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