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