task_queues(task_queues),
4316 _term(n_workers, task_queues),
4317 _strong_roots_scope(strong_roots_scope) { }
4318
4319 OopTaskQueueSet* task_queues() { return _task_queues; }
4320
4321 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4322
4323 ParallelTaskTerminator* terminator() { return &_term; }
4324 uint n_workers() { return _n_workers; }
4325
4326 void work(uint worker_id);
4327
4328 private:
4329 // ... of dirty cards in old space
4330 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4331 ParMarkRefsIntoAndScanClosure* cl);
4332
4333 // ... work stealing for the above
4334 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4335 };
4336
4337 class RemarkCLDClosure : public CLDClosure {
4338 CLDToOopClosure _cm_closure;
4339 public:
4340 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {}
4341 void do_cld(ClassLoaderData* cld) {
4342 // Check if we have modified any oops in the CLD during the concurrent marking.
4343 if (cld->has_accumulated_modified_oops()) {
4344 cld->clear_accumulated_modified_oops();
4345
4346 // We could have transfered the current modified marks to the accumulated marks,
4347 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4348 } else if (cld->has_modified_oops()) {
4349 // Don't clear anything, this info is needed by the next young collection.
4350 } else {
4351 // No modified oops in the ClassLoaderData.
4352 return;
4353 }
4354
4355 // The klass has modified fields, need to scan the klass.
4356 _cm_closure.do_cld(cld);
4357 }
4358 };
4359
4360 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4361 ParNewGeneration* young_gen = _collector->_young_gen;
4362 ContiguousSpace* eden_space = young_gen->eden();
4363 ContiguousSpace* from_space = young_gen->from();
4364 ContiguousSpace* to_space = young_gen->to();
4365
4366 HeapWord** eca = _collector->_eden_chunk_array;
4367 size_t ect = _collector->_eden_chunk_index;
4368 HeapWord** sca = _collector->_survivor_chunk_array;
4369 size_t sct = _collector->_survivor_chunk_index;
4370
4371 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4372 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4373
4374 do_young_space_rescan(cl, to_space, NULL, 0);
4375 do_young_space_rescan(cl, from_space, sca, sct);
4376 do_young_space_rescan(cl, eden_space, eca, ect);
4377 }
4378
4379 // work_queue(i) is passed to the closure
4380 // ParMarkRefsIntoAndScanClosure. The "i" parameter
4381 // also is passed to do_dirty_card_rescan_tasks() and to
4382 // do_work_steal() to select the i-th task_queue.
4383
4384 void CMSParRemarkTask::work(uint worker_id) {
4385 elapsedTimer _timer;
4386 ResourceMark rm;
4387 HandleMark hm;
4388
4389 // ---------- rescan from roots --------------
4390 _timer.start();
4391 CMSHeap* heap = CMSHeap::heap();
4392 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4393 _collector->_span, _collector->ref_processor(),
4394 &(_collector->_markBitMap),
4395 work_queue(worker_id));
4396
4397 // Rescan young gen roots first since these are likely
4398 // coarsely partitioned and may, on that account, constitute
4399 // the critical path; thus, it's best to start off that
4400 // work first.
4401 // ---------- young gen roots --------------
4402 {
4403 work_on_young_gen_roots(&par_mrias_cl);
4404 _timer.stop();
4405 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4406 }
4407
4408 // ---------- remaining roots --------------
4409 _timer.reset();
4410 _timer.start();
4411 heap->cms_process_roots(_strong_roots_scope,
4412 false, // yg was scanned above
4413 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4414 _collector->should_unload_classes(),
4415 &par_mrias_cl,
4416 NULL, // The dirty klasses will be handled below
4417 &_par_state_string);
4418
4419 assert(_collector->should_unload_classes()
4420 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4421 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4422 _timer.stop();
4423 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4424
4425 // ---------- unhandled CLD scanning ----------
4426 if (worker_id == 0) { // Single threaded at the moment.
4427 _timer.reset();
4428 _timer.start();
4429
4430 // Scan all new class loader data objects and new dependencies that were
4431 // introduced during concurrent marking.
4432 ResourceMark rm;
4433 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4434 for (int i = 0; i < array->length(); i++) {
4435 par_mrias_cl.do_cld_nv(array->at(i));
4436 }
4437
4438 // We don't need to keep track of new CLDs anymore.
4439 ClassLoaderDataGraph::remember_new_clds(false);
4440
4441 _timer.stop();
4442 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4443 }
4444
4445 // We might have added oops to ClassLoaderData::_handles during the
4446 // concurrent marking phase. These oops do not always point to newly allocated objects
4447 // that are guaranteed to be kept alive. Hence,
4448 // we do have to revisit the _handles block during the remark phase.
4449
4450 // ---------- dirty CLD scanning ----------
4451 if (worker_id == 0) { // Single threaded at the moment.
4452 _timer.reset();
4453 _timer.start();
4454
4455 // Scan all classes that was dirtied during the concurrent marking phase.
4456 RemarkCLDClosure remark_closure(&par_mrias_cl);
4457 ClassLoaderDataGraph::cld_do(&remark_closure);
4458
4459 _timer.stop();
4460 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4461 }
4462
4463
4464 // ---------- rescan dirty cards ------------
4465 _timer.reset();
4466 _timer.start();
4467
4468 // Do the rescan tasks for each of the two spaces
4469 // (cms_space) in turn.
4470 // "worker_id" is passed to select the task_queue for "worker_id"
4471 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4472 _timer.stop();
4473 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4474
4475 // ---------- steal work from other threads ...
4476 // ---------- ... and drain overflow list.
4477 _timer.reset();
4478 _timer.start();
4479 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4480 _timer.stop();
4481 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4482 }
4483
4484 void
4485 CMSParMarkTask::do_young_space_rescan(
4486 OopsInGenClosure* cl, ContiguousSpace* space,
4487 HeapWord** chunk_array, size_t chunk_top) {
4488 // Until all tasks completed:
4489 // . claim an unclaimed task
4490 // . compute region boundaries corresponding to task claimed
4491 // using chunk_array
4492 // . par_oop_iterate(cl) over that region
4493
4494 ResourceMark rm;
4495 HandleMark hm;
4496
4497 SequentialSubTasksDone* pst = space->par_seq_tasks();
4498
4499 uint nth_task = 0;
4500 uint n_tasks = pst->n_tasks();
4501
4502 if (n_tasks > 0) {
4503 assert(pst->valid(), "Uninitialized use?");
4504 HeapWord *start, *end;
4505 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4506 // We claimed task # nth_task; compute its boundaries.
4507 if (chunk_top == 0) { // no samples were taken
4508 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4509 start = space->bottom();
4510 end = space->top();
4511 } else if (nth_task == 0) {
4512 start = space->bottom();
4513 end = chunk_array[nth_task];
4514 } else if (nth_task < (uint)chunk_top) {
4515 assert(nth_task >= 1, "Control point invariant");
4516 start = chunk_array[nth_task - 1];
4517 end = chunk_array[nth_task];
4518 } else {
4519 assert(nth_task == (uint)chunk_top, "Control point invariant");
4520 start = chunk_array[chunk_top - 1];
4521 end = space->top();
4522 }
4523 MemRegion mr(start, end);
4524 // Verify that mr is in space
4525 assert(mr.is_empty() || space->used_region().contains(mr),
4526 "Should be in space");
4527 // Verify that "start" is an object boundary
4528 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4529 "Should be an oop");
4530 space->par_oop_iterate(mr, cl);
4531 }
4532 pst->all_tasks_completed();
4533 }
4534 }
4535
4536 void
4537 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4538 CompactibleFreeListSpace* sp, int i,
4539 ParMarkRefsIntoAndScanClosure* cl) {
4540 // Until all tasks completed:
4541 // . claim an unclaimed task
4542 // . compute region boundaries corresponding to task claimed
4543 // . transfer dirty bits ct->mut for that region
4544 // . apply rescanclosure to dirty mut bits for that region
4545
4546 ResourceMark rm;
4547 HandleMark hm;
4548
4549 OopTaskQueue* work_q = work_queue(i);
4550 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4551 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4552 // CAUTION: This closure has state that persists across calls to
4553 // the work method dirty_range_iterate_clear() in that it has
4554 // embedded in it a (subtype of) UpwardsObjectClosure. The
4555 // use of that state in the embedded UpwardsObjectClosure instance
4556 // assumes that the cards are always iterated (even if in parallel
4557 // by several threads) in monotonically increasing order per each
4558 // thread. This is true of the implementation below which picks
4559 // card ranges (chunks) in monotonically increasing order globally
4560 // and, a-fortiori, in monotonically increasing order per thread
4561 // (the latter order being a subsequence of the former).
4562 // If the work code below is ever reorganized into a more chaotic
4563 // work-partitioning form than the current "sequential tasks"
4564 // paradigm, the use of that persistent state will have to be
4565 // revisited and modified appropriately. See also related
4566 // bug 4756801 work on which should examine this code to make
4567 // sure that the changes there do not run counter to the
4568 // assumptions made here and necessary for correctness and
4569 // efficiency. Note also that this code might yield inefficient
4570 // behavior in the case of very large objects that span one or
4571 // more work chunks. Such objects would potentially be scanned
4572 // several times redundantly. Work on 4756801 should try and
4573 // address that performance anomaly if at all possible. XXX
4574 MemRegion full_span = _collector->_span;
4575 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4576 MarkFromDirtyCardsClosure
4577 greyRescanClosure(_collector, full_span, // entire span of interest
4578 sp, bm, work_q, cl);
4579
4580 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4581 assert(pst->valid(), "Uninitialized use?");
4582 uint nth_task = 0;
4583 const int alignment = CardTable::card_size * BitsPerWord;
4584 MemRegion span = sp->used_region();
4585 HeapWord* start_addr = span.start();
4586 HeapWord* end_addr = align_up(span.end(), alignment);
4587 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4588 assert(is_aligned(start_addr, alignment), "Check alignment");
4589 assert(is_aligned(chunk_size, alignment), "Check alignment");
4590
4591 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4592 // Having claimed the nth_task, compute corresponding mem-region,
4593 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4594 // The alignment restriction ensures that we do not need any
4595 // synchronization with other gang-workers while setting or
4596 // clearing bits in thus chunk of the MUT.
4597 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4598 start_addr + (nth_task+1)*chunk_size);
4599 // The last chunk's end might be way beyond end of the
4600 // used region. In that case pull back appropriately.
4601 if (this_span.end() > end_addr) {
4602 this_span.set_end(end_addr);
4603 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4604 }
4605 // Iterate over the dirty cards covering this chunk, marking them
4606 // precleaned, and setting the corresponding bits in the mod union
4607 // table. Since we have been careful to partition at Card and MUT-word
4608 // boundaries no synchronization is needed between parallel threads.
4609 _collector->_ct->dirty_card_iterate(this_span,
4610 &modUnionClosure);
4611
4612 // Having transferred these marks into the modUnionTable,
4613 // rescan the marked objects on the dirty cards in the modUnionTable.
4614 // Even if this is at a synchronous collection, the initial marking
4615 // may have been done during an asynchronous collection so there
4616 // may be dirty bits in the mod-union table.
4617 _collector->_modUnionTable.dirty_range_iterate_clear(
4618 this_span, &greyRescanClosure);
4619 _collector->_modUnionTable.verifyNoOneBitsInRange(
4620 this_span.start(),
4621 this_span.end());
4622 }
4623 pst->all_tasks_completed(); // declare that i am done
4624 }
4625
4626 // . see if we can share work_queues with ParNew? XXX
4627 void
4628 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4629 int* seed) {
4630 OopTaskQueue* work_q = work_queue(i);
4631 NOT_PRODUCT(int num_steals = 0;)
4632 oop obj_to_scan;
4633 CMSBitMap* bm = &(_collector->_markBitMap);
4634
4635 while (true) {
4636 // Completely finish any left over work from (an) earlier round(s)
4637 cl->trim_queue(0);
4638 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4639 (size_t)ParGCDesiredObjsFromOverflowList);
4640 // Now check if there's any work in the overflow list
4641 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4642 // only affects the number of attempts made to get work from the
4643 // overflow list and does not affect the number of workers. Just
4644 // pass ParallelGCThreads so this behavior is unchanged.
4645 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4646 work_q,
4647 ParallelGCThreads)) {
4648 // found something in global overflow list;
4649 // not yet ready to go stealing work from others.
4650 // We'd like to assert(work_q->size() != 0, ...)
4651 // because we just took work from the overflow list,
4652 // but of course we can't since all of that could have
4653 // been already stolen from us.
4654 // "He giveth and He taketh away."
4655 continue;
4656 }
4657 // Verify that we have no work before we resort to stealing
4658 assert(work_q->size() == 0, "Have work, shouldn't steal");
4659 // Try to steal from other queues that have work
4660 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4661 NOT_PRODUCT(num_steals++;)
4662 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4663 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4664 // Do scanning work
4665 obj_to_scan->oop_iterate(cl);
4666 // Loop around, finish this work, and try to steal some more
4667 } else if (terminator()->offer_termination()) {
4668 break; // nirvana from the infinite cycle
4669 }
4670 }
4671 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4672 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4673 "Else our work is not yet done");
4674 }
4675
4676 // Record object boundaries in _eden_chunk_array by sampling the eden
4677 // top in the slow-path eden object allocation code path and record
4678 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4679 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4680 // sampling in sample_eden() that activates during the part of the
4681 // preclean phase.
4682 void CMSCollector::sample_eden_chunk() {
4683 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4684 if (_eden_chunk_lock->try_lock()) {
4685 // Record a sample. This is the critical section. The contents
4686 // of the _eden_chunk_array have to be non-decreasing in the
4687 // address order.
4688 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4689 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4690 "Unexpected state of Eden");
4691 if (_eden_chunk_index == 0 ||
4692 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4693 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4694 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4695 _eden_chunk_index++; // commit sample
4696 }
4697 _eden_chunk_lock->unlock();
4698 }
4699 }
4700 }
4701
4702 // Return a thread-local PLAB recording array, as appropriate.
4703 void* CMSCollector::get_data_recorder(int thr_num) {
4704 if (_survivor_plab_array != NULL &&
4705 (CMSPLABRecordAlways ||
4706 (_collectorState > Marking && _collectorState < FinalMarking))) {
4707 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4708 ChunkArray* ca = &_survivor_plab_array[thr_num];
4709 ca->reset(); // clear it so that fresh data is recorded
4710 return (void*) ca;
4711 } else {
4712 return NULL;
4713 }
4714 }
4715
4716 // Reset all the thread-local PLAB recording arrays
4717 void CMSCollector::reset_survivor_plab_arrays() {
4718 for (uint i = 0; i < ParallelGCThreads; i++) {
4719 _survivor_plab_array[i].reset();
4720 }
4721 }
4722
4723 // Merge the per-thread plab arrays into the global survivor chunk
4724 // array which will provide the partitioning of the survivor space
4725 // for CMS initial scan and rescan.
4726 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4727 int no_of_gc_threads) {
4728 assert(_survivor_plab_array != NULL, "Error");
4729 assert(_survivor_chunk_array != NULL, "Error");
4730 assert(_collectorState == FinalMarking ||
4731 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4732 for (int j = 0; j < no_of_gc_threads; j++) {
4733 _cursor[j] = 0;
4734 }
4735 HeapWord* top = surv->top();
4736 size_t i;
4737 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4738 HeapWord* min_val = top; // Higher than any PLAB address
4739 uint min_tid = 0; // position of min_val this round
4740 for (int j = 0; j < no_of_gc_threads; j++) {
4741 ChunkArray* cur_sca = &_survivor_plab_array[j];
4742 if (_cursor[j] == cur_sca->end()) {
4743 continue;
4744 }
4745 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4746 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4747 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4748 if (cur_val < min_val) {
4749 min_tid = j;
4750 min_val = cur_val;
4751 } else {
4752 assert(cur_val < top, "All recorded addresses should be less");
4753 }
4754 }
4755 // At this point min_val and min_tid are respectively
4756 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4757 // and the thread (j) that witnesses that address.
4758 // We record this address in the _survivor_chunk_array[i]
4759 // and increment _cursor[min_tid] prior to the next round i.
4760 if (min_val == top) {
4761 break;
4762 }
4763 _survivor_chunk_array[i] = min_val;
4764 _cursor[min_tid]++;
4765 }
4766 // We are all done; record the size of the _survivor_chunk_array
4767 _survivor_chunk_index = i; // exclusive: [0, i)
4768 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4769 // Verify that we used up all the recorded entries
4770 #ifdef ASSERT
4771 size_t total = 0;
4772 for (int j = 0; j < no_of_gc_threads; j++) {
4773 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4774 total += _cursor[j];
4775 }
4776 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4777 // Check that the merged array is in sorted order
4778 if (total > 0) {
4779 for (size_t i = 0; i < total - 1; i++) {
4780 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4781 i, p2i(_survivor_chunk_array[i]));
4782 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4783 "Not sorted");
4784 }
4785 }
4786 #endif // ASSERT
4787 }
4788
4789 // Set up the space's par_seq_tasks structure for work claiming
4790 // for parallel initial scan and rescan of young gen.
4791 // See ParRescanTask where this is currently used.
4792 void
4793 CMSCollector::
4794 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4795 assert(n_threads > 0, "Unexpected n_threads argument");
4796
4797 // Eden space
4798 if (!_young_gen->eden()->is_empty()) {
4799 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4800 assert(!pst->valid(), "Clobbering existing data?");
4801 // Each valid entry in [0, _eden_chunk_index) represents a task.
4802 size_t n_tasks = _eden_chunk_index + 1;
4803 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4804 // Sets the condition for completion of the subtask (how many threads
4805 // need to finish in order to be done).
4806 pst->set_n_threads(n_threads);
4807 pst->set_n_tasks((int)n_tasks);
4808 }
4809
4810 // Merge the survivor plab arrays into _survivor_chunk_array
4811 if (_survivor_plab_array != NULL) {
4812 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4813 } else {
4814 assert(_survivor_chunk_index == 0, "Error");
4815 }
4816
4817 // To space
4818 {
4819 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4820 assert(!pst->valid(), "Clobbering existing data?");
4821 // Sets the condition for completion of the subtask (how many threads
4822 // need to finish in order to be done).
4823 pst->set_n_threads(n_threads);
4824 pst->set_n_tasks(1);
4825 assert(pst->valid(), "Error");
4826 }
4827
4828 // From space
4829 {
4830 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4831 assert(!pst->valid(), "Clobbering existing data?");
4832 size_t n_tasks = _survivor_chunk_index + 1;
4833 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4834 // Sets the condition for completion of the subtask (how many threads
4835 // need to finish in order to be done).
4836 pst->set_n_threads(n_threads);
4837 pst->set_n_tasks((int)n_tasks);
4838 assert(pst->valid(), "Error");
4839 }
4840 }
4841
4842 // Parallel version of remark
4843 void CMSCollector::do_remark_parallel() {
4844 CMSHeap* heap = CMSHeap::heap();
4845 WorkGang* workers = heap->workers();
4846 assert(workers != NULL, "Need parallel worker threads.");
4847 // Choose to use the number of GC workers most recently set
4848 // into "active_workers".
4849 uint n_workers = workers->active_workers();
4850
4851 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4852
4853 StrongRootsScope srs(n_workers);
4854
4855 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4856
4857 // We won't be iterating over the cards in the card table updating
4858 // the younger_gen cards, so we shouldn't call the following else
4859 // the verification code as well as subsequent younger_refs_iterate
4860 // code would get confused. XXX
4861 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4862
4863 // The young gen rescan work will not be done as part of
4864 // process_roots (which currently doesn't know how to
4865 // parallelize such a scan), but rather will be broken up into
4866 // a set of parallel tasks (via the sampling that the [abortable]
4867 // preclean phase did of eden, plus the [two] tasks of
4868 // scanning the [two] survivor spaces. Further fine-grain
4869 // parallelization of the scanning of the survivor spaces
4870 // themselves, and of precleaning of the young gen itself
4871 // is deferred to the future.
4872 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4873
4874 // The dirty card rescan work is broken up into a "sequence"
4875 // of parallel tasks (per constituent space) that are dynamically
4876 // claimed by the parallel threads.
4877 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4878
4879 // It turns out that even when we're using 1 thread, doing the work in a
4880 // separate thread causes wide variance in run times. We can't help this
4881 // in the multi-threaded case, but we special-case n=1 here to get
4882 // repeatable measurements of the 1-thread overhead of the parallel code.
4883 if (n_workers > 1) {
4884 // Make refs discovery MT-safe, if it isn't already: it may not
4885 // necessarily be so, since it's possible that we are doing
4886 // ST marking.
4887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4888 workers->run_task(&tsk);
4889 } else {
4890 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4891 tsk.work(0);
4892 }
4893
4894 // restore, single-threaded for now, any preserved marks
4895 // as a result of work_q overflow
4896 restore_preserved_marks_if_any();
4897 }
4898
4899 // Non-parallel version of remark
4900 void CMSCollector::do_remark_non_parallel() {
4901 ResourceMark rm;
4902 HandleMark hm;
4903 CMSHeap* heap = CMSHeap::heap();
4904 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4905
4906 MarkRefsIntoAndScanClosure
4907 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4908 &_markStack, this,
4909 false /* should_yield */, false /* not precleaning */);
4910 MarkFromDirtyCardsClosure
4911 markFromDirtyCardsClosure(this, _span,
4912 NULL, // space is set further below
4913 &_markBitMap, &_markStack, &mrias_cl);
4914 {
4915 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4916 // Iterate over the dirty cards, setting the corresponding bits in the
4917 // mod union table.
4918 {
4919 ModUnionClosure modUnionClosure(&_modUnionTable);
4920 _ct->dirty_card_iterate(_cmsGen->used_region(),
4921 &modUnionClosure);
4922 }
4923 // Having transferred these marks into the modUnionTable, we just need
4924 // to rescan the marked objects on the dirty cards in the modUnionTable.
4925 // The initial marking may have been done during an asynchronous
4926 // collection so there may be dirty bits in the mod-union table.
4927 const int alignment = CardTable::card_size * BitsPerWord;
4928 {
4929 // ... First handle dirty cards in CMS gen
4930 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4931 MemRegion ur = _cmsGen->used_region();
4932 HeapWord* lb = ur.start();
4933 HeapWord* ub = align_up(ur.end(), alignment);
4934 MemRegion cms_span(lb, ub);
4935 _modUnionTable.dirty_range_iterate_clear(cms_span,
4936 &markFromDirtyCardsClosure);
4937 verify_work_stacks_empty();
4938 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4939 }
4940 }
4941 if (VerifyDuringGC &&
4942 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4943 HandleMark hm; // Discard invalid handles created during verification
4944 Universe::verify();
4945 }
4946 {
4947 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4948
4949 verify_work_stacks_empty();
4950
4951 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4952 StrongRootsScope srs(1);
4953
4954 heap->cms_process_roots(&srs,
4955 true, // young gen as roots
4956 GenCollectedHeap::ScanningOption(roots_scanning_options()),
4957 should_unload_classes(),
4958 &mrias_cl,
4959 NULL); // The dirty klasses will be handled below
4960
4961 assert(should_unload_classes()
4962 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4963 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4964 }
4965
4966 {
4967 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4968
4969 verify_work_stacks_empty();
4970
4971 // Scan all class loader data objects that might have been introduced
4972 // during concurrent marking.
4973 ResourceMark rm;
4974 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4975 for (int i = 0; i < array->length(); i++) {
4976 mrias_cl.do_cld_nv(array->at(i));
4977 }
4978
4979 // We don't need to keep track of new CLDs anymore.
4980 ClassLoaderDataGraph::remember_new_clds(false);
4981
4982 verify_work_stacks_empty();
4983 }
4984
4985 // We might have added oops to ClassLoaderData::_handles during the
4986 // concurrent marking phase. These oops do not point to newly allocated objects
4987 // that are guaranteed to be kept alive. Hence,
4988 // we do have to revisit the _handles block during the remark phase.
4989 {
4990 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm);
4991
4992 verify_work_stacks_empty();
4993
4994 RemarkCLDClosure remark_closure(&mrias_cl);
4995 ClassLoaderDataGraph::cld_do(&remark_closure);
4996
4997 verify_work_stacks_empty();
4998 }
4999
5000 verify_work_stacks_empty();
5001 // Restore evacuated mark words, if any, used for overflow list links
5002 restore_preserved_marks_if_any();
5003
5004 verify_overflow_empty();
5005 }
5006
5007 ////////////////////////////////////////////////////////
5008 // Parallel Reference Processing Task Proxy Class
5009 ////////////////////////////////////////////////////////
5010 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5011 OopTaskQueueSet* _queues;
5012 ParallelTaskTerminator _terminator;
5013 public:
5014 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5015 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5016 ParallelTaskTerminator* terminator() { return &_terminator; }
5017 OopTaskQueueSet* queues() { return _queues; }
5018 };
5019
5020 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5021 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5022 CMSCollector* _collector;
5023 CMSBitMap* _mark_bit_map;
5024 const MemRegion _span;
5025 ProcessTask& _task;
5026
5027 public:
5028 CMSRefProcTaskProxy(ProcessTask& task,
5029 CMSCollector* collector,
5030 const MemRegion& span,
5031 CMSBitMap* mark_bit_map,
5032 AbstractWorkGang* workers,
5033 OopTaskQueueSet* task_queues):
5034 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5035 task_queues,
5036 workers->active_workers()),
5037 _task(task),
5038 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5039 {
5040 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5041 "Inconsistency in _span");
5042 }
5043
5044 OopTaskQueueSet* task_queues() { return queues(); }
5045
5046 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5047
5048 void do_work_steal(int i,
5049 CMSParDrainMarkingStackClosure* drain,
5050 CMSParKeepAliveClosure* keep_alive,
5051 int* seed);
5052
5053 virtual void work(uint worker_id);
5054 };
5055
5056 void CMSRefProcTaskProxy::work(uint worker_id) {
5057 ResourceMark rm;
5058 HandleMark hm;
5059 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5060 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5061 _mark_bit_map,
5062 work_queue(worker_id));
5063 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5064 _mark_bit_map,
5065 work_queue(worker_id));
5066 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5067 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5068 if (_task.marks_oops_alive()) {
5069 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5070 _collector->hash_seed(worker_id));
5071 }
5072 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5073 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5074 }
5075
5076 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5077 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5078 _span(span),
5079 _bit_map(bit_map),
5080 _work_queue(work_queue),
5081 _mark_and_push(collector, span, bit_map, work_queue),
5082 _low_water_mark(MIN2((work_queue->max_elems()/4),
5083 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5084 { }
5085
5086 // . see if we can share work_queues with ParNew? XXX
5087 void CMSRefProcTaskProxy::do_work_steal(int i,
5088 CMSParDrainMarkingStackClosure* drain,
5089 CMSParKeepAliveClosure* keep_alive,
5090 int* seed) {
5091 OopTaskQueue* work_q = work_queue(i);
5092 NOT_PRODUCT(int num_steals = 0;)
5093 oop obj_to_scan;
5094
5095 while (true) {
5096 // Completely finish any left over work from (an) earlier round(s)
5097 drain->trim_queue(0);
5098 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5099 (size_t)ParGCDesiredObjsFromOverflowList);
5100 // Now check if there's any work in the overflow list
5101 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5102 // only affects the number of attempts made to get work from the
5103 // overflow list and does not affect the number of workers. Just
5104 // pass ParallelGCThreads so this behavior is unchanged.
5105 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5106 work_q,
5107 ParallelGCThreads)) {
5108 // Found something in global overflow list;
5109 // not yet ready to go stealing work from others.
5110 // We'd like to assert(work_q->size() != 0, ...)
5111 // because we just took work from the overflow list,
5112 // but of course we can't, since all of that might have
5113 // been already stolen from us.
5114 continue;
5115 }
5116 // Verify that we have no work before we resort to stealing
5117 assert(work_q->size() == 0, "Have work, shouldn't steal");
5118 // Try to steal from other queues that have work
5119 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5120 NOT_PRODUCT(num_steals++;)
5121 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5122 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5123 // Do scanning work
5124 obj_to_scan->oop_iterate(keep_alive);
5125 // Loop around, finish this work, and try to steal some more
5126 } else if (terminator()->offer_termination()) {
5127 break; // nirvana from the infinite cycle
5128 }
5129 }
5130 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5131 }
5132
5133 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5134 {
5135 CMSHeap* heap = CMSHeap::heap();
5136 WorkGang* workers = heap->workers();
5137 assert(workers != NULL, "Need parallel worker threads.");
5138 CMSRefProcTaskProxy rp_task(task, &_collector,
5139 _collector.ref_processor_span(),
5140 _collector.markBitMap(),
5141 workers, _collector.task_queues());
5142 workers->run_task(&rp_task);
5143 }
5144
5145 void CMSCollector::refProcessingWork() {
5146 ResourceMark rm;
5147 HandleMark hm;
5148
5149 ReferenceProcessor* rp = ref_processor();
5150 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal");
5151 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5152 // Process weak references.
5153 rp->setup_policy(false);
5154 verify_work_stacks_empty();
5155
5156 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues());
5157 {
5158 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5159
5160 // Setup keep_alive and complete closures.
5161 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5162 &_markStack, false /* !preclean */);
5163 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5164 _span, &_markBitMap, &_markStack,
5165 &cmsKeepAliveClosure, false /* !preclean */);
5166
5167 ReferenceProcessorStats stats;
5168 if (rp->processing_is_mt()) {
5169 // Set the degree of MT here. If the discovery is done MT, there
5170 // may have been a different number of threads doing the discovery
5171 // and a different number of discovered lists may have Ref objects.
5172 // That is OK as long as the Reference lists are balanced (see
5173 // balance_all_queues() and balance_queues()).
5174 CMSHeap* heap = CMSHeap::heap();
5175 uint active_workers = ParallelGCThreads;
5176 WorkGang* workers = heap->workers();
5177 if (workers != NULL) {
5178 active_workers = workers->active_workers();
5179 // The expectation is that active_workers will have already
5180 // been set to a reasonable value. If it has not been set,
5181 // investigate.
5182 assert(active_workers > 0, "Should have been set during scavenge");
5183 }
5184 rp->set_active_mt_degree(active_workers);
5185 CMSRefProcTaskExecutor task_executor(*this);
5186 stats = rp->process_discovered_references(&_is_alive_closure,
5187 &cmsKeepAliveClosure,
5188 &cmsDrainMarkingStackClosure,
5189 &task_executor,
5190 &pt);
5191 } else {
5192 stats = rp->process_discovered_references(&_is_alive_closure,
5193 &cmsKeepAliveClosure,
5194 &cmsDrainMarkingStackClosure,
5195 NULL,
5196 &pt);
5197 }
5198 _gc_tracer_cm->report_gc_reference_stats(stats);
5199 pt.print_all_references();
5200 }
5201
5202 // This is the point where the entire marking should have completed.
5203 verify_work_stacks_empty();
5204
5205 {
5206 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm);
5207 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl);
5208 }
5209
5210 if (should_unload_classes()) {
5211 {
5212 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5213
5214 // Unload classes and purge the SystemDictionary.
5215 bool purged_class = SystemDictionary::do_unloading(_gc_timer_cm);
5216
5217 // Unload nmethods.
5218 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5219
5220 // Prune dead klasses from subklass/sibling/implementor lists.
5221 Klass::clean_weak_klass_links(purged_class);
5222 }
5223
5224 {
5225 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5226 // Clean up unreferenced symbols in symbol table.
5227 SymbolTable::unlink();
5228 }
5229
5230 {
5231 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5232 // Delete entries for dead interned strings.
5233 StringTable::unlink(&_is_alive_closure);
5234 }
5235 }
5236
5237 // Restore any preserved marks as a result of mark stack or
5238 // work queue overflow
5239 restore_preserved_marks_if_any(); // done single-threaded for now
5240
5241 rp->set_enqueuing_is_done(true);
5242 rp->verify_no_references_recorded();
5243 }
5244
5245 #ifndef PRODUCT
5246 void CMSCollector::check_correct_thread_executing() {
5247 Thread* t = Thread::current();
5248 // Only the VM thread or the CMS thread should be here.
5249 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5250 "Unexpected thread type");
5251 // If this is the vm thread, the foreground process
5252 // should not be waiting. Note that _foregroundGCIsActive is
5253 // true while the foreground collector is waiting.
5254 if (_foregroundGCShouldWait) {
5255 // We cannot be the VM thread
5256 assert(t->is_ConcurrentGC_thread(),
5257 "Should be CMS thread");
5258 } else {
5259 // We can be the CMS thread only if we are in a stop-world
5260 // phase of CMS collection.
5261 if (t->is_ConcurrentGC_thread()) {
5262 assert(_collectorState == InitialMarking ||
5263 _collectorState == FinalMarking,
5264 "Should be a stop-world phase");
5265 // The CMS thread should be holding the CMS_token.
5266 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5267 "Potential interference with concurrently "
5268 "executing VM thread");
5269 }
5270 }
5271 }
5272 #endif
5273
5274 void CMSCollector::sweep() {
5275 assert(_collectorState == Sweeping, "just checking");
5276 check_correct_thread_executing();
5277 verify_work_stacks_empty();
5278 verify_overflow_empty();
5279 increment_sweep_count();
5280 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
5281
5282 _inter_sweep_timer.stop();
5283 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5284
5285 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5286 _intra_sweep_timer.reset();
5287 _intra_sweep_timer.start();
5288 {
5289 GCTraceCPUTime tcpu;
5290 CMSPhaseAccounting pa(this, "Concurrent Sweep");
5291 // First sweep the old gen
5292 {
5293 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5294 bitMapLock());
5295 sweepWork(_cmsGen);
5296 }
5297
5298 // Update Universe::_heap_*_at_gc figures.
5299 // We need all the free list locks to make the abstract state
5300 // transition from Sweeping to Resetting. See detailed note
5301 // further below.
5302 {
5303 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5304 // Update heap occupancy information which is used as
5305 // input to soft ref clearing policy at the next gc.
5306 Universe::update_heap_info_at_gc();
5307 _collectorState = Resizing;
5308 }
5309 }
5310 verify_work_stacks_empty();
5311 verify_overflow_empty();
5312
5313 if (should_unload_classes()) {
5314 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5315 // requires that the virtual spaces are stable and not deleted.
5316 ClassLoaderDataGraph::set_should_purge(true);
5317 }
5318
5319 _intra_sweep_timer.stop();
5320 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5321
5322 _inter_sweep_timer.reset();
5323 _inter_sweep_timer.start();
5324
5325 // We need to use a monotonically non-decreasing time in ms
5326 // or we will see time-warp warnings and os::javaTimeMillis()
5327 // does not guarantee monotonicity.
5328 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5329 update_time_of_last_gc(now);
5330
5331 // NOTE on abstract state transitions:
5332 // Mutators allocate-live and/or mark the mod-union table dirty
5333 // based on the state of the collection. The former is done in
5334 // the interval [Marking, Sweeping] and the latter in the interval
5335 // [Marking, Sweeping). Thus the transitions into the Marking state
5336 // and out of the Sweeping state must be synchronously visible
5337 // globally to the mutators.
5338 // The transition into the Marking state happens with the world
5339 // stopped so the mutators will globally see it. Sweeping is
5340 // done asynchronously by the background collector so the transition
5341 // from the Sweeping state to the Resizing state must be done
5342 // under the freelistLock (as is the check for whether to
5343 // allocate-live and whether to dirty the mod-union table).
5344 assert(_collectorState == Resizing, "Change of collector state to"
5345 " Resizing must be done under the freelistLocks (plural)");
5346
5347 // Now that sweeping has been completed, we clear
5348 // the incremental_collection_failed flag,
5349 // thus inviting a younger gen collection to promote into
5350 // this generation. If such a promotion may still fail,
5351 // the flag will be set again when a young collection is
5352 // attempted.
5353 CMSHeap* heap = CMSHeap::heap();
5354 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5355 heap->update_full_collections_completed(_collection_count_start);
5356 }
5357
5358 // FIX ME!!! Looks like this belongs in CFLSpace, with
5359 // CMSGen merely delegating to it.
5360 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5361 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5362 HeapWord* minAddr = _cmsSpace->bottom();
5363 HeapWord* largestAddr =
5364 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5365 if (largestAddr == NULL) {
5366 // The dictionary appears to be empty. In this case
5367 // try to coalesce at the end of the heap.
5368 largestAddr = _cmsSpace->end();
5369 }
5370 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5371 size_t nearLargestOffset =
5372 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5373 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5374 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5375 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5376 }
5377
5378 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5379 return addr >= _cmsSpace->nearLargestChunk();
5380 }
5381
5382 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5383 return _cmsSpace->find_chunk_at_end();
5384 }
5385
5386 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5387 bool full) {
5388 // If the young generation has been collected, gather any statistics
5389 // that are of interest at this point.
5390 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation);
5391 if (!full && current_is_young) {
5392 // Gather statistics on the young generation collection.
5393 collector()->stats().record_gc0_end(used());
5394 }
5395 }
5396
5397 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5398 // We iterate over the space(s) underlying this generation,
5399 // checking the mark bit map to see if the bits corresponding
5400 // to specific blocks are marked or not. Blocks that are
5401 // marked are live and are not swept up. All remaining blocks
5402 // are swept up, with coalescing on-the-fly as we sweep up
5403 // contiguous free and/or garbage blocks:
5404 // We need to ensure that the sweeper synchronizes with allocators
5405 // and stop-the-world collectors. In particular, the following
5406 // locks are used:
5407 // . CMS token: if this is held, a stop the world collection cannot occur
5408 // . freelistLock: if this is held no allocation can occur from this
5409 // generation by another thread
5410 // . bitMapLock: if this is held, no other thread can access or update
5411 //
5412
5413 // Note that we need to hold the freelistLock if we use
5414 // block iterate below; else the iterator might go awry if
5415 // a mutator (or promotion) causes block contents to change
5416 // (for instance if the allocator divvies up a block).
5417 // If we hold the free list lock, for all practical purposes
5418 // young generation GC's can't occur (they'll usually need to
5419 // promote), so we might as well prevent all young generation
5420 // GC's while we do a sweeping step. For the same reason, we might
5421 // as well take the bit map lock for the entire duration
5422
5423 // check that we hold the requisite locks
5424 assert(have_cms_token(), "Should hold cms token");
5425 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5426 assert_lock_strong(old_gen->freelistLock());
5427 assert_lock_strong(bitMapLock());
5428
5429 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5430 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5431 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5432 _inter_sweep_estimate.padded_average(),
5433 _intra_sweep_estimate.padded_average());
5434 old_gen->setNearLargestChunk();
5435
5436 {
5437 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5438 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5439 // We need to free-up/coalesce garbage/blocks from a
5440 // co-terminal free run. This is done in the SweepClosure
5441 // destructor; so, do not remove this scope, else the
5442 // end-of-sweep-census below will be off by a little bit.
5443 }
5444 old_gen->cmsSpace()->sweep_completed();
5445 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5446 if (should_unload_classes()) { // unloaded classes this cycle,
5447 _concurrent_cycles_since_last_unload = 0; // ... reset count
5448 } else { // did not unload classes,
5449 _concurrent_cycles_since_last_unload++; // ... increment count
5450 }
5451 }
5452
5453 // Reset CMS data structures (for now just the marking bit map)
5454 // preparatory for the next cycle.
5455 void CMSCollector::reset_concurrent() {
5456 CMSTokenSyncWithLocks ts(true, bitMapLock());
5457
5458 // If the state is not "Resetting", the foreground thread
5459 // has done a collection and the resetting.
5460 if (_collectorState != Resetting) {
5461 assert(_collectorState == Idling, "The state should only change"
5462 " because the foreground collector has finished the collection");
5463 return;
5464 }
5465
5466 {
5467 // Clear the mark bitmap (no grey objects to start with)
5468 // for the next cycle.
5469 GCTraceCPUTime tcpu;
5470 CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5471
5472 HeapWord* curAddr = _markBitMap.startWord();
5473 while (curAddr < _markBitMap.endWord()) {
5474 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5475 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5476 _markBitMap.clear_large_range(chunk);
5477 if (ConcurrentMarkSweepThread::should_yield() &&
5478 !foregroundGCIsActive() &&
5479 CMSYield) {
5480 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5481 "CMS thread should hold CMS token");
5482 assert_lock_strong(bitMapLock());
5483 bitMapLock()->unlock();
5484 ConcurrentMarkSweepThread::desynchronize(true);
5485 stopTimer();
5486 incrementYields();
5487
5488 // See the comment in coordinator_yield()
5489 for (unsigned i = 0; i < CMSYieldSleepCount &&
5490 ConcurrentMarkSweepThread::should_yield() &&
5491 !CMSCollector::foregroundGCIsActive(); ++i) {
5492 os::sleep(Thread::current(), 1, false);
5493 }
5494
5495 ConcurrentMarkSweepThread::synchronize(true);
5496 bitMapLock()->lock_without_safepoint_check();
5497 startTimer();
5498 }
5499 curAddr = chunk.end();
5500 }
5501 // A successful mostly concurrent collection has been done.
5502 // Because only the full (i.e., concurrent mode failure) collections
5503 // are being measured for gc overhead limits, clean the "near" flag
5504 // and count.
5505 size_policy()->reset_gc_overhead_limit_count();
5506 _collectorState = Idling;
5507 }
5508
5509 register_gc_end();
5510 }
5511
5512 // Same as above but for STW paths
5513 void CMSCollector::reset_stw() {
5514 // already have the lock
5515 assert(_collectorState == Resetting, "just checking");
5516 assert_lock_strong(bitMapLock());
5517 GCIdMark gc_id_mark(_cmsThread->gc_id());
5518 _markBitMap.clear_all();
5519 _collectorState = Idling;
5520 register_gc_end();
5521 }
5522
5523 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5524 GCTraceCPUTime tcpu;
5525 TraceCollectorStats tcs_cgc(cgc_counters());
5526
5527 switch (op) {
5528 case CMS_op_checkpointRootsInitial: {
5529 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5530 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5531 checkpointRootsInitial();
5532 break;
5533 }
5534 case CMS_op_checkpointRootsFinal: {
5535 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5536 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5537 checkpointRootsFinal();
5538 break;
5539 }
5540 default:
5541 fatal("No such CMS_op");
5542 }
5543 }
5544
5545 #ifndef PRODUCT
5546 size_t const CMSCollector::skip_header_HeapWords() {
5547 return FreeChunk::header_size();
5548 }
5549
5550 // Try and collect here conditions that should hold when
5551 // CMS thread is exiting. The idea is that the foreground GC
5552 // thread should not be blocked if it wants to terminate
5553 // the CMS thread and yet continue to run the VM for a while
5554 // after that.
5555 void CMSCollector::verify_ok_to_terminate() const {
5556 assert(Thread::current()->is_ConcurrentGC_thread(),
5557 "should be called by CMS thread");
5558 assert(!_foregroundGCShouldWait, "should be false");
5559 // We could check here that all the various low-level locks
5560 // are not held by the CMS thread, but that is overkill; see
5561 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5562 // is checked.
5563 }
5564 #endif
5565
5566 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5567 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5568 "missing Printezis mark?");
5569 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5570 size_t size = pointer_delta(nextOneAddr + 1, addr);
5571 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5572 "alignment problem");
5573 assert(size >= 3, "Necessary for Printezis marks to work");
5574 return size;
5575 }
5576
5577 // A variant of the above (block_size_using_printezis_bits()) except
5578 // that we return 0 if the P-bits are not yet set.
5579 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5580 if (_markBitMap.isMarked(addr + 1)) {
5581 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5582 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5583 size_t size = pointer_delta(nextOneAddr + 1, addr);
5584 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5585 "alignment problem");
5586 assert(size >= 3, "Necessary for Printezis marks to work");
5587 return size;
5588 }
5589 return 0;
5590 }
5591
5592 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5593 size_t sz = 0;
5594 oop p = (oop)addr;
5595 if (p->klass_or_null_acquire() != NULL) {
5596 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5597 } else {
5598 sz = block_size_using_printezis_bits(addr);
5599 }
5600 assert(sz > 0, "size must be nonzero");
5601 HeapWord* next_block = addr + sz;
5602 HeapWord* next_card = align_up(next_block, CardTable::card_size);
5603 assert(align_down((uintptr_t)addr, CardTable::card_size) <
5604 align_down((uintptr_t)next_card, CardTable::card_size),
5605 "must be different cards");
5606 return next_card;
5607 }
5608
5609
5610 // CMS Bit Map Wrapper /////////////////////////////////////////
5611
5612 // Construct a CMS bit map infrastructure, but don't create the
5613 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5614 // further below.
5615 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5616 _bm(),
5617 _shifter(shifter),
5618 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5619 Monitor::_safepoint_check_sometimes) : NULL)
5620 {
5621 _bmStartWord = 0;
5622 _bmWordSize = 0;
5623 }
5624
5625 bool CMSBitMap::allocate(MemRegion mr) {
5626 _bmStartWord = mr.start();
5627 _bmWordSize = mr.word_size();
5628 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5629 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5630 if (!brs.is_reserved()) {
5631 log_warning(gc)("CMS bit map allocation failure");
5632 return false;
5633 }
5634 // For now we'll just commit all of the bit map up front.
5635 // Later on we'll try to be more parsimonious with swap.
5636 if (!_virtual_space.initialize(brs, brs.size())) {
5637 log_warning(gc)("CMS bit map backing store failure");
5638 return false;
5639 }
5640 assert(_virtual_space.committed_size() == brs.size(),
5641 "didn't reserve backing store for all of CMS bit map?");
5642 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5643 _bmWordSize, "inconsistency in bit map sizing");
5644 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5645
5646 // bm.clear(); // can we rely on getting zero'd memory? verify below
5647 assert(isAllClear(),
5648 "Expected zero'd memory from ReservedSpace constructor");
5649 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5650 "consistency check");
5651 return true;
5652 }
5653
5654 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5655 HeapWord *next_addr, *end_addr, *last_addr;
5656 assert_locked();
5657 assert(covers(mr), "out-of-range error");
5658 // XXX assert that start and end are appropriately aligned
5659 for (next_addr = mr.start(), end_addr = mr.end();
5660 next_addr < end_addr; next_addr = last_addr) {
5661 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5662 last_addr = dirty_region.end();
5663 if (!dirty_region.is_empty()) {
5664 cl->do_MemRegion(dirty_region);
5665 } else {
5666 assert(last_addr == end_addr, "program logic");
5667 return;
5668 }
5669 }
5670 }
5671
5672 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5673 _bm.print_on_error(st, prefix);
5674 }
5675
5676 #ifndef PRODUCT
5677 void CMSBitMap::assert_locked() const {
5678 CMSLockVerifier::assert_locked(lock());
5679 }
5680
5681 bool CMSBitMap::covers(MemRegion mr) const {
5682 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5683 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5684 "size inconsistency");
5685 return (mr.start() >= _bmStartWord) &&
5686 (mr.end() <= endWord());
5687 }
5688
5689 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5690 return (start >= _bmStartWord && (start + size) <= endWord());
5691 }
5692
5693 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5694 // verify that there are no 1 bits in the interval [left, right)
5695 FalseBitMapClosure falseBitMapClosure;
5696 iterate(&falseBitMapClosure, left, right);
5697 }
5698
5699 void CMSBitMap::region_invariant(MemRegion mr)
5700 {
5701 assert_locked();
5702 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5703 assert(!mr.is_empty(), "unexpected empty region");
5704 assert(covers(mr), "mr should be covered by bit map");
5705 // convert address range into offset range
5706 size_t start_ofs = heapWordToOffset(mr.start());
5707 // Make sure that end() is appropriately aligned
5708 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))),
5709 "Misaligned mr.end()");
5710 size_t end_ofs = heapWordToOffset(mr.end());
5711 assert(end_ofs > start_ofs, "Should mark at least one bit");
5712 }
5713
5714 #endif
5715
5716 bool CMSMarkStack::allocate(size_t size) {
5717 // allocate a stack of the requisite depth
5718 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5719 size * sizeof(oop)));
5720 if (!rs.is_reserved()) {
5721 log_warning(gc)("CMSMarkStack allocation failure");
5722 return false;
5723 }
5724 if (!_virtual_space.initialize(rs, rs.size())) {
5725 log_warning(gc)("CMSMarkStack backing store failure");
5726 return false;
5727 }
5728 assert(_virtual_space.committed_size() == rs.size(),
5729 "didn't reserve backing store for all of CMS stack?");
5730 _base = (oop*)(_virtual_space.low());
5731 _index = 0;
5732 _capacity = size;
5733 NOT_PRODUCT(_max_depth = 0);
5734 return true;
5735 }
5736
5737 // XXX FIX ME !!! In the MT case we come in here holding a
5738 // leaf lock. For printing we need to take a further lock
5739 // which has lower rank. We need to recalibrate the two
5740 // lock-ranks involved in order to be able to print the
5741 // messages below. (Or defer the printing to the caller.
5742 // For now we take the expedient path of just disabling the
5743 // messages for the problematic case.)
5744 void CMSMarkStack::expand() {
5745 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5746 if (_capacity == MarkStackSizeMax) {
5747 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5748 // We print a warning message only once per CMS cycle.
5749 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5750 }
5751 return;
5752 }
5753 // Double capacity if possible
5754 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5755 // Do not give up existing stack until we have managed to
5756 // get the double capacity that we desired.
5757 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5758 new_capacity * sizeof(oop)));
5759 if (rs.is_reserved()) {
5760 // Release the backing store associated with old stack
5761 _virtual_space.release();
5762 // Reinitialize virtual space for new stack
5763 if (!_virtual_space.initialize(rs, rs.size())) {
5764 fatal("Not enough swap for expanded marking stack");
5765 }
5766 _base = (oop*)(_virtual_space.low());
5767 _index = 0;
5768 _capacity = new_capacity;
5769 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5770 // Failed to double capacity, continue;
5771 // we print a detail message only once per CMS cycle.
5772 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5773 _capacity / K, new_capacity / K);
5774 }
5775 }
5776
5777
5778 // Closures
5779 // XXX: there seems to be a lot of code duplication here;
5780 // should refactor and consolidate common code.
5781
5782 // This closure is used to mark refs into the CMS generation in
5783 // the CMS bit map. Called at the first checkpoint. This closure
5784 // assumes that we do not need to re-mark dirty cards; if the CMS
5785 // generation on which this is used is not an oldest
5786 // generation then this will lose younger_gen cards!
5787
5788 MarkRefsIntoClosure::MarkRefsIntoClosure(
5789 MemRegion span, CMSBitMap* bitMap):
5790 _span(span),
5791 _bitMap(bitMap)
5792 {
5793 assert(ref_discoverer() == NULL, "deliberately left NULL");
5794 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5795 }
5796
5797 void MarkRefsIntoClosure::do_oop(oop obj) {
5798 // if p points into _span, then mark corresponding bit in _markBitMap
5799 assert(oopDesc::is_oop(obj), "expected an oop");
5800 HeapWord* addr = (HeapWord*)obj;
5801 if (_span.contains(addr)) {
5802 // this should be made more efficient
5803 _bitMap->mark(addr);
5804 }
5805 }
5806
5807 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5808 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5809
5810 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5811 MemRegion span, CMSBitMap* bitMap):
5812 _span(span),
5813 _bitMap(bitMap)
5814 {
5815 assert(ref_discoverer() == NULL, "deliberately left NULL");
5816 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5817 }
5818
5819 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5820 // if p points into _span, then mark corresponding bit in _markBitMap
5821 assert(oopDesc::is_oop(obj), "expected an oop");
5822 HeapWord* addr = (HeapWord*)obj;
5823 if (_span.contains(addr)) {
5824 // this should be made more efficient
5825 _bitMap->par_mark(addr);
5826 }
5827 }
5828
5829 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5830 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5831
5832 // A variant of the above, used for CMS marking verification.
5833 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5834 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5835 _span(span),
5836 _verification_bm(verification_bm),
5837 _cms_bm(cms_bm)
5838 {
5839 assert(ref_discoverer() == NULL, "deliberately left NULL");
5840 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5841 }
5842
5843 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5844 // if p points into _span, then mark corresponding bit in _markBitMap
5845 assert(oopDesc::is_oop(obj), "expected an oop");
5846 HeapWord* addr = (HeapWord*)obj;
5847 if (_span.contains(addr)) {
5848 _verification_bm->mark(addr);
5849 if (!_cms_bm->isMarked(addr)) {
5850 Log(gc, verify) log;
5851 ResourceMark rm;
5852 LogStream ls(log.error());
5853 oop(addr)->print_on(&ls);
5854 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5855 fatal("... aborting");
5856 }
5857 }
5858 }
5859
5860 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5861 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5862
5863 //////////////////////////////////////////////////
5864 // MarkRefsIntoAndScanClosure
5865 //////////////////////////////////////////////////
5866
5867 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5868 ReferenceDiscoverer* rd,
5869 CMSBitMap* bit_map,
5870 CMSBitMap* mod_union_table,
5871 CMSMarkStack* mark_stack,
5872 CMSCollector* collector,
5873 bool should_yield,
5874 bool concurrent_precleaning):
5875 _collector(collector),
5876 _span(span),
5877 _bit_map(bit_map),
5878 _mark_stack(mark_stack),
5879 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table,
5880 mark_stack, concurrent_precleaning),
5881 _yield(should_yield),
5882 _concurrent_precleaning(concurrent_precleaning),
5883 _freelistLock(NULL)
5884 {
5885 // FIXME: Should initialize in base class constructor.
5886 assert(rd != NULL, "ref_discoverer shouldn't be NULL");
5887 set_ref_discoverer_internal(rd);
5888 }
5889
5890 // This closure is used to mark refs into the CMS generation at the
5891 // second (final) checkpoint, and to scan and transitively follow
5892 // the unmarked oops. It is also used during the concurrent precleaning
5893 // phase while scanning objects on dirty cards in the CMS generation.
5894 // The marks are made in the marking bit map and the marking stack is
5895 // used for keeping the (newly) grey objects during the scan.
5896 // The parallel version (Par_...) appears further below.
5897 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5898 if (obj != NULL) {
5899 assert(oopDesc::is_oop(obj), "expected an oop");
5900 HeapWord* addr = (HeapWord*)obj;
5901 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5902 assert(_collector->overflow_list_is_empty(),
5903 "overflow list should be empty");
5904 if (_span.contains(addr) &&
5905 !_bit_map->isMarked(addr)) {
5906 // mark bit map (object is now grey)
5907 _bit_map->mark(addr);
5908 // push on marking stack (stack should be empty), and drain the
5909 // stack by applying this closure to the oops in the oops popped
5910 // from the stack (i.e. blacken the grey objects)
5911 bool res = _mark_stack->push(obj);
5912 assert(res, "Should have space to push on empty stack");
5913 do {
5914 oop new_oop = _mark_stack->pop();
5915 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5916 assert(_bit_map->isMarked((HeapWord*)new_oop),
5917 "only grey objects on this stack");
5918 // iterate over the oops in this oop, marking and pushing
5919 // the ones in CMS heap (i.e. in _span).
5920 new_oop->oop_iterate(&_pushAndMarkClosure);
5921 // check if it's time to yield
5922 do_yield_check();
5923 } while (!_mark_stack->isEmpty() ||
5924 (!_concurrent_precleaning && take_from_overflow_list()));
5925 // if marking stack is empty, and we are not doing this
5926 // during precleaning, then check the overflow list
5927 }
5928 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5929 assert(_collector->overflow_list_is_empty(),
5930 "overflow list was drained above");
5931
5932 assert(_collector->no_preserved_marks(),
5933 "All preserved marks should have been restored above");
5934 }
5935 }
5936
5937 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5938 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5939
5940 void MarkRefsIntoAndScanClosure::do_yield_work() {
5941 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5942 "CMS thread should hold CMS token");
5943 assert_lock_strong(_freelistLock);
5944 assert_lock_strong(_bit_map->lock());
5945 // relinquish the free_list_lock and bitMaplock()
5946 _bit_map->lock()->unlock();
5947 _freelistLock->unlock();
5948 ConcurrentMarkSweepThread::desynchronize(true);
5949 _collector->stopTimer();
5950 _collector->incrementYields();
5951
5952 // See the comment in coordinator_yield()
5953 for (unsigned i = 0;
5954 i < CMSYieldSleepCount &&
5955 ConcurrentMarkSweepThread::should_yield() &&
5956 !CMSCollector::foregroundGCIsActive();
5957 ++i) {
5958 os::sleep(Thread::current(), 1, false);
5959 }
5960
5961 ConcurrentMarkSweepThread::synchronize(true);
5962 _freelistLock->lock_without_safepoint_check();
5963 _bit_map->lock()->lock_without_safepoint_check();
5964 _collector->startTimer();
5965 }
5966
5967 ///////////////////////////////////////////////////////////
5968 // ParMarkRefsIntoAndScanClosure: a parallel version of
5969 // MarkRefsIntoAndScanClosure
5970 ///////////////////////////////////////////////////////////
5971 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
5972 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd,
5973 CMSBitMap* bit_map, OopTaskQueue* work_queue):
5974 _span(span),
5975 _bit_map(bit_map),
5976 _work_queue(work_queue),
5977 _low_water_mark(MIN2((work_queue->max_elems()/4),
5978 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
5979 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue)
5980 {
5981 // FIXME: Should initialize in base class constructor.
5982 assert(rd != NULL, "ref_discoverer shouldn't be NULL");
5983 set_ref_discoverer_internal(rd);
5984 }
5985
5986 // This closure is used to mark refs into the CMS generation at the
5987 // second (final) checkpoint, and to scan and transitively follow
5988 // the unmarked oops. The marks are made in the marking bit map and
5989 // the work_queue is used for keeping the (newly) grey objects during
5990 // the scan phase whence they are also available for stealing by parallel
5991 // threads. Since the marking bit map is shared, updates are
5992 // synchronized (via CAS).
5993 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
5994 if (obj != NULL) {
5995 // Ignore mark word because this could be an already marked oop
5996 // that may be chained at the end of the overflow list.
5997 assert(oopDesc::is_oop(obj, true), "expected an oop");
5998 HeapWord* addr = (HeapWord*)obj;
5999 if (_span.contains(addr) &&
6000 !_bit_map->isMarked(addr)) {
6001 // mark bit map (object will become grey):
6002 // It is possible for several threads to be
6003 // trying to "claim" this object concurrently;
6004 // the unique thread that succeeds in marking the
6005 // object first will do the subsequent push on
6006 // to the work queue (or overflow list).
6007 if (_bit_map->par_mark(addr)) {
6008 // push on work_queue (which may not be empty), and trim the
6009 // queue to an appropriate length by applying this closure to
6010 // the oops in the oops popped from the stack (i.e. blacken the
6011 // grey objects)
6012 bool res = _work_queue->push(obj);
6013 assert(res, "Low water mark should be less than capacity?");
6014 trim_queue(_low_water_mark);
6015 } // Else, another thread claimed the object
6016 }
6017 }
6018 }
6019
6020 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6021 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6022
6023 // This closure is used to rescan the marked objects on the dirty cards
6024 // in the mod union table and the card table proper.
6025 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6026 oop p, MemRegion mr) {
6027
6028 size_t size = 0;
6029 HeapWord* addr = (HeapWord*)p;
6030 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6031 assert(_span.contains(addr), "we are scanning the CMS generation");
6032 // check if it's time to yield
6033 if (do_yield_check()) {
6034 // We yielded for some foreground stop-world work,
6035 // and we have been asked to abort this ongoing preclean cycle.
6036 return 0;
6037 }
6038 if (_bitMap->isMarked(addr)) {
6039 // it's marked; is it potentially uninitialized?
6040 if (p->klass_or_null_acquire() != NULL) {
6041 // an initialized object; ignore mark word in verification below
6042 // since we are running concurrent with mutators
6043 assert(oopDesc::is_oop(p, true), "should be an oop");
6044 if (p->is_objArray()) {
6045 // objArrays are precisely marked; restrict scanning
6046 // to dirty cards only.
6047 size = CompactibleFreeListSpace::adjustObjectSize(
6048 p->oop_iterate_size(_scanningClosure, mr));
6049 } else {
6050 // A non-array may have been imprecisely marked; we need
6051 // to scan object in its entirety.
6052 size = CompactibleFreeListSpace::adjustObjectSize(
6053 p->oop_iterate_size(_scanningClosure));
6054 }
6055 #ifdef ASSERT
6056 size_t direct_size =
6057 CompactibleFreeListSpace::adjustObjectSize(p->size());
6058 assert(size == direct_size, "Inconsistency in size");
6059 assert(size >= 3, "Necessary for Printezis marks to work");
6060 HeapWord* start_pbit = addr + 1;
6061 HeapWord* end_pbit = addr + size - 1;
6062 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6063 "inconsistent Printezis mark");
6064 // Verify inner mark bits (between Printezis bits) are clear,
6065 // but don't repeat if there are multiple dirty regions for
6066 // the same object, to avoid potential O(N^2) performance.
6067 if (addr != _last_scanned_object) {
6068 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6069 _last_scanned_object = addr;
6070 }
6071 #endif // ASSERT
6072 } else {
6073 // An uninitialized object.
6074 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6075 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6076 size = pointer_delta(nextOneAddr + 1, addr);
6077 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6078 "alignment problem");
6079 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6080 // will dirty the card when the klass pointer is installed in the
6081 // object (signaling the completion of initialization).
6082 }
6083 } else {
6084 // Either a not yet marked object or an uninitialized object
6085 if (p->klass_or_null_acquire() == NULL) {
6086 // An uninitialized object, skip to the next card, since
6087 // we may not be able to read its P-bits yet.
6088 assert(size == 0, "Initial value");
6089 } else {
6090 // An object not (yet) reached by marking: we merely need to
6091 // compute its size so as to go look at the next block.
6092 assert(oopDesc::is_oop(p, true), "should be an oop");
6093 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6094 }
6095 }
6096 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6097 return size;
6098 }
6099
6100 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6101 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6102 "CMS thread should hold CMS token");
6103 assert_lock_strong(_freelistLock);
6104 assert_lock_strong(_bitMap->lock());
6105 // relinquish the free_list_lock and bitMaplock()
6106 _bitMap->lock()->unlock();
6107 _freelistLock->unlock();
6108 ConcurrentMarkSweepThread::desynchronize(true);
6109 _collector->stopTimer();
6110 _collector->incrementYields();
6111
6112 // See the comment in coordinator_yield()
6113 for (unsigned i = 0; i < CMSYieldSleepCount &&
6114 ConcurrentMarkSweepThread::should_yield() &&
6115 !CMSCollector::foregroundGCIsActive(); ++i) {
6116 os::sleep(Thread::current(), 1, false);
6117 }
6118
6119 ConcurrentMarkSweepThread::synchronize(true);
6120 _freelistLock->lock_without_safepoint_check();
6121 _bitMap->lock()->lock_without_safepoint_check();
6122 _collector->startTimer();
6123 }
6124
6125
6126 //////////////////////////////////////////////////////////////////
6127 // SurvivorSpacePrecleanClosure
6128 //////////////////////////////////////////////////////////////////
6129 // This (single-threaded) closure is used to preclean the oops in
6130 // the survivor spaces.
6131 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6132
6133 HeapWord* addr = (HeapWord*)p;
6134 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6135 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6136 assert(p->klass_or_null() != NULL, "object should be initialized");
6137 // an initialized object; ignore mark word in verification below
6138 // since we are running concurrent with mutators
6139 assert(oopDesc::is_oop(p, true), "should be an oop");
6140 // Note that we do not yield while we iterate over
6141 // the interior oops of p, pushing the relevant ones
6142 // on our marking stack.
6143 size_t size = p->oop_iterate_size(_scanning_closure);
6144 do_yield_check();
6145 // Observe that below, we do not abandon the preclean
6146 // phase as soon as we should; rather we empty the
6147 // marking stack before returning. This is to satisfy
6148 // some existing assertions. In general, it may be a
6149 // good idea to abort immediately and complete the marking
6150 // from the grey objects at a later time.
6151 while (!_mark_stack->isEmpty()) {
6152 oop new_oop = _mark_stack->pop();
6153 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6154 assert(_bit_map->isMarked((HeapWord*)new_oop),
6155 "only grey objects on this stack");
6156 // iterate over the oops in this oop, marking and pushing
6157 // the ones in CMS heap (i.e. in _span).
6158 new_oop->oop_iterate(_scanning_closure);
6159 // check if it's time to yield
6160 do_yield_check();
6161 }
6162 unsigned int after_count =
6163 CMSHeap::heap()->total_collections();
6164 bool abort = (_before_count != after_count) ||
6165 _collector->should_abort_preclean();
6166 return abort ? 0 : size;
6167 }
6168
6169 void SurvivorSpacePrecleanClosure::do_yield_work() {
6170 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6171 "CMS thread should hold CMS token");
6172 assert_lock_strong(_bit_map->lock());
6173 // Relinquish the bit map lock
6174 _bit_map->lock()->unlock();
6175 ConcurrentMarkSweepThread::desynchronize(true);
6176 _collector->stopTimer();
6177 _collector->incrementYields();
6178
6179 // See the comment in coordinator_yield()
6180 for (unsigned i = 0; i < CMSYieldSleepCount &&
6181 ConcurrentMarkSweepThread::should_yield() &&
6182 !CMSCollector::foregroundGCIsActive(); ++i) {
6183 os::sleep(Thread::current(), 1, false);
6184 }
6185
6186 ConcurrentMarkSweepThread::synchronize(true);
6187 _bit_map->lock()->lock_without_safepoint_check();
6188 _collector->startTimer();
6189 }
6190
6191 // This closure is used to rescan the marked objects on the dirty cards
6192 // in the mod union table and the card table proper. In the parallel
6193 // case, although the bitMap is shared, we do a single read so the
6194 // isMarked() query is "safe".
6195 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6196 // Ignore mark word because we are running concurrent with mutators
6197 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6198 HeapWord* addr = (HeapWord*)p;
6199 assert(_span.contains(addr), "we are scanning the CMS generation");
6200 bool is_obj_array = false;
6201 #ifdef ASSERT
6202 if (!_parallel) {
6203 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6204 assert(_collector->overflow_list_is_empty(),
6205 "overflow list should be empty");
6206
6207 }
6208 #endif // ASSERT
6209 if (_bit_map->isMarked(addr)) {
6210 // Obj arrays are precisely marked, non-arrays are not;
6211 // so we scan objArrays precisely and non-arrays in their
6212 // entirety.
6213 if (p->is_objArray()) {
6214 is_obj_array = true;
6215 if (_parallel) {
6216 p->oop_iterate(_par_scan_closure, mr);
6217 } else {
6218 p->oop_iterate(_scan_closure, mr);
6219 }
6220 } else {
6221 if (_parallel) {
6222 p->oop_iterate(_par_scan_closure);
6223 } else {
6224 p->oop_iterate(_scan_closure);
6225 }
6226 }
6227 }
6228 #ifdef ASSERT
6229 if (!_parallel) {
6230 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6231 assert(_collector->overflow_list_is_empty(),
6232 "overflow list should be empty");
6233
6234 }
6235 #endif // ASSERT
6236 return is_obj_array;
6237 }
6238
6239 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6240 MemRegion span,
6241 CMSBitMap* bitMap, CMSMarkStack* markStack,
6242 bool should_yield, bool verifying):
6243 _collector(collector),
6244 _span(span),
6245 _bitMap(bitMap),
6246 _mut(&collector->_modUnionTable),
6247 _markStack(markStack),
6248 _yield(should_yield),
6249 _skipBits(0)
6250 {
6251 assert(_markStack->isEmpty(), "stack should be empty");
6252 _finger = _bitMap->startWord();
6253 _threshold = _finger;
6254 assert(_collector->_restart_addr == NULL, "Sanity check");
6255 assert(_span.contains(_finger), "Out of bounds _finger?");
6256 DEBUG_ONLY(_verifying = verifying;)
6257 }
6258
6259 void MarkFromRootsClosure::reset(HeapWord* addr) {
6260 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6261 assert(_span.contains(addr), "Out of bounds _finger?");
6262 _finger = addr;
6263 _threshold = align_up(_finger, CardTable::card_size);
6264 }
6265
6266 // Should revisit to see if this should be restructured for
6267 // greater efficiency.
6268 bool MarkFromRootsClosure::do_bit(size_t offset) {
6269 if (_skipBits > 0) {
6270 _skipBits--;
6271 return true;
6272 }
6273 // convert offset into a HeapWord*
6274 HeapWord* addr = _bitMap->startWord() + offset;
6275 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6276 "address out of range");
6277 assert(_bitMap->isMarked(addr), "tautology");
6278 if (_bitMap->isMarked(addr+1)) {
6279 // this is an allocated but not yet initialized object
6280 assert(_skipBits == 0, "tautology");
6281 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6282 oop p = oop(addr);
6283 if (p->klass_or_null_acquire() == NULL) {
6284 DEBUG_ONLY(if (!_verifying) {)
6285 // We re-dirty the cards on which this object lies and increase
6286 // the _threshold so that we'll come back to scan this object
6287 // during the preclean or remark phase. (CMSCleanOnEnter)
6288 if (CMSCleanOnEnter) {
6289 size_t sz = _collector->block_size_using_printezis_bits(addr);
6290 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6291 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6292 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6293 // Bump _threshold to end_card_addr; note that
6294 // _threshold cannot possibly exceed end_card_addr, anyhow.
6295 // This prevents future clearing of the card as the scan proceeds
6296 // to the right.
6297 assert(_threshold <= end_card_addr,
6298 "Because we are just scanning into this object");
6299 if (_threshold < end_card_addr) {
6300 _threshold = end_card_addr;
6301 }
6302 if (p->klass_or_null_acquire() != NULL) {
6303 // Redirty the range of cards...
6304 _mut->mark_range(redirty_range);
6305 } // ...else the setting of klass will dirty the card anyway.
6306 }
6307 DEBUG_ONLY(})
6308 return true;
6309 }
6310 }
6311 scanOopsInOop(addr);
6312 return true;
6313 }
6314
6315 // We take a break if we've been at this for a while,
6316 // so as to avoid monopolizing the locks involved.
6317 void MarkFromRootsClosure::do_yield_work() {
6318 // First give up the locks, then yield, then re-lock
6319 // We should probably use a constructor/destructor idiom to
6320 // do this unlock/lock or modify the MutexUnlocker class to
6321 // serve our purpose. XXX
6322 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6323 "CMS thread should hold CMS token");
6324 assert_lock_strong(_bitMap->lock());
6325 _bitMap->lock()->unlock();
6326 ConcurrentMarkSweepThread::desynchronize(true);
6327 _collector->stopTimer();
6328 _collector->incrementYields();
6329
6330 // See the comment in coordinator_yield()
6331 for (unsigned i = 0; i < CMSYieldSleepCount &&
6332 ConcurrentMarkSweepThread::should_yield() &&
6333 !CMSCollector::foregroundGCIsActive(); ++i) {
6334 os::sleep(Thread::current(), 1, false);
6335 }
6336
6337 ConcurrentMarkSweepThread::synchronize(true);
6338 _bitMap->lock()->lock_without_safepoint_check();
6339 _collector->startTimer();
6340 }
6341
6342 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6343 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6344 assert(_markStack->isEmpty(),
6345 "should drain stack to limit stack usage");
6346 // convert ptr to an oop preparatory to scanning
6347 oop obj = oop(ptr);
6348 // Ignore mark word in verification below, since we
6349 // may be running concurrent with mutators.
6350 assert(oopDesc::is_oop(obj, true), "should be an oop");
6351 assert(_finger <= ptr, "_finger runneth ahead");
6352 // advance the finger to right end of this object
6353 _finger = ptr + obj->size();
6354 assert(_finger > ptr, "we just incremented it above");
6355 // On large heaps, it may take us some time to get through
6356 // the marking phase. During
6357 // this time it's possible that a lot of mutations have
6358 // accumulated in the card table and the mod union table --
6359 // these mutation records are redundant until we have
6360 // actually traced into the corresponding card.
6361 // Here, we check whether advancing the finger would make
6362 // us cross into a new card, and if so clear corresponding
6363 // cards in the MUT (preclean them in the card-table in the
6364 // future).
6365
6366 DEBUG_ONLY(if (!_verifying) {)
6367 // The clean-on-enter optimization is disabled by default,
6368 // until we fix 6178663.
6369 if (CMSCleanOnEnter && (_finger > _threshold)) {
6370 // [_threshold, _finger) represents the interval
6371 // of cards to be cleared in MUT (or precleaned in card table).
6372 // The set of cards to be cleared is all those that overlap
6373 // with the interval [_threshold, _finger); note that
6374 // _threshold is always kept card-aligned but _finger isn't
6375 // always card-aligned.
6376 HeapWord* old_threshold = _threshold;
6377 assert(is_aligned(old_threshold, CardTable::card_size),
6378 "_threshold should always be card-aligned");
6379 _threshold = align_up(_finger, CardTable::card_size);
6380 MemRegion mr(old_threshold, _threshold);
6381 assert(!mr.is_empty(), "Control point invariant");
6382 assert(_span.contains(mr), "Should clear within span");
6383 _mut->clear_range(mr);
6384 }
6385 DEBUG_ONLY(})
6386 // Note: the finger doesn't advance while we drain
6387 // the stack below.
6388 PushOrMarkClosure pushOrMarkClosure(_collector,
6389 _span, _bitMap, _markStack,
6390 _finger, this);
6391 bool res = _markStack->push(obj);
6392 assert(res, "Empty non-zero size stack should have space for single push");
6393 while (!_markStack->isEmpty()) {
6394 oop new_oop = _markStack->pop();
6395 // Skip verifying header mark word below because we are
6396 // running concurrent with mutators.
6397 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6398 // now scan this oop's oops
6399 new_oop->oop_iterate(&pushOrMarkClosure);
6400 do_yield_check();
6401 }
6402 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6403 }
6404
6405 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6406 CMSCollector* collector, MemRegion span,
6407 CMSBitMap* bit_map,
6408 OopTaskQueue* work_queue,
6409 CMSMarkStack* overflow_stack):
6410 _collector(collector),
6411 _whole_span(collector->_span),
6412 _span(span),
6413 _bit_map(bit_map),
6414 _mut(&collector->_modUnionTable),
6415 _work_queue(work_queue),
6416 _overflow_stack(overflow_stack),
6417 _skip_bits(0),
6418 _task(task)
6419 {
6420 assert(_work_queue->size() == 0, "work_queue should be empty");
6421 _finger = span.start();
6422 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6423 assert(_span.contains(_finger), "Out of bounds _finger?");
6424 }
6425
6426 // Should revisit to see if this should be restructured for
6427 // greater efficiency.
6428 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6429 if (_skip_bits > 0) {
6430 _skip_bits--;
6431 return true;
6432 }
6433 // convert offset into a HeapWord*
6434 HeapWord* addr = _bit_map->startWord() + offset;
6435 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6436 "address out of range");
6437 assert(_bit_map->isMarked(addr), "tautology");
6438 if (_bit_map->isMarked(addr+1)) {
6439 // this is an allocated object that might not yet be initialized
6440 assert(_skip_bits == 0, "tautology");
6441 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6442 oop p = oop(addr);
6443 if (p->klass_or_null_acquire() == NULL) {
6444 // in the case of Clean-on-Enter optimization, redirty card
6445 // and avoid clearing card by increasing the threshold.
6446 return true;
6447 }
6448 }
6449 scan_oops_in_oop(addr);
6450 return true;
6451 }
6452
6453 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6454 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6455 // Should we assert that our work queue is empty or
6456 // below some drain limit?
6457 assert(_work_queue->size() == 0,
6458 "should drain stack to limit stack usage");
6459 // convert ptr to an oop preparatory to scanning
6460 oop obj = oop(ptr);
6461 // Ignore mark word in verification below, since we
6462 // may be running concurrent with mutators.
6463 assert(oopDesc::is_oop(obj, true), "should be an oop");
6464 assert(_finger <= ptr, "_finger runneth ahead");
6465 // advance the finger to right end of this object
6466 _finger = ptr + obj->size();
6467 assert(_finger > ptr, "we just incremented it above");
6468 // On large heaps, it may take us some time to get through
6469 // the marking phase. During
6470 // this time it's possible that a lot of mutations have
6471 // accumulated in the card table and the mod union table --
6472 // these mutation records are redundant until we have
6473 // actually traced into the corresponding card.
6474 // Here, we check whether advancing the finger would make
6475 // us cross into a new card, and if so clear corresponding
6476 // cards in the MUT (preclean them in the card-table in the
6477 // future).
6478
6479 // The clean-on-enter optimization is disabled by default,
6480 // until we fix 6178663.
6481 if (CMSCleanOnEnter && (_finger > _threshold)) {
6482 // [_threshold, _finger) represents the interval
6483 // of cards to be cleared in MUT (or precleaned in card table).
6484 // The set of cards to be cleared is all those that overlap
6485 // with the interval [_threshold, _finger); note that
6486 // _threshold is always kept card-aligned but _finger isn't
6487 // always card-aligned.
6488 HeapWord* old_threshold = _threshold;
6489 assert(is_aligned(old_threshold, CardTable::card_size),
6490 "_threshold should always be card-aligned");
6491 _threshold = align_up(_finger, CardTable::card_size);
6492 MemRegion mr(old_threshold, _threshold);
6493 assert(!mr.is_empty(), "Control point invariant");
6494 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6495 _mut->clear_range(mr);
6496 }
6497
6498 // Note: the local finger doesn't advance while we drain
6499 // the stack below, but the global finger sure can and will.
6500 HeapWord* volatile* gfa = _task->global_finger_addr();
6501 ParPushOrMarkClosure pushOrMarkClosure(_collector,
6502 _span, _bit_map,
6503 _work_queue,
6504 _overflow_stack,
6505 _finger,
6506 gfa, this);
6507 bool res = _work_queue->push(obj); // overflow could occur here
6508 assert(res, "Will hold once we use workqueues");
6509 while (true) {
6510 oop new_oop;
6511 if (!_work_queue->pop_local(new_oop)) {
6512 // We emptied our work_queue; check if there's stuff that can
6513 // be gotten from the overflow stack.
6514 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6515 _overflow_stack, _work_queue)) {
6516 do_yield_check();
6517 continue;
6518 } else { // done
6519 break;
6520 }
6521 }
6522 // Skip verifying header mark word below because we are
6523 // running concurrent with mutators.
6524 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6525 // now scan this oop's oops
6526 new_oop->oop_iterate(&pushOrMarkClosure);
6527 do_yield_check();
6528 }
6529 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6530 }
6531
6532 // Yield in response to a request from VM Thread or
6533 // from mutators.
6534 void ParMarkFromRootsClosure::do_yield_work() {
6535 assert(_task != NULL, "sanity");
6536 _task->yield();
6537 }
6538
6539 // A variant of the above used for verifying CMS marking work.
6540 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6541 MemRegion span,
6542 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6543 CMSMarkStack* mark_stack):
6544 _collector(collector),
6545 _span(span),
6546 _verification_bm(verification_bm),
6547 _cms_bm(cms_bm),
6548 _mark_stack(mark_stack),
6549 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6550 mark_stack)
6551 {
6552 assert(_mark_stack->isEmpty(), "stack should be empty");
6553 _finger = _verification_bm->startWord();
6554 assert(_collector->_restart_addr == NULL, "Sanity check");
6555 assert(_span.contains(_finger), "Out of bounds _finger?");
6556 }
6557
6558 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6559 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6560 assert(_span.contains(addr), "Out of bounds _finger?");
6561 _finger = addr;
6562 }
6563
6564 // Should revisit to see if this should be restructured for
6565 // greater efficiency.
6566 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6567 // convert offset into a HeapWord*
6568 HeapWord* addr = _verification_bm->startWord() + offset;
6569 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6570 "address out of range");
6571 assert(_verification_bm->isMarked(addr), "tautology");
6572 assert(_cms_bm->isMarked(addr), "tautology");
6573
6574 assert(_mark_stack->isEmpty(),
6575 "should drain stack to limit stack usage");
6576 // convert addr to an oop preparatory to scanning
6577 oop obj = oop(addr);
6578 assert(oopDesc::is_oop(obj), "should be an oop");
6579 assert(_finger <= addr, "_finger runneth ahead");
6580 // advance the finger to right end of this object
6581 _finger = addr + obj->size();
6582 assert(_finger > addr, "we just incremented it above");
6583 // Note: the finger doesn't advance while we drain
6584 // the stack below.
6585 bool res = _mark_stack->push(obj);
6586 assert(res, "Empty non-zero size stack should have space for single push");
6587 while (!_mark_stack->isEmpty()) {
6588 oop new_oop = _mark_stack->pop();
6589 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6590 // now scan this oop's oops
6591 new_oop->oop_iterate(&_pam_verify_closure);
6592 }
6593 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6594 return true;
6595 }
6596
6597 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6598 CMSCollector* collector, MemRegion span,
6599 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6600 CMSMarkStack* mark_stack):
6601 MetadataAwareOopClosure(collector->ref_processor()),
6602 _collector(collector),
6603 _span(span),
6604 _verification_bm(verification_bm),
6605 _cms_bm(cms_bm),
6606 _mark_stack(mark_stack)
6607 { }
6608
6609 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) {
6610 oop obj = RawAccess<>::oop_load(p);
6611 do_oop(obj);
6612 }
6613
6614 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6615 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6616
6617 // Upon stack overflow, we discard (part of) the stack,
6618 // remembering the least address amongst those discarded
6619 // in CMSCollector's _restart_address.
6620 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6621 // Remember the least grey address discarded
6622 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6623 _collector->lower_restart_addr(ra);
6624 _mark_stack->reset(); // discard stack contents
6625 _mark_stack->expand(); // expand the stack if possible
6626 }
6627
6628 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6629 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6630 HeapWord* addr = (HeapWord*)obj;
6631 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6632 // Oop lies in _span and isn't yet grey or black
6633 _verification_bm->mark(addr); // now grey
6634 if (!_cms_bm->isMarked(addr)) {
6635 Log(gc, verify) log;
6636 ResourceMark rm;
6637 LogStream ls(log.error());
6638 oop(addr)->print_on(&ls);
6639 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6640 fatal("... aborting");
6641 }
6642
6643 if (!_mark_stack->push(obj)) { // stack overflow
6644 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6645 assert(_mark_stack->isFull(), "Else push should have succeeded");
6646 handle_stack_overflow(addr);
6647 }
6648 // anything including and to the right of _finger
6649 // will be scanned as we iterate over the remainder of the
6650 // bit map
6651 }
6652 }
6653
6654 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6655 MemRegion span,
6656 CMSBitMap* bitMap, CMSMarkStack* markStack,
6657 HeapWord* finger, MarkFromRootsClosure* parent) :
6658 MetadataAwareOopClosure(collector->ref_processor()),
6659 _collector(collector),
6660 _span(span),
6661 _bitMap(bitMap),
6662 _markStack(markStack),
6663 _finger(finger),
6664 _parent(parent)
6665 { }
6666
6667 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6668 MemRegion span,
6669 CMSBitMap* bit_map,
6670 OopTaskQueue* work_queue,
6671 CMSMarkStack* overflow_stack,
6672 HeapWord* finger,
6673 HeapWord* volatile* global_finger_addr,
6674 ParMarkFromRootsClosure* parent) :
6675 MetadataAwareOopClosure(collector->ref_processor()),
6676 _collector(collector),
6677 _whole_span(collector->_span),
6678 _span(span),
6679 _bit_map(bit_map),
6680 _work_queue(work_queue),
6681 _overflow_stack(overflow_stack),
6682 _finger(finger),
6683 _global_finger_addr(global_finger_addr),
6684 _parent(parent)
6685 { }
6686
6687 // Assumes thread-safe access by callers, who are
6688 // responsible for mutual exclusion.
6689 void CMSCollector::lower_restart_addr(HeapWord* low) {
6690 assert(_span.contains(low), "Out of bounds addr");
6691 if (_restart_addr == NULL) {
6692 _restart_addr = low;
6693 } else {
6694 _restart_addr = MIN2(_restart_addr, low);
6695 }
6696 }
6697
6698 // Upon stack overflow, we discard (part of) the stack,
6699 // remembering the least address amongst those discarded
6700 // in CMSCollector's _restart_address.
6701 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6702 // Remember the least grey address discarded
6703 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6704 _collector->lower_restart_addr(ra);
6705 _markStack->reset(); // discard stack contents
6706 _markStack->expand(); // expand the stack if possible
6707 }
6708
6709 // Upon stack overflow, we discard (part of) the stack,
6710 // remembering the least address amongst those discarded
6711 // in CMSCollector's _restart_address.
6712 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6713 // We need to do this under a mutex to prevent other
6714 // workers from interfering with the work done below.
6715 MutexLockerEx ml(_overflow_stack->par_lock(),
6716 Mutex::_no_safepoint_check_flag);
6717 // Remember the least grey address discarded
6718 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6719 _collector->lower_restart_addr(ra);
6720 _overflow_stack->reset(); // discard stack contents
6721 _overflow_stack->expand(); // expand the stack if possible
6722 }
6723
6724 void PushOrMarkClosure::do_oop(oop obj) {
6725 // Ignore mark word because we are running concurrent with mutators.
6726 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6727 HeapWord* addr = (HeapWord*)obj;
6728 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6729 // Oop lies in _span and isn't yet grey or black
6730 _bitMap->mark(addr); // now grey
6731 if (addr < _finger) {
6732 // the bit map iteration has already either passed, or
6733 // sampled, this bit in the bit map; we'll need to
6734 // use the marking stack to scan this oop's oops.
6735 bool simulate_overflow = false;
6736 NOT_PRODUCT(
6737 if (CMSMarkStackOverflowALot &&
6738 _collector->simulate_overflow()) {
6739 // simulate a stack overflow
6740 simulate_overflow = true;
6741 }
6742 )
6743 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6744 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6745 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6746 handle_stack_overflow(addr);
6747 }
6748 }
6749 // anything including and to the right of _finger
6750 // will be scanned as we iterate over the remainder of the
6751 // bit map
6752 do_yield_check();
6753 }
6754 }
6755
6756 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
6757 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6758
6759 void ParPushOrMarkClosure::do_oop(oop obj) {
6760 // Ignore mark word because we are running concurrent with mutators.
6761 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6762 HeapWord* addr = (HeapWord*)obj;
6763 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6764 // Oop lies in _span and isn't yet grey or black
6765 // We read the global_finger (volatile read) strictly after marking oop
6766 bool res = _bit_map->par_mark(addr); // now grey
6767 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6768 // Should we push this marked oop on our stack?
6769 // -- if someone else marked it, nothing to do
6770 // -- if target oop is above global finger nothing to do
6771 // -- if target oop is in chunk and above local finger
6772 // then nothing to do
6773 // -- else push on work queue
6774 if ( !res // someone else marked it, they will deal with it
6775 || (addr >= *gfa) // will be scanned in a later task
6776 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6777 return;
6778 }
6779 // the bit map iteration has already either passed, or
6780 // sampled, this bit in the bit map; we'll need to
6781 // use the marking stack to scan this oop's oops.
6782 bool simulate_overflow = false;
6783 NOT_PRODUCT(
6784 if (CMSMarkStackOverflowALot &&
6785 _collector->simulate_overflow()) {
6786 // simulate a stack overflow
6787 simulate_overflow = true;
6788 }
6789 )
6790 if (simulate_overflow ||
6791 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6792 // stack overflow
6793 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6794 // We cannot assert that the overflow stack is full because
6795 // it may have been emptied since.
6796 assert(simulate_overflow ||
6797 _work_queue->size() == _work_queue->max_elems(),
6798 "Else push should have succeeded");
6799 handle_stack_overflow(addr);
6800 }
6801 do_yield_check();
6802 }
6803 }
6804
6805 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6806 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6807
6808 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6809 MemRegion span,
6810 ReferenceDiscoverer* rd,
6811 CMSBitMap* bit_map,
6812 CMSBitMap* mod_union_table,
6813 CMSMarkStack* mark_stack,
6814 bool concurrent_precleaning):
6815 MetadataAwareOopClosure(rd),
6816 _collector(collector),
6817 _span(span),
6818 _bit_map(bit_map),
6819 _mod_union_table(mod_union_table),
6820 _mark_stack(mark_stack),
6821 _concurrent_precleaning(concurrent_precleaning)
6822 {
6823 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL");
6824 }
6825
6826 // Grey object rescan during pre-cleaning and second checkpoint phases --
6827 // the non-parallel version (the parallel version appears further below.)
6828 void PushAndMarkClosure::do_oop(oop obj) {
6829 // Ignore mark word verification. If during concurrent precleaning,
6830 // the object monitor may be locked. If during the checkpoint
6831 // phases, the object may already have been reached by a different
6832 // path and may be at the end of the global overflow list (so
6833 // the mark word may be NULL).
6834 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6835 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6836 HeapWord* addr = (HeapWord*)obj;
6837 // Check if oop points into the CMS generation
6838 // and is not marked
6839 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6840 // a white object ...
6841 _bit_map->mark(addr); // ... now grey
6842 // push on the marking stack (grey set)
6843 bool simulate_overflow = false;
6844 NOT_PRODUCT(
6845 if (CMSMarkStackOverflowALot &&
6846 _collector->simulate_overflow()) {
6847 // simulate a stack overflow
6848 simulate_overflow = true;
6849 }
6850 )
6851 if (simulate_overflow || !_mark_stack->push(obj)) {
6852 if (_concurrent_precleaning) {
6853 // During precleaning we can just dirty the appropriate card(s)
6854 // in the mod union table, thus ensuring that the object remains
6855 // in the grey set and continue. In the case of object arrays
6856 // we need to dirty all of the cards that the object spans,
6857 // since the rescan of object arrays will be limited to the
6858 // dirty cards.
6859 // Note that no one can be interfering with us in this action
6860 // of dirtying the mod union table, so no locking or atomics
6861 // are required.
6862 if (obj->is_objArray()) {
6863 size_t sz = obj->size();
6864 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6865 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6866 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6867 _mod_union_table->mark_range(redirty_range);
6868 } else {
6869 _mod_union_table->mark(addr);
6870 }
6871 _collector->_ser_pmc_preclean_ovflw++;
6872 } else {
6873 // During the remark phase, we need to remember this oop
6874 // in the overflow list.
6875 _collector->push_on_overflow_list(obj);
6876 _collector->_ser_pmc_remark_ovflw++;
6877 }
6878 }
6879 }
6880 }
6881
6882 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6883 MemRegion span,
6884 ReferenceDiscoverer* rd,
6885 CMSBitMap* bit_map,
6886 OopTaskQueue* work_queue):
6887 MetadataAwareOopClosure(rd),
6888 _collector(collector),
6889 _span(span),
6890 _bit_map(bit_map),
6891 _work_queue(work_queue)
6892 {
6893 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL");
6894 }
6895
6896 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
6897 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6898
6899 // Grey object rescan during second checkpoint phase --
6900 // the parallel version.
6901 void ParPushAndMarkClosure::do_oop(oop obj) {
6902 // In the assert below, we ignore the mark word because
6903 // this oop may point to an already visited object that is
6904 // on the overflow stack (in which case the mark word has
6905 // been hijacked for chaining into the overflow stack --
6906 // if this is the last object in the overflow stack then
6907 // its mark word will be NULL). Because this object may
6908 // have been subsequently popped off the global overflow
6909 // stack, and the mark word possibly restored to the prototypical
6910 // value, by the time we get to examined this failing assert in
6911 // the debugger, is_oop_or_null(false) may subsequently start
6912 // to hold.
6913 assert(oopDesc::is_oop_or_null(obj, true),
6914 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6915 HeapWord* addr = (HeapWord*)obj;
6916 // Check if oop points into the CMS generation
6917 // and is not marked
6918 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6919 // a white object ...
6920 // If we manage to "claim" the object, by being the
6921 // first thread to mark it, then we push it on our
6922 // marking stack
6923 if (_bit_map->par_mark(addr)) { // ... now grey
6924 // push on work queue (grey set)
6925 bool simulate_overflow = false;
6926 NOT_PRODUCT(
6927 if (CMSMarkStackOverflowALot &&
6928 _collector->par_simulate_overflow()) {
6929 // simulate a stack overflow
6930 simulate_overflow = true;
6931 }
6932 )
6933 if (simulate_overflow || !_work_queue->push(obj)) {
6934 _collector->par_push_on_overflow_list(obj);
6935 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
6936 }
6937 } // Else, some other thread got there first
6938 }
6939 }
6940
6941 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6942 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6943
6944 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6945 Mutex* bml = _collector->bitMapLock();
6946 assert_lock_strong(bml);
6947 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6948 "CMS thread should hold CMS token");
6949
6950 bml->unlock();
6951 ConcurrentMarkSweepThread::desynchronize(true);
6952
6953 _collector->stopTimer();
6954 _collector->incrementYields();
6955
6956 // See the comment in coordinator_yield()
6957 for (unsigned i = 0; i < CMSYieldSleepCount &&
6958 ConcurrentMarkSweepThread::should_yield() &&
6959 !CMSCollector::foregroundGCIsActive(); ++i) {
6960 os::sleep(Thread::current(), 1, false);
6961 }
6962
6963 ConcurrentMarkSweepThread::synchronize(true);
6964 bml->lock();
6965
6966 _collector->startTimer();
6967 }
6968
6969 bool CMSPrecleanRefsYieldClosure::should_return() {
6970 if (ConcurrentMarkSweepThread::should_yield()) {
6971 do_yield_work();
6972 }
6973 return _collector->foregroundGCIsActive();
6974 }
6975
6976 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
6977 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0,
6978 "mr should be aligned to start at a card boundary");
6979 // We'd like to assert:
6980 // assert(mr.word_size()%CardTable::card_size_in_words == 0,
6981 // "mr should be a range of cards");
6982 // However, that would be too strong in one case -- the last
6983 // partition ends at _unallocated_block which, in general, can be
6984 // an arbitrary boundary, not necessarily card aligned.
6985 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words;
6986 _space->object_iterate_mem(mr, &_scan_cl);
6987 }
6988
6989 SweepClosure::SweepClosure(CMSCollector* collector,
6990 ConcurrentMarkSweepGeneration* g,
6991 CMSBitMap* bitMap, bool should_yield) :
6992 _collector(collector),
6993 _g(g),
6994 _sp(g->cmsSpace()),
6995 _limit(_sp->sweep_limit()),
6996 _freelistLock(_sp->freelistLock()),
6997 _bitMap(bitMap),
6998 _yield(should_yield),
6999 _inFreeRange(false), // No free range at beginning of sweep
7000 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7001 _lastFreeRangeCoalesced(false),
7002 _freeFinger(g->used_region().start())
7003 {
7004 NOT_PRODUCT(
7005 _numObjectsFreed = 0;
7006 _numWordsFreed = 0;
7007 _numObjectsLive = 0;
7008 _numWordsLive = 0;
7009 _numObjectsAlreadyFree = 0;
7010 _numWordsAlreadyFree = 0;
7011 _last_fc = NULL;
7012
7013 _sp->initializeIndexedFreeListArrayReturnedBytes();
7014 _sp->dictionary()->initialize_dict_returned_bytes();
7015 )
7016 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7017 "sweep _limit out of bounds");
7018 log_develop_trace(gc, sweep)("====================");
7019 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7020 }
7021
7022 void SweepClosure::print_on(outputStream* st) const {
7023 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7024 p2i(_sp->bottom()), p2i(_sp->end()));
7025 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7026 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7027 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7028 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7029 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7030 }
7031
7032 #ifndef PRODUCT
7033 // Assertion checking only: no useful work in product mode --
7034 // however, if any of the flags below become product flags,
7035 // you may need to review this code to see if it needs to be
7036 // enabled in product mode.
7037 SweepClosure::~SweepClosure() {
7038 assert_lock_strong(_freelistLock);
7039 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7040 "sweep _limit out of bounds");
7041 if (inFreeRange()) {
7042 Log(gc, sweep) log;
7043 log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7044 ResourceMark rm;
7045 LogStream ls(log.error());
7046 print_on(&ls);
7047 ShouldNotReachHere();
7048 }
7049
7050 if (log_is_enabled(Debug, gc, sweep)) {
7051 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7052 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7053 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7054 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7055 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7056 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7057 }
7058
7059 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7060 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7061 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7062 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7063 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes",
7064 returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7065 }
7066 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7067 log_develop_trace(gc, sweep)("================");
7068 }
7069 #endif // PRODUCT
7070
7071 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7072 bool freeRangeInFreeLists) {
7073 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7074 p2i(freeFinger), freeRangeInFreeLists);
7075 assert(!inFreeRange(), "Trampling existing free range");
7076 set_inFreeRange(true);
7077 set_lastFreeRangeCoalesced(false);
7078
7079 set_freeFinger(freeFinger);
7080 set_freeRangeInFreeLists(freeRangeInFreeLists);
7081 if (CMSTestInFreeList) {
7082 if (freeRangeInFreeLists) {
7083 FreeChunk* fc = (FreeChunk*) freeFinger;
7084 assert(fc->is_free(), "A chunk on the free list should be free.");
7085 assert(fc->size() > 0, "Free range should have a size");
7086 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7087 }
7088 }
7089 }
7090
7091 // Note that the sweeper runs concurrently with mutators. Thus,
7092 // it is possible for direct allocation in this generation to happen
7093 // in the middle of the sweep. Note that the sweeper also coalesces
7094 // contiguous free blocks. Thus, unless the sweeper and the allocator
7095 // synchronize appropriately freshly allocated blocks may get swept up.
7096 // This is accomplished by the sweeper locking the free lists while
7097 // it is sweeping. Thus blocks that are determined to be free are
7098 // indeed free. There is however one additional complication:
7099 // blocks that have been allocated since the final checkpoint and
7100 // mark, will not have been marked and so would be treated as
7101 // unreachable and swept up. To prevent this, the allocator marks
7102 // the bit map when allocating during the sweep phase. This leads,
7103 // however, to a further complication -- objects may have been allocated
7104 // but not yet initialized -- in the sense that the header isn't yet
7105 // installed. The sweeper can not then determine the size of the block
7106 // in order to skip over it. To deal with this case, we use a technique
7107 // (due to Printezis) to encode such uninitialized block sizes in the
7108 // bit map. Since the bit map uses a bit per every HeapWord, but the
7109 // CMS generation has a minimum object size of 3 HeapWords, it follows
7110 // that "normal marks" won't be adjacent in the bit map (there will
7111 // always be at least two 0 bits between successive 1 bits). We make use
7112 // of these "unused" bits to represent uninitialized blocks -- the bit
7113 // corresponding to the start of the uninitialized object and the next
7114 // bit are both set. Finally, a 1 bit marks the end of the object that
7115 // started with the two consecutive 1 bits to indicate its potentially
7116 // uninitialized state.
7117
7118 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7119 FreeChunk* fc = (FreeChunk*)addr;
7120 size_t res;
7121
7122 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7123 // than "addr == _limit" because although _limit was a block boundary when
7124 // we started the sweep, it may no longer be one because heap expansion
7125 // may have caused us to coalesce the block ending at the address _limit
7126 // with a newly expanded chunk (this happens when _limit was set to the
7127 // previous _end of the space), so we may have stepped past _limit:
7128 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7129 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7130 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7131 "sweep _limit out of bounds");
7132 assert(addr < _sp->end(), "addr out of bounds");
7133 // Flush any free range we might be holding as a single
7134 // coalesced chunk to the appropriate free list.
7135 if (inFreeRange()) {
7136 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7137 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger()));
7138 flush_cur_free_chunk(freeFinger(),
7139 pointer_delta(addr, freeFinger()));
7140 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7141 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7142 lastFreeRangeCoalesced() ? 1 : 0);
7143 }
7144
7145 // help the iterator loop finish
7146 return pointer_delta(_sp->end(), addr);
7147 }
7148
7149 assert(addr < _limit, "sweep invariant");
7150 // check if we should yield
7151 do_yield_check(addr);
7152 if (fc->is_free()) {
7153 // Chunk that is already free
7154 res = fc->size();
7155 do_already_free_chunk(fc);
7156 debug_only(_sp->verifyFreeLists());
7157 // If we flush the chunk at hand in lookahead_and_flush()
7158 // and it's coalesced with a preceding chunk, then the
7159 // process of "mangling" the payload of the coalesced block
7160 // will cause erasure of the size information from the
7161 // (erstwhile) header of all the coalesced blocks but the
7162 // first, so the first disjunct in the assert will not hold
7163 // in that specific case (in which case the second disjunct
7164 // will hold).
7165 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7166 "Otherwise the size info doesn't change at this step");
7167 NOT_PRODUCT(
7168 _numObjectsAlreadyFree++;
7169 _numWordsAlreadyFree += res;
7170 )
7171 NOT_PRODUCT(_last_fc = fc;)
7172 } else if (!_bitMap->isMarked(addr)) {
7173 // Chunk is fresh garbage
7174 res = do_garbage_chunk(fc);
7175 debug_only(_sp->verifyFreeLists());
7176 NOT_PRODUCT(
7177 _numObjectsFreed++;
7178 _numWordsFreed += res;
7179 )
7180 } else {
7181 // Chunk that is alive.
7182 res = do_live_chunk(fc);
7183 debug_only(_sp->verifyFreeLists());
7184 NOT_PRODUCT(
7185 _numObjectsLive++;
7186 _numWordsLive += res;
7187 )
7188 }
7189 return res;
7190 }
7191
7192 // For the smart allocation, record following
7193 // split deaths - a free chunk is removed from its free list because
7194 // it is being split into two or more chunks.
7195 // split birth - a free chunk is being added to its free list because
7196 // a larger free chunk has been split and resulted in this free chunk.
7197 // coal death - a free chunk is being removed from its free list because
7198 // it is being coalesced into a large free chunk.
7199 // coal birth - a free chunk is being added to its free list because
7200 // it was created when two or more free chunks where coalesced into
7201 // this free chunk.
7202 //
7203 // These statistics are used to determine the desired number of free
7204 // chunks of a given size. The desired number is chosen to be relative
7205 // to the end of a CMS sweep. The desired number at the end of a sweep
7206 // is the
7207 // count-at-end-of-previous-sweep (an amount that was enough)
7208 // - count-at-beginning-of-current-sweep (the excess)
7209 // + split-births (gains in this size during interval)
7210 // - split-deaths (demands on this size during interval)
7211 // where the interval is from the end of one sweep to the end of the
7212 // next.
7213 //
7214 // When sweeping the sweeper maintains an accumulated chunk which is
7215 // the chunk that is made up of chunks that have been coalesced. That
7216 // will be termed the left-hand chunk. A new chunk of garbage that
7217 // is being considered for coalescing will be referred to as the
7218 // right-hand chunk.
7219 //
7220 // When making a decision on whether to coalesce a right-hand chunk with
7221 // the current left-hand chunk, the current count vs. the desired count
7222 // of the left-hand chunk is considered. Also if the right-hand chunk
7223 // is near the large chunk at the end of the heap (see
7224 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7225 // left-hand chunk is coalesced.
7226 //
7227 // When making a decision about whether to split a chunk, the desired count
7228 // vs. the current count of the candidate to be split is also considered.
7229 // If the candidate is underpopulated (currently fewer chunks than desired)
7230 // a chunk of an overpopulated (currently more chunks than desired) size may
7231 // be chosen. The "hint" associated with a free list, if non-null, points
7232 // to a free list which may be overpopulated.
7233 //
7234
7235 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7236 const size_t size = fc->size();
7237 // Chunks that cannot be coalesced are not in the
7238 // free lists.
7239 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7240 assert(_sp->verify_chunk_in_free_list(fc),
7241 "free chunk should be in free lists");
7242 }
7243 // a chunk that is already free, should not have been
7244 // marked in the bit map
7245 HeapWord* const addr = (HeapWord*) fc;
7246 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7247 // Verify that the bit map has no bits marked between
7248 // addr and purported end of this block.
7249 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7250
7251 // Some chunks cannot be coalesced under any circumstances.
7252 // See the definition of cantCoalesce().
7253 if (!fc->cantCoalesce()) {
7254 // This chunk can potentially be coalesced.
7255 // All the work is done in
7256 do_post_free_or_garbage_chunk(fc, size);
7257 // Note that if the chunk is not coalescable (the else arm
7258 // below), we unconditionally flush, without needing to do
7259 // a "lookahead," as we do below.
7260 if (inFreeRange()) lookahead_and_flush(fc, size);
7261 } else {
7262 // Code path common to both original and adaptive free lists.
7263
7264 // cant coalesce with previous block; this should be treated
7265 // as the end of a free run if any
7266 if (inFreeRange()) {
7267 // we kicked some butt; time to pick up the garbage
7268 assert(freeFinger() < addr, "freeFinger points too high");
7269 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7270 }
7271 // else, nothing to do, just continue
7272 }
7273 }
7274
7275 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7276 // This is a chunk of garbage. It is not in any free list.
7277 // Add it to a free list or let it possibly be coalesced into
7278 // a larger chunk.
7279 HeapWord* const addr = (HeapWord*) fc;
7280 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7281
7282 // Verify that the bit map has no bits marked between
7283 // addr and purported end of just dead object.
7284 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7285 do_post_free_or_garbage_chunk(fc, size);
7286
7287 assert(_limit >= addr + size,
7288 "A freshly garbage chunk can't possibly straddle over _limit");
7289 if (inFreeRange()) lookahead_and_flush(fc, size);
7290 return size;
7291 }
7292
7293 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7294 HeapWord* addr = (HeapWord*) fc;
7295 // The sweeper has just found a live object. Return any accumulated
7296 // left hand chunk to the free lists.
7297 if (inFreeRange()) {
7298 assert(freeFinger() < addr, "freeFinger points too high");
7299 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7300 }
7301
7302 // This object is live: we'd normally expect this to be
7303 // an oop, and like to assert the following:
7304 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7305 // However, as we commented above, this may be an object whose
7306 // header hasn't yet been initialized.
7307 size_t size;
7308 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7309 if (_bitMap->isMarked(addr + 1)) {
7310 // Determine the size from the bit map, rather than trying to
7311 // compute it from the object header.
7312 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7313 size = pointer_delta(nextOneAddr + 1, addr);
7314 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7315 "alignment problem");
7316
7317 #ifdef ASSERT
7318 if (oop(addr)->klass_or_null_acquire() != NULL) {
7319 // Ignore mark word because we are running concurrent with mutators
7320 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7321 assert(size ==
7322 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7323 "P-mark and computed size do not agree");
7324 }
7325 #endif
7326
7327 } else {
7328 // This should be an initialized object that's alive.
7329 assert(oop(addr)->klass_or_null_acquire() != NULL,
7330 "Should be an initialized object");
7331 // Ignore mark word because we are running concurrent with mutators
7332 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7333 // Verify that the bit map has no bits marked between
7334 // addr and purported end of this block.
7335 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7336 assert(size >= 3, "Necessary for Printezis marks to work");
7337 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7338 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7339 }
7340 return size;
7341 }
7342
7343 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7344 size_t chunkSize) {
7345 // do_post_free_or_garbage_chunk() should only be called in the case
7346 // of the adaptive free list allocator.
7347 const bool fcInFreeLists = fc->is_free();
7348 assert((HeapWord*)fc <= _limit, "sweep invariant");
7349 if (CMSTestInFreeList && fcInFreeLists) {
7350 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7351 }
7352
7353 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7354
7355 HeapWord* const fc_addr = (HeapWord*) fc;
7356
7357 bool coalesce = false;
7358 const size_t left = pointer_delta(fc_addr, freeFinger());
7359 const size_t right = chunkSize;
7360 switch (FLSCoalescePolicy) {
7361 // numeric value forms a coalition aggressiveness metric
7362 case 0: { // never coalesce
7363 coalesce = false;
7364 break;
7365 }
7366 case 1: { // coalesce if left & right chunks on overpopulated lists
7367 coalesce = _sp->coalOverPopulated(left) &&
7368 _sp->coalOverPopulated(right);
7369 break;
7370 }
7371 case 2: { // coalesce if left chunk on overpopulated list (default)
7372 coalesce = _sp->coalOverPopulated(left);
7373 break;
7374 }
7375 case 3: { // coalesce if left OR right chunk on overpopulated list
7376 coalesce = _sp->coalOverPopulated(left) ||
7377 _sp->coalOverPopulated(right);
7378 break;
7379 }
7380 case 4: { // always coalesce
7381 coalesce = true;
7382 break;
7383 }
7384 default:
7385 ShouldNotReachHere();
7386 }
7387
7388 // Should the current free range be coalesced?
7389 // If the chunk is in a free range and either we decided to coalesce above
7390 // or the chunk is near the large block at the end of the heap
7391 // (isNearLargestChunk() returns true), then coalesce this chunk.
7392 const bool doCoalesce = inFreeRange()
7393 && (coalesce || _g->isNearLargestChunk(fc_addr));
7394 if (doCoalesce) {
7395 // Coalesce the current free range on the left with the new
7396 // chunk on the right. If either is on a free list,
7397 // it must be removed from the list and stashed in the closure.
7398 if (freeRangeInFreeLists()) {
7399 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7400 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7401 "Size of free range is inconsistent with chunk size.");
7402 if (CMSTestInFreeList) {
7403 assert(_sp->verify_chunk_in_free_list(ffc),
7404 "Chunk is not in free lists");
7405 }
7406 _sp->coalDeath(ffc->size());
7407 _sp->removeFreeChunkFromFreeLists(ffc);
7408 set_freeRangeInFreeLists(false);
7409 }
7410 if (fcInFreeLists) {
7411 _sp->coalDeath(chunkSize);
7412 assert(fc->size() == chunkSize,
7413 "The chunk has the wrong size or is not in the free lists");
7414 _sp->removeFreeChunkFromFreeLists(fc);
7415 }
7416 set_lastFreeRangeCoalesced(true);
7417 print_free_block_coalesced(fc);
7418 } else { // not in a free range and/or should not coalesce
7419 // Return the current free range and start a new one.
7420 if (inFreeRange()) {
7421 // In a free range but cannot coalesce with the right hand chunk.
7422 // Put the current free range into the free lists.
7423 flush_cur_free_chunk(freeFinger(),
7424 pointer_delta(fc_addr, freeFinger()));
7425 }
7426 // Set up for new free range. Pass along whether the right hand
7427 // chunk is in the free lists.
7428 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7429 }
7430 }
7431
7432 // Lookahead flush:
7433 // If we are tracking a free range, and this is the last chunk that
7434 // we'll look at because its end crosses past _limit, we'll preemptively
7435 // flush it along with any free range we may be holding on to. Note that
7436 // this can be the case only for an already free or freshly garbage
7437 // chunk. If this block is an object, it can never straddle
7438 // over _limit. The "straddling" occurs when _limit is set at
7439 // the previous end of the space when this cycle started, and
7440 // a subsequent heap expansion caused the previously co-terminal
7441 // free block to be coalesced with the newly expanded portion,
7442 // thus rendering _limit a non-block-boundary making it dangerous
7443 // for the sweeper to step over and examine.
7444 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7445 assert(inFreeRange(), "Should only be called if currently in a free range.");
7446 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7447 assert(_sp->used_region().contains(eob - 1),
7448 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7449 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7450 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7451 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7452 if (eob >= _limit) {
7453 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7454 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7455 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7456 "[" PTR_FORMAT "," PTR_FORMAT ")",
7457 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7458 // Return the storage we are tracking back into the free lists.
7459 log_develop_trace(gc, sweep)("Flushing ... ");
7460 assert(freeFinger() < eob, "Error");
7461 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7462 }
7463 }
7464
7465 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7466 assert(inFreeRange(), "Should only be called if currently in a free range.");
7467 assert(size > 0,
7468 "A zero sized chunk cannot be added to the free lists.");
7469 if (!freeRangeInFreeLists()) {
7470 if (CMSTestInFreeList) {
7471 FreeChunk* fc = (FreeChunk*) chunk;
7472 fc->set_size(size);
7473 assert(!_sp->verify_chunk_in_free_list(fc),
7474 "chunk should not be in free lists yet");
7475 }
7476 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7477 // A new free range is going to be starting. The current
7478 // free range has not been added to the free lists yet or
7479 // was removed so add it back.
7480 // If the current free range was coalesced, then the death
7481 // of the free range was recorded. Record a birth now.
7482 if (lastFreeRangeCoalesced()) {
7483 _sp->coalBirth(size);
7484 }
7485 _sp->addChunkAndRepairOffsetTable(chunk, size,
7486 lastFreeRangeCoalesced());
7487 } else {
7488 log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7489 }
7490 set_inFreeRange(false);
7491 set_freeRangeInFreeLists(false);
7492 }
7493
7494 // We take a break if we've been at this for a while,
7495 // so as to avoid monopolizing the locks involved.
7496 void SweepClosure::do_yield_work(HeapWord* addr) {
7497 // Return current free chunk being used for coalescing (if any)
7498 // to the appropriate freelist. After yielding, the next
7499 // free block encountered will start a coalescing range of
7500 // free blocks. If the next free block is adjacent to the
7501 // chunk just flushed, they will need to wait for the next
7502 // sweep to be coalesced.
7503 if (inFreeRange()) {
7504 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7505 }
7506
7507 // First give up the locks, then yield, then re-lock.
7508 // We should probably use a constructor/destructor idiom to
7509 // do this unlock/lock or modify the MutexUnlocker class to
7510 // serve our purpose. XXX
7511 assert_lock_strong(_bitMap->lock());
7512 assert_lock_strong(_freelistLock);
7513 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7514 "CMS thread should hold CMS token");
7515 _bitMap->lock()->unlock();
7516 _freelistLock->unlock();
7517 ConcurrentMarkSweepThread::desynchronize(true);
7518 _collector->stopTimer();
7519 _collector->incrementYields();
7520
7521 // See the comment in coordinator_yield()
7522 for (unsigned i = 0; i < CMSYieldSleepCount &&
7523 ConcurrentMarkSweepThread::should_yield() &&
7524 !CMSCollector::foregroundGCIsActive(); ++i) {
7525 os::sleep(Thread::current(), 1, false);
7526 }
7527
7528 ConcurrentMarkSweepThread::synchronize(true);
7529 _freelistLock->lock();
7530 _bitMap->lock()->lock_without_safepoint_check();
7531 _collector->startTimer();
7532 }
7533
7534 #ifndef PRODUCT
7535 // This is actually very useful in a product build if it can
7536 // be called from the debugger. Compile it into the product
7537 // as needed.
7538 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7539 return debug_cms_space->verify_chunk_in_free_list(fc);
7540 }
7541 #endif
7542
7543 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7544 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7545 p2i(fc), fc->size());
7546 }
7547
7548 // CMSIsAliveClosure
7549 bool CMSIsAliveClosure::do_object_b(oop obj) {
7550 HeapWord* addr = (HeapWord*)obj;
7551 return addr != NULL &&
7552 (!_span.contains(addr) || _bit_map->isMarked(addr));
7553 }
7554
7555
7556 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7557 MemRegion span,
7558 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7559 bool cpc):
7560 _collector(collector),
7561 _span(span),
7562 _bit_map(bit_map),
7563 _mark_stack(mark_stack),
7564 _concurrent_precleaning(cpc) {
7565 assert(!_span.is_empty(), "Empty span could spell trouble");
7566 }
7567
7568
7569 // CMSKeepAliveClosure: the serial version
7570 void CMSKeepAliveClosure::do_oop(oop obj) {
7571 HeapWord* addr = (HeapWord*)obj;
7572 if (_span.contains(addr) &&
7573 !_bit_map->isMarked(addr)) {
7574 _bit_map->mark(addr);
7575 bool simulate_overflow = false;
7576 NOT_PRODUCT(
7577 if (CMSMarkStackOverflowALot &&
7578 _collector->simulate_overflow()) {
7579 // simulate a stack overflow
7580 simulate_overflow = true;
7581 }
7582 )
7583 if (simulate_overflow || !_mark_stack->push(obj)) {
7584 if (_concurrent_precleaning) {
7585 // We dirty the overflown object and let the remark
7586 // phase deal with it.
7587 assert(_collector->overflow_list_is_empty(), "Error");
7588 // In the case of object arrays, we need to dirty all of
7589 // the cards that the object spans. No locking or atomics
7590 // are needed since no one else can be mutating the mod union
7591 // table.
7592 if (obj->is_objArray()) {
7593 size_t sz = obj->size();
7594 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
7595 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7596 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7597 _collector->_modUnionTable.mark_range(redirty_range);
7598 } else {
7599 _collector->_modUnionTable.mark(addr);
7600 }
7601 _collector->_ser_kac_preclean_ovflw++;
7602 } else {
7603 _collector->push_on_overflow_list(obj);
7604 _collector->_ser_kac_ovflw++;
7605 }
7606 }
7607 }
7608 }
7609
7610 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7611 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7612
7613 // CMSParKeepAliveClosure: a parallel version of the above.
7614 // The work queues are private to each closure (thread),
7615 // but (may be) available for stealing by other threads.
7616 void CMSParKeepAliveClosure::do_oop(oop obj) {
7617 HeapWord* addr = (HeapWord*)obj;
7618 if (_span.contains(addr) &&
7619 !_bit_map->isMarked(addr)) {
7620 // In general, during recursive tracing, several threads
7621 // may be concurrently getting here; the first one to
7622 // "tag" it, claims it.
7623 if (_bit_map->par_mark(addr)) {
7624 bool res = _work_queue->push(obj);
7625 assert(res, "Low water mark should be much less than capacity");
7626 // Do a recursive trim in the hope that this will keep
7627 // stack usage lower, but leave some oops for potential stealers
7628 trim_queue(_low_water_mark);
7629 } // Else, another thread got there first
7630 }
7631 }
7632
7633 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7634 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7635
7636 void CMSParKeepAliveClosure::trim_queue(uint max) {
7637 while (_work_queue->size() > max) {
7638 oop new_oop;
7639 if (_work_queue->pop_local(new_oop)) {
7640 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7641 assert(_bit_map->isMarked((HeapWord*)new_oop),
7642 "no white objects on this stack!");
7643 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7644 // iterate over the oops in this oop, marking and pushing
7645 // the ones in CMS heap (i.e. in _span).
7646 new_oop->oop_iterate(&_mark_and_push);
7647 }
7648 }
7649 }
7650
7651 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7652 CMSCollector* collector,
7653 MemRegion span, CMSBitMap* bit_map,
7654 OopTaskQueue* work_queue):
7655 _collector(collector),
7656 _span(span),
7657 _bit_map(bit_map),
7658 _work_queue(work_queue) { }
7659
7660 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7661 HeapWord* addr = (HeapWord*)obj;
7662 if (_span.contains(addr) &&
7663 !_bit_map->isMarked(addr)) {
7664 if (_bit_map->par_mark(addr)) {
7665 bool simulate_overflow = false;
7666 NOT_PRODUCT(
7667 if (CMSMarkStackOverflowALot &&
7668 _collector->par_simulate_overflow()) {
7669 // simulate a stack overflow
7670 simulate_overflow = true;
7671 }
7672 )
7673 if (simulate_overflow || !_work_queue->push(obj)) {
7674 _collector->par_push_on_overflow_list(obj);
7675 _collector->_par_kac_ovflw++;
7676 }
7677 } // Else another thread got there already
7678 }
7679 }
7680
7681 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7682 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7683
7684 //////////////////////////////////////////////////////////////////
7685 // CMSExpansionCause /////////////////////////////
7686 //////////////////////////////////////////////////////////////////
7687 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7688 switch (cause) {
7689 case _no_expansion:
7690 return "No expansion";
7691 case _satisfy_free_ratio:
7692 return "Free ratio";
7693 case _satisfy_promotion:
7694 return "Satisfy promotion";
7695 case _satisfy_allocation:
7696 return "allocation";
7697 case _allocate_par_lab:
7698 return "Par LAB";
7699 case _allocate_par_spooling_space:
7700 return "Par Spooling Space";
7701 case _adaptive_size_policy:
7702 return "Ergonomics";
7703 default:
7704 return "unknown";
7705 }
7706 }
7707
7708 void CMSDrainMarkingStackClosure::do_void() {
7709 // the max number to take from overflow list at a time
7710 const size_t num = _mark_stack->capacity()/4;
7711 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7712 "Overflow list should be NULL during concurrent phases");
7713 while (!_mark_stack->isEmpty() ||
7714 // if stack is empty, check the overflow list
7715 _collector->take_from_overflow_list(num, _mark_stack)) {
7716 oop obj = _mark_stack->pop();
7717 HeapWord* addr = (HeapWord*)obj;
7718 assert(_span.contains(addr), "Should be within span");
7719 assert(_bit_map->isMarked(addr), "Should be marked");
7720 assert(oopDesc::is_oop(obj), "Should be an oop");
7721 obj->oop_iterate(_keep_alive);
7722 }
7723 }
7724
7725 void CMSParDrainMarkingStackClosure::do_void() {
7726 // drain queue
7727 trim_queue(0);
7728 }
7729
7730 // Trim our work_queue so its length is below max at return
7731 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7732 while (_work_queue->size() > max) {
7733 oop new_oop;
7734 if (_work_queue->pop_local(new_oop)) {
7735 assert(oopDesc::is_oop(new_oop), "Expected an oop");
7736 assert(_bit_map->isMarked((HeapWord*)new_oop),
7737 "no white objects on this stack!");
7738 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7739 // iterate over the oops in this oop, marking and pushing
7740 // the ones in CMS heap (i.e. in _span).
7741 new_oop->oop_iterate(&_mark_and_push);
7742 }
7743 }
7744 }
7745
7746 ////////////////////////////////////////////////////////////////////
7747 // Support for Marking Stack Overflow list handling and related code
7748 ////////////////////////////////////////////////////////////////////
7749 // Much of the following code is similar in shape and spirit to the
7750 // code used in ParNewGC. We should try and share that code
7751 // as much as possible in the future.
7752
7753 #ifndef PRODUCT
7754 // Debugging support for CMSStackOverflowALot
7755
7756 // It's OK to call this multi-threaded; the worst thing
7757 // that can happen is that we'll get a bunch of closely
7758 // spaced simulated overflows, but that's OK, in fact
7759 // probably good as it would exercise the overflow code
7760 // under contention.
7761 bool CMSCollector::simulate_overflow() {
7762 if (_overflow_counter-- <= 0) { // just being defensive
7763 _overflow_counter = CMSMarkStackOverflowInterval;
7764 return true;
7765 } else {
7766 return false;
7767 }
7768 }
7769
7770 bool CMSCollector::par_simulate_overflow() {
7771 return simulate_overflow();
7772 }
7773 #endif
7774
7775 // Single-threaded
7776 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7777 assert(stack->isEmpty(), "Expected precondition");
7778 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7779 size_t i = num;
7780 oop cur = _overflow_list;
7781 const markOop proto = markOopDesc::prototype();
7782 NOT_PRODUCT(ssize_t n = 0;)
7783 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7784 next = oop(cur->mark_raw());
7785 cur->set_mark_raw(proto); // until proven otherwise
7786 assert(oopDesc::is_oop(cur), "Should be an oop");
7787 bool res = stack->push(cur);
7788 assert(res, "Bit off more than can chew?");
7789 NOT_PRODUCT(n++;)
7790 }
7791 _overflow_list = cur;
7792 #ifndef PRODUCT
7793 assert(_num_par_pushes >= n, "Too many pops?");
7794 _num_par_pushes -=n;
7795 #endif
7796 return !stack->isEmpty();
7797 }
7798
7799 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
7800 // (MT-safe) Get a prefix of at most "num" from the list.
7801 // The overflow list is chained through the mark word of
7802 // each object in the list. We fetch the entire list,
7803 // break off a prefix of the right size and return the
7804 // remainder. If other threads try to take objects from
7805 // the overflow list at that time, they will wait for
7806 // some time to see if data becomes available. If (and
7807 // only if) another thread places one or more object(s)
7808 // on the global list before we have returned the suffix
7809 // to the global list, we will walk down our local list
7810 // to find its end and append the global list to
7811 // our suffix before returning it. This suffix walk can
7812 // prove to be expensive (quadratic in the amount of traffic)
7813 // when there are many objects in the overflow list and
7814 // there is much producer-consumer contention on the list.
7815 // *NOTE*: The overflow list manipulation code here and
7816 // in ParNewGeneration:: are very similar in shape,
7817 // except that in the ParNew case we use the old (from/eden)
7818 // copy of the object to thread the list via its klass word.
7819 // Because of the common code, if you make any changes in
7820 // the code below, please check the ParNew version to see if
7821 // similar changes might be needed.
7822 // CR 6797058 has been filed to consolidate the common code.
7823 bool CMSCollector::par_take_from_overflow_list(size_t num,
7824 OopTaskQueue* work_q,
7825 int no_of_gc_threads) {
7826 assert(work_q->size() == 0, "First empty local work queue");
7827 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7828 if (_overflow_list == NULL) {
7829 return false;
7830 }
7831 // Grab the entire list; we'll put back a suffix
7832 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7833 Thread* tid = Thread::current();
7834 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7835 // set to ParallelGCThreads.
7836 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7837 size_t sleep_time_millis = MAX2((size_t)1, num/100);
7838 // If the list is busy, we spin for a short while,
7839 // sleeping between attempts to get the list.
7840 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7841 os::sleep(tid, sleep_time_millis, false);
7842 if (_overflow_list == NULL) {
7843 // Nothing left to take
7844 return false;
7845 } else if (_overflow_list != BUSY) {
7846 // Try and grab the prefix
7847 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7848 }
7849 }
7850 // If the list was found to be empty, or we spun long
7851 // enough, we give up and return empty-handed. If we leave
7852 // the list in the BUSY state below, it must be the case that
7853 // some other thread holds the overflow list and will set it
7854 // to a non-BUSY state in the future.
7855 if (prefix == NULL || prefix == BUSY) {
7856 // Nothing to take or waited long enough
7857 if (prefix == NULL) {
7858 // Write back the NULL in case we overwrote it with BUSY above
7859 // and it is still the same value.
7860 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7861 }
7862 return false;
7863 }
7864 assert(prefix != NULL && prefix != BUSY, "Error");
7865 size_t i = num;
7866 oop cur = prefix;
7867 // Walk down the first "num" objects, unless we reach the end.
7868 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--);
7869 if (cur->mark_raw() == NULL) {
7870 // We have "num" or fewer elements in the list, so there
7871 // is nothing to return to the global list.
7872 // Write back the NULL in lieu of the BUSY we wrote
7873 // above, if it is still the same value.
7874 if (_overflow_list == BUSY) {
7875 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7876 }
7877 } else {
7878 // Chop off the suffix and return it to the global list.
7879 assert(cur->mark_raw() != BUSY, "Error");
7880 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list
7881 cur->set_mark_raw(NULL); // break off suffix
7882 // It's possible that the list is still in the empty(busy) state
7883 // we left it in a short while ago; in that case we may be
7884 // able to place back the suffix without incurring the cost
7885 // of a walk down the list.
7886 oop observed_overflow_list = _overflow_list;
7887 oop cur_overflow_list = observed_overflow_list;
7888 bool attached = false;
7889 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7890 observed_overflow_list =
7891 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7892 if (cur_overflow_list == observed_overflow_list) {
7893 attached = true;
7894 break;
7895 } else cur_overflow_list = observed_overflow_list;
7896 }
7897 if (!attached) {
7898 // Too bad, someone else sneaked in (at least) an element; we'll need
7899 // to do a splice. Find tail of suffix so we can prepend suffix to global
7900 // list.
7901 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw()));
7902 oop suffix_tail = cur;
7903 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL,
7904 "Tautology");
7905 observed_overflow_list = _overflow_list;
7906 do {
7907 cur_overflow_list = observed_overflow_list;
7908 if (cur_overflow_list != BUSY) {
7909 // Do the splice ...
7910 suffix_tail->set_mark_raw(markOop(cur_overflow_list));
7911 } else { // cur_overflow_list == BUSY
7912 suffix_tail->set_mark_raw(NULL);
7913 }
7914 // ... and try to place spliced list back on overflow_list ...
7915 observed_overflow_list =
7916 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7917 } while (cur_overflow_list != observed_overflow_list);
7918 // ... until we have succeeded in doing so.
7919 }
7920 }
7921
7922 // Push the prefix elements on work_q
7923 assert(prefix != NULL, "control point invariant");
7924 const markOop proto = markOopDesc::prototype();
7925 oop next;
7926 NOT_PRODUCT(ssize_t n = 0;)
7927 for (cur = prefix; cur != NULL; cur = next) {
7928 next = oop(cur->mark_raw());
7929 cur->set_mark_raw(proto); // until proven otherwise
7930 assert(oopDesc::is_oop(cur), "Should be an oop");
7931 bool res = work_q->push(cur);
7932 assert(res, "Bit off more than we can chew?");
7933 NOT_PRODUCT(n++;)
7934 }
7935 #ifndef PRODUCT
7936 assert(_num_par_pushes >= n, "Too many pops?");
7937 Atomic::sub(n, &_num_par_pushes);
7938 #endif
7939 return true;
7940 }
7941
7942 // Single-threaded
7943 void CMSCollector::push_on_overflow_list(oop p) {
7944 NOT_PRODUCT(_num_par_pushes++;)
7945 assert(oopDesc::is_oop(p), "Not an oop");
7946 preserve_mark_if_necessary(p);
7947 p->set_mark_raw((markOop)_overflow_list);
7948 _overflow_list = p;
7949 }
7950
7951 // Multi-threaded; use CAS to prepend to overflow list
7952 void CMSCollector::par_push_on_overflow_list(oop p) {
7953 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);)
7954 assert(oopDesc::is_oop(p), "Not an oop");
7955 par_preserve_mark_if_necessary(p);
7956 oop observed_overflow_list = _overflow_list;
7957 oop cur_overflow_list;
7958 do {
7959 cur_overflow_list = observed_overflow_list;
7960 if (cur_overflow_list != BUSY) {
7961 p->set_mark_raw(markOop(cur_overflow_list));
7962 } else {
7963 p->set_mark_raw(NULL);
7964 }
7965 observed_overflow_list =
7966 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list);
7967 } while (cur_overflow_list != observed_overflow_list);
7968 }
7969 #undef BUSY
7970
7971 // Single threaded
7972 // General Note on GrowableArray: pushes may silently fail
7973 // because we are (temporarily) out of C-heap for expanding
7974 // the stack. The problem is quite ubiquitous and affects
7975 // a lot of code in the JVM. The prudent thing for GrowableArray
7976 // to do (for now) is to exit with an error. However, that may
7977 // be too draconian in some cases because the caller may be
7978 // able to recover without much harm. For such cases, we
7979 // should probably introduce a "soft_push" method which returns
7980 // an indication of success or failure with the assumption that
7981 // the caller may be able to recover from a failure; code in
7982 // the VM can then be changed, incrementally, to deal with such
7983 // failures where possible, thus, incrementally hardening the VM
7984 // in such low resource situations.
7985 void CMSCollector::preserve_mark_work(oop p, markOop m) {
7986 _preserved_oop_stack.push(p);
7987 _preserved_mark_stack.push(m);
7988 assert(m == p->mark_raw(), "Mark word changed");
7989 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
7990 "bijection");
7991 }
7992
7993 // Single threaded
7994 void CMSCollector::preserve_mark_if_necessary(oop p) {
7995 markOop m = p->mark_raw();
7996 if (m->must_be_preserved(p)) {
7997 preserve_mark_work(p, m);
7998 }
7999 }
8000
8001 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8002 markOop m = p->mark_raw();
8003 if (m->must_be_preserved(p)) {
8004 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8005 // Even though we read the mark word without holding
8006 // the lock, we are assured that it will not change
8007 // because we "own" this oop, so no other thread can
8008 // be trying to push it on the overflow list; see
8009 // the assertion in preserve_mark_work() that checks
8010 // that m == p->mark_raw().
8011 preserve_mark_work(p, m);
8012 }
8013 }
8014
8015 // We should be able to do this multi-threaded,
8016 // a chunk of stack being a task (this is
8017 // correct because each oop only ever appears
8018 // once in the overflow list. However, it's
8019 // not very easy to completely overlap this with
8020 // other operations, so will generally not be done
8021 // until all work's been completed. Because we
8022 // expect the preserved oop stack (set) to be small,
8023 // it's probably fine to do this single-threaded.
8024 // We can explore cleverer concurrent/overlapped/parallel
8025 // processing of preserved marks if we feel the
8026 // need for this in the future. Stack overflow should
8027 // be so rare in practice and, when it happens, its
8028 // effect on performance so great that this will
8029 // likely just be in the noise anyway.
8030 void CMSCollector::restore_preserved_marks_if_any() {
8031 assert(SafepointSynchronize::is_at_safepoint(),
8032 "world should be stopped");
8033 assert(Thread::current()->is_ConcurrentGC_thread() ||
8034 Thread::current()->is_VM_thread(),
8035 "should be single-threaded");
8036 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8037 "bijection");
8038
8039 while (!_preserved_oop_stack.is_empty()) {
8040 oop p = _preserved_oop_stack.pop();
8041 assert(oopDesc::is_oop(p), "Should be an oop");
8042 assert(_span.contains(p), "oop should be in _span");
8043 assert(p->mark_raw() == markOopDesc::prototype(),
8044 "Set when taken from overflow list");
8045 markOop m = _preserved_mark_stack.pop();
8046 p->set_mark_raw(m);
8047 }
8048 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8049 "stacks were cleared above");
8050 }
8051
8052 #ifndef PRODUCT
8053 bool CMSCollector::no_preserved_marks() const {
8054 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8055 }
8056 #endif
8057
8058 // Transfer some number of overflown objects to usual marking
8059 // stack. Return true if some objects were transferred.
8060 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8061 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8062 (size_t)ParGCDesiredObjsFromOverflowList);
8063
8064 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8065 assert(_collector->overflow_list_is_empty() || res,
8066 "If list is not empty, we should have taken something");
8067 assert(!res || !_mark_stack->isEmpty(),
8068 "If we took something, it should now be on our stack");
8069 return res;
8070 }
8071
8072 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8073 size_t res = _sp->block_size_no_stall(addr, _collector);
8074 if (_sp->block_is_obj(addr)) {
8075 if (_live_bit_map->isMarked(addr)) {
8076 // It can't have been dead in a previous cycle
8077 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8078 } else {
8079 _dead_bit_map->mark(addr); // mark the dead object
8080 }
8081 }
8082 // Could be 0, if the block size could not be computed without stalling.
8083 return res;
8084 }
8085
8086 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8087 GCMemoryManager* manager = CMSHeap::heap()->old_manager();
8088 switch (phase) {
8089 case CMSCollector::InitialMarking:
8090 initialize(manager /* GC manager */ ,
8091 cause /* cause of the GC */,
8092 true /* recordGCBeginTime */,
8093 true /* recordPreGCUsage */,
8094 false /* recordPeakUsage */,
8095 false /* recordPostGCusage */,
8096 true /* recordAccumulatedGCTime */,
8097 false /* recordGCEndTime */,
8098 false /* countCollection */ );
8099 break;
8100
8101 case CMSCollector::FinalMarking:
8102 initialize(manager /* GC manager */ ,
8103 cause /* cause of the GC */,
8104 false /* recordGCBeginTime */,
8105 false /* recordPreGCUsage */,
8106 false /* recordPeakUsage */,
8107 false /* recordPostGCusage */,
8108 true /* recordAccumulatedGCTime */,
8109 false /* recordGCEndTime */,
8110 false /* countCollection */ );
8111 break;
8112
8113 case CMSCollector::Sweeping:
8114 initialize(manager /* GC manager */ ,
8115 cause /* cause of the GC */,
8116 false /* recordGCBeginTime */,
8117 false /* recordPreGCUsage */,
8118 true /* recordPeakUsage */,
8119 true /* recordPostGCusage */,
8120 false /* recordAccumulatedGCTime */,
8121 true /* recordGCEndTime */,
8122 true /* countCollection */ );
8123 break;
8124
8125 default:
8126 ShouldNotReachHere();
8127 }
8128 }