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/vmCMSOperations.hpp"
40 #include "gc/serial/genMarkSweep.hpp"
41 #include "gc/serial/tenuredGeneration.hpp"
42 #include "gc/shared/adaptiveSizePolicy.hpp"
43 #include "gc/shared/cardGeneration.inline.hpp"
44 #include "gc/shared/cardTableRS.hpp"
45 #include "gc/shared/collectedHeap.inline.hpp"
46 #include "gc/shared/collectorCounters.hpp"
47 #include "gc/shared/collectorPolicy.hpp"
48 #include "gc/shared/gcLocker.hpp"
49 #include "gc/shared/gcPolicyCounters.hpp"
50 #include "gc/shared/gcTimer.hpp"
51 #include "gc/shared/gcTrace.hpp"
52 #include "gc/shared/gcTraceTime.inline.hpp"
53 #include "gc/shared/genCollectedHeap.hpp"
54 #include "gc/shared/genOopClosures.inline.hpp"
55 #include "gc/shared/isGCActiveMark.hpp"
56 #include "gc/shared/referencePolicy.hpp"
57 #include "gc/shared/strongRootsScope.hpp"
58 #include "gc/shared/taskqueue.inline.hpp"
59 #include "gc/shared/weakProcessor.hpp"
60 #include "logging/log.hpp"
61 #include "logging/logStream.hpp"
62 #include "memory/allocation.hpp"
63 #include "memory/binaryTreeDictionary.inline.hpp"
64 #include "memory/iterator.inline.hpp"
65 #include "memory/padded.hpp"
66 #include "memory/resourceArea.hpp"
67 #include "oops/access.inline.hpp"
68 #include "oops/oop.inline.hpp"
69 #include "prims/jvmtiExport.hpp"
70 #include "runtime/atomic.hpp"
71 #include "runtime/flags/flagSetting.hpp"
72 #include "runtime/globals_extension.hpp"
73 #include "runtime/handles.inline.hpp"
74 #include "runtime/java.hpp"
75 #include "runtime/orderAccess.inline.hpp"
76 #include "runtime/timer.hpp"
77 #include "runtime/vmThread.hpp"
78 #include "services/memoryService.hpp"
79 #include "services/runtimeService.hpp"
80 #include "utilities/align.hpp"
81 #include "utilities/stack.inline.hpp"
82
83 // statics
84 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
85 bool CMSCollector::_full_gc_requested = false;
86 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
87
88 //////////////////////////////////////////////////////////////////
89 // In support of CMS/VM thread synchronization
90 //////////////////////////////////////////////////////////////////
91 // We split use of the CGC_lock into 2 "levels".
92 // The low-level locking is of the usual CGC_lock monitor. We introduce
93 // a higher level "token" (hereafter "CMS token") built on top of the
94 // low level monitor (hereafter "CGC lock").
95 // The token-passing protocol gives priority to the VM thread. The
96 // CMS-lock doesn't provide any fairness guarantees, but clients
97 // should ensure that it is only held for very short, bounded
98 // durations.
99 //
100 // When either of the CMS thread or the VM thread is involved in
101 // collection operations during which it does not want the other
102 // thread to interfere, it obtains the CMS token.
103 //
104 // If either thread tries to get the token while the other has
105 // it, that thread waits. However, if the VM thread and CMS thread
106 // both want the token, then the VM thread gets priority while the
107 // CMS thread waits. This ensures, for instance, that the "concurrent"
108 // phases of the CMS thread's work do not block out the VM thread
109 // for long periods of time as the CMS thread continues to hog
110 // the token. (See bug 4616232).
111 //
112 // The baton-passing functions are, however, controlled by the
113 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
114 // and here the low-level CMS lock, not the high level token,
115 // ensures mutual exclusion.
116 //
117 // Two important conditions that we have to satisfy:
118 // 1. if a thread does a low-level wait on the CMS lock, then it
119 // relinquishes the CMS token if it were holding that token
120 // when it acquired the low-level CMS lock.
121 // 2. any low-level notifications on the low-level lock
122 // should only be sent when a thread has relinquished the token.
123 //
124 // In the absence of either property, we'd have potential deadlock.
125 //
126 // We protect each of the CMS (concurrent and sequential) phases
127 // with the CMS _token_, not the CMS _lock_.
128 //
129 // The only code protected by CMS lock is the token acquisition code
130 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
131 // baton-passing code.
132 //
133 // Unfortunately, i couldn't come up with a good abstraction to factor and
134 // hide the naked CGC_lock manipulation in the baton-passing code
135 // further below. That's something we should try to do. Also, the proof
136 // of correctness of this 2-level locking scheme is far from obvious,
137 // and potentially quite slippery. We have an uneasy suspicion, for instance,
138 // that there may be a theoretical possibility of delay/starvation in the
139 // low-level lock/wait/notify scheme used for the baton-passing because of
140 // potential interference with the priority scheme embodied in the
141 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
142 // invocation further below and marked with "XXX 20011219YSR".
143 // Indeed, as we note elsewhere, this may become yet more slippery
144 // in the presence of multiple CMS and/or multiple VM threads. XXX
145
146 class CMSTokenSync: public StackObj {
147 private:
148 bool _is_cms_thread;
149 public:
150 CMSTokenSync(bool is_cms_thread):
151 _is_cms_thread(is_cms_thread) {
152 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
153 "Incorrect argument to constructor");
154 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
155 }
156
157 ~CMSTokenSync() {
158 assert(_is_cms_thread ?
159 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
160 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
161 "Incorrect state");
162 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
163 }
164 };
165
166 // Convenience class that does a CMSTokenSync, and then acquires
167 // upto three locks.
168 class CMSTokenSyncWithLocks: public CMSTokenSync {
169 private:
170 // Note: locks are acquired in textual declaration order
171 // and released in the opposite order
172 MutexLockerEx _locker1, _locker2, _locker3;
173 public:
174 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
175 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
176 CMSTokenSync(is_cms_thread),
177 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
178 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
179 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
180 { }
181 };
182
183
184 //////////////////////////////////////////////////////////////////
185 // Concurrent Mark-Sweep Generation /////////////////////////////
186 //////////////////////////////////////////////////////////////////
187
188 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
189
190 // This struct contains per-thread things necessary to support parallel
191 // young-gen collection.
192 class CMSParGCThreadState: public CHeapObj<mtGC> {
193 public:
194 CompactibleFreeListSpaceLAB lab;
195 PromotionInfo promo;
196
197 // Constructor.
198 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
199 promo.setSpace(cfls);
200 }
201 };
202
203 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
204 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
205 CardGeneration(rs, initial_byte_size, ct),
206 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
207 _did_compact(false)
208 {
209 HeapWord* bottom = (HeapWord*) _virtual_space.low();
210 HeapWord* end = (HeapWord*) _virtual_space.high();
211
212 _direct_allocated_words = 0;
213 NOT_PRODUCT(
214 _numObjectsPromoted = 0;
215 _numWordsPromoted = 0;
216 _numObjectsAllocated = 0;
217 _numWordsAllocated = 0;
218 )
219
220 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
221 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
222 _cmsSpace->_old_gen = this;
223
224 _gc_stats = new CMSGCStats();
225
226 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
227 // offsets match. The ability to tell free chunks from objects
228 // depends on this property.
229 debug_only(
230 FreeChunk* junk = NULL;
231 assert(UseCompressedClassPointers ||
232 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
233 "Offset of FreeChunk::_prev within FreeChunk must match"
234 " that of OopDesc::_klass within OopDesc");
235 )
236
237 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
238 for (uint i = 0; i < ParallelGCThreads; i++) {
239 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
240 }
241
242 _incremental_collection_failed = false;
243 // The "dilatation_factor" is the expansion that can occur on
244 // account of the fact that the minimum object size in the CMS
245 // generation may be larger than that in, say, a contiguous young
246 // generation.
247 // Ideally, in the calculation below, we'd compute the dilatation
248 // factor as: MinChunkSize/(promoting_gen's min object size)
249 // Since we do not have such a general query interface for the
250 // promoting generation, we'll instead just use the minimum
251 // object size (which today is a header's worth of space);
252 // note that all arithmetic is in units of HeapWords.
253 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
254 assert(_dilatation_factor >= 1.0, "from previous assert");
255 }
256
257
258 // The field "_initiating_occupancy" represents the occupancy percentage
259 // at which we trigger a new collection cycle. Unless explicitly specified
260 // via CMSInitiatingOccupancyFraction (argument "io" below), it
261 // is calculated by:
262 //
263 // Let "f" be MinHeapFreeRatio in
264 //
265 // _initiating_occupancy = 100-f +
266 // f * (CMSTriggerRatio/100)
267 // where CMSTriggerRatio is the argument "tr" below.
268 //
269 // That is, if we assume the heap is at its desired maximum occupancy at the
270 // end of a collection, we let CMSTriggerRatio of the (purported) free
271 // space be allocated before initiating a new collection cycle.
272 //
273 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
274 assert(io <= 100 && tr <= 100, "Check the arguments");
275 if (io >= 0) {
276 _initiating_occupancy = (double)io / 100.0;
277 } else {
278 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
279 (double)(tr * MinHeapFreeRatio) / 100.0)
280 / 100.0;
281 }
282 }
283
284 void ConcurrentMarkSweepGeneration::ref_processor_init() {
285 assert(collector() != NULL, "no collector");
286 collector()->ref_processor_init();
287 }
288
289 void CMSCollector::ref_processor_init() {
290 if (_ref_processor == NULL) {
291 // Allocate and initialize a reference processor
292 _ref_processor =
293 new ReferenceProcessor(&_span_discoverer, // span
294 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
295 ParallelGCThreads, // mt processing degree
296 _cmsGen->refs_discovery_is_mt(), // mt discovery
297 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
298 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
299 &_is_alive_closure); // closure for liveness info
300 // Initialize the _ref_processor field of CMSGen
301 _cmsGen->set_ref_processor(_ref_processor);
302
303 }
304 }
305
306 AdaptiveSizePolicy* CMSCollector::size_policy() {
307 return CMSHeap::heap()->size_policy();
308 }
309
310 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
311
312 const char* gen_name = "old";
313 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy();
314 // Generation Counters - generation 1, 1 subspace
315 _gen_counters = new GenerationCounters(gen_name, 1, 1,
316 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
317
318 _space_counters = new GSpaceCounters(gen_name, 0,
319 _virtual_space.reserved_size(),
320 this, _gen_counters);
321 }
322
323 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
324 _cms_gen(cms_gen)
325 {
326 assert(alpha <= 100, "bad value");
327 _saved_alpha = alpha;
328
329 // Initialize the alphas to the bootstrap value of 100.
330 _gc0_alpha = _cms_alpha = 100;
331
332 _cms_begin_time.update();
333 _cms_end_time.update();
334
335 _gc0_duration = 0.0;
336 _gc0_period = 0.0;
337 _gc0_promoted = 0;
338
339 _cms_duration = 0.0;
340 _cms_period = 0.0;
341 _cms_allocated = 0;
342
343 _cms_used_at_gc0_begin = 0;
344 _cms_used_at_gc0_end = 0;
345 _allow_duty_cycle_reduction = false;
346 _valid_bits = 0;
347 }
348
349 double CMSStats::cms_free_adjustment_factor(size_t free) const {
350 // TBD: CR 6909490
351 return 1.0;
352 }
353
354 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
355 }
356
357 // If promotion failure handling is on use
358 // the padded average size of the promotion for each
359 // young generation collection.
360 double CMSStats::time_until_cms_gen_full() const {
361 size_t cms_free = _cms_gen->cmsSpace()->free();
362 CMSHeap* heap = CMSHeap::heap();
363 size_t expected_promotion = MIN2(heap->young_gen()->capacity(),
364 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
365 if (cms_free > expected_promotion) {
366 // Start a cms collection if there isn't enough space to promote
367 // for the next young collection. Use the padded average as
368 // a safety factor.
369 cms_free -= expected_promotion;
370
371 // Adjust by the safety factor.
372 double cms_free_dbl = (double)cms_free;
373 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
374 // Apply a further correction factor which tries to adjust
375 // for recent occurance of concurrent mode failures.
376 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
377 cms_free_dbl = cms_free_dbl * cms_adjustment;
378
379 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
380 cms_free, expected_promotion);
381 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
382 // Add 1 in case the consumption rate goes to zero.
383 return cms_free_dbl / (cms_consumption_rate() + 1.0);
384 }
385 return 0.0;
386 }
387
388 // Compare the duration of the cms collection to the
389 // time remaining before the cms generation is empty.
390 // Note that the time from the start of the cms collection
391 // to the start of the cms sweep (less than the total
392 // duration of the cms collection) can be used. This
393 // has been tried and some applications experienced
394 // promotion failures early in execution. This was
395 // possibly because the averages were not accurate
396 // enough at the beginning.
397 double CMSStats::time_until_cms_start() const {
398 // We add "gc0_period" to the "work" calculation
399 // below because this query is done (mostly) at the
400 // end of a scavenge, so we need to conservatively
401 // account for that much possible delay
402 // in the query so as to avoid concurrent mode failures
403 // due to starting the collection just a wee bit too
404 // late.
405 double work = cms_duration() + gc0_period();
406 double deadline = time_until_cms_gen_full();
407 // If a concurrent mode failure occurred recently, we want to be
408 // more conservative and halve our expected time_until_cms_gen_full()
409 if (work > deadline) {
410 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
411 cms_duration(), gc0_period(), time_until_cms_gen_full());
412 return 0.0;
413 }
414 return work - deadline;
415 }
416
417 #ifndef PRODUCT
418 void CMSStats::print_on(outputStream *st) const {
419 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
420 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
421 gc0_duration(), gc0_period(), gc0_promoted());
422 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
423 cms_duration(), cms_period(), cms_allocated());
424 st->print(",cms_since_beg=%g,cms_since_end=%g",
425 cms_time_since_begin(), cms_time_since_end());
426 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
427 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
428
429 if (valid()) {
430 st->print(",promo_rate=%g,cms_alloc_rate=%g",
431 promotion_rate(), cms_allocation_rate());
432 st->print(",cms_consumption_rate=%g,time_until_full=%g",
433 cms_consumption_rate(), time_until_cms_gen_full());
434 }
435 st->cr();
436 }
437 #endif // #ifndef PRODUCT
438
439 CMSCollector::CollectorState CMSCollector::_collectorState =
440 CMSCollector::Idling;
441 bool CMSCollector::_foregroundGCIsActive = false;
442 bool CMSCollector::_foregroundGCShouldWait = false;
443
444 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
445 CardTableRS* ct,
446 ConcurrentMarkSweepPolicy* cp):
447 _cmsGen(cmsGen),
448 // Adjust span to cover old (cms) gen
449 _span(cmsGen->reserved()),
450 _ct(ct),
451 _span_discoverer(_span),
452 _ref_processor(NULL), // will be set later
453 _conc_workers(NULL), // may be set later
454 _abort_preclean(false),
455 _start_sampling(false),
456 _between_prologue_and_epilogue(false),
457 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
458 _modUnionTable((CardTable::card_shift - LogHeapWordSize),
459 -1 /* lock-free */, "No_lock" /* dummy */),
460 _modUnionClosurePar(&_modUnionTable),
461 // Construct the is_alive_closure with _span & markBitMap
462 _is_alive_closure(_span, &_markBitMap),
463 _restart_addr(NULL),
464 _overflow_list(NULL),
465 _stats(cmsGen),
466 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
467 //verify that this lock should be acquired with safepoint check.
468 Monitor::_safepoint_check_sometimes)),
469 _eden_chunk_array(NULL), // may be set in ctor body
470 _eden_chunk_capacity(0), // -- ditto --
471 _eden_chunk_index(0), // -- ditto --
472 _survivor_plab_array(NULL), // -- ditto --
473 _survivor_chunk_array(NULL), // -- ditto --
474 _survivor_chunk_capacity(0), // -- ditto --
475 _survivor_chunk_index(0), // -- ditto --
476 _ser_pmc_preclean_ovflw(0),
477 _ser_kac_preclean_ovflw(0),
478 _ser_pmc_remark_ovflw(0),
479 _par_pmc_remark_ovflw(0),
480 _ser_kac_ovflw(0),
481 _par_kac_ovflw(0),
482 #ifndef PRODUCT
483 _num_par_pushes(0),
484 #endif
485 _collection_count_start(0),
486 _verifying(false),
487 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
488 _completed_initialization(false),
489 _collector_policy(cp),
490 _should_unload_classes(CMSClassUnloadingEnabled),
491 _concurrent_cycles_since_last_unload(0),
492 _roots_scanning_options(GenCollectedHeap::SO_None),
493 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
494 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
495 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
496 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
497 _cms_start_registered(false)
498 {
499 // Now expand the span and allocate the collection support structures
500 // (MUT, marking bit map etc.) to cover both generations subject to
501 // collection.
502
503 // For use by dirty card to oop closures.
504 _cmsGen->cmsSpace()->set_collector(this);
505
506 // Allocate MUT and marking bit map
507 {
508 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
509 if (!_markBitMap.allocate(_span)) {
510 log_warning(gc)("Failed to allocate CMS Bit Map");
511 return;
512 }
513 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
514 }
515 {
516 _modUnionTable.allocate(_span);
517 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
518 }
519
520 if (!_markStack.allocate(MarkStackSize)) {
521 log_warning(gc)("Failed to allocate CMS Marking Stack");
522 return;
523 }
524
525 // Support for multi-threaded concurrent phases
526 if (CMSConcurrentMTEnabled) {
527 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
528 // just for now
529 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
530 }
531 if (ConcGCThreads > 1) {
532 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
533 ConcGCThreads, true);
534 if (_conc_workers == NULL) {
535 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
536 CMSConcurrentMTEnabled = false;
537 } else {
538 _conc_workers->initialize_workers();
539 }
540 } else {
541 CMSConcurrentMTEnabled = false;
542 }
543 }
544 if (!CMSConcurrentMTEnabled) {
545 ConcGCThreads = 0;
546 } else {
547 // Turn off CMSCleanOnEnter optimization temporarily for
548 // the MT case where it's not fixed yet; see 6178663.
549 CMSCleanOnEnter = false;
550 }
551 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
552 "Inconsistency");
553 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
554 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
555
556 // Parallel task queues; these are shared for the
557 // concurrent and stop-world phases of CMS, but
558 // are not shared with parallel scavenge (ParNew).
559 {
560 uint i;
561 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
562
563 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
564 || ParallelRefProcEnabled)
565 && num_queues > 0) {
566 _task_queues = new OopTaskQueueSet(num_queues);
567 if (_task_queues == NULL) {
568 log_warning(gc)("task_queues allocation failure.");
569 return;
570 }
571 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
572 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
573 for (i = 0; i < num_queues; i++) {
574 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
575 if (q == NULL) {
576 log_warning(gc)("work_queue allocation failure.");
577 return;
578 }
579 _task_queues->register_queue(i, q);
580 }
581 for (i = 0; i < num_queues; i++) {
582 _task_queues->queue(i)->initialize();
583 _hash_seed[i] = 17; // copied from ParNew
584 }
585 }
586 }
587
588 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
589
590 // Clip CMSBootstrapOccupancy between 0 and 100.
591 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
592
593 // Now tell CMS generations the identity of their collector
594 ConcurrentMarkSweepGeneration::set_collector(this);
595
596 // Create & start a CMS thread for this CMS collector
597 _cmsThread = ConcurrentMarkSweepThread::start(this);
598 assert(cmsThread() != NULL, "CMS Thread should have been created");
599 assert(cmsThread()->collector() == this,
600 "CMS Thread should refer to this gen");
601 assert(CGC_lock != NULL, "Where's the CGC_lock?");
602
603 // Support for parallelizing young gen rescan
604 CMSHeap* heap = CMSHeap::heap();
605 assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
606 _young_gen = (ParNewGeneration*)heap->young_gen();
607 if (heap->supports_inline_contig_alloc()) {
608 _top_addr = heap->top_addr();
609 _end_addr = heap->end_addr();
610 assert(_young_gen != NULL, "no _young_gen");
611 _eden_chunk_index = 0;
612 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
613 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
614 }
615
616 // Support for parallelizing survivor space rescan
617 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
618 const size_t max_plab_samples =
619 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
620
621 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
622 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
623 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
624 _survivor_chunk_capacity = max_plab_samples;
625 for (uint i = 0; i < ParallelGCThreads; i++) {
626 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
627 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
628 assert(cur->end() == 0, "Should be 0");
629 assert(cur->array() == vec, "Should be vec");
630 assert(cur->capacity() == max_plab_samples, "Error");
631 }
632 }
633
634 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
635 _gc_counters = new CollectorCounters("CMS", 1);
636 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2);
637 _completed_initialization = true;
638 _inter_sweep_timer.start(); // start of time
639 }
640
641 const char* ConcurrentMarkSweepGeneration::name() const {
642 return "concurrent mark-sweep generation";
643 }
644 void ConcurrentMarkSweepGeneration::update_counters() {
645 if (UsePerfData) {
646 _space_counters->update_all();
647 _gen_counters->update_all();
648 }
649 }
650
651 // this is an optimized version of update_counters(). it takes the
652 // used value as a parameter rather than computing it.
653 //
654 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
655 if (UsePerfData) {
656 _space_counters->update_used(used);
657 _space_counters->update_capacity();
658 _gen_counters->update_all();
659 }
660 }
661
662 void ConcurrentMarkSweepGeneration::print() const {
663 Generation::print();
664 cmsSpace()->print();
665 }
666
667 #ifndef PRODUCT
668 void ConcurrentMarkSweepGeneration::print_statistics() {
669 cmsSpace()->printFLCensus(0);
670 }
671 #endif
672
673 size_t
674 ConcurrentMarkSweepGeneration::contiguous_available() const {
675 // dld proposes an improvement in precision here. If the committed
676 // part of the space ends in a free block we should add that to
677 // uncommitted size in the calculation below. Will make this
678 // change later, staying with the approximation below for the
679 // time being. -- ysr.
680 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
681 }
682
683 size_t
684 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
685 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
686 }
687
688 size_t ConcurrentMarkSweepGeneration::max_available() const {
689 return free() + _virtual_space.uncommitted_size();
690 }
691
692 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
693 size_t available = max_available();
694 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
695 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
696 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
697 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
698 return res;
699 }
700
701 // At a promotion failure dump information on block layout in heap
702 // (cms old generation).
703 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
704 Log(gc, promotion) log;
705 if (log.is_trace()) {
706 LogStream ls(log.trace());
707 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
708 }
709 }
710
711 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
712 // Clear the promotion information. These pointers can be adjusted
713 // along with all the other pointers into the heap but
714 // compaction is expected to be a rare event with
715 // a heap using cms so don't do it without seeing the need.
716 for (uint i = 0; i < ParallelGCThreads; i++) {
717 _par_gc_thread_states[i]->promo.reset();
718 }
719 }
720
721 void ConcurrentMarkSweepGeneration::compute_new_size() {
722 assert_locked_or_safepoint(Heap_lock);
723
724 // If incremental collection failed, we just want to expand
725 // to the limit.
726 if (incremental_collection_failed()) {
727 clear_incremental_collection_failed();
728 grow_to_reserved();
729 return;
730 }
731
732 // The heap has been compacted but not reset yet.
733 // Any metric such as free() or used() will be incorrect.
734
735 CardGeneration::compute_new_size();
736
737 // Reset again after a possible resizing
738 if (did_compact()) {
739 cmsSpace()->reset_after_compaction();
740 }
741 }
742
743 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
744 assert_locked_or_safepoint(Heap_lock);
745
746 // If incremental collection failed, we just want to expand
747 // to the limit.
748 if (incremental_collection_failed()) {
749 clear_incremental_collection_failed();
750 grow_to_reserved();
751 return;
752 }
753
754 double free_percentage = ((double) free()) / capacity();
755 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
756 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
757
758 // compute expansion delta needed for reaching desired free percentage
759 if (free_percentage < desired_free_percentage) {
760 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
761 assert(desired_capacity >= capacity(), "invalid expansion size");
762 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
763 Log(gc) log;
764 if (log.is_trace()) {
765 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
766 log.trace("From compute_new_size: ");
767 log.trace(" Free fraction %f", free_percentage);
768 log.trace(" Desired free fraction %f", desired_free_percentage);
769 log.trace(" Maximum free fraction %f", maximum_free_percentage);
770 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000);
771 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
772 CMSHeap* heap = CMSHeap::heap();
773 assert(heap->is_old_gen(this), "The CMS generation should always be the old generation");
774 size_t young_size = heap->young_gen()->capacity();
775 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000);
776 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
777 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
778 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
779 }
780 // safe if expansion fails
781 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
782 log.trace(" Expanded free fraction %f", ((double) free()) / capacity());
783 } else {
784 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
785 assert(desired_capacity <= capacity(), "invalid expansion size");
786 size_t shrink_bytes = capacity() - desired_capacity;
787 // Don't shrink unless the delta is greater than the minimum shrink we want
788 if (shrink_bytes >= MinHeapDeltaBytes) {
789 shrink_free_list_by(shrink_bytes);
790 }
791 }
792 }
793
794 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
795 return cmsSpace()->freelistLock();
796 }
797
798 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
799 CMSSynchronousYieldRequest yr;
800 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
801 return have_lock_and_allocate(size, tlab);
802 }
803
804 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
805 bool tlab /* ignored */) {
806 assert_lock_strong(freelistLock());
807 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
808 HeapWord* res = cmsSpace()->allocate(adjustedSize);
809 // Allocate the object live (grey) if the background collector has
810 // started marking. This is necessary because the marker may
811 // have passed this address and consequently this object will
812 // not otherwise be greyed and would be incorrectly swept up.
813 // Note that if this object contains references, the writing
814 // of those references will dirty the card containing this object
815 // allowing the object to be blackened (and its references scanned)
816 // either during a preclean phase or at the final checkpoint.
817 if (res != NULL) {
818 // We may block here with an uninitialized object with
819 // its mark-bit or P-bits not yet set. Such objects need
820 // to be safely navigable by block_start().
821 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
822 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
823 collector()->direct_allocated(res, adjustedSize);
824 _direct_allocated_words += adjustedSize;
825 // allocation counters
826 NOT_PRODUCT(
827 _numObjectsAllocated++;
828 _numWordsAllocated += (int)adjustedSize;
829 )
830 }
831 return res;
832 }
833
834 // In the case of direct allocation by mutators in a generation that
835 // is being concurrently collected, the object must be allocated
836 // live (grey) if the background collector has started marking.
837 // This is necessary because the marker may
838 // have passed this address and consequently this object will
839 // not otherwise be greyed and would be incorrectly swept up.
840 // Note that if this object contains references, the writing
841 // of those references will dirty the card containing this object
842 // allowing the object to be blackened (and its references scanned)
843 // either during a preclean phase or at the final checkpoint.
844 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
845 assert(_markBitMap.covers(start, size), "Out of bounds");
846 if (_collectorState >= Marking) {
847 MutexLockerEx y(_markBitMap.lock(),
848 Mutex::_no_safepoint_check_flag);
849 // [see comments preceding SweepClosure::do_blk() below for details]
850 //
851 // Can the P-bits be deleted now? JJJ
852 //
853 // 1. need to mark the object as live so it isn't collected
854 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
855 // 3. need to mark the end of the object so marking, precleaning or sweeping
856 // can skip over uninitialized or unparsable objects. An allocated
857 // object is considered uninitialized for our purposes as long as
858 // its klass word is NULL. All old gen objects are parsable
859 // as soon as they are initialized.)
860 _markBitMap.mark(start); // object is live
861 _markBitMap.mark(start + 1); // object is potentially uninitialized?
862 _markBitMap.mark(start + size - 1);
863 // mark end of object
864 }
865 // check that oop looks uninitialized
866 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
867 }
868
869 void CMSCollector::promoted(bool par, HeapWord* start,
870 bool is_obj_array, size_t obj_size) {
871 assert(_markBitMap.covers(start), "Out of bounds");
872 // See comment in direct_allocated() about when objects should
873 // be allocated live.
874 if (_collectorState >= Marking) {
875 // we already hold the marking bit map lock, taken in
876 // the prologue
877 if (par) {
878 _markBitMap.par_mark(start);
879 } else {
880 _markBitMap.mark(start);
881 }
882 // We don't need to mark the object as uninitialized (as
883 // in direct_allocated above) because this is being done with the
884 // world stopped and the object will be initialized by the
885 // time the marking, precleaning or sweeping get to look at it.
886 // But see the code for copying objects into the CMS generation,
887 // where we need to ensure that concurrent readers of the
888 // block offset table are able to safely navigate a block that
889 // is in flux from being free to being allocated (and in
890 // transition while being copied into) and subsequently
891 // becoming a bona-fide object when the copy/promotion is complete.
892 assert(SafepointSynchronize::is_at_safepoint(),
893 "expect promotion only at safepoints");
894
895 if (_collectorState < Sweeping) {
896 // Mark the appropriate cards in the modUnionTable, so that
897 // this object gets scanned before the sweep. If this is
898 // not done, CMS generation references in the object might
899 // not get marked.
900 // For the case of arrays, which are otherwise precisely
901 // marked, we need to dirty the entire array, not just its head.
902 if (is_obj_array) {
903 // The [par_]mark_range() method expects mr.end() below to
904 // be aligned to the granularity of a bit's representation
905 // in the heap. In the case of the MUT below, that's a
906 // card size.
907 MemRegion mr(start,
908 align_up(start + obj_size,
909 CardTable::card_size /* bytes */));
910 if (par) {
911 _modUnionTable.par_mark_range(mr);
912 } else {
913 _modUnionTable.mark_range(mr);
914 }
915 } else { // not an obj array; we can just mark the head
916 if (par) {
917 _modUnionTable.par_mark(start);
918 } else {
919 _modUnionTable.mark(start);
920 }
921 }
922 }
923 }
924 }
925
926 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
927 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
928 // allocate, copy and if necessary update promoinfo --
929 // delegate to underlying space.
930 assert_lock_strong(freelistLock());
931
932 #ifndef PRODUCT
933 if (CMSHeap::heap()->promotion_should_fail()) {
934 return NULL;
935 }
936 #endif // #ifndef PRODUCT
937
938 oop res = _cmsSpace->promote(obj, obj_size);
939 if (res == NULL) {
940 // expand and retry
941 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
942 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
943 // Since this is the old generation, we don't try to promote
944 // into a more senior generation.
945 res = _cmsSpace->promote(obj, obj_size);
946 }
947 if (res != NULL) {
948 // See comment in allocate() about when objects should
949 // be allocated live.
950 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below");
951 collector()->promoted(false, // Not parallel
952 (HeapWord*)res, obj->is_objArray(), obj_size);
953 // promotion counters
954 NOT_PRODUCT(
955 _numObjectsPromoted++;
956 _numWordsPromoted +=
957 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
958 )
959 }
960 return res;
961 }
962
963
964 // IMPORTANT: Notes on object size recognition in CMS.
965 // ---------------------------------------------------
966 // A block of storage in the CMS generation is always in
967 // one of three states. A free block (FREE), an allocated
968 // object (OBJECT) whose size() method reports the correct size,
969 // and an intermediate state (TRANSIENT) in which its size cannot
970 // be accurately determined.
971 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
972 // -----------------------------------------------------
973 // FREE: klass_word & 1 == 1; mark_word holds block size
974 //
975 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
976 // obj->size() computes correct size
977 //
978 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
979 //
980 // STATE IDENTIFICATION: (64 bit+COOPS)
981 // ------------------------------------
982 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
983 //
984 // OBJECT: klass_word installed; klass_word != 0;
985 // obj->size() computes correct size
986 //
987 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
988 //
989 //
990 // STATE TRANSITION DIAGRAM
991 //
992 // mut / parnew mut / parnew
993 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
994 // ^ |
995 // |------------------------ DEAD <------------------------------------|
996 // sweep mut
997 //
998 // While a block is in TRANSIENT state its size cannot be determined
999 // so readers will either need to come back later or stall until
1000 // the size can be determined. Note that for the case of direct
1001 // allocation, P-bits, when available, may be used to determine the
1002 // size of an object that may not yet have been initialized.
1003
1004 // Things to support parallel young-gen collection.
1005 oop
1006 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1007 oop old, markOop m,
1008 size_t word_sz) {
1009 #ifndef PRODUCT
1010 if (CMSHeap::heap()->promotion_should_fail()) {
1011 return NULL;
1012 }
1013 #endif // #ifndef PRODUCT
1014
1015 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1016 PromotionInfo* promoInfo = &ps->promo;
1017 // if we are tracking promotions, then first ensure space for
1018 // promotion (including spooling space for saving header if necessary).
1019 // then allocate and copy, then track promoted info if needed.
1020 // When tracking (see PromotionInfo::track()), the mark word may
1021 // be displaced and in this case restoration of the mark word
1022 // occurs in the (oop_since_save_marks_)iterate phase.
1023 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1024 // Out of space for allocating spooling buffers;
1025 // try expanding and allocating spooling buffers.
1026 if (!expand_and_ensure_spooling_space(promoInfo)) {
1027 return NULL;
1028 }
1029 }
1030 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1031 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1032 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1033 if (obj_ptr == NULL) {
1034 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1035 if (obj_ptr == NULL) {
1036 return NULL;
1037 }
1038 }
1039 oop obj = oop(obj_ptr);
1040 OrderAccess::storestore();
1041 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1042 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1043 // IMPORTANT: See note on object initialization for CMS above.
1044 // Otherwise, copy the object. Here we must be careful to insert the
1045 // klass pointer last, since this marks the block as an allocated object.
1046 // Except with compressed oops it's the mark word.
1047 HeapWord* old_ptr = (HeapWord*)old;
1048 // Restore the mark word copied above.
1049 obj->set_mark_raw(m);
1050 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1051 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1052 OrderAccess::storestore();
1053
1054 if (UseCompressedClassPointers) {
1055 // Copy gap missed by (aligned) header size calculation below
1056 obj->set_klass_gap(old->klass_gap());
1057 }
1058 if (word_sz > (size_t)oopDesc::header_size()) {
1059 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1060 obj_ptr + oopDesc::header_size(),
1061 word_sz - oopDesc::header_size());
1062 }
1063
1064 // Now we can track the promoted object, if necessary. We take care
1065 // to delay the transition from uninitialized to full object
1066 // (i.e., insertion of klass pointer) until after, so that it
1067 // atomically becomes a promoted object.
1068 if (promoInfo->tracking()) {
1069 promoInfo->track((PromotedObject*)obj, old->klass());
1070 }
1071 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1072 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1073 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below");
1074
1075 // Finally, install the klass pointer (this should be volatile).
1076 OrderAccess::storestore();
1077 obj->set_klass(old->klass());
1078 // We should now be able to calculate the right size for this object
1079 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1080
1081 collector()->promoted(true, // parallel
1082 obj_ptr, old->is_objArray(), word_sz);
1083
1084 NOT_PRODUCT(
1085 Atomic::inc(&_numObjectsPromoted);
1086 Atomic::add(alloc_sz, &_numWordsPromoted);
1087 )
1088
1089 return obj;
1090 }
1091
1092 void
1093 ConcurrentMarkSweepGeneration::
1094 par_promote_alloc_done(int thread_num) {
1095 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1096 ps->lab.retire(thread_num);
1097 }
1098
1099 void
1100 ConcurrentMarkSweepGeneration::
1101 par_oop_since_save_marks_iterate_done(int thread_num) {
1102 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1103 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1104 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1105
1106 // Because card-scanning has been completed, subsequent phases
1107 // (e.g., reference processing) will not need to recognize which
1108 // objects have been promoted during this GC. So, we can now disable
1109 // promotion tracking.
1110 ps->promo.stopTrackingPromotions();
1111 }
1112
1113 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1114 size_t size,
1115 bool tlab)
1116 {
1117 // We allow a STW collection only if a full
1118 // collection was requested.
1119 return full || should_allocate(size, tlab); // FIX ME !!!
1120 // This and promotion failure handling are connected at the
1121 // hip and should be fixed by untying them.
1122 }
1123
1124 bool CMSCollector::shouldConcurrentCollect() {
1125 LogTarget(Trace, gc) log;
1126
1127 if (_full_gc_requested) {
1128 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)");
1129 return true;
1130 }
1131
1132 FreelistLocker x(this);
1133 // ------------------------------------------------------------------
1134 // Print out lots of information which affects the initiation of
1135 // a collection.
1136 if (log.is_enabled() && stats().valid()) {
1137 log.print("CMSCollector shouldConcurrentCollect: ");
1138
1139 LogStream out(log);
1140 stats().print_on(&out);
1141
1142 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1143 log.print("free=" SIZE_FORMAT, _cmsGen->free());
1144 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1145 log.print("promotion_rate=%g", stats().promotion_rate());
1146 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1147 log.print("occupancy=%3.7f", _cmsGen->occupancy());
1148 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1149 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1150 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1151 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1152 }
1153 // ------------------------------------------------------------------
1154
1155 // If the estimated time to complete a cms collection (cms_duration())
1156 // is less than the estimated time remaining until the cms generation
1157 // is full, start a collection.
1158 if (!UseCMSInitiatingOccupancyOnly) {
1159 if (stats().valid()) {
1160 if (stats().time_until_cms_start() == 0.0) {
1161 return true;
1162 }
1163 } else {
1164 // We want to conservatively collect somewhat early in order
1165 // to try and "bootstrap" our CMS/promotion statistics;
1166 // this branch will not fire after the first successful CMS
1167 // collection because the stats should then be valid.
1168 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1169 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1170 _cmsGen->occupancy(), _bootstrap_occupancy);
1171 return true;
1172 }
1173 }
1174 }
1175
1176 // Otherwise, we start a collection cycle if
1177 // old gen want a collection cycle started. Each may use
1178 // an appropriate criterion for making this decision.
1179 // XXX We need to make sure that the gen expansion
1180 // criterion dovetails well with this. XXX NEED TO FIX THIS
1181 if (_cmsGen->should_concurrent_collect()) {
1182 log.print("CMS old gen initiated");
1183 return true;
1184 }
1185
1186 // We start a collection if we believe an incremental collection may fail;
1187 // this is not likely to be productive in practice because it's probably too
1188 // late anyway.
1189 CMSHeap* heap = CMSHeap::heap();
1190 if (heap->incremental_collection_will_fail(true /* consult_young */)) {
1191 log.print("CMSCollector: collect because incremental collection will fail ");
1192 return true;
1193 }
1194
1195 if (MetaspaceGC::should_concurrent_collect()) {
1196 log.print("CMSCollector: collect for metadata allocation ");
1197 return true;
1198 }
1199
1200 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1201 if (CMSTriggerInterval >= 0) {
1202 if (CMSTriggerInterval == 0) {
1203 // Trigger always
1204 return true;
1205 }
1206
1207 // Check the CMS time since begin (we do not check the stats validity
1208 // as we want to be able to trigger the first CMS cycle as well)
1209 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1210 if (stats().valid()) {
1211 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1212 stats().cms_time_since_begin());
1213 } else {
1214 log.print("CMSCollector: collect because of trigger interval (first collection)");
1215 }
1216 return true;
1217 }
1218 }
1219
1220 return false;
1221 }
1222
1223 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1224
1225 // Clear _expansion_cause fields of constituent generations
1226 void CMSCollector::clear_expansion_cause() {
1227 _cmsGen->clear_expansion_cause();
1228 }
1229
1230 // We should be conservative in starting a collection cycle. To
1231 // start too eagerly runs the risk of collecting too often in the
1232 // extreme. To collect too rarely falls back on full collections,
1233 // which works, even if not optimum in terms of concurrent work.
1234 // As a work around for too eagerly collecting, use the flag
1235 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1236 // giving the user an easily understandable way of controlling the
1237 // collections.
1238 // We want to start a new collection cycle if any of the following
1239 // conditions hold:
1240 // . our current occupancy exceeds the configured initiating occupancy
1241 // for this generation, or
1242 // . we recently needed to expand this space and have not, since that
1243 // expansion, done a collection of this generation, or
1244 // . the underlying space believes that it may be a good idea to initiate
1245 // a concurrent collection (this may be based on criteria such as the
1246 // following: the space uses linear allocation and linear allocation is
1247 // going to fail, or there is believed to be excessive fragmentation in
1248 // the generation, etc... or ...
1249 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1250 // the case of the old generation; see CR 6543076):
1251 // we may be approaching a point at which allocation requests may fail because
1252 // we will be out of sufficient free space given allocation rate estimates.]
1253 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1254
1255 assert_lock_strong(freelistLock());
1256 if (occupancy() > initiating_occupancy()) {
1257 log_trace(gc)(" %s: collect because of occupancy %f / %f ",
1258 short_name(), occupancy(), initiating_occupancy());
1259 return true;
1260 }
1261 if (UseCMSInitiatingOccupancyOnly) {
1262 return false;
1263 }
1264 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1265 log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1266 return true;
1267 }
1268 return false;
1269 }
1270
1271 void ConcurrentMarkSweepGeneration::collect(bool full,
1272 bool clear_all_soft_refs,
1273 size_t size,
1274 bool tlab)
1275 {
1276 collector()->collect(full, clear_all_soft_refs, size, tlab);
1277 }
1278
1279 void CMSCollector::collect(bool full,
1280 bool clear_all_soft_refs,
1281 size_t size,
1282 bool tlab)
1283 {
1284 // The following "if" branch is present for defensive reasons.
1285 // In the current uses of this interface, it can be replaced with:
1286 // assert(!GCLocker.is_active(), "Can't be called otherwise");
1287 // But I am not placing that assert here to allow future
1288 // generality in invoking this interface.
1289 if (GCLocker::is_active()) {
1290 // A consistency test for GCLocker
1291 assert(GCLocker::needs_gc(), "Should have been set already");
1292 // Skip this foreground collection, instead
1293 // expanding the heap if necessary.
1294 // Need the free list locks for the call to free() in compute_new_size()
1295 compute_new_size();
1296 return;
1297 }
1298 acquire_control_and_collect(full, clear_all_soft_refs);
1299 }
1300
1301 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1302 CMSHeap* heap = CMSHeap::heap();
1303 unsigned int gc_count = heap->total_full_collections();
1304 if (gc_count == full_gc_count) {
1305 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1306 _full_gc_requested = true;
1307 _full_gc_cause = cause;
1308 CGC_lock->notify(); // nudge CMS thread
1309 } else {
1310 assert(gc_count > full_gc_count, "Error: causal loop");
1311 }
1312 }
1313
1314 bool CMSCollector::is_external_interruption() {
1315 GCCause::Cause cause = CMSHeap::heap()->gc_cause();
1316 return GCCause::is_user_requested_gc(cause) ||
1317 GCCause::is_serviceability_requested_gc(cause);
1318 }
1319
1320 void CMSCollector::report_concurrent_mode_interruption() {
1321 if (is_external_interruption()) {
1322 log_debug(gc)("Concurrent mode interrupted");
1323 } else {
1324 log_debug(gc)("Concurrent mode failure");
1325 _gc_tracer_cm->report_concurrent_mode_failure();
1326 }
1327 }
1328
1329
1330 // The foreground and background collectors need to coordinate in order
1331 // to make sure that they do not mutually interfere with CMS collections.
1332 // When a background collection is active,
1333 // the foreground collector may need to take over (preempt) and
1334 // synchronously complete an ongoing collection. Depending on the
1335 // frequency of the background collections and the heap usage
1336 // of the application, this preemption can be seldom or frequent.
1337 // There are only certain
1338 // points in the background collection that the "collection-baton"
1339 // can be passed to the foreground collector.
1340 //
1341 // The foreground collector will wait for the baton before
1342 // starting any part of the collection. The foreground collector
1343 // will only wait at one location.
1344 //
1345 // The background collector will yield the baton before starting a new
1346 // phase of the collection (e.g., before initial marking, marking from roots,
1347 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1348 // of the loop which switches the phases. The background collector does some
1349 // of the phases (initial mark, final re-mark) with the world stopped.
1350 // Because of locking involved in stopping the world,
1351 // the foreground collector should not block waiting for the background
1352 // collector when it is doing a stop-the-world phase. The background
1353 // collector will yield the baton at an additional point just before
1354 // it enters a stop-the-world phase. Once the world is stopped, the
1355 // background collector checks the phase of the collection. If the
1356 // phase has not changed, it proceeds with the collection. If the
1357 // phase has changed, it skips that phase of the collection. See
1358 // the comments on the use of the Heap_lock in collect_in_background().
1359 //
1360 // Variable used in baton passing.
1361 // _foregroundGCIsActive - Set to true by the foreground collector when
1362 // it wants the baton. The foreground clears it when it has finished
1363 // the collection.
1364 // _foregroundGCShouldWait - Set to true by the background collector
1365 // when it is running. The foreground collector waits while
1366 // _foregroundGCShouldWait is true.
1367 // CGC_lock - monitor used to protect access to the above variables
1368 // and to notify the foreground and background collectors.
1369 // _collectorState - current state of the CMS collection.
1370 //
1371 // The foreground collector
1372 // acquires the CGC_lock
1373 // sets _foregroundGCIsActive
1374 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1375 // various locks acquired in preparation for the collection
1376 // are released so as not to block the background collector
1377 // that is in the midst of a collection
1378 // proceeds with the collection
1379 // clears _foregroundGCIsActive
1380 // returns
1381 //
1382 // The background collector in a loop iterating on the phases of the
1383 // collection
1384 // acquires the CGC_lock
1385 // sets _foregroundGCShouldWait
1386 // if _foregroundGCIsActive is set
1387 // clears _foregroundGCShouldWait, notifies _CGC_lock
1388 // waits on _CGC_lock for _foregroundGCIsActive to become false
1389 // and exits the loop.
1390 // otherwise
1391 // proceed with that phase of the collection
1392 // if the phase is a stop-the-world phase,
1393 // yield the baton once more just before enqueueing
1394 // the stop-world CMS operation (executed by the VM thread).
1395 // returns after all phases of the collection are done
1396 //
1397
1398 void CMSCollector::acquire_control_and_collect(bool full,
1399 bool clear_all_soft_refs) {
1400 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1401 assert(!Thread::current()->is_ConcurrentGC_thread(),
1402 "shouldn't try to acquire control from self!");
1403
1404 // Start the protocol for acquiring control of the
1405 // collection from the background collector (aka CMS thread).
1406 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1407 "VM thread should have CMS token");
1408 // Remember the possibly interrupted state of an ongoing
1409 // concurrent collection
1410 CollectorState first_state = _collectorState;
1411
1412 // Signal to a possibly ongoing concurrent collection that
1413 // we want to do a foreground collection.
1414 _foregroundGCIsActive = true;
1415
1416 // release locks and wait for a notify from the background collector
1417 // releasing the locks in only necessary for phases which
1418 // do yields to improve the granularity of the collection.
1419 assert_lock_strong(bitMapLock());
1420 // We need to lock the Free list lock for the space that we are
1421 // currently collecting.
1422 assert(haveFreelistLocks(), "Must be holding free list locks");
1423 bitMapLock()->unlock();
1424 releaseFreelistLocks();
1425 {
1426 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1427 if (_foregroundGCShouldWait) {
1428 // We are going to be waiting for action for the CMS thread;
1429 // it had better not be gone (for instance at shutdown)!
1430 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1431 "CMS thread must be running");
1432 // Wait here until the background collector gives us the go-ahead
1433 ConcurrentMarkSweepThread::clear_CMS_flag(
1434 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1435 // Get a possibly blocked CMS thread going:
1436 // Note that we set _foregroundGCIsActive true above,
1437 // without protection of the CGC_lock.
1438 CGC_lock->notify();
1439 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1440 "Possible deadlock");
1441 while (_foregroundGCShouldWait) {
1442 // wait for notification
1443 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1444 // Possibility of delay/starvation here, since CMS token does
1445 // not know to give priority to VM thread? Actually, i think
1446 // there wouldn't be any delay/starvation, but the proof of
1447 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1448 }
1449 ConcurrentMarkSweepThread::set_CMS_flag(
1450 ConcurrentMarkSweepThread::CMS_vm_has_token);
1451 }
1452 }
1453 // The CMS_token is already held. Get back the other locks.
1454 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1455 "VM thread should have CMS token");
1456 getFreelistLocks();
1457 bitMapLock()->lock_without_safepoint_check();
1458 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1459 p2i(Thread::current()), first_state);
1460 log_debug(gc, state)(" gets control with state %d", _collectorState);
1461
1462 // Inform cms gen if this was due to partial collection failing.
1463 // The CMS gen may use this fact to determine its expansion policy.
1464 CMSHeap* heap = CMSHeap::heap();
1465 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) {
1466 assert(!_cmsGen->incremental_collection_failed(),
1467 "Should have been noticed, reacted to and cleared");
1468 _cmsGen->set_incremental_collection_failed();
1469 }
1470
1471 if (first_state > Idling) {
1472 report_concurrent_mode_interruption();
1473 }
1474
1475 set_did_compact(true);
1476
1477 // If the collection is being acquired from the background
1478 // collector, there may be references on the discovered
1479 // references lists. Abandon those references, since some
1480 // of them may have become unreachable after concurrent
1481 // discovery; the STW compacting collector will redo discovery
1482 // more precisely, without being subject to floating garbage.
1483 // Leaving otherwise unreachable references in the discovered
1484 // lists would require special handling.
1485 ref_processor()->disable_discovery();
1486 ref_processor()->abandon_partial_discovery();
1487 ref_processor()->verify_no_references_recorded();
1488
1489 if (first_state > Idling) {
1490 save_heap_summary();
1491 }
1492
1493 do_compaction_work(clear_all_soft_refs);
1494
1495 // Has the GC time limit been exceeded?
1496 size_t max_eden_size = _young_gen->max_eden_size();
1497 GCCause::Cause gc_cause = heap->gc_cause();
1498 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1499 _young_gen->eden()->used(),
1500 _cmsGen->max_capacity(),
1501 max_eden_size,
1502 full,
1503 gc_cause,
1504 heap->soft_ref_policy());
1505
1506 // Reset the expansion cause, now that we just completed
1507 // a collection cycle.
1508 clear_expansion_cause();
1509 _foregroundGCIsActive = false;
1510 return;
1511 }
1512
1513 // Resize the tenured generation
1514 // after obtaining the free list locks for the
1515 // two generations.
1516 void CMSCollector::compute_new_size() {
1517 assert_locked_or_safepoint(Heap_lock);
1518 FreelistLocker z(this);
1519 MetaspaceGC::compute_new_size();
1520 _cmsGen->compute_new_size_free_list();
1521 }
1522
1523 // A work method used by the foreground collector to do
1524 // a mark-sweep-compact.
1525 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1526 CMSHeap* heap = CMSHeap::heap();
1527
1528 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1529 gc_timer->register_gc_start();
1530
1531 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1532 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start());
1533
1534 heap->pre_full_gc_dump(gc_timer);
1535
1536 GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1537
1538 // Temporarily widen the span of the weak reference processing to
1539 // the entire heap.
1540 MemRegion new_span(CMSHeap::heap()->reserved_region());
1541 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1542 // Temporarily, clear the "is_alive_non_header" field of the
1543 // reference processor.
1544 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1545 // Temporarily make reference _processing_ single threaded (non-MT).
1546 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1547 // Temporarily make refs discovery atomic
1548 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1549 // Temporarily make reference _discovery_ single threaded (non-MT)
1550 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1551
1552 ref_processor()->set_enqueuing_is_done(false);
1553 ref_processor()->enable_discovery();
1554 ref_processor()->setup_policy(clear_all_soft_refs);
1555 // If an asynchronous collection finishes, the _modUnionTable is
1556 // all clear. If we are assuming the collection from an asynchronous
1557 // collection, clear the _modUnionTable.
1558 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1559 "_modUnionTable should be clear if the baton was not passed");
1560 _modUnionTable.clear_all();
1561 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(),
1562 "mod union for klasses should be clear if the baton was passed");
1563 _ct->cld_rem_set()->clear_mod_union();
1564
1565
1566 // We must adjust the allocation statistics being maintained
1567 // in the free list space. We do so by reading and clearing
1568 // the sweep timer and updating the block flux rate estimates below.
1569 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1570 if (_inter_sweep_timer.is_active()) {
1571 _inter_sweep_timer.stop();
1572 // Note that we do not use this sample to update the _inter_sweep_estimate.
1573 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1574 _inter_sweep_estimate.padded_average(),
1575 _intra_sweep_estimate.padded_average());
1576 }
1577
1578 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1579 #ifdef ASSERT
1580 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1581 size_t free_size = cms_space->free();
1582 assert(free_size ==
1583 pointer_delta(cms_space->end(), cms_space->compaction_top())
1584 * HeapWordSize,
1585 "All the free space should be compacted into one chunk at top");
1586 assert(cms_space->dictionary()->total_chunk_size(
1587 debug_only(cms_space->freelistLock())) == 0 ||
1588 cms_space->totalSizeInIndexedFreeLists() == 0,
1589 "All the free space should be in a single chunk");
1590 size_t num = cms_space->totalCount();
1591 assert((free_size == 0 && num == 0) ||
1592 (free_size > 0 && (num == 1 || num == 2)),
1593 "There should be at most 2 free chunks after compaction");
1594 #endif // ASSERT
1595 _collectorState = Resetting;
1596 assert(_restart_addr == NULL,
1597 "Should have been NULL'd before baton was passed");
1598 reset_stw();
1599 _cmsGen->reset_after_compaction();
1600 _concurrent_cycles_since_last_unload = 0;
1601
1602 // Clear any data recorded in the PLAB chunk arrays.
1603 if (_survivor_plab_array != NULL) {
1604 reset_survivor_plab_arrays();
1605 }
1606
1607 // Adjust the per-size allocation stats for the next epoch.
1608 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1609 // Restart the "inter sweep timer" for the next epoch.
1610 _inter_sweep_timer.reset();
1611 _inter_sweep_timer.start();
1612
1613 // No longer a need to do a concurrent collection for Metaspace.
1614 MetaspaceGC::set_should_concurrent_collect(false);
1615
1616 heap->post_full_gc_dump(gc_timer);
1617
1618 gc_timer->register_gc_end();
1619
1620 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1621
1622 // For a mark-sweep-compact, compute_new_size() will be called
1623 // in the heap's do_collection() method.
1624 }
1625
1626 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1627 Log(gc, heap) log;
1628 if (!log.is_trace()) {
1629 return;
1630 }
1631
1632 ContiguousSpace* eden_space = _young_gen->eden();
1633 ContiguousSpace* from_space = _young_gen->from();
1634 ContiguousSpace* to_space = _young_gen->to();
1635 // Eden
1636 if (_eden_chunk_array != NULL) {
1637 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1638 p2i(eden_space->bottom()), p2i(eden_space->top()),
1639 p2i(eden_space->end()), eden_space->capacity());
1640 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1641 _eden_chunk_index, _eden_chunk_capacity);
1642 for (size_t i = 0; i < _eden_chunk_index; i++) {
1643 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1644 }
1645 }
1646 // Survivor
1647 if (_survivor_chunk_array != NULL) {
1648 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1649 p2i(from_space->bottom()), p2i(from_space->top()),
1650 p2i(from_space->end()), from_space->capacity());
1651 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1652 _survivor_chunk_index, _survivor_chunk_capacity);
1653 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1654 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1655 }
1656 }
1657 }
1658
1659 void CMSCollector::getFreelistLocks() const {
1660 // Get locks for all free lists in all generations that this
1661 // collector is responsible for
1662 _cmsGen->freelistLock()->lock_without_safepoint_check();
1663 }
1664
1665 void CMSCollector::releaseFreelistLocks() const {
1666 // Release locks for all free lists in all generations that this
1667 // collector is responsible for
1668 _cmsGen->freelistLock()->unlock();
1669 }
1670
1671 bool CMSCollector::haveFreelistLocks() const {
1672 // Check locks for all free lists in all generations that this
1673 // collector is responsible for
1674 assert_lock_strong(_cmsGen->freelistLock());
1675 PRODUCT_ONLY(ShouldNotReachHere());
1676 return true;
1677 }
1678
1679 // A utility class that is used by the CMS collector to
1680 // temporarily "release" the foreground collector from its
1681 // usual obligation to wait for the background collector to
1682 // complete an ongoing phase before proceeding.
1683 class ReleaseForegroundGC: public StackObj {
1684 private:
1685 CMSCollector* _c;
1686 public:
1687 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1688 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1689 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1690 // allow a potentially blocked foreground collector to proceed
1691 _c->_foregroundGCShouldWait = false;
1692 if (_c->_foregroundGCIsActive) {
1693 CGC_lock->notify();
1694 }
1695 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1696 "Possible deadlock");
1697 }
1698
1699 ~ReleaseForegroundGC() {
1700 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1701 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1702 _c->_foregroundGCShouldWait = true;
1703 }
1704 };
1705
1706 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1707 assert(Thread::current()->is_ConcurrentGC_thread(),
1708 "A CMS asynchronous collection is only allowed on a CMS thread.");
1709
1710 CMSHeap* heap = CMSHeap::heap();
1711 {
1712 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1713 MutexLockerEx hl(Heap_lock, safepoint_check);
1714 FreelistLocker fll(this);
1715 MutexLockerEx x(CGC_lock, safepoint_check);
1716 if (_foregroundGCIsActive) {
1717 // The foreground collector is. Skip this
1718 // background collection.
1719 assert(!_foregroundGCShouldWait, "Should be clear");
1720 return;
1721 } else {
1722 assert(_collectorState == Idling, "Should be idling before start.");
1723 _collectorState = InitialMarking;
1724 register_gc_start(cause);
1725 // Reset the expansion cause, now that we are about to begin
1726 // a new cycle.
1727 clear_expansion_cause();
1728
1729 // Clear the MetaspaceGC flag since a concurrent collection
1730 // is starting but also clear it after the collection.
1731 MetaspaceGC::set_should_concurrent_collect(false);
1732 }
1733 // Decide if we want to enable class unloading as part of the
1734 // ensuing concurrent GC cycle.
1735 update_should_unload_classes();
1736 _full_gc_requested = false; // acks all outstanding full gc requests
1737 _full_gc_cause = GCCause::_no_gc;
1738 // Signal that we are about to start a collection
1739 heap->increment_total_full_collections(); // ... starting a collection cycle
1740 _collection_count_start = heap->total_full_collections();
1741 }
1742
1743 size_t prev_used = _cmsGen->used();
1744
1745 // The change of the collection state is normally done at this level;
1746 // the exceptions are phases that are executed while the world is
1747 // stopped. For those phases the change of state is done while the
1748 // world is stopped. For baton passing purposes this allows the
1749 // background collector to finish the phase and change state atomically.
1750 // The foreground collector cannot wait on a phase that is done
1751 // while the world is stopped because the foreground collector already
1752 // has the world stopped and would deadlock.
1753 while (_collectorState != Idling) {
1754 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1755 p2i(Thread::current()), _collectorState);
1756 // The foreground collector
1757 // holds the Heap_lock throughout its collection.
1758 // holds the CMS token (but not the lock)
1759 // except while it is waiting for the background collector to yield.
1760 //
1761 // The foreground collector should be blocked (not for long)
1762 // if the background collector is about to start a phase
1763 // executed with world stopped. If the background
1764 // collector has already started such a phase, the
1765 // foreground collector is blocked waiting for the
1766 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1767 // are executed in the VM thread.
1768 //
1769 // The locking order is
1770 // PendingListLock (PLL) -- if applicable (FinalMarking)
1771 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1772 // CMS token (claimed in
1773 // stop_world_and_do() -->
1774 // safepoint_synchronize() -->
1775 // CMSThread::synchronize())
1776
1777 {
1778 // Check if the FG collector wants us to yield.
1779 CMSTokenSync x(true); // is cms thread
1780 if (waitForForegroundGC()) {
1781 // We yielded to a foreground GC, nothing more to be
1782 // done this round.
1783 assert(_foregroundGCShouldWait == false, "We set it to false in "
1784 "waitForForegroundGC()");
1785 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1786 p2i(Thread::current()), _collectorState);
1787 return;
1788 } else {
1789 // The background collector can run but check to see if the
1790 // foreground collector has done a collection while the
1791 // background collector was waiting to get the CGC_lock
1792 // above. If yes, break so that _foregroundGCShouldWait
1793 // is cleared before returning.
1794 if (_collectorState == Idling) {
1795 break;
1796 }
1797 }
1798 }
1799
1800 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1801 "should be waiting");
1802
1803 switch (_collectorState) {
1804 case InitialMarking:
1805 {
1806 ReleaseForegroundGC x(this);
1807 stats().record_cms_begin();
1808 VM_CMS_Initial_Mark initial_mark_op(this);
1809 VMThread::execute(&initial_mark_op);
1810 }
1811 // The collector state may be any legal state at this point
1812 // since the background collector may have yielded to the
1813 // foreground collector.
1814 break;
1815 case Marking:
1816 // initial marking in checkpointRootsInitialWork has been completed
1817 if (markFromRoots()) { // we were successful
1818 assert(_collectorState == Precleaning, "Collector state should "
1819 "have changed");
1820 } else {
1821 assert(_foregroundGCIsActive, "Internal state inconsistency");
1822 }
1823 break;
1824 case Precleaning:
1825 // marking from roots in markFromRoots has been completed
1826 preclean();
1827 assert(_collectorState == AbortablePreclean ||
1828 _collectorState == FinalMarking,
1829 "Collector state should have changed");
1830 break;
1831 case AbortablePreclean:
1832 abortable_preclean();
1833 assert(_collectorState == FinalMarking, "Collector state should "
1834 "have changed");
1835 break;
1836 case FinalMarking:
1837 {
1838 ReleaseForegroundGC x(this);
1839
1840 VM_CMS_Final_Remark final_remark_op(this);
1841 VMThread::execute(&final_remark_op);
1842 }
1843 assert(_foregroundGCShouldWait, "block post-condition");
1844 break;
1845 case Sweeping:
1846 // final marking in checkpointRootsFinal has been completed
1847 sweep();
1848 assert(_collectorState == Resizing, "Collector state change "
1849 "to Resizing must be done under the free_list_lock");
1850
1851 case Resizing: {
1852 // Sweeping has been completed...
1853 // At this point the background collection has completed.
1854 // Don't move the call to compute_new_size() down
1855 // into code that might be executed if the background
1856 // collection was preempted.
1857 {
1858 ReleaseForegroundGC x(this); // unblock FG collection
1859 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1860 CMSTokenSync z(true); // not strictly needed.
1861 if (_collectorState == Resizing) {
1862 compute_new_size();
1863 save_heap_summary();
1864 _collectorState = Resetting;
1865 } else {
1866 assert(_collectorState == Idling, "The state should only change"
1867 " because the foreground collector has finished the collection");
1868 }
1869 }
1870 break;
1871 }
1872 case Resetting:
1873 // CMS heap resizing has been completed
1874 reset_concurrent();
1875 assert(_collectorState == Idling, "Collector state should "
1876 "have changed");
1877
1878 MetaspaceGC::set_should_concurrent_collect(false);
1879
1880 stats().record_cms_end();
1881 // Don't move the concurrent_phases_end() and compute_new_size()
1882 // calls to here because a preempted background collection
1883 // has it's state set to "Resetting".
1884 break;
1885 case Idling:
1886 default:
1887 ShouldNotReachHere();
1888 break;
1889 }
1890 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1891 p2i(Thread::current()), _collectorState);
1892 assert(_foregroundGCShouldWait, "block post-condition");
1893 }
1894
1895 // Should this be in gc_epilogue?
1896 heap->counters()->update_counters();
1897
1898 {
1899 // Clear _foregroundGCShouldWait and, in the event that the
1900 // foreground collector is waiting, notify it, before
1901 // returning.
1902 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1903 _foregroundGCShouldWait = false;
1904 if (_foregroundGCIsActive) {
1905 CGC_lock->notify();
1906 }
1907 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1908 "Possible deadlock");
1909 }
1910 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1911 p2i(Thread::current()), _collectorState);
1912 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)",
1913 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1914 }
1915
1916 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1917 _cms_start_registered = true;
1918 _gc_timer_cm->register_gc_start();
1919 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1920 }
1921
1922 void CMSCollector::register_gc_end() {
1923 if (_cms_start_registered) {
1924 report_heap_summary(GCWhen::AfterGC);
1925
1926 _gc_timer_cm->register_gc_end();
1927 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1928 _cms_start_registered = false;
1929 }
1930 }
1931
1932 void CMSCollector::save_heap_summary() {
1933 CMSHeap* heap = CMSHeap::heap();
1934 _last_heap_summary = heap->create_heap_summary();
1935 _last_metaspace_summary = heap->create_metaspace_summary();
1936 }
1937
1938 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1939 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1940 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1941 }
1942
1943 bool CMSCollector::waitForForegroundGC() {
1944 bool res = false;
1945 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1946 "CMS thread should have CMS token");
1947 // Block the foreground collector until the
1948 // background collectors decides whether to
1949 // yield.
1950 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1951 _foregroundGCShouldWait = true;
1952 if (_foregroundGCIsActive) {
1953 // The background collector yields to the
1954 // foreground collector and returns a value
1955 // indicating that it has yielded. The foreground
1956 // collector can proceed.
1957 res = true;
1958 _foregroundGCShouldWait = false;
1959 ConcurrentMarkSweepThread::clear_CMS_flag(
1960 ConcurrentMarkSweepThread::CMS_cms_has_token);
1961 ConcurrentMarkSweepThread::set_CMS_flag(
1962 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1963 // Get a possibly blocked foreground thread going
1964 CGC_lock->notify();
1965 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1966 p2i(Thread::current()), _collectorState);
1967 while (_foregroundGCIsActive) {
1968 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1969 }
1970 ConcurrentMarkSweepThread::set_CMS_flag(
1971 ConcurrentMarkSweepThread::CMS_cms_has_token);
1972 ConcurrentMarkSweepThread::clear_CMS_flag(
1973 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1974 }
1975 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1976 p2i(Thread::current()), _collectorState);
1977 return res;
1978 }
1979
1980 // Because of the need to lock the free lists and other structures in
1981 // the collector, common to all the generations that the collector is
1982 // collecting, we need the gc_prologues of individual CMS generations
1983 // delegate to their collector. It may have been simpler had the
1984 // current infrastructure allowed one to call a prologue on a
1985 // collector. In the absence of that we have the generation's
1986 // prologue delegate to the collector, which delegates back
1987 // some "local" work to a worker method in the individual generations
1988 // that it's responsible for collecting, while itself doing any
1989 // work common to all generations it's responsible for. A similar
1990 // comment applies to the gc_epilogue()'s.
1991 // The role of the variable _between_prologue_and_epilogue is to
1992 // enforce the invocation protocol.
1993 void CMSCollector::gc_prologue(bool full) {
1994 // Call gc_prologue_work() for the CMSGen
1995 // we are responsible for.
1996
1997 // The following locking discipline assumes that we are only called
1998 // when the world is stopped.
1999 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2000
2001 // The CMSCollector prologue must call the gc_prologues for the
2002 // "generations" that it's responsible
2003 // for.
2004
2005 assert( Thread::current()->is_VM_thread()
2006 || ( CMSScavengeBeforeRemark
2007 && Thread::current()->is_ConcurrentGC_thread()),
2008 "Incorrect thread type for prologue execution");
2009
2010 if (_between_prologue_and_epilogue) {
2011 // We have already been invoked; this is a gc_prologue delegation
2012 // from yet another CMS generation that we are responsible for, just
2013 // ignore it since all relevant work has already been done.
2014 return;
2015 }
2016
2017 // set a bit saying prologue has been called; cleared in epilogue
2018 _between_prologue_and_epilogue = true;
2019 // Claim locks for common data structures, then call gc_prologue_work()
2020 // for each CMSGen.
2021
2022 getFreelistLocks(); // gets free list locks on constituent spaces
2023 bitMapLock()->lock_without_safepoint_check();
2024
2025 // Should call gc_prologue_work() for all cms gens we are responsible for
2026 bool duringMarking = _collectorState >= Marking
2027 && _collectorState < Sweeping;
2028
2029 // The young collections clear the modified oops state, which tells if
2030 // there are any modified oops in the class. The remark phase also needs
2031 // that information. Tell the young collection to save the union of all
2032 // modified klasses.
2033 if (duringMarking) {
2034 _ct->cld_rem_set()->set_accumulate_modified_oops(true);
2035 }
2036
2037 bool registerClosure = duringMarking;
2038
2039 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2040
2041 if (!full) {
2042 stats().record_gc0_begin();
2043 }
2044 }
2045
2046 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2047
2048 _capacity_at_prologue = capacity();
2049 _used_at_prologue = used();
2050
2051 // We enable promotion tracking so that card-scanning can recognize
2052 // which objects have been promoted during this GC and skip them.
2053 for (uint i = 0; i < ParallelGCThreads; i++) {
2054 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2055 }
2056
2057 // Delegate to CMScollector which knows how to coordinate between
2058 // this and any other CMS generations that it is responsible for
2059 // collecting.
2060 collector()->gc_prologue(full);
2061 }
2062
2063 // This is a "private" interface for use by this generation's CMSCollector.
2064 // Not to be called directly by any other entity (for instance,
2065 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2066 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2067 bool registerClosure, ModUnionClosure* modUnionClosure) {
2068 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2069 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2070 "Should be NULL");
2071 if (registerClosure) {
2072 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2073 }
2074 cmsSpace()->gc_prologue();
2075 // Clear stat counters
2076 NOT_PRODUCT(
2077 assert(_numObjectsPromoted == 0, "check");
2078 assert(_numWordsPromoted == 0, "check");
2079 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2080 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2081 _numObjectsAllocated = 0;
2082 _numWordsAllocated = 0;
2083 )
2084 }
2085
2086 void CMSCollector::gc_epilogue(bool full) {
2087 // The following locking discipline assumes that we are only called
2088 // when the world is stopped.
2089 assert(SafepointSynchronize::is_at_safepoint(),
2090 "world is stopped assumption");
2091
2092 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2093 // if linear allocation blocks need to be appropriately marked to allow the
2094 // the blocks to be parsable. We also check here whether we need to nudge the
2095 // CMS collector thread to start a new cycle (if it's not already active).
2096 assert( Thread::current()->is_VM_thread()
2097 || ( CMSScavengeBeforeRemark
2098 && Thread::current()->is_ConcurrentGC_thread()),
2099 "Incorrect thread type for epilogue execution");
2100
2101 if (!_between_prologue_and_epilogue) {
2102 // We have already been invoked; this is a gc_epilogue delegation
2103 // from yet another CMS generation that we are responsible for, just
2104 // ignore it since all relevant work has already been done.
2105 return;
2106 }
2107 assert(haveFreelistLocks(), "must have freelist locks");
2108 assert_lock_strong(bitMapLock());
2109
2110 _ct->cld_rem_set()->set_accumulate_modified_oops(false);
2111
2112 _cmsGen->gc_epilogue_work(full);
2113
2114 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2115 // in case sampling was not already enabled, enable it
2116 _start_sampling = true;
2117 }
2118 // reset _eden_chunk_array so sampling starts afresh
2119 _eden_chunk_index = 0;
2120
2121 size_t cms_used = _cmsGen->cmsSpace()->used();
2122
2123 // update performance counters - this uses a special version of
2124 // update_counters() that allows the utilization to be passed as a
2125 // parameter, avoiding multiple calls to used().
2126 //
2127 _cmsGen->update_counters(cms_used);
2128
2129 bitMapLock()->unlock();
2130 releaseFreelistLocks();
2131
2132 if (!CleanChunkPoolAsync) {
2133 Chunk::clean_chunk_pool();
2134 }
2135
2136 set_did_compact(false);
2137 _between_prologue_and_epilogue = false; // ready for next cycle
2138 }
2139
2140 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2141 collector()->gc_epilogue(full);
2142
2143 // When using ParNew, promotion tracking should have already been
2144 // disabled. However, the prologue (which enables promotion
2145 // tracking) and epilogue are called irrespective of the type of
2146 // GC. So they will also be called before and after Full GCs, during
2147 // which promotion tracking will not be explicitly disabled. So,
2148 // it's safer to also disable it here too (to be symmetric with
2149 // enabling it in the prologue).
2150 for (uint i = 0; i < ParallelGCThreads; i++) {
2151 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2152 }
2153 }
2154
2155 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2156 assert(!incremental_collection_failed(), "Should have been cleared");
2157 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2158 cmsSpace()->gc_epilogue();
2159 // Print stat counters
2160 NOT_PRODUCT(
2161 assert(_numObjectsAllocated == 0, "check");
2162 assert(_numWordsAllocated == 0, "check");
2163 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2164 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2165 _numObjectsPromoted = 0;
2166 _numWordsPromoted = 0;
2167 )
2168
2169 // Call down the chain in contiguous_available needs the freelistLock
2170 // so print this out before releasing the freeListLock.
2171 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2172 }
2173
2174 #ifndef PRODUCT
2175 bool CMSCollector::have_cms_token() {
2176 Thread* thr = Thread::current();
2177 if (thr->is_VM_thread()) {
2178 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2179 } else if (thr->is_ConcurrentGC_thread()) {
2180 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2181 } else if (thr->is_GC_task_thread()) {
2182 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2183 ParGCRareEvent_lock->owned_by_self();
2184 }
2185 return false;
2186 }
2187
2188 // Check reachability of the given heap address in CMS generation,
2189 // treating all other generations as roots.
2190 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2191 // We could "guarantee" below, rather than assert, but I'll
2192 // leave these as "asserts" so that an adventurous debugger
2193 // could try this in the product build provided some subset of
2194 // the conditions were met, provided they were interested in the
2195 // results and knew that the computation below wouldn't interfere
2196 // with other concurrent computations mutating the structures
2197 // being read or written.
2198 assert(SafepointSynchronize::is_at_safepoint(),
2199 "Else mutations in object graph will make answer suspect");
2200 assert(have_cms_token(), "Should hold cms token");
2201 assert(haveFreelistLocks(), "must hold free list locks");
2202 assert_lock_strong(bitMapLock());
2203
2204 // Clear the marking bit map array before starting, but, just
2205 // for kicks, first report if the given address is already marked
2206 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2207 _markBitMap.isMarked(addr) ? "" : " not");
2208
2209 if (verify_after_remark()) {
2210 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2211 bool result = verification_mark_bm()->isMarked(addr);
2212 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2213 result ? "IS" : "is NOT");
2214 return result;
2215 } else {
2216 tty->print_cr("Could not compute result");
2217 return false;
2218 }
2219 }
2220 #endif
2221
2222 void
2223 CMSCollector::print_on_error(outputStream* st) {
2224 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2225 if (collector != NULL) {
2226 CMSBitMap* bitmap = &collector->_markBitMap;
2227 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2228 bitmap->print_on_error(st, " Bits: ");
2229
2230 st->cr();
2231
2232 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2233 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2234 mut_bitmap->print_on_error(st, " Bits: ");
2235 }
2236 }
2237
2238 ////////////////////////////////////////////////////////
2239 // CMS Verification Support
2240 ////////////////////////////////////////////////////////
2241 // Following the remark phase, the following invariant
2242 // should hold -- each object in the CMS heap which is
2243 // marked in markBitMap() should be marked in the verification_mark_bm().
2244
2245 class VerifyMarkedClosure: public BitMapClosure {
2246 CMSBitMap* _marks;
2247 bool _failed;
2248
2249 public:
2250 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2251
2252 bool do_bit(size_t offset) {
2253 HeapWord* addr = _marks->offsetToHeapWord(offset);
2254 if (!_marks->isMarked(addr)) {
2255 Log(gc, verify) log;
2256 ResourceMark rm;
2257 LogStream ls(log.error());
2258 oop(addr)->print_on(&ls);
2259 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2260 _failed = true;
2261 }
2262 return true;
2263 }
2264
2265 bool failed() { return _failed; }
2266 };
2267
2268 bool CMSCollector::verify_after_remark() {
2269 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2270 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2271 static bool init = false;
2272
2273 assert(SafepointSynchronize::is_at_safepoint(),
2274 "Else mutations in object graph will make answer suspect");
2275 assert(have_cms_token(),
2276 "Else there may be mutual interference in use of "
2277 " verification data structures");
2278 assert(_collectorState > Marking && _collectorState <= Sweeping,
2279 "Else marking info checked here may be obsolete");
2280 assert(haveFreelistLocks(), "must hold free list locks");
2281 assert_lock_strong(bitMapLock());
2282
2283
2284 // Allocate marking bit map if not already allocated
2285 if (!init) { // first time
2286 if (!verification_mark_bm()->allocate(_span)) {
2287 return false;
2288 }
2289 init = true;
2290 }
2291
2292 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2293
2294 // Turn off refs discovery -- so we will be tracing through refs.
2295 // This is as intended, because by this time
2296 // GC must already have cleared any refs that need to be cleared,
2297 // and traced those that need to be marked; moreover,
2298 // the marking done here is not going to interfere in any
2299 // way with the marking information used by GC.
2300 NoRefDiscovery no_discovery(ref_processor());
2301
2302 #if COMPILER2_OR_JVMCI
2303 DerivedPointerTableDeactivate dpt_deact;
2304 #endif
2305
2306 // Clear any marks from a previous round
2307 verification_mark_bm()->clear_all();
2308 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2309 verify_work_stacks_empty();
2310
2311 CMSHeap* heap = CMSHeap::heap();
2312 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2313 // Update the saved marks which may affect the root scans.
2314 heap->save_marks();
2315
2316 if (CMSRemarkVerifyVariant == 1) {
2317 // In this first variant of verification, we complete
2318 // all marking, then check if the new marks-vector is
2319 // a subset of the CMS marks-vector.
2320 verify_after_remark_work_1();
2321 } else {
2322 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2323 // In this second variant of verification, we flag an error
2324 // (i.e. an object reachable in the new marks-vector not reachable
2325 // in the CMS marks-vector) immediately, also indicating the
2326 // identify of an object (A) that references the unmarked object (B) --
2327 // presumably, a mutation to A failed to be picked up by preclean/remark?
2328 verify_after_remark_work_2();
2329 }
2330
2331 return true;
2332 }
2333
2334 void CMSCollector::verify_after_remark_work_1() {
2335 ResourceMark rm;
2336 HandleMark hm;
2337 CMSHeap* heap = CMSHeap::heap();
2338
2339 // Get a clear set of claim bits for the roots processing to work with.
2340 ClassLoaderDataGraph::clear_claimed_marks();
2341
2342 // Mark from roots one level into CMS
2343 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2344 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2345
2346 {
2347 StrongRootsScope srs(1);
2348
2349 heap->cms_process_roots(&srs,
2350 true, // young gen as roots
2351 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2352 should_unload_classes(),
2353 ¬Older,
2354 NULL);
2355 }
2356
2357 // Now mark from the roots
2358 MarkFromRootsClosure markFromRootsClosure(this, _span,
2359 verification_mark_bm(), verification_mark_stack(),
2360 false /* don't yield */, true /* verifying */);
2361 assert(_restart_addr == NULL, "Expected pre-condition");
2362 verification_mark_bm()->iterate(&markFromRootsClosure);
2363 while (_restart_addr != NULL) {
2364 // Deal with stack overflow: by restarting at the indicated
2365 // address.
2366 HeapWord* ra = _restart_addr;
2367 markFromRootsClosure.reset(ra);
2368 _restart_addr = NULL;
2369 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2370 }
2371 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2372 verify_work_stacks_empty();
2373
2374 // Marking completed -- now verify that each bit marked in
2375 // verification_mark_bm() is also marked in markBitMap(); flag all
2376 // errors by printing corresponding objects.
2377 VerifyMarkedClosure vcl(markBitMap());
2378 verification_mark_bm()->iterate(&vcl);
2379 if (vcl.failed()) {
2380 Log(gc, verify) log;
2381 log.error("Failed marking verification after remark");
2382 ResourceMark rm;
2383 LogStream ls(log.error());
2384 heap->print_on(&ls);
2385 fatal("CMS: failed marking verification after remark");
2386 }
2387 }
2388
2389 class VerifyCLDOopsCLDClosure : public CLDClosure {
2390 class VerifyCLDOopsClosure : public OopClosure {
2391 CMSBitMap* _bitmap;
2392 public:
2393 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2394 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2395 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2396 } _oop_closure;
2397 public:
2398 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2399 void do_cld(ClassLoaderData* cld) {
2400 cld->oops_do(&_oop_closure, false, false);
2401 }
2402 };
2403
2404 void CMSCollector::verify_after_remark_work_2() {
2405 ResourceMark rm;
2406 HandleMark hm;
2407 CMSHeap* heap = CMSHeap::heap();
2408
2409 // Get a clear set of claim bits for the roots processing to work with.
2410 ClassLoaderDataGraph::clear_claimed_marks();
2411
2412 // Mark from roots one level into CMS
2413 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2414 markBitMap());
2415 CLDToOopClosure cld_closure(¬Older, true);
2416
2417 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2418
2419 {
2420 StrongRootsScope srs(1);
2421
2422 heap->cms_process_roots(&srs,
2423 true, // young gen as roots
2424 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2425 should_unload_classes(),
2426 ¬Older,
2427 &cld_closure);
2428 }
2429
2430 // Now mark from the roots
2431 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2432 verification_mark_bm(), markBitMap(), verification_mark_stack());
2433 assert(_restart_addr == NULL, "Expected pre-condition");
2434 verification_mark_bm()->iterate(&markFromRootsClosure);
2435 while (_restart_addr != NULL) {
2436 // Deal with stack overflow: by restarting at the indicated
2437 // address.
2438 HeapWord* ra = _restart_addr;
2439 markFromRootsClosure.reset(ra);
2440 _restart_addr = NULL;
2441 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2442 }
2443 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2444 verify_work_stacks_empty();
2445
2446 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm());
2447 ClassLoaderDataGraph::cld_do(&verify_cld_oops);
2448
2449 // Marking completed -- now verify that each bit marked in
2450 // verification_mark_bm() is also marked in markBitMap(); flag all
2451 // errors by printing corresponding objects.
2452 VerifyMarkedClosure vcl(markBitMap());
2453 verification_mark_bm()->iterate(&vcl);
2454 assert(!vcl.failed(), "Else verification above should not have succeeded");
2455 }
2456
2457 void ConcurrentMarkSweepGeneration::save_marks() {
2458 // delegate to CMS space
2459 cmsSpace()->save_marks();
2460 }
2461
2462 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2463 return cmsSpace()->no_allocs_since_save_marks();
2464 }
2465
2466 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2467 \
2468 void ConcurrentMarkSweepGeneration:: \
2469 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2470 cl->set_generation(this); \
2471 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2472 cl->reset_generation(); \
2473 save_marks(); \
2474 }
2475
2476 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2477
2478 void
2479 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2480 if (freelistLock()->owned_by_self()) {
2481 Generation::oop_iterate(cl);
2482 } else {
2483 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2484 Generation::oop_iterate(cl);
2485 }
2486 }
2487
2488 void
2489 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2490 if (freelistLock()->owned_by_self()) {
2491 Generation::object_iterate(cl);
2492 } else {
2493 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2494 Generation::object_iterate(cl);
2495 }
2496 }
2497
2498 void
2499 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2500 if (freelistLock()->owned_by_self()) {
2501 Generation::safe_object_iterate(cl);
2502 } else {
2503 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2504 Generation::safe_object_iterate(cl);
2505 }
2506 }
2507
2508 void
2509 ConcurrentMarkSweepGeneration::post_compact() {
2510 }
2511
2512 void
2513 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2514 // Fix the linear allocation blocks to look like free blocks.
2515
2516 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2517 // are not called when the heap is verified during universe initialization and
2518 // at vm shutdown.
2519 if (freelistLock()->owned_by_self()) {
2520 cmsSpace()->prepare_for_verify();
2521 } else {
2522 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2523 cmsSpace()->prepare_for_verify();
2524 }
2525 }
2526
2527 void
2528 ConcurrentMarkSweepGeneration::verify() {
2529 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2530 // are not called when the heap is verified during universe initialization and
2531 // at vm shutdown.
2532 if (freelistLock()->owned_by_self()) {
2533 cmsSpace()->verify();
2534 } else {
2535 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2536 cmsSpace()->verify();
2537 }
2538 }
2539
2540 void CMSCollector::verify() {
2541 _cmsGen->verify();
2542 }
2543
2544 #ifndef PRODUCT
2545 bool CMSCollector::overflow_list_is_empty() const {
2546 assert(_num_par_pushes >= 0, "Inconsistency");
2547 if (_overflow_list == NULL) {
2548 assert(_num_par_pushes == 0, "Inconsistency");
2549 }
2550 return _overflow_list == NULL;
2551 }
2552
2553 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2554 // merely consolidate assertion checks that appear to occur together frequently.
2555 void CMSCollector::verify_work_stacks_empty() const {
2556 assert(_markStack.isEmpty(), "Marking stack should be empty");
2557 assert(overflow_list_is_empty(), "Overflow list should be empty");
2558 }
2559
2560 void CMSCollector::verify_overflow_empty() const {
2561 assert(overflow_list_is_empty(), "Overflow list should be empty");
2562 assert(no_preserved_marks(), "No preserved marks");
2563 }
2564 #endif // PRODUCT
2565
2566 // Decide if we want to enable class unloading as part of the
2567 // ensuing concurrent GC cycle. We will collect and
2568 // unload classes if it's the case that:
2569 // (a) class unloading is enabled at the command line, and
2570 // (b) old gen is getting really full
2571 // NOTE: Provided there is no change in the state of the heap between
2572 // calls to this method, it should have idempotent results. Moreover,
2573 // its results should be monotonically increasing (i.e. going from 0 to 1,
2574 // but not 1 to 0) between successive calls between which the heap was
2575 // not collected. For the implementation below, it must thus rely on
2576 // the property that concurrent_cycles_since_last_unload()
2577 // will not decrease unless a collection cycle happened and that
2578 // _cmsGen->is_too_full() are
2579 // themselves also monotonic in that sense. See check_monotonicity()
2580 // below.
2581 void CMSCollector::update_should_unload_classes() {
2582 _should_unload_classes = false;
2583 if (CMSClassUnloadingEnabled) {
2584 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2585 CMSClassUnloadingMaxInterval)
2586 || _cmsGen->is_too_full();
2587 }
2588 }
2589
2590 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2591 bool res = should_concurrent_collect();
2592 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2593 return res;
2594 }
2595
2596 void CMSCollector::setup_cms_unloading_and_verification_state() {
2597 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2598 || VerifyBeforeExit;
2599 const int rso = GenCollectedHeap::SO_AllCodeCache;
2600
2601 // We set the proper root for this CMS cycle here.
2602 if (should_unload_classes()) { // Should unload classes this cycle
2603 remove_root_scanning_option(rso); // Shrink the root set appropriately
2604 set_verifying(should_verify); // Set verification state for this cycle
2605 return; // Nothing else needs to be done at this time
2606 }
2607
2608 // Not unloading classes this cycle
2609 assert(!should_unload_classes(), "Inconsistency!");
2610
2611 // If we are not unloading classes then add SO_AllCodeCache to root
2612 // scanning options.
2613 add_root_scanning_option(rso);
2614
2615 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2616 set_verifying(true);
2617 } else if (verifying() && !should_verify) {
2618 // We were verifying, but some verification flags got disabled.
2619 set_verifying(false);
2620 // Exclude symbols, strings and code cache elements from root scanning to
2621 // reduce IM and RM pauses.
2622 remove_root_scanning_option(rso);
2623 }
2624 }
2625
2626
2627 #ifndef PRODUCT
2628 HeapWord* CMSCollector::block_start(const void* p) const {
2629 const HeapWord* addr = (HeapWord*)p;
2630 if (_span.contains(p)) {
2631 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2632 return _cmsGen->cmsSpace()->block_start(p);
2633 }
2634 }
2635 return NULL;
2636 }
2637 #endif
2638
2639 HeapWord*
2640 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2641 bool tlab,
2642 bool parallel) {
2643 CMSSynchronousYieldRequest yr;
2644 assert(!tlab, "Can't deal with TLAB allocation");
2645 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2646 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2647 if (GCExpandToAllocateDelayMillis > 0) {
2648 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2649 }
2650 return have_lock_and_allocate(word_size, tlab);
2651 }
2652
2653 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2654 size_t bytes,
2655 size_t expand_bytes,
2656 CMSExpansionCause::Cause cause)
2657 {
2658
2659 bool success = expand(bytes, expand_bytes);
2660
2661 // remember why we expanded; this information is used
2662 // by shouldConcurrentCollect() when making decisions on whether to start
2663 // a new CMS cycle.
2664 if (success) {
2665 set_expansion_cause(cause);
2666 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause));
2667 }
2668 }
2669
2670 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2671 HeapWord* res = NULL;
2672 MutexLocker x(ParGCRareEvent_lock);
2673 while (true) {
2674 // Expansion by some other thread might make alloc OK now:
2675 res = ps->lab.alloc(word_sz);
2676 if (res != NULL) return res;
2677 // If there's not enough expansion space available, give up.
2678 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2679 return NULL;
2680 }
2681 // Otherwise, we try expansion.
2682 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2683 // Now go around the loop and try alloc again;
2684 // A competing par_promote might beat us to the expansion space,
2685 // so we may go around the loop again if promotion fails again.
2686 if (GCExpandToAllocateDelayMillis > 0) {
2687 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2688 }
2689 }
2690 }
2691
2692
2693 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2694 PromotionInfo* promo) {
2695 MutexLocker x(ParGCRareEvent_lock);
2696 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2697 while (true) {
2698 // Expansion by some other thread might make alloc OK now:
2699 if (promo->ensure_spooling_space()) {
2700 assert(promo->has_spooling_space(),
2701 "Post-condition of successful ensure_spooling_space()");
2702 return true;
2703 }
2704 // If there's not enough expansion space available, give up.
2705 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2706 return false;
2707 }
2708 // Otherwise, we try expansion.
2709 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2710 // Now go around the loop and try alloc again;
2711 // A competing allocation might beat us to the expansion space,
2712 // so we may go around the loop again if allocation fails again.
2713 if (GCExpandToAllocateDelayMillis > 0) {
2714 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2715 }
2716 }
2717 }
2718
2719 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2720 // Only shrink if a compaction was done so that all the free space
2721 // in the generation is in a contiguous block at the end.
2722 if (did_compact()) {
2723 CardGeneration::shrink(bytes);
2724 }
2725 }
2726
2727 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2728 assert_locked_or_safepoint(Heap_lock);
2729 }
2730
2731 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2732 assert_locked_or_safepoint(Heap_lock);
2733 assert_lock_strong(freelistLock());
2734 log_trace(gc)("Shrinking of CMS not yet implemented");
2735 return;
2736 }
2737
2738
2739 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2740 // phases.
2741 class CMSPhaseAccounting: public StackObj {
2742 public:
2743 CMSPhaseAccounting(CMSCollector *collector,
2744 const char *title);
2745 ~CMSPhaseAccounting();
2746
2747 private:
2748 CMSCollector *_collector;
2749 const char *_title;
2750 GCTraceConcTime(Info, gc) _trace_time;
2751
2752 public:
2753 // Not MT-safe; so do not pass around these StackObj's
2754 // where they may be accessed by other threads.
2755 double wallclock_millis() {
2756 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2757 }
2758 };
2759
2760 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2761 const char *title) :
2762 _collector(collector), _title(title), _trace_time(title) {
2763
2764 _collector->resetYields();
2765 _collector->resetTimer();
2766 _collector->startTimer();
2767 _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2768 }
2769
2770 CMSPhaseAccounting::~CMSPhaseAccounting() {
2771 _collector->gc_timer_cm()->register_gc_concurrent_end();
2772 _collector->stopTimer();
2773 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2774 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2775 }
2776
2777 // CMS work
2778
2779 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2780 class CMSParMarkTask : public AbstractGangTask {
2781 protected:
2782 CMSCollector* _collector;
2783 uint _n_workers;
2784 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2785 AbstractGangTask(name),
2786 _collector(collector),
2787 _n_workers(n_workers) {}
2788 // Work method in support of parallel rescan ... of young gen spaces
2789 void do_young_space_rescan(OopsInGenClosure* cl,
2790 ContiguousSpace* space,
2791 HeapWord** chunk_array, size_t chunk_top);
2792 void work_on_young_gen_roots(OopsInGenClosure* cl);
2793 };
2794
2795 // Parallel initial mark task
2796 class CMSParInitialMarkTask: public CMSParMarkTask {
2797 StrongRootsScope* _strong_roots_scope;
2798 public:
2799 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2800 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2801 _strong_roots_scope(strong_roots_scope) {}
2802 void work(uint worker_id);
2803 };
2804
2805 // Checkpoint the roots into this generation from outside
2806 // this generation. [Note this initial checkpoint need only
2807 // be approximate -- we'll do a catch up phase subsequently.]
2808 void CMSCollector::checkpointRootsInitial() {
2809 assert(_collectorState == InitialMarking, "Wrong collector state");
2810 check_correct_thread_executing();
2811 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
2812
2813 save_heap_summary();
2814 report_heap_summary(GCWhen::BeforeGC);
2815
2816 ReferenceProcessor* rp = ref_processor();
2817 assert(_restart_addr == NULL, "Control point invariant");
2818 {
2819 // acquire locks for subsequent manipulations
2820 MutexLockerEx x(bitMapLock(),
2821 Mutex::_no_safepoint_check_flag);
2822 checkpointRootsInitialWork();
2823 // enable ("weak") refs discovery
2824 rp->enable_discovery();
2825 _collectorState = Marking;
2826 }
2827 }
2828
2829 void CMSCollector::checkpointRootsInitialWork() {
2830 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2831 assert(_collectorState == InitialMarking, "just checking");
2832
2833 // Already have locks.
2834 assert_lock_strong(bitMapLock());
2835 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2836
2837 // Setup the verification and class unloading state for this
2838 // CMS collection cycle.
2839 setup_cms_unloading_and_verification_state();
2840
2841 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2842
2843 // Reset all the PLAB chunk arrays if necessary.
2844 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2845 reset_survivor_plab_arrays();
2846 }
2847
2848 ResourceMark rm;
2849 HandleMark hm;
2850
2851 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2852 CMSHeap* heap = CMSHeap::heap();
2853
2854 verify_work_stacks_empty();
2855 verify_overflow_empty();
2856
2857 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2858 // Update the saved marks which may affect the root scans.
2859 heap->save_marks();
2860
2861 // weak reference processing has not started yet.
2862 ref_processor()->set_enqueuing_is_done(false);
2863
2864 // Need to remember all newly created CLDs,
2865 // so that we can guarantee that the remark finds them.
2866 ClassLoaderDataGraph::remember_new_clds(true);
2867
2868 // Whenever a CLD is found, it will be claimed before proceeding to mark
2869 // the klasses. The claimed marks need to be cleared before marking starts.
2870 ClassLoaderDataGraph::clear_claimed_marks();
2871
2872 print_eden_and_survivor_chunk_arrays();
2873
2874 {
2875 #if COMPILER2_OR_JVMCI
2876 DerivedPointerTableDeactivate dpt_deact;
2877 #endif
2878 if (CMSParallelInitialMarkEnabled) {
2879 // The parallel version.
2880 WorkGang* workers = heap->workers();
2881 assert(workers != NULL, "Need parallel worker threads.");
2882 uint n_workers = workers->active_workers();
2883
2884 StrongRootsScope srs(n_workers);
2885
2886 CMSParInitialMarkTask tsk(this, &srs, n_workers);
2887 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2888 // If the total workers is greater than 1, then multiple workers
2889 // may be used at some time and the initialization has been set
2890 // such that the single threaded path cannot be used.
2891 if (workers->total_workers() > 1) {
2892 workers->run_task(&tsk);
2893 } else {
2894 tsk.work(0);
2895 }
2896 } else {
2897 // The serial version.
2898 CLDToOopClosure cld_closure(¬Older, true);
2899 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2900
2901 StrongRootsScope srs(1);
2902
2903 heap->cms_process_roots(&srs,
2904 true, // young gen as roots
2905 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2906 should_unload_classes(),
2907 ¬Older,
2908 &cld_closure);
2909 }
2910 }
2911
2912 // Clear mod-union table; it will be dirtied in the prologue of
2913 // CMS generation per each young generation collection.
2914
2915 assert(_modUnionTable.isAllClear(),
2916 "Was cleared in most recent final checkpoint phase"
2917 " or no bits are set in the gc_prologue before the start of the next "
2918 "subsequent marking phase.");
2919
2920 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be");
2921
2922 // Save the end of the used_region of the constituent generations
2923 // to be used to limit the extent of sweep in each generation.
2924 save_sweep_limits();
2925 verify_overflow_empty();
2926 }
2927
2928 bool CMSCollector::markFromRoots() {
2929 // we might be tempted to assert that:
2930 // assert(!SafepointSynchronize::is_at_safepoint(),
2931 // "inconsistent argument?");
2932 // However that wouldn't be right, because it's possible that
2933 // a safepoint is indeed in progress as a young generation
2934 // stop-the-world GC happens even as we mark in this generation.
2935 assert(_collectorState == Marking, "inconsistent state?");
2936 check_correct_thread_executing();
2937 verify_overflow_empty();
2938
2939 // Weak ref discovery note: We may be discovering weak
2940 // refs in this generation concurrent (but interleaved) with
2941 // weak ref discovery by the young generation collector.
2942
2943 CMSTokenSyncWithLocks ts(true, bitMapLock());
2944 GCTraceCPUTime tcpu;
2945 CMSPhaseAccounting pa(this, "Concurrent Mark");
2946 bool res = markFromRootsWork();
2947 if (res) {
2948 _collectorState = Precleaning;
2949 } else { // We failed and a foreground collection wants to take over
2950 assert(_foregroundGCIsActive, "internal state inconsistency");
2951 assert(_restart_addr == NULL, "foreground will restart from scratch");
2952 log_debug(gc)("bailing out to foreground collection");
2953 }
2954 verify_overflow_empty();
2955 return res;
2956 }
2957
2958 bool CMSCollector::markFromRootsWork() {
2959 // iterate over marked bits in bit map, doing a full scan and mark
2960 // from these roots using the following algorithm:
2961 // . if oop is to the right of the current scan pointer,
2962 // mark corresponding bit (we'll process it later)
2963 // . else (oop is to left of current scan pointer)
2964 // push oop on marking stack
2965 // . drain the marking stack
2966
2967 // Note that when we do a marking step we need to hold the
2968 // bit map lock -- recall that direct allocation (by mutators)
2969 // and promotion (by the young generation collector) is also
2970 // marking the bit map. [the so-called allocate live policy.]
2971 // Because the implementation of bit map marking is not
2972 // robust wrt simultaneous marking of bits in the same word,
2973 // we need to make sure that there is no such interference
2974 // between concurrent such updates.
2975
2976 // already have locks
2977 assert_lock_strong(bitMapLock());
2978
2979 verify_work_stacks_empty();
2980 verify_overflow_empty();
2981 bool result = false;
2982 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2983 result = do_marking_mt();
2984 } else {
2985 result = do_marking_st();
2986 }
2987 return result;
2988 }
2989
2990 // Forward decl
2991 class CMSConcMarkingTask;
2992
2993 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2994 CMSCollector* _collector;
2995 CMSConcMarkingTask* _task;
2996 public:
2997 virtual void yield();
2998
2999 // "n_threads" is the number of threads to be terminated.
3000 // "queue_set" is a set of work queues of other threads.
3001 // "collector" is the CMS collector associated with this task terminator.
3002 // "yield" indicates whether we need the gang as a whole to yield.
3003 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3004 ParallelTaskTerminator(n_threads, queue_set),
3005 _collector(collector) { }
3006
3007 void set_task(CMSConcMarkingTask* task) {
3008 _task = task;
3009 }
3010 };
3011
3012 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3013 CMSConcMarkingTask* _task;
3014 public:
3015 bool should_exit_termination();
3016 void set_task(CMSConcMarkingTask* task) {
3017 _task = task;
3018 }
3019 };
3020
3021 // MT Concurrent Marking Task
3022 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3023 CMSCollector* _collector;
3024 uint _n_workers; // requested/desired # workers
3025 bool _result;
3026 CompactibleFreeListSpace* _cms_space;
3027 char _pad_front[64]; // padding to ...
3028 HeapWord* volatile _global_finger; // ... avoid sharing cache line
3029 char _pad_back[64];
3030 HeapWord* _restart_addr;
3031
3032 // Exposed here for yielding support
3033 Mutex* const _bit_map_lock;
3034
3035 // The per thread work queues, available here for stealing
3036 OopTaskQueueSet* _task_queues;
3037
3038 // Termination (and yielding) support
3039 CMSConcMarkingTerminator _term;
3040 CMSConcMarkingTerminatorTerminator _term_term;
3041
3042 public:
3043 CMSConcMarkingTask(CMSCollector* collector,
3044 CompactibleFreeListSpace* cms_space,
3045 YieldingFlexibleWorkGang* workers,
3046 OopTaskQueueSet* task_queues):
3047 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3048 _collector(collector),
3049 _cms_space(cms_space),
3050 _n_workers(0), _result(true),
3051 _task_queues(task_queues),
3052 _term(_n_workers, task_queues, _collector),
3053 _bit_map_lock(collector->bitMapLock())
3054 {
3055 _requested_size = _n_workers;
3056 _term.set_task(this);
3057 _term_term.set_task(this);
3058 _restart_addr = _global_finger = _cms_space->bottom();
3059 }
3060
3061
3062 OopTaskQueueSet* task_queues() { return _task_queues; }
3063
3064 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3065
3066 HeapWord* volatile* global_finger_addr() { return &_global_finger; }
3067
3068 CMSConcMarkingTerminator* terminator() { return &_term; }
3069
3070 virtual void set_for_termination(uint active_workers) {
3071 terminator()->reset_for_reuse(active_workers);
3072 }
3073
3074 void work(uint worker_id);
3075 bool should_yield() {
3076 return ConcurrentMarkSweepThread::should_yield()
3077 && !_collector->foregroundGCIsActive();
3078 }
3079
3080 virtual void coordinator_yield(); // stuff done by coordinator
3081 bool result() { return _result; }
3082
3083 void reset(HeapWord* ra) {
3084 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3085 _restart_addr = _global_finger = ra;
3086 _term.reset_for_reuse();
3087 }
3088
3089 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3090 OopTaskQueue* work_q);
3091
3092 private:
3093 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3094 void do_work_steal(int i);
3095 void bump_global_finger(HeapWord* f);
3096 };
3097
3098 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3099 assert(_task != NULL, "Error");
3100 return _task->yielding();
3101 // Note that we do not need the disjunct || _task->should_yield() above
3102 // because we want terminating threads to yield only if the task
3103 // is already in the midst of yielding, which happens only after at least one
3104 // thread has yielded.
3105 }
3106
3107 void CMSConcMarkingTerminator::yield() {
3108 if (_task->should_yield()) {
3109 _task->yield();
3110 } else {
3111 ParallelTaskTerminator::yield();
3112 }
3113 }
3114
3115 ////////////////////////////////////////////////////////////////
3116 // Concurrent Marking Algorithm Sketch
3117 ////////////////////////////////////////////////////////////////
3118 // Until all tasks exhausted (both spaces):
3119 // -- claim next available chunk
3120 // -- bump global finger via CAS
3121 // -- find first object that starts in this chunk
3122 // and start scanning bitmap from that position
3123 // -- scan marked objects for oops
3124 // -- CAS-mark target, and if successful:
3125 // . if target oop is above global finger (volatile read)
3126 // nothing to do
3127 // . if target oop is in chunk and above local finger
3128 // then nothing to do
3129 // . else push on work-queue
3130 // -- Deal with possible overflow issues:
3131 // . local work-queue overflow causes stuff to be pushed on
3132 // global (common) overflow queue
3133 // . always first empty local work queue
3134 // . then get a batch of oops from global work queue if any
3135 // . then do work stealing
3136 // -- When all tasks claimed (both spaces)
3137 // and local work queue empty,
3138 // then in a loop do:
3139 // . check global overflow stack; steal a batch of oops and trace
3140 // . try to steal from other threads oif GOS is empty
3141 // . if neither is available, offer termination
3142 // -- Terminate and return result
3143 //
3144 void CMSConcMarkingTask::work(uint worker_id) {
3145 elapsedTimer _timer;
3146 ResourceMark rm;
3147 HandleMark hm;
3148
3149 DEBUG_ONLY(_collector->verify_overflow_empty();)
3150
3151 // Before we begin work, our work queue should be empty
3152 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3153 // Scan the bitmap covering _cms_space, tracing through grey objects.
3154 _timer.start();
3155 do_scan_and_mark(worker_id, _cms_space);
3156 _timer.stop();
3157 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3158
3159 // ... do work stealing
3160 _timer.reset();
3161 _timer.start();
3162 do_work_steal(worker_id);
3163 _timer.stop();
3164 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3165 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3166 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3167 // Note that under the current task protocol, the
3168 // following assertion is true even of the spaces
3169 // expanded since the completion of the concurrent
3170 // marking. XXX This will likely change under a strict
3171 // ABORT semantics.
3172 // After perm removal the comparison was changed to
3173 // greater than or equal to from strictly greater than.
3174 // Before perm removal the highest address sweep would
3175 // have been at the end of perm gen but now is at the
3176 // end of the tenured gen.
3177 assert(_global_finger >= _cms_space->end(),
3178 "All tasks have been completed");
3179 DEBUG_ONLY(_collector->verify_overflow_empty();)
3180 }
3181
3182 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3183 HeapWord* read = _global_finger;
3184 HeapWord* cur = read;
3185 while (f > read) {
3186 cur = read;
3187 read = Atomic::cmpxchg(f, &_global_finger, cur);
3188 if (cur == read) {
3189 // our cas succeeded
3190 assert(_global_finger >= f, "protocol consistency");
3191 break;
3192 }
3193 }
3194 }
3195
3196 // This is really inefficient, and should be redone by
3197 // using (not yet available) block-read and -write interfaces to the
3198 // stack and the work_queue. XXX FIX ME !!!
3199 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3200 OopTaskQueue* work_q) {
3201 // Fast lock-free check
3202 if (ovflw_stk->length() == 0) {
3203 return false;
3204 }
3205 assert(work_q->size() == 0, "Shouldn't steal");
3206 MutexLockerEx ml(ovflw_stk->par_lock(),
3207 Mutex::_no_safepoint_check_flag);
3208 // Grab up to 1/4 the size of the work queue
3209 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3210 (size_t)ParGCDesiredObjsFromOverflowList);
3211 num = MIN2(num, ovflw_stk->length());
3212 for (int i = (int) num; i > 0; i--) {
3213 oop cur = ovflw_stk->pop();
3214 assert(cur != NULL, "Counted wrong?");
3215 work_q->push(cur);
3216 }
3217 return num > 0;
3218 }
3219
3220 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3221 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3222 int n_tasks = pst->n_tasks();
3223 // We allow that there may be no tasks to do here because
3224 // we are restarting after a stack overflow.
3225 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3226 uint nth_task = 0;
3227
3228 HeapWord* aligned_start = sp->bottom();
3229 if (sp->used_region().contains(_restart_addr)) {
3230 // Align down to a card boundary for the start of 0th task
3231 // for this space.
3232 aligned_start = align_down(_restart_addr, CardTable::card_size);
3233 }
3234
3235 size_t chunk_size = sp->marking_task_size();
3236 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3237 // Having claimed the nth task in this space,
3238 // compute the chunk that it corresponds to:
3239 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3240 aligned_start + (nth_task+1)*chunk_size);
3241 // Try and bump the global finger via a CAS;
3242 // note that we need to do the global finger bump
3243 // _before_ taking the intersection below, because
3244 // the task corresponding to that region will be
3245 // deemed done even if the used_region() expands
3246 // because of allocation -- as it almost certainly will
3247 // during start-up while the threads yield in the
3248 // closure below.
3249 HeapWord* finger = span.end();
3250 bump_global_finger(finger); // atomically
3251 // There are null tasks here corresponding to chunks
3252 // beyond the "top" address of the space.
3253 span = span.intersection(sp->used_region());
3254 if (!span.is_empty()) { // Non-null task
3255 HeapWord* prev_obj;
3256 assert(!span.contains(_restart_addr) || nth_task == 0,
3257 "Inconsistency");
3258 if (nth_task == 0) {
3259 // For the 0th task, we'll not need to compute a block_start.
3260 if (span.contains(_restart_addr)) {
3261 // In the case of a restart because of stack overflow,
3262 // we might additionally skip a chunk prefix.
3263 prev_obj = _restart_addr;
3264 } else {
3265 prev_obj = span.start();
3266 }
3267 } else {
3268 // We want to skip the first object because
3269 // the protocol is to scan any object in its entirety
3270 // that _starts_ in this span; a fortiori, any
3271 // object starting in an earlier span is scanned
3272 // as part of an earlier claimed task.
3273 // Below we use the "careful" version of block_start
3274 // so we do not try to navigate uninitialized objects.
3275 prev_obj = sp->block_start_careful(span.start());
3276 // Below we use a variant of block_size that uses the
3277 // Printezis bits to avoid waiting for allocated
3278 // objects to become initialized/parsable.
3279 while (prev_obj < span.start()) {
3280 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3281 if (sz > 0) {
3282 prev_obj += sz;
3283 } else {
3284 // In this case we may end up doing a bit of redundant
3285 // scanning, but that appears unavoidable, short of
3286 // locking the free list locks; see bug 6324141.
3287 break;
3288 }
3289 }
3290 }
3291 if (prev_obj < span.end()) {
3292 MemRegion my_span = MemRegion(prev_obj, span.end());
3293 // Do the marking work within a non-empty span --
3294 // the last argument to the constructor indicates whether the
3295 // iteration should be incremental with periodic yields.
3296 ParMarkFromRootsClosure cl(this, _collector, my_span,
3297 &_collector->_markBitMap,
3298 work_queue(i),
3299 &_collector->_markStack);
3300 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3301 } // else nothing to do for this task
3302 } // else nothing to do for this task
3303 }
3304 // We'd be tempted to assert here that since there are no
3305 // more tasks left to claim in this space, the global_finger
3306 // must exceed space->top() and a fortiori space->end(). However,
3307 // that would not quite be correct because the bumping of
3308 // global_finger occurs strictly after the claiming of a task,
3309 // so by the time we reach here the global finger may not yet
3310 // have been bumped up by the thread that claimed the last
3311 // task.
3312 pst->all_tasks_completed();
3313 }
3314
3315 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3316 private:
3317 CMSCollector* _collector;
3318 CMSConcMarkingTask* _task;
3319 MemRegion _span;
3320 CMSBitMap* _bit_map;
3321 CMSMarkStack* _overflow_stack;
3322 OopTaskQueue* _work_queue;
3323 protected:
3324 DO_OOP_WORK_DEFN
3325 public:
3326 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3327 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3328 MetadataAwareOopClosure(collector->ref_processor()),
3329 _collector(collector),
3330 _task(task),
3331 _span(collector->_span),
3332 _work_queue(work_queue),
3333 _bit_map(bit_map),
3334 _overflow_stack(overflow_stack)
3335 { }
3336 virtual void do_oop(oop* p);
3337 virtual void do_oop(narrowOop* p);
3338
3339 void trim_queue(size_t max);
3340 void handle_stack_overflow(HeapWord* lost);
3341 void do_yield_check() {
3342 if (_task->should_yield()) {
3343 _task->yield();
3344 }
3345 }
3346 };
3347
3348 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3349
3350 // Grey object scanning during work stealing phase --
3351 // the salient assumption here is that any references
3352 // that are in these stolen objects being scanned must
3353 // already have been initialized (else they would not have
3354 // been published), so we do not need to check for
3355 // uninitialized objects before pushing here.
3356 void ParConcMarkingClosure::do_oop(oop obj) {
3357 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3358 HeapWord* addr = (HeapWord*)obj;
3359 // Check if oop points into the CMS generation
3360 // and is not marked
3361 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3362 // a white object ...
3363 // If we manage to "claim" the object, by being the
3364 // first thread to mark it, then we push it on our
3365 // marking stack
3366 if (_bit_map->par_mark(addr)) { // ... now grey
3367 // push on work queue (grey set)
3368 bool simulate_overflow = false;
3369 NOT_PRODUCT(
3370 if (CMSMarkStackOverflowALot &&
3371 _collector->simulate_overflow()) {
3372 // simulate a stack overflow
3373 simulate_overflow = true;
3374 }
3375 )
3376 if (simulate_overflow ||
3377 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3378 // stack overflow
3379 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3380 // We cannot assert that the overflow stack is full because
3381 // it may have been emptied since.
3382 assert(simulate_overflow ||
3383 _work_queue->size() == _work_queue->max_elems(),
3384 "Else push should have succeeded");
3385 handle_stack_overflow(addr);
3386 }
3387 } // Else, some other thread got there first
3388 do_yield_check();
3389 }
3390 }
3391
3392 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); }
3393 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3394
3395 void ParConcMarkingClosure::trim_queue(size_t max) {
3396 while (_work_queue->size() > max) {
3397 oop new_oop;
3398 if (_work_queue->pop_local(new_oop)) {
3399 assert(oopDesc::is_oop(new_oop), "Should be an oop");
3400 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3401 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3402 new_oop->oop_iterate(this); // do_oop() above
3403 do_yield_check();
3404 }
3405 }
3406 }
3407
3408 // Upon stack overflow, we discard (part of) the stack,
3409 // remembering the least address amongst those discarded
3410 // in CMSCollector's _restart_address.
3411 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3412 // We need to do this under a mutex to prevent other
3413 // workers from interfering with the work done below.
3414 MutexLockerEx ml(_overflow_stack->par_lock(),
3415 Mutex::_no_safepoint_check_flag);
3416 // Remember the least grey address discarded
3417 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3418 _collector->lower_restart_addr(ra);
3419 _overflow_stack->reset(); // discard stack contents
3420 _overflow_stack->expand(); // expand the stack if possible
3421 }
3422
3423
3424 void CMSConcMarkingTask::do_work_steal(int i) {
3425 OopTaskQueue* work_q = work_queue(i);
3426 oop obj_to_scan;
3427 CMSBitMap* bm = &(_collector->_markBitMap);
3428 CMSMarkStack* ovflw = &(_collector->_markStack);
3429 int* seed = _collector->hash_seed(i);
3430 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3431 while (true) {
3432 cl.trim_queue(0);
3433 assert(work_q->size() == 0, "Should have been emptied above");
3434 if (get_work_from_overflow_stack(ovflw, work_q)) {
3435 // Can't assert below because the work obtained from the
3436 // overflow stack may already have been stolen from us.
3437 // assert(work_q->size() > 0, "Work from overflow stack");
3438 continue;
3439 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3440 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3441 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3442 obj_to_scan->oop_iterate(&cl);
3443 } else if (terminator()->offer_termination(&_term_term)) {
3444 assert(work_q->size() == 0, "Impossible!");
3445 break;
3446 } else if (yielding() || should_yield()) {
3447 yield();
3448 }
3449 }
3450 }
3451
3452 // This is run by the CMS (coordinator) thread.
3453 void CMSConcMarkingTask::coordinator_yield() {
3454 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3455 "CMS thread should hold CMS token");
3456 // First give up the locks, then yield, then re-lock
3457 // We should probably use a constructor/destructor idiom to
3458 // do this unlock/lock or modify the MutexUnlocker class to
3459 // serve our purpose. XXX
3460 assert_lock_strong(_bit_map_lock);
3461 _bit_map_lock->unlock();
3462 ConcurrentMarkSweepThread::desynchronize(true);
3463 _collector->stopTimer();
3464 _collector->incrementYields();
3465
3466 // It is possible for whichever thread initiated the yield request
3467 // not to get a chance to wake up and take the bitmap lock between
3468 // this thread releasing it and reacquiring it. So, while the
3469 // should_yield() flag is on, let's sleep for a bit to give the
3470 // other thread a chance to wake up. The limit imposed on the number
3471 // of iterations is defensive, to avoid any unforseen circumstances
3472 // putting us into an infinite loop. Since it's always been this
3473 // (coordinator_yield()) method that was observed to cause the
3474 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3475 // which is by default non-zero. For the other seven methods that
3476 // also perform the yield operation, as are using a different
3477 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3478 // can enable the sleeping for those methods too, if necessary.
3479 // See 6442774.
3480 //
3481 // We really need to reconsider the synchronization between the GC
3482 // thread and the yield-requesting threads in the future and we
3483 // should really use wait/notify, which is the recommended
3484 // way of doing this type of interaction. Additionally, we should
3485 // consolidate the eight methods that do the yield operation and they
3486 // are almost identical into one for better maintainability and
3487 // readability. See 6445193.
3488 //
3489 // Tony 2006.06.29
3490 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3491 ConcurrentMarkSweepThread::should_yield() &&
3492 !CMSCollector::foregroundGCIsActive(); ++i) {
3493 os::sleep(Thread::current(), 1, false);
3494 }
3495
3496 ConcurrentMarkSweepThread::synchronize(true);
3497 _bit_map_lock->lock_without_safepoint_check();
3498 _collector->startTimer();
3499 }
3500
3501 bool CMSCollector::do_marking_mt() {
3502 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3503 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3504 conc_workers()->active_workers(),
3505 Threads::number_of_non_daemon_threads());
3506 num_workers = conc_workers()->update_active_workers(num_workers);
3507 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3508
3509 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3510
3511 CMSConcMarkingTask tsk(this,
3512 cms_space,
3513 conc_workers(),
3514 task_queues());
3515
3516 // Since the actual number of workers we get may be different
3517 // from the number we requested above, do we need to do anything different
3518 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3519 // class?? XXX
3520 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3521
3522 // Refs discovery is already non-atomic.
3523 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3524 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3525 conc_workers()->start_task(&tsk);
3526 while (tsk.yielded()) {
3527 tsk.coordinator_yield();
3528 conc_workers()->continue_task(&tsk);
3529 }
3530 // If the task was aborted, _restart_addr will be non-NULL
3531 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3532 while (_restart_addr != NULL) {
3533 // XXX For now we do not make use of ABORTED state and have not
3534 // yet implemented the right abort semantics (even in the original
3535 // single-threaded CMS case). That needs some more investigation
3536 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3537 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3538 // If _restart_addr is non-NULL, a marking stack overflow
3539 // occurred; we need to do a fresh marking iteration from the
3540 // indicated restart address.
3541 if (_foregroundGCIsActive) {
3542 // We may be running into repeated stack overflows, having
3543 // reached the limit of the stack size, while making very
3544 // slow forward progress. It may be best to bail out and
3545 // let the foreground collector do its job.
3546 // Clear _restart_addr, so that foreground GC
3547 // works from scratch. This avoids the headache of
3548 // a "rescan" which would otherwise be needed because
3549 // of the dirty mod union table & card table.
3550 _restart_addr = NULL;
3551 return false;
3552 }
3553 // Adjust the task to restart from _restart_addr
3554 tsk.reset(_restart_addr);
3555 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3556 _restart_addr);
3557 _restart_addr = NULL;
3558 // Get the workers going again
3559 conc_workers()->start_task(&tsk);
3560 while (tsk.yielded()) {
3561 tsk.coordinator_yield();
3562 conc_workers()->continue_task(&tsk);
3563 }
3564 }
3565 assert(tsk.completed(), "Inconsistency");
3566 assert(tsk.result() == true, "Inconsistency");
3567 return true;
3568 }
3569
3570 bool CMSCollector::do_marking_st() {
3571 ResourceMark rm;
3572 HandleMark hm;
3573
3574 // Temporarily make refs discovery single threaded (non-MT)
3575 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3576 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3577 &_markStack, CMSYield);
3578 // the last argument to iterate indicates whether the iteration
3579 // should be incremental with periodic yields.
3580 _markBitMap.iterate(&markFromRootsClosure);
3581 // If _restart_addr is non-NULL, a marking stack overflow
3582 // occurred; we need to do a fresh iteration from the
3583 // indicated restart address.
3584 while (_restart_addr != NULL) {
3585 if (_foregroundGCIsActive) {
3586 // We may be running into repeated stack overflows, having
3587 // reached the limit of the stack size, while making very
3588 // slow forward progress. It may be best to bail out and
3589 // let the foreground collector do its job.
3590 // Clear _restart_addr, so that foreground GC
3591 // works from scratch. This avoids the headache of
3592 // a "rescan" which would otherwise be needed because
3593 // of the dirty mod union table & card table.
3594 _restart_addr = NULL;
3595 return false; // indicating failure to complete marking
3596 }
3597 // Deal with stack overflow:
3598 // we restart marking from _restart_addr
3599 HeapWord* ra = _restart_addr;
3600 markFromRootsClosure.reset(ra);
3601 _restart_addr = NULL;
3602 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3603 }
3604 return true;
3605 }
3606
3607 void CMSCollector::preclean() {
3608 check_correct_thread_executing();
3609 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3610 verify_work_stacks_empty();
3611 verify_overflow_empty();
3612 _abort_preclean = false;
3613 if (CMSPrecleaningEnabled) {
3614 if (!CMSEdenChunksRecordAlways) {
3615 _eden_chunk_index = 0;
3616 }
3617 size_t used = get_eden_used();
3618 size_t capacity = get_eden_capacity();
3619 // Don't start sampling unless we will get sufficiently
3620 // many samples.
3621 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3622 * CMSScheduleRemarkEdenPenetration)) {
3623 _start_sampling = true;
3624 } else {
3625 _start_sampling = false;
3626 }
3627 GCTraceCPUTime tcpu;
3628 CMSPhaseAccounting pa(this, "Concurrent Preclean");
3629 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3630 }
3631 CMSTokenSync x(true); // is cms thread
3632 if (CMSPrecleaningEnabled) {
3633 sample_eden();
3634 _collectorState = AbortablePreclean;
3635 } else {
3636 _collectorState = FinalMarking;
3637 }
3638 verify_work_stacks_empty();
3639 verify_overflow_empty();
3640 }
3641
3642 // Try and schedule the remark such that young gen
3643 // occupancy is CMSScheduleRemarkEdenPenetration %.
3644 void CMSCollector::abortable_preclean() {
3645 check_correct_thread_executing();
3646 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3647 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3648
3649 // If Eden's current occupancy is below this threshold,
3650 // immediately schedule the remark; else preclean
3651 // past the next scavenge in an effort to
3652 // schedule the pause as described above. By choosing
3653 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3654 // we will never do an actual abortable preclean cycle.
3655 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3656 GCTraceCPUTime tcpu;
3657 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3658 // We need more smarts in the abortable preclean
3659 // loop below to deal with cases where allocation
3660 // in young gen is very very slow, and our precleaning
3661 // is running a losing race against a horde of
3662 // mutators intent on flooding us with CMS updates
3663 // (dirty cards).
3664 // One, admittedly dumb, strategy is to give up
3665 // after a certain number of abortable precleaning loops
3666 // or after a certain maximum time. We want to make
3667 // this smarter in the next iteration.
3668 // XXX FIX ME!!! YSR
3669 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3670 while (!(should_abort_preclean() ||
3671 ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3672 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3673 cumworkdone += workdone;
3674 loops++;
3675 // Voluntarily terminate abortable preclean phase if we have
3676 // been at it for too long.
3677 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3678 loops >= CMSMaxAbortablePrecleanLoops) {
3679 log_debug(gc)(" CMS: abort preclean due to loops ");
3680 break;
3681 }
3682 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3683 log_debug(gc)(" CMS: abort preclean due to time ");
3684 break;
3685 }
3686 // If we are doing little work each iteration, we should
3687 // take a short break.
3688 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3689 // Sleep for some time, waiting for work to accumulate
3690 stopTimer();
3691 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3692 startTimer();
3693 waited++;
3694 }
3695 }
3696 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3697 loops, waited, cumworkdone);
3698 }
3699 CMSTokenSync x(true); // is cms thread
3700 if (_collectorState != Idling) {
3701 assert(_collectorState == AbortablePreclean,
3702 "Spontaneous state transition?");
3703 _collectorState = FinalMarking;
3704 } // Else, a foreground collection completed this CMS cycle.
3705 return;
3706 }
3707
3708 // Respond to an Eden sampling opportunity
3709 void CMSCollector::sample_eden() {
3710 // Make sure a young gc cannot sneak in between our
3711 // reading and recording of a sample.
3712 assert(Thread::current()->is_ConcurrentGC_thread(),
3713 "Only the cms thread may collect Eden samples");
3714 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3715 "Should collect samples while holding CMS token");
3716 if (!_start_sampling) {
3717 return;
3718 }
3719 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3720 // is populated by the young generation.
3721 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3722 if (_eden_chunk_index < _eden_chunk_capacity) {
3723 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3724 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3725 "Unexpected state of Eden");
3726 // We'd like to check that what we just sampled is an oop-start address;
3727 // however, we cannot do that here since the object may not yet have been
3728 // initialized. So we'll instead do the check when we _use_ this sample
3729 // later.
3730 if (_eden_chunk_index == 0 ||
3731 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3732 _eden_chunk_array[_eden_chunk_index-1])
3733 >= CMSSamplingGrain)) {
3734 _eden_chunk_index++; // commit sample
3735 }
3736 }
3737 }
3738 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3739 size_t used = get_eden_used();
3740 size_t capacity = get_eden_capacity();
3741 assert(used <= capacity, "Unexpected state of Eden");
3742 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3743 _abort_preclean = true;
3744 }
3745 }
3746 }
3747
3748 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3749 assert(_collectorState == Precleaning ||
3750 _collectorState == AbortablePreclean, "incorrect state");
3751 ResourceMark rm;
3752 HandleMark hm;
3753
3754 // Precleaning is currently not MT but the reference processor
3755 // may be set for MT. Disable it temporarily here.
3756 ReferenceProcessor* rp = ref_processor();
3757 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3758
3759 // Do one pass of scrubbing the discovered reference lists
3760 // to remove any reference objects with strongly-reachable
3761 // referents.
3762 if (clean_refs) {
3763 CMSPrecleanRefsYieldClosure yield_cl(this);
3764 assert(_span_discoverer.span().equals(_span), "Spans should be equal");
3765 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3766 &_markStack, true /* preclean */);
3767 CMSDrainMarkingStackClosure complete_trace(this,
3768 _span, &_markBitMap, &_markStack,
3769 &keep_alive, true /* preclean */);
3770
3771 // We don't want this step to interfere with a young
3772 // collection because we don't want to take CPU
3773 // or memory bandwidth away from the young GC threads
3774 // (which may be as many as there are CPUs).
3775 // Note that we don't need to protect ourselves from
3776 // interference with mutators because they can't
3777 // manipulate the discovered reference lists nor affect
3778 // the computed reachability of the referents, the
3779 // only properties manipulated by the precleaning
3780 // of these reference lists.
3781 stopTimer();
3782 CMSTokenSyncWithLocks x(true /* is cms thread */,
3783 bitMapLock());
3784 startTimer();
3785 sample_eden();
3786
3787 // The following will yield to allow foreground
3788 // collection to proceed promptly. XXX YSR:
3789 // The code in this method may need further
3790 // tweaking for better performance and some restructuring
3791 // for cleaner interfaces.
3792 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3793 rp->preclean_discovered_references(
3794 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3795 gc_timer);
3796 }
3797
3798 if (clean_survivor) { // preclean the active survivor space(s)
3799 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3800 &_markBitMap, &_modUnionTable,
3801 &_markStack, true /* precleaning phase */);
3802 stopTimer();
3803 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3804 bitMapLock());
3805 startTimer();
3806 unsigned int before_count =
3807 CMSHeap::heap()->total_collections();
3808 SurvivorSpacePrecleanClosure
3809 sss_cl(this, _span, &_markBitMap, &_markStack,
3810 &pam_cl, before_count, CMSYield);
3811 _young_gen->from()->object_iterate_careful(&sss_cl);
3812 _young_gen->to()->object_iterate_careful(&sss_cl);
3813 }
3814 MarkRefsIntoAndScanClosure
3815 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3816 &_markStack, this, CMSYield,
3817 true /* precleaning phase */);
3818 // CAUTION: The following closure has persistent state that may need to
3819 // be reset upon a decrease in the sequence of addresses it
3820 // processes.
3821 ScanMarkedObjectsAgainCarefullyClosure
3822 smoac_cl(this, _span,
3823 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3824
3825 // Preclean dirty cards in ModUnionTable and CardTable using
3826 // appropriate convergence criterion;
3827 // repeat CMSPrecleanIter times unless we find that
3828 // we are losing.
3829 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3830 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3831 "Bad convergence multiplier");
3832 assert(CMSPrecleanThreshold >= 100,
3833 "Unreasonably low CMSPrecleanThreshold");
3834
3835 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3836 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3837 numIter < CMSPrecleanIter;
3838 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3839 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3840 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3841 // Either there are very few dirty cards, so re-mark
3842 // pause will be small anyway, or our pre-cleaning isn't
3843 // that much faster than the rate at which cards are being
3844 // dirtied, so we might as well stop and re-mark since
3845 // precleaning won't improve our re-mark time by much.
3846 if (curNumCards <= CMSPrecleanThreshold ||
3847 (numIter > 0 &&
3848 (curNumCards * CMSPrecleanDenominator >
3849 lastNumCards * CMSPrecleanNumerator))) {
3850 numIter++;
3851 cumNumCards += curNumCards;
3852 break;
3853 }
3854 }
3855
3856 preclean_cld(&mrias_cl, _cmsGen->freelistLock());
3857
3858 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3859 cumNumCards += curNumCards;
3860 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3861 curNumCards, cumNumCards, numIter);
3862 return cumNumCards; // as a measure of useful work done
3863 }
3864
3865 // PRECLEANING NOTES:
3866 // Precleaning involves:
3867 // . reading the bits of the modUnionTable and clearing the set bits.
3868 // . For the cards corresponding to the set bits, we scan the
3869 // objects on those cards. This means we need the free_list_lock
3870 // so that we can safely iterate over the CMS space when scanning
3871 // for oops.
3872 // . When we scan the objects, we'll be both reading and setting
3873 // marks in the marking bit map, so we'll need the marking bit map.
3874 // . For protecting _collector_state transitions, we take the CGC_lock.
3875 // Note that any races in the reading of of card table entries by the
3876 // CMS thread on the one hand and the clearing of those entries by the
3877 // VM thread or the setting of those entries by the mutator threads on the
3878 // other are quite benign. However, for efficiency it makes sense to keep
3879 // the VM thread from racing with the CMS thread while the latter is
3880 // dirty card info to the modUnionTable. We therefore also use the
3881 // CGC_lock to protect the reading of the card table and the mod union
3882 // table by the CM thread.
3883 // . We run concurrently with mutator updates, so scanning
3884 // needs to be done carefully -- we should not try to scan
3885 // potentially uninitialized objects.
3886 //
3887 // Locking strategy: While holding the CGC_lock, we scan over and
3888 // reset a maximal dirty range of the mod union / card tables, then lock
3889 // the free_list_lock and bitmap lock to do a full marking, then
3890 // release these locks; and repeat the cycle. This allows for a
3891 // certain amount of fairness in the sharing of these locks between
3892 // the CMS collector on the one hand, and the VM thread and the
3893 // mutators on the other.
3894
3895 // NOTE: preclean_mod_union_table() and preclean_card_table()
3896 // further below are largely identical; if you need to modify
3897 // one of these methods, please check the other method too.
3898
3899 size_t CMSCollector::preclean_mod_union_table(
3900 ConcurrentMarkSweepGeneration* old_gen,
3901 ScanMarkedObjectsAgainCarefullyClosure* cl) {
3902 verify_work_stacks_empty();
3903 verify_overflow_empty();
3904
3905 // strategy: starting with the first card, accumulate contiguous
3906 // ranges of dirty cards; clear these cards, then scan the region
3907 // covered by these cards.
3908
3909 // Since all of the MUT is committed ahead, we can just use
3910 // that, in case the generations expand while we are precleaning.
3911 // It might also be fine to just use the committed part of the
3912 // generation, but we might potentially miss cards when the
3913 // generation is rapidly expanding while we are in the midst
3914 // of precleaning.
3915 HeapWord* startAddr = old_gen->reserved().start();
3916 HeapWord* endAddr = old_gen->reserved().end();
3917
3918 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
3919
3920 size_t numDirtyCards, cumNumDirtyCards;
3921 HeapWord *nextAddr, *lastAddr;
3922 for (cumNumDirtyCards = numDirtyCards = 0,
3923 nextAddr = lastAddr = startAddr;
3924 nextAddr < endAddr;
3925 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3926
3927 ResourceMark rm;
3928 HandleMark hm;
3929
3930 MemRegion dirtyRegion;
3931 {
3932 stopTimer();
3933 // Potential yield point
3934 CMSTokenSync ts(true);
3935 startTimer();
3936 sample_eden();
3937 // Get dirty region starting at nextOffset (inclusive),
3938 // simultaneously clearing it.
3939 dirtyRegion =
3940 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3941 assert(dirtyRegion.start() >= nextAddr,
3942 "returned region inconsistent?");
3943 }
3944 // Remember where the next search should begin.
3945 // The returned region (if non-empty) is a right open interval,
3946 // so lastOffset is obtained from the right end of that
3947 // interval.
3948 lastAddr = dirtyRegion.end();
3949 // Should do something more transparent and less hacky XXX
3950 numDirtyCards =
3951 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3952
3953 // We'll scan the cards in the dirty region (with periodic
3954 // yields for foreground GC as needed).
3955 if (!dirtyRegion.is_empty()) {
3956 assert(numDirtyCards > 0, "consistency check");
3957 HeapWord* stop_point = NULL;
3958 stopTimer();
3959 // Potential yield point
3960 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3961 bitMapLock());
3962 startTimer();
3963 {
3964 verify_work_stacks_empty();
3965 verify_overflow_empty();
3966 sample_eden();
3967 stop_point =
3968 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3969 }
3970 if (stop_point != NULL) {
3971 // The careful iteration stopped early either because it found an
3972 // uninitialized object, or because we were in the midst of an
3973 // "abortable preclean", which should now be aborted. Redirty
3974 // the bits corresponding to the partially-scanned or unscanned
3975 // cards. We'll either restart at the next block boundary or
3976 // abort the preclean.
3977 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3978 "Should only be AbortablePreclean.");
3979 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3980 if (should_abort_preclean()) {
3981 break; // out of preclean loop
3982 } else {
3983 // Compute the next address at which preclean should pick up;
3984 // might need bitMapLock in order to read P-bits.
3985 lastAddr = next_card_start_after_block(stop_point);
3986 }
3987 }
3988 } else {
3989 assert(lastAddr == endAddr, "consistency check");
3990 assert(numDirtyCards == 0, "consistency check");
3991 break;
3992 }
3993 }
3994 verify_work_stacks_empty();
3995 verify_overflow_empty();
3996 return cumNumDirtyCards;
3997 }
3998
3999 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4000 // below are largely identical; if you need to modify
4001 // one of these methods, please check the other method too.
4002
4003 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4004 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4005 // strategy: it's similar to precleamModUnionTable above, in that
4006 // we accumulate contiguous ranges of dirty cards, mark these cards
4007 // precleaned, then scan the region covered by these cards.
4008 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4009 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4010
4011 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4012
4013 size_t numDirtyCards, cumNumDirtyCards;
4014 HeapWord *lastAddr, *nextAddr;
4015
4016 for (cumNumDirtyCards = numDirtyCards = 0,
4017 nextAddr = lastAddr = startAddr;
4018 nextAddr < endAddr;
4019 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4020
4021 ResourceMark rm;
4022 HandleMark hm;
4023
4024 MemRegion dirtyRegion;
4025 {
4026 // See comments in "Precleaning notes" above on why we
4027 // do this locking. XXX Could the locking overheads be
4028 // too high when dirty cards are sparse? [I don't think so.]
4029 stopTimer();
4030 CMSTokenSync x(true); // is cms thread
4031 startTimer();
4032 sample_eden();
4033 // Get and clear dirty region from card table
4034 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr),
4035 true,
4036 CardTable::precleaned_card_val());
4037
4038 assert(dirtyRegion.start() >= nextAddr,
4039 "returned region inconsistent?");
4040 }
4041 lastAddr = dirtyRegion.end();
4042 numDirtyCards =
4043 dirtyRegion.word_size()/CardTable::card_size_in_words;
4044
4045 if (!dirtyRegion.is_empty()) {
4046 stopTimer();
4047 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4048 startTimer();
4049 sample_eden();
4050 verify_work_stacks_empty();
4051 verify_overflow_empty();
4052 HeapWord* stop_point =
4053 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4054 if (stop_point != NULL) {
4055 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4056 "Should only be AbortablePreclean.");
4057 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4058 if (should_abort_preclean()) {
4059 break; // out of preclean loop
4060 } else {
4061 // Compute the next address at which preclean should pick up.
4062 lastAddr = next_card_start_after_block(stop_point);
4063 }
4064 }
4065 } else {
4066 break;
4067 }
4068 }
4069 verify_work_stacks_empty();
4070 verify_overflow_empty();
4071 return cumNumDirtyCards;
4072 }
4073
4074 class PrecleanCLDClosure : public CLDClosure {
4075 MetadataAwareOopsInGenClosure* _cm_closure;
4076 public:
4077 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {}
4078 void do_cld(ClassLoaderData* cld) {
4079 if (cld->has_accumulated_modified_oops()) {
4080 cld->clear_accumulated_modified_oops();
4081
4082 _cm_closure->do_cld(cld);
4083 }
4084 }
4085 };
4086
4087 // The freelist lock is needed to prevent asserts, is it really needed?
4088 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4089
4090 cl->set_freelistLock(freelistLock);
4091
4092 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4093
4094 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4095 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4096 PrecleanCLDClosure preclean_closure(cl);
4097 ClassLoaderDataGraph::cld_do(&preclean_closure);
4098
4099 verify_work_stacks_empty();
4100 verify_overflow_empty();
4101 }
4102
4103 void CMSCollector::checkpointRootsFinal() {
4104 assert(_collectorState == FinalMarking, "incorrect state transition?");
4105 check_correct_thread_executing();
4106 // world is stopped at this checkpoint
4107 assert(SafepointSynchronize::is_at_safepoint(),
4108 "world should be stopped");
4109 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
4110
4111 verify_work_stacks_empty();
4112 verify_overflow_empty();
4113
4114 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4115 _young_gen->used() / K, _young_gen->capacity() / K);
4116 {
4117 if (CMSScavengeBeforeRemark) {
4118 CMSHeap* heap = CMSHeap::heap();
4119 // Temporarily set flag to false, GCH->do_collection will
4120 // expect it to be false and set to true
4121 FlagSetting fl(heap->_is_gc_active, false);
4122
4123 heap->do_collection(true, // full (i.e. force, see below)
4124 false, // !clear_all_soft_refs
4125 0, // size
4126 false, // is_tlab
4127 GenCollectedHeap::YoungGen // type
4128 );
4129 }
4130 FreelistLocker x(this);
4131 MutexLockerEx y(bitMapLock(),
4132 Mutex::_no_safepoint_check_flag);
4133 checkpointRootsFinalWork();
4134 }
4135 verify_work_stacks_empty();
4136 verify_overflow_empty();
4137 }
4138
4139 void CMSCollector::checkpointRootsFinalWork() {
4140 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4141
4142 assert(haveFreelistLocks(), "must have free list locks");
4143 assert_lock_strong(bitMapLock());
4144
4145 ResourceMark rm;
4146 HandleMark hm;
4147
4148 CMSHeap* heap = CMSHeap::heap();
4149
4150 if (should_unload_classes()) {
4151 CodeCache::gc_prologue();
4152 }
4153 assert(haveFreelistLocks(), "must have free list locks");
4154 assert_lock_strong(bitMapLock());
4155
4156 // We might assume that we need not fill TLAB's when
4157 // CMSScavengeBeforeRemark is set, because we may have just done
4158 // a scavenge which would have filled all TLAB's -- and besides
4159 // Eden would be empty. This however may not always be the case --
4160 // for instance although we asked for a scavenge, it may not have
4161 // happened because of a JNI critical section. We probably need
4162 // a policy for deciding whether we can in that case wait until
4163 // the critical section releases and then do the remark following
4164 // the scavenge, and skip it here. In the absence of that policy,
4165 // or of an indication of whether the scavenge did indeed occur,
4166 // we cannot rely on TLAB's having been filled and must do
4167 // so here just in case a scavenge did not happen.
4168 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them
4169 // Update the saved marks which may affect the root scans.
4170 heap->save_marks();
4171
4172 print_eden_and_survivor_chunk_arrays();
4173
4174 {
4175 #if COMPILER2_OR_JVMCI
4176 DerivedPointerTableDeactivate dpt_deact;
4177 #endif
4178
4179 // Note on the role of the mod union table:
4180 // Since the marker in "markFromRoots" marks concurrently with
4181 // mutators, it is possible for some reachable objects not to have been
4182 // scanned. For instance, an only reference to an object A was
4183 // placed in object B after the marker scanned B. Unless B is rescanned,
4184 // A would be collected. Such updates to references in marked objects
4185 // are detected via the mod union table which is the set of all cards
4186 // dirtied since the first checkpoint in this GC cycle and prior to
4187 // the most recent young generation GC, minus those cleaned up by the
4188 // concurrent precleaning.
4189 if (CMSParallelRemarkEnabled) {
4190 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4191 do_remark_parallel();
4192 } else {
4193 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4194 do_remark_non_parallel();
4195 }
4196 }
4197 verify_work_stacks_empty();
4198 verify_overflow_empty();
4199
4200 {
4201 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4202 refProcessingWork();
4203 }
4204 verify_work_stacks_empty();
4205 verify_overflow_empty();
4206
4207 if (should_unload_classes()) {
4208 CodeCache::gc_epilogue();
4209 }
4210 JvmtiExport::gc_epilogue();
4211
4212 // If we encountered any (marking stack / work queue) overflow
4213 // events during the current CMS cycle, take appropriate
4214 // remedial measures, where possible, so as to try and avoid
4215 // recurrence of that condition.
4216 assert(_markStack.isEmpty(), "No grey objects");
4217 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4218 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4219 if (ser_ovflw > 0) {
4220 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4221 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4222 _markStack.expand();
4223 _ser_pmc_remark_ovflw = 0;
4224 _ser_pmc_preclean_ovflw = 0;
4225 _ser_kac_preclean_ovflw = 0;
4226 _ser_kac_ovflw = 0;
4227 }
4228 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4229 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4230 _par_pmc_remark_ovflw, _par_kac_ovflw);
4231 _par_pmc_remark_ovflw = 0;
4232 _par_kac_ovflw = 0;
4233 }
4234 if (_markStack._hit_limit > 0) {
4235 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4236 _markStack._hit_limit);
4237 }
4238 if (_markStack._failed_double > 0) {
4239 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4240 _markStack._failed_double, _markStack.capacity());
4241 }
4242 _markStack._hit_limit = 0;
4243 _markStack._failed_double = 0;
4244
4245 if ((VerifyAfterGC || VerifyDuringGC) &&
4246 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4247 verify_after_remark();
4248 }
4249
4250 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4251
4252 // Change under the freelistLocks.
4253 _collectorState = Sweeping;
4254 // Call isAllClear() under bitMapLock
4255 assert(_modUnionTable.isAllClear(),
4256 "Should be clear by end of the final marking");
4257 assert(_ct->cld_rem_set()->mod_union_is_clear(),
4258 "Should be clear by end of the final marking");
4259 }
4260
4261 void CMSParInitialMarkTask::work(uint worker_id) {
4262 elapsedTimer _timer;
4263 ResourceMark rm;
4264 HandleMark hm;
4265
4266 // ---------- scan from roots --------------
4267 _timer.start();
4268 CMSHeap* heap = CMSHeap::heap();
4269 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4270
4271 // ---------- young gen roots --------------
4272 {
4273 work_on_young_gen_roots(&par_mri_cl);
4274 _timer.stop();
4275 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4276 }
4277
4278 // ---------- remaining roots --------------
4279 _timer.reset();
4280 _timer.start();
4281
4282 CLDToOopClosure cld_closure(&par_mri_cl, true);
4283
4284 heap->cms_process_roots(_strong_roots_scope,
4285 false, // yg was scanned above
4286 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4287 _collector->should_unload_classes(),
4288 &par_mri_cl,
4289 &cld_closure);
4290 assert(_collector->should_unload_classes()
4291 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4292 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4293 _timer.stop();
4294 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4295 }
4296
4297 // Parallel remark task
4298 class CMSParRemarkTask: public CMSParMarkTask {
4299 CompactibleFreeListSpace* _cms_space;
4300
4301 // The per-thread work queues, available here for stealing.
4302 OopTaskQueueSet* _task_queues;
4303 ParallelTaskTerminator _term;
4304 StrongRootsScope* _strong_roots_scope;
4305
4306 public:
4307 // A value of 0 passed to n_workers will cause the number of
4308 // workers to be taken from the active workers in the work gang.
4309 CMSParRemarkTask(CMSCollector* collector,
4310 CompactibleFreeListSpace* cms_space,
4311 uint n_workers, WorkGang* workers,
4312 OopTaskQueueSet* task_queues,
4313 StrongRootsScope* strong_roots_scope):
4314 CMSParMarkTask("Rescan roots and grey objects in parallel",
4315 collector, n_workers),
4316 _cms_space(cms_space),
4317 _task_queues(task_queues),
4318 _term(n_workers, task_queues),
4319 _strong_roots_scope(strong_roots_scope) { }
4320
4321 OopTaskQueueSet* task_queues() { return _task_queues; }
4322
4323 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4324
4325 ParallelTaskTerminator* terminator() { return &_term; }
4326 uint n_workers() { return _n_workers; }
4327
4328 void work(uint worker_id);
4329
4330 private:
4331 // ... of dirty cards in old space
4332 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4333 ParMarkRefsIntoAndScanClosure* cl);
4334
4335 // ... work stealing for the above
4336 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4337 };
4338
4339 class RemarkCLDClosure : public CLDClosure {
4340 CLDToOopClosure _cm_closure;
4341 public:
4342 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {}
4343 void do_cld(ClassLoaderData* cld) {
4344 // Check if we have modified any oops in the CLD during the concurrent marking.
4345 if (cld->has_accumulated_modified_oops()) {
4346 cld->clear_accumulated_modified_oops();
4347
4348 // We could have transfered the current modified marks to the accumulated marks,
4349 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4350 } else if (cld->has_modified_oops()) {
4351 // Don't clear anything, this info is needed by the next young collection.
4352 } else {
4353 // No modified oops in the ClassLoaderData.
4354 return;
4355 }
4356
4357 // The klass has modified fields, need to scan the klass.
4358 _cm_closure.do_cld(cld);
4359 }
4360 };
4361
4362 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4363 ParNewGeneration* young_gen = _collector->_young_gen;
4364 ContiguousSpace* eden_space = young_gen->eden();
4365 ContiguousSpace* from_space = young_gen->from();
4366 ContiguousSpace* to_space = young_gen->to();
4367
4368 HeapWord** eca = _collector->_eden_chunk_array;
4369 size_t ect = _collector->_eden_chunk_index;
4370 HeapWord** sca = _collector->_survivor_chunk_array;
4371 size_t sct = _collector->_survivor_chunk_index;
4372
4373 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4374 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4375
4376 do_young_space_rescan(cl, to_space, NULL, 0);
4377 do_young_space_rescan(cl, from_space, sca, sct);
4378 do_young_space_rescan(cl, eden_space, eca, ect);
4379 }
4380
4381 // work_queue(i) is passed to the closure
4382 // ParMarkRefsIntoAndScanClosure. The "i" parameter
4383 // also is passed to do_dirty_card_rescan_tasks() and to
4384 // do_work_steal() to select the i-th task_queue.
4385
4386 void CMSParRemarkTask::work(uint worker_id) {
4387 elapsedTimer _timer;
4388 ResourceMark rm;
4389 HandleMark hm;
4390
4391 // ---------- rescan from roots --------------
4392 _timer.start();
4393 CMSHeap* heap = CMSHeap::heap();
4394 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4395 _collector->_span, _collector->ref_processor(),
4396 &(_collector->_markBitMap),
4397 work_queue(worker_id));
4398
4399 // Rescan young gen roots first since these are likely
4400 // coarsely partitioned and may, on that account, constitute
4401 // the critical path; thus, it's best to start off that
4402 // work first.
4403 // ---------- young gen roots --------------
4404 {
4405 work_on_young_gen_roots(&par_mrias_cl);
4406 _timer.stop();
4407 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4408 }
4409
4410 // ---------- remaining roots --------------
4411 _timer.reset();
4412 _timer.start();
4413 heap->cms_process_roots(_strong_roots_scope,
4414 false, // yg was scanned above
4415 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4416 _collector->should_unload_classes(),
4417 &par_mrias_cl,
4418 NULL); // The dirty klasses will be handled below
4419
4420 assert(_collector->should_unload_classes()
4421 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4422 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4423 _timer.stop();
4424 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4425
4426 // ---------- unhandled CLD scanning ----------
4427 if (worker_id == 0) { // Single threaded at the moment.
4428 _timer.reset();
4429 _timer.start();
4430
4431 // Scan all new class loader data objects and new dependencies that were
4432 // introduced during concurrent marking.
4433 ResourceMark rm;
4434 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4435 for (int i = 0; i < array->length(); i++) {
4436 par_mrias_cl.do_cld_nv(array->at(i));
4437 }
4438
4439 // We don't need to keep track of new CLDs anymore.
4440 ClassLoaderDataGraph::remember_new_clds(false);
4441
4442 _timer.stop();
4443 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4444 }
4445
4446 // We might have added oops to ClassLoaderData::_handles during the
4447 // concurrent marking phase. These oops do not always point to newly allocated objects
4448 // that are guaranteed to be kept alive. Hence,
4449 // we do have to revisit the _handles block during the remark phase.
4450
4451 // ---------- dirty CLD scanning ----------
4452 if (worker_id == 0) { // Single threaded at the moment.
4453 _timer.reset();
4454 _timer.start();
4455
4456 // Scan all classes that was dirtied during the concurrent marking phase.
4457 RemarkCLDClosure remark_closure(&par_mrias_cl);
4458 ClassLoaderDataGraph::cld_do(&remark_closure);
4459
4460 _timer.stop();
4461 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4462 }
4463
4464
4465 // ---------- rescan dirty cards ------------
4466 _timer.reset();
4467 _timer.start();
4468
4469 // Do the rescan tasks for each of the two spaces
4470 // (cms_space) in turn.
4471 // "worker_id" is passed to select the task_queue for "worker_id"
4472 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4473 _timer.stop();
4474 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4475
4476 // ---------- steal work from other threads ...
4477 // ---------- ... and drain overflow list.
4478 _timer.reset();
4479 _timer.start();
4480 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4481 _timer.stop();
4482 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4483 }
4484
4485 void
4486 CMSParMarkTask::do_young_space_rescan(
4487 OopsInGenClosure* cl, ContiguousSpace* space,
4488 HeapWord** chunk_array, size_t chunk_top) {
4489 // Until all tasks completed:
4490 // . claim an unclaimed task
4491 // . compute region boundaries corresponding to task claimed
4492 // using chunk_array
4493 // . par_oop_iterate(cl) over that region
4494
4495 ResourceMark rm;
4496 HandleMark hm;
4497
4498 SequentialSubTasksDone* pst = space->par_seq_tasks();
4499
4500 uint nth_task = 0;
4501 uint n_tasks = pst->n_tasks();
4502
4503 if (n_tasks > 0) {
4504 assert(pst->valid(), "Uninitialized use?");
4505 HeapWord *start, *end;
4506 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4507 // We claimed task # nth_task; compute its boundaries.
4508 if (chunk_top == 0) { // no samples were taken
4509 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4510 start = space->bottom();
4511 end = space->top();
4512 } else if (nth_task == 0) {
4513 start = space->bottom();
4514 end = chunk_array[nth_task];
4515 } else if (nth_task < (uint)chunk_top) {
4516 assert(nth_task >= 1, "Control point invariant");
4517 start = chunk_array[nth_task - 1];
4518 end = chunk_array[nth_task];
4519 } else {
4520 assert(nth_task == (uint)chunk_top, "Control point invariant");
4521 start = chunk_array[chunk_top - 1];
4522 end = space->top();
4523 }
4524 MemRegion mr(start, end);
4525 // Verify that mr is in space
4526 assert(mr.is_empty() || space->used_region().contains(mr),
4527 "Should be in space");
4528 // Verify that "start" is an object boundary
4529 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4530 "Should be an oop");
4531 space->par_oop_iterate(mr, cl);
4532 }
4533 pst->all_tasks_completed();
4534 }
4535 }
4536
4537 void
4538 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4539 CompactibleFreeListSpace* sp, int i,
4540 ParMarkRefsIntoAndScanClosure* cl) {
4541 // Until all tasks completed:
4542 // . claim an unclaimed task
4543 // . compute region boundaries corresponding to task claimed
4544 // . transfer dirty bits ct->mut for that region
4545 // . apply rescanclosure to dirty mut bits for that region
4546
4547 ResourceMark rm;
4548 HandleMark hm;
4549
4550 OopTaskQueue* work_q = work_queue(i);
4551 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4552 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4553 // CAUTION: This closure has state that persists across calls to
4554 // the work method dirty_range_iterate_clear() in that it has
4555 // embedded in it a (subtype of) UpwardsObjectClosure. The
4556 // use of that state in the embedded UpwardsObjectClosure instance
4557 // assumes that the cards are always iterated (even if in parallel
4558 // by several threads) in monotonically increasing order per each
4559 // thread. This is true of the implementation below which picks
4560 // card ranges (chunks) in monotonically increasing order globally
4561 // and, a-fortiori, in monotonically increasing order per thread
4562 // (the latter order being a subsequence of the former).
4563 // If the work code below is ever reorganized into a more chaotic
4564 // work-partitioning form than the current "sequential tasks"
4565 // paradigm, the use of that persistent state will have to be
4566 // revisited and modified appropriately. See also related
4567 // bug 4756801 work on which should examine this code to make
4568 // sure that the changes there do not run counter to the
4569 // assumptions made here and necessary for correctness and
4570 // efficiency. Note also that this code might yield inefficient
4571 // behavior in the case of very large objects that span one or
4572 // more work chunks. Such objects would potentially be scanned
4573 // several times redundantly. Work on 4756801 should try and
4574 // address that performance anomaly if at all possible. XXX
4575 MemRegion full_span = _collector->_span;
4576 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4577 MarkFromDirtyCardsClosure
4578 greyRescanClosure(_collector, full_span, // entire span of interest
4579 sp, bm, work_q, cl);
4580
4581 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4582 assert(pst->valid(), "Uninitialized use?");
4583 uint nth_task = 0;
4584 const int alignment = CardTable::card_size * BitsPerWord;
4585 MemRegion span = sp->used_region();
4586 HeapWord* start_addr = span.start();
4587 HeapWord* end_addr = align_up(span.end(), alignment);
4588 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4589 assert(is_aligned(start_addr, alignment), "Check alignment");
4590 assert(is_aligned(chunk_size, alignment), "Check alignment");
4591
4592 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4593 // Having claimed the nth_task, compute corresponding mem-region,
4594 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4595 // The alignment restriction ensures that we do not need any
4596 // synchronization with other gang-workers while setting or
4597 // clearing bits in thus chunk of the MUT.
4598 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4599 start_addr + (nth_task+1)*chunk_size);
4600 // The last chunk's end might be way beyond end of the
4601 // used region. In that case pull back appropriately.
4602 if (this_span.end() > end_addr) {
4603 this_span.set_end(end_addr);
4604 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4605 }
4606 // Iterate over the dirty cards covering this chunk, marking them
4607 // precleaned, and setting the corresponding bits in the mod union
4608 // table. Since we have been careful to partition at Card and MUT-word
4609 // boundaries no synchronization is needed between parallel threads.
4610 _collector->_ct->dirty_card_iterate(this_span,
4611 &modUnionClosure);
4612
4613 // Having transferred these marks into the modUnionTable,
4614 // rescan the marked objects on the dirty cards in the modUnionTable.
4615 // Even if this is at a synchronous collection, the initial marking
4616 // may have been done during an asynchronous collection so there
4617 // may be dirty bits in the mod-union table.
4618 _collector->_modUnionTable.dirty_range_iterate_clear(
4619 this_span, &greyRescanClosure);
4620 _collector->_modUnionTable.verifyNoOneBitsInRange(
4621 this_span.start(),
4622 this_span.end());
4623 }
4624 pst->all_tasks_completed(); // declare that i am done
4625 }
4626
4627 // . see if we can share work_queues with ParNew? XXX
4628 void
4629 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4630 int* seed) {
4631 OopTaskQueue* work_q = work_queue(i);
4632 NOT_PRODUCT(int num_steals = 0;)
4633 oop obj_to_scan;
4634 CMSBitMap* bm = &(_collector->_markBitMap);
4635
4636 while (true) {
4637 // Completely finish any left over work from (an) earlier round(s)
4638 cl->trim_queue(0);
4639 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4640 (size_t)ParGCDesiredObjsFromOverflowList);
4641 // Now check if there's any work in the overflow list
4642 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4643 // only affects the number of attempts made to get work from the
4644 // overflow list and does not affect the number of workers. Just
4645 // pass ParallelGCThreads so this behavior is unchanged.
4646 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4647 work_q,
4648 ParallelGCThreads)) {
4649 // found something in global overflow list;
4650 // not yet ready to go stealing work from others.
4651 // We'd like to assert(work_q->size() != 0, ...)
4652 // because we just took work from the overflow list,
4653 // but of course we can't since all of that could have
4654 // been already stolen from us.
4655 // "He giveth and He taketh away."
4656 continue;
4657 }
4658 // Verify that we have no work before we resort to stealing
4659 assert(work_q->size() == 0, "Have work, shouldn't steal");
4660 // Try to steal from other queues that have work
4661 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4662 NOT_PRODUCT(num_steals++;)
4663 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4664 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4665 // Do scanning work
4666 obj_to_scan->oop_iterate(cl);
4667 // Loop around, finish this work, and try to steal some more
4668 } else if (terminator()->offer_termination()) {
4669 break; // nirvana from the infinite cycle
4670 }
4671 }
4672 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4673 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4674 "Else our work is not yet done");
4675 }
4676
4677 // Record object boundaries in _eden_chunk_array by sampling the eden
4678 // top in the slow-path eden object allocation code path and record
4679 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4680 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4681 // sampling in sample_eden() that activates during the part of the
4682 // preclean phase.
4683 void CMSCollector::sample_eden_chunk() {
4684 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4685 if (_eden_chunk_lock->try_lock()) {
4686 // Record a sample. This is the critical section. The contents
4687 // of the _eden_chunk_array have to be non-decreasing in the
4688 // address order.
4689 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4690 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4691 "Unexpected state of Eden");
4692 if (_eden_chunk_index == 0 ||
4693 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4694 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4695 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4696 _eden_chunk_index++; // commit sample
4697 }
4698 _eden_chunk_lock->unlock();
4699 }
4700 }
4701 }
4702
4703 // Return a thread-local PLAB recording array, as appropriate.
4704 void* CMSCollector::get_data_recorder(int thr_num) {
4705 if (_survivor_plab_array != NULL &&
4706 (CMSPLABRecordAlways ||
4707 (_collectorState > Marking && _collectorState < FinalMarking))) {
4708 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4709 ChunkArray* ca = &_survivor_plab_array[thr_num];
4710 ca->reset(); // clear it so that fresh data is recorded
4711 return (void*) ca;
4712 } else {
4713 return NULL;
4714 }
4715 }
4716
4717 // Reset all the thread-local PLAB recording arrays
4718 void CMSCollector::reset_survivor_plab_arrays() {
4719 for (uint i = 0; i < ParallelGCThreads; i++) {
4720 _survivor_plab_array[i].reset();
4721 }
4722 }
4723
4724 // Merge the per-thread plab arrays into the global survivor chunk
4725 // array which will provide the partitioning of the survivor space
4726 // for CMS initial scan and rescan.
4727 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4728 int no_of_gc_threads) {
4729 assert(_survivor_plab_array != NULL, "Error");
4730 assert(_survivor_chunk_array != NULL, "Error");
4731 assert(_collectorState == FinalMarking ||
4732 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4733 for (int j = 0; j < no_of_gc_threads; j++) {
4734 _cursor[j] = 0;
4735 }
4736 HeapWord* top = surv->top();
4737 size_t i;
4738 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4739 HeapWord* min_val = top; // Higher than any PLAB address
4740 uint min_tid = 0; // position of min_val this round
4741 for (int j = 0; j < no_of_gc_threads; j++) {
4742 ChunkArray* cur_sca = &_survivor_plab_array[j];
4743 if (_cursor[j] == cur_sca->end()) {
4744 continue;
4745 }
4746 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4747 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4748 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4749 if (cur_val < min_val) {
4750 min_tid = j;
4751 min_val = cur_val;
4752 } else {
4753 assert(cur_val < top, "All recorded addresses should be less");
4754 }
4755 }
4756 // At this point min_val and min_tid are respectively
4757 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4758 // and the thread (j) that witnesses that address.
4759 // We record this address in the _survivor_chunk_array[i]
4760 // and increment _cursor[min_tid] prior to the next round i.
4761 if (min_val == top) {
4762 break;
4763 }
4764 _survivor_chunk_array[i] = min_val;
4765 _cursor[min_tid]++;
4766 }
4767 // We are all done; record the size of the _survivor_chunk_array
4768 _survivor_chunk_index = i; // exclusive: [0, i)
4769 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4770 // Verify that we used up all the recorded entries
4771 #ifdef ASSERT
4772 size_t total = 0;
4773 for (int j = 0; j < no_of_gc_threads; j++) {
4774 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4775 total += _cursor[j];
4776 }
4777 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4778 // Check that the merged array is in sorted order
4779 if (total > 0) {
4780 for (size_t i = 0; i < total - 1; i++) {
4781 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4782 i, p2i(_survivor_chunk_array[i]));
4783 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4784 "Not sorted");
4785 }
4786 }
4787 #endif // ASSERT
4788 }
4789
4790 // Set up the space's par_seq_tasks structure for work claiming
4791 // for parallel initial scan and rescan of young gen.
4792 // See ParRescanTask where this is currently used.
4793 void
4794 CMSCollector::
4795 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4796 assert(n_threads > 0, "Unexpected n_threads argument");
4797
4798 // Eden space
4799 if (!_young_gen->eden()->is_empty()) {
4800 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4801 assert(!pst->valid(), "Clobbering existing data?");
4802 // Each valid entry in [0, _eden_chunk_index) represents a task.
4803 size_t n_tasks = _eden_chunk_index + 1;
4804 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4805 // Sets the condition for completion of the subtask (how many threads
4806 // need to finish in order to be done).
4807 pst->set_n_threads(n_threads);
4808 pst->set_n_tasks((int)n_tasks);
4809 }
4810
4811 // Merge the survivor plab arrays into _survivor_chunk_array
4812 if (_survivor_plab_array != NULL) {
4813 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4814 } else {
4815 assert(_survivor_chunk_index == 0, "Error");
4816 }
4817
4818 // To space
4819 {
4820 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4821 assert(!pst->valid(), "Clobbering existing data?");
4822 // Sets the condition for completion of the subtask (how many threads
4823 // need to finish in order to be done).
4824 pst->set_n_threads(n_threads);
4825 pst->set_n_tasks(1);
4826 assert(pst->valid(), "Error");
4827 }
4828
4829 // From space
4830 {
4831 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4832 assert(!pst->valid(), "Clobbering existing data?");
4833 size_t n_tasks = _survivor_chunk_index + 1;
4834 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4835 // Sets the condition for completion of the subtask (how many threads
4836 // need to finish in order to be done).
4837 pst->set_n_threads(n_threads);
4838 pst->set_n_tasks((int)n_tasks);
4839 assert(pst->valid(), "Error");
4840 }
4841 }
4842
4843 // Parallel version of remark
4844 void CMSCollector::do_remark_parallel() {
4845 CMSHeap* heap = CMSHeap::heap();
4846 WorkGang* workers = heap->workers();
4847 assert(workers != NULL, "Need parallel worker threads.");
4848 // Choose to use the number of GC workers most recently set
4849 // into "active_workers".
4850 uint n_workers = workers->active_workers();
4851
4852 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4853
4854 StrongRootsScope srs(n_workers);
4855
4856 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4857
4858 // We won't be iterating over the cards in the card table updating
4859 // the younger_gen cards, so we shouldn't call the following else
4860 // the verification code as well as subsequent younger_refs_iterate
4861 // code would get confused. XXX
4862 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4863
4864 // The young gen rescan work will not be done as part of
4865 // process_roots (which currently doesn't know how to
4866 // parallelize such a scan), but rather will be broken up into
4867 // a set of parallel tasks (via the sampling that the [abortable]
4868 // preclean phase did of eden, plus the [two] tasks of
4869 // scanning the [two] survivor spaces. Further fine-grain
4870 // parallelization of the scanning of the survivor spaces
4871 // themselves, and of precleaning of the young gen itself
4872 // is deferred to the future.
4873 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4874
4875 // The dirty card rescan work is broken up into a "sequence"
4876 // of parallel tasks (per constituent space) that are dynamically
4877 // claimed by the parallel threads.
4878 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4879
4880 // It turns out that even when we're using 1 thread, doing the work in a
4881 // separate thread causes wide variance in run times. We can't help this
4882 // in the multi-threaded case, but we special-case n=1 here to get
4883 // repeatable measurements of the 1-thread overhead of the parallel code.
4884 if (n_workers > 1) {
4885 // Make refs discovery MT-safe, if it isn't already: it may not
4886 // necessarily be so, since it's possible that we are doing
4887 // ST marking.
4888 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4889 workers->run_task(&tsk);
4890 } else {
4891 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4892 tsk.work(0);
4893 }
4894
4895 // restore, single-threaded for now, any preserved marks
4896 // as a result of work_q overflow
4897 restore_preserved_marks_if_any();
4898 }
4899
4900 // Non-parallel version of remark
4901 void CMSCollector::do_remark_non_parallel() {
4902 ResourceMark rm;
4903 HandleMark hm;
4904 CMSHeap* heap = CMSHeap::heap();
4905 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4906
4907 MarkRefsIntoAndScanClosure
4908 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4909 &_markStack, this,
4910 false /* should_yield */, false /* not precleaning */);
4911 MarkFromDirtyCardsClosure
4912 markFromDirtyCardsClosure(this, _span,
4913 NULL, // space is set further below
4914 &_markBitMap, &_markStack, &mrias_cl);
4915 {
4916 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4917 // Iterate over the dirty cards, setting the corresponding bits in the
4918 // mod union table.
4919 {
4920 ModUnionClosure modUnionClosure(&_modUnionTable);
4921 _ct->dirty_card_iterate(_cmsGen->used_region(),
4922 &modUnionClosure);
4923 }
4924 // Having transferred these marks into the modUnionTable, we just need
4925 // to rescan the marked objects on the dirty cards in the modUnionTable.
4926 // The initial marking may have been done during an asynchronous
4927 // collection so there may be dirty bits in the mod-union table.
4928 const int alignment = CardTable::card_size * BitsPerWord;
4929 {
4930 // ... First handle dirty cards in CMS gen
4931 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4932 MemRegion ur = _cmsGen->used_region();
4933 HeapWord* lb = ur.start();
4934 HeapWord* ub = align_up(ur.end(), alignment);
4935 MemRegion cms_span(lb, ub);
4936 _modUnionTable.dirty_range_iterate_clear(cms_span,
4937 &markFromDirtyCardsClosure);
4938 verify_work_stacks_empty();
4939 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4940 }
4941 }
4942 if (VerifyDuringGC &&
4943 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4944 HandleMark hm; // Discard invalid handles created during verification
4945 Universe::verify();
4946 }
4947 {
4948 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4949
4950 verify_work_stacks_empty();
4951
4952 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4953 StrongRootsScope srs(1);
4954
4955 heap->cms_process_roots(&srs,
4956 true, // young gen as roots
4957 GenCollectedHeap::ScanningOption(roots_scanning_options()),
4958 should_unload_classes(),
4959 &mrias_cl,
4960 NULL); // The dirty klasses will be handled below
4961
4962 assert(should_unload_classes()
4963 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4964 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4965 }
4966
4967 {
4968 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4969
4970 verify_work_stacks_empty();
4971
4972 // Scan all class loader data objects that might have been introduced
4973 // during concurrent marking.
4974 ResourceMark rm;
4975 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4976 for (int i = 0; i < array->length(); i++) {
4977 mrias_cl.do_cld_nv(array->at(i));
4978 }
4979
4980 // We don't need to keep track of new CLDs anymore.
4981 ClassLoaderDataGraph::remember_new_clds(false);
4982
4983 verify_work_stacks_empty();
4984 }
4985
4986 // We might have added oops to ClassLoaderData::_handles during the
4987 // concurrent marking phase. These oops do not point to newly allocated objects
4988 // that are guaranteed to be kept alive. Hence,
4989 // we do have to revisit the _handles block during the remark phase.
4990 {
4991 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm);
4992
4993 verify_work_stacks_empty();
4994
4995 RemarkCLDClosure remark_closure(&mrias_cl);
4996 ClassLoaderDataGraph::cld_do(&remark_closure);
4997
4998 verify_work_stacks_empty();
4999 }
5000
5001 verify_work_stacks_empty();
5002 // Restore evacuated mark words, if any, used for overflow list links
5003 restore_preserved_marks_if_any();
5004
5005 verify_overflow_empty();
5006 }
5007
5008 ////////////////////////////////////////////////////////
5009 // Parallel Reference Processing Task Proxy Class
5010 ////////////////////////////////////////////////////////
5011 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5012 OopTaskQueueSet* _queues;
5013 ParallelTaskTerminator _terminator;
5014 public:
5015 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5016 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5017 ParallelTaskTerminator* terminator() { return &_terminator; }
5018 OopTaskQueueSet* queues() { return _queues; }
5019 };
5020
5021 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5022 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5023 CMSCollector* _collector;
5024 CMSBitMap* _mark_bit_map;
5025 const MemRegion _span;
5026 ProcessTask& _task;
5027
5028 public:
5029 CMSRefProcTaskProxy(ProcessTask& task,
5030 CMSCollector* collector,
5031 const MemRegion& span,
5032 CMSBitMap* mark_bit_map,
5033 AbstractWorkGang* workers,
5034 OopTaskQueueSet* task_queues):
5035 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5036 task_queues,
5037 workers->active_workers()),
5038 _task(task),
5039 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5040 {
5041 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5042 "Inconsistency in _span");
5043 }
5044
5045 OopTaskQueueSet* task_queues() { return queues(); }
5046
5047 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5048
5049 void do_work_steal(int i,
5050 CMSParDrainMarkingStackClosure* drain,
5051 CMSParKeepAliveClosure* keep_alive,
5052 int* seed);
5053
5054 virtual void work(uint worker_id);
5055 };
5056
5057 void CMSRefProcTaskProxy::work(uint worker_id) {
5058 ResourceMark rm;
5059 HandleMark hm;
5060 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5061 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5062 _mark_bit_map,
5063 work_queue(worker_id));
5064 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5065 _mark_bit_map,
5066 work_queue(worker_id));
5067 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5068 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5069 if (_task.marks_oops_alive()) {
5070 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5071 _collector->hash_seed(worker_id));
5072 }
5073 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5074 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5075 }
5076
5077 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5078 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5079 EnqueueTask& _task;
5080
5081 public:
5082 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5083 : AbstractGangTask("Enqueue reference objects in parallel"),
5084 _task(task)
5085 { }
5086
5087 virtual void work(uint worker_id)
5088 {
5089 _task.work(worker_id);
5090 }
5091 };
5092
5093 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5094 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5095 _span(span),
5096 _bit_map(bit_map),
5097 _work_queue(work_queue),
5098 _mark_and_push(collector, span, bit_map, work_queue),
5099 _low_water_mark(MIN2((work_queue->max_elems()/4),
5100 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5101 { }
5102
5103 // . see if we can share work_queues with ParNew? XXX
5104 void CMSRefProcTaskProxy::do_work_steal(int i,
5105 CMSParDrainMarkingStackClosure* drain,
5106 CMSParKeepAliveClosure* keep_alive,
5107 int* seed) {
5108 OopTaskQueue* work_q = work_queue(i);
5109 NOT_PRODUCT(int num_steals = 0;)
5110 oop obj_to_scan;
5111
5112 while (true) {
5113 // Completely finish any left over work from (an) earlier round(s)
5114 drain->trim_queue(0);
5115 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5116 (size_t)ParGCDesiredObjsFromOverflowList);
5117 // Now check if there's any work in the overflow list
5118 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5119 // only affects the number of attempts made to get work from the
5120 // overflow list and does not affect the number of workers. Just
5121 // pass ParallelGCThreads so this behavior is unchanged.
5122 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5123 work_q,
5124 ParallelGCThreads)) {
5125 // Found something in global overflow list;
5126 // not yet ready to go stealing work from others.
5127 // We'd like to assert(work_q->size() != 0, ...)
5128 // because we just took work from the overflow list,
5129 // but of course we can't, since all of that might have
5130 // been already stolen from us.
5131 continue;
5132 }
5133 // Verify that we have no work before we resort to stealing
5134 assert(work_q->size() == 0, "Have work, shouldn't steal");
5135 // Try to steal from other queues that have work
5136 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5137 NOT_PRODUCT(num_steals++;)
5138 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5139 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5140 // Do scanning work
5141 obj_to_scan->oop_iterate(keep_alive);
5142 // Loop around, finish this work, and try to steal some more
5143 } else if (terminator()->offer_termination()) {
5144 break; // nirvana from the infinite cycle
5145 }
5146 }
5147 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5148 }
5149
5150 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5151 {
5152 CMSHeap* heap = CMSHeap::heap();
5153 WorkGang* workers = heap->workers();
5154 assert(workers != NULL, "Need parallel worker threads.");
5155 CMSRefProcTaskProxy rp_task(task, &_collector,
5156 _collector.ref_processor_span(),
5157 _collector.markBitMap(),
5158 workers, _collector.task_queues());
5159 workers->run_task(&rp_task);
5160 }
5161
5162 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5163 {
5164
5165 CMSHeap* heap = CMSHeap::heap();
5166 WorkGang* workers = heap->workers();
5167 assert(workers != NULL, "Need parallel worker threads.");
5168 CMSRefEnqueueTaskProxy enq_task(task);
5169 workers->run_task(&enq_task);
5170 }
5171
5172 void CMSCollector::refProcessingWork() {
5173 ResourceMark rm;
5174 HandleMark hm;
5175
5176 ReferenceProcessor* rp = ref_processor();
5177 assert(_span_discoverer.span().equals(_span), "Spans should be equal");
5178 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5179 // Process weak references.
5180 rp->setup_policy(false);
5181 verify_work_stacks_empty();
5182
5183 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
5184 {
5185 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5186
5187 // Setup keep_alive and complete closures.
5188 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5189 &_markStack, false /* !preclean */);
5190 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5191 _span, &_markBitMap, &_markStack,
5192 &cmsKeepAliveClosure, false /* !preclean */);
5193
5194 ReferenceProcessorStats stats;
5195 if (rp->processing_is_mt()) {
5196 // Set the degree of MT here. If the discovery is done MT, there
5197 // may have been a different number of threads doing the discovery
5198 // and a different number of discovered lists may have Ref objects.
5199 // That is OK as long as the Reference lists are balanced (see
5200 // balance_all_queues() and balance_queues()).
5201 CMSHeap* heap = CMSHeap::heap();
5202 uint active_workers = ParallelGCThreads;
5203 WorkGang* workers = heap->workers();
5204 if (workers != NULL) {
5205 active_workers = workers->active_workers();
5206 // The expectation is that active_workers will have already
5207 // been set to a reasonable value. If it has not been set,
5208 // investigate.
5209 assert(active_workers > 0, "Should have been set during scavenge");
5210 }
5211 rp->set_active_mt_degree(active_workers);
5212 CMSRefProcTaskExecutor task_executor(*this);
5213 stats = rp->process_discovered_references(&_is_alive_closure,
5214 &cmsKeepAliveClosure,
5215 &cmsDrainMarkingStackClosure,
5216 &task_executor,
5217 &pt);
5218 } else {
5219 stats = rp->process_discovered_references(&_is_alive_closure,
5220 &cmsKeepAliveClosure,
5221 &cmsDrainMarkingStackClosure,
5222 NULL,
5223 &pt);
5224 }
5225 _gc_tracer_cm->report_gc_reference_stats(stats);
5226 pt.print_all_references();
5227 }
5228
5229 // This is the point where the entire marking should have completed.
5230 verify_work_stacks_empty();
5231
5232 {
5233 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm);
5234 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl);
5235 }
5236
5237 if (should_unload_classes()) {
5238 {
5239 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5240
5241 // Unload classes and purge the SystemDictionary.
5242 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm);
5243
5244 // Unload nmethods.
5245 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5246
5247 // Prune dead klasses from subklass/sibling/implementor lists.
5248 Klass::clean_weak_klass_links();
5249 }
5250
5251 {
5252 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5253 // Clean up unreferenced symbols in symbol table.
5254 SymbolTable::unlink();
5255 }
5256
5257 {
5258 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5259 // Delete entries for dead interned strings.
5260 StringTable::unlink(&_is_alive_closure);
5261 }
5262 }
5263
5264 // Restore any preserved marks as a result of mark stack or
5265 // work queue overflow
5266 restore_preserved_marks_if_any(); // done single-threaded for now
5267
5268 rp->set_enqueuing_is_done(true);
5269 if (rp->processing_is_mt()) {
5270 rp->balance_all_queues();
5271 CMSRefProcTaskExecutor task_executor(*this);
5272 rp->enqueue_discovered_references(&task_executor, &pt);
5273 } else {
5274 rp->enqueue_discovered_references(NULL, &pt);
5275 }
5276 rp->verify_no_references_recorded();
5277 pt.print_enqueue_phase();
5278 assert(!rp->discovery_enabled(), "should have been disabled");
5279 }
5280
5281 #ifndef PRODUCT
5282 void CMSCollector::check_correct_thread_executing() {
5283 Thread* t = Thread::current();
5284 // Only the VM thread or the CMS thread should be here.
5285 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5286 "Unexpected thread type");
5287 // If this is the vm thread, the foreground process
5288 // should not be waiting. Note that _foregroundGCIsActive is
5289 // true while the foreground collector is waiting.
5290 if (_foregroundGCShouldWait) {
5291 // We cannot be the VM thread
5292 assert(t->is_ConcurrentGC_thread(),
5293 "Should be CMS thread");
5294 } else {
5295 // We can be the CMS thread only if we are in a stop-world
5296 // phase of CMS collection.
5297 if (t->is_ConcurrentGC_thread()) {
5298 assert(_collectorState == InitialMarking ||
5299 _collectorState == FinalMarking,
5300 "Should be a stop-world phase");
5301 // The CMS thread should be holding the CMS_token.
5302 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5303 "Potential interference with concurrently "
5304 "executing VM thread");
5305 }
5306 }
5307 }
5308 #endif
5309
5310 void CMSCollector::sweep() {
5311 assert(_collectorState == Sweeping, "just checking");
5312 check_correct_thread_executing();
5313 verify_work_stacks_empty();
5314 verify_overflow_empty();
5315 increment_sweep_count();
5316 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
5317
5318 _inter_sweep_timer.stop();
5319 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5320
5321 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5322 _intra_sweep_timer.reset();
5323 _intra_sweep_timer.start();
5324 {
5325 GCTraceCPUTime tcpu;
5326 CMSPhaseAccounting pa(this, "Concurrent Sweep");
5327 // First sweep the old gen
5328 {
5329 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5330 bitMapLock());
5331 sweepWork(_cmsGen);
5332 }
5333
5334 // Update Universe::_heap_*_at_gc figures.
5335 // We need all the free list locks to make the abstract state
5336 // transition from Sweeping to Resetting. See detailed note
5337 // further below.
5338 {
5339 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5340 // Update heap occupancy information which is used as
5341 // input to soft ref clearing policy at the next gc.
5342 Universe::update_heap_info_at_gc();
5343 _collectorState = Resizing;
5344 }
5345 }
5346 verify_work_stacks_empty();
5347 verify_overflow_empty();
5348
5349 if (should_unload_classes()) {
5350 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5351 // requires that the virtual spaces are stable and not deleted.
5352 ClassLoaderDataGraph::set_should_purge(true);
5353 }
5354
5355 _intra_sweep_timer.stop();
5356 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5357
5358 _inter_sweep_timer.reset();
5359 _inter_sweep_timer.start();
5360
5361 // We need to use a monotonically non-decreasing time in ms
5362 // or we will see time-warp warnings and os::javaTimeMillis()
5363 // does not guarantee monotonicity.
5364 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5365 update_time_of_last_gc(now);
5366
5367 // NOTE on abstract state transitions:
5368 // Mutators allocate-live and/or mark the mod-union table dirty
5369 // based on the state of the collection. The former is done in
5370 // the interval [Marking, Sweeping] and the latter in the interval
5371 // [Marking, Sweeping). Thus the transitions into the Marking state
5372 // and out of the Sweeping state must be synchronously visible
5373 // globally to the mutators.
5374 // The transition into the Marking state happens with the world
5375 // stopped so the mutators will globally see it. Sweeping is
5376 // done asynchronously by the background collector so the transition
5377 // from the Sweeping state to the Resizing state must be done
5378 // under the freelistLock (as is the check for whether to
5379 // allocate-live and whether to dirty the mod-union table).
5380 assert(_collectorState == Resizing, "Change of collector state to"
5381 " Resizing must be done under the freelistLocks (plural)");
5382
5383 // Now that sweeping has been completed, we clear
5384 // the incremental_collection_failed flag,
5385 // thus inviting a younger gen collection to promote into
5386 // this generation. If such a promotion may still fail,
5387 // the flag will be set again when a young collection is
5388 // attempted.
5389 CMSHeap* heap = CMSHeap::heap();
5390 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5391 heap->update_full_collections_completed(_collection_count_start);
5392 }
5393
5394 // FIX ME!!! Looks like this belongs in CFLSpace, with
5395 // CMSGen merely delegating to it.
5396 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5397 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5398 HeapWord* minAddr = _cmsSpace->bottom();
5399 HeapWord* largestAddr =
5400 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5401 if (largestAddr == NULL) {
5402 // The dictionary appears to be empty. In this case
5403 // try to coalesce at the end of the heap.
5404 largestAddr = _cmsSpace->end();
5405 }
5406 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5407 size_t nearLargestOffset =
5408 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5409 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5410 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5411 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5412 }
5413
5414 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5415 return addr >= _cmsSpace->nearLargestChunk();
5416 }
5417
5418 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5419 return _cmsSpace->find_chunk_at_end();
5420 }
5421
5422 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5423 bool full) {
5424 // If the young generation has been collected, gather any statistics
5425 // that are of interest at this point.
5426 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation);
5427 if (!full && current_is_young) {
5428 // Gather statistics on the young generation collection.
5429 collector()->stats().record_gc0_end(used());
5430 }
5431 }
5432
5433 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5434 // We iterate over the space(s) underlying this generation,
5435 // checking the mark bit map to see if the bits corresponding
5436 // to specific blocks are marked or not. Blocks that are
5437 // marked are live and are not swept up. All remaining blocks
5438 // are swept up, with coalescing on-the-fly as we sweep up
5439 // contiguous free and/or garbage blocks:
5440 // We need to ensure that the sweeper synchronizes with allocators
5441 // and stop-the-world collectors. In particular, the following
5442 // locks are used:
5443 // . CMS token: if this is held, a stop the world collection cannot occur
5444 // . freelistLock: if this is held no allocation can occur from this
5445 // generation by another thread
5446 // . bitMapLock: if this is held, no other thread can access or update
5447 //
5448
5449 // Note that we need to hold the freelistLock if we use
5450 // block iterate below; else the iterator might go awry if
5451 // a mutator (or promotion) causes block contents to change
5452 // (for instance if the allocator divvies up a block).
5453 // If we hold the free list lock, for all practical purposes
5454 // young generation GC's can't occur (they'll usually need to
5455 // promote), so we might as well prevent all young generation
5456 // GC's while we do a sweeping step. For the same reason, we might
5457 // as well take the bit map lock for the entire duration
5458
5459 // check that we hold the requisite locks
5460 assert(have_cms_token(), "Should hold cms token");
5461 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5462 assert_lock_strong(old_gen->freelistLock());
5463 assert_lock_strong(bitMapLock());
5464
5465 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5466 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5467 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5468 _inter_sweep_estimate.padded_average(),
5469 _intra_sweep_estimate.padded_average());
5470 old_gen->setNearLargestChunk();
5471
5472 {
5473 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5474 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5475 // We need to free-up/coalesce garbage/blocks from a
5476 // co-terminal free run. This is done in the SweepClosure
5477 // destructor; so, do not remove this scope, else the
5478 // end-of-sweep-census below will be off by a little bit.
5479 }
5480 old_gen->cmsSpace()->sweep_completed();
5481 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5482 if (should_unload_classes()) { // unloaded classes this cycle,
5483 _concurrent_cycles_since_last_unload = 0; // ... reset count
5484 } else { // did not unload classes,
5485 _concurrent_cycles_since_last_unload++; // ... increment count
5486 }
5487 }
5488
5489 // Reset CMS data structures (for now just the marking bit map)
5490 // preparatory for the next cycle.
5491 void CMSCollector::reset_concurrent() {
5492 CMSTokenSyncWithLocks ts(true, bitMapLock());
5493
5494 // If the state is not "Resetting", the foreground thread
5495 // has done a collection and the resetting.
5496 if (_collectorState != Resetting) {
5497 assert(_collectorState == Idling, "The state should only change"
5498 " because the foreground collector has finished the collection");
5499 return;
5500 }
5501
5502 {
5503 // Clear the mark bitmap (no grey objects to start with)
5504 // for the next cycle.
5505 GCTraceCPUTime tcpu;
5506 CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5507
5508 HeapWord* curAddr = _markBitMap.startWord();
5509 while (curAddr < _markBitMap.endWord()) {
5510 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5511 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5512 _markBitMap.clear_large_range(chunk);
5513 if (ConcurrentMarkSweepThread::should_yield() &&
5514 !foregroundGCIsActive() &&
5515 CMSYield) {
5516 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5517 "CMS thread should hold CMS token");
5518 assert_lock_strong(bitMapLock());
5519 bitMapLock()->unlock();
5520 ConcurrentMarkSweepThread::desynchronize(true);
5521 stopTimer();
5522 incrementYields();
5523
5524 // See the comment in coordinator_yield()
5525 for (unsigned i = 0; i < CMSYieldSleepCount &&
5526 ConcurrentMarkSweepThread::should_yield() &&
5527 !CMSCollector::foregroundGCIsActive(); ++i) {
5528 os::sleep(Thread::current(), 1, false);
5529 }
5530
5531 ConcurrentMarkSweepThread::synchronize(true);
5532 bitMapLock()->lock_without_safepoint_check();
5533 startTimer();
5534 }
5535 curAddr = chunk.end();
5536 }
5537 // A successful mostly concurrent collection has been done.
5538 // Because only the full (i.e., concurrent mode failure) collections
5539 // are being measured for gc overhead limits, clean the "near" flag
5540 // and count.
5541 size_policy()->reset_gc_overhead_limit_count();
5542 _collectorState = Idling;
5543 }
5544
5545 register_gc_end();
5546 }
5547
5548 // Same as above but for STW paths
5549 void CMSCollector::reset_stw() {
5550 // already have the lock
5551 assert(_collectorState == Resetting, "just checking");
5552 assert_lock_strong(bitMapLock());
5553 GCIdMark gc_id_mark(_cmsThread->gc_id());
5554 _markBitMap.clear_all();
5555 _collectorState = Idling;
5556 register_gc_end();
5557 }
5558
5559 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5560 GCTraceCPUTime tcpu;
5561 TraceCollectorStats tcs_cgc(cgc_counters());
5562
5563 switch (op) {
5564 case CMS_op_checkpointRootsInitial: {
5565 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5566 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5567 checkpointRootsInitial();
5568 break;
5569 }
5570 case CMS_op_checkpointRootsFinal: {
5571 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5572 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5573 checkpointRootsFinal();
5574 break;
5575 }
5576 default:
5577 fatal("No such CMS_op");
5578 }
5579 }
5580
5581 #ifndef PRODUCT
5582 size_t const CMSCollector::skip_header_HeapWords() {
5583 return FreeChunk::header_size();
5584 }
5585
5586 // Try and collect here conditions that should hold when
5587 // CMS thread is exiting. The idea is that the foreground GC
5588 // thread should not be blocked if it wants to terminate
5589 // the CMS thread and yet continue to run the VM for a while
5590 // after that.
5591 void CMSCollector::verify_ok_to_terminate() const {
5592 assert(Thread::current()->is_ConcurrentGC_thread(),
5593 "should be called by CMS thread");
5594 assert(!_foregroundGCShouldWait, "should be false");
5595 // We could check here that all the various low-level locks
5596 // are not held by the CMS thread, but that is overkill; see
5597 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5598 // is checked.
5599 }
5600 #endif
5601
5602 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5603 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5604 "missing Printezis mark?");
5605 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5606 size_t size = pointer_delta(nextOneAddr + 1, addr);
5607 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5608 "alignment problem");
5609 assert(size >= 3, "Necessary for Printezis marks to work");
5610 return size;
5611 }
5612
5613 // A variant of the above (block_size_using_printezis_bits()) except
5614 // that we return 0 if the P-bits are not yet set.
5615 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5616 if (_markBitMap.isMarked(addr + 1)) {
5617 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5618 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5619 size_t size = pointer_delta(nextOneAddr + 1, addr);
5620 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5621 "alignment problem");
5622 assert(size >= 3, "Necessary for Printezis marks to work");
5623 return size;
5624 }
5625 return 0;
5626 }
5627
5628 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5629 size_t sz = 0;
5630 oop p = (oop)addr;
5631 if (p->klass_or_null_acquire() != NULL) {
5632 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5633 } else {
5634 sz = block_size_using_printezis_bits(addr);
5635 }
5636 assert(sz > 0, "size must be nonzero");
5637 HeapWord* next_block = addr + sz;
5638 HeapWord* next_card = align_up(next_block, CardTable::card_size);
5639 assert(align_down((uintptr_t)addr, CardTable::card_size) <
5640 align_down((uintptr_t)next_card, CardTable::card_size),
5641 "must be different cards");
5642 return next_card;
5643 }
5644
5645
5646 // CMS Bit Map Wrapper /////////////////////////////////////////
5647
5648 // Construct a CMS bit map infrastructure, but don't create the
5649 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5650 // further below.
5651 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5652 _bm(),
5653 _shifter(shifter),
5654 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5655 Monitor::_safepoint_check_sometimes) : NULL)
5656 {
5657 _bmStartWord = 0;
5658 _bmWordSize = 0;
5659 }
5660
5661 bool CMSBitMap::allocate(MemRegion mr) {
5662 _bmStartWord = mr.start();
5663 _bmWordSize = mr.word_size();
5664 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5665 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5666 if (!brs.is_reserved()) {
5667 log_warning(gc)("CMS bit map allocation failure");
5668 return false;
5669 }
5670 // For now we'll just commit all of the bit map up front.
5671 // Later on we'll try to be more parsimonious with swap.
5672 if (!_virtual_space.initialize(brs, brs.size())) {
5673 log_warning(gc)("CMS bit map backing store failure");
5674 return false;
5675 }
5676 assert(_virtual_space.committed_size() == brs.size(),
5677 "didn't reserve backing store for all of CMS bit map?");
5678 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5679 _bmWordSize, "inconsistency in bit map sizing");
5680 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5681
5682 // bm.clear(); // can we rely on getting zero'd memory? verify below
5683 assert(isAllClear(),
5684 "Expected zero'd memory from ReservedSpace constructor");
5685 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5686 "consistency check");
5687 return true;
5688 }
5689
5690 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5691 HeapWord *next_addr, *end_addr, *last_addr;
5692 assert_locked();
5693 assert(covers(mr), "out-of-range error");
5694 // XXX assert that start and end are appropriately aligned
5695 for (next_addr = mr.start(), end_addr = mr.end();
5696 next_addr < end_addr; next_addr = last_addr) {
5697 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5698 last_addr = dirty_region.end();
5699 if (!dirty_region.is_empty()) {
5700 cl->do_MemRegion(dirty_region);
5701 } else {
5702 assert(last_addr == end_addr, "program logic");
5703 return;
5704 }
5705 }
5706 }
5707
5708 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5709 _bm.print_on_error(st, prefix);
5710 }
5711
5712 #ifndef PRODUCT
5713 void CMSBitMap::assert_locked() const {
5714 CMSLockVerifier::assert_locked(lock());
5715 }
5716
5717 bool CMSBitMap::covers(MemRegion mr) const {
5718 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5719 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5720 "size inconsistency");
5721 return (mr.start() >= _bmStartWord) &&
5722 (mr.end() <= endWord());
5723 }
5724
5725 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5726 return (start >= _bmStartWord && (start + size) <= endWord());
5727 }
5728
5729 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5730 // verify that there are no 1 bits in the interval [left, right)
5731 FalseBitMapClosure falseBitMapClosure;
5732 iterate(&falseBitMapClosure, left, right);
5733 }
5734
5735 void CMSBitMap::region_invariant(MemRegion mr)
5736 {
5737 assert_locked();
5738 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5739 assert(!mr.is_empty(), "unexpected empty region");
5740 assert(covers(mr), "mr should be covered by bit map");
5741 // convert address range into offset range
5742 size_t start_ofs = heapWordToOffset(mr.start());
5743 // Make sure that end() is appropriately aligned
5744 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))),
5745 "Misaligned mr.end()");
5746 size_t end_ofs = heapWordToOffset(mr.end());
5747 assert(end_ofs > start_ofs, "Should mark at least one bit");
5748 }
5749
5750 #endif
5751
5752 bool CMSMarkStack::allocate(size_t size) {
5753 // allocate a stack of the requisite depth
5754 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5755 size * sizeof(oop)));
5756 if (!rs.is_reserved()) {
5757 log_warning(gc)("CMSMarkStack allocation failure");
5758 return false;
5759 }
5760 if (!_virtual_space.initialize(rs, rs.size())) {
5761 log_warning(gc)("CMSMarkStack backing store failure");
5762 return false;
5763 }
5764 assert(_virtual_space.committed_size() == rs.size(),
5765 "didn't reserve backing store for all of CMS stack?");
5766 _base = (oop*)(_virtual_space.low());
5767 _index = 0;
5768 _capacity = size;
5769 NOT_PRODUCT(_max_depth = 0);
5770 return true;
5771 }
5772
5773 // XXX FIX ME !!! In the MT case we come in here holding a
5774 // leaf lock. For printing we need to take a further lock
5775 // which has lower rank. We need to recalibrate the two
5776 // lock-ranks involved in order to be able to print the
5777 // messages below. (Or defer the printing to the caller.
5778 // For now we take the expedient path of just disabling the
5779 // messages for the problematic case.)
5780 void CMSMarkStack::expand() {
5781 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5782 if (_capacity == MarkStackSizeMax) {
5783 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5784 // We print a warning message only once per CMS cycle.
5785 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5786 }
5787 return;
5788 }
5789 // Double capacity if possible
5790 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5791 // Do not give up existing stack until we have managed to
5792 // get the double capacity that we desired.
5793 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5794 new_capacity * sizeof(oop)));
5795 if (rs.is_reserved()) {
5796 // Release the backing store associated with old stack
5797 _virtual_space.release();
5798 // Reinitialize virtual space for new stack
5799 if (!_virtual_space.initialize(rs, rs.size())) {
5800 fatal("Not enough swap for expanded marking stack");
5801 }
5802 _base = (oop*)(_virtual_space.low());
5803 _index = 0;
5804 _capacity = new_capacity;
5805 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5806 // Failed to double capacity, continue;
5807 // we print a detail message only once per CMS cycle.
5808 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5809 _capacity / K, new_capacity / K);
5810 }
5811 }
5812
5813
5814 // Closures
5815 // XXX: there seems to be a lot of code duplication here;
5816 // should refactor and consolidate common code.
5817
5818 // This closure is used to mark refs into the CMS generation in
5819 // the CMS bit map. Called at the first checkpoint. This closure
5820 // assumes that we do not need to re-mark dirty cards; if the CMS
5821 // generation on which this is used is not an oldest
5822 // generation then this will lose younger_gen cards!
5823
5824 MarkRefsIntoClosure::MarkRefsIntoClosure(
5825 MemRegion span, CMSBitMap* bitMap):
5826 _span(span),
5827 _bitMap(bitMap)
5828 {
5829 assert(ref_discoverer() == NULL, "deliberately left NULL");
5830 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5831 }
5832
5833 void MarkRefsIntoClosure::do_oop(oop obj) {
5834 // if p points into _span, then mark corresponding bit in _markBitMap
5835 assert(oopDesc::is_oop(obj), "expected an oop");
5836 HeapWord* addr = (HeapWord*)obj;
5837 if (_span.contains(addr)) {
5838 // this should be made more efficient
5839 _bitMap->mark(addr);
5840 }
5841 }
5842
5843 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5844 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5845
5846 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5847 MemRegion span, CMSBitMap* bitMap):
5848 _span(span),
5849 _bitMap(bitMap)
5850 {
5851 assert(ref_discoverer() == NULL, "deliberately left NULL");
5852 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5853 }
5854
5855 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5856 // if p points into _span, then mark corresponding bit in _markBitMap
5857 assert(oopDesc::is_oop(obj), "expected an oop");
5858 HeapWord* addr = (HeapWord*)obj;
5859 if (_span.contains(addr)) {
5860 // this should be made more efficient
5861 _bitMap->par_mark(addr);
5862 }
5863 }
5864
5865 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5866 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5867
5868 // A variant of the above, used for CMS marking verification.
5869 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5870 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5871 _span(span),
5872 _verification_bm(verification_bm),
5873 _cms_bm(cms_bm)
5874 {
5875 assert(ref_discoverer() == NULL, "deliberately left NULL");
5876 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5877 }
5878
5879 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5880 // if p points into _span, then mark corresponding bit in _markBitMap
5881 assert(oopDesc::is_oop(obj), "expected an oop");
5882 HeapWord* addr = (HeapWord*)obj;
5883 if (_span.contains(addr)) {
5884 _verification_bm->mark(addr);
5885 if (!_cms_bm->isMarked(addr)) {
5886 Log(gc, verify) log;
5887 ResourceMark rm;
5888 LogStream ls(log.error());
5889 oop(addr)->print_on(&ls);
5890 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5891 fatal("... aborting");
5892 }
5893 }
5894 }
5895
5896 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5897 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5898
5899 //////////////////////////////////////////////////
5900 // MarkRefsIntoAndScanClosure
5901 //////////////////////////////////////////////////
5902
5903 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5904 ReferenceDiscoverer* rd,
5905 CMSBitMap* bit_map,
5906 CMSBitMap* mod_union_table,
5907 CMSMarkStack* mark_stack,
5908 CMSCollector* collector,
5909 bool should_yield,
5910 bool concurrent_precleaning):
5911 _collector(collector),
5912 _span(span),
5913 _bit_map(bit_map),
5914 _mark_stack(mark_stack),
5915 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table,
5916 mark_stack, concurrent_precleaning),
5917 _yield(should_yield),
5918 _concurrent_precleaning(concurrent_precleaning),
5919 _freelistLock(NULL)
5920 {
5921 // FIXME: Should initialize in base class constructor.
5922 assert(rd != NULL, "ref_discoverer shouldn't be NULL");
5923 set_ref_discoverer_internal(rd);
5924 }
5925
5926 // This closure is used to mark refs into the CMS generation at the
5927 // second (final) checkpoint, and to scan and transitively follow
5928 // the unmarked oops. It is also used during the concurrent precleaning
5929 // phase while scanning objects on dirty cards in the CMS generation.
5930 // The marks are made in the marking bit map and the marking stack is
5931 // used for keeping the (newly) grey objects during the scan.
5932 // The parallel version (Par_...) appears further below.
5933 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5934 if (obj != NULL) {
5935 assert(oopDesc::is_oop(obj), "expected an oop");
5936 HeapWord* addr = (HeapWord*)obj;
5937 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5938 assert(_collector->overflow_list_is_empty(),
5939 "overflow list should be empty");
5940 if (_span.contains(addr) &&
5941 !_bit_map->isMarked(addr)) {
5942 // mark bit map (object is now grey)
5943 _bit_map->mark(addr);
5944 // push on marking stack (stack should be empty), and drain the
5945 // stack by applying this closure to the oops in the oops popped
5946 // from the stack (i.e. blacken the grey objects)
5947 bool res = _mark_stack->push(obj);
5948 assert(res, "Should have space to push on empty stack");
5949 do {
5950 oop new_oop = _mark_stack->pop();
5951 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5952 assert(_bit_map->isMarked((HeapWord*)new_oop),
5953 "only grey objects on this stack");
5954 // iterate over the oops in this oop, marking and pushing
5955 // the ones in CMS heap (i.e. in _span).
5956 new_oop->oop_iterate(&_pushAndMarkClosure);
5957 // check if it's time to yield
5958 do_yield_check();
5959 } while (!_mark_stack->isEmpty() ||
5960 (!_concurrent_precleaning && take_from_overflow_list()));
5961 // if marking stack is empty, and we are not doing this
5962 // during precleaning, then check the overflow list
5963 }
5964 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5965 assert(_collector->overflow_list_is_empty(),
5966 "overflow list was drained above");
5967
5968 assert(_collector->no_preserved_marks(),
5969 "All preserved marks should have been restored above");
5970 }
5971 }
5972
5973 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5974 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5975
5976 void MarkRefsIntoAndScanClosure::do_yield_work() {
5977 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5978 "CMS thread should hold CMS token");
5979 assert_lock_strong(_freelistLock);
5980 assert_lock_strong(_bit_map->lock());
5981 // relinquish the free_list_lock and bitMaplock()
5982 _bit_map->lock()->unlock();
5983 _freelistLock->unlock();
5984 ConcurrentMarkSweepThread::desynchronize(true);
5985 _collector->stopTimer();
5986 _collector->incrementYields();
5987
5988 // See the comment in coordinator_yield()
5989 for (unsigned i = 0;
5990 i < CMSYieldSleepCount &&
5991 ConcurrentMarkSweepThread::should_yield() &&
5992 !CMSCollector::foregroundGCIsActive();
5993 ++i) {
5994 os::sleep(Thread::current(), 1, false);
5995 }
5996
5997 ConcurrentMarkSweepThread::synchronize(true);
5998 _freelistLock->lock_without_safepoint_check();
5999 _bit_map->lock()->lock_without_safepoint_check();
6000 _collector->startTimer();
6001 }
6002
6003 ///////////////////////////////////////////////////////////
6004 // ParMarkRefsIntoAndScanClosure: a parallel version of
6005 // MarkRefsIntoAndScanClosure
6006 ///////////////////////////////////////////////////////////
6007 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6008 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd,
6009 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6010 _span(span),
6011 _bit_map(bit_map),
6012 _work_queue(work_queue),
6013 _low_water_mark(MIN2((work_queue->max_elems()/4),
6014 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6015 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue)
6016 {
6017 // FIXME: Should initialize in base class constructor.
6018 assert(rd != NULL, "ref_discoverer shouldn't be NULL");
6019 set_ref_discoverer_internal(rd);
6020 }
6021
6022 // This closure is used to mark refs into the CMS generation at the
6023 // second (final) checkpoint, and to scan and transitively follow
6024 // the unmarked oops. The marks are made in the marking bit map and
6025 // the work_queue is used for keeping the (newly) grey objects during
6026 // the scan phase whence they are also available for stealing by parallel
6027 // threads. Since the marking bit map is shared, updates are
6028 // synchronized (via CAS).
6029 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6030 if (obj != NULL) {
6031 // Ignore mark word because this could be an already marked oop
6032 // that may be chained at the end of the overflow list.
6033 assert(oopDesc::is_oop(obj, true), "expected an oop");
6034 HeapWord* addr = (HeapWord*)obj;
6035 if (_span.contains(addr) &&
6036 !_bit_map->isMarked(addr)) {
6037 // mark bit map (object will become grey):
6038 // It is possible for several threads to be
6039 // trying to "claim" this object concurrently;
6040 // the unique thread that succeeds in marking the
6041 // object first will do the subsequent push on
6042 // to the work queue (or overflow list).
6043 if (_bit_map->par_mark(addr)) {
6044 // push on work_queue (which may not be empty), and trim the
6045 // queue to an appropriate length by applying this closure to
6046 // the oops in the oops popped from the stack (i.e. blacken the
6047 // grey objects)
6048 bool res = _work_queue->push(obj);
6049 assert(res, "Low water mark should be less than capacity?");
6050 trim_queue(_low_water_mark);
6051 } // Else, another thread claimed the object
6052 }
6053 }
6054 }
6055
6056 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6057 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6058
6059 // This closure is used to rescan the marked objects on the dirty cards
6060 // in the mod union table and the card table proper.
6061 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6062 oop p, MemRegion mr) {
6063
6064 size_t size = 0;
6065 HeapWord* addr = (HeapWord*)p;
6066 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6067 assert(_span.contains(addr), "we are scanning the CMS generation");
6068 // check if it's time to yield
6069 if (do_yield_check()) {
6070 // We yielded for some foreground stop-world work,
6071 // and we have been asked to abort this ongoing preclean cycle.
6072 return 0;
6073 }
6074 if (_bitMap->isMarked(addr)) {
6075 // it's marked; is it potentially uninitialized?
6076 if (p->klass_or_null_acquire() != NULL) {
6077 // an initialized object; ignore mark word in verification below
6078 // since we are running concurrent with mutators
6079 assert(oopDesc::is_oop(p, true), "should be an oop");
6080 if (p->is_objArray()) {
6081 // objArrays are precisely marked; restrict scanning
6082 // to dirty cards only.
6083 size = CompactibleFreeListSpace::adjustObjectSize(
6084 p->oop_iterate_size(_scanningClosure, mr));
6085 } else {
6086 // A non-array may have been imprecisely marked; we need
6087 // to scan object in its entirety.
6088 size = CompactibleFreeListSpace::adjustObjectSize(
6089 p->oop_iterate_size(_scanningClosure));
6090 }
6091 #ifdef ASSERT
6092 size_t direct_size =
6093 CompactibleFreeListSpace::adjustObjectSize(p->size());
6094 assert(size == direct_size, "Inconsistency in size");
6095 assert(size >= 3, "Necessary for Printezis marks to work");
6096 HeapWord* start_pbit = addr + 1;
6097 HeapWord* end_pbit = addr + size - 1;
6098 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6099 "inconsistent Printezis mark");
6100 // Verify inner mark bits (between Printezis bits) are clear,
6101 // but don't repeat if there are multiple dirty regions for
6102 // the same object, to avoid potential O(N^2) performance.
6103 if (addr != _last_scanned_object) {
6104 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6105 _last_scanned_object = addr;
6106 }
6107 #endif // ASSERT
6108 } else {
6109 // An uninitialized object.
6110 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6111 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6112 size = pointer_delta(nextOneAddr + 1, addr);
6113 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6114 "alignment problem");
6115 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6116 // will dirty the card when the klass pointer is installed in the
6117 // object (signaling the completion of initialization).
6118 }
6119 } else {
6120 // Either a not yet marked object or an uninitialized object
6121 if (p->klass_or_null_acquire() == NULL) {
6122 // An uninitialized object, skip to the next card, since
6123 // we may not be able to read its P-bits yet.
6124 assert(size == 0, "Initial value");
6125 } else {
6126 // An object not (yet) reached by marking: we merely need to
6127 // compute its size so as to go look at the next block.
6128 assert(oopDesc::is_oop(p, true), "should be an oop");
6129 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6130 }
6131 }
6132 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6133 return size;
6134 }
6135
6136 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6137 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6138 "CMS thread should hold CMS token");
6139 assert_lock_strong(_freelistLock);
6140 assert_lock_strong(_bitMap->lock());
6141 // relinquish the free_list_lock and bitMaplock()
6142 _bitMap->lock()->unlock();
6143 _freelistLock->unlock();
6144 ConcurrentMarkSweepThread::desynchronize(true);
6145 _collector->stopTimer();
6146 _collector->incrementYields();
6147
6148 // See the comment in coordinator_yield()
6149 for (unsigned i = 0; i < CMSYieldSleepCount &&
6150 ConcurrentMarkSweepThread::should_yield() &&
6151 !CMSCollector::foregroundGCIsActive(); ++i) {
6152 os::sleep(Thread::current(), 1, false);
6153 }
6154
6155 ConcurrentMarkSweepThread::synchronize(true);
6156 _freelistLock->lock_without_safepoint_check();
6157 _bitMap->lock()->lock_without_safepoint_check();
6158 _collector->startTimer();
6159 }
6160
6161
6162 //////////////////////////////////////////////////////////////////
6163 // SurvivorSpacePrecleanClosure
6164 //////////////////////////////////////////////////////////////////
6165 // This (single-threaded) closure is used to preclean the oops in
6166 // the survivor spaces.
6167 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6168
6169 HeapWord* addr = (HeapWord*)p;
6170 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6171 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6172 assert(p->klass_or_null() != NULL, "object should be initialized");
6173 // an initialized object; ignore mark word in verification below
6174 // since we are running concurrent with mutators
6175 assert(oopDesc::is_oop(p, true), "should be an oop");
6176 // Note that we do not yield while we iterate over
6177 // the interior oops of p, pushing the relevant ones
6178 // on our marking stack.
6179 size_t size = p->oop_iterate_size(_scanning_closure);
6180 do_yield_check();
6181 // Observe that below, we do not abandon the preclean
6182 // phase as soon as we should; rather we empty the
6183 // marking stack before returning. This is to satisfy
6184 // some existing assertions. In general, it may be a
6185 // good idea to abort immediately and complete the marking
6186 // from the grey objects at a later time.
6187 while (!_mark_stack->isEmpty()) {
6188 oop new_oop = _mark_stack->pop();
6189 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6190 assert(_bit_map->isMarked((HeapWord*)new_oop),
6191 "only grey objects on this stack");
6192 // iterate over the oops in this oop, marking and pushing
6193 // the ones in CMS heap (i.e. in _span).
6194 new_oop->oop_iterate(_scanning_closure);
6195 // check if it's time to yield
6196 do_yield_check();
6197 }
6198 unsigned int after_count =
6199 CMSHeap::heap()->total_collections();
6200 bool abort = (_before_count != after_count) ||
6201 _collector->should_abort_preclean();
6202 return abort ? 0 : size;
6203 }
6204
6205 void SurvivorSpacePrecleanClosure::do_yield_work() {
6206 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6207 "CMS thread should hold CMS token");
6208 assert_lock_strong(_bit_map->lock());
6209 // Relinquish the bit map lock
6210 _bit_map->lock()->unlock();
6211 ConcurrentMarkSweepThread::desynchronize(true);
6212 _collector->stopTimer();
6213 _collector->incrementYields();
6214
6215 // See the comment in coordinator_yield()
6216 for (unsigned i = 0; i < CMSYieldSleepCount &&
6217 ConcurrentMarkSweepThread::should_yield() &&
6218 !CMSCollector::foregroundGCIsActive(); ++i) {
6219 os::sleep(Thread::current(), 1, false);
6220 }
6221
6222 ConcurrentMarkSweepThread::synchronize(true);
6223 _bit_map->lock()->lock_without_safepoint_check();
6224 _collector->startTimer();
6225 }
6226
6227 // This closure is used to rescan the marked objects on the dirty cards
6228 // in the mod union table and the card table proper. In the parallel
6229 // case, although the bitMap is shared, we do a single read so the
6230 // isMarked() query is "safe".
6231 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6232 // Ignore mark word because we are running concurrent with mutators
6233 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6234 HeapWord* addr = (HeapWord*)p;
6235 assert(_span.contains(addr), "we are scanning the CMS generation");
6236 bool is_obj_array = false;
6237 #ifdef ASSERT
6238 if (!_parallel) {
6239 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6240 assert(_collector->overflow_list_is_empty(),
6241 "overflow list should be empty");
6242
6243 }
6244 #endif // ASSERT
6245 if (_bit_map->isMarked(addr)) {
6246 // Obj arrays are precisely marked, non-arrays are not;
6247 // so we scan objArrays precisely and non-arrays in their
6248 // entirety.
6249 if (p->is_objArray()) {
6250 is_obj_array = true;
6251 if (_parallel) {
6252 p->oop_iterate(_par_scan_closure, mr);
6253 } else {
6254 p->oop_iterate(_scan_closure, mr);
6255 }
6256 } else {
6257 if (_parallel) {
6258 p->oop_iterate(_par_scan_closure);
6259 } else {
6260 p->oop_iterate(_scan_closure);
6261 }
6262 }
6263 }
6264 #ifdef ASSERT
6265 if (!_parallel) {
6266 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6267 assert(_collector->overflow_list_is_empty(),
6268 "overflow list should be empty");
6269
6270 }
6271 #endif // ASSERT
6272 return is_obj_array;
6273 }
6274
6275 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6276 MemRegion span,
6277 CMSBitMap* bitMap, CMSMarkStack* markStack,
6278 bool should_yield, bool verifying):
6279 _collector(collector),
6280 _span(span),
6281 _bitMap(bitMap),
6282 _mut(&collector->_modUnionTable),
6283 _markStack(markStack),
6284 _yield(should_yield),
6285 _skipBits(0)
6286 {
6287 assert(_markStack->isEmpty(), "stack should be empty");
6288 _finger = _bitMap->startWord();
6289 _threshold = _finger;
6290 assert(_collector->_restart_addr == NULL, "Sanity check");
6291 assert(_span.contains(_finger), "Out of bounds _finger?");
6292 DEBUG_ONLY(_verifying = verifying;)
6293 }
6294
6295 void MarkFromRootsClosure::reset(HeapWord* addr) {
6296 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6297 assert(_span.contains(addr), "Out of bounds _finger?");
6298 _finger = addr;
6299 _threshold = align_up(_finger, CardTable::card_size);
6300 }
6301
6302 // Should revisit to see if this should be restructured for
6303 // greater efficiency.
6304 bool MarkFromRootsClosure::do_bit(size_t offset) {
6305 if (_skipBits > 0) {
6306 _skipBits--;
6307 return true;
6308 }
6309 // convert offset into a HeapWord*
6310 HeapWord* addr = _bitMap->startWord() + offset;
6311 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6312 "address out of range");
6313 assert(_bitMap->isMarked(addr), "tautology");
6314 if (_bitMap->isMarked(addr+1)) {
6315 // this is an allocated but not yet initialized object
6316 assert(_skipBits == 0, "tautology");
6317 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6318 oop p = oop(addr);
6319 if (p->klass_or_null_acquire() == NULL) {
6320 DEBUG_ONLY(if (!_verifying) {)
6321 // We re-dirty the cards on which this object lies and increase
6322 // the _threshold so that we'll come back to scan this object
6323 // during the preclean or remark phase. (CMSCleanOnEnter)
6324 if (CMSCleanOnEnter) {
6325 size_t sz = _collector->block_size_using_printezis_bits(addr);
6326 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6327 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6328 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6329 // Bump _threshold to end_card_addr; note that
6330 // _threshold cannot possibly exceed end_card_addr, anyhow.
6331 // This prevents future clearing of the card as the scan proceeds
6332 // to the right.
6333 assert(_threshold <= end_card_addr,
6334 "Because we are just scanning into this object");
6335 if (_threshold < end_card_addr) {
6336 _threshold = end_card_addr;
6337 }
6338 if (p->klass_or_null_acquire() != NULL) {
6339 // Redirty the range of cards...
6340 _mut->mark_range(redirty_range);
6341 } // ...else the setting of klass will dirty the card anyway.
6342 }
6343 DEBUG_ONLY(})
6344 return true;
6345 }
6346 }
6347 scanOopsInOop(addr);
6348 return true;
6349 }
6350
6351 // We take a break if we've been at this for a while,
6352 // so as to avoid monopolizing the locks involved.
6353 void MarkFromRootsClosure::do_yield_work() {
6354 // First give up the locks, then yield, then re-lock
6355 // We should probably use a constructor/destructor idiom to
6356 // do this unlock/lock or modify the MutexUnlocker class to
6357 // serve our purpose. XXX
6358 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6359 "CMS thread should hold CMS token");
6360 assert_lock_strong(_bitMap->lock());
6361 _bitMap->lock()->unlock();
6362 ConcurrentMarkSweepThread::desynchronize(true);
6363 _collector->stopTimer();
6364 _collector->incrementYields();
6365
6366 // See the comment in coordinator_yield()
6367 for (unsigned i = 0; i < CMSYieldSleepCount &&
6368 ConcurrentMarkSweepThread::should_yield() &&
6369 !CMSCollector::foregroundGCIsActive(); ++i) {
6370 os::sleep(Thread::current(), 1, false);
6371 }
6372
6373 ConcurrentMarkSweepThread::synchronize(true);
6374 _bitMap->lock()->lock_without_safepoint_check();
6375 _collector->startTimer();
6376 }
6377
6378 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6379 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6380 assert(_markStack->isEmpty(),
6381 "should drain stack to limit stack usage");
6382 // convert ptr to an oop preparatory to scanning
6383 oop obj = oop(ptr);
6384 // Ignore mark word in verification below, since we
6385 // may be running concurrent with mutators.
6386 assert(oopDesc::is_oop(obj, true), "should be an oop");
6387 assert(_finger <= ptr, "_finger runneth ahead");
6388 // advance the finger to right end of this object
6389 _finger = ptr + obj->size();
6390 assert(_finger > ptr, "we just incremented it above");
6391 // On large heaps, it may take us some time to get through
6392 // the marking phase. During
6393 // this time it's possible that a lot of mutations have
6394 // accumulated in the card table and the mod union table --
6395 // these mutation records are redundant until we have
6396 // actually traced into the corresponding card.
6397 // Here, we check whether advancing the finger would make
6398 // us cross into a new card, and if so clear corresponding
6399 // cards in the MUT (preclean them in the card-table in the
6400 // future).
6401
6402 DEBUG_ONLY(if (!_verifying) {)
6403 // The clean-on-enter optimization is disabled by default,
6404 // until we fix 6178663.
6405 if (CMSCleanOnEnter && (_finger > _threshold)) {
6406 // [_threshold, _finger) represents the interval
6407 // of cards to be cleared in MUT (or precleaned in card table).
6408 // The set of cards to be cleared is all those that overlap
6409 // with the interval [_threshold, _finger); note that
6410 // _threshold is always kept card-aligned but _finger isn't
6411 // always card-aligned.
6412 HeapWord* old_threshold = _threshold;
6413 assert(is_aligned(old_threshold, CardTable::card_size),
6414 "_threshold should always be card-aligned");
6415 _threshold = align_up(_finger, CardTable::card_size);
6416 MemRegion mr(old_threshold, _threshold);
6417 assert(!mr.is_empty(), "Control point invariant");
6418 assert(_span.contains(mr), "Should clear within span");
6419 _mut->clear_range(mr);
6420 }
6421 DEBUG_ONLY(})
6422 // Note: the finger doesn't advance while we drain
6423 // the stack below.
6424 PushOrMarkClosure pushOrMarkClosure(_collector,
6425 _span, _bitMap, _markStack,
6426 _finger, this);
6427 bool res = _markStack->push(obj);
6428 assert(res, "Empty non-zero size stack should have space for single push");
6429 while (!_markStack->isEmpty()) {
6430 oop new_oop = _markStack->pop();
6431 // Skip verifying header mark word below because we are
6432 // running concurrent with mutators.
6433 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6434 // now scan this oop's oops
6435 new_oop->oop_iterate(&pushOrMarkClosure);
6436 do_yield_check();
6437 }
6438 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6439 }
6440
6441 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6442 CMSCollector* collector, MemRegion span,
6443 CMSBitMap* bit_map,
6444 OopTaskQueue* work_queue,
6445 CMSMarkStack* overflow_stack):
6446 _collector(collector),
6447 _whole_span(collector->_span),
6448 _span(span),
6449 _bit_map(bit_map),
6450 _mut(&collector->_modUnionTable),
6451 _work_queue(work_queue),
6452 _overflow_stack(overflow_stack),
6453 _skip_bits(0),
6454 _task(task)
6455 {
6456 assert(_work_queue->size() == 0, "work_queue should be empty");
6457 _finger = span.start();
6458 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6459 assert(_span.contains(_finger), "Out of bounds _finger?");
6460 }
6461
6462 // Should revisit to see if this should be restructured for
6463 // greater efficiency.
6464 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6465 if (_skip_bits > 0) {
6466 _skip_bits--;
6467 return true;
6468 }
6469 // convert offset into a HeapWord*
6470 HeapWord* addr = _bit_map->startWord() + offset;
6471 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6472 "address out of range");
6473 assert(_bit_map->isMarked(addr), "tautology");
6474 if (_bit_map->isMarked(addr+1)) {
6475 // this is an allocated object that might not yet be initialized
6476 assert(_skip_bits == 0, "tautology");
6477 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6478 oop p = oop(addr);
6479 if (p->klass_or_null_acquire() == NULL) {
6480 // in the case of Clean-on-Enter optimization, redirty card
6481 // and avoid clearing card by increasing the threshold.
6482 return true;
6483 }
6484 }
6485 scan_oops_in_oop(addr);
6486 return true;
6487 }
6488
6489 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6490 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6491 // Should we assert that our work queue is empty or
6492 // below some drain limit?
6493 assert(_work_queue->size() == 0,
6494 "should drain stack to limit stack usage");
6495 // convert ptr to an oop preparatory to scanning
6496 oop obj = oop(ptr);
6497 // Ignore mark word in verification below, since we
6498 // may be running concurrent with mutators.
6499 assert(oopDesc::is_oop(obj, true), "should be an oop");
6500 assert(_finger <= ptr, "_finger runneth ahead");
6501 // advance the finger to right end of this object
6502 _finger = ptr + obj->size();
6503 assert(_finger > ptr, "we just incremented it above");
6504 // On large heaps, it may take us some time to get through
6505 // the marking phase. During
6506 // this time it's possible that a lot of mutations have
6507 // accumulated in the card table and the mod union table --
6508 // these mutation records are redundant until we have
6509 // actually traced into the corresponding card.
6510 // Here, we check whether advancing the finger would make
6511 // us cross into a new card, and if so clear corresponding
6512 // cards in the MUT (preclean them in the card-table in the
6513 // future).
6514
6515 // The clean-on-enter optimization is disabled by default,
6516 // until we fix 6178663.
6517 if (CMSCleanOnEnter && (_finger > _threshold)) {
6518 // [_threshold, _finger) represents the interval
6519 // of cards to be cleared in MUT (or precleaned in card table).
6520 // The set of cards to be cleared is all those that overlap
6521 // with the interval [_threshold, _finger); note that
6522 // _threshold is always kept card-aligned but _finger isn't
6523 // always card-aligned.
6524 HeapWord* old_threshold = _threshold;
6525 assert(is_aligned(old_threshold, CardTable::card_size),
6526 "_threshold should always be card-aligned");
6527 _threshold = align_up(_finger, CardTable::card_size);
6528 MemRegion mr(old_threshold, _threshold);
6529 assert(!mr.is_empty(), "Control point invariant");
6530 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6531 _mut->clear_range(mr);
6532 }
6533
6534 // Note: the local finger doesn't advance while we drain
6535 // the stack below, but the global finger sure can and will.
6536 HeapWord* volatile* gfa = _task->global_finger_addr();
6537 ParPushOrMarkClosure pushOrMarkClosure(_collector,
6538 _span, _bit_map,
6539 _work_queue,
6540 _overflow_stack,
6541 _finger,
6542 gfa, this);
6543 bool res = _work_queue->push(obj); // overflow could occur here
6544 assert(res, "Will hold once we use workqueues");
6545 while (true) {
6546 oop new_oop;
6547 if (!_work_queue->pop_local(new_oop)) {
6548 // We emptied our work_queue; check if there's stuff that can
6549 // be gotten from the overflow stack.
6550 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6551 _overflow_stack, _work_queue)) {
6552 do_yield_check();
6553 continue;
6554 } else { // done
6555 break;
6556 }
6557 }
6558 // Skip verifying header mark word below because we are
6559 // running concurrent with mutators.
6560 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6561 // now scan this oop's oops
6562 new_oop->oop_iterate(&pushOrMarkClosure);
6563 do_yield_check();
6564 }
6565 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6566 }
6567
6568 // Yield in response to a request from VM Thread or
6569 // from mutators.
6570 void ParMarkFromRootsClosure::do_yield_work() {
6571 assert(_task != NULL, "sanity");
6572 _task->yield();
6573 }
6574
6575 // A variant of the above used for verifying CMS marking work.
6576 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6577 MemRegion span,
6578 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6579 CMSMarkStack* mark_stack):
6580 _collector(collector),
6581 _span(span),
6582 _verification_bm(verification_bm),
6583 _cms_bm(cms_bm),
6584 _mark_stack(mark_stack),
6585 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6586 mark_stack)
6587 {
6588 assert(_mark_stack->isEmpty(), "stack should be empty");
6589 _finger = _verification_bm->startWord();
6590 assert(_collector->_restart_addr == NULL, "Sanity check");
6591 assert(_span.contains(_finger), "Out of bounds _finger?");
6592 }
6593
6594 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6595 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6596 assert(_span.contains(addr), "Out of bounds _finger?");
6597 _finger = addr;
6598 }
6599
6600 // Should revisit to see if this should be restructured for
6601 // greater efficiency.
6602 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6603 // convert offset into a HeapWord*
6604 HeapWord* addr = _verification_bm->startWord() + offset;
6605 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6606 "address out of range");
6607 assert(_verification_bm->isMarked(addr), "tautology");
6608 assert(_cms_bm->isMarked(addr), "tautology");
6609
6610 assert(_mark_stack->isEmpty(),
6611 "should drain stack to limit stack usage");
6612 // convert addr to an oop preparatory to scanning
6613 oop obj = oop(addr);
6614 assert(oopDesc::is_oop(obj), "should be an oop");
6615 assert(_finger <= addr, "_finger runneth ahead");
6616 // advance the finger to right end of this object
6617 _finger = addr + obj->size();
6618 assert(_finger > addr, "we just incremented it above");
6619 // Note: the finger doesn't advance while we drain
6620 // the stack below.
6621 bool res = _mark_stack->push(obj);
6622 assert(res, "Empty non-zero size stack should have space for single push");
6623 while (!_mark_stack->isEmpty()) {
6624 oop new_oop = _mark_stack->pop();
6625 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6626 // now scan this oop's oops
6627 new_oop->oop_iterate(&_pam_verify_closure);
6628 }
6629 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6630 return true;
6631 }
6632
6633 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6634 CMSCollector* collector, MemRegion span,
6635 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6636 CMSMarkStack* mark_stack):
6637 MetadataAwareOopClosure(collector->ref_processor()),
6638 _collector(collector),
6639 _span(span),
6640 _verification_bm(verification_bm),
6641 _cms_bm(cms_bm),
6642 _mark_stack(mark_stack)
6643 { }
6644
6645 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) {
6646 oop obj = RawAccess<>::oop_load(p);
6647 do_oop(obj);
6648 }
6649
6650 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6651 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6652
6653 // Upon stack overflow, we discard (part of) the stack,
6654 // remembering the least address amongst those discarded
6655 // in CMSCollector's _restart_address.
6656 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6657 // Remember the least grey address discarded
6658 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6659 _collector->lower_restart_addr(ra);
6660 _mark_stack->reset(); // discard stack contents
6661 _mark_stack->expand(); // expand the stack if possible
6662 }
6663
6664 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6665 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6666 HeapWord* addr = (HeapWord*)obj;
6667 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6668 // Oop lies in _span and isn't yet grey or black
6669 _verification_bm->mark(addr); // now grey
6670 if (!_cms_bm->isMarked(addr)) {
6671 Log(gc, verify) log;
6672 ResourceMark rm;
6673 LogStream ls(log.error());
6674 oop(addr)->print_on(&ls);
6675 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6676 fatal("... aborting");
6677 }
6678
6679 if (!_mark_stack->push(obj)) { // stack overflow
6680 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6681 assert(_mark_stack->isFull(), "Else push should have succeeded");
6682 handle_stack_overflow(addr);
6683 }
6684 // anything including and to the right of _finger
6685 // will be scanned as we iterate over the remainder of the
6686 // bit map
6687 }
6688 }
6689
6690 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6691 MemRegion span,
6692 CMSBitMap* bitMap, CMSMarkStack* markStack,
6693 HeapWord* finger, MarkFromRootsClosure* parent) :
6694 MetadataAwareOopClosure(collector->ref_processor()),
6695 _collector(collector),
6696 _span(span),
6697 _bitMap(bitMap),
6698 _markStack(markStack),
6699 _finger(finger),
6700 _parent(parent)
6701 { }
6702
6703 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6704 MemRegion span,
6705 CMSBitMap* bit_map,
6706 OopTaskQueue* work_queue,
6707 CMSMarkStack* overflow_stack,
6708 HeapWord* finger,
6709 HeapWord* volatile* global_finger_addr,
6710 ParMarkFromRootsClosure* parent) :
6711 MetadataAwareOopClosure(collector->ref_processor()),
6712 _collector(collector),
6713 _whole_span(collector->_span),
6714 _span(span),
6715 _bit_map(bit_map),
6716 _work_queue(work_queue),
6717 _overflow_stack(overflow_stack),
6718 _finger(finger),
6719 _global_finger_addr(global_finger_addr),
6720 _parent(parent)
6721 { }
6722
6723 // Assumes thread-safe access by callers, who are
6724 // responsible for mutual exclusion.
6725 void CMSCollector::lower_restart_addr(HeapWord* low) {
6726 assert(_span.contains(low), "Out of bounds addr");
6727 if (_restart_addr == NULL) {
6728 _restart_addr = low;
6729 } else {
6730 _restart_addr = MIN2(_restart_addr, low);
6731 }
6732 }
6733
6734 // Upon stack overflow, we discard (part of) the stack,
6735 // remembering the least address amongst those discarded
6736 // in CMSCollector's _restart_address.
6737 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6738 // Remember the least grey address discarded
6739 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6740 _collector->lower_restart_addr(ra);
6741 _markStack->reset(); // discard stack contents
6742 _markStack->expand(); // expand the stack if possible
6743 }
6744
6745 // Upon stack overflow, we discard (part of) the stack,
6746 // remembering the least address amongst those discarded
6747 // in CMSCollector's _restart_address.
6748 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6749 // We need to do this under a mutex to prevent other
6750 // workers from interfering with the work done below.
6751 MutexLockerEx ml(_overflow_stack->par_lock(),
6752 Mutex::_no_safepoint_check_flag);
6753 // Remember the least grey address discarded
6754 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6755 _collector->lower_restart_addr(ra);
6756 _overflow_stack->reset(); // discard stack contents
6757 _overflow_stack->expand(); // expand the stack if possible
6758 }
6759
6760 void PushOrMarkClosure::do_oop(oop obj) {
6761 // Ignore mark word because we are running concurrent with mutators.
6762 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6763 HeapWord* addr = (HeapWord*)obj;
6764 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6765 // Oop lies in _span and isn't yet grey or black
6766 _bitMap->mark(addr); // now grey
6767 if (addr < _finger) {
6768 // the bit map iteration has already either passed, or
6769 // sampled, this bit in the bit map; we'll need to
6770 // use the marking stack to scan this oop's oops.
6771 bool simulate_overflow = false;
6772 NOT_PRODUCT(
6773 if (CMSMarkStackOverflowALot &&
6774 _collector->simulate_overflow()) {
6775 // simulate a stack overflow
6776 simulate_overflow = true;
6777 }
6778 )
6779 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6780 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6781 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6782 handle_stack_overflow(addr);
6783 }
6784 }
6785 // anything including and to the right of _finger
6786 // will be scanned as we iterate over the remainder of the
6787 // bit map
6788 do_yield_check();
6789 }
6790 }
6791
6792 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
6793 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6794
6795 void ParPushOrMarkClosure::do_oop(oop obj) {
6796 // Ignore mark word because we are running concurrent with mutators.
6797 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6798 HeapWord* addr = (HeapWord*)obj;
6799 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6800 // Oop lies in _span and isn't yet grey or black
6801 // We read the global_finger (volatile read) strictly after marking oop
6802 bool res = _bit_map->par_mark(addr); // now grey
6803 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6804 // Should we push this marked oop on our stack?
6805 // -- if someone else marked it, nothing to do
6806 // -- if target oop is above global finger nothing to do
6807 // -- if target oop is in chunk and above local finger
6808 // then nothing to do
6809 // -- else push on work queue
6810 if ( !res // someone else marked it, they will deal with it
6811 || (addr >= *gfa) // will be scanned in a later task
6812 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6813 return;
6814 }
6815 // the bit map iteration has already either passed, or
6816 // sampled, this bit in the bit map; we'll need to
6817 // use the marking stack to scan this oop's oops.
6818 bool simulate_overflow = false;
6819 NOT_PRODUCT(
6820 if (CMSMarkStackOverflowALot &&
6821 _collector->simulate_overflow()) {
6822 // simulate a stack overflow
6823 simulate_overflow = true;
6824 }
6825 )
6826 if (simulate_overflow ||
6827 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6828 // stack overflow
6829 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6830 // We cannot assert that the overflow stack is full because
6831 // it may have been emptied since.
6832 assert(simulate_overflow ||
6833 _work_queue->size() == _work_queue->max_elems(),
6834 "Else push should have succeeded");
6835 handle_stack_overflow(addr);
6836 }
6837 do_yield_check();
6838 }
6839 }
6840
6841 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6842 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6843
6844 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6845 MemRegion span,
6846 ReferenceDiscoverer* rd,
6847 CMSBitMap* bit_map,
6848 CMSBitMap* mod_union_table,
6849 CMSMarkStack* mark_stack,
6850 bool concurrent_precleaning):
6851 MetadataAwareOopClosure(rd),
6852 _collector(collector),
6853 _span(span),
6854 _bit_map(bit_map),
6855 _mod_union_table(mod_union_table),
6856 _mark_stack(mark_stack),
6857 _concurrent_precleaning(concurrent_precleaning)
6858 {
6859 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL");
6860 }
6861
6862 // Grey object rescan during pre-cleaning and second checkpoint phases --
6863 // the non-parallel version (the parallel version appears further below.)
6864 void PushAndMarkClosure::do_oop(oop obj) {
6865 // Ignore mark word verification. If during concurrent precleaning,
6866 // the object monitor may be locked. If during the checkpoint
6867 // phases, the object may already have been reached by a different
6868 // path and may be at the end of the global overflow list (so
6869 // the mark word may be NULL).
6870 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6871 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6872 HeapWord* addr = (HeapWord*)obj;
6873 // Check if oop points into the CMS generation
6874 // and is not marked
6875 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6876 // a white object ...
6877 _bit_map->mark(addr); // ... now grey
6878 // push on the marking stack (grey set)
6879 bool simulate_overflow = false;
6880 NOT_PRODUCT(
6881 if (CMSMarkStackOverflowALot &&
6882 _collector->simulate_overflow()) {
6883 // simulate a stack overflow
6884 simulate_overflow = true;
6885 }
6886 )
6887 if (simulate_overflow || !_mark_stack->push(obj)) {
6888 if (_concurrent_precleaning) {
6889 // During precleaning we can just dirty the appropriate card(s)
6890 // in the mod union table, thus ensuring that the object remains
6891 // in the grey set and continue. In the case of object arrays
6892 // we need to dirty all of the cards that the object spans,
6893 // since the rescan of object arrays will be limited to the
6894 // dirty cards.
6895 // Note that no one can be interfering with us in this action
6896 // of dirtying the mod union table, so no locking or atomics
6897 // are required.
6898 if (obj->is_objArray()) {
6899 size_t sz = obj->size();
6900 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6901 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6902 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6903 _mod_union_table->mark_range(redirty_range);
6904 } else {
6905 _mod_union_table->mark(addr);
6906 }
6907 _collector->_ser_pmc_preclean_ovflw++;
6908 } else {
6909 // During the remark phase, we need to remember this oop
6910 // in the overflow list.
6911 _collector->push_on_overflow_list(obj);
6912 _collector->_ser_pmc_remark_ovflw++;
6913 }
6914 }
6915 }
6916 }
6917
6918 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6919 MemRegion span,
6920 ReferenceDiscoverer* rd,
6921 CMSBitMap* bit_map,
6922 OopTaskQueue* work_queue):
6923 MetadataAwareOopClosure(rd),
6924 _collector(collector),
6925 _span(span),
6926 _bit_map(bit_map),
6927 _work_queue(work_queue)
6928 {
6929 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL");
6930 }
6931
6932 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
6933 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6934
6935 // Grey object rescan during second checkpoint phase --
6936 // the parallel version.
6937 void ParPushAndMarkClosure::do_oop(oop obj) {
6938 // In the assert below, we ignore the mark word because
6939 // this oop may point to an already visited object that is
6940 // on the overflow stack (in which case the mark word has
6941 // been hijacked for chaining into the overflow stack --
6942 // if this is the last object in the overflow stack then
6943 // its mark word will be NULL). Because this object may
6944 // have been subsequently popped off the global overflow
6945 // stack, and the mark word possibly restored to the prototypical
6946 // value, by the time we get to examined this failing assert in
6947 // the debugger, is_oop_or_null(false) may subsequently start
6948 // to hold.
6949 assert(oopDesc::is_oop_or_null(obj, true),
6950 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6951 HeapWord* addr = (HeapWord*)obj;
6952 // Check if oop points into the CMS generation
6953 // and is not marked
6954 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6955 // a white object ...
6956 // If we manage to "claim" the object, by being the
6957 // first thread to mark it, then we push it on our
6958 // marking stack
6959 if (_bit_map->par_mark(addr)) { // ... now grey
6960 // push on work queue (grey set)
6961 bool simulate_overflow = false;
6962 NOT_PRODUCT(
6963 if (CMSMarkStackOverflowALot &&
6964 _collector->par_simulate_overflow()) {
6965 // simulate a stack overflow
6966 simulate_overflow = true;
6967 }
6968 )
6969 if (simulate_overflow || !_work_queue->push(obj)) {
6970 _collector->par_push_on_overflow_list(obj);
6971 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
6972 }
6973 } // Else, some other thread got there first
6974 }
6975 }
6976
6977 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6978 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6979
6980 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6981 Mutex* bml = _collector->bitMapLock();
6982 assert_lock_strong(bml);
6983 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6984 "CMS thread should hold CMS token");
6985
6986 bml->unlock();
6987 ConcurrentMarkSweepThread::desynchronize(true);
6988
6989 _collector->stopTimer();
6990 _collector->incrementYields();
6991
6992 // See the comment in coordinator_yield()
6993 for (unsigned i = 0; i < CMSYieldSleepCount &&
6994 ConcurrentMarkSweepThread::should_yield() &&
6995 !CMSCollector::foregroundGCIsActive(); ++i) {
6996 os::sleep(Thread::current(), 1, false);
6997 }
6998
6999 ConcurrentMarkSweepThread::synchronize(true);
7000 bml->lock();
7001
7002 _collector->startTimer();
7003 }
7004
7005 bool CMSPrecleanRefsYieldClosure::should_return() {
7006 if (ConcurrentMarkSweepThread::should_yield()) {
7007 do_yield_work();
7008 }
7009 return _collector->foregroundGCIsActive();
7010 }
7011
7012 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7013 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0,
7014 "mr should be aligned to start at a card boundary");
7015 // We'd like to assert:
7016 // assert(mr.word_size()%CardTable::card_size_in_words == 0,
7017 // "mr should be a range of cards");
7018 // However, that would be too strong in one case -- the last
7019 // partition ends at _unallocated_block which, in general, can be
7020 // an arbitrary boundary, not necessarily card aligned.
7021 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words;
7022 _space->object_iterate_mem(mr, &_scan_cl);
7023 }
7024
7025 SweepClosure::SweepClosure(CMSCollector* collector,
7026 ConcurrentMarkSweepGeneration* g,
7027 CMSBitMap* bitMap, bool should_yield) :
7028 _collector(collector),
7029 _g(g),
7030 _sp(g->cmsSpace()),
7031 _limit(_sp->sweep_limit()),
7032 _freelistLock(_sp->freelistLock()),
7033 _bitMap(bitMap),
7034 _yield(should_yield),
7035 _inFreeRange(false), // No free range at beginning of sweep
7036 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7037 _lastFreeRangeCoalesced(false),
7038 _freeFinger(g->used_region().start())
7039 {
7040 NOT_PRODUCT(
7041 _numObjectsFreed = 0;
7042 _numWordsFreed = 0;
7043 _numObjectsLive = 0;
7044 _numWordsLive = 0;
7045 _numObjectsAlreadyFree = 0;
7046 _numWordsAlreadyFree = 0;
7047 _last_fc = NULL;
7048
7049 _sp->initializeIndexedFreeListArrayReturnedBytes();
7050 _sp->dictionary()->initialize_dict_returned_bytes();
7051 )
7052 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7053 "sweep _limit out of bounds");
7054 log_develop_trace(gc, sweep)("====================");
7055 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7056 }
7057
7058 void SweepClosure::print_on(outputStream* st) const {
7059 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7060 p2i(_sp->bottom()), p2i(_sp->end()));
7061 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7062 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7063 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7064 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7065 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7066 }
7067
7068 #ifndef PRODUCT
7069 // Assertion checking only: no useful work in product mode --
7070 // however, if any of the flags below become product flags,
7071 // you may need to review this code to see if it needs to be
7072 // enabled in product mode.
7073 SweepClosure::~SweepClosure() {
7074 assert_lock_strong(_freelistLock);
7075 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7076 "sweep _limit out of bounds");
7077 if (inFreeRange()) {
7078 Log(gc, sweep) log;
7079 log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7080 ResourceMark rm;
7081 LogStream ls(log.error());
7082 print_on(&ls);
7083 ShouldNotReachHere();
7084 }
7085
7086 if (log_is_enabled(Debug, gc, sweep)) {
7087 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7088 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7089 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7090 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7091 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7092 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7093 }
7094
7095 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7096 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7097 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7098 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7099 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes",
7100 returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7101 }
7102 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7103 log_develop_trace(gc, sweep)("================");
7104 }
7105 #endif // PRODUCT
7106
7107 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7108 bool freeRangeInFreeLists) {
7109 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7110 p2i(freeFinger), freeRangeInFreeLists);
7111 assert(!inFreeRange(), "Trampling existing free range");
7112 set_inFreeRange(true);
7113 set_lastFreeRangeCoalesced(false);
7114
7115 set_freeFinger(freeFinger);
7116 set_freeRangeInFreeLists(freeRangeInFreeLists);
7117 if (CMSTestInFreeList) {
7118 if (freeRangeInFreeLists) {
7119 FreeChunk* fc = (FreeChunk*) freeFinger;
7120 assert(fc->is_free(), "A chunk on the free list should be free.");
7121 assert(fc->size() > 0, "Free range should have a size");
7122 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7123 }
7124 }
7125 }
7126
7127 // Note that the sweeper runs concurrently with mutators. Thus,
7128 // it is possible for direct allocation in this generation to happen
7129 // in the middle of the sweep. Note that the sweeper also coalesces
7130 // contiguous free blocks. Thus, unless the sweeper and the allocator
7131 // synchronize appropriately freshly allocated blocks may get swept up.
7132 // This is accomplished by the sweeper locking the free lists while
7133 // it is sweeping. Thus blocks that are determined to be free are
7134 // indeed free. There is however one additional complication:
7135 // blocks that have been allocated since the final checkpoint and
7136 // mark, will not have been marked and so would be treated as
7137 // unreachable and swept up. To prevent this, the allocator marks
7138 // the bit map when allocating during the sweep phase. This leads,
7139 // however, to a further complication -- objects may have been allocated
7140 // but not yet initialized -- in the sense that the header isn't yet
7141 // installed. The sweeper can not then determine the size of the block
7142 // in order to skip over it. To deal with this case, we use a technique
7143 // (due to Printezis) to encode such uninitialized block sizes in the
7144 // bit map. Since the bit map uses a bit per every HeapWord, but the
7145 // CMS generation has a minimum object size of 3 HeapWords, it follows
7146 // that "normal marks" won't be adjacent in the bit map (there will
7147 // always be at least two 0 bits between successive 1 bits). We make use
7148 // of these "unused" bits to represent uninitialized blocks -- the bit
7149 // corresponding to the start of the uninitialized object and the next
7150 // bit are both set. Finally, a 1 bit marks the end of the object that
7151 // started with the two consecutive 1 bits to indicate its potentially
7152 // uninitialized state.
7153
7154 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7155 FreeChunk* fc = (FreeChunk*)addr;
7156 size_t res;
7157
7158 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7159 // than "addr == _limit" because although _limit was a block boundary when
7160 // we started the sweep, it may no longer be one because heap expansion
7161 // may have caused us to coalesce the block ending at the address _limit
7162 // with a newly expanded chunk (this happens when _limit was set to the
7163 // previous _end of the space), so we may have stepped past _limit:
7164 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7165 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7166 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7167 "sweep _limit out of bounds");
7168 assert(addr < _sp->end(), "addr out of bounds");
7169 // Flush any free range we might be holding as a single
7170 // coalesced chunk to the appropriate free list.
7171 if (inFreeRange()) {
7172 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7173 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7174 flush_cur_free_chunk(freeFinger(),
7175 pointer_delta(addr, freeFinger()));
7176 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7177 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7178 lastFreeRangeCoalesced() ? 1 : 0);
7179 }
7180
7181 // help the iterator loop finish
7182 return pointer_delta(_sp->end(), addr);
7183 }
7184
7185 assert(addr < _limit, "sweep invariant");
7186 // check if we should yield
7187 do_yield_check(addr);
7188 if (fc->is_free()) {
7189 // Chunk that is already free
7190 res = fc->size();
7191 do_already_free_chunk(fc);
7192 debug_only(_sp->verifyFreeLists());
7193 // If we flush the chunk at hand in lookahead_and_flush()
7194 // and it's coalesced with a preceding chunk, then the
7195 // process of "mangling" the payload of the coalesced block
7196 // will cause erasure of the size information from the
7197 // (erstwhile) header of all the coalesced blocks but the
7198 // first, so the first disjunct in the assert will not hold
7199 // in that specific case (in which case the second disjunct
7200 // will hold).
7201 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7202 "Otherwise the size info doesn't change at this step");
7203 NOT_PRODUCT(
7204 _numObjectsAlreadyFree++;
7205 _numWordsAlreadyFree += res;
7206 )
7207 NOT_PRODUCT(_last_fc = fc;)
7208 } else if (!_bitMap->isMarked(addr)) {
7209 // Chunk is fresh garbage
7210 res = do_garbage_chunk(fc);
7211 debug_only(_sp->verifyFreeLists());
7212 NOT_PRODUCT(
7213 _numObjectsFreed++;
7214 _numWordsFreed += res;
7215 )
7216 } else {
7217 // Chunk that is alive.
7218 res = do_live_chunk(fc);
7219 debug_only(_sp->verifyFreeLists());
7220 NOT_PRODUCT(
7221 _numObjectsLive++;
7222 _numWordsLive += res;
7223 )
7224 }
7225 return res;
7226 }
7227
7228 // For the smart allocation, record following
7229 // split deaths - a free chunk is removed from its free list because
7230 // it is being split into two or more chunks.
7231 // split birth - a free chunk is being added to its free list because
7232 // a larger free chunk has been split and resulted in this free chunk.
7233 // coal death - a free chunk is being removed from its free list because
7234 // it is being coalesced into a large free chunk.
7235 // coal birth - a free chunk is being added to its free list because
7236 // it was created when two or more free chunks where coalesced into
7237 // this free chunk.
7238 //
7239 // These statistics are used to determine the desired number of free
7240 // chunks of a given size. The desired number is chosen to be relative
7241 // to the end of a CMS sweep. The desired number at the end of a sweep
7242 // is the
7243 // count-at-end-of-previous-sweep (an amount that was enough)
7244 // - count-at-beginning-of-current-sweep (the excess)
7245 // + split-births (gains in this size during interval)
7246 // - split-deaths (demands on this size during interval)
7247 // where the interval is from the end of one sweep to the end of the
7248 // next.
7249 //
7250 // When sweeping the sweeper maintains an accumulated chunk which is
7251 // the chunk that is made up of chunks that have been coalesced. That
7252 // will be termed the left-hand chunk. A new chunk of garbage that
7253 // is being considered for coalescing will be referred to as the
7254 // right-hand chunk.
7255 //
7256 // When making a decision on whether to coalesce a right-hand chunk with
7257 // the current left-hand chunk, the current count vs. the desired count
7258 // of the left-hand chunk is considered. Also if the right-hand chunk
7259 // is near the large chunk at the end of the heap (see
7260 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7261 // left-hand chunk is coalesced.
7262 //
7263 // When making a decision about whether to split a chunk, the desired count
7264 // vs. the current count of the candidate to be split is also considered.
7265 // If the candidate is underpopulated (currently fewer chunks than desired)
7266 // a chunk of an overpopulated (currently more chunks than desired) size may
7267 // be chosen. The "hint" associated with a free list, if non-null, points
7268 // to a free list which may be overpopulated.
7269 //
7270
7271 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7272 const size_t size = fc->size();
7273 // Chunks that cannot be coalesced are not in the
7274 // free lists.
7275 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7276 assert(_sp->verify_chunk_in_free_list(fc),
7277 "free chunk should be in free lists");
7278 }
7279 // a chunk that is already free, should not have been
7280 // marked in the bit map
7281 HeapWord* const addr = (HeapWord*) fc;
7282 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7283 // Verify that the bit map has no bits marked between
7284 // addr and purported end of this block.
7285 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7286
7287 // Some chunks cannot be coalesced under any circumstances.
7288 // See the definition of cantCoalesce().
7289 if (!fc->cantCoalesce()) {
7290 // This chunk can potentially be coalesced.
7291 // All the work is done in
7292 do_post_free_or_garbage_chunk(fc, size);
7293 // Note that if the chunk is not coalescable (the else arm
7294 // below), we unconditionally flush, without needing to do
7295 // a "lookahead," as we do below.
7296 if (inFreeRange()) lookahead_and_flush(fc, size);
7297 } else {
7298 // Code path common to both original and adaptive free lists.
7299
7300 // cant coalesce with previous block; this should be treated
7301 // as the end of a free run if any
7302 if (inFreeRange()) {
7303 // we kicked some butt; time to pick up the garbage
7304 assert(freeFinger() < addr, "freeFinger points too high");
7305 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7306 }
7307 // else, nothing to do, just continue
7308 }
7309 }
7310
7311 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7312 // This is a chunk of garbage. It is not in any free list.
7313 // Add it to a free list or let it possibly be coalesced into
7314 // a larger chunk.
7315 HeapWord* const addr = (HeapWord*) fc;
7316 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7317
7318 // Verify that the bit map has no bits marked between
7319 // addr and purported end of just dead object.
7320 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7321 do_post_free_or_garbage_chunk(fc, size);
7322
7323 assert(_limit >= addr + size,
7324 "A freshly garbage chunk can't possibly straddle over _limit");
7325 if (inFreeRange()) lookahead_and_flush(fc, size);
7326 return size;
7327 }
7328
7329 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7330 HeapWord* addr = (HeapWord*) fc;
7331 // The sweeper has just found a live object. Return any accumulated
7332 // left hand chunk to the free lists.
7333 if (inFreeRange()) {
7334 assert(freeFinger() < addr, "freeFinger points too high");
7335 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7336 }
7337
7338 // This object is live: we'd normally expect this to be
7339 // an oop, and like to assert the following:
7340 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7341 // However, as we commented above, this may be an object whose
7342 // header hasn't yet been initialized.
7343 size_t size;
7344 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7345 if (_bitMap->isMarked(addr + 1)) {
7346 // Determine the size from the bit map, rather than trying to
7347 // compute it from the object header.
7348 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7349 size = pointer_delta(nextOneAddr + 1, addr);
7350 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7351 "alignment problem");
7352
7353 #ifdef ASSERT
7354 if (oop(addr)->klass_or_null_acquire() != NULL) {
7355 // Ignore mark word because we are running concurrent with mutators
7356 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7357 assert(size ==
7358 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7359 "P-mark and computed size do not agree");
7360 }
7361 #endif
7362
7363 } else {
7364 // This should be an initialized object that's alive.
7365 assert(oop(addr)->klass_or_null_acquire() != NULL,
7366 "Should be an initialized object");
7367 // Ignore mark word because we are running concurrent with mutators
7368 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7369 // Verify that the bit map has no bits marked between
7370 // addr and purported end of this block.
7371 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7372 assert(size >= 3, "Necessary for Printezis marks to work");
7373 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7374 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7375 }
7376 return size;
7377 }
7378
7379 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7380 size_t chunkSize) {
7381 // do_post_free_or_garbage_chunk() should only be called in the case
7382 // of the adaptive free list allocator.
7383 const bool fcInFreeLists = fc->is_free();
7384 assert((HeapWord*)fc <= _limit, "sweep invariant");
7385 if (CMSTestInFreeList && fcInFreeLists) {
7386 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7387 }
7388
7389 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7390
7391 HeapWord* const fc_addr = (HeapWord*) fc;
7392
7393 bool coalesce = false;
7394 const size_t left = pointer_delta(fc_addr, freeFinger());
7395 const size_t right = chunkSize;
7396 switch (FLSCoalescePolicy) {
7397 // numeric value forms a coalition aggressiveness metric
7398 case 0: { // never coalesce
7399 coalesce = false;
7400 break;
7401 }
7402 case 1: { // coalesce if left & right chunks on overpopulated lists
7403 coalesce = _sp->coalOverPopulated(left) &&
7404 _sp->coalOverPopulated(right);
7405 break;
7406 }
7407 case 2: { // coalesce if left chunk on overpopulated list (default)
7408 coalesce = _sp->coalOverPopulated(left);
7409 break;
7410 }
7411 case 3: { // coalesce if left OR right chunk on overpopulated list
7412 coalesce = _sp->coalOverPopulated(left) ||
7413 _sp->coalOverPopulated(right);
7414 break;
7415 }
7416 case 4: { // always coalesce
7417 coalesce = true;
7418 break;
7419 }
7420 default:
7421 ShouldNotReachHere();
7422 }
7423
7424 // Should the current free range be coalesced?
7425 // If the chunk is in a free range and either we decided to coalesce above
7426 // or the chunk is near the large block at the end of the heap
7427 // (isNearLargestChunk() returns true), then coalesce this chunk.
7428 const bool doCoalesce = inFreeRange()
7429 && (coalesce || _g->isNearLargestChunk(fc_addr));
7430 if (doCoalesce) {
7431 // Coalesce the current free range on the left with the new
7432 // chunk on the right. If either is on a free list,
7433 // it must be removed from the list and stashed in the closure.
7434 if (freeRangeInFreeLists()) {
7435 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7436 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7437 "Size of free range is inconsistent with chunk size.");
7438 if (CMSTestInFreeList) {
7439 assert(_sp->verify_chunk_in_free_list(ffc),
7440 "Chunk is not in free lists");
7441 }
7442 _sp->coalDeath(ffc->size());
7443 _sp->removeFreeChunkFromFreeLists(ffc);
7444 set_freeRangeInFreeLists(false);
7445 }
7446 if (fcInFreeLists) {
7447 _sp->coalDeath(chunkSize);
7448 assert(fc->size() == chunkSize,
7449 "The chunk has the wrong size or is not in the free lists");
7450 _sp->removeFreeChunkFromFreeLists(fc);
7451 }
7452 set_lastFreeRangeCoalesced(true);
7453 print_free_block_coalesced(fc);
7454 } else { // not in a free range and/or should not coalesce
7455 // Return the current free range and start a new one.
7456 if (inFreeRange()) {
7457 // In a free range but cannot coalesce with the right hand chunk.
7458 // Put the current free range into the free lists.
7459 flush_cur_free_chunk(freeFinger(),
7460 pointer_delta(fc_addr, freeFinger()));
7461 }
7462 // Set up for new free range. Pass along whether the right hand
7463 // chunk is in the free lists.
7464 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7465 }
7466 }
7467
7468 // Lookahead flush:
7469 // If we are tracking a free range, and this is the last chunk that
7470 // we'll look at because its end crosses past _limit, we'll preemptively
7471 // flush it along with any free range we may be holding on to. Note that
7472 // this can be the case only for an already free or freshly garbage
7473 // chunk. If this block is an object, it can never straddle
7474 // over _limit. The "straddling" occurs when _limit is set at
7475 // the previous end of the space when this cycle started, and
7476 // a subsequent heap expansion caused the previously co-terminal
7477 // free block to be coalesced with the newly expanded portion,
7478 // thus rendering _limit a non-block-boundary making it dangerous
7479 // for the sweeper to step over and examine.
7480 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7481 assert(inFreeRange(), "Should only be called if currently in a free range.");
7482 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7483 assert(_sp->used_region().contains(eob - 1),
7484 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7485 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7486 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7487 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7488 if (eob >= _limit) {
7489 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7490 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7491 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7492 "[" PTR_FORMAT "," PTR_FORMAT ")",
7493 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7494 // Return the storage we are tracking back into the free lists.
7495 log_develop_trace(gc, sweep)("Flushing ... ");
7496 assert(freeFinger() < eob, "Error");
7497 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7498 }
7499 }
7500
7501 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7502 assert(inFreeRange(), "Should only be called if currently in a free range.");
7503 assert(size > 0,
7504 "A zero sized chunk cannot be added to the free lists.");
7505 if (!freeRangeInFreeLists()) {
7506 if (CMSTestInFreeList) {
7507 FreeChunk* fc = (FreeChunk*) chunk;
7508 fc->set_size(size);
7509 assert(!_sp->verify_chunk_in_free_list(fc),
7510 "chunk should not be in free lists yet");
7511 }
7512 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7513 // A new free range is going to be starting. The current
7514 // free range has not been added to the free lists yet or
7515 // was removed so add it back.
7516 // If the current free range was coalesced, then the death
7517 // of the free range was recorded. Record a birth now.
7518 if (lastFreeRangeCoalesced()) {
7519 _sp->coalBirth(size);
7520 }
7521 _sp->addChunkAndRepairOffsetTable(chunk, size,
7522 lastFreeRangeCoalesced());
7523 } else {
7524 log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7525 }
7526 set_inFreeRange(false);
7527 set_freeRangeInFreeLists(false);
7528 }
7529
7530 // We take a break if we've been at this for a while,
7531 // so as to avoid monopolizing the locks involved.
7532 void SweepClosure::do_yield_work(HeapWord* addr) {
7533 // Return current free chunk being used for coalescing (if any)
7534 // to the appropriate freelist. After yielding, the next
7535 // free block encountered will start a coalescing range of
7536 // free blocks. If the next free block is adjacent to the
7537 // chunk just flushed, they will need to wait for the next
7538 // sweep to be coalesced.
7539 if (inFreeRange()) {
7540 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7541 }
7542
7543 // First give up the locks, then yield, then re-lock.
7544 // We should probably use a constructor/destructor idiom to
7545 // do this unlock/lock or modify the MutexUnlocker class to
7546 // serve our purpose. XXX
7547 assert_lock_strong(_bitMap->lock());
7548 assert_lock_strong(_freelistLock);
7549 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7550 "CMS thread should hold CMS token");
7551 _bitMap->lock()->unlock();
7552 _freelistLock->unlock();
7553 ConcurrentMarkSweepThread::desynchronize(true);
7554 _collector->stopTimer();
7555 _collector->incrementYields();
7556
7557 // See the comment in coordinator_yield()
7558 for (unsigned i = 0; i < CMSYieldSleepCount &&
7559 ConcurrentMarkSweepThread::should_yield() &&
7560 !CMSCollector::foregroundGCIsActive(); ++i) {
7561 os::sleep(Thread::current(), 1, false);
7562 }
7563
7564 ConcurrentMarkSweepThread::synchronize(true);
7565 _freelistLock->lock();
7566 _bitMap->lock()->lock_without_safepoint_check();
7567 _collector->startTimer();
7568 }
7569
7570 #ifndef PRODUCT
7571 // This is actually very useful in a product build if it can
7572 // be called from the debugger. Compile it into the product
7573 // as needed.
7574 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7575 return debug_cms_space->verify_chunk_in_free_list(fc);
7576 }
7577 #endif
7578
7579 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7580 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7581 p2i(fc), fc->size());
7582 }
7583
7584 // CMSIsAliveClosure
7585 bool CMSIsAliveClosure::do_object_b(oop obj) {
7586 HeapWord* addr = (HeapWord*)obj;
7587 return addr != NULL &&
7588 (!_span.contains(addr) || _bit_map->isMarked(addr));
7589 }
7590
7591
7592 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7593 MemRegion span,
7594 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7595 bool cpc):
7596 _collector(collector),
7597 _span(span),
7598 _bit_map(bit_map),
7599 _mark_stack(mark_stack),
7600 _concurrent_precleaning(cpc) {
7601 assert(!_span.is_empty(), "Empty span could spell trouble");
7602 }
7603
7604
7605 // CMSKeepAliveClosure: the serial version
7606 void CMSKeepAliveClosure::do_oop(oop obj) {
7607 HeapWord* addr = (HeapWord*)obj;
7608 if (_span.contains(addr) &&
7609 !_bit_map->isMarked(addr)) {
7610 _bit_map->mark(addr);
7611 bool simulate_overflow = false;
7612 NOT_PRODUCT(
7613 if (CMSMarkStackOverflowALot &&
7614 _collector->simulate_overflow()) {
7615 // simulate a stack overflow
7616 simulate_overflow = true;
7617 }
7618 )
7619 if (simulate_overflow || !_mark_stack->push(obj)) {
7620 if (_concurrent_precleaning) {
7621 // We dirty the overflown object and let the remark
7622 // phase deal with it.
7623 assert(_collector->overflow_list_is_empty(), "Error");
7624 // In the case of object arrays, we need to dirty all of
7625 // the cards that the object spans. No locking or atomics
7626 // are needed since no one else can be mutating the mod union
7627 // table.
7628 if (obj->is_objArray()) {
7629 size_t sz = obj->size();
7630 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
7631 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7632 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7633 _collector->_modUnionTable.mark_range(redirty_range);
7634 } else {
7635 _collector->_modUnionTable.mark(addr);
7636 }
7637 _collector->_ser_kac_preclean_ovflw++;
7638 } else {
7639 _collector->push_on_overflow_list(obj);
7640 _collector->_ser_kac_ovflw++;
7641 }
7642 }
7643 }
7644 }
7645
7646 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7647 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7648
7649 // CMSParKeepAliveClosure: a parallel version of the above.
7650 // The work queues are private to each closure (thread),
7651 // but (may be) available for stealing by other threads.
7652 void CMSParKeepAliveClosure::do_oop(oop obj) {
7653 HeapWord* addr = (HeapWord*)obj;
7654 if (_span.contains(addr) &&
7655 !_bit_map->isMarked(addr)) {
7656 // In general, during recursive tracing, several threads
7657 // may be concurrently getting here; the first one to
7658 // "tag" it, claims it.
7659 if (_bit_map->par_mark(addr)) {
7660 bool res = _work_queue->push(obj);
7661 assert(res, "Low water mark should be much less than capacity");
7662 // Do a recursive trim in the hope that this will keep
7663 // stack usage lower, but leave some oops for potential stealers
7664 trim_queue(_low_water_mark);
7665 } // Else, another thread got there first
7666 }
7667 }
7668
7669 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7670 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7671
7672 void CMSParKeepAliveClosure::trim_queue(uint max) {
7673 while (_work_queue->size() > max) {
7674 oop new_oop;
7675 if (_work_queue->pop_local(new_oop)) {
7676 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7677 assert(_bit_map->isMarked((HeapWord*)new_oop),
7678 "no white objects on this stack!");
7679 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7680 // iterate over the oops in this oop, marking and pushing
7681 // the ones in CMS heap (i.e. in _span).
7682 new_oop->oop_iterate(&_mark_and_push);
7683 }
7684 }
7685 }
7686
7687 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7688 CMSCollector* collector,
7689 MemRegion span, CMSBitMap* bit_map,
7690 OopTaskQueue* work_queue):
7691 _collector(collector),
7692 _span(span),
7693 _bit_map(bit_map),
7694 _work_queue(work_queue) { }
7695
7696 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7697 HeapWord* addr = (HeapWord*)obj;
7698 if (_span.contains(addr) &&
7699 !_bit_map->isMarked(addr)) {
7700 if (_bit_map->par_mark(addr)) {
7701 bool simulate_overflow = false;
7702 NOT_PRODUCT(
7703 if (CMSMarkStackOverflowALot &&
7704 _collector->par_simulate_overflow()) {
7705 // simulate a stack overflow
7706 simulate_overflow = true;
7707 }
7708 )
7709 if (simulate_overflow || !_work_queue->push(obj)) {
7710 _collector->par_push_on_overflow_list(obj);
7711 _collector->_par_kac_ovflw++;
7712 }
7713 } // Else another thread got there already
7714 }
7715 }
7716
7717 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7718 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7719
7720 //////////////////////////////////////////////////////////////////
7721 // CMSExpansionCause /////////////////////////////
7722 //////////////////////////////////////////////////////////////////
7723 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7724 switch (cause) {
7725 case _no_expansion:
7726 return "No expansion";
7727 case _satisfy_free_ratio:
7728 return "Free ratio";
7729 case _satisfy_promotion:
7730 return "Satisfy promotion";
7731 case _satisfy_allocation:
7732 return "allocation";
7733 case _allocate_par_lab:
7734 return "Par LAB";
7735 case _allocate_par_spooling_space:
7736 return "Par Spooling Space";
7737 case _adaptive_size_policy:
7738 return "Ergonomics";
7739 default:
7740 return "unknown";
7741 }
7742 }
7743
7744 void CMSDrainMarkingStackClosure::do_void() {
7745 // the max number to take from overflow list at a time
7746 const size_t num = _mark_stack->capacity()/4;
7747 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7748 "Overflow list should be NULL during concurrent phases");
7749 while (!_mark_stack->isEmpty() ||
7750 // if stack is empty, check the overflow list
7751 _collector->take_from_overflow_list(num, _mark_stack)) {
7752 oop obj = _mark_stack->pop();
7753 HeapWord* addr = (HeapWord*)obj;
7754 assert(_span.contains(addr), "Should be within span");
7755 assert(_bit_map->isMarked(addr), "Should be marked");
7756 assert(oopDesc::is_oop(obj), "Should be an oop");
7757 obj->oop_iterate(_keep_alive);
7758 }
7759 }
7760
7761 void CMSParDrainMarkingStackClosure::do_void() {
7762 // drain queue
7763 trim_queue(0);
7764 }
7765
7766 // Trim our work_queue so its length is below max at return
7767 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7768 while (_work_queue->size() > max) {
7769 oop new_oop;
7770 if (_work_queue->pop_local(new_oop)) {
7771 assert(oopDesc::is_oop(new_oop), "Expected an oop");
7772 assert(_bit_map->isMarked((HeapWord*)new_oop),
7773 "no white objects on this stack!");
7774 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7775 // iterate over the oops in this oop, marking and pushing
7776 // the ones in CMS heap (i.e. in _span).
7777 new_oop->oop_iterate(&_mark_and_push);
7778 }
7779 }
7780 }
7781
7782 ////////////////////////////////////////////////////////////////////
7783 // Support for Marking Stack Overflow list handling and related code
7784 ////////////////////////////////////////////////////////////////////
7785 // Much of the following code is similar in shape and spirit to the
7786 // code used in ParNewGC. We should try and share that code
7787 // as much as possible in the future.
7788
7789 #ifndef PRODUCT
7790 // Debugging support for CMSStackOverflowALot
7791
7792 // It's OK to call this multi-threaded; the worst thing
7793 // that can happen is that we'll get a bunch of closely
7794 // spaced simulated overflows, but that's OK, in fact
7795 // probably good as it would exercise the overflow code
7796 // under contention.
7797 bool CMSCollector::simulate_overflow() {
7798 if (_overflow_counter-- <= 0) { // just being defensive
7799 _overflow_counter = CMSMarkStackOverflowInterval;
7800 return true;
7801 } else {
7802 return false;
7803 }
7804 }
7805
7806 bool CMSCollector::par_simulate_overflow() {
7807 return simulate_overflow();
7808 }
7809 #endif
7810
7811 // Single-threaded
7812 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7813 assert(stack->isEmpty(), "Expected precondition");
7814 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7815 size_t i = num;
7816 oop cur = _overflow_list;
7817 const markOop proto = markOopDesc::prototype();
7818 NOT_PRODUCT(ssize_t n = 0;)
7819 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7820 next = oop(cur->mark_raw());
7821 cur->set_mark_raw(proto); // until proven otherwise
7822 assert(oopDesc::is_oop(cur), "Should be an oop");
7823 bool res = stack->push(cur);
7824 assert(res, "Bit off more than can chew?");
7825 NOT_PRODUCT(n++;)
7826 }
7827 _overflow_list = cur;
7828 #ifndef PRODUCT
7829 assert(_num_par_pushes >= n, "Too many pops?");
7830 _num_par_pushes -=n;
7831 #endif
7832 return !stack->isEmpty();
7833 }
7834
7835 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
7836 // (MT-safe) Get a prefix of at most "num" from the list.
7837 // The overflow list is chained through the mark word of
7838 // each object in the list. We fetch the entire list,
7839 // break off a prefix of the right size and return the
7840 // remainder. If other threads try to take objects from
7841 // the overflow list at that time, they will wait for
7842 // some time to see if data becomes available. If (and
7843 // only if) another thread places one or more object(s)
7844 // on the global list before we have returned the suffix
7845 // to the global list, we will walk down our local list
7846 // to find its end and append the global list to
7847 // our suffix before returning it. This suffix walk can
7848 // prove to be expensive (quadratic in the amount of traffic)
7849 // when there are many objects in the overflow list and
7850 // there is much producer-consumer contention on the list.
7851 // *NOTE*: The overflow list manipulation code here and
7852 // in ParNewGeneration:: are very similar in shape,
7853 // except that in the ParNew case we use the old (from/eden)
7854 // copy of the object to thread the list via its klass word.
7855 // Because of the common code, if you make any changes in
7856 // the code below, please check the ParNew version to see if
7857 // similar changes might be needed.
7858 // CR 6797058 has been filed to consolidate the common code.
7859 bool CMSCollector::par_take_from_overflow_list(size_t num,
7860 OopTaskQueue* work_q,
7861 int no_of_gc_threads) {
7862 assert(work_q->size() == 0, "First empty local work queue");
7863 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7864 if (_overflow_list == NULL) {
7865 return false;
7866 }
7867 // Grab the entire list; we'll put back a suffix
7868 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7869 Thread* tid = Thread::current();
7870 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7871 // set to ParallelGCThreads.
7872 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7873 size_t sleep_time_millis = MAX2((size_t)1, num/100);
7874 // If the list is busy, we spin for a short while,
7875 // sleeping between attempts to get the list.
7876 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7877 os::sleep(tid, sleep_time_millis, false);
7878 if (_overflow_list == NULL) {
7879 // Nothing left to take
7880 return false;
7881 } else if (_overflow_list != BUSY) {
7882 // Try and grab the prefix
7883 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7884 }
7885 }
7886 // If the list was found to be empty, or we spun long
7887 // enough, we give up and return empty-handed. If we leave
7888 // the list in the BUSY state below, it must be the case that
7889 // some other thread holds the overflow list and will set it
7890 // to a non-BUSY state in the future.
7891 if (prefix == NULL || prefix == BUSY) {
7892 // Nothing to take or waited long enough
7893 if (prefix == NULL) {
7894 // Write back the NULL in case we overwrote it with BUSY above
7895 // and it is still the same value.
7896 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7897 }
7898 return false;
7899 }
7900 assert(prefix != NULL && prefix != BUSY, "Error");
7901 size_t i = num;
7902 oop cur = prefix;
7903 // Walk down the first "num" objects, unless we reach the end.
7904 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--);
7905 if (cur->mark_raw() == NULL) {
7906 // We have "num" or fewer elements in the list, so there
7907 // is nothing to return to the global list.
7908 // Write back the NULL in lieu of the BUSY we wrote
7909 // above, if it is still the same value.
7910 if (_overflow_list == BUSY) {
7911 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7912 }
7913 } else {
7914 // Chop off the suffix and return it to the global list.
7915 assert(cur->mark_raw() != BUSY, "Error");
7916 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list
7917 cur->set_mark_raw(NULL); // break off suffix
7918 // It's possible that the list is still in the empty(busy) state
7919 // we left it in a short while ago; in that case we may be
7920 // able to place back the suffix without incurring the cost
7921 // of a walk down the list.
7922 oop observed_overflow_list = _overflow_list;
7923 oop cur_overflow_list = observed_overflow_list;
7924 bool attached = false;
7925 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7926 observed_overflow_list =
7927 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7928 if (cur_overflow_list == observed_overflow_list) {
7929 attached = true;
7930 break;
7931 } else cur_overflow_list = observed_overflow_list;
7932 }
7933 if (!attached) {
7934 // Too bad, someone else sneaked in (at least) an element; we'll need
7935 // to do a splice. Find tail of suffix so we can prepend suffix to global
7936 // list.
7937 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw()));
7938 oop suffix_tail = cur;
7939 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL,
7940 "Tautology");
7941 observed_overflow_list = _overflow_list;
7942 do {
7943 cur_overflow_list = observed_overflow_list;
7944 if (cur_overflow_list != BUSY) {
7945 // Do the splice ...
7946 suffix_tail->set_mark_raw(markOop(cur_overflow_list));
7947 } else { // cur_overflow_list == BUSY
7948 suffix_tail->set_mark_raw(NULL);
7949 }
7950 // ... and try to place spliced list back on overflow_list ...
7951 observed_overflow_list =
7952 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7953 } while (cur_overflow_list != observed_overflow_list);
7954 // ... until we have succeeded in doing so.
7955 }
7956 }
7957
7958 // Push the prefix elements on work_q
7959 assert(prefix != NULL, "control point invariant");
7960 const markOop proto = markOopDesc::prototype();
7961 oop next;
7962 NOT_PRODUCT(ssize_t n = 0;)
7963 for (cur = prefix; cur != NULL; cur = next) {
7964 next = oop(cur->mark_raw());
7965 cur->set_mark_raw(proto); // until proven otherwise
7966 assert(oopDesc::is_oop(cur), "Should be an oop");
7967 bool res = work_q->push(cur);
7968 assert(res, "Bit off more than we can chew?");
7969 NOT_PRODUCT(n++;)
7970 }
7971 #ifndef PRODUCT
7972 assert(_num_par_pushes >= n, "Too many pops?");
7973 Atomic::sub(n, &_num_par_pushes);
7974 #endif
7975 return true;
7976 }
7977
7978 // Single-threaded
7979 void CMSCollector::push_on_overflow_list(oop p) {
7980 NOT_PRODUCT(_num_par_pushes++;)
7981 assert(oopDesc::is_oop(p), "Not an oop");
7982 preserve_mark_if_necessary(p);
7983 p->set_mark_raw((markOop)_overflow_list);
7984 _overflow_list = p;
7985 }
7986
7987 // Multi-threaded; use CAS to prepend to overflow list
7988 void CMSCollector::par_push_on_overflow_list(oop p) {
7989 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);)
7990 assert(oopDesc::is_oop(p), "Not an oop");
7991 par_preserve_mark_if_necessary(p);
7992 oop observed_overflow_list = _overflow_list;
7993 oop cur_overflow_list;
7994 do {
7995 cur_overflow_list = observed_overflow_list;
7996 if (cur_overflow_list != BUSY) {
7997 p->set_mark_raw(markOop(cur_overflow_list));
7998 } else {
7999 p->set_mark_raw(NULL);
8000 }
8001 observed_overflow_list =
8002 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list);
8003 } while (cur_overflow_list != observed_overflow_list);
8004 }
8005 #undef BUSY
8006
8007 // Single threaded
8008 // General Note on GrowableArray: pushes may silently fail
8009 // because we are (temporarily) out of C-heap for expanding
8010 // the stack. The problem is quite ubiquitous and affects
8011 // a lot of code in the JVM. The prudent thing for GrowableArray
8012 // to do (for now) is to exit with an error. However, that may
8013 // be too draconian in some cases because the caller may be
8014 // able to recover without much harm. For such cases, we
8015 // should probably introduce a "soft_push" method which returns
8016 // an indication of success or failure with the assumption that
8017 // the caller may be able to recover from a failure; code in
8018 // the VM can then be changed, incrementally, to deal with such
8019 // failures where possible, thus, incrementally hardening the VM
8020 // in such low resource situations.
8021 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8022 _preserved_oop_stack.push(p);
8023 _preserved_mark_stack.push(m);
8024 assert(m == p->mark_raw(), "Mark word changed");
8025 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8026 "bijection");
8027 }
8028
8029 // Single threaded
8030 void CMSCollector::preserve_mark_if_necessary(oop p) {
8031 markOop m = p->mark_raw();
8032 if (m->must_be_preserved(p)) {
8033 preserve_mark_work(p, m);
8034 }
8035 }
8036
8037 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8038 markOop m = p->mark_raw();
8039 if (m->must_be_preserved(p)) {
8040 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8041 // Even though we read the mark word without holding
8042 // the lock, we are assured that it will not change
8043 // because we "own" this oop, so no other thread can
8044 // be trying to push it on the overflow list; see
8045 // the assertion in preserve_mark_work() that checks
8046 // that m == p->mark_raw().
8047 preserve_mark_work(p, m);
8048 }
8049 }
8050
8051 // We should be able to do this multi-threaded,
8052 // a chunk of stack being a task (this is
8053 // correct because each oop only ever appears
8054 // once in the overflow list. However, it's
8055 // not very easy to completely overlap this with
8056 // other operations, so will generally not be done
8057 // until all work's been completed. Because we
8058 // expect the preserved oop stack (set) to be small,
8059 // it's probably fine to do this single-threaded.
8060 // We can explore cleverer concurrent/overlapped/parallel
8061 // processing of preserved marks if we feel the
8062 // need for this in the future. Stack overflow should
8063 // be so rare in practice and, when it happens, its
8064 // effect on performance so great that this will
8065 // likely just be in the noise anyway.
8066 void CMSCollector::restore_preserved_marks_if_any() {
8067 assert(SafepointSynchronize::is_at_safepoint(),
8068 "world should be stopped");
8069 assert(Thread::current()->is_ConcurrentGC_thread() ||
8070 Thread::current()->is_VM_thread(),
8071 "should be single-threaded");
8072 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8073 "bijection");
8074
8075 while (!_preserved_oop_stack.is_empty()) {
8076 oop p = _preserved_oop_stack.pop();
8077 assert(oopDesc::is_oop(p), "Should be an oop");
8078 assert(_span.contains(p), "oop should be in _span");
8079 assert(p->mark_raw() == markOopDesc::prototype(),
8080 "Set when taken from overflow list");
8081 markOop m = _preserved_mark_stack.pop();
8082 p->set_mark_raw(m);
8083 }
8084 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8085 "stacks were cleared above");
8086 }
8087
8088 #ifndef PRODUCT
8089 bool CMSCollector::no_preserved_marks() const {
8090 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8091 }
8092 #endif
8093
8094 // Transfer some number of overflown objects to usual marking
8095 // stack. Return true if some objects were transferred.
8096 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8097 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8098 (size_t)ParGCDesiredObjsFromOverflowList);
8099
8100 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8101 assert(_collector->overflow_list_is_empty() || res,
8102 "If list is not empty, we should have taken something");
8103 assert(!res || !_mark_stack->isEmpty(),
8104 "If we took something, it should now be on our stack");
8105 return res;
8106 }
8107
8108 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8109 size_t res = _sp->block_size_no_stall(addr, _collector);
8110 if (_sp->block_is_obj(addr)) {
8111 if (_live_bit_map->isMarked(addr)) {
8112 // It can't have been dead in a previous cycle
8113 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8114 } else {
8115 _dead_bit_map->mark(addr); // mark the dead object
8116 }
8117 }
8118 // Could be 0, if the block size could not be computed without stalling.
8119 return res;
8120 }
8121
8122 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8123 GCMemoryManager* manager = CMSHeap::heap()->old_manager();
8124 switch (phase) {
8125 case CMSCollector::InitialMarking:
8126 initialize(manager /* GC manager */ ,
8127 cause /* cause of the GC */,
8128 true /* recordGCBeginTime */,
8129 true /* recordPreGCUsage */,
8130 false /* recordPeakUsage */,
8131 false /* recordPostGCusage */,
8132 true /* recordAccumulatedGCTime */,
8133 false /* recordGCEndTime */,
8134 false /* countCollection */ );
8135 break;
8136
8137 case CMSCollector::FinalMarking:
8138 initialize(manager /* GC manager */ ,
8139 cause /* cause of the GC */,
8140 false /* recordGCBeginTime */,
8141 false /* recordPreGCUsage */,
8142 false /* recordPeakUsage */,
8143 false /* recordPostGCusage */,
8144 true /* recordAccumulatedGCTime */,
8145 false /* recordGCEndTime */,
8146 false /* countCollection */ );
8147 break;
8148
8149 case CMSCollector::Sweeping:
8150 initialize(manager /* GC manager */ ,
8151 cause /* cause of the GC */,
8152 false /* recordGCBeginTime */,
8153 false /* recordPreGCUsage */,
8154 true /* recordPeakUsage */,
8155 true /* recordPostGCusage */,
8156 false /* recordAccumulatedGCTime */,
8157 true /* recordGCEndTime */,
8158 true /* countCollection */ );
8159 break;
8160
8161 default:
8162 ShouldNotReachHere();
8163 }
8164 }