1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc/cms/cmsCollectorPolicy.hpp"
31 #include "gc/cms/cmsOopClosures.inline.hpp"
32 #include "gc/cms/compactibleFreeListSpace.hpp"
33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
34 #include "gc/cms/concurrentMarkSweepThread.hpp"
35 #include "gc/cms/parNewGeneration.hpp"
36 #include "gc/cms/vmCMSOperations.hpp"
37 #include "gc/serial/genMarkSweep.hpp"
38 #include "gc/serial/tenuredGeneration.hpp"
39 #include "gc/shared/adaptiveSizePolicy.hpp"
40 #include "gc/shared/cardGeneration.inline.hpp"
41 #include "gc/shared/cardTableRS.hpp"
42 #include "gc/shared/collectedHeap.inline.hpp"
43 #include "gc/shared/collectorCounters.hpp"
44 #include "gc/shared/collectorPolicy.hpp"
45 #include "gc/shared/gcLocker.inline.hpp"
46 #include "gc/shared/gcPolicyCounters.hpp"
47 #include "gc/shared/gcTimer.hpp"
48 #include "gc/shared/gcTrace.hpp"
49 #include "gc/shared/gcTraceTime.hpp"
50 #include "gc/shared/genCollectedHeap.hpp"
51 #include "gc/shared/genOopClosures.inline.hpp"
52 #include "gc/shared/isGCActiveMark.hpp"
53 #include "gc/shared/referencePolicy.hpp"
54 #include "gc/shared/strongRootsScope.hpp"
55 #include "gc/shared/taskqueue.inline.hpp"
56 #include "memory/allocation.hpp"
57 #include "memory/iterator.inline.hpp"
58 #include "memory/padded.hpp"
59 #include "memory/resourceArea.hpp"
60 #include "oops/oop.inline.hpp"
61 #include "prims/jvmtiExport.hpp"
62 #include "runtime/atomic.inline.hpp"
63 #include "runtime/globals_extension.hpp"
64 #include "runtime/handles.inline.hpp"
65 #include "runtime/java.hpp"
66 #include "runtime/orderAccess.inline.hpp"
67 #include "runtime/vmThread.hpp"
68 #include "services/memoryService.hpp"
69 #include "services/runtimeService.hpp"
70 #include "utilities/stack.inline.hpp"
71
72 // statics
73 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
74 bool CMSCollector::_full_gc_requested = false;
75 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
76
77 //////////////////////////////////////////////////////////////////
78 // In support of CMS/VM thread synchronization
79 //////////////////////////////////////////////////////////////////
80 // We split use of the CGC_lock into 2 "levels".
81 // The low-level locking is of the usual CGC_lock monitor. We introduce
82 // a higher level "token" (hereafter "CMS token") built on top of the
83 // low level monitor (hereafter "CGC lock").
84 // The token-passing protocol gives priority to the VM thread. The
85 // CMS-lock doesn't provide any fairness guarantees, but clients
86 // should ensure that it is only held for very short, bounded
87 // durations.
88 //
89 // When either of the CMS thread or the VM thread is involved in
90 // collection operations during which it does not want the other
91 // thread to interfere, it obtains the CMS token.
92 //
93 // If either thread tries to get the token while the other has
94 // it, that thread waits. However, if the VM thread and CMS thread
95 // both want the token, then the VM thread gets priority while the
96 // CMS thread waits. This ensures, for instance, that the "concurrent"
97 // phases of the CMS thread's work do not block out the VM thread
98 // for long periods of time as the CMS thread continues to hog
99 // the token. (See bug 4616232).
100 //
101 // The baton-passing functions are, however, controlled by the
102 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
103 // and here the low-level CMS lock, not the high level token,
104 // ensures mutual exclusion.
105 //
106 // Two important conditions that we have to satisfy:
107 // 1. if a thread does a low-level wait on the CMS lock, then it
108 // relinquishes the CMS token if it were holding that token
109 // when it acquired the low-level CMS lock.
110 // 2. any low-level notifications on the low-level lock
111 // should only be sent when a thread has relinquished the token.
112 //
113 // In the absence of either property, we'd have potential deadlock.
114 //
115 // We protect each of the CMS (concurrent and sequential) phases
116 // with the CMS _token_, not the CMS _lock_.
117 //
118 // The only code protected by CMS lock is the token acquisition code
119 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
120 // baton-passing code.
121 //
122 // Unfortunately, i couldn't come up with a good abstraction to factor and
123 // hide the naked CGC_lock manipulation in the baton-passing code
124 // further below. That's something we should try to do. Also, the proof
125 // of correctness of this 2-level locking scheme is far from obvious,
126 // and potentially quite slippery. We have an uneasy suspicion, for instance,
127 // that there may be a theoretical possibility of delay/starvation in the
128 // low-level lock/wait/notify scheme used for the baton-passing because of
129 // potential interference with the priority scheme embodied in the
130 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
131 // invocation further below and marked with "XXX 20011219YSR".
132 // Indeed, as we note elsewhere, this may become yet more slippery
133 // in the presence of multiple CMS and/or multiple VM threads. XXX
134
135 class CMSTokenSync: public StackObj {
136 private:
137 bool _is_cms_thread;
138 public:
139 CMSTokenSync(bool is_cms_thread):
140 _is_cms_thread(is_cms_thread) {
141 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
142 "Incorrect argument to constructor");
143 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
144 }
145
146 ~CMSTokenSync() {
147 assert(_is_cms_thread ?
148 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
149 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
150 "Incorrect state");
151 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
152 }
153 };
154
155 // Convenience class that does a CMSTokenSync, and then acquires
156 // upto three locks.
157 class CMSTokenSyncWithLocks: public CMSTokenSync {
158 private:
159 // Note: locks are acquired in textual declaration order
160 // and released in the opposite order
161 MutexLockerEx _locker1, _locker2, _locker3;
162 public:
163 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
164 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
165 CMSTokenSync(is_cms_thread),
166 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
167 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
168 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
169 { }
170 };
171
172
173 //////////////////////////////////////////////////////////////////
174 // Concurrent Mark-Sweep Generation /////////////////////////////
175 //////////////////////////////////////////////////////////////////
176
177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
178
179 // This struct contains per-thread things necessary to support parallel
180 // young-gen collection.
181 class CMSParGCThreadState: public CHeapObj<mtGC> {
182 public:
183 CFLS_LAB lab;
184 PromotionInfo promo;
185
186 // Constructor.
187 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
188 promo.setSpace(cfls);
189 }
190 };
191
192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
193 ReservedSpace rs, size_t initial_byte_size, int level,
194 CardTableRS* ct, bool use_adaptive_freelists,
195 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
196 CardGeneration(rs, initial_byte_size, level, ct),
197 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
198 _did_compact(false)
199 {
200 HeapWord* bottom = (HeapWord*) _virtual_space.low();
201 HeapWord* end = (HeapWord*) _virtual_space.high();
202
203 _direct_allocated_words = 0;
204 NOT_PRODUCT(
205 _numObjectsPromoted = 0;
206 _numWordsPromoted = 0;
207 _numObjectsAllocated = 0;
208 _numWordsAllocated = 0;
209 )
210
211 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
212 use_adaptive_freelists,
213 dictionaryChoice);
214 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
215 _cmsSpace->_gen = this;
216
217 _gc_stats = new CMSGCStats();
218
219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
220 // offsets match. The ability to tell free chunks from objects
221 // depends on this property.
222 debug_only(
223 FreeChunk* junk = NULL;
224 assert(UseCompressedClassPointers ||
225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
226 "Offset of FreeChunk::_prev within FreeChunk must match"
227 " that of OopDesc::_klass within OopDesc");
228 )
229
230 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
231 for (uint i = 0; i < ParallelGCThreads; i++) {
232 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
233 }
234
235 _incremental_collection_failed = false;
236 // The "dilatation_factor" is the expansion that can occur on
237 // account of the fact that the minimum object size in the CMS
238 // generation may be larger than that in, say, a contiguous young
239 // generation.
240 // Ideally, in the calculation below, we'd compute the dilatation
241 // factor as: MinChunkSize/(promoting_gen's min object size)
242 // Since we do not have such a general query interface for the
243 // promoting generation, we'll instead just use the minimum
244 // object size (which today is a header's worth of space);
245 // note that all arithmetic is in units of HeapWords.
246 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
247 assert(_dilatation_factor >= 1.0, "from previous assert");
248 }
249
250
251 // The field "_initiating_occupancy" represents the occupancy percentage
252 // at which we trigger a new collection cycle. Unless explicitly specified
253 // via CMSInitiatingOccupancyFraction (argument "io" below), it
254 // is calculated by:
255 //
256 // Let "f" be MinHeapFreeRatio in
257 //
258 // _initiating_occupancy = 100-f +
259 // f * (CMSTriggerRatio/100)
260 // where CMSTriggerRatio is the argument "tr" below.
261 //
262 // That is, if we assume the heap is at its desired maximum occupancy at the
263 // end of a collection, we let CMSTriggerRatio of the (purported) free
264 // space be allocated before initiating a new collection cycle.
265 //
266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
267 assert(io <= 100 && tr <= 100, "Check the arguments");
268 if (io >= 0) {
269 _initiating_occupancy = (double)io / 100.0;
270 } else {
271 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
272 (double)(tr * MinHeapFreeRatio) / 100.0)
273 / 100.0;
274 }
275 }
276
277 void ConcurrentMarkSweepGeneration::ref_processor_init() {
278 assert(collector() != NULL, "no collector");
279 collector()->ref_processor_init();
280 }
281
282 void CMSCollector::ref_processor_init() {
283 if (_ref_processor == NULL) {
284 // Allocate and initialize a reference processor
285 _ref_processor =
286 new ReferenceProcessor(_span, // span
287 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
288 (int) ParallelGCThreads, // mt processing degree
289 _cmsGen->refs_discovery_is_mt(), // mt discovery
290 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
291 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
292 &_is_alive_closure); // closure for liveness info
293 // Initialize the _ref_processor field of CMSGen
294 _cmsGen->set_ref_processor(_ref_processor);
295
296 }
297 }
298
299 AdaptiveSizePolicy* CMSCollector::size_policy() {
300 GenCollectedHeap* gch = GenCollectedHeap::heap();
301 return gch->gen_policy()->size_policy();
302 }
303
304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
305
306 const char* gen_name = "old";
307 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy();
308
309 // Generation Counters - generation 1, 1 subspace
310 _gen_counters = new GenerationCounters(gen_name, 1, 1,
311 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
312
313 _space_counters = new GSpaceCounters(gen_name, 0,
314 _virtual_space.reserved_size(),
315 this, _gen_counters);
316 }
317
318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
319 _cms_gen(cms_gen)
320 {
321 assert(alpha <= 100, "bad value");
322 _saved_alpha = alpha;
323
324 // Initialize the alphas to the bootstrap value of 100.
325 _gc0_alpha = _cms_alpha = 100;
326
327 _cms_begin_time.update();
328 _cms_end_time.update();
329
330 _gc0_duration = 0.0;
331 _gc0_period = 0.0;
332 _gc0_promoted = 0;
333
334 _cms_duration = 0.0;
335 _cms_period = 0.0;
336 _cms_allocated = 0;
337
338 _cms_used_at_gc0_begin = 0;
339 _cms_used_at_gc0_end = 0;
340 _allow_duty_cycle_reduction = false;
341 _valid_bits = 0;
342 }
343
344 double CMSStats::cms_free_adjustment_factor(size_t free) const {
345 // TBD: CR 6909490
346 return 1.0;
347 }
348
349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
350 }
351
352 // If promotion failure handling is on use
353 // the padded average size of the promotion for each
354 // young generation collection.
355 double CMSStats::time_until_cms_gen_full() const {
356 size_t cms_free = _cms_gen->cmsSpace()->free();
357 GenCollectedHeap* gch = GenCollectedHeap::heap();
358 size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
359 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
360 if (cms_free > expected_promotion) {
361 // Start a cms collection if there isn't enough space to promote
362 // for the next minor collection. Use the padded average as
363 // a safety factor.
364 cms_free -= expected_promotion;
365
366 // Adjust by the safety factor.
367 double cms_free_dbl = (double)cms_free;
368 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
369 // Apply a further correction factor which tries to adjust
370 // for recent occurance of concurrent mode failures.
371 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
372 cms_free_dbl = cms_free_dbl * cms_adjustment;
373
374 if (PrintGCDetails && Verbose) {
375 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
376 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
377 cms_free, expected_promotion);
378 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
379 cms_free_dbl, cms_consumption_rate() + 1.0);
380 }
381 // Add 1 in case the consumption rate goes to zero.
382 return cms_free_dbl / (cms_consumption_rate() + 1.0);
383 }
384 return 0.0;
385 }
386
387 // Compare the duration of the cms collection to the
388 // time remaining before the cms generation is empty.
389 // Note that the time from the start of the cms collection
390 // to the start of the cms sweep (less than the total
391 // duration of the cms collection) can be used. This
392 // has been tried and some applications experienced
393 // promotion failures early in execution. This was
394 // possibly because the averages were not accurate
395 // enough at the beginning.
396 double CMSStats::time_until_cms_start() const {
397 // We add "gc0_period" to the "work" calculation
398 // below because this query is done (mostly) at the
399 // end of a scavenge, so we need to conservatively
400 // account for that much possible delay
401 // in the query so as to avoid concurrent mode failures
402 // due to starting the collection just a wee bit too
403 // late.
404 double work = cms_duration() + gc0_period();
405 double deadline = time_until_cms_gen_full();
406 // If a concurrent mode failure occurred recently, we want to be
407 // more conservative and halve our expected time_until_cms_gen_full()
408 if (work > deadline) {
409 if (Verbose && PrintGCDetails) {
410 gclog_or_tty->print(
411 " CMSCollector: collect because of anticipated promotion "
412 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
413 gc0_period(), time_until_cms_gen_full());
414 }
415 return 0.0;
416 }
417 return work - deadline;
418 }
419
420 #ifndef PRODUCT
421 void CMSStats::print_on(outputStream *st) const {
422 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
423 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
424 gc0_duration(), gc0_period(), gc0_promoted());
425 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
426 cms_duration(), cms_period(), cms_allocated());
427 st->print(",cms_since_beg=%g,cms_since_end=%g",
428 cms_time_since_begin(), cms_time_since_end());
429 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
430 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
431
432 if (valid()) {
433 st->print(",promo_rate=%g,cms_alloc_rate=%g",
434 promotion_rate(), cms_allocation_rate());
435 st->print(",cms_consumption_rate=%g,time_until_full=%g",
436 cms_consumption_rate(), time_until_cms_gen_full());
437 }
438 st->print(" ");
439 }
440 #endif // #ifndef PRODUCT
441
442 CMSCollector::CollectorState CMSCollector::_collectorState =
443 CMSCollector::Idling;
444 bool CMSCollector::_foregroundGCIsActive = false;
445 bool CMSCollector::_foregroundGCShouldWait = false;
446
447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
448 CardTableRS* ct,
449 ConcurrentMarkSweepPolicy* cp):
450 _cmsGen(cmsGen),
451 _ct(ct),
452 _ref_processor(NULL), // will be set later
453 _conc_workers(NULL), // may be set later
454 _abort_preclean(false),
455 _start_sampling(false),
456 _between_prologue_and_epilogue(false),
457 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
458 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
459 -1 /* lock-free */, "No_lock" /* dummy */),
460 _modUnionClosurePar(&_modUnionTable),
461 // Adjust my span to cover old (cms) gen
462 _span(cmsGen->reserved()),
463 // Construct the is_alive_closure with _span & markBitMap
464 _is_alive_closure(_span, &_markBitMap),
465 _restart_addr(NULL),
466 _overflow_list(NULL),
467 _stats(cmsGen),
468 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
469 //verify that this lock should be acquired with safepoint check.
470 Monitor::_safepoint_check_sometimes)),
471 _eden_chunk_array(NULL), // may be set in ctor body
472 _eden_chunk_capacity(0), // -- ditto --
473 _eden_chunk_index(0), // -- ditto --
474 _survivor_plab_array(NULL), // -- ditto --
475 _survivor_chunk_array(NULL), // -- ditto --
476 _survivor_chunk_capacity(0), // -- ditto --
477 _survivor_chunk_index(0), // -- ditto --
478 _ser_pmc_preclean_ovflw(0),
479 _ser_kac_preclean_ovflw(0),
480 _ser_pmc_remark_ovflw(0),
481 _par_pmc_remark_ovflw(0),
482 _ser_kac_ovflw(0),
483 _par_kac_ovflw(0),
484 #ifndef PRODUCT
485 _num_par_pushes(0),
486 #endif
487 _collection_count_start(0),
488 _verifying(false),
489 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
490 _completed_initialization(false),
491 _collector_policy(cp),
492 _should_unload_classes(CMSClassUnloadingEnabled),
493 _concurrent_cycles_since_last_unload(0),
494 _roots_scanning_options(GenCollectedHeap::SO_None),
495 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
496 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
497 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
498 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
499 _cms_start_registered(false)
500 {
501 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
502 ExplicitGCInvokesConcurrent = true;
503 }
504 // Now expand the span and allocate the collection support structures
505 // (MUT, marking bit map etc.) to cover both generations subject to
506 // collection.
507
508 // For use by dirty card to oop closures.
509 _cmsGen->cmsSpace()->set_collector(this);
510
511 // Allocate MUT and marking bit map
512 {
513 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
514 if (!_markBitMap.allocate(_span)) {
515 warning("Failed to allocate CMS Bit Map");
516 return;
517 }
518 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
519 }
520 {
521 _modUnionTable.allocate(_span);
522 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
523 }
524
525 if (!_markStack.allocate(MarkStackSize)) {
526 warning("Failed to allocate CMS Marking Stack");
527 return;
528 }
529
530 // Support for multi-threaded concurrent phases
531 if (CMSConcurrentMTEnabled) {
532 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
533 // just for now
534 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
535 }
536 if (ConcGCThreads > 1) {
537 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
538 ConcGCThreads, true);
539 if (_conc_workers == NULL) {
540 warning("GC/CMS: _conc_workers allocation failure: "
541 "forcing -CMSConcurrentMTEnabled");
542 CMSConcurrentMTEnabled = false;
543 } else {
544 _conc_workers->initialize_workers();
545 }
546 } else {
547 CMSConcurrentMTEnabled = false;
548 }
549 }
550 if (!CMSConcurrentMTEnabled) {
551 ConcGCThreads = 0;
552 } else {
553 // Turn off CMSCleanOnEnter optimization temporarily for
554 // the MT case where it's not fixed yet; see 6178663.
555 CMSCleanOnEnter = false;
556 }
557 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
558 "Inconsistency");
559
560 // Parallel task queues; these are shared for the
561 // concurrent and stop-world phases of CMS, but
562 // are not shared with parallel scavenge (ParNew).
563 {
564 uint i;
565 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
566
567 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
568 || ParallelRefProcEnabled)
569 && num_queues > 0) {
570 _task_queues = new OopTaskQueueSet(num_queues);
571 if (_task_queues == NULL) {
572 warning("task_queues allocation failure.");
573 return;
574 }
575 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
576 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
577 for (i = 0; i < num_queues; i++) {
578 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
579 if (q == NULL) {
580 warning("work_queue allocation failure.");
581 return;
582 }
583 _task_queues->register_queue(i, q);
584 }
585 for (i = 0; i < num_queues; i++) {
586 _task_queues->queue(i)->initialize();
587 _hash_seed[i] = 17; // copied from ParNew
588 }
589 }
590 }
591
592 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
593
594 // Clip CMSBootstrapOccupancy between 0 and 100.
595 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
596
597 // Now tell CMS generations the identity of their collector
598 ConcurrentMarkSweepGeneration::set_collector(this);
599
600 // Create & start a CMS thread for this CMS collector
601 _cmsThread = ConcurrentMarkSweepThread::start(this);
602 assert(cmsThread() != NULL, "CMS Thread should have been created");
603 assert(cmsThread()->collector() == this,
604 "CMS Thread should refer to this gen");
605 assert(CGC_lock != NULL, "Where's the CGC_lock?");
606
607 // Support for parallelizing young gen rescan
608 GenCollectedHeap* gch = GenCollectedHeap::heap();
609 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
610 _young_gen = (ParNewGeneration*)gch->young_gen();
611 if (gch->supports_inline_contig_alloc()) {
612 _top_addr = gch->top_addr();
613 _end_addr = gch->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 // The 2*K (default MinTLABSize) is large enough to allow smooth striping of work
623 // and avoids being linked to unusual MinTLABSize set on the command line
624 const size_t max_plab_samples = ((DefNewGeneration*)_young_gen)->max_survivor_size() / (2 * K);
625
626 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
627 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
628 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
629 _survivor_chunk_capacity = 2*max_plab_samples;
630 for (uint i = 0; i < ParallelGCThreads; i++) {
631 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
632 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
633 assert(cur->end() == 0, "Should be 0");
634 assert(cur->array() == vec, "Should be vec");
635 assert(cur->capacity() == max_plab_samples, "Error");
636 }
637 }
638
639 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
640 _gc_counters = new CollectorCounters("CMS", 1);
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 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
678 GenCollectedHeap* gch = GenCollectedHeap::heap();
679 if (PrintGCDetails) {
680 if (Verbose) {
681 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
682 level(), short_name(), s, used(), capacity());
683 } else {
684 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
685 level(), short_name(), s, used() / K, capacity() / K);
686 }
687 }
688 if (Verbose) {
689 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
690 gch->used(), gch->capacity());
691 } else {
692 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
693 gch->used() / K, gch->capacity() / K);
694 }
695 }
696
697 size_t
698 ConcurrentMarkSweepGeneration::contiguous_available() const {
699 // dld proposes an improvement in precision here. If the committed
700 // part of the space ends in a free block we should add that to
701 // uncommitted size in the calculation below. Will make this
702 // change later, staying with the approximation below for the
703 // time being. -- ysr.
704 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
705 }
706
707 size_t
708 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
709 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
710 }
711
712 size_t ConcurrentMarkSweepGeneration::max_available() const {
713 return free() + _virtual_space.uncommitted_size();
714 }
715
716 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
717 size_t available = max_available();
718 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
719 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
720 if (Verbose && PrintGCDetails) {
721 gclog_or_tty->print_cr(
722 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
723 "max_promo("SIZE_FORMAT")",
724 res? "":" not", available, res? ">=":"<",
725 av_promo, max_promotion_in_bytes);
726 }
727 return res;
728 }
729
730 // At a promotion failure dump information on block layout in heap
731 // (cms old generation).
732 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
733 if (CMSDumpAtPromotionFailure) {
734 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
735 }
736 }
737
738 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
739 // Clear the promotion information. These pointers can be adjusted
740 // along with all the other pointers into the heap but
741 // compaction is expected to be a rare event with
742 // a heap using cms so don't do it without seeing the need.
743 for (uint i = 0; i < ParallelGCThreads; i++) {
744 _par_gc_thread_states[i]->promo.reset();
745 }
746 }
747
748 void ConcurrentMarkSweepGeneration::compute_new_size() {
749 assert_locked_or_safepoint(Heap_lock);
750
751 // If incremental collection failed, we just want to expand
752 // to the limit.
753 if (incremental_collection_failed()) {
754 clear_incremental_collection_failed();
755 grow_to_reserved();
756 return;
757 }
758
759 // The heap has been compacted but not reset yet.
760 // Any metric such as free() or used() will be incorrect.
761
762 CardGeneration::compute_new_size();
763
764 // Reset again after a possible resizing
765 if (did_compact()) {
766 cmsSpace()->reset_after_compaction();
767 }
768 }
769
770 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
771 assert_locked_or_safepoint(Heap_lock);
772
773 // If incremental collection failed, we just want to expand
774 // to the limit.
775 if (incremental_collection_failed()) {
776 clear_incremental_collection_failed();
777 grow_to_reserved();
778 return;
779 }
780
781 double free_percentage = ((double) free()) / capacity();
782 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
783 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
784
785 // compute expansion delta needed for reaching desired free percentage
786 if (free_percentage < desired_free_percentage) {
787 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
788 assert(desired_capacity >= capacity(), "invalid expansion size");
789 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
790 if (PrintGCDetails && Verbose) {
791 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
792 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
793 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
794 gclog_or_tty->print_cr(" Desired free fraction %f",
795 desired_free_percentage);
796 gclog_or_tty->print_cr(" Maximum free fraction %f",
797 maximum_free_percentage);
798 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000);
799 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
800 desired_capacity/1000);
801 int prev_level = level() - 1;
802 if (prev_level >= 0) {
803 size_t prev_size = 0;
804 GenCollectedHeap* gch = GenCollectedHeap::heap();
805 Generation* prev_gen = gch->young_gen();
806 prev_size = prev_gen->capacity();
807 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
808 prev_size/1000);
809 }
810 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
811 unsafe_max_alloc_nogc()/1000);
812 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
813 contiguous_available()/1000);
814 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
815 expand_bytes);
816 }
817 // safe if expansion fails
818 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
819 if (PrintGCDetails && Verbose) {
820 gclog_or_tty->print_cr(" Expanded free fraction %f",
821 ((double) free()) / capacity());
822 }
823 } else {
824 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
825 assert(desired_capacity <= capacity(), "invalid expansion size");
826 size_t shrink_bytes = capacity() - desired_capacity;
827 // Don't shrink unless the delta is greater than the minimum shrink we want
828 if (shrink_bytes >= MinHeapDeltaBytes) {
829 shrink_free_list_by(shrink_bytes);
830 }
831 }
832 }
833
834 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
835 return cmsSpace()->freelistLock();
836 }
837
838 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
839 bool tlab) {
840 CMSSynchronousYieldRequest yr;
841 MutexLockerEx x(freelistLock(),
842 Mutex::_no_safepoint_check_flag);
843 return have_lock_and_allocate(size, tlab);
844 }
845
846 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
847 bool tlab /* ignored */) {
848 assert_lock_strong(freelistLock());
849 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
850 HeapWord* res = cmsSpace()->allocate(adjustedSize);
851 // Allocate the object live (grey) if the background collector has
852 // started marking. This is necessary because the marker may
853 // have passed this address and consequently this object will
854 // not otherwise be greyed and would be incorrectly swept up.
855 // Note that if this object contains references, the writing
856 // of those references will dirty the card containing this object
857 // allowing the object to be blackened (and its references scanned)
858 // either during a preclean phase or at the final checkpoint.
859 if (res != NULL) {
860 // We may block here with an uninitialized object with
861 // its mark-bit or P-bits not yet set. Such objects need
862 // to be safely navigable by block_start().
863 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
864 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
865 collector()->direct_allocated(res, adjustedSize);
866 _direct_allocated_words += adjustedSize;
867 // allocation counters
868 NOT_PRODUCT(
869 _numObjectsAllocated++;
870 _numWordsAllocated += (int)adjustedSize;
871 )
872 }
873 return res;
874 }
875
876 // In the case of direct allocation by mutators in a generation that
877 // is being concurrently collected, the object must be allocated
878 // live (grey) if the background collector has started marking.
879 // This is necessary because the marker may
880 // have passed this address and consequently this object will
881 // not otherwise be greyed and would be incorrectly swept up.
882 // Note that if this object contains references, the writing
883 // of those references will dirty the card containing this object
884 // allowing the object to be blackened (and its references scanned)
885 // either during a preclean phase or at the final checkpoint.
886 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
887 assert(_markBitMap.covers(start, size), "Out of bounds");
888 if (_collectorState >= Marking) {
889 MutexLockerEx y(_markBitMap.lock(),
890 Mutex::_no_safepoint_check_flag);
891 // [see comments preceding SweepClosure::do_blk() below for details]
892 //
893 // Can the P-bits be deleted now? JJJ
894 //
895 // 1. need to mark the object as live so it isn't collected
896 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
897 // 3. need to mark the end of the object so marking, precleaning or sweeping
898 // can skip over uninitialized or unparsable objects. An allocated
899 // object is considered uninitialized for our purposes as long as
900 // its klass word is NULL. All old gen objects are parsable
901 // as soon as they are initialized.)
902 _markBitMap.mark(start); // object is live
903 _markBitMap.mark(start + 1); // object is potentially uninitialized?
904 _markBitMap.mark(start + size - 1);
905 // mark end of object
906 }
907 // check that oop looks uninitialized
908 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
909 }
910
911 void CMSCollector::promoted(bool par, HeapWord* start,
912 bool is_obj_array, size_t obj_size) {
913 assert(_markBitMap.covers(start), "Out of bounds");
914 // See comment in direct_allocated() about when objects should
915 // be allocated live.
916 if (_collectorState >= Marking) {
917 // we already hold the marking bit map lock, taken in
918 // the prologue
919 if (par) {
920 _markBitMap.par_mark(start);
921 } else {
922 _markBitMap.mark(start);
923 }
924 // We don't need to mark the object as uninitialized (as
925 // in direct_allocated above) because this is being done with the
926 // world stopped and the object will be initialized by the
927 // time the marking, precleaning or sweeping get to look at it.
928 // But see the code for copying objects into the CMS generation,
929 // where we need to ensure that concurrent readers of the
930 // block offset table are able to safely navigate a block that
931 // is in flux from being free to being allocated (and in
932 // transition while being copied into) and subsequently
933 // becoming a bona-fide object when the copy/promotion is complete.
934 assert(SafepointSynchronize::is_at_safepoint(),
935 "expect promotion only at safepoints");
936
937 if (_collectorState < Sweeping) {
938 // Mark the appropriate cards in the modUnionTable, so that
939 // this object gets scanned before the sweep. If this is
940 // not done, CMS generation references in the object might
941 // not get marked.
942 // For the case of arrays, which are otherwise precisely
943 // marked, we need to dirty the entire array, not just its head.
944 if (is_obj_array) {
945 // The [par_]mark_range() method expects mr.end() below to
946 // be aligned to the granularity of a bit's representation
947 // in the heap. In the case of the MUT below, that's a
948 // card size.
949 MemRegion mr(start,
950 (HeapWord*)round_to((intptr_t)(start + obj_size),
951 CardTableModRefBS::card_size /* bytes */));
952 if (par) {
953 _modUnionTable.par_mark_range(mr);
954 } else {
955 _modUnionTable.mark_range(mr);
956 }
957 } else { // not an obj array; we can just mark the head
958 if (par) {
959 _modUnionTable.par_mark(start);
960 } else {
961 _modUnionTable.mark(start);
962 }
963 }
964 }
965 }
966 }
967
968 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
969 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
970 // allocate, copy and if necessary update promoinfo --
971 // delegate to underlying space.
972 assert_lock_strong(freelistLock());
973
974 #ifndef PRODUCT
975 if (GenCollectedHeap::heap()->promotion_should_fail()) {
976 return NULL;
977 }
978 #endif // #ifndef PRODUCT
979
980 oop res = _cmsSpace->promote(obj, obj_size);
981 if (res == NULL) {
982 // expand and retry
983 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
984 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
985 // Since this is the old generation, we don't try to promote
986 // into a more senior generation.
987 res = _cmsSpace->promote(obj, obj_size);
988 }
989 if (res != NULL) {
990 // See comment in allocate() about when objects should
991 // be allocated live.
992 assert(obj->is_oop(), "Will dereference klass pointer below");
993 collector()->promoted(false, // Not parallel
994 (HeapWord*)res, obj->is_objArray(), obj_size);
995 // promotion counters
996 NOT_PRODUCT(
997 _numObjectsPromoted++;
998 _numWordsPromoted +=
999 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1000 )
1001 }
1002 return res;
1003 }
1004
1005
1006 // IMPORTANT: Notes on object size recognition in CMS.
1007 // ---------------------------------------------------
1008 // A block of storage in the CMS generation is always in
1009 // one of three states. A free block (FREE), an allocated
1010 // object (OBJECT) whose size() method reports the correct size,
1011 // and an intermediate state (TRANSIENT) in which its size cannot
1012 // be accurately determined.
1013 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1014 // -----------------------------------------------------
1015 // FREE: klass_word & 1 == 1; mark_word holds block size
1016 //
1017 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1018 // obj->size() computes correct size
1019 //
1020 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1021 //
1022 // STATE IDENTIFICATION: (64 bit+COOPS)
1023 // ------------------------------------
1024 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1025 //
1026 // OBJECT: klass_word installed; klass_word != 0;
1027 // obj->size() computes correct size
1028 //
1029 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1030 //
1031 //
1032 // STATE TRANSITION DIAGRAM
1033 //
1034 // mut / parnew mut / parnew
1035 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1036 // ^ |
1037 // |------------------------ DEAD <------------------------------------|
1038 // sweep mut
1039 //
1040 // While a block is in TRANSIENT state its size cannot be determined
1041 // so readers will either need to come back later or stall until
1042 // the size can be determined. Note that for the case of direct
1043 // allocation, P-bits, when available, may be used to determine the
1044 // size of an object that may not yet have been initialized.
1045
1046 // Things to support parallel young-gen collection.
1047 oop
1048 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1049 oop old, markOop m,
1050 size_t word_sz) {
1051 #ifndef PRODUCT
1052 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1053 return NULL;
1054 }
1055 #endif // #ifndef PRODUCT
1056
1057 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1058 PromotionInfo* promoInfo = &ps->promo;
1059 // if we are tracking promotions, then first ensure space for
1060 // promotion (including spooling space for saving header if necessary).
1061 // then allocate and copy, then track promoted info if needed.
1062 // When tracking (see PromotionInfo::track()), the mark word may
1063 // be displaced and in this case restoration of the mark word
1064 // occurs in the (oop_since_save_marks_)iterate phase.
1065 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1066 // Out of space for allocating spooling buffers;
1067 // try expanding and allocating spooling buffers.
1068 if (!expand_and_ensure_spooling_space(promoInfo)) {
1069 return NULL;
1070 }
1071 }
1072 assert(promoInfo->has_spooling_space(), "Control point invariant");
1073 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1074 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1075 if (obj_ptr == NULL) {
1076 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1077 if (obj_ptr == NULL) {
1078 return NULL;
1079 }
1080 }
1081 oop obj = oop(obj_ptr);
1082 OrderAccess::storestore();
1083 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1084 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1085 // IMPORTANT: See note on object initialization for CMS above.
1086 // Otherwise, copy the object. Here we must be careful to insert the
1087 // klass pointer last, since this marks the block as an allocated object.
1088 // Except with compressed oops it's the mark word.
1089 HeapWord* old_ptr = (HeapWord*)old;
1090 // Restore the mark word copied above.
1091 obj->set_mark(m);
1092 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1093 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1094 OrderAccess::storestore();
1095
1096 if (UseCompressedClassPointers) {
1097 // Copy gap missed by (aligned) header size calculation below
1098 obj->set_klass_gap(old->klass_gap());
1099 }
1100 if (word_sz > (size_t)oopDesc::header_size()) {
1101 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1102 obj_ptr + oopDesc::header_size(),
1103 word_sz - oopDesc::header_size());
1104 }
1105
1106 // Now we can track the promoted object, if necessary. We take care
1107 // to delay the transition from uninitialized to full object
1108 // (i.e., insertion of klass pointer) until after, so that it
1109 // atomically becomes a promoted object.
1110 if (promoInfo->tracking()) {
1111 promoInfo->track((PromotedObject*)obj, old->klass());
1112 }
1113 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1114 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1115 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1116
1117 // Finally, install the klass pointer (this should be volatile).
1118 OrderAccess::storestore();
1119 obj->set_klass(old->klass());
1120 // We should now be able to calculate the right size for this object
1121 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1122
1123 collector()->promoted(true, // parallel
1124 obj_ptr, old->is_objArray(), word_sz);
1125
1126 NOT_PRODUCT(
1127 Atomic::inc_ptr(&_numObjectsPromoted);
1128 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1129 )
1130
1131 return obj;
1132 }
1133
1134 void
1135 ConcurrentMarkSweepGeneration::
1136 par_promote_alloc_done(int thread_num) {
1137 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1138 ps->lab.retire(thread_num);
1139 }
1140
1141 void
1142 ConcurrentMarkSweepGeneration::
1143 par_oop_since_save_marks_iterate_done(int thread_num) {
1144 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1145 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1146 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1147 }
1148
1149 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1150 size_t size,
1151 bool tlab)
1152 {
1153 // We allow a STW collection only if a full
1154 // collection was requested.
1155 return full || should_allocate(size, tlab); // FIX ME !!!
1156 // This and promotion failure handling are connected at the
1157 // hip and should be fixed by untying them.
1158 }
1159
1160 bool CMSCollector::shouldConcurrentCollect() {
1161 if (_full_gc_requested) {
1162 if (Verbose && PrintGCDetails) {
1163 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1164 " gc request (or gc_locker)");
1165 }
1166 return true;
1167 }
1168
1169 FreelistLocker x(this);
1170 // ------------------------------------------------------------------
1171 // Print out lots of information which affects the initiation of
1172 // a collection.
1173 if (PrintCMSInitiationStatistics && stats().valid()) {
1174 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1175 gclog_or_tty->stamp();
1176 gclog_or_tty->cr();
1177 stats().print_on(gclog_or_tty);
1178 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1179 stats().time_until_cms_gen_full());
1180 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1181 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1182 _cmsGen->contiguous_available());
1183 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1184 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1185 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1186 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1187 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1188 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1189 gclog_or_tty->print_cr("metadata initialized %d",
1190 MetaspaceGC::should_concurrent_collect());
1191 }
1192 // ------------------------------------------------------------------
1193
1194 // If the estimated time to complete a cms collection (cms_duration())
1195 // is less than the estimated time remaining until the cms generation
1196 // is full, start a collection.
1197 if (!UseCMSInitiatingOccupancyOnly) {
1198 if (stats().valid()) {
1199 if (stats().time_until_cms_start() == 0.0) {
1200 return true;
1201 }
1202 } else {
1203 // We want to conservatively collect somewhat early in order
1204 // to try and "bootstrap" our CMS/promotion statistics;
1205 // this branch will not fire after the first successful CMS
1206 // collection because the stats should then be valid.
1207 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1208 if (Verbose && PrintGCDetails) {
1209 gclog_or_tty->print_cr(
1210 " CMSCollector: collect for bootstrapping statistics:"
1211 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1212 _bootstrap_occupancy);
1213 }
1214 return true;
1215 }
1216 }
1217 }
1218
1219 // Otherwise, we start a collection cycle if
1220 // old gen want a collection cycle started. Each may use
1221 // an appropriate criterion for making this decision.
1222 // XXX We need to make sure that the gen expansion
1223 // criterion dovetails well with this. XXX NEED TO FIX THIS
1224 if (_cmsGen->should_concurrent_collect()) {
1225 if (Verbose && PrintGCDetails) {
1226 gclog_or_tty->print_cr("CMS old gen initiated");
1227 }
1228 return true;
1229 }
1230
1231 // We start a collection if we believe an incremental collection may fail;
1232 // this is not likely to be productive in practice because it's probably too
1233 // late anyway.
1234 GenCollectedHeap* gch = GenCollectedHeap::heap();
1235 assert(gch->collector_policy()->is_generation_policy(),
1236 "You may want to check the correctness of the following");
1237 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1238 if (Verbose && PrintGCDetails) {
1239 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1240 }
1241 return true;
1242 }
1243
1244 if (MetaspaceGC::should_concurrent_collect()) {
1245 if (Verbose && PrintGCDetails) {
1246 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1247 }
1248 return true;
1249 }
1250
1251 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1252 if (CMSTriggerInterval >= 0) {
1253 if (CMSTriggerInterval == 0) {
1254 // Trigger always
1255 return true;
1256 }
1257
1258 // Check the CMS time since begin (we do not check the stats validity
1259 // as we want to be able to trigger the first CMS cycle as well)
1260 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1261 if (Verbose && PrintGCDetails) {
1262 if (stats().valid()) {
1263 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1264 stats().cms_time_since_begin());
1265 } else {
1266 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1267 }
1268 }
1269 return true;
1270 }
1271 }
1272
1273 return false;
1274 }
1275
1276 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1277
1278 // Clear _expansion_cause fields of constituent generations
1279 void CMSCollector::clear_expansion_cause() {
1280 _cmsGen->clear_expansion_cause();
1281 }
1282
1283 // We should be conservative in starting a collection cycle. To
1284 // start too eagerly runs the risk of collecting too often in the
1285 // extreme. To collect too rarely falls back on full collections,
1286 // which works, even if not optimum in terms of concurrent work.
1287 // As a work around for too eagerly collecting, use the flag
1288 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1289 // giving the user an easily understandable way of controlling the
1290 // collections.
1291 // We want to start a new collection cycle if any of the following
1292 // conditions hold:
1293 // . our current occupancy exceeds the configured initiating occupancy
1294 // for this generation, or
1295 // . we recently needed to expand this space and have not, since that
1296 // expansion, done a collection of this generation, or
1297 // . the underlying space believes that it may be a good idea to initiate
1298 // a concurrent collection (this may be based on criteria such as the
1299 // following: the space uses linear allocation and linear allocation is
1300 // going to fail, or there is believed to be excessive fragmentation in
1301 // the generation, etc... or ...
1302 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1303 // the case of the old generation; see CR 6543076):
1304 // we may be approaching a point at which allocation requests may fail because
1305 // we will be out of sufficient free space given allocation rate estimates.]
1306 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1307
1308 assert_lock_strong(freelistLock());
1309 if (occupancy() > initiating_occupancy()) {
1310 if (PrintGCDetails && Verbose) {
1311 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1312 short_name(), occupancy(), initiating_occupancy());
1313 }
1314 return true;
1315 }
1316 if (UseCMSInitiatingOccupancyOnly) {
1317 return false;
1318 }
1319 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1320 if (PrintGCDetails && Verbose) {
1321 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1322 short_name());
1323 }
1324 return true;
1325 }
1326 if (_cmsSpace->should_concurrent_collect()) {
1327 if (PrintGCDetails && Verbose) {
1328 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1329 short_name());
1330 }
1331 return true;
1332 }
1333 return false;
1334 }
1335
1336 void ConcurrentMarkSweepGeneration::collect(bool full,
1337 bool clear_all_soft_refs,
1338 size_t size,
1339 bool tlab)
1340 {
1341 collector()->collect(full, clear_all_soft_refs, size, tlab);
1342 }
1343
1344 void CMSCollector::collect(bool full,
1345 bool clear_all_soft_refs,
1346 size_t size,
1347 bool tlab)
1348 {
1349 // The following "if" branch is present for defensive reasons.
1350 // In the current uses of this interface, it can be replaced with:
1351 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1352 // But I am not placing that assert here to allow future
1353 // generality in invoking this interface.
1354 if (GC_locker::is_active()) {
1355 // A consistency test for GC_locker
1356 assert(GC_locker::needs_gc(), "Should have been set already");
1357 // Skip this foreground collection, instead
1358 // expanding the heap if necessary.
1359 // Need the free list locks for the call to free() in compute_new_size()
1360 compute_new_size();
1361 return;
1362 }
1363 acquire_control_and_collect(full, clear_all_soft_refs);
1364 }
1365
1366 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1367 GenCollectedHeap* gch = GenCollectedHeap::heap();
1368 unsigned int gc_count = gch->total_full_collections();
1369 if (gc_count == full_gc_count) {
1370 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1371 _full_gc_requested = true;
1372 _full_gc_cause = cause;
1373 CGC_lock->notify(); // nudge CMS thread
1374 } else {
1375 assert(gc_count > full_gc_count, "Error: causal loop");
1376 }
1377 }
1378
1379 bool CMSCollector::is_external_interruption() {
1380 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1381 return GCCause::is_user_requested_gc(cause) ||
1382 GCCause::is_serviceability_requested_gc(cause);
1383 }
1384
1385 void CMSCollector::report_concurrent_mode_interruption() {
1386 if (is_external_interruption()) {
1387 if (PrintGCDetails) {
1388 gclog_or_tty->print(" (concurrent mode interrupted)");
1389 }
1390 } else {
1391 if (PrintGCDetails) {
1392 gclog_or_tty->print(" (concurrent mode failure)");
1393 }
1394 _gc_tracer_cm->report_concurrent_mode_failure();
1395 }
1396 }
1397
1398
1399 // The foreground and background collectors need to coordinate in order
1400 // to make sure that they do not mutually interfere with CMS collections.
1401 // When a background collection is active,
1402 // the foreground collector may need to take over (preempt) and
1403 // synchronously complete an ongoing collection. Depending on the
1404 // frequency of the background collections and the heap usage
1405 // of the application, this preemption can be seldom or frequent.
1406 // There are only certain
1407 // points in the background collection that the "collection-baton"
1408 // can be passed to the foreground collector.
1409 //
1410 // The foreground collector will wait for the baton before
1411 // starting any part of the collection. The foreground collector
1412 // will only wait at one location.
1413 //
1414 // The background collector will yield the baton before starting a new
1415 // phase of the collection (e.g., before initial marking, marking from roots,
1416 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1417 // of the loop which switches the phases. The background collector does some
1418 // of the phases (initial mark, final re-mark) with the world stopped.
1419 // Because of locking involved in stopping the world,
1420 // the foreground collector should not block waiting for the background
1421 // collector when it is doing a stop-the-world phase. The background
1422 // collector will yield the baton at an additional point just before
1423 // it enters a stop-the-world phase. Once the world is stopped, the
1424 // background collector checks the phase of the collection. If the
1425 // phase has not changed, it proceeds with the collection. If the
1426 // phase has changed, it skips that phase of the collection. See
1427 // the comments on the use of the Heap_lock in collect_in_background().
1428 //
1429 // Variable used in baton passing.
1430 // _foregroundGCIsActive - Set to true by the foreground collector when
1431 // it wants the baton. The foreground clears it when it has finished
1432 // the collection.
1433 // _foregroundGCShouldWait - Set to true by the background collector
1434 // when it is running. The foreground collector waits while
1435 // _foregroundGCShouldWait is true.
1436 // CGC_lock - monitor used to protect access to the above variables
1437 // and to notify the foreground and background collectors.
1438 // _collectorState - current state of the CMS collection.
1439 //
1440 // The foreground collector
1441 // acquires the CGC_lock
1442 // sets _foregroundGCIsActive
1443 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1444 // various locks acquired in preparation for the collection
1445 // are released so as not to block the background collector
1446 // that is in the midst of a collection
1447 // proceeds with the collection
1448 // clears _foregroundGCIsActive
1449 // returns
1450 //
1451 // The background collector in a loop iterating on the phases of the
1452 // collection
1453 // acquires the CGC_lock
1454 // sets _foregroundGCShouldWait
1455 // if _foregroundGCIsActive is set
1456 // clears _foregroundGCShouldWait, notifies _CGC_lock
1457 // waits on _CGC_lock for _foregroundGCIsActive to become false
1458 // and exits the loop.
1459 // otherwise
1460 // proceed with that phase of the collection
1461 // if the phase is a stop-the-world phase,
1462 // yield the baton once more just before enqueueing
1463 // the stop-world CMS operation (executed by the VM thread).
1464 // returns after all phases of the collection are done
1465 //
1466
1467 void CMSCollector::acquire_control_and_collect(bool full,
1468 bool clear_all_soft_refs) {
1469 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1470 assert(!Thread::current()->is_ConcurrentGC_thread(),
1471 "shouldn't try to acquire control from self!");
1472
1473 // Start the protocol for acquiring control of the
1474 // collection from the background collector (aka CMS thread).
1475 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1476 "VM thread should have CMS token");
1477 // Remember the possibly interrupted state of an ongoing
1478 // concurrent collection
1479 CollectorState first_state = _collectorState;
1480
1481 // Signal to a possibly ongoing concurrent collection that
1482 // we want to do a foreground collection.
1483 _foregroundGCIsActive = true;
1484
1485 // release locks and wait for a notify from the background collector
1486 // releasing the locks in only necessary for phases which
1487 // do yields to improve the granularity of the collection.
1488 assert_lock_strong(bitMapLock());
1489 // We need to lock the Free list lock for the space that we are
1490 // currently collecting.
1491 assert(haveFreelistLocks(), "Must be holding free list locks");
1492 bitMapLock()->unlock();
1493 releaseFreelistLocks();
1494 {
1495 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1496 if (_foregroundGCShouldWait) {
1497 // We are going to be waiting for action for the CMS thread;
1498 // it had better not be gone (for instance at shutdown)!
1499 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1500 "CMS thread must be running");
1501 // Wait here until the background collector gives us the go-ahead
1502 ConcurrentMarkSweepThread::clear_CMS_flag(
1503 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1504 // Get a possibly blocked CMS thread going:
1505 // Note that we set _foregroundGCIsActive true above,
1506 // without protection of the CGC_lock.
1507 CGC_lock->notify();
1508 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1509 "Possible deadlock");
1510 while (_foregroundGCShouldWait) {
1511 // wait for notification
1512 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1513 // Possibility of delay/starvation here, since CMS token does
1514 // not know to give priority to VM thread? Actually, i think
1515 // there wouldn't be any delay/starvation, but the proof of
1516 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1517 }
1518 ConcurrentMarkSweepThread::set_CMS_flag(
1519 ConcurrentMarkSweepThread::CMS_vm_has_token);
1520 }
1521 }
1522 // The CMS_token is already held. Get back the other locks.
1523 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1524 "VM thread should have CMS token");
1525 getFreelistLocks();
1526 bitMapLock()->lock_without_safepoint_check();
1527 if (TraceCMSState) {
1528 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1529 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1530 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1531 }
1532
1533 // Inform cms gen if this was due to partial collection failing.
1534 // The CMS gen may use this fact to determine its expansion policy.
1535 GenCollectedHeap* gch = GenCollectedHeap::heap();
1536 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1537 assert(!_cmsGen->incremental_collection_failed(),
1538 "Should have been noticed, reacted to and cleared");
1539 _cmsGen->set_incremental_collection_failed();
1540 }
1541
1542 if (first_state > Idling) {
1543 report_concurrent_mode_interruption();
1544 }
1545
1546 set_did_compact(true);
1547
1548 // If the collection is being acquired from the background
1549 // collector, there may be references on the discovered
1550 // references lists. Abandon those references, since some
1551 // of them may have become unreachable after concurrent
1552 // discovery; the STW compacting collector will redo discovery
1553 // more precisely, without being subject to floating garbage.
1554 // Leaving otherwise unreachable references in the discovered
1555 // lists would require special handling.
1556 ref_processor()->disable_discovery();
1557 ref_processor()->abandon_partial_discovery();
1558 ref_processor()->verify_no_references_recorded();
1559
1560 if (first_state > Idling) {
1561 save_heap_summary();
1562 }
1563
1564 do_compaction_work(clear_all_soft_refs);
1565
1566 // Has the GC time limit been exceeded?
1567 size_t max_eden_size = _young_gen->max_capacity() -
1568 _young_gen->to()->capacity() -
1569 _young_gen->from()->capacity();
1570 GCCause::Cause gc_cause = gch->gc_cause();
1571 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1572 _young_gen->eden()->used(),
1573 _cmsGen->max_capacity(),
1574 max_eden_size,
1575 full,
1576 gc_cause,
1577 gch->collector_policy());
1578
1579 // Reset the expansion cause, now that we just completed
1580 // a collection cycle.
1581 clear_expansion_cause();
1582 _foregroundGCIsActive = false;
1583 return;
1584 }
1585
1586 // Resize the tenured generation
1587 // after obtaining the free list locks for the
1588 // two generations.
1589 void CMSCollector::compute_new_size() {
1590 assert_locked_or_safepoint(Heap_lock);
1591 FreelistLocker z(this);
1592 MetaspaceGC::compute_new_size();
1593 _cmsGen->compute_new_size_free_list();
1594 }
1595
1596 // A work method used by the foreground collector to do
1597 // a mark-sweep-compact.
1598 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1599 GenCollectedHeap* gch = GenCollectedHeap::heap();
1600
1601 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1602 gc_timer->register_gc_start();
1603
1604 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1605 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1606
1607 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1608
1609 // Temporarily widen the span of the weak reference processing to
1610 // the entire heap.
1611 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1612 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1613 // Temporarily, clear the "is_alive_non_header" field of the
1614 // reference processor.
1615 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1616 // Temporarily make reference _processing_ single threaded (non-MT).
1617 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1618 // Temporarily make refs discovery atomic
1619 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1620 // Temporarily make reference _discovery_ single threaded (non-MT)
1621 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1622
1623 ref_processor()->set_enqueuing_is_done(false);
1624 ref_processor()->enable_discovery();
1625 ref_processor()->setup_policy(clear_all_soft_refs);
1626 // If an asynchronous collection finishes, the _modUnionTable is
1627 // all clear. If we are assuming the collection from an asynchronous
1628 // collection, clear the _modUnionTable.
1629 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1630 "_modUnionTable should be clear if the baton was not passed");
1631 _modUnionTable.clear_all();
1632 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1633 "mod union for klasses should be clear if the baton was passed");
1634 _ct->klass_rem_set()->clear_mod_union();
1635
1636 // We must adjust the allocation statistics being maintained
1637 // in the free list space. We do so by reading and clearing
1638 // the sweep timer and updating the block flux rate estimates below.
1639 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1640 if (_inter_sweep_timer.is_active()) {
1641 _inter_sweep_timer.stop();
1642 // Note that we do not use this sample to update the _inter_sweep_estimate.
1643 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1644 _inter_sweep_estimate.padded_average(),
1645 _intra_sweep_estimate.padded_average());
1646 }
1647
1648 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1649 ref_processor(), clear_all_soft_refs);
1650 #ifdef ASSERT
1651 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1652 size_t free_size = cms_space->free();
1653 assert(free_size ==
1654 pointer_delta(cms_space->end(), cms_space->compaction_top())
1655 * HeapWordSize,
1656 "All the free space should be compacted into one chunk at top");
1657 assert(cms_space->dictionary()->total_chunk_size(
1658 debug_only(cms_space->freelistLock())) == 0 ||
1659 cms_space->totalSizeInIndexedFreeLists() == 0,
1660 "All the free space should be in a single chunk");
1661 size_t num = cms_space->totalCount();
1662 assert((free_size == 0 && num == 0) ||
1663 (free_size > 0 && (num == 1 || num == 2)),
1664 "There should be at most 2 free chunks after compaction");
1665 #endif // ASSERT
1666 _collectorState = Resetting;
1667 assert(_restart_addr == NULL,
1668 "Should have been NULL'd before baton was passed");
1669 reset(false /* == !concurrent */);
1670 _cmsGen->reset_after_compaction();
1671 _concurrent_cycles_since_last_unload = 0;
1672
1673 // Clear any data recorded in the PLAB chunk arrays.
1674 if (_survivor_plab_array != NULL) {
1675 reset_survivor_plab_arrays();
1676 }
1677
1678 // Adjust the per-size allocation stats for the next epoch.
1679 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1680 // Restart the "inter sweep timer" for the next epoch.
1681 _inter_sweep_timer.reset();
1682 _inter_sweep_timer.start();
1683
1684 gc_timer->register_gc_end();
1685
1686 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1687
1688 // For a mark-sweep-compact, compute_new_size() will be called
1689 // in the heap's do_collection() method.
1690 }
1691
1692 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1693 ContiguousSpace* eden_space = _young_gen->eden();
1694 ContiguousSpace* from_space = _young_gen->from();
1695 ContiguousSpace* to_space = _young_gen->to();
1696 // Eden
1697 if (_eden_chunk_array != NULL) {
1698 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1699 p2i(eden_space->bottom()), p2i(eden_space->top()),
1700 p2i(eden_space->end()), eden_space->capacity());
1701 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1702 "_eden_chunk_capacity=" SIZE_FORMAT,
1703 _eden_chunk_index, _eden_chunk_capacity);
1704 for (size_t i = 0; i < _eden_chunk_index; i++) {
1705 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1706 i, p2i(_eden_chunk_array[i]));
1707 }
1708 }
1709 // Survivor
1710 if (_survivor_chunk_array != NULL) {
1711 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1712 p2i(from_space->bottom()), p2i(from_space->top()),
1713 p2i(from_space->end()), from_space->capacity());
1714 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1715 "_survivor_chunk_capacity=" SIZE_FORMAT,
1716 _survivor_chunk_index, _survivor_chunk_capacity);
1717 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1718 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1719 i, p2i(_survivor_chunk_array[i]));
1720 }
1721 }
1722 }
1723
1724 void CMSCollector::getFreelistLocks() const {
1725 // Get locks for all free lists in all generations that this
1726 // collector is responsible for
1727 _cmsGen->freelistLock()->lock_without_safepoint_check();
1728 }
1729
1730 void CMSCollector::releaseFreelistLocks() const {
1731 // Release locks for all free lists in all generations that this
1732 // collector is responsible for
1733 _cmsGen->freelistLock()->unlock();
1734 }
1735
1736 bool CMSCollector::haveFreelistLocks() const {
1737 // Check locks for all free lists in all generations that this
1738 // collector is responsible for
1739 assert_lock_strong(_cmsGen->freelistLock());
1740 PRODUCT_ONLY(ShouldNotReachHere());
1741 return true;
1742 }
1743
1744 // A utility class that is used by the CMS collector to
1745 // temporarily "release" the foreground collector from its
1746 // usual obligation to wait for the background collector to
1747 // complete an ongoing phase before proceeding.
1748 class ReleaseForegroundGC: public StackObj {
1749 private:
1750 CMSCollector* _c;
1751 public:
1752 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1753 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1754 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1755 // allow a potentially blocked foreground collector to proceed
1756 _c->_foregroundGCShouldWait = false;
1757 if (_c->_foregroundGCIsActive) {
1758 CGC_lock->notify();
1759 }
1760 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1761 "Possible deadlock");
1762 }
1763
1764 ~ReleaseForegroundGC() {
1765 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1766 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1767 _c->_foregroundGCShouldWait = true;
1768 }
1769 };
1770
1771 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1772 assert(Thread::current()->is_ConcurrentGC_thread(),
1773 "A CMS asynchronous collection is only allowed on a CMS thread.");
1774
1775 GenCollectedHeap* gch = GenCollectedHeap::heap();
1776 {
1777 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1778 MutexLockerEx hl(Heap_lock, safepoint_check);
1779 FreelistLocker fll(this);
1780 MutexLockerEx x(CGC_lock, safepoint_check);
1781 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1782 // The foreground collector is active or we're
1783 // not using asynchronous collections. Skip this
1784 // background collection.
1785 assert(!_foregroundGCShouldWait, "Should be clear");
1786 return;
1787 } else {
1788 assert(_collectorState == Idling, "Should be idling before start.");
1789 _collectorState = InitialMarking;
1790 register_gc_start(cause);
1791 // Reset the expansion cause, now that we are about to begin
1792 // a new cycle.
1793 clear_expansion_cause();
1794
1795 // Clear the MetaspaceGC flag since a concurrent collection
1796 // is starting but also clear it after the collection.
1797 MetaspaceGC::set_should_concurrent_collect(false);
1798 }
1799 // Decide if we want to enable class unloading as part of the
1800 // ensuing concurrent GC cycle.
1801 update_should_unload_classes();
1802 _full_gc_requested = false; // acks all outstanding full gc requests
1803 _full_gc_cause = GCCause::_no_gc;
1804 // Signal that we are about to start a collection
1805 gch->increment_total_full_collections(); // ... starting a collection cycle
1806 _collection_count_start = gch->total_full_collections();
1807 }
1808
1809 // Used for PrintGC
1810 size_t prev_used;
1811 if (PrintGC && Verbose) {
1812 prev_used = _cmsGen->used();
1813 }
1814
1815 // The change of the collection state is normally done at this level;
1816 // the exceptions are phases that are executed while the world is
1817 // stopped. For those phases the change of state is done while the
1818 // world is stopped. For baton passing purposes this allows the
1819 // background collector to finish the phase and change state atomically.
1820 // The foreground collector cannot wait on a phase that is done
1821 // while the world is stopped because the foreground collector already
1822 // has the world stopped and would deadlock.
1823 while (_collectorState != Idling) {
1824 if (TraceCMSState) {
1825 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1826 p2i(Thread::current()), _collectorState);
1827 }
1828 // The foreground collector
1829 // holds the Heap_lock throughout its collection.
1830 // holds the CMS token (but not the lock)
1831 // except while it is waiting for the background collector to yield.
1832 //
1833 // The foreground collector should be blocked (not for long)
1834 // if the background collector is about to start a phase
1835 // executed with world stopped. If the background
1836 // collector has already started such a phase, the
1837 // foreground collector is blocked waiting for the
1838 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1839 // are executed in the VM thread.
1840 //
1841 // The locking order is
1842 // PendingListLock (PLL) -- if applicable (FinalMarking)
1843 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1844 // CMS token (claimed in
1845 // stop_world_and_do() -->
1846 // safepoint_synchronize() -->
1847 // CMSThread::synchronize())
1848
1849 {
1850 // Check if the FG collector wants us to yield.
1851 CMSTokenSync x(true); // is cms thread
1852 if (waitForForegroundGC()) {
1853 // We yielded to a foreground GC, nothing more to be
1854 // done this round.
1855 assert(_foregroundGCShouldWait == false, "We set it to false in "
1856 "waitForForegroundGC()");
1857 if (TraceCMSState) {
1858 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1859 " exiting collection CMS state %d",
1860 p2i(Thread::current()), _collectorState);
1861 }
1862 return;
1863 } else {
1864 // The background collector can run but check to see if the
1865 // foreground collector has done a collection while the
1866 // background collector was waiting to get the CGC_lock
1867 // above. If yes, break so that _foregroundGCShouldWait
1868 // is cleared before returning.
1869 if (_collectorState == Idling) {
1870 break;
1871 }
1872 }
1873 }
1874
1875 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1876 "should be waiting");
1877
1878 switch (_collectorState) {
1879 case InitialMarking:
1880 {
1881 ReleaseForegroundGC x(this);
1882 stats().record_cms_begin();
1883 VM_CMS_Initial_Mark initial_mark_op(this);
1884 VMThread::execute(&initial_mark_op);
1885 }
1886 // The collector state may be any legal state at this point
1887 // since the background collector may have yielded to the
1888 // foreground collector.
1889 break;
1890 case Marking:
1891 // initial marking in checkpointRootsInitialWork has been completed
1892 if (markFromRoots()) { // we were successful
1893 assert(_collectorState == Precleaning, "Collector state should "
1894 "have changed");
1895 } else {
1896 assert(_foregroundGCIsActive, "Internal state inconsistency");
1897 }
1898 break;
1899 case Precleaning:
1900 // marking from roots in markFromRoots has been completed
1901 preclean();
1902 assert(_collectorState == AbortablePreclean ||
1903 _collectorState == FinalMarking,
1904 "Collector state should have changed");
1905 break;
1906 case AbortablePreclean:
1907 abortable_preclean();
1908 assert(_collectorState == FinalMarking, "Collector state should "
1909 "have changed");
1910 break;
1911 case FinalMarking:
1912 {
1913 ReleaseForegroundGC x(this);
1914
1915 VM_CMS_Final_Remark final_remark_op(this);
1916 VMThread::execute(&final_remark_op);
1917 }
1918 assert(_foregroundGCShouldWait, "block post-condition");
1919 break;
1920 case Sweeping:
1921 // final marking in checkpointRootsFinal has been completed
1922 sweep();
1923 assert(_collectorState == Resizing, "Collector state change "
1924 "to Resizing must be done under the free_list_lock");
1925
1926 case Resizing: {
1927 // Sweeping has been completed...
1928 // At this point the background collection has completed.
1929 // Don't move the call to compute_new_size() down
1930 // into code that might be executed if the background
1931 // collection was preempted.
1932 {
1933 ReleaseForegroundGC x(this); // unblock FG collection
1934 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1935 CMSTokenSync z(true); // not strictly needed.
1936 if (_collectorState == Resizing) {
1937 compute_new_size();
1938 save_heap_summary();
1939 _collectorState = Resetting;
1940 } else {
1941 assert(_collectorState == Idling, "The state should only change"
1942 " because the foreground collector has finished the collection");
1943 }
1944 }
1945 break;
1946 }
1947 case Resetting:
1948 // CMS heap resizing has been completed
1949 reset(true);
1950 assert(_collectorState == Idling, "Collector state should "
1951 "have changed");
1952
1953 MetaspaceGC::set_should_concurrent_collect(false);
1954
1955 stats().record_cms_end();
1956 // Don't move the concurrent_phases_end() and compute_new_size()
1957 // calls to here because a preempted background collection
1958 // has it's state set to "Resetting".
1959 break;
1960 case Idling:
1961 default:
1962 ShouldNotReachHere();
1963 break;
1964 }
1965 if (TraceCMSState) {
1966 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1967 p2i(Thread::current()), _collectorState);
1968 }
1969 assert(_foregroundGCShouldWait, "block post-condition");
1970 }
1971
1972 // Should this be in gc_epilogue?
1973 collector_policy()->counters()->update_counters();
1974
1975 {
1976 // Clear _foregroundGCShouldWait and, in the event that the
1977 // foreground collector is waiting, notify it, before
1978 // returning.
1979 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1980 _foregroundGCShouldWait = false;
1981 if (_foregroundGCIsActive) {
1982 CGC_lock->notify();
1983 }
1984 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1985 "Possible deadlock");
1986 }
1987 if (TraceCMSState) {
1988 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1989 " exiting collection CMS state %d",
1990 p2i(Thread::current()), _collectorState);
1991 }
1992 if (PrintGC && Verbose) {
1993 _cmsGen->print_heap_change(prev_used);
1994 }
1995 }
1996
1997 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1998 _cms_start_registered = true;
1999 _gc_timer_cm->register_gc_start();
2000 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2001 }
2002
2003 void CMSCollector::register_gc_end() {
2004 if (_cms_start_registered) {
2005 report_heap_summary(GCWhen::AfterGC);
2006
2007 _gc_timer_cm->register_gc_end();
2008 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2009 _cms_start_registered = false;
2010 }
2011 }
2012
2013 void CMSCollector::save_heap_summary() {
2014 GenCollectedHeap* gch = GenCollectedHeap::heap();
2015 _last_heap_summary = gch->create_heap_summary();
2016 _last_metaspace_summary = gch->create_metaspace_summary();
2017 }
2018
2019 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2020 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2021 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2022 }
2023
2024 bool CMSCollector::waitForForegroundGC() {
2025 bool res = false;
2026 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2027 "CMS thread should have CMS token");
2028 // Block the foreground collector until the
2029 // background collectors decides whether to
2030 // yield.
2031 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2032 _foregroundGCShouldWait = true;
2033 if (_foregroundGCIsActive) {
2034 // The background collector yields to the
2035 // foreground collector and returns a value
2036 // indicating that it has yielded. The foreground
2037 // collector can proceed.
2038 res = true;
2039 _foregroundGCShouldWait = false;
2040 ConcurrentMarkSweepThread::clear_CMS_flag(
2041 ConcurrentMarkSweepThread::CMS_cms_has_token);
2042 ConcurrentMarkSweepThread::set_CMS_flag(
2043 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2044 // Get a possibly blocked foreground thread going
2045 CGC_lock->notify();
2046 if (TraceCMSState) {
2047 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2048 p2i(Thread::current()), _collectorState);
2049 }
2050 while (_foregroundGCIsActive) {
2051 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2052 }
2053 ConcurrentMarkSweepThread::set_CMS_flag(
2054 ConcurrentMarkSweepThread::CMS_cms_has_token);
2055 ConcurrentMarkSweepThread::clear_CMS_flag(
2056 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2057 }
2058 if (TraceCMSState) {
2059 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2060 p2i(Thread::current()), _collectorState);
2061 }
2062 return res;
2063 }
2064
2065 // Because of the need to lock the free lists and other structures in
2066 // the collector, common to all the generations that the collector is
2067 // collecting, we need the gc_prologues of individual CMS generations
2068 // delegate to their collector. It may have been simpler had the
2069 // current infrastructure allowed one to call a prologue on a
2070 // collector. In the absence of that we have the generation's
2071 // prologue delegate to the collector, which delegates back
2072 // some "local" work to a worker method in the individual generations
2073 // that it's responsible for collecting, while itself doing any
2074 // work common to all generations it's responsible for. A similar
2075 // comment applies to the gc_epilogue()'s.
2076 // The role of the variable _between_prologue_and_epilogue is to
2077 // enforce the invocation protocol.
2078 void CMSCollector::gc_prologue(bool full) {
2079 // Call gc_prologue_work() for the CMSGen
2080 // we are responsible for.
2081
2082 // The following locking discipline assumes that we are only called
2083 // when the world is stopped.
2084 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2085
2086 // The CMSCollector prologue must call the gc_prologues for the
2087 // "generations" that it's responsible
2088 // for.
2089
2090 assert( Thread::current()->is_VM_thread()
2091 || ( CMSScavengeBeforeRemark
2092 && Thread::current()->is_ConcurrentGC_thread()),
2093 "Incorrect thread type for prologue execution");
2094
2095 if (_between_prologue_and_epilogue) {
2096 // We have already been invoked; this is a gc_prologue delegation
2097 // from yet another CMS generation that we are responsible for, just
2098 // ignore it since all relevant work has already been done.
2099 return;
2100 }
2101
2102 // set a bit saying prologue has been called; cleared in epilogue
2103 _between_prologue_and_epilogue = true;
2104 // Claim locks for common data structures, then call gc_prologue_work()
2105 // for each CMSGen.
2106
2107 getFreelistLocks(); // gets free list locks on constituent spaces
2108 bitMapLock()->lock_without_safepoint_check();
2109
2110 // Should call gc_prologue_work() for all cms gens we are responsible for
2111 bool duringMarking = _collectorState >= Marking
2112 && _collectorState < Sweeping;
2113
2114 // The young collections clear the modified oops state, which tells if
2115 // there are any modified oops in the class. The remark phase also needs
2116 // that information. Tell the young collection to save the union of all
2117 // modified klasses.
2118 if (duringMarking) {
2119 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2120 }
2121
2122 bool registerClosure = duringMarking;
2123
2124 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2125
2126 if (!full) {
2127 stats().record_gc0_begin();
2128 }
2129 }
2130
2131 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2132
2133 _capacity_at_prologue = capacity();
2134 _used_at_prologue = used();
2135
2136 // Delegate to CMScollector which knows how to coordinate between
2137 // this and any other CMS generations that it is responsible for
2138 // collecting.
2139 collector()->gc_prologue(full);
2140 }
2141
2142 // This is a "private" interface for use by this generation's CMSCollector.
2143 // Not to be called directly by any other entity (for instance,
2144 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2145 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2146 bool registerClosure, ModUnionClosure* modUnionClosure) {
2147 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2148 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2149 "Should be NULL");
2150 if (registerClosure) {
2151 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2152 }
2153 cmsSpace()->gc_prologue();
2154 // Clear stat counters
2155 NOT_PRODUCT(
2156 assert(_numObjectsPromoted == 0, "check");
2157 assert(_numWordsPromoted == 0, "check");
2158 if (Verbose && PrintGC) {
2159 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2160 SIZE_FORMAT" bytes concurrently",
2161 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2162 }
2163 _numObjectsAllocated = 0;
2164 _numWordsAllocated = 0;
2165 )
2166 }
2167
2168 void CMSCollector::gc_epilogue(bool full) {
2169 // The following locking discipline assumes that we are only called
2170 // when the world is stopped.
2171 assert(SafepointSynchronize::is_at_safepoint(),
2172 "world is stopped assumption");
2173
2174 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2175 // if linear allocation blocks need to be appropriately marked to allow the
2176 // the blocks to be parsable. We also check here whether we need to nudge the
2177 // CMS collector thread to start a new cycle (if it's not already active).
2178 assert( Thread::current()->is_VM_thread()
2179 || ( CMSScavengeBeforeRemark
2180 && Thread::current()->is_ConcurrentGC_thread()),
2181 "Incorrect thread type for epilogue execution");
2182
2183 if (!_between_prologue_and_epilogue) {
2184 // We have already been invoked; this is a gc_epilogue delegation
2185 // from yet another CMS generation that we are responsible for, just
2186 // ignore it since all relevant work has already been done.
2187 return;
2188 }
2189 assert(haveFreelistLocks(), "must have freelist locks");
2190 assert_lock_strong(bitMapLock());
2191
2192 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2193
2194 _cmsGen->gc_epilogue_work(full);
2195
2196 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2197 // in case sampling was not already enabled, enable it
2198 _start_sampling = true;
2199 }
2200 // reset _eden_chunk_array so sampling starts afresh
2201 _eden_chunk_index = 0;
2202
2203 size_t cms_used = _cmsGen->cmsSpace()->used();
2204
2205 // update performance counters - this uses a special version of
2206 // update_counters() that allows the utilization to be passed as a
2207 // parameter, avoiding multiple calls to used().
2208 //
2209 _cmsGen->update_counters(cms_used);
2210
2211 bitMapLock()->unlock();
2212 releaseFreelistLocks();
2213
2214 if (!CleanChunkPoolAsync) {
2215 Chunk::clean_chunk_pool();
2216 }
2217
2218 set_did_compact(false);
2219 _between_prologue_and_epilogue = false; // ready for next cycle
2220 }
2221
2222 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2223 collector()->gc_epilogue(full);
2224
2225 // Also reset promotion tracking in par gc thread states.
2226 for (uint i = 0; i < ParallelGCThreads; i++) {
2227 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2228 }
2229 }
2230
2231 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2232 assert(!incremental_collection_failed(), "Should have been cleared");
2233 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2234 cmsSpace()->gc_epilogue();
2235 // Print stat counters
2236 NOT_PRODUCT(
2237 assert(_numObjectsAllocated == 0, "check");
2238 assert(_numWordsAllocated == 0, "check");
2239 if (Verbose && PrintGC) {
2240 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2241 SIZE_FORMAT" bytes",
2242 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2243 }
2244 _numObjectsPromoted = 0;
2245 _numWordsPromoted = 0;
2246 )
2247
2248 if (PrintGC && Verbose) {
2249 // Call down the chain in contiguous_available needs the freelistLock
2250 // so print this out before releasing the freeListLock.
2251 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2252 contiguous_available());
2253 }
2254 }
2255
2256 #ifndef PRODUCT
2257 bool CMSCollector::have_cms_token() {
2258 Thread* thr = Thread::current();
2259 if (thr->is_VM_thread()) {
2260 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2261 } else if (thr->is_ConcurrentGC_thread()) {
2262 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2263 } else if (thr->is_GC_task_thread()) {
2264 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2265 ParGCRareEvent_lock->owned_by_self();
2266 }
2267 return false;
2268 }
2269 #endif
2270
2271 // Check reachability of the given heap address in CMS generation,
2272 // treating all other generations as roots.
2273 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2274 // We could "guarantee" below, rather than assert, but I'll
2275 // leave these as "asserts" so that an adventurous debugger
2276 // could try this in the product build provided some subset of
2277 // the conditions were met, provided they were interested in the
2278 // results and knew that the computation below wouldn't interfere
2279 // with other concurrent computations mutating the structures
2280 // being read or written.
2281 assert(SafepointSynchronize::is_at_safepoint(),
2282 "Else mutations in object graph will make answer suspect");
2283 assert(have_cms_token(), "Should hold cms token");
2284 assert(haveFreelistLocks(), "must hold free list locks");
2285 assert_lock_strong(bitMapLock());
2286
2287 // Clear the marking bit map array before starting, but, just
2288 // for kicks, first report if the given address is already marked
2289 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2290 _markBitMap.isMarked(addr) ? "" : " not");
2291
2292 if (verify_after_remark()) {
2293 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2294 bool result = verification_mark_bm()->isMarked(addr);
2295 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2296 result ? "IS" : "is NOT");
2297 return result;
2298 } else {
2299 gclog_or_tty->print_cr("Could not compute result");
2300 return false;
2301 }
2302 }
2303
2304
2305 void
2306 CMSCollector::print_on_error(outputStream* st) {
2307 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2308 if (collector != NULL) {
2309 CMSBitMap* bitmap = &collector->_markBitMap;
2310 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2311 bitmap->print_on_error(st, " Bits: ");
2312
2313 st->cr();
2314
2315 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2316 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2317 mut_bitmap->print_on_error(st, " Bits: ");
2318 }
2319 }
2320
2321 ////////////////////////////////////////////////////////
2322 // CMS Verification Support
2323 ////////////////////////////////////////////////////////
2324 // Following the remark phase, the following invariant
2325 // should hold -- each object in the CMS heap which is
2326 // marked in markBitMap() should be marked in the verification_mark_bm().
2327
2328 class VerifyMarkedClosure: public BitMapClosure {
2329 CMSBitMap* _marks;
2330 bool _failed;
2331
2332 public:
2333 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2334
2335 bool do_bit(size_t offset) {
2336 HeapWord* addr = _marks->offsetToHeapWord(offset);
2337 if (!_marks->isMarked(addr)) {
2338 oop(addr)->print_on(gclog_or_tty);
2339 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr));
2340 _failed = true;
2341 }
2342 return true;
2343 }
2344
2345 bool failed() { return _failed; }
2346 };
2347
2348 bool CMSCollector::verify_after_remark(bool silent) {
2349 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2350 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2351 static bool init = false;
2352
2353 assert(SafepointSynchronize::is_at_safepoint(),
2354 "Else mutations in object graph will make answer suspect");
2355 assert(have_cms_token(),
2356 "Else there may be mutual interference in use of "
2357 " verification data structures");
2358 assert(_collectorState > Marking && _collectorState <= Sweeping,
2359 "Else marking info checked here may be obsolete");
2360 assert(haveFreelistLocks(), "must hold free list locks");
2361 assert_lock_strong(bitMapLock());
2362
2363
2364 // Allocate marking bit map if not already allocated
2365 if (!init) { // first time
2366 if (!verification_mark_bm()->allocate(_span)) {
2367 return false;
2368 }
2369 init = true;
2370 }
2371
2372 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2373
2374 // Turn off refs discovery -- so we will be tracing through refs.
2375 // This is as intended, because by this time
2376 // GC must already have cleared any refs that need to be cleared,
2377 // and traced those that need to be marked; moreover,
2378 // the marking done here is not going to interfere in any
2379 // way with the marking information used by GC.
2380 NoRefDiscovery no_discovery(ref_processor());
2381
2382 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2383
2384 // Clear any marks from a previous round
2385 verification_mark_bm()->clear_all();
2386 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2387 verify_work_stacks_empty();
2388
2389 GenCollectedHeap* gch = GenCollectedHeap::heap();
2390 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2391 // Update the saved marks which may affect the root scans.
2392 gch->save_marks();
2393
2394 if (CMSRemarkVerifyVariant == 1) {
2395 // In this first variant of verification, we complete
2396 // all marking, then check if the new marks-vector is
2397 // a subset of the CMS marks-vector.
2398 verify_after_remark_work_1();
2399 } else if (CMSRemarkVerifyVariant == 2) {
2400 // In this second variant of verification, we flag an error
2401 // (i.e. an object reachable in the new marks-vector not reachable
2402 // in the CMS marks-vector) immediately, also indicating the
2403 // identify of an object (A) that references the unmarked object (B) --
2404 // presumably, a mutation to A failed to be picked up by preclean/remark?
2405 verify_after_remark_work_2();
2406 } else {
2407 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2408 CMSRemarkVerifyVariant);
2409 }
2410 if (!silent) gclog_or_tty->print(" done] ");
2411 return true;
2412 }
2413
2414 void CMSCollector::verify_after_remark_work_1() {
2415 ResourceMark rm;
2416 HandleMark hm;
2417 GenCollectedHeap* gch = GenCollectedHeap::heap();
2418
2419 // Get a clear set of claim bits for the roots processing to work with.
2420 ClassLoaderDataGraph::clear_claimed_marks();
2421
2422 // Mark from roots one level into CMS
2423 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2424 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2425
2426 {
2427 StrongRootsScope srs(1);
2428
2429 gch->gen_process_roots(&srs,
2430 _cmsGen->level(),
2431 true, // younger gens are roots
2432 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2433 should_unload_classes(),
2434 ¬Older,
2435 NULL,
2436 NULL);
2437 }
2438
2439 // Now mark from the roots
2440 MarkFromRootsClosure markFromRootsClosure(this, _span,
2441 verification_mark_bm(), verification_mark_stack(),
2442 false /* don't yield */, true /* verifying */);
2443 assert(_restart_addr == NULL, "Expected pre-condition");
2444 verification_mark_bm()->iterate(&markFromRootsClosure);
2445 while (_restart_addr != NULL) {
2446 // Deal with stack overflow: by restarting at the indicated
2447 // address.
2448 HeapWord* ra = _restart_addr;
2449 markFromRootsClosure.reset(ra);
2450 _restart_addr = NULL;
2451 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2452 }
2453 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2454 verify_work_stacks_empty();
2455
2456 // Marking completed -- now verify that each bit marked in
2457 // verification_mark_bm() is also marked in markBitMap(); flag all
2458 // errors by printing corresponding objects.
2459 VerifyMarkedClosure vcl(markBitMap());
2460 verification_mark_bm()->iterate(&vcl);
2461 if (vcl.failed()) {
2462 gclog_or_tty->print("Verification failed");
2463 gch->print_on(gclog_or_tty);
2464 fatal("CMS: failed marking verification after remark");
2465 }
2466 }
2467
2468 class VerifyKlassOopsKlassClosure : public KlassClosure {
2469 class VerifyKlassOopsClosure : public OopClosure {
2470 CMSBitMap* _bitmap;
2471 public:
2472 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2473 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2474 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2475 } _oop_closure;
2476 public:
2477 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2478 void do_klass(Klass* k) {
2479 k->oops_do(&_oop_closure);
2480 }
2481 };
2482
2483 void CMSCollector::verify_after_remark_work_2() {
2484 ResourceMark rm;
2485 HandleMark hm;
2486 GenCollectedHeap* gch = GenCollectedHeap::heap();
2487
2488 // Get a clear set of claim bits for the roots processing to work with.
2489 ClassLoaderDataGraph::clear_claimed_marks();
2490
2491 // Mark from roots one level into CMS
2492 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2493 markBitMap());
2494 CLDToOopClosure cld_closure(¬Older, true);
2495
2496 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2497
2498 {
2499 StrongRootsScope srs(1);
2500
2501 gch->gen_process_roots(&srs,
2502 _cmsGen->level(),
2503 true, // younger gens are roots
2504 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2505 should_unload_classes(),
2506 ¬Older,
2507 NULL,
2508 &cld_closure);
2509 }
2510
2511 // Now mark from the roots
2512 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2513 verification_mark_bm(), markBitMap(), verification_mark_stack());
2514 assert(_restart_addr == NULL, "Expected pre-condition");
2515 verification_mark_bm()->iterate(&markFromRootsClosure);
2516 while (_restart_addr != NULL) {
2517 // Deal with stack overflow: by restarting at the indicated
2518 // address.
2519 HeapWord* ra = _restart_addr;
2520 markFromRootsClosure.reset(ra);
2521 _restart_addr = NULL;
2522 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2523 }
2524 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2525 verify_work_stacks_empty();
2526
2527 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2528 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2529
2530 // Marking completed -- now verify that each bit marked in
2531 // verification_mark_bm() is also marked in markBitMap(); flag all
2532 // errors by printing corresponding objects.
2533 VerifyMarkedClosure vcl(markBitMap());
2534 verification_mark_bm()->iterate(&vcl);
2535 assert(!vcl.failed(), "Else verification above should not have succeeded");
2536 }
2537
2538 void ConcurrentMarkSweepGeneration::save_marks() {
2539 // delegate to CMS space
2540 cmsSpace()->save_marks();
2541 for (uint i = 0; i < ParallelGCThreads; i++) {
2542 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2543 }
2544 }
2545
2546 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2547 return cmsSpace()->no_allocs_since_save_marks();
2548 }
2549
2550 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2551 \
2552 void ConcurrentMarkSweepGeneration:: \
2553 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2554 cl->set_generation(this); \
2555 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2556 cl->reset_generation(); \
2557 save_marks(); \
2558 }
2559
2560 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2561
2562 void
2563 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2564 if (freelistLock()->owned_by_self()) {
2565 Generation::oop_iterate(cl);
2566 } else {
2567 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2568 Generation::oop_iterate(cl);
2569 }
2570 }
2571
2572 void
2573 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2574 if (freelistLock()->owned_by_self()) {
2575 Generation::object_iterate(cl);
2576 } else {
2577 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2578 Generation::object_iterate(cl);
2579 }
2580 }
2581
2582 void
2583 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2584 if (freelistLock()->owned_by_self()) {
2585 Generation::safe_object_iterate(cl);
2586 } else {
2587 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2588 Generation::safe_object_iterate(cl);
2589 }
2590 }
2591
2592 void
2593 ConcurrentMarkSweepGeneration::post_compact() {
2594 }
2595
2596 void
2597 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2598 // Fix the linear allocation blocks to look like free blocks.
2599
2600 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2601 // are not called when the heap is verified during universe initialization and
2602 // at vm shutdown.
2603 if (freelistLock()->owned_by_self()) {
2604 cmsSpace()->prepare_for_verify();
2605 } else {
2606 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2607 cmsSpace()->prepare_for_verify();
2608 }
2609 }
2610
2611 void
2612 ConcurrentMarkSweepGeneration::verify() {
2613 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2614 // are not called when the heap is verified during universe initialization and
2615 // at vm shutdown.
2616 if (freelistLock()->owned_by_self()) {
2617 cmsSpace()->verify();
2618 } else {
2619 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2620 cmsSpace()->verify();
2621 }
2622 }
2623
2624 void CMSCollector::verify() {
2625 _cmsGen->verify();
2626 }
2627
2628 #ifndef PRODUCT
2629 bool CMSCollector::overflow_list_is_empty() const {
2630 assert(_num_par_pushes >= 0, "Inconsistency");
2631 if (_overflow_list == NULL) {
2632 assert(_num_par_pushes == 0, "Inconsistency");
2633 }
2634 return _overflow_list == NULL;
2635 }
2636
2637 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2638 // merely consolidate assertion checks that appear to occur together frequently.
2639 void CMSCollector::verify_work_stacks_empty() const {
2640 assert(_markStack.isEmpty(), "Marking stack should be empty");
2641 assert(overflow_list_is_empty(), "Overflow list should be empty");
2642 }
2643
2644 void CMSCollector::verify_overflow_empty() const {
2645 assert(overflow_list_is_empty(), "Overflow list should be empty");
2646 assert(no_preserved_marks(), "No preserved marks");
2647 }
2648 #endif // PRODUCT
2649
2650 // Decide if we want to enable class unloading as part of the
2651 // ensuing concurrent GC cycle. We will collect and
2652 // unload classes if it's the case that:
2653 // (1) an explicit gc request has been made and the flag
2654 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2655 // (2) (a) class unloading is enabled at the command line, and
2656 // (b) old gen is getting really full
2657 // NOTE: Provided there is no change in the state of the heap between
2658 // calls to this method, it should have idempotent results. Moreover,
2659 // its results should be monotonically increasing (i.e. going from 0 to 1,
2660 // but not 1 to 0) between successive calls between which the heap was
2661 // not collected. For the implementation below, it must thus rely on
2662 // the property that concurrent_cycles_since_last_unload()
2663 // will not decrease unless a collection cycle happened and that
2664 // _cmsGen->is_too_full() are
2665 // themselves also monotonic in that sense. See check_monotonicity()
2666 // below.
2667 void CMSCollector::update_should_unload_classes() {
2668 _should_unload_classes = false;
2669 // Condition 1 above
2670 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2671 _should_unload_classes = true;
2672 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2673 // Disjuncts 2.b.(i,ii,iii) above
2674 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2675 CMSClassUnloadingMaxInterval)
2676 || _cmsGen->is_too_full();
2677 }
2678 }
2679
2680 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2681 bool res = should_concurrent_collect();
2682 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2683 return res;
2684 }
2685
2686 void CMSCollector::setup_cms_unloading_and_verification_state() {
2687 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2688 || VerifyBeforeExit;
2689 const int rso = GenCollectedHeap::SO_AllCodeCache;
2690
2691 // We set the proper root for this CMS cycle here.
2692 if (should_unload_classes()) { // Should unload classes this cycle
2693 remove_root_scanning_option(rso); // Shrink the root set appropriately
2694 set_verifying(should_verify); // Set verification state for this cycle
2695 return; // Nothing else needs to be done at this time
2696 }
2697
2698 // Not unloading classes this cycle
2699 assert(!should_unload_classes(), "Inconsistency!");
2700
2701 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2702 // Include symbols, strings and code cache elements to prevent their resurrection.
2703 add_root_scanning_option(rso);
2704 set_verifying(true);
2705 } else if (verifying() && !should_verify) {
2706 // We were verifying, but some verification flags got disabled.
2707 set_verifying(false);
2708 // Exclude symbols, strings and code cache elements from root scanning to
2709 // reduce IM and RM pauses.
2710 remove_root_scanning_option(rso);
2711 }
2712 }
2713
2714
2715 #ifndef PRODUCT
2716 HeapWord* CMSCollector::block_start(const void* p) const {
2717 const HeapWord* addr = (HeapWord*)p;
2718 if (_span.contains(p)) {
2719 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2720 return _cmsGen->cmsSpace()->block_start(p);
2721 }
2722 }
2723 return NULL;
2724 }
2725 #endif
2726
2727 HeapWord*
2728 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2729 bool tlab,
2730 bool parallel) {
2731 CMSSynchronousYieldRequest yr;
2732 assert(!tlab, "Can't deal with TLAB allocation");
2733 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2734 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2735 if (GCExpandToAllocateDelayMillis > 0) {
2736 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2737 }
2738 return have_lock_and_allocate(word_size, tlab);
2739 }
2740
2741 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2742 size_t bytes,
2743 size_t expand_bytes,
2744 CMSExpansionCause::Cause cause)
2745 {
2746
2747 bool success = expand(bytes, expand_bytes);
2748
2749 // remember why we expanded; this information is used
2750 // by shouldConcurrentCollect() when making decisions on whether to start
2751 // a new CMS cycle.
2752 if (success) {
2753 set_expansion_cause(cause);
2754 if (PrintGCDetails && Verbose) {
2755 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2756 CMSExpansionCause::to_string(cause));
2757 }
2758 }
2759 }
2760
2761 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2762 HeapWord* res = NULL;
2763 MutexLocker x(ParGCRareEvent_lock);
2764 while (true) {
2765 // Expansion by some other thread might make alloc OK now:
2766 res = ps->lab.alloc(word_sz);
2767 if (res != NULL) return res;
2768 // If there's not enough expansion space available, give up.
2769 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2770 return NULL;
2771 }
2772 // Otherwise, we try expansion.
2773 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2774 // Now go around the loop and try alloc again;
2775 // A competing par_promote might beat us to the expansion space,
2776 // so we may go around the loop again if promotion fails again.
2777 if (GCExpandToAllocateDelayMillis > 0) {
2778 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2779 }
2780 }
2781 }
2782
2783
2784 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2785 PromotionInfo* promo) {
2786 MutexLocker x(ParGCRareEvent_lock);
2787 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2788 while (true) {
2789 // Expansion by some other thread might make alloc OK now:
2790 if (promo->ensure_spooling_space()) {
2791 assert(promo->has_spooling_space(),
2792 "Post-condition of successful ensure_spooling_space()");
2793 return true;
2794 }
2795 // If there's not enough expansion space available, give up.
2796 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2797 return false;
2798 }
2799 // Otherwise, we try expansion.
2800 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2801 // Now go around the loop and try alloc again;
2802 // A competing allocation might beat us to the expansion space,
2803 // so we may go around the loop again if allocation fails again.
2804 if (GCExpandToAllocateDelayMillis > 0) {
2805 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2806 }
2807 }
2808 }
2809
2810 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2811 // Only shrink if a compaction was done so that all the free space
2812 // in the generation is in a contiguous block at the end.
2813 if (did_compact()) {
2814 CardGeneration::shrink(bytes);
2815 }
2816 }
2817
2818 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2819 assert_locked_or_safepoint(Heap_lock);
2820 }
2821
2822 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2823 assert_locked_or_safepoint(Heap_lock);
2824 assert_lock_strong(freelistLock());
2825 if (PrintGCDetails && Verbose) {
2826 warning("Shrinking of CMS not yet implemented");
2827 }
2828 return;
2829 }
2830
2831
2832 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2833 // phases.
2834 class CMSPhaseAccounting: public StackObj {
2835 public:
2836 CMSPhaseAccounting(CMSCollector *collector,
2837 const char *phase,
2838 const GCId gc_id,
2839 bool print_cr = true);
2840 ~CMSPhaseAccounting();
2841
2842 private:
2843 CMSCollector *_collector;
2844 const char *_phase;
2845 elapsedTimer _wallclock;
2846 bool _print_cr;
2847 const GCId _gc_id;
2848
2849 public:
2850 // Not MT-safe; so do not pass around these StackObj's
2851 // where they may be accessed by other threads.
2852 jlong wallclock_millis() {
2853 assert(_wallclock.is_active(), "Wall clock should not stop");
2854 _wallclock.stop(); // to record time
2855 jlong ret = _wallclock.milliseconds();
2856 _wallclock.start(); // restart
2857 return ret;
2858 }
2859 };
2860
2861 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2862 const char *phase,
2863 const GCId gc_id,
2864 bool print_cr) :
2865 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2866
2867 if (PrintCMSStatistics != 0) {
2868 _collector->resetYields();
2869 }
2870 if (PrintGCDetails) {
2871 gclog_or_tty->gclog_stamp(_gc_id);
2872 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2873 _collector->cmsGen()->short_name(), _phase);
2874 }
2875 _collector->resetTimer();
2876 _wallclock.start();
2877 _collector->startTimer();
2878 }
2879
2880 CMSPhaseAccounting::~CMSPhaseAccounting() {
2881 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2882 _collector->stopTimer();
2883 _wallclock.stop();
2884 if (PrintGCDetails) {
2885 gclog_or_tty->gclog_stamp(_gc_id);
2886 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2887 _collector->cmsGen()->short_name(),
2888 _phase, _collector->timerValue(), _wallclock.seconds());
2889 if (_print_cr) {
2890 gclog_or_tty->cr();
2891 }
2892 if (PrintCMSStatistics != 0) {
2893 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2894 _collector->yields());
2895 }
2896 }
2897 }
2898
2899 // CMS work
2900
2901 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2902 class CMSParMarkTask : public AbstractGangTask {
2903 protected:
2904 CMSCollector* _collector;
2905 uint _n_workers;
2906 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2907 AbstractGangTask(name),
2908 _collector(collector),
2909 _n_workers(n_workers) {}
2910 // Work method in support of parallel rescan ... of young gen spaces
2911 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2912 ContiguousSpace* space,
2913 HeapWord** chunk_array, size_t chunk_top);
2914 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2915 };
2916
2917 // Parallel initial mark task
2918 class CMSParInitialMarkTask: public CMSParMarkTask {
2919 StrongRootsScope* _strong_roots_scope;
2920 public:
2921 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2922 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2923 _strong_roots_scope(strong_roots_scope) {}
2924 void work(uint worker_id);
2925 };
2926
2927 // Checkpoint the roots into this generation from outside
2928 // this generation. [Note this initial checkpoint need only
2929 // be approximate -- we'll do a catch up phase subsequently.]
2930 void CMSCollector::checkpointRootsInitial() {
2931 assert(_collectorState == InitialMarking, "Wrong collector state");
2932 check_correct_thread_executing();
2933 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2934
2935 save_heap_summary();
2936 report_heap_summary(GCWhen::BeforeGC);
2937
2938 ReferenceProcessor* rp = ref_processor();
2939 assert(_restart_addr == NULL, "Control point invariant");
2940 {
2941 // acquire locks for subsequent manipulations
2942 MutexLockerEx x(bitMapLock(),
2943 Mutex::_no_safepoint_check_flag);
2944 checkpointRootsInitialWork();
2945 // enable ("weak") refs discovery
2946 rp->enable_discovery();
2947 _collectorState = Marking;
2948 }
2949 }
2950
2951 void CMSCollector::checkpointRootsInitialWork() {
2952 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2953 assert(_collectorState == InitialMarking, "just checking");
2954
2955 // If there has not been a GC[n-1] since last GC[n] cycle completed,
2956 // precede our marking with a collection of all
2957 // younger generations to keep floating garbage to a minimum.
2958 // XXX: we won't do this for now -- it's an optimization to be done later.
2959
2960 // already have locks
2961 assert_lock_strong(bitMapLock());
2962 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2963
2964 // Setup the verification and class unloading state for this
2965 // CMS collection cycle.
2966 setup_cms_unloading_and_verification_state();
2967
2968 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2969 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
2970
2971 // Reset all the PLAB chunk arrays if necessary.
2972 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2973 reset_survivor_plab_arrays();
2974 }
2975
2976 ResourceMark rm;
2977 HandleMark hm;
2978
2979 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2980 GenCollectedHeap* gch = GenCollectedHeap::heap();
2981
2982 verify_work_stacks_empty();
2983 verify_overflow_empty();
2984
2985 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2986 // Update the saved marks which may affect the root scans.
2987 gch->save_marks();
2988
2989 // weak reference processing has not started yet.
2990 ref_processor()->set_enqueuing_is_done(false);
2991
2992 // Need to remember all newly created CLDs,
2993 // so that we can guarantee that the remark finds them.
2994 ClassLoaderDataGraph::remember_new_clds(true);
2995
2996 // Whenever a CLD is found, it will be claimed before proceeding to mark
2997 // the klasses. The claimed marks need to be cleared before marking starts.
2998 ClassLoaderDataGraph::clear_claimed_marks();
2999
3000 if (CMSPrintEdenSurvivorChunks) {
3001 print_eden_and_survivor_chunk_arrays();
3002 }
3003
3004 {
3005 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3006 if (CMSParallelInitialMarkEnabled) {
3007 // The parallel version.
3008 FlexibleWorkGang* workers = gch->workers();
3009 assert(workers != NULL, "Need parallel worker threads.");
3010 uint n_workers = workers->active_workers();
3011
3012 StrongRootsScope srs(n_workers);
3013
3014 CMSParInitialMarkTask tsk(this, &srs, n_workers);
3015 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3016 if (n_workers > 1) {
3017 workers->run_task(&tsk);
3018 } else {
3019 tsk.work(0);
3020 }
3021 } else {
3022 // The serial version.
3023 CLDToOopClosure cld_closure(¬Older, true);
3024 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3025
3026 StrongRootsScope srs(1);
3027
3028 gch->gen_process_roots(&srs,
3029 _cmsGen->level(),
3030 true, // younger gens are roots
3031 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3032 should_unload_classes(),
3033 ¬Older,
3034 NULL,
3035 &cld_closure);
3036 }
3037 }
3038
3039 // Clear mod-union table; it will be dirtied in the prologue of
3040 // CMS generation per each younger generation collection.
3041
3042 assert(_modUnionTable.isAllClear(),
3043 "Was cleared in most recent final checkpoint phase"
3044 " or no bits are set in the gc_prologue before the start of the next "
3045 "subsequent marking phase.");
3046
3047 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3048
3049 // Save the end of the used_region of the constituent generations
3050 // to be used to limit the extent of sweep in each generation.
3051 save_sweep_limits();
3052 verify_overflow_empty();
3053 }
3054
3055 bool CMSCollector::markFromRoots() {
3056 // we might be tempted to assert that:
3057 // assert(!SafepointSynchronize::is_at_safepoint(),
3058 // "inconsistent argument?");
3059 // However that wouldn't be right, because it's possible that
3060 // a safepoint is indeed in progress as a younger generation
3061 // stop-the-world GC happens even as we mark in this generation.
3062 assert(_collectorState == Marking, "inconsistent state?");
3063 check_correct_thread_executing();
3064 verify_overflow_empty();
3065
3066 // Weak ref discovery note: We may be discovering weak
3067 // refs in this generation concurrent (but interleaved) with
3068 // weak ref discovery by a younger generation collector.
3069
3070 CMSTokenSyncWithLocks ts(true, bitMapLock());
3071 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3072 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3073 bool res = markFromRootsWork();
3074 if (res) {
3075 _collectorState = Precleaning;
3076 } else { // We failed and a foreground collection wants to take over
3077 assert(_foregroundGCIsActive, "internal state inconsistency");
3078 assert(_restart_addr == NULL, "foreground will restart from scratch");
3079 if (PrintGCDetails) {
3080 gclog_or_tty->print_cr("bailing out to foreground collection");
3081 }
3082 }
3083 verify_overflow_empty();
3084 return res;
3085 }
3086
3087 bool CMSCollector::markFromRootsWork() {
3088 // iterate over marked bits in bit map, doing a full scan and mark
3089 // from these roots using the following algorithm:
3090 // . if oop is to the right of the current scan pointer,
3091 // mark corresponding bit (we'll process it later)
3092 // . else (oop is to left of current scan pointer)
3093 // push oop on marking stack
3094 // . drain the marking stack
3095
3096 // Note that when we do a marking step we need to hold the
3097 // bit map lock -- recall that direct allocation (by mutators)
3098 // and promotion (by younger generation collectors) is also
3099 // marking the bit map. [the so-called allocate live policy.]
3100 // Because the implementation of bit map marking is not
3101 // robust wrt simultaneous marking of bits in the same word,
3102 // we need to make sure that there is no such interference
3103 // between concurrent such updates.
3104
3105 // already have locks
3106 assert_lock_strong(bitMapLock());
3107
3108 verify_work_stacks_empty();
3109 verify_overflow_empty();
3110 bool result = false;
3111 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3112 result = do_marking_mt();
3113 } else {
3114 result = do_marking_st();
3115 }
3116 return result;
3117 }
3118
3119 // Forward decl
3120 class CMSConcMarkingTask;
3121
3122 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3123 CMSCollector* _collector;
3124 CMSConcMarkingTask* _task;
3125 public:
3126 virtual void yield();
3127
3128 // "n_threads" is the number of threads to be terminated.
3129 // "queue_set" is a set of work queues of other threads.
3130 // "collector" is the CMS collector associated with this task terminator.
3131 // "yield" indicates whether we need the gang as a whole to yield.
3132 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3133 ParallelTaskTerminator(n_threads, queue_set),
3134 _collector(collector) { }
3135
3136 void set_task(CMSConcMarkingTask* task) {
3137 _task = task;
3138 }
3139 };
3140
3141 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3142 CMSConcMarkingTask* _task;
3143 public:
3144 bool should_exit_termination();
3145 void set_task(CMSConcMarkingTask* task) {
3146 _task = task;
3147 }
3148 };
3149
3150 // MT Concurrent Marking Task
3151 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3152 CMSCollector* _collector;
3153 uint _n_workers; // requested/desired # workers
3154 bool _result;
3155 CompactibleFreeListSpace* _cms_space;
3156 char _pad_front[64]; // padding to ...
3157 HeapWord* _global_finger; // ... avoid sharing cache line
3158 char _pad_back[64];
3159 HeapWord* _restart_addr;
3160
3161 // Exposed here for yielding support
3162 Mutex* const _bit_map_lock;
3163
3164 // The per thread work queues, available here for stealing
3165 OopTaskQueueSet* _task_queues;
3166
3167 // Termination (and yielding) support
3168 CMSConcMarkingTerminator _term;
3169 CMSConcMarkingTerminatorTerminator _term_term;
3170
3171 public:
3172 CMSConcMarkingTask(CMSCollector* collector,
3173 CompactibleFreeListSpace* cms_space,
3174 YieldingFlexibleWorkGang* workers,
3175 OopTaskQueueSet* task_queues):
3176 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3177 _collector(collector),
3178 _cms_space(cms_space),
3179 _n_workers(0), _result(true),
3180 _task_queues(task_queues),
3181 _term(_n_workers, task_queues, _collector),
3182 _bit_map_lock(collector->bitMapLock())
3183 {
3184 _requested_size = _n_workers;
3185 _term.set_task(this);
3186 _term_term.set_task(this);
3187 _restart_addr = _global_finger = _cms_space->bottom();
3188 }
3189
3190
3191 OopTaskQueueSet* task_queues() { return _task_queues; }
3192
3193 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3194
3195 HeapWord** global_finger_addr() { return &_global_finger; }
3196
3197 CMSConcMarkingTerminator* terminator() { return &_term; }
3198
3199 virtual void set_for_termination(uint active_workers) {
3200 terminator()->reset_for_reuse(active_workers);
3201 }
3202
3203 void work(uint worker_id);
3204 bool should_yield() {
3205 return ConcurrentMarkSweepThread::should_yield()
3206 && !_collector->foregroundGCIsActive();
3207 }
3208
3209 virtual void coordinator_yield(); // stuff done by coordinator
3210 bool result() { return _result; }
3211
3212 void reset(HeapWord* ra) {
3213 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3214 _restart_addr = _global_finger = ra;
3215 _term.reset_for_reuse();
3216 }
3217
3218 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3219 OopTaskQueue* work_q);
3220
3221 private:
3222 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3223 void do_work_steal(int i);
3224 void bump_global_finger(HeapWord* f);
3225 };
3226
3227 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3228 assert(_task != NULL, "Error");
3229 return _task->yielding();
3230 // Note that we do not need the disjunct || _task->should_yield() above
3231 // because we want terminating threads to yield only if the task
3232 // is already in the midst of yielding, which happens only after at least one
3233 // thread has yielded.
3234 }
3235
3236 void CMSConcMarkingTerminator::yield() {
3237 if (_task->should_yield()) {
3238 _task->yield();
3239 } else {
3240 ParallelTaskTerminator::yield();
3241 }
3242 }
3243
3244 ////////////////////////////////////////////////////////////////
3245 // Concurrent Marking Algorithm Sketch
3246 ////////////////////////////////////////////////////////////////
3247 // Until all tasks exhausted (both spaces):
3248 // -- claim next available chunk
3249 // -- bump global finger via CAS
3250 // -- find first object that starts in this chunk
3251 // and start scanning bitmap from that position
3252 // -- scan marked objects for oops
3253 // -- CAS-mark target, and if successful:
3254 // . if target oop is above global finger (volatile read)
3255 // nothing to do
3256 // . if target oop is in chunk and above local finger
3257 // then nothing to do
3258 // . else push on work-queue
3259 // -- Deal with possible overflow issues:
3260 // . local work-queue overflow causes stuff to be pushed on
3261 // global (common) overflow queue
3262 // . always first empty local work queue
3263 // . then get a batch of oops from global work queue if any
3264 // . then do work stealing
3265 // -- When all tasks claimed (both spaces)
3266 // and local work queue empty,
3267 // then in a loop do:
3268 // . check global overflow stack; steal a batch of oops and trace
3269 // . try to steal from other threads oif GOS is empty
3270 // . if neither is available, offer termination
3271 // -- Terminate and return result
3272 //
3273 void CMSConcMarkingTask::work(uint worker_id) {
3274 elapsedTimer _timer;
3275 ResourceMark rm;
3276 HandleMark hm;
3277
3278 DEBUG_ONLY(_collector->verify_overflow_empty();)
3279
3280 // Before we begin work, our work queue should be empty
3281 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3282 // Scan the bitmap covering _cms_space, tracing through grey objects.
3283 _timer.start();
3284 do_scan_and_mark(worker_id, _cms_space);
3285 _timer.stop();
3286 if (PrintCMSStatistics != 0) {
3287 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3288 worker_id, _timer.seconds());
3289 // XXX: need xxx/xxx type of notation, two timers
3290 }
3291
3292 // ... do work stealing
3293 _timer.reset();
3294 _timer.start();
3295 do_work_steal(worker_id);
3296 _timer.stop();
3297 if (PrintCMSStatistics != 0) {
3298 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3299 worker_id, _timer.seconds());
3300 // XXX: need xxx/xxx type of notation, two timers
3301 }
3302 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3303 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3304 // Note that under the current task protocol, the
3305 // following assertion is true even of the spaces
3306 // expanded since the completion of the concurrent
3307 // marking. XXX This will likely change under a strict
3308 // ABORT semantics.
3309 // After perm removal the comparison was changed to
3310 // greater than or equal to from strictly greater than.
3311 // Before perm removal the highest address sweep would
3312 // have been at the end of perm gen but now is at the
3313 // end of the tenured gen.
3314 assert(_global_finger >= _cms_space->end(),
3315 "All tasks have been completed");
3316 DEBUG_ONLY(_collector->verify_overflow_empty();)
3317 }
3318
3319 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3320 HeapWord* read = _global_finger;
3321 HeapWord* cur = read;
3322 while (f > read) {
3323 cur = read;
3324 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3325 if (cur == read) {
3326 // our cas succeeded
3327 assert(_global_finger >= f, "protocol consistency");
3328 break;
3329 }
3330 }
3331 }
3332
3333 // This is really inefficient, and should be redone by
3334 // using (not yet available) block-read and -write interfaces to the
3335 // stack and the work_queue. XXX FIX ME !!!
3336 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3337 OopTaskQueue* work_q) {
3338 // Fast lock-free check
3339 if (ovflw_stk->length() == 0) {
3340 return false;
3341 }
3342 assert(work_q->size() == 0, "Shouldn't steal");
3343 MutexLockerEx ml(ovflw_stk->par_lock(),
3344 Mutex::_no_safepoint_check_flag);
3345 // Grab up to 1/4 the size of the work queue
3346 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3347 (size_t)ParGCDesiredObjsFromOverflowList);
3348 num = MIN2(num, ovflw_stk->length());
3349 for (int i = (int) num; i > 0; i--) {
3350 oop cur = ovflw_stk->pop();
3351 assert(cur != NULL, "Counted wrong?");
3352 work_q->push(cur);
3353 }
3354 return num > 0;
3355 }
3356
3357 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3358 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3359 int n_tasks = pst->n_tasks();
3360 // We allow that there may be no tasks to do here because
3361 // we are restarting after a stack overflow.
3362 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3363 uint nth_task = 0;
3364
3365 HeapWord* aligned_start = sp->bottom();
3366 if (sp->used_region().contains(_restart_addr)) {
3367 // Align down to a card boundary for the start of 0th task
3368 // for this space.
3369 aligned_start =
3370 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3371 CardTableModRefBS::card_size);
3372 }
3373
3374 size_t chunk_size = sp->marking_task_size();
3375 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3376 // Having claimed the nth task in this space,
3377 // compute the chunk that it corresponds to:
3378 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3379 aligned_start + (nth_task+1)*chunk_size);
3380 // Try and bump the global finger via a CAS;
3381 // note that we need to do the global finger bump
3382 // _before_ taking the intersection below, because
3383 // the task corresponding to that region will be
3384 // deemed done even if the used_region() expands
3385 // because of allocation -- as it almost certainly will
3386 // during start-up while the threads yield in the
3387 // closure below.
3388 HeapWord* finger = span.end();
3389 bump_global_finger(finger); // atomically
3390 // There are null tasks here corresponding to chunks
3391 // beyond the "top" address of the space.
3392 span = span.intersection(sp->used_region());
3393 if (!span.is_empty()) { // Non-null task
3394 HeapWord* prev_obj;
3395 assert(!span.contains(_restart_addr) || nth_task == 0,
3396 "Inconsistency");
3397 if (nth_task == 0) {
3398 // For the 0th task, we'll not need to compute a block_start.
3399 if (span.contains(_restart_addr)) {
3400 // In the case of a restart because of stack overflow,
3401 // we might additionally skip a chunk prefix.
3402 prev_obj = _restart_addr;
3403 } else {
3404 prev_obj = span.start();
3405 }
3406 } else {
3407 // We want to skip the first object because
3408 // the protocol is to scan any object in its entirety
3409 // that _starts_ in this span; a fortiori, any
3410 // object starting in an earlier span is scanned
3411 // as part of an earlier claimed task.
3412 // Below we use the "careful" version of block_start
3413 // so we do not try to navigate uninitialized objects.
3414 prev_obj = sp->block_start_careful(span.start());
3415 // Below we use a variant of block_size that uses the
3416 // Printezis bits to avoid waiting for allocated
3417 // objects to become initialized/parsable.
3418 while (prev_obj < span.start()) {
3419 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3420 if (sz > 0) {
3421 prev_obj += sz;
3422 } else {
3423 // In this case we may end up doing a bit of redundant
3424 // scanning, but that appears unavoidable, short of
3425 // locking the free list locks; see bug 6324141.
3426 break;
3427 }
3428 }
3429 }
3430 if (prev_obj < span.end()) {
3431 MemRegion my_span = MemRegion(prev_obj, span.end());
3432 // Do the marking work within a non-empty span --
3433 // the last argument to the constructor indicates whether the
3434 // iteration should be incremental with periodic yields.
3435 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3436 &_collector->_markBitMap,
3437 work_queue(i),
3438 &_collector->_markStack);
3439 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3440 } // else nothing to do for this task
3441 } // else nothing to do for this task
3442 }
3443 // We'd be tempted to assert here that since there are no
3444 // more tasks left to claim in this space, the global_finger
3445 // must exceed space->top() and a fortiori space->end(). However,
3446 // that would not quite be correct because the bumping of
3447 // global_finger occurs strictly after the claiming of a task,
3448 // so by the time we reach here the global finger may not yet
3449 // have been bumped up by the thread that claimed the last
3450 // task.
3451 pst->all_tasks_completed();
3452 }
3453
3454 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3455 private:
3456 CMSCollector* _collector;
3457 CMSConcMarkingTask* _task;
3458 MemRegion _span;
3459 CMSBitMap* _bit_map;
3460 CMSMarkStack* _overflow_stack;
3461 OopTaskQueue* _work_queue;
3462 protected:
3463 DO_OOP_WORK_DEFN
3464 public:
3465 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3466 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3467 MetadataAwareOopClosure(collector->ref_processor()),
3468 _collector(collector),
3469 _task(task),
3470 _span(collector->_span),
3471 _work_queue(work_queue),
3472 _bit_map(bit_map),
3473 _overflow_stack(overflow_stack)
3474 { }
3475 virtual void do_oop(oop* p);
3476 virtual void do_oop(narrowOop* p);
3477
3478 void trim_queue(size_t max);
3479 void handle_stack_overflow(HeapWord* lost);
3480 void do_yield_check() {
3481 if (_task->should_yield()) {
3482 _task->yield();
3483 }
3484 }
3485 };
3486
3487 // Grey object scanning during work stealing phase --
3488 // the salient assumption here is that any references
3489 // that are in these stolen objects being scanned must
3490 // already have been initialized (else they would not have
3491 // been published), so we do not need to check for
3492 // uninitialized objects before pushing here.
3493 void Par_ConcMarkingClosure::do_oop(oop obj) {
3494 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3495 HeapWord* addr = (HeapWord*)obj;
3496 // Check if oop points into the CMS generation
3497 // and is not marked
3498 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3499 // a white object ...
3500 // If we manage to "claim" the object, by being the
3501 // first thread to mark it, then we push it on our
3502 // marking stack
3503 if (_bit_map->par_mark(addr)) { // ... now grey
3504 // push on work queue (grey set)
3505 bool simulate_overflow = false;
3506 NOT_PRODUCT(
3507 if (CMSMarkStackOverflowALot &&
3508 _collector->simulate_overflow()) {
3509 // simulate a stack overflow
3510 simulate_overflow = true;
3511 }
3512 )
3513 if (simulate_overflow ||
3514 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3515 // stack overflow
3516 if (PrintCMSStatistics != 0) {
3517 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3518 SIZE_FORMAT, _overflow_stack->capacity());
3519 }
3520 // We cannot assert that the overflow stack is full because
3521 // it may have been emptied since.
3522 assert(simulate_overflow ||
3523 _work_queue->size() == _work_queue->max_elems(),
3524 "Else push should have succeeded");
3525 handle_stack_overflow(addr);
3526 }
3527 } // Else, some other thread got there first
3528 do_yield_check();
3529 }
3530 }
3531
3532 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3533 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3534
3535 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3536 while (_work_queue->size() > max) {
3537 oop new_oop;
3538 if (_work_queue->pop_local(new_oop)) {
3539 assert(new_oop->is_oop(), "Should be an oop");
3540 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3541 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3542 new_oop->oop_iterate(this); // do_oop() above
3543 do_yield_check();
3544 }
3545 }
3546 }
3547
3548 // Upon stack overflow, we discard (part of) the stack,
3549 // remembering the least address amongst those discarded
3550 // in CMSCollector's _restart_address.
3551 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3552 // We need to do this under a mutex to prevent other
3553 // workers from interfering with the work done below.
3554 MutexLockerEx ml(_overflow_stack->par_lock(),
3555 Mutex::_no_safepoint_check_flag);
3556 // Remember the least grey address discarded
3557 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3558 _collector->lower_restart_addr(ra);
3559 _overflow_stack->reset(); // discard stack contents
3560 _overflow_stack->expand(); // expand the stack if possible
3561 }
3562
3563
3564 void CMSConcMarkingTask::do_work_steal(int i) {
3565 OopTaskQueue* work_q = work_queue(i);
3566 oop obj_to_scan;
3567 CMSBitMap* bm = &(_collector->_markBitMap);
3568 CMSMarkStack* ovflw = &(_collector->_markStack);
3569 int* seed = _collector->hash_seed(i);
3570 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3571 while (true) {
3572 cl.trim_queue(0);
3573 assert(work_q->size() == 0, "Should have been emptied above");
3574 if (get_work_from_overflow_stack(ovflw, work_q)) {
3575 // Can't assert below because the work obtained from the
3576 // overflow stack may already have been stolen from us.
3577 // assert(work_q->size() > 0, "Work from overflow stack");
3578 continue;
3579 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3580 assert(obj_to_scan->is_oop(), "Should be an oop");
3581 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3582 obj_to_scan->oop_iterate(&cl);
3583 } else if (terminator()->offer_termination(&_term_term)) {
3584 assert(work_q->size() == 0, "Impossible!");
3585 break;
3586 } else if (yielding() || should_yield()) {
3587 yield();
3588 }
3589 }
3590 }
3591
3592 // This is run by the CMS (coordinator) thread.
3593 void CMSConcMarkingTask::coordinator_yield() {
3594 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3595 "CMS thread should hold CMS token");
3596 // First give up the locks, then yield, then re-lock
3597 // We should probably use a constructor/destructor idiom to
3598 // do this unlock/lock or modify the MutexUnlocker class to
3599 // serve our purpose. XXX
3600 assert_lock_strong(_bit_map_lock);
3601 _bit_map_lock->unlock();
3602 ConcurrentMarkSweepThread::desynchronize(true);
3603 _collector->stopTimer();
3604 if (PrintCMSStatistics != 0) {
3605 _collector->incrementYields();
3606 }
3607
3608 // It is possible for whichever thread initiated the yield request
3609 // not to get a chance to wake up and take the bitmap lock between
3610 // this thread releasing it and reacquiring it. So, while the
3611 // should_yield() flag is on, let's sleep for a bit to give the
3612 // other thread a chance to wake up. The limit imposed on the number
3613 // of iterations is defensive, to avoid any unforseen circumstances
3614 // putting us into an infinite loop. Since it's always been this
3615 // (coordinator_yield()) method that was observed to cause the
3616 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3617 // which is by default non-zero. For the other seven methods that
3618 // also perform the yield operation, as are using a different
3619 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3620 // can enable the sleeping for those methods too, if necessary.
3621 // See 6442774.
3622 //
3623 // We really need to reconsider the synchronization between the GC
3624 // thread and the yield-requesting threads in the future and we
3625 // should really use wait/notify, which is the recommended
3626 // way of doing this type of interaction. Additionally, we should
3627 // consolidate the eight methods that do the yield operation and they
3628 // are almost identical into one for better maintainability and
3629 // readability. See 6445193.
3630 //
3631 // Tony 2006.06.29
3632 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3633 ConcurrentMarkSweepThread::should_yield() &&
3634 !CMSCollector::foregroundGCIsActive(); ++i) {
3635 os::sleep(Thread::current(), 1, false);
3636 }
3637
3638 ConcurrentMarkSweepThread::synchronize(true);
3639 _bit_map_lock->lock_without_safepoint_check();
3640 _collector->startTimer();
3641 }
3642
3643 bool CMSCollector::do_marking_mt() {
3644 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3645 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3646 conc_workers()->active_workers(),
3647 Threads::number_of_non_daemon_threads());
3648 conc_workers()->set_active_workers(num_workers);
3649
3650 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3651
3652 CMSConcMarkingTask tsk(this,
3653 cms_space,
3654 conc_workers(),
3655 task_queues());
3656
3657 // Since the actual number of workers we get may be different
3658 // from the number we requested above, do we need to do anything different
3659 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3660 // class?? XXX
3661 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3662
3663 // Refs discovery is already non-atomic.
3664 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3665 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3666 conc_workers()->start_task(&tsk);
3667 while (tsk.yielded()) {
3668 tsk.coordinator_yield();
3669 conc_workers()->continue_task(&tsk);
3670 }
3671 // If the task was aborted, _restart_addr will be non-NULL
3672 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3673 while (_restart_addr != NULL) {
3674 // XXX For now we do not make use of ABORTED state and have not
3675 // yet implemented the right abort semantics (even in the original
3676 // single-threaded CMS case). That needs some more investigation
3677 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3678 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3679 // If _restart_addr is non-NULL, a marking stack overflow
3680 // occurred; we need to do a fresh marking iteration from the
3681 // indicated restart address.
3682 if (_foregroundGCIsActive) {
3683 // We may be running into repeated stack overflows, having
3684 // reached the limit of the stack size, while making very
3685 // slow forward progress. It may be best to bail out and
3686 // let the foreground collector do its job.
3687 // Clear _restart_addr, so that foreground GC
3688 // works from scratch. This avoids the headache of
3689 // a "rescan" which would otherwise be needed because
3690 // of the dirty mod union table & card table.
3691 _restart_addr = NULL;
3692 return false;
3693 }
3694 // Adjust the task to restart from _restart_addr
3695 tsk.reset(_restart_addr);
3696 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3697 _restart_addr);
3698 _restart_addr = NULL;
3699 // Get the workers going again
3700 conc_workers()->start_task(&tsk);
3701 while (tsk.yielded()) {
3702 tsk.coordinator_yield();
3703 conc_workers()->continue_task(&tsk);
3704 }
3705 }
3706 assert(tsk.completed(), "Inconsistency");
3707 assert(tsk.result() == true, "Inconsistency");
3708 return true;
3709 }
3710
3711 bool CMSCollector::do_marking_st() {
3712 ResourceMark rm;
3713 HandleMark hm;
3714
3715 // Temporarily make refs discovery single threaded (non-MT)
3716 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3717 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3718 &_markStack, CMSYield);
3719 // the last argument to iterate indicates whether the iteration
3720 // should be incremental with periodic yields.
3721 _markBitMap.iterate(&markFromRootsClosure);
3722 // If _restart_addr is non-NULL, a marking stack overflow
3723 // occurred; we need to do a fresh iteration from the
3724 // indicated restart address.
3725 while (_restart_addr != NULL) {
3726 if (_foregroundGCIsActive) {
3727 // We may be running into repeated stack overflows, having
3728 // reached the limit of the stack size, while making very
3729 // slow forward progress. It may be best to bail out and
3730 // let the foreground collector do its job.
3731 // Clear _restart_addr, so that foreground GC
3732 // works from scratch. This avoids the headache of
3733 // a "rescan" which would otherwise be needed because
3734 // of the dirty mod union table & card table.
3735 _restart_addr = NULL;
3736 return false; // indicating failure to complete marking
3737 }
3738 // Deal with stack overflow:
3739 // we restart marking from _restart_addr
3740 HeapWord* ra = _restart_addr;
3741 markFromRootsClosure.reset(ra);
3742 _restart_addr = NULL;
3743 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3744 }
3745 return true;
3746 }
3747
3748 void CMSCollector::preclean() {
3749 check_correct_thread_executing();
3750 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3751 verify_work_stacks_empty();
3752 verify_overflow_empty();
3753 _abort_preclean = false;
3754 if (CMSPrecleaningEnabled) {
3755 if (!CMSEdenChunksRecordAlways) {
3756 _eden_chunk_index = 0;
3757 }
3758 size_t used = get_eden_used();
3759 size_t capacity = get_eden_capacity();
3760 // Don't start sampling unless we will get sufficiently
3761 // many samples.
3762 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3763 * CMSScheduleRemarkEdenPenetration)) {
3764 _start_sampling = true;
3765 } else {
3766 _start_sampling = false;
3767 }
3768 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3769 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3770 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3771 }
3772 CMSTokenSync x(true); // is cms thread
3773 if (CMSPrecleaningEnabled) {
3774 sample_eden();
3775 _collectorState = AbortablePreclean;
3776 } else {
3777 _collectorState = FinalMarking;
3778 }
3779 verify_work_stacks_empty();
3780 verify_overflow_empty();
3781 }
3782
3783 // Try and schedule the remark such that young gen
3784 // occupancy is CMSScheduleRemarkEdenPenetration %.
3785 void CMSCollector::abortable_preclean() {
3786 check_correct_thread_executing();
3787 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3788 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3789
3790 // If Eden's current occupancy is below this threshold,
3791 // immediately schedule the remark; else preclean
3792 // past the next scavenge in an effort to
3793 // schedule the pause as described above. By choosing
3794 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3795 // we will never do an actual abortable preclean cycle.
3796 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3797 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3798 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3799 // We need more smarts in the abortable preclean
3800 // loop below to deal with cases where allocation
3801 // in young gen is very very slow, and our precleaning
3802 // is running a losing race against a horde of
3803 // mutators intent on flooding us with CMS updates
3804 // (dirty cards).
3805 // One, admittedly dumb, strategy is to give up
3806 // after a certain number of abortable precleaning loops
3807 // or after a certain maximum time. We want to make
3808 // this smarter in the next iteration.
3809 // XXX FIX ME!!! YSR
3810 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3811 while (!(should_abort_preclean() ||
3812 ConcurrentMarkSweepThread::should_terminate())) {
3813 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3814 cumworkdone += workdone;
3815 loops++;
3816 // Voluntarily terminate abortable preclean phase if we have
3817 // been at it for too long.
3818 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3819 loops >= CMSMaxAbortablePrecleanLoops) {
3820 if (PrintGCDetails) {
3821 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3822 }
3823 break;
3824 }
3825 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3826 if (PrintGCDetails) {
3827 gclog_or_tty->print(" CMS: abort preclean due to time ");
3828 }
3829 break;
3830 }
3831 // If we are doing little work each iteration, we should
3832 // take a short break.
3833 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3834 // Sleep for some time, waiting for work to accumulate
3835 stopTimer();
3836 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3837 startTimer();
3838 waited++;
3839 }
3840 }
3841 if (PrintCMSStatistics > 0) {
3842 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3843 loops, waited, cumworkdone);
3844 }
3845 }
3846 CMSTokenSync x(true); // is cms thread
3847 if (_collectorState != Idling) {
3848 assert(_collectorState == AbortablePreclean,
3849 "Spontaneous state transition?");
3850 _collectorState = FinalMarking;
3851 } // Else, a foreground collection completed this CMS cycle.
3852 return;
3853 }
3854
3855 // Respond to an Eden sampling opportunity
3856 void CMSCollector::sample_eden() {
3857 // Make sure a young gc cannot sneak in between our
3858 // reading and recording of a sample.
3859 assert(Thread::current()->is_ConcurrentGC_thread(),
3860 "Only the cms thread may collect Eden samples");
3861 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3862 "Should collect samples while holding CMS token");
3863 if (!_start_sampling) {
3864 return;
3865 }
3866 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3867 // is populated by the young generation.
3868 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3869 if (_eden_chunk_index < _eden_chunk_capacity) {
3870 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3871 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3872 "Unexpected state of Eden");
3873 // We'd like to check that what we just sampled is an oop-start address;
3874 // however, we cannot do that here since the object may not yet have been
3875 // initialized. So we'll instead do the check when we _use_ this sample
3876 // later.
3877 if (_eden_chunk_index == 0 ||
3878 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3879 _eden_chunk_array[_eden_chunk_index-1])
3880 >= CMSSamplingGrain)) {
3881 _eden_chunk_index++; // commit sample
3882 }
3883 }
3884 }
3885 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3886 size_t used = get_eden_used();
3887 size_t capacity = get_eden_capacity();
3888 assert(used <= capacity, "Unexpected state of Eden");
3889 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3890 _abort_preclean = true;
3891 }
3892 }
3893 }
3894
3895
3896 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3897 assert(_collectorState == Precleaning ||
3898 _collectorState == AbortablePreclean, "incorrect state");
3899 ResourceMark rm;
3900 HandleMark hm;
3901
3902 // Precleaning is currently not MT but the reference processor
3903 // may be set for MT. Disable it temporarily here.
3904 ReferenceProcessor* rp = ref_processor();
3905 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3906
3907 // Do one pass of scrubbing the discovered reference lists
3908 // to remove any reference objects with strongly-reachable
3909 // referents.
3910 if (clean_refs) {
3911 CMSPrecleanRefsYieldClosure yield_cl(this);
3912 assert(rp->span().equals(_span), "Spans should be equal");
3913 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3914 &_markStack, true /* preclean */);
3915 CMSDrainMarkingStackClosure complete_trace(this,
3916 _span, &_markBitMap, &_markStack,
3917 &keep_alive, true /* preclean */);
3918
3919 // We don't want this step to interfere with a young
3920 // collection because we don't want to take CPU
3921 // or memory bandwidth away from the young GC threads
3922 // (which may be as many as there are CPUs).
3923 // Note that we don't need to protect ourselves from
3924 // interference with mutators because they can't
3925 // manipulate the discovered reference lists nor affect
3926 // the computed reachability of the referents, the
3927 // only properties manipulated by the precleaning
3928 // of these reference lists.
3929 stopTimer();
3930 CMSTokenSyncWithLocks x(true /* is cms thread */,
3931 bitMapLock());
3932 startTimer();
3933 sample_eden();
3934
3935 // The following will yield to allow foreground
3936 // collection to proceed promptly. XXX YSR:
3937 // The code in this method may need further
3938 // tweaking for better performance and some restructuring
3939 // for cleaner interfaces.
3940 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3941 rp->preclean_discovered_references(
3942 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3943 gc_timer, _gc_tracer_cm->gc_id());
3944 }
3945
3946 if (clean_survivor) { // preclean the active survivor space(s)
3947 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3948 &_markBitMap, &_modUnionTable,
3949 &_markStack, true /* precleaning phase */);
3950 stopTimer();
3951 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3952 bitMapLock());
3953 startTimer();
3954 unsigned int before_count =
3955 GenCollectedHeap::heap()->total_collections();
3956 SurvivorSpacePrecleanClosure
3957 sss_cl(this, _span, &_markBitMap, &_markStack,
3958 &pam_cl, before_count, CMSYield);
3959 _young_gen->from()->object_iterate_careful(&sss_cl);
3960 _young_gen->to()->object_iterate_careful(&sss_cl);
3961 }
3962 MarkRefsIntoAndScanClosure
3963 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3964 &_markStack, this, CMSYield,
3965 true /* precleaning phase */);
3966 // CAUTION: The following closure has persistent state that may need to
3967 // be reset upon a decrease in the sequence of addresses it
3968 // processes.
3969 ScanMarkedObjectsAgainCarefullyClosure
3970 smoac_cl(this, _span,
3971 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3972
3973 // Preclean dirty cards in ModUnionTable and CardTable using
3974 // appropriate convergence criterion;
3975 // repeat CMSPrecleanIter times unless we find that
3976 // we are losing.
3977 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3978 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3979 "Bad convergence multiplier");
3980 assert(CMSPrecleanThreshold >= 100,
3981 "Unreasonably low CMSPrecleanThreshold");
3982
3983 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3984 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3985 numIter < CMSPrecleanIter;
3986 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3987 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3988 if (Verbose && PrintGCDetails) {
3989 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3990 }
3991 // Either there are very few dirty cards, so re-mark
3992 // pause will be small anyway, or our pre-cleaning isn't
3993 // that much faster than the rate at which cards are being
3994 // dirtied, so we might as well stop and re-mark since
3995 // precleaning won't improve our re-mark time by much.
3996 if (curNumCards <= CMSPrecleanThreshold ||
3997 (numIter > 0 &&
3998 (curNumCards * CMSPrecleanDenominator >
3999 lastNumCards * CMSPrecleanNumerator))) {
4000 numIter++;
4001 cumNumCards += curNumCards;
4002 break;
4003 }
4004 }
4005
4006 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4007
4008 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4009 cumNumCards += curNumCards;
4010 if (PrintGCDetails && PrintCMSStatistics != 0) {
4011 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4012 curNumCards, cumNumCards, numIter);
4013 }
4014 return cumNumCards; // as a measure of useful work done
4015 }
4016
4017 // PRECLEANING NOTES:
4018 // Precleaning involves:
4019 // . reading the bits of the modUnionTable and clearing the set bits.
4020 // . For the cards corresponding to the set bits, we scan the
4021 // objects on those cards. This means we need the free_list_lock
4022 // so that we can safely iterate over the CMS space when scanning
4023 // for oops.
4024 // . When we scan the objects, we'll be both reading and setting
4025 // marks in the marking bit map, so we'll need the marking bit map.
4026 // . For protecting _collector_state transitions, we take the CGC_lock.
4027 // Note that any races in the reading of of card table entries by the
4028 // CMS thread on the one hand and the clearing of those entries by the
4029 // VM thread or the setting of those entries by the mutator threads on the
4030 // other are quite benign. However, for efficiency it makes sense to keep
4031 // the VM thread from racing with the CMS thread while the latter is
4032 // dirty card info to the modUnionTable. We therefore also use the
4033 // CGC_lock to protect the reading of the card table and the mod union
4034 // table by the CM thread.
4035 // . We run concurrently with mutator updates, so scanning
4036 // needs to be done carefully -- we should not try to scan
4037 // potentially uninitialized objects.
4038 //
4039 // Locking strategy: While holding the CGC_lock, we scan over and
4040 // reset a maximal dirty range of the mod union / card tables, then lock
4041 // the free_list_lock and bitmap lock to do a full marking, then
4042 // release these locks; and repeat the cycle. This allows for a
4043 // certain amount of fairness in the sharing of these locks between
4044 // the CMS collector on the one hand, and the VM thread and the
4045 // mutators on the other.
4046
4047 // NOTE: preclean_mod_union_table() and preclean_card_table()
4048 // further below are largely identical; if you need to modify
4049 // one of these methods, please check the other method too.
4050
4051 size_t CMSCollector::preclean_mod_union_table(
4052 ConcurrentMarkSweepGeneration* gen,
4053 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4054 verify_work_stacks_empty();
4055 verify_overflow_empty();
4056
4057 // strategy: starting with the first card, accumulate contiguous
4058 // ranges of dirty cards; clear these cards, then scan the region
4059 // covered by these cards.
4060
4061 // Since all of the MUT is committed ahead, we can just use
4062 // that, in case the generations expand while we are precleaning.
4063 // It might also be fine to just use the committed part of the
4064 // generation, but we might potentially miss cards when the
4065 // generation is rapidly expanding while we are in the midst
4066 // of precleaning.
4067 HeapWord* startAddr = gen->reserved().start();
4068 HeapWord* endAddr = gen->reserved().end();
4069
4070 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4071
4072 size_t numDirtyCards, cumNumDirtyCards;
4073 HeapWord *nextAddr, *lastAddr;
4074 for (cumNumDirtyCards = numDirtyCards = 0,
4075 nextAddr = lastAddr = startAddr;
4076 nextAddr < endAddr;
4077 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4078
4079 ResourceMark rm;
4080 HandleMark hm;
4081
4082 MemRegion dirtyRegion;
4083 {
4084 stopTimer();
4085 // Potential yield point
4086 CMSTokenSync ts(true);
4087 startTimer();
4088 sample_eden();
4089 // Get dirty region starting at nextOffset (inclusive),
4090 // simultaneously clearing it.
4091 dirtyRegion =
4092 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4093 assert(dirtyRegion.start() >= nextAddr,
4094 "returned region inconsistent?");
4095 }
4096 // Remember where the next search should begin.
4097 // The returned region (if non-empty) is a right open interval,
4098 // so lastOffset is obtained from the right end of that
4099 // interval.
4100 lastAddr = dirtyRegion.end();
4101 // Should do something more transparent and less hacky XXX
4102 numDirtyCards =
4103 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4104
4105 // We'll scan the cards in the dirty region (with periodic
4106 // yields for foreground GC as needed).
4107 if (!dirtyRegion.is_empty()) {
4108 assert(numDirtyCards > 0, "consistency check");
4109 HeapWord* stop_point = NULL;
4110 stopTimer();
4111 // Potential yield point
4112 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4113 bitMapLock());
4114 startTimer();
4115 {
4116 verify_work_stacks_empty();
4117 verify_overflow_empty();
4118 sample_eden();
4119 stop_point =
4120 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4121 }
4122 if (stop_point != NULL) {
4123 // The careful iteration stopped early either because it found an
4124 // uninitialized object, or because we were in the midst of an
4125 // "abortable preclean", which should now be aborted. Redirty
4126 // the bits corresponding to the partially-scanned or unscanned
4127 // cards. We'll either restart at the next block boundary or
4128 // abort the preclean.
4129 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4130 "Should only be AbortablePreclean.");
4131 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4132 if (should_abort_preclean()) {
4133 break; // out of preclean loop
4134 } else {
4135 // Compute the next address at which preclean should pick up;
4136 // might need bitMapLock in order to read P-bits.
4137 lastAddr = next_card_start_after_block(stop_point);
4138 }
4139 }
4140 } else {
4141 assert(lastAddr == endAddr, "consistency check");
4142 assert(numDirtyCards == 0, "consistency check");
4143 break;
4144 }
4145 }
4146 verify_work_stacks_empty();
4147 verify_overflow_empty();
4148 return cumNumDirtyCards;
4149 }
4150
4151 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4152 // below are largely identical; if you need to modify
4153 // one of these methods, please check the other method too.
4154
4155 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4156 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4157 // strategy: it's similar to precleamModUnionTable above, in that
4158 // we accumulate contiguous ranges of dirty cards, mark these cards
4159 // precleaned, then scan the region covered by these cards.
4160 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4161 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4162
4163 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4164
4165 size_t numDirtyCards, cumNumDirtyCards;
4166 HeapWord *lastAddr, *nextAddr;
4167
4168 for (cumNumDirtyCards = numDirtyCards = 0,
4169 nextAddr = lastAddr = startAddr;
4170 nextAddr < endAddr;
4171 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4172
4173 ResourceMark rm;
4174 HandleMark hm;
4175
4176 MemRegion dirtyRegion;
4177 {
4178 // See comments in "Precleaning notes" above on why we
4179 // do this locking. XXX Could the locking overheads be
4180 // too high when dirty cards are sparse? [I don't think so.]
4181 stopTimer();
4182 CMSTokenSync x(true); // is cms thread
4183 startTimer();
4184 sample_eden();
4185 // Get and clear dirty region from card table
4186 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4187 MemRegion(nextAddr, endAddr),
4188 true,
4189 CardTableModRefBS::precleaned_card_val());
4190
4191 assert(dirtyRegion.start() >= nextAddr,
4192 "returned region inconsistent?");
4193 }
4194 lastAddr = dirtyRegion.end();
4195 numDirtyCards =
4196 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4197
4198 if (!dirtyRegion.is_empty()) {
4199 stopTimer();
4200 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4201 startTimer();
4202 sample_eden();
4203 verify_work_stacks_empty();
4204 verify_overflow_empty();
4205 HeapWord* stop_point =
4206 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4207 if (stop_point != NULL) {
4208 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4209 "Should only be AbortablePreclean.");
4210 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4211 if (should_abort_preclean()) {
4212 break; // out of preclean loop
4213 } else {
4214 // Compute the next address at which preclean should pick up.
4215 lastAddr = next_card_start_after_block(stop_point);
4216 }
4217 }
4218 } else {
4219 break;
4220 }
4221 }
4222 verify_work_stacks_empty();
4223 verify_overflow_empty();
4224 return cumNumDirtyCards;
4225 }
4226
4227 class PrecleanKlassClosure : public KlassClosure {
4228 KlassToOopClosure _cm_klass_closure;
4229 public:
4230 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4231 void do_klass(Klass* k) {
4232 if (k->has_accumulated_modified_oops()) {
4233 k->clear_accumulated_modified_oops();
4234
4235 _cm_klass_closure.do_klass(k);
4236 }
4237 }
4238 };
4239
4240 // The freelist lock is needed to prevent asserts, is it really needed?
4241 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4242
4243 cl->set_freelistLock(freelistLock);
4244
4245 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4246
4247 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4248 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4249 PrecleanKlassClosure preclean_klass_closure(cl);
4250 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4251
4252 verify_work_stacks_empty();
4253 verify_overflow_empty();
4254 }
4255
4256 void CMSCollector::checkpointRootsFinal() {
4257 assert(_collectorState == FinalMarking, "incorrect state transition?");
4258 check_correct_thread_executing();
4259 // world is stopped at this checkpoint
4260 assert(SafepointSynchronize::is_at_safepoint(),
4261 "world should be stopped");
4262 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4263
4264 verify_work_stacks_empty();
4265 verify_overflow_empty();
4266
4267 if (PrintGCDetails) {
4268 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4269 _young_gen->used() / K,
4270 _young_gen->capacity() / K);
4271 }
4272 {
4273 if (CMSScavengeBeforeRemark) {
4274 GenCollectedHeap* gch = GenCollectedHeap::heap();
4275 // Temporarily set flag to false, GCH->do_collection will
4276 // expect it to be false and set to true
4277 FlagSetting fl(gch->_is_gc_active, false);
4278 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4279 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4280 int level = _cmsGen->level() - 1;
4281 if (level >= 0) {
4282 gch->do_collection(true, // full (i.e. force, see below)
4283 false, // !clear_all_soft_refs
4284 0, // size
4285 false, // is_tlab
4286 level // max_level
4287 );
4288 }
4289 }
4290 FreelistLocker x(this);
4291 MutexLockerEx y(bitMapLock(),
4292 Mutex::_no_safepoint_check_flag);
4293 checkpointRootsFinalWork();
4294 }
4295 verify_work_stacks_empty();
4296 verify_overflow_empty();
4297 }
4298
4299 void CMSCollector::checkpointRootsFinalWork() {
4300 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4301
4302 assert(haveFreelistLocks(), "must have free list locks");
4303 assert_lock_strong(bitMapLock());
4304
4305 ResourceMark rm;
4306 HandleMark hm;
4307
4308 GenCollectedHeap* gch = GenCollectedHeap::heap();
4309
4310 if (should_unload_classes()) {
4311 CodeCache::gc_prologue();
4312 }
4313 assert(haveFreelistLocks(), "must have free list locks");
4314 assert_lock_strong(bitMapLock());
4315
4316 // We might assume that we need not fill TLAB's when
4317 // CMSScavengeBeforeRemark is set, because we may have just done
4318 // a scavenge which would have filled all TLAB's -- and besides
4319 // Eden would be empty. This however may not always be the case --
4320 // for instance although we asked for a scavenge, it may not have
4321 // happened because of a JNI critical section. We probably need
4322 // a policy for deciding whether we can in that case wait until
4323 // the critical section releases and then do the remark following
4324 // the scavenge, and skip it here. In the absence of that policy,
4325 // or of an indication of whether the scavenge did indeed occur,
4326 // we cannot rely on TLAB's having been filled and must do
4327 // so here just in case a scavenge did not happen.
4328 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4329 // Update the saved marks which may affect the root scans.
4330 gch->save_marks();
4331
4332 if (CMSPrintEdenSurvivorChunks) {
4333 print_eden_and_survivor_chunk_arrays();
4334 }
4335
4336 {
4337 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4338
4339 // Note on the role of the mod union table:
4340 // Since the marker in "markFromRoots" marks concurrently with
4341 // mutators, it is possible for some reachable objects not to have been
4342 // scanned. For instance, an only reference to an object A was
4343 // placed in object B after the marker scanned B. Unless B is rescanned,
4344 // A would be collected. Such updates to references in marked objects
4345 // are detected via the mod union table which is the set of all cards
4346 // dirtied since the first checkpoint in this GC cycle and prior to
4347 // the most recent young generation GC, minus those cleaned up by the
4348 // concurrent precleaning.
4349 if (CMSParallelRemarkEnabled) {
4350 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4351 do_remark_parallel();
4352 } else {
4353 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4354 _gc_timer_cm, _gc_tracer_cm->gc_id());
4355 do_remark_non_parallel();
4356 }
4357 }
4358 verify_work_stacks_empty();
4359 verify_overflow_empty();
4360
4361 {
4362 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4363 refProcessingWork();
4364 }
4365 verify_work_stacks_empty();
4366 verify_overflow_empty();
4367
4368 if (should_unload_classes()) {
4369 CodeCache::gc_epilogue();
4370 }
4371 JvmtiExport::gc_epilogue();
4372
4373 // If we encountered any (marking stack / work queue) overflow
4374 // events during the current CMS cycle, take appropriate
4375 // remedial measures, where possible, so as to try and avoid
4376 // recurrence of that condition.
4377 assert(_markStack.isEmpty(), "No grey objects");
4378 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4379 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4380 if (ser_ovflw > 0) {
4381 if (PrintCMSStatistics != 0) {
4382 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4383 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4384 ", kac_preclean="SIZE_FORMAT")",
4385 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4386 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4387 }
4388 _markStack.expand();
4389 _ser_pmc_remark_ovflw = 0;
4390 _ser_pmc_preclean_ovflw = 0;
4391 _ser_kac_preclean_ovflw = 0;
4392 _ser_kac_ovflw = 0;
4393 }
4394 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4395 if (PrintCMSStatistics != 0) {
4396 gclog_or_tty->print_cr("Work queue overflow (benign) "
4397 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4398 _par_pmc_remark_ovflw, _par_kac_ovflw);
4399 }
4400 _par_pmc_remark_ovflw = 0;
4401 _par_kac_ovflw = 0;
4402 }
4403 if (PrintCMSStatistics != 0) {
4404 if (_markStack._hit_limit > 0) {
4405 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4406 _markStack._hit_limit);
4407 }
4408 if (_markStack._failed_double > 0) {
4409 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4410 " current capacity "SIZE_FORMAT,
4411 _markStack._failed_double,
4412 _markStack.capacity());
4413 }
4414 }
4415 _markStack._hit_limit = 0;
4416 _markStack._failed_double = 0;
4417
4418 if ((VerifyAfterGC || VerifyDuringGC) &&
4419 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4420 verify_after_remark();
4421 }
4422
4423 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4424
4425 // Change under the freelistLocks.
4426 _collectorState = Sweeping;
4427 // Call isAllClear() under bitMapLock
4428 assert(_modUnionTable.isAllClear(),
4429 "Should be clear by end of the final marking");
4430 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4431 "Should be clear by end of the final marking");
4432 }
4433
4434 void CMSParInitialMarkTask::work(uint worker_id) {
4435 elapsedTimer _timer;
4436 ResourceMark rm;
4437 HandleMark hm;
4438
4439 // ---------- scan from roots --------------
4440 _timer.start();
4441 GenCollectedHeap* gch = GenCollectedHeap::heap();
4442 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4443
4444 // ---------- young gen roots --------------
4445 {
4446 work_on_young_gen_roots(worker_id, &par_mri_cl);
4447 _timer.stop();
4448 if (PrintCMSStatistics != 0) {
4449 gclog_or_tty->print_cr(
4450 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4451 worker_id, _timer.seconds());
4452 }
4453 }
4454
4455 // ---------- remaining roots --------------
4456 _timer.reset();
4457 _timer.start();
4458
4459 CLDToOopClosure cld_closure(&par_mri_cl, true);
4460
4461 gch->gen_process_roots(_strong_roots_scope,
4462 _collector->_cmsGen->level(),
4463 false, // yg was scanned above
4464 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4465 _collector->should_unload_classes(),
4466 &par_mri_cl,
4467 NULL,
4468 &cld_closure);
4469 assert(_collector->should_unload_classes()
4470 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4471 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4472 _timer.stop();
4473 if (PrintCMSStatistics != 0) {
4474 gclog_or_tty->print_cr(
4475 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4476 worker_id, _timer.seconds());
4477 }
4478 }
4479
4480 // Parallel remark task
4481 class CMSParRemarkTask: public CMSParMarkTask {
4482 CompactibleFreeListSpace* _cms_space;
4483
4484 // The per-thread work queues, available here for stealing.
4485 OopTaskQueueSet* _task_queues;
4486 ParallelTaskTerminator _term;
4487 StrongRootsScope* _strong_roots_scope;
4488
4489 public:
4490 // A value of 0 passed to n_workers will cause the number of
4491 // workers to be taken from the active workers in the work gang.
4492 CMSParRemarkTask(CMSCollector* collector,
4493 CompactibleFreeListSpace* cms_space,
4494 uint n_workers, FlexibleWorkGang* workers,
4495 OopTaskQueueSet* task_queues,
4496 StrongRootsScope* strong_roots_scope):
4497 CMSParMarkTask("Rescan roots and grey objects in parallel",
4498 collector, n_workers),
4499 _cms_space(cms_space),
4500 _task_queues(task_queues),
4501 _term(n_workers, task_queues),
4502 _strong_roots_scope(strong_roots_scope) { }
4503
4504 OopTaskQueueSet* task_queues() { return _task_queues; }
4505
4506 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4507
4508 ParallelTaskTerminator* terminator() { return &_term; }
4509 uint n_workers() { return _n_workers; }
4510
4511 void work(uint worker_id);
4512
4513 private:
4514 // ... of dirty cards in old space
4515 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4516 Par_MarkRefsIntoAndScanClosure* cl);
4517
4518 // ... work stealing for the above
4519 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4520 };
4521
4522 class RemarkKlassClosure : public KlassClosure {
4523 KlassToOopClosure _cm_klass_closure;
4524 public:
4525 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4526 void do_klass(Klass* k) {
4527 // Check if we have modified any oops in the Klass during the concurrent marking.
4528 if (k->has_accumulated_modified_oops()) {
4529 k->clear_accumulated_modified_oops();
4530
4531 // We could have transfered the current modified marks to the accumulated marks,
4532 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4533 } else if (k->has_modified_oops()) {
4534 // Don't clear anything, this info is needed by the next young collection.
4535 } else {
4536 // No modified oops in the Klass.
4537 return;
4538 }
4539
4540 // The klass has modified fields, need to scan the klass.
4541 _cm_klass_closure.do_klass(k);
4542 }
4543 };
4544
4545 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4546 ParNewGeneration* young_gen = _collector->_young_gen;
4547 ContiguousSpace* eden_space = young_gen->eden();
4548 ContiguousSpace* from_space = young_gen->from();
4549 ContiguousSpace* to_space = young_gen->to();
4550
4551 HeapWord** eca = _collector->_eden_chunk_array;
4552 size_t ect = _collector->_eden_chunk_index;
4553 HeapWord** sca = _collector->_survivor_chunk_array;
4554 size_t sct = _collector->_survivor_chunk_index;
4555
4556 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4557 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4558
4559 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4560 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4561 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4562 }
4563
4564 // work_queue(i) is passed to the closure
4565 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4566 // also is passed to do_dirty_card_rescan_tasks() and to
4567 // do_work_steal() to select the i-th task_queue.
4568
4569 void CMSParRemarkTask::work(uint worker_id) {
4570 elapsedTimer _timer;
4571 ResourceMark rm;
4572 HandleMark hm;
4573
4574 // ---------- rescan from roots --------------
4575 _timer.start();
4576 GenCollectedHeap* gch = GenCollectedHeap::heap();
4577 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4578 _collector->_span, _collector->ref_processor(),
4579 &(_collector->_markBitMap),
4580 work_queue(worker_id));
4581
4582 // Rescan young gen roots first since these are likely
4583 // coarsely partitioned and may, on that account, constitute
4584 // the critical path; thus, it's best to start off that
4585 // work first.
4586 // ---------- young gen roots --------------
4587 {
4588 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4589 _timer.stop();
4590 if (PrintCMSStatistics != 0) {
4591 gclog_or_tty->print_cr(
4592 "Finished young gen rescan work in %dth thread: %3.3f sec",
4593 worker_id, _timer.seconds());
4594 }
4595 }
4596
4597 // ---------- remaining roots --------------
4598 _timer.reset();
4599 _timer.start();
4600 gch->gen_process_roots(_strong_roots_scope,
4601 _collector->_cmsGen->level(),
4602 false, // yg was scanned above
4603 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4604 _collector->should_unload_classes(),
4605 &par_mrias_cl,
4606 NULL,
4607 NULL); // The dirty klasses will be handled below
4608
4609 assert(_collector->should_unload_classes()
4610 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4611 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4612 _timer.stop();
4613 if (PrintCMSStatistics != 0) {
4614 gclog_or_tty->print_cr(
4615 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4616 worker_id, _timer.seconds());
4617 }
4618
4619 // ---------- unhandled CLD scanning ----------
4620 if (worker_id == 0) { // Single threaded at the moment.
4621 _timer.reset();
4622 _timer.start();
4623
4624 // Scan all new class loader data objects and new dependencies that were
4625 // introduced during concurrent marking.
4626 ResourceMark rm;
4627 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4628 for (int i = 0; i < array->length(); i++) {
4629 par_mrias_cl.do_class_loader_data(array->at(i));
4630 }
4631
4632 // We don't need to keep track of new CLDs anymore.
4633 ClassLoaderDataGraph::remember_new_clds(false);
4634
4635 _timer.stop();
4636 if (PrintCMSStatistics != 0) {
4637 gclog_or_tty->print_cr(
4638 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4639 worker_id, _timer.seconds());
4640 }
4641 }
4642
4643 // ---------- dirty klass scanning ----------
4644 if (worker_id == 0) { // Single threaded at the moment.
4645 _timer.reset();
4646 _timer.start();
4647
4648 // Scan all classes that was dirtied during the concurrent marking phase.
4649 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4650 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4651
4652 _timer.stop();
4653 if (PrintCMSStatistics != 0) {
4654 gclog_or_tty->print_cr(
4655 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4656 worker_id, _timer.seconds());
4657 }
4658 }
4659
4660 // We might have added oops to ClassLoaderData::_handles during the
4661 // concurrent marking phase. These oops point to newly allocated objects
4662 // that are guaranteed to be kept alive. Either by the direct allocation
4663 // code, or when the young collector processes the roots. Hence,
4664 // we don't have to revisit the _handles block during the remark phase.
4665
4666 // ---------- rescan dirty cards ------------
4667 _timer.reset();
4668 _timer.start();
4669
4670 // Do the rescan tasks for each of the two spaces
4671 // (cms_space) in turn.
4672 // "worker_id" is passed to select the task_queue for "worker_id"
4673 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4674 _timer.stop();
4675 if (PrintCMSStatistics != 0) {
4676 gclog_or_tty->print_cr(
4677 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4678 worker_id, _timer.seconds());
4679 }
4680
4681 // ---------- steal work from other threads ...
4682 // ---------- ... and drain overflow list.
4683 _timer.reset();
4684 _timer.start();
4685 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4686 _timer.stop();
4687 if (PrintCMSStatistics != 0) {
4688 gclog_or_tty->print_cr(
4689 "Finished work stealing in %dth thread: %3.3f sec",
4690 worker_id, _timer.seconds());
4691 }
4692 }
4693
4694 // Note that parameter "i" is not used.
4695 void
4696 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4697 OopsInGenClosure* cl, ContiguousSpace* space,
4698 HeapWord** chunk_array, size_t chunk_top) {
4699 // Until all tasks completed:
4700 // . claim an unclaimed task
4701 // . compute region boundaries corresponding to task claimed
4702 // using chunk_array
4703 // . par_oop_iterate(cl) over that region
4704
4705 ResourceMark rm;
4706 HandleMark hm;
4707
4708 SequentialSubTasksDone* pst = space->par_seq_tasks();
4709
4710 uint nth_task = 0;
4711 uint n_tasks = pst->n_tasks();
4712
4713 if (n_tasks > 0) {
4714 assert(pst->valid(), "Uninitialized use?");
4715 HeapWord *start, *end;
4716 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4717 // We claimed task # nth_task; compute its boundaries.
4718 if (chunk_top == 0) { // no samples were taken
4719 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4720 start = space->bottom();
4721 end = space->top();
4722 } else if (nth_task == 0) {
4723 start = space->bottom();
4724 end = chunk_array[nth_task];
4725 } else if (nth_task < (uint)chunk_top) {
4726 assert(nth_task >= 1, "Control point invariant");
4727 start = chunk_array[nth_task - 1];
4728 end = chunk_array[nth_task];
4729 } else {
4730 assert(nth_task == (uint)chunk_top, "Control point invariant");
4731 start = chunk_array[chunk_top - 1];
4732 end = space->top();
4733 }
4734 MemRegion mr(start, end);
4735 // Verify that mr is in space
4736 assert(mr.is_empty() || space->used_region().contains(mr),
4737 "Should be in space");
4738 // Verify that "start" is an object boundary
4739 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4740 "Should be an oop");
4741 space->par_oop_iterate(mr, cl);
4742 }
4743 pst->all_tasks_completed();
4744 }
4745 }
4746
4747 void
4748 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4749 CompactibleFreeListSpace* sp, int i,
4750 Par_MarkRefsIntoAndScanClosure* cl) {
4751 // Until all tasks completed:
4752 // . claim an unclaimed task
4753 // . compute region boundaries corresponding to task claimed
4754 // . transfer dirty bits ct->mut for that region
4755 // . apply rescanclosure to dirty mut bits for that region
4756
4757 ResourceMark rm;
4758 HandleMark hm;
4759
4760 OopTaskQueue* work_q = work_queue(i);
4761 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4762 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4763 // CAUTION: This closure has state that persists across calls to
4764 // the work method dirty_range_iterate_clear() in that it has
4765 // embedded in it a (subtype of) UpwardsObjectClosure. The
4766 // use of that state in the embedded UpwardsObjectClosure instance
4767 // assumes that the cards are always iterated (even if in parallel
4768 // by several threads) in monotonically increasing order per each
4769 // thread. This is true of the implementation below which picks
4770 // card ranges (chunks) in monotonically increasing order globally
4771 // and, a-fortiori, in monotonically increasing order per thread
4772 // (the latter order being a subsequence of the former).
4773 // If the work code below is ever reorganized into a more chaotic
4774 // work-partitioning form than the current "sequential tasks"
4775 // paradigm, the use of that persistent state will have to be
4776 // revisited and modified appropriately. See also related
4777 // bug 4756801 work on which should examine this code to make
4778 // sure that the changes there do not run counter to the
4779 // assumptions made here and necessary for correctness and
4780 // efficiency. Note also that this code might yield inefficient
4781 // behavior in the case of very large objects that span one or
4782 // more work chunks. Such objects would potentially be scanned
4783 // several times redundantly. Work on 4756801 should try and
4784 // address that performance anomaly if at all possible. XXX
4785 MemRegion full_span = _collector->_span;
4786 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4787 MarkFromDirtyCardsClosure
4788 greyRescanClosure(_collector, full_span, // entire span of interest
4789 sp, bm, work_q, cl);
4790
4791 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4792 assert(pst->valid(), "Uninitialized use?");
4793 uint nth_task = 0;
4794 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4795 MemRegion span = sp->used_region();
4796 HeapWord* start_addr = span.start();
4797 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4798 alignment);
4799 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4800 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4801 start_addr, "Check alignment");
4802 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4803 chunk_size, "Check alignment");
4804
4805 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4806 // Having claimed the nth_task, compute corresponding mem-region,
4807 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4808 // The alignment restriction ensures that we do not need any
4809 // synchronization with other gang-workers while setting or
4810 // clearing bits in thus chunk of the MUT.
4811 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4812 start_addr + (nth_task+1)*chunk_size);
4813 // The last chunk's end might be way beyond end of the
4814 // used region. In that case pull back appropriately.
4815 if (this_span.end() > end_addr) {
4816 this_span.set_end(end_addr);
4817 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4818 }
4819 // Iterate over the dirty cards covering this chunk, marking them
4820 // precleaned, and setting the corresponding bits in the mod union
4821 // table. Since we have been careful to partition at Card and MUT-word
4822 // boundaries no synchronization is needed between parallel threads.
4823 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4824 &modUnionClosure);
4825
4826 // Having transferred these marks into the modUnionTable,
4827 // rescan the marked objects on the dirty cards in the modUnionTable.
4828 // Even if this is at a synchronous collection, the initial marking
4829 // may have been done during an asynchronous collection so there
4830 // may be dirty bits in the mod-union table.
4831 _collector->_modUnionTable.dirty_range_iterate_clear(
4832 this_span, &greyRescanClosure);
4833 _collector->_modUnionTable.verifyNoOneBitsInRange(
4834 this_span.start(),
4835 this_span.end());
4836 }
4837 pst->all_tasks_completed(); // declare that i am done
4838 }
4839
4840 // . see if we can share work_queues with ParNew? XXX
4841 void
4842 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4843 int* seed) {
4844 OopTaskQueue* work_q = work_queue(i);
4845 NOT_PRODUCT(int num_steals = 0;)
4846 oop obj_to_scan;
4847 CMSBitMap* bm = &(_collector->_markBitMap);
4848
4849 while (true) {
4850 // Completely finish any left over work from (an) earlier round(s)
4851 cl->trim_queue(0);
4852 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4853 (size_t)ParGCDesiredObjsFromOverflowList);
4854 // Now check if there's any work in the overflow list
4855 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4856 // only affects the number of attempts made to get work from the
4857 // overflow list and does not affect the number of workers. Just
4858 // pass ParallelGCThreads so this behavior is unchanged.
4859 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4860 work_q,
4861 ParallelGCThreads)) {
4862 // found something in global overflow list;
4863 // not yet ready to go stealing work from others.
4864 // We'd like to assert(work_q->size() != 0, ...)
4865 // because we just took work from the overflow list,
4866 // but of course we can't since all of that could have
4867 // been already stolen from us.
4868 // "He giveth and He taketh away."
4869 continue;
4870 }
4871 // Verify that we have no work before we resort to stealing
4872 assert(work_q->size() == 0, "Have work, shouldn't steal");
4873 // Try to steal from other queues that have work
4874 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4875 NOT_PRODUCT(num_steals++;)
4876 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4877 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4878 // Do scanning work
4879 obj_to_scan->oop_iterate(cl);
4880 // Loop around, finish this work, and try to steal some more
4881 } else if (terminator()->offer_termination()) {
4882 break; // nirvana from the infinite cycle
4883 }
4884 }
4885 NOT_PRODUCT(
4886 if (PrintCMSStatistics != 0) {
4887 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4888 }
4889 )
4890 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4891 "Else our work is not yet done");
4892 }
4893
4894 // Record object boundaries in _eden_chunk_array by sampling the eden
4895 // top in the slow-path eden object allocation code path and record
4896 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4897 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4898 // sampling in sample_eden() that activates during the part of the
4899 // preclean phase.
4900 void CMSCollector::sample_eden_chunk() {
4901 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4902 if (_eden_chunk_lock->try_lock()) {
4903 // Record a sample. This is the critical section. The contents
4904 // of the _eden_chunk_array have to be non-decreasing in the
4905 // address order.
4906 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4907 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4908 "Unexpected state of Eden");
4909 if (_eden_chunk_index == 0 ||
4910 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4911 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4912 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4913 _eden_chunk_index++; // commit sample
4914 }
4915 _eden_chunk_lock->unlock();
4916 }
4917 }
4918 }
4919
4920 // Return a thread-local PLAB recording array, as appropriate.
4921 void* CMSCollector::get_data_recorder(int thr_num) {
4922 if (_survivor_plab_array != NULL &&
4923 (CMSPLABRecordAlways ||
4924 (_collectorState > Marking && _collectorState < FinalMarking))) {
4925 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4926 ChunkArray* ca = &_survivor_plab_array[thr_num];
4927 ca->reset(); // clear it so that fresh data is recorded
4928 return (void*) ca;
4929 } else {
4930 return NULL;
4931 }
4932 }
4933
4934 // Reset all the thread-local PLAB recording arrays
4935 void CMSCollector::reset_survivor_plab_arrays() {
4936 for (uint i = 0; i < ParallelGCThreads; i++) {
4937 _survivor_plab_array[i].reset();
4938 }
4939 }
4940
4941 // Merge the per-thread plab arrays into the global survivor chunk
4942 // array which will provide the partitioning of the survivor space
4943 // for CMS initial scan and rescan.
4944 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4945 int no_of_gc_threads) {
4946 assert(_survivor_plab_array != NULL, "Error");
4947 assert(_survivor_chunk_array != NULL, "Error");
4948 assert(_collectorState == FinalMarking ||
4949 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4950 for (int j = 0; j < no_of_gc_threads; j++) {
4951 _cursor[j] = 0;
4952 }
4953 HeapWord* top = surv->top();
4954 size_t i;
4955 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4956 HeapWord* min_val = top; // Higher than any PLAB address
4957 uint min_tid = 0; // position of min_val this round
4958 for (int j = 0; j < no_of_gc_threads; j++) {
4959 ChunkArray* cur_sca = &_survivor_plab_array[j];
4960 if (_cursor[j] == cur_sca->end()) {
4961 continue;
4962 }
4963 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4964 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4965 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4966 if (cur_val < min_val) {
4967 min_tid = j;
4968 min_val = cur_val;
4969 } else {
4970 assert(cur_val < top, "All recorded addresses should be less");
4971 }
4972 }
4973 // At this point min_val and min_tid are respectively
4974 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4975 // and the thread (j) that witnesses that address.
4976 // We record this address in the _survivor_chunk_array[i]
4977 // and increment _cursor[min_tid] prior to the next round i.
4978 if (min_val == top) {
4979 break;
4980 }
4981 _survivor_chunk_array[i] = min_val;
4982 _cursor[min_tid]++;
4983 }
4984 // We are all done; record the size of the _survivor_chunk_array
4985 _survivor_chunk_index = i; // exclusive: [0, i)
4986 if (PrintCMSStatistics > 0) {
4987 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4988 }
4989 // Verify that we used up all the recorded entries
4990 #ifdef ASSERT
4991 size_t total = 0;
4992 for (int j = 0; j < no_of_gc_threads; j++) {
4993 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4994 total += _cursor[j];
4995 }
4996 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4997 // Check that the merged array is in sorted order
4998 if (total > 0) {
4999 for (size_t i = 0; i < total - 1; i++) {
5000 if (PrintCMSStatistics > 0) {
5001 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5002 i, p2i(_survivor_chunk_array[i]));
5003 }
5004 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5005 "Not sorted");
5006 }
5007 }
5008 #endif // ASSERT
5009 }
5010
5011 // Set up the space's par_seq_tasks structure for work claiming
5012 // for parallel initial scan and rescan of young gen.
5013 // See ParRescanTask where this is currently used.
5014 void
5015 CMSCollector::
5016 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5017 assert(n_threads > 0, "Unexpected n_threads argument");
5018
5019 // Eden space
5020 if (!_young_gen->eden()->is_empty()) {
5021 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5022 assert(!pst->valid(), "Clobbering existing data?");
5023 // Each valid entry in [0, _eden_chunk_index) represents a task.
5024 size_t n_tasks = _eden_chunk_index + 1;
5025 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5026 // Sets the condition for completion of the subtask (how many threads
5027 // need to finish in order to be done).
5028 pst->set_n_threads(n_threads);
5029 pst->set_n_tasks((int)n_tasks);
5030 }
5031
5032 // Merge the survivor plab arrays into _survivor_chunk_array
5033 if (_survivor_plab_array != NULL) {
5034 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5035 } else {
5036 assert(_survivor_chunk_index == 0, "Error");
5037 }
5038
5039 // To space
5040 {
5041 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5042 assert(!pst->valid(), "Clobbering existing data?");
5043 // Sets the condition for completion of the subtask (how many threads
5044 // need to finish in order to be done).
5045 pst->set_n_threads(n_threads);
5046 pst->set_n_tasks(1);
5047 assert(pst->valid(), "Error");
5048 }
5049
5050 // From space
5051 {
5052 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5053 assert(!pst->valid(), "Clobbering existing data?");
5054 size_t n_tasks = _survivor_chunk_index + 1;
5055 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5056 // Sets the condition for completion of the subtask (how many threads
5057 // need to finish in order to be done).
5058 pst->set_n_threads(n_threads);
5059 pst->set_n_tasks((int)n_tasks);
5060 assert(pst->valid(), "Error");
5061 }
5062 }
5063
5064 // Parallel version of remark
5065 void CMSCollector::do_remark_parallel() {
5066 GenCollectedHeap* gch = GenCollectedHeap::heap();
5067 FlexibleWorkGang* workers = gch->workers();
5068 assert(workers != NULL, "Need parallel worker threads.");
5069 // Choose to use the number of GC workers most recently set
5070 // into "active_workers".
5071 uint n_workers = workers->active_workers();
5072
5073 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5074
5075 StrongRootsScope srs(n_workers);
5076
5077 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
5078
5079 // We won't be iterating over the cards in the card table updating
5080 // the younger_gen cards, so we shouldn't call the following else
5081 // the verification code as well as subsequent younger_refs_iterate
5082 // code would get confused. XXX
5083 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5084
5085 // The young gen rescan work will not be done as part of
5086 // process_roots (which currently doesn't know how to
5087 // parallelize such a scan), but rather will be broken up into
5088 // a set of parallel tasks (via the sampling that the [abortable]
5089 // preclean phase did of eden, plus the [two] tasks of
5090 // scanning the [two] survivor spaces. Further fine-grain
5091 // parallelization of the scanning of the survivor spaces
5092 // themselves, and of precleaning of the younger gen itself
5093 // is deferred to the future.
5094 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5095
5096 // The dirty card rescan work is broken up into a "sequence"
5097 // of parallel tasks (per constituent space) that are dynamically
5098 // claimed by the parallel threads.
5099 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5100
5101 // It turns out that even when we're using 1 thread, doing the work in a
5102 // separate thread causes wide variance in run times. We can't help this
5103 // in the multi-threaded case, but we special-case n=1 here to get
5104 // repeatable measurements of the 1-thread overhead of the parallel code.
5105 if (n_workers > 1) {
5106 // Make refs discovery MT-safe, if it isn't already: it may not
5107 // necessarily be so, since it's possible that we are doing
5108 // ST marking.
5109 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5110 workers->run_task(&tsk);
5111 } else {
5112 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5113 tsk.work(0);
5114 }
5115
5116 // restore, single-threaded for now, any preserved marks
5117 // as a result of work_q overflow
5118 restore_preserved_marks_if_any();
5119 }
5120
5121 // Non-parallel version of remark
5122 void CMSCollector::do_remark_non_parallel() {
5123 ResourceMark rm;
5124 HandleMark hm;
5125 GenCollectedHeap* gch = GenCollectedHeap::heap();
5126 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5127
5128 MarkRefsIntoAndScanClosure
5129 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5130 &_markStack, this,
5131 false /* should_yield */, false /* not precleaning */);
5132 MarkFromDirtyCardsClosure
5133 markFromDirtyCardsClosure(this, _span,
5134 NULL, // space is set further below
5135 &_markBitMap, &_markStack, &mrias_cl);
5136 {
5137 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5138 // Iterate over the dirty cards, setting the corresponding bits in the
5139 // mod union table.
5140 {
5141 ModUnionClosure modUnionClosure(&_modUnionTable);
5142 _ct->ct_bs()->dirty_card_iterate(
5143 _cmsGen->used_region(),
5144 &modUnionClosure);
5145 }
5146 // Having transferred these marks into the modUnionTable, we just need
5147 // to rescan the marked objects on the dirty cards in the modUnionTable.
5148 // The initial marking may have been done during an asynchronous
5149 // collection so there may be dirty bits in the mod-union table.
5150 const int alignment =
5151 CardTableModRefBS::card_size * BitsPerWord;
5152 {
5153 // ... First handle dirty cards in CMS gen
5154 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5155 MemRegion ur = _cmsGen->used_region();
5156 HeapWord* lb = ur.start();
5157 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5158 MemRegion cms_span(lb, ub);
5159 _modUnionTable.dirty_range_iterate_clear(cms_span,
5160 &markFromDirtyCardsClosure);
5161 verify_work_stacks_empty();
5162 if (PrintCMSStatistics != 0) {
5163 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5164 markFromDirtyCardsClosure.num_dirty_cards());
5165 }
5166 }
5167 }
5168 if (VerifyDuringGC &&
5169 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5170 HandleMark hm; // Discard invalid handles created during verification
5171 Universe::verify();
5172 }
5173 {
5174 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5175
5176 verify_work_stacks_empty();
5177
5178 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5179 StrongRootsScope srs(1);
5180
5181 gch->gen_process_roots(&srs,
5182 _cmsGen->level(),
5183 true, // younger gens as roots
5184 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5185 should_unload_classes(),
5186 &mrias_cl,
5187 NULL,
5188 NULL); // The dirty klasses will be handled below
5189
5190 assert(should_unload_classes()
5191 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5192 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5193 }
5194
5195 {
5196 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5197
5198 verify_work_stacks_empty();
5199
5200 // Scan all class loader data objects that might have been introduced
5201 // during concurrent marking.
5202 ResourceMark rm;
5203 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5204 for (int i = 0; i < array->length(); i++) {
5205 mrias_cl.do_class_loader_data(array->at(i));
5206 }
5207
5208 // We don't need to keep track of new CLDs anymore.
5209 ClassLoaderDataGraph::remember_new_clds(false);
5210
5211 verify_work_stacks_empty();
5212 }
5213
5214 {
5215 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5216
5217 verify_work_stacks_empty();
5218
5219 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5220 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5221
5222 verify_work_stacks_empty();
5223 }
5224
5225 // We might have added oops to ClassLoaderData::_handles during the
5226 // concurrent marking phase. These oops point to newly allocated objects
5227 // that are guaranteed to be kept alive. Either by the direct allocation
5228 // code, or when the young collector processes the roots. Hence,
5229 // we don't have to revisit the _handles block during the remark phase.
5230
5231 verify_work_stacks_empty();
5232 // Restore evacuated mark words, if any, used for overflow list links
5233 if (!CMSOverflowEarlyRestoration) {
5234 restore_preserved_marks_if_any();
5235 }
5236 verify_overflow_empty();
5237 }
5238
5239 ////////////////////////////////////////////////////////
5240 // Parallel Reference Processing Task Proxy Class
5241 ////////////////////////////////////////////////////////
5242 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5243 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5244 CMSCollector* _collector;
5245 CMSBitMap* _mark_bit_map;
5246 const MemRegion _span;
5247 ProcessTask& _task;
5248
5249 public:
5250 CMSRefProcTaskProxy(ProcessTask& task,
5251 CMSCollector* collector,
5252 const MemRegion& span,
5253 CMSBitMap* mark_bit_map,
5254 AbstractWorkGang* workers,
5255 OopTaskQueueSet* task_queues):
5256 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5257 task_queues,
5258 workers->active_workers()),
5259 _task(task),
5260 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5261 {
5262 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5263 "Inconsistency in _span");
5264 }
5265
5266 OopTaskQueueSet* task_queues() { return queues(); }
5267
5268 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5269
5270 void do_work_steal(int i,
5271 CMSParDrainMarkingStackClosure* drain,
5272 CMSParKeepAliveClosure* keep_alive,
5273 int* seed);
5274
5275 virtual void work(uint worker_id);
5276 };
5277
5278 void CMSRefProcTaskProxy::work(uint worker_id) {
5279 ResourceMark rm;
5280 HandleMark hm;
5281 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5282 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5283 _mark_bit_map,
5284 work_queue(worker_id));
5285 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5286 _mark_bit_map,
5287 work_queue(worker_id));
5288 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5289 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5290 if (_task.marks_oops_alive()) {
5291 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5292 _collector->hash_seed(worker_id));
5293 }
5294 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5295 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5296 }
5297
5298 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5299 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5300 EnqueueTask& _task;
5301
5302 public:
5303 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5304 : AbstractGangTask("Enqueue reference objects in parallel"),
5305 _task(task)
5306 { }
5307
5308 virtual void work(uint worker_id)
5309 {
5310 _task.work(worker_id);
5311 }
5312 };
5313
5314 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5315 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5316 _span(span),
5317 _bit_map(bit_map),
5318 _work_queue(work_queue),
5319 _mark_and_push(collector, span, bit_map, work_queue),
5320 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5321 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5322 { }
5323
5324 // . see if we can share work_queues with ParNew? XXX
5325 void CMSRefProcTaskProxy::do_work_steal(int i,
5326 CMSParDrainMarkingStackClosure* drain,
5327 CMSParKeepAliveClosure* keep_alive,
5328 int* seed) {
5329 OopTaskQueue* work_q = work_queue(i);
5330 NOT_PRODUCT(int num_steals = 0;)
5331 oop obj_to_scan;
5332
5333 while (true) {
5334 // Completely finish any left over work from (an) earlier round(s)
5335 drain->trim_queue(0);
5336 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5337 (size_t)ParGCDesiredObjsFromOverflowList);
5338 // Now check if there's any work in the overflow list
5339 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5340 // only affects the number of attempts made to get work from the
5341 // overflow list and does not affect the number of workers. Just
5342 // pass ParallelGCThreads so this behavior is unchanged.
5343 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5344 work_q,
5345 ParallelGCThreads)) {
5346 // Found something in global overflow list;
5347 // not yet ready to go stealing work from others.
5348 // We'd like to assert(work_q->size() != 0, ...)
5349 // because we just took work from the overflow list,
5350 // but of course we can't, since all of that might have
5351 // been already stolen from us.
5352 continue;
5353 }
5354 // Verify that we have no work before we resort to stealing
5355 assert(work_q->size() == 0, "Have work, shouldn't steal");
5356 // Try to steal from other queues that have work
5357 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5358 NOT_PRODUCT(num_steals++;)
5359 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5360 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5361 // Do scanning work
5362 obj_to_scan->oop_iterate(keep_alive);
5363 // Loop around, finish this work, and try to steal some more
5364 } else if (terminator()->offer_termination()) {
5365 break; // nirvana from the infinite cycle
5366 }
5367 }
5368 NOT_PRODUCT(
5369 if (PrintCMSStatistics != 0) {
5370 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5371 }
5372 )
5373 }
5374
5375 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5376 {
5377 GenCollectedHeap* gch = GenCollectedHeap::heap();
5378 FlexibleWorkGang* workers = gch->workers();
5379 assert(workers != NULL, "Need parallel worker threads.");
5380 CMSRefProcTaskProxy rp_task(task, &_collector,
5381 _collector.ref_processor()->span(),
5382 _collector.markBitMap(),
5383 workers, _collector.task_queues());
5384 workers->run_task(&rp_task);
5385 }
5386
5387 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5388 {
5389
5390 GenCollectedHeap* gch = GenCollectedHeap::heap();
5391 FlexibleWorkGang* workers = gch->workers();
5392 assert(workers != NULL, "Need parallel worker threads.");
5393 CMSRefEnqueueTaskProxy enq_task(task);
5394 workers->run_task(&enq_task);
5395 }
5396
5397 void CMSCollector::refProcessingWork() {
5398 ResourceMark rm;
5399 HandleMark hm;
5400
5401 ReferenceProcessor* rp = ref_processor();
5402 assert(rp->span().equals(_span), "Spans should be equal");
5403 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5404 // Process weak references.
5405 rp->setup_policy(false);
5406 verify_work_stacks_empty();
5407
5408 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5409 &_markStack, false /* !preclean */);
5410 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5411 _span, &_markBitMap, &_markStack,
5412 &cmsKeepAliveClosure, false /* !preclean */);
5413 {
5414 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5415
5416 ReferenceProcessorStats stats;
5417 if (rp->processing_is_mt()) {
5418 // Set the degree of MT here. If the discovery is done MT, there
5419 // may have been a different number of threads doing the discovery
5420 // and a different number of discovered lists may have Ref objects.
5421 // That is OK as long as the Reference lists are balanced (see
5422 // balance_all_queues() and balance_queues()).
5423 GenCollectedHeap* gch = GenCollectedHeap::heap();
5424 uint active_workers = ParallelGCThreads;
5425 FlexibleWorkGang* workers = gch->workers();
5426 if (workers != NULL) {
5427 active_workers = workers->active_workers();
5428 // The expectation is that active_workers will have already
5429 // been set to a reasonable value. If it has not been set,
5430 // investigate.
5431 assert(active_workers > 0, "Should have been set during scavenge");
5432 }
5433 rp->set_active_mt_degree(active_workers);
5434 CMSRefProcTaskExecutor task_executor(*this);
5435 stats = rp->process_discovered_references(&_is_alive_closure,
5436 &cmsKeepAliveClosure,
5437 &cmsDrainMarkingStackClosure,
5438 &task_executor,
5439 _gc_timer_cm,
5440 _gc_tracer_cm->gc_id());
5441 } else {
5442 stats = rp->process_discovered_references(&_is_alive_closure,
5443 &cmsKeepAliveClosure,
5444 &cmsDrainMarkingStackClosure,
5445 NULL,
5446 _gc_timer_cm,
5447 _gc_tracer_cm->gc_id());
5448 }
5449 _gc_tracer_cm->report_gc_reference_stats(stats);
5450
5451 }
5452
5453 // This is the point where the entire marking should have completed.
5454 verify_work_stacks_empty();
5455
5456 if (should_unload_classes()) {
5457 {
5458 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5459
5460 // Unload classes and purge the SystemDictionary.
5461 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5462
5463 // Unload nmethods.
5464 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5465
5466 // Prune dead klasses from subklass/sibling/implementor lists.
5467 Klass::clean_weak_klass_links(&_is_alive_closure);
5468 }
5469
5470 {
5471 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5472 // Clean up unreferenced symbols in symbol table.
5473 SymbolTable::unlink();
5474 }
5475
5476 {
5477 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5478 // Delete entries for dead interned strings.
5479 StringTable::unlink(&_is_alive_closure);
5480 }
5481 }
5482
5483
5484 // Restore any preserved marks as a result of mark stack or
5485 // work queue overflow
5486 restore_preserved_marks_if_any(); // done single-threaded for now
5487
5488 rp->set_enqueuing_is_done(true);
5489 if (rp->processing_is_mt()) {
5490 rp->balance_all_queues();
5491 CMSRefProcTaskExecutor task_executor(*this);
5492 rp->enqueue_discovered_references(&task_executor);
5493 } else {
5494 rp->enqueue_discovered_references(NULL);
5495 }
5496 rp->verify_no_references_recorded();
5497 assert(!rp->discovery_enabled(), "should have been disabled");
5498 }
5499
5500 #ifndef PRODUCT
5501 void CMSCollector::check_correct_thread_executing() {
5502 Thread* t = Thread::current();
5503 // Only the VM thread or the CMS thread should be here.
5504 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5505 "Unexpected thread type");
5506 // If this is the vm thread, the foreground process
5507 // should not be waiting. Note that _foregroundGCIsActive is
5508 // true while the foreground collector is waiting.
5509 if (_foregroundGCShouldWait) {
5510 // We cannot be the VM thread
5511 assert(t->is_ConcurrentGC_thread(),
5512 "Should be CMS thread");
5513 } else {
5514 // We can be the CMS thread only if we are in a stop-world
5515 // phase of CMS collection.
5516 if (t->is_ConcurrentGC_thread()) {
5517 assert(_collectorState == InitialMarking ||
5518 _collectorState == FinalMarking,
5519 "Should be a stop-world phase");
5520 // The CMS thread should be holding the CMS_token.
5521 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5522 "Potential interference with concurrently "
5523 "executing VM thread");
5524 }
5525 }
5526 }
5527 #endif
5528
5529 void CMSCollector::sweep() {
5530 assert(_collectorState == Sweeping, "just checking");
5531 check_correct_thread_executing();
5532 verify_work_stacks_empty();
5533 verify_overflow_empty();
5534 increment_sweep_count();
5535 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5536
5537 _inter_sweep_timer.stop();
5538 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5539
5540 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5541 _intra_sweep_timer.reset();
5542 _intra_sweep_timer.start();
5543 {
5544 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5545 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5546 // First sweep the old gen
5547 {
5548 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5549 bitMapLock());
5550 sweepWork(_cmsGen);
5551 }
5552
5553 // Update Universe::_heap_*_at_gc figures.
5554 // We need all the free list locks to make the abstract state
5555 // transition from Sweeping to Resetting. See detailed note
5556 // further below.
5557 {
5558 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5559 // Update heap occupancy information which is used as
5560 // input to soft ref clearing policy at the next gc.
5561 Universe::update_heap_info_at_gc();
5562 _collectorState = Resizing;
5563 }
5564 }
5565 verify_work_stacks_empty();
5566 verify_overflow_empty();
5567
5568 if (should_unload_classes()) {
5569 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5570 // requires that the virtual spaces are stable and not deleted.
5571 ClassLoaderDataGraph::set_should_purge(true);
5572 }
5573
5574 _intra_sweep_timer.stop();
5575 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5576
5577 _inter_sweep_timer.reset();
5578 _inter_sweep_timer.start();
5579
5580 // We need to use a monotonically non-decreasing time in ms
5581 // or we will see time-warp warnings and os::javaTimeMillis()
5582 // does not guarantee monotonicity.
5583 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5584 update_time_of_last_gc(now);
5585
5586 // NOTE on abstract state transitions:
5587 // Mutators allocate-live and/or mark the mod-union table dirty
5588 // based on the state of the collection. The former is done in
5589 // the interval [Marking, Sweeping] and the latter in the interval
5590 // [Marking, Sweeping). Thus the transitions into the Marking state
5591 // and out of the Sweeping state must be synchronously visible
5592 // globally to the mutators.
5593 // The transition into the Marking state happens with the world
5594 // stopped so the mutators will globally see it. Sweeping is
5595 // done asynchronously by the background collector so the transition
5596 // from the Sweeping state to the Resizing state must be done
5597 // under the freelistLock (as is the check for whether to
5598 // allocate-live and whether to dirty the mod-union table).
5599 assert(_collectorState == Resizing, "Change of collector state to"
5600 " Resizing must be done under the freelistLocks (plural)");
5601
5602 // Now that sweeping has been completed, we clear
5603 // the incremental_collection_failed flag,
5604 // thus inviting a younger gen collection to promote into
5605 // this generation. If such a promotion may still fail,
5606 // the flag will be set again when a young collection is
5607 // attempted.
5608 GenCollectedHeap* gch = GenCollectedHeap::heap();
5609 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5610 gch->update_full_collections_completed(_collection_count_start);
5611 }
5612
5613 // FIX ME!!! Looks like this belongs in CFLSpace, with
5614 // CMSGen merely delegating to it.
5615 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5616 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5617 HeapWord* minAddr = _cmsSpace->bottom();
5618 HeapWord* largestAddr =
5619 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5620 if (largestAddr == NULL) {
5621 // The dictionary appears to be empty. In this case
5622 // try to coalesce at the end of the heap.
5623 largestAddr = _cmsSpace->end();
5624 }
5625 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5626 size_t nearLargestOffset =
5627 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5628 if (PrintFLSStatistics != 0) {
5629 gclog_or_tty->print_cr(
5630 "CMS: Large Block: " PTR_FORMAT ";"
5631 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5632 p2i(largestAddr),
5633 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5634 }
5635 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5636 }
5637
5638 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5639 return addr >= _cmsSpace->nearLargestChunk();
5640 }
5641
5642 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5643 return _cmsSpace->find_chunk_at_end();
5644 }
5645
5646 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5647 bool full) {
5648 // The next lower level has been collected. Gather any statistics
5649 // that are of interest at this point.
5650 if (!full && (current_level + 1) == level()) {
5651 // Gather statistics on the young generation collection.
5652 collector()->stats().record_gc0_end(used());
5653 }
5654 }
5655
5656 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) {
5657 // We iterate over the space(s) underlying this generation,
5658 // checking the mark bit map to see if the bits corresponding
5659 // to specific blocks are marked or not. Blocks that are
5660 // marked are live and are not swept up. All remaining blocks
5661 // are swept up, with coalescing on-the-fly as we sweep up
5662 // contiguous free and/or garbage blocks:
5663 // We need to ensure that the sweeper synchronizes with allocators
5664 // and stop-the-world collectors. In particular, the following
5665 // locks are used:
5666 // . CMS token: if this is held, a stop the world collection cannot occur
5667 // . freelistLock: if this is held no allocation can occur from this
5668 // generation by another thread
5669 // . bitMapLock: if this is held, no other thread can access or update
5670 //
5671
5672 // Note that we need to hold the freelistLock if we use
5673 // block iterate below; else the iterator might go awry if
5674 // a mutator (or promotion) causes block contents to change
5675 // (for instance if the allocator divvies up a block).
5676 // If we hold the free list lock, for all practical purposes
5677 // young generation GC's can't occur (they'll usually need to
5678 // promote), so we might as well prevent all young generation
5679 // GC's while we do a sweeping step. For the same reason, we might
5680 // as well take the bit map lock for the entire duration
5681
5682 // check that we hold the requisite locks
5683 assert(have_cms_token(), "Should hold cms token");
5684 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5685 assert_lock_strong(gen->freelistLock());
5686 assert_lock_strong(bitMapLock());
5687
5688 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5689 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5690 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5691 _inter_sweep_estimate.padded_average(),
5692 _intra_sweep_estimate.padded_average());
5693 gen->setNearLargestChunk();
5694
5695 {
5696 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield);
5697 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5698 // We need to free-up/coalesce garbage/blocks from a
5699 // co-terminal free run. This is done in the SweepClosure
5700 // destructor; so, do not remove this scope, else the
5701 // end-of-sweep-census below will be off by a little bit.
5702 }
5703 gen->cmsSpace()->sweep_completed();
5704 gen->cmsSpace()->endSweepFLCensus(sweep_count());
5705 if (should_unload_classes()) { // unloaded classes this cycle,
5706 _concurrent_cycles_since_last_unload = 0; // ... reset count
5707 } else { // did not unload classes,
5708 _concurrent_cycles_since_last_unload++; // ... increment count
5709 }
5710 }
5711
5712 // Reset CMS data structures (for now just the marking bit map)
5713 // preparatory for the next cycle.
5714 void CMSCollector::reset(bool concurrent) {
5715 if (concurrent) {
5716 CMSTokenSyncWithLocks ts(true, bitMapLock());
5717
5718 // If the state is not "Resetting", the foreground thread
5719 // has done a collection and the resetting.
5720 if (_collectorState != Resetting) {
5721 assert(_collectorState == Idling, "The state should only change"
5722 " because the foreground collector has finished the collection");
5723 return;
5724 }
5725
5726 // Clear the mark bitmap (no grey objects to start with)
5727 // for the next cycle.
5728 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5729 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5730
5731 HeapWord* curAddr = _markBitMap.startWord();
5732 while (curAddr < _markBitMap.endWord()) {
5733 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5734 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5735 _markBitMap.clear_large_range(chunk);
5736 if (ConcurrentMarkSweepThread::should_yield() &&
5737 !foregroundGCIsActive() &&
5738 CMSYield) {
5739 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5740 "CMS thread should hold CMS token");
5741 assert_lock_strong(bitMapLock());
5742 bitMapLock()->unlock();
5743 ConcurrentMarkSweepThread::desynchronize(true);
5744 stopTimer();
5745 if (PrintCMSStatistics != 0) {
5746 incrementYields();
5747 }
5748
5749 // See the comment in coordinator_yield()
5750 for (unsigned i = 0; i < CMSYieldSleepCount &&
5751 ConcurrentMarkSweepThread::should_yield() &&
5752 !CMSCollector::foregroundGCIsActive(); ++i) {
5753 os::sleep(Thread::current(), 1, false);
5754 }
5755
5756 ConcurrentMarkSweepThread::synchronize(true);
5757 bitMapLock()->lock_without_safepoint_check();
5758 startTimer();
5759 }
5760 curAddr = chunk.end();
5761 }
5762 // A successful mostly concurrent collection has been done.
5763 // Because only the full (i.e., concurrent mode failure) collections
5764 // are being measured for gc overhead limits, clean the "near" flag
5765 // and count.
5766 size_policy()->reset_gc_overhead_limit_count();
5767 _collectorState = Idling;
5768 } else {
5769 // already have the lock
5770 assert(_collectorState == Resetting, "just checking");
5771 assert_lock_strong(bitMapLock());
5772 _markBitMap.clear_all();
5773 _collectorState = Idling;
5774 }
5775
5776 register_gc_end();
5777 }
5778
5779 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5780 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5781 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5782 TraceCollectorStats tcs(counters());
5783
5784 switch (op) {
5785 case CMS_op_checkpointRootsInitial: {
5786 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5787 checkpointRootsInitial();
5788 if (PrintGC) {
5789 _cmsGen->printOccupancy("initial-mark");
5790 }
5791 break;
5792 }
5793 case CMS_op_checkpointRootsFinal: {
5794 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5795 checkpointRootsFinal();
5796 if (PrintGC) {
5797 _cmsGen->printOccupancy("remark");
5798 }
5799 break;
5800 }
5801 default:
5802 fatal("No such CMS_op");
5803 }
5804 }
5805
5806 #ifndef PRODUCT
5807 size_t const CMSCollector::skip_header_HeapWords() {
5808 return FreeChunk::header_size();
5809 }
5810
5811 // Try and collect here conditions that should hold when
5812 // CMS thread is exiting. The idea is that the foreground GC
5813 // thread should not be blocked if it wants to terminate
5814 // the CMS thread and yet continue to run the VM for a while
5815 // after that.
5816 void CMSCollector::verify_ok_to_terminate() const {
5817 assert(Thread::current()->is_ConcurrentGC_thread(),
5818 "should be called by CMS thread");
5819 assert(!_foregroundGCShouldWait, "should be false");
5820 // We could check here that all the various low-level locks
5821 // are not held by the CMS thread, but that is overkill; see
5822 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5823 // is checked.
5824 }
5825 #endif
5826
5827 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5828 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5829 "missing Printezis mark?");
5830 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5831 size_t size = pointer_delta(nextOneAddr + 1, addr);
5832 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5833 "alignment problem");
5834 assert(size >= 3, "Necessary for Printezis marks to work");
5835 return size;
5836 }
5837
5838 // A variant of the above (block_size_using_printezis_bits()) except
5839 // that we return 0 if the P-bits are not yet set.
5840 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5841 if (_markBitMap.isMarked(addr + 1)) {
5842 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5843 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5844 size_t size = pointer_delta(nextOneAddr + 1, addr);
5845 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5846 "alignment problem");
5847 assert(size >= 3, "Necessary for Printezis marks to work");
5848 return size;
5849 }
5850 return 0;
5851 }
5852
5853 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5854 size_t sz = 0;
5855 oop p = (oop)addr;
5856 if (p->klass_or_null() != NULL) {
5857 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5858 } else {
5859 sz = block_size_using_printezis_bits(addr);
5860 }
5861 assert(sz > 0, "size must be nonzero");
5862 HeapWord* next_block = addr + sz;
5863 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5864 CardTableModRefBS::card_size);
5865 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5866 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5867 "must be different cards");
5868 return next_card;
5869 }
5870
5871
5872 // CMS Bit Map Wrapper /////////////////////////////////////////
5873
5874 // Construct a CMS bit map infrastructure, but don't create the
5875 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5876 // further below.
5877 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5878 _bm(),
5879 _shifter(shifter),
5880 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5881 Monitor::_safepoint_check_sometimes) : NULL)
5882 {
5883 _bmStartWord = 0;
5884 _bmWordSize = 0;
5885 }
5886
5887 bool CMSBitMap::allocate(MemRegion mr) {
5888 _bmStartWord = mr.start();
5889 _bmWordSize = mr.word_size();
5890 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5891 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5892 if (!brs.is_reserved()) {
5893 warning("CMS bit map allocation failure");
5894 return false;
5895 }
5896 // For now we'll just commit all of the bit map up front.
5897 // Later on we'll try to be more parsimonious with swap.
5898 if (!_virtual_space.initialize(brs, brs.size())) {
5899 warning("CMS bit map backing store failure");
5900 return false;
5901 }
5902 assert(_virtual_space.committed_size() == brs.size(),
5903 "didn't reserve backing store for all of CMS bit map?");
5904 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5905 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5906 _bmWordSize, "inconsistency in bit map sizing");
5907 _bm.set_size(_bmWordSize >> _shifter);
5908
5909 // bm.clear(); // can we rely on getting zero'd memory? verify below
5910 assert(isAllClear(),
5911 "Expected zero'd memory from ReservedSpace constructor");
5912 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5913 "consistency check");
5914 return true;
5915 }
5916
5917 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5918 HeapWord *next_addr, *end_addr, *last_addr;
5919 assert_locked();
5920 assert(covers(mr), "out-of-range error");
5921 // XXX assert that start and end are appropriately aligned
5922 for (next_addr = mr.start(), end_addr = mr.end();
5923 next_addr < end_addr; next_addr = last_addr) {
5924 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5925 last_addr = dirty_region.end();
5926 if (!dirty_region.is_empty()) {
5927 cl->do_MemRegion(dirty_region);
5928 } else {
5929 assert(last_addr == end_addr, "program logic");
5930 return;
5931 }
5932 }
5933 }
5934
5935 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5936 _bm.print_on_error(st, prefix);
5937 }
5938
5939 #ifndef PRODUCT
5940 void CMSBitMap::assert_locked() const {
5941 CMSLockVerifier::assert_locked(lock());
5942 }
5943
5944 bool CMSBitMap::covers(MemRegion mr) const {
5945 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5946 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5947 "size inconsistency");
5948 return (mr.start() >= _bmStartWord) &&
5949 (mr.end() <= endWord());
5950 }
5951
5952 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5953 return (start >= _bmStartWord && (start + size) <= endWord());
5954 }
5955
5956 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5957 // verify that there are no 1 bits in the interval [left, right)
5958 FalseBitMapClosure falseBitMapClosure;
5959 iterate(&falseBitMapClosure, left, right);
5960 }
5961
5962 void CMSBitMap::region_invariant(MemRegion mr)
5963 {
5964 assert_locked();
5965 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5966 assert(!mr.is_empty(), "unexpected empty region");
5967 assert(covers(mr), "mr should be covered by bit map");
5968 // convert address range into offset range
5969 size_t start_ofs = heapWordToOffset(mr.start());
5970 // Make sure that end() is appropriately aligned
5971 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5972 (1 << (_shifter+LogHeapWordSize))),
5973 "Misaligned mr.end()");
5974 size_t end_ofs = heapWordToOffset(mr.end());
5975 assert(end_ofs > start_ofs, "Should mark at least one bit");
5976 }
5977
5978 #endif
5979
5980 bool CMSMarkStack::allocate(size_t size) {
5981 // allocate a stack of the requisite depth
5982 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5983 size * sizeof(oop)));
5984 if (!rs.is_reserved()) {
5985 warning("CMSMarkStack allocation failure");
5986 return false;
5987 }
5988 if (!_virtual_space.initialize(rs, rs.size())) {
5989 warning("CMSMarkStack backing store failure");
5990 return false;
5991 }
5992 assert(_virtual_space.committed_size() == rs.size(),
5993 "didn't reserve backing store for all of CMS stack?");
5994 _base = (oop*)(_virtual_space.low());
5995 _index = 0;
5996 _capacity = size;
5997 NOT_PRODUCT(_max_depth = 0);
5998 return true;
5999 }
6000
6001 // XXX FIX ME !!! In the MT case we come in here holding a
6002 // leaf lock. For printing we need to take a further lock
6003 // which has lower rank. We need to recalibrate the two
6004 // lock-ranks involved in order to be able to print the
6005 // messages below. (Or defer the printing to the caller.
6006 // For now we take the expedient path of just disabling the
6007 // messages for the problematic case.)
6008 void CMSMarkStack::expand() {
6009 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6010 if (_capacity == MarkStackSizeMax) {
6011 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6012 // We print a warning message only once per CMS cycle.
6013 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6014 }
6015 return;
6016 }
6017 // Double capacity if possible
6018 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6019 // Do not give up existing stack until we have managed to
6020 // get the double capacity that we desired.
6021 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6022 new_capacity * sizeof(oop)));
6023 if (rs.is_reserved()) {
6024 // Release the backing store associated with old stack
6025 _virtual_space.release();
6026 // Reinitialize virtual space for new stack
6027 if (!_virtual_space.initialize(rs, rs.size())) {
6028 fatal("Not enough swap for expanded marking stack");
6029 }
6030 _base = (oop*)(_virtual_space.low());
6031 _index = 0;
6032 _capacity = new_capacity;
6033 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6034 // Failed to double capacity, continue;
6035 // we print a detail message only once per CMS cycle.
6036 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6037 SIZE_FORMAT"K",
6038 _capacity / K, new_capacity / K);
6039 }
6040 }
6041
6042
6043 // Closures
6044 // XXX: there seems to be a lot of code duplication here;
6045 // should refactor and consolidate common code.
6046
6047 // This closure is used to mark refs into the CMS generation in
6048 // the CMS bit map. Called at the first checkpoint. This closure
6049 // assumes that we do not need to re-mark dirty cards; if the CMS
6050 // generation on which this is used is not an oldest
6051 // generation then this will lose younger_gen cards!
6052
6053 MarkRefsIntoClosure::MarkRefsIntoClosure(
6054 MemRegion span, CMSBitMap* bitMap):
6055 _span(span),
6056 _bitMap(bitMap)
6057 {
6058 assert(_ref_processor == NULL, "deliberately left NULL");
6059 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6060 }
6061
6062 void MarkRefsIntoClosure::do_oop(oop obj) {
6063 // if p points into _span, then mark corresponding bit in _markBitMap
6064 assert(obj->is_oop(), "expected an oop");
6065 HeapWord* addr = (HeapWord*)obj;
6066 if (_span.contains(addr)) {
6067 // this should be made more efficient
6068 _bitMap->mark(addr);
6069 }
6070 }
6071
6072 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6073 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6074
6075 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6076 MemRegion span, CMSBitMap* bitMap):
6077 _span(span),
6078 _bitMap(bitMap)
6079 {
6080 assert(_ref_processor == NULL, "deliberately left NULL");
6081 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6082 }
6083
6084 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6085 // if p points into _span, then mark corresponding bit in _markBitMap
6086 assert(obj->is_oop(), "expected an oop");
6087 HeapWord* addr = (HeapWord*)obj;
6088 if (_span.contains(addr)) {
6089 // this should be made more efficient
6090 _bitMap->par_mark(addr);
6091 }
6092 }
6093
6094 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6095 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6096
6097 // A variant of the above, used for CMS marking verification.
6098 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6099 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6100 _span(span),
6101 _verification_bm(verification_bm),
6102 _cms_bm(cms_bm)
6103 {
6104 assert(_ref_processor == NULL, "deliberately left NULL");
6105 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6106 }
6107
6108 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6109 // if p points into _span, then mark corresponding bit in _markBitMap
6110 assert(obj->is_oop(), "expected an oop");
6111 HeapWord* addr = (HeapWord*)obj;
6112 if (_span.contains(addr)) {
6113 _verification_bm->mark(addr);
6114 if (!_cms_bm->isMarked(addr)) {
6115 oop(addr)->print();
6116 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6117 fatal("... aborting");
6118 }
6119 }
6120 }
6121
6122 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6123 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6124
6125 //////////////////////////////////////////////////
6126 // MarkRefsIntoAndScanClosure
6127 //////////////////////////////////////////////////
6128
6129 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6130 ReferenceProcessor* rp,
6131 CMSBitMap* bit_map,
6132 CMSBitMap* mod_union_table,
6133 CMSMarkStack* mark_stack,
6134 CMSCollector* collector,
6135 bool should_yield,
6136 bool concurrent_precleaning):
6137 _collector(collector),
6138 _span(span),
6139 _bit_map(bit_map),
6140 _mark_stack(mark_stack),
6141 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6142 mark_stack, concurrent_precleaning),
6143 _yield(should_yield),
6144 _concurrent_precleaning(concurrent_precleaning),
6145 _freelistLock(NULL)
6146 {
6147 _ref_processor = rp;
6148 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6149 }
6150
6151 // This closure is used to mark refs into the CMS generation at the
6152 // second (final) checkpoint, and to scan and transitively follow
6153 // the unmarked oops. It is also used during the concurrent precleaning
6154 // phase while scanning objects on dirty cards in the CMS generation.
6155 // The marks are made in the marking bit map and the marking stack is
6156 // used for keeping the (newly) grey objects during the scan.
6157 // The parallel version (Par_...) appears further below.
6158 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6159 if (obj != NULL) {
6160 assert(obj->is_oop(), "expected an oop");
6161 HeapWord* addr = (HeapWord*)obj;
6162 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6163 assert(_collector->overflow_list_is_empty(),
6164 "overflow list should be empty");
6165 if (_span.contains(addr) &&
6166 !_bit_map->isMarked(addr)) {
6167 // mark bit map (object is now grey)
6168 _bit_map->mark(addr);
6169 // push on marking stack (stack should be empty), and drain the
6170 // stack by applying this closure to the oops in the oops popped
6171 // from the stack (i.e. blacken the grey objects)
6172 bool res = _mark_stack->push(obj);
6173 assert(res, "Should have space to push on empty stack");
6174 do {
6175 oop new_oop = _mark_stack->pop();
6176 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6177 assert(_bit_map->isMarked((HeapWord*)new_oop),
6178 "only grey objects on this stack");
6179 // iterate over the oops in this oop, marking and pushing
6180 // the ones in CMS heap (i.e. in _span).
6181 new_oop->oop_iterate(&_pushAndMarkClosure);
6182 // check if it's time to yield
6183 do_yield_check();
6184 } while (!_mark_stack->isEmpty() ||
6185 (!_concurrent_precleaning && take_from_overflow_list()));
6186 // if marking stack is empty, and we are not doing this
6187 // during precleaning, then check the overflow list
6188 }
6189 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6190 assert(_collector->overflow_list_is_empty(),
6191 "overflow list was drained above");
6192 // We could restore evacuated mark words, if any, used for
6193 // overflow list links here because the overflow list is
6194 // provably empty here. That would reduce the maximum
6195 // size requirements for preserved_{oop,mark}_stack.
6196 // But we'll just postpone it until we are all done
6197 // so we can just stream through.
6198 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6199 _collector->restore_preserved_marks_if_any();
6200 assert(_collector->no_preserved_marks(), "No preserved marks");
6201 }
6202 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6203 "All preserved marks should have been restored above");
6204 }
6205 }
6206
6207 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6208 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6209
6210 void MarkRefsIntoAndScanClosure::do_yield_work() {
6211 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6212 "CMS thread should hold CMS token");
6213 assert_lock_strong(_freelistLock);
6214 assert_lock_strong(_bit_map->lock());
6215 // relinquish the free_list_lock and bitMaplock()
6216 _bit_map->lock()->unlock();
6217 _freelistLock->unlock();
6218 ConcurrentMarkSweepThread::desynchronize(true);
6219 _collector->stopTimer();
6220 if (PrintCMSStatistics != 0) {
6221 _collector->incrementYields();
6222 }
6223
6224 // See the comment in coordinator_yield()
6225 for (unsigned i = 0;
6226 i < CMSYieldSleepCount &&
6227 ConcurrentMarkSweepThread::should_yield() &&
6228 !CMSCollector::foregroundGCIsActive();
6229 ++i) {
6230 os::sleep(Thread::current(), 1, false);
6231 }
6232
6233 ConcurrentMarkSweepThread::synchronize(true);
6234 _freelistLock->lock_without_safepoint_check();
6235 _bit_map->lock()->lock_without_safepoint_check();
6236 _collector->startTimer();
6237 }
6238
6239 ///////////////////////////////////////////////////////////
6240 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6241 // MarkRefsIntoAndScanClosure
6242 ///////////////////////////////////////////////////////////
6243 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6244 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6245 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6246 _span(span),
6247 _bit_map(bit_map),
6248 _work_queue(work_queue),
6249 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6250 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6251 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6252 {
6253 _ref_processor = rp;
6254 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6255 }
6256
6257 // This closure is used to mark refs into the CMS generation at the
6258 // second (final) checkpoint, and to scan and transitively follow
6259 // the unmarked oops. The marks are made in the marking bit map and
6260 // the work_queue is used for keeping the (newly) grey objects during
6261 // the scan phase whence they are also available for stealing by parallel
6262 // threads. Since the marking bit map is shared, updates are
6263 // synchronized (via CAS).
6264 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6265 if (obj != NULL) {
6266 // Ignore mark word because this could be an already marked oop
6267 // that may be chained at the end of the overflow list.
6268 assert(obj->is_oop(true), "expected an oop");
6269 HeapWord* addr = (HeapWord*)obj;
6270 if (_span.contains(addr) &&
6271 !_bit_map->isMarked(addr)) {
6272 // mark bit map (object will become grey):
6273 // It is possible for several threads to be
6274 // trying to "claim" this object concurrently;
6275 // the unique thread that succeeds in marking the
6276 // object first will do the subsequent push on
6277 // to the work queue (or overflow list).
6278 if (_bit_map->par_mark(addr)) {
6279 // push on work_queue (which may not be empty), and trim the
6280 // queue to an appropriate length by applying this closure to
6281 // the oops in the oops popped from the stack (i.e. blacken the
6282 // grey objects)
6283 bool res = _work_queue->push(obj);
6284 assert(res, "Low water mark should be less than capacity?");
6285 trim_queue(_low_water_mark);
6286 } // Else, another thread claimed the object
6287 }
6288 }
6289 }
6290
6291 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6292 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6293
6294 // This closure is used to rescan the marked objects on the dirty cards
6295 // in the mod union table and the card table proper.
6296 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6297 oop p, MemRegion mr) {
6298
6299 size_t size = 0;
6300 HeapWord* addr = (HeapWord*)p;
6301 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6302 assert(_span.contains(addr), "we are scanning the CMS generation");
6303 // check if it's time to yield
6304 if (do_yield_check()) {
6305 // We yielded for some foreground stop-world work,
6306 // and we have been asked to abort this ongoing preclean cycle.
6307 return 0;
6308 }
6309 if (_bitMap->isMarked(addr)) {
6310 // it's marked; is it potentially uninitialized?
6311 if (p->klass_or_null() != NULL) {
6312 // an initialized object; ignore mark word in verification below
6313 // since we are running concurrent with mutators
6314 assert(p->is_oop(true), "should be an oop");
6315 if (p->is_objArray()) {
6316 // objArrays are precisely marked; restrict scanning
6317 // to dirty cards only.
6318 size = CompactibleFreeListSpace::adjustObjectSize(
6319 p->oop_iterate(_scanningClosure, mr));
6320 } else {
6321 // A non-array may have been imprecisely marked; we need
6322 // to scan object in its entirety.
6323 size = CompactibleFreeListSpace::adjustObjectSize(
6324 p->oop_iterate(_scanningClosure));
6325 }
6326 #ifdef ASSERT
6327 size_t direct_size =
6328 CompactibleFreeListSpace::adjustObjectSize(p->size());
6329 assert(size == direct_size, "Inconsistency in size");
6330 assert(size >= 3, "Necessary for Printezis marks to work");
6331 if (!_bitMap->isMarked(addr+1)) {
6332 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6333 } else {
6334 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6335 assert(_bitMap->isMarked(addr+size-1),
6336 "inconsistent Printezis mark");
6337 }
6338 #endif // ASSERT
6339 } else {
6340 // An uninitialized object.
6341 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6342 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6343 size = pointer_delta(nextOneAddr + 1, addr);
6344 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6345 "alignment problem");
6346 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6347 // will dirty the card when the klass pointer is installed in the
6348 // object (signaling the completion of initialization).
6349 }
6350 } else {
6351 // Either a not yet marked object or an uninitialized object
6352 if (p->klass_or_null() == NULL) {
6353 // An uninitialized object, skip to the next card, since
6354 // we may not be able to read its P-bits yet.
6355 assert(size == 0, "Initial value");
6356 } else {
6357 // An object not (yet) reached by marking: we merely need to
6358 // compute its size so as to go look at the next block.
6359 assert(p->is_oop(true), "should be an oop");
6360 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6361 }
6362 }
6363 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6364 return size;
6365 }
6366
6367 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6368 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6369 "CMS thread should hold CMS token");
6370 assert_lock_strong(_freelistLock);
6371 assert_lock_strong(_bitMap->lock());
6372 // relinquish the free_list_lock and bitMaplock()
6373 _bitMap->lock()->unlock();
6374 _freelistLock->unlock();
6375 ConcurrentMarkSweepThread::desynchronize(true);
6376 _collector->stopTimer();
6377 if (PrintCMSStatistics != 0) {
6378 _collector->incrementYields();
6379 }
6380
6381 // See the comment in coordinator_yield()
6382 for (unsigned i = 0; i < CMSYieldSleepCount &&
6383 ConcurrentMarkSweepThread::should_yield() &&
6384 !CMSCollector::foregroundGCIsActive(); ++i) {
6385 os::sleep(Thread::current(), 1, false);
6386 }
6387
6388 ConcurrentMarkSweepThread::synchronize(true);
6389 _freelistLock->lock_without_safepoint_check();
6390 _bitMap->lock()->lock_without_safepoint_check();
6391 _collector->startTimer();
6392 }
6393
6394
6395 //////////////////////////////////////////////////////////////////
6396 // SurvivorSpacePrecleanClosure
6397 //////////////////////////////////////////////////////////////////
6398 // This (single-threaded) closure is used to preclean the oops in
6399 // the survivor spaces.
6400 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6401
6402 HeapWord* addr = (HeapWord*)p;
6403 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6404 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6405 assert(p->klass_or_null() != NULL, "object should be initialized");
6406 // an initialized object; ignore mark word in verification below
6407 // since we are running concurrent with mutators
6408 assert(p->is_oop(true), "should be an oop");
6409 // Note that we do not yield while we iterate over
6410 // the interior oops of p, pushing the relevant ones
6411 // on our marking stack.
6412 size_t size = p->oop_iterate(_scanning_closure);
6413 do_yield_check();
6414 // Observe that below, we do not abandon the preclean
6415 // phase as soon as we should; rather we empty the
6416 // marking stack before returning. This is to satisfy
6417 // some existing assertions. In general, it may be a
6418 // good idea to abort immediately and complete the marking
6419 // from the grey objects at a later time.
6420 while (!_mark_stack->isEmpty()) {
6421 oop new_oop = _mark_stack->pop();
6422 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6423 assert(_bit_map->isMarked((HeapWord*)new_oop),
6424 "only grey objects on this stack");
6425 // iterate over the oops in this oop, marking and pushing
6426 // the ones in CMS heap (i.e. in _span).
6427 new_oop->oop_iterate(_scanning_closure);
6428 // check if it's time to yield
6429 do_yield_check();
6430 }
6431 unsigned int after_count =
6432 GenCollectedHeap::heap()->total_collections();
6433 bool abort = (_before_count != after_count) ||
6434 _collector->should_abort_preclean();
6435 return abort ? 0 : size;
6436 }
6437
6438 void SurvivorSpacePrecleanClosure::do_yield_work() {
6439 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6440 "CMS thread should hold CMS token");
6441 assert_lock_strong(_bit_map->lock());
6442 // Relinquish the bit map lock
6443 _bit_map->lock()->unlock();
6444 ConcurrentMarkSweepThread::desynchronize(true);
6445 _collector->stopTimer();
6446 if (PrintCMSStatistics != 0) {
6447 _collector->incrementYields();
6448 }
6449
6450 // See the comment in coordinator_yield()
6451 for (unsigned i = 0; i < CMSYieldSleepCount &&
6452 ConcurrentMarkSweepThread::should_yield() &&
6453 !CMSCollector::foregroundGCIsActive(); ++i) {
6454 os::sleep(Thread::current(), 1, false);
6455 }
6456
6457 ConcurrentMarkSweepThread::synchronize(true);
6458 _bit_map->lock()->lock_without_safepoint_check();
6459 _collector->startTimer();
6460 }
6461
6462 // This closure is used to rescan the marked objects on the dirty cards
6463 // in the mod union table and the card table proper. In the parallel
6464 // case, although the bitMap is shared, we do a single read so the
6465 // isMarked() query is "safe".
6466 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6467 // Ignore mark word because we are running concurrent with mutators
6468 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6469 HeapWord* addr = (HeapWord*)p;
6470 assert(_span.contains(addr), "we are scanning the CMS generation");
6471 bool is_obj_array = false;
6472 #ifdef ASSERT
6473 if (!_parallel) {
6474 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6475 assert(_collector->overflow_list_is_empty(),
6476 "overflow list should be empty");
6477
6478 }
6479 #endif // ASSERT
6480 if (_bit_map->isMarked(addr)) {
6481 // Obj arrays are precisely marked, non-arrays are not;
6482 // so we scan objArrays precisely and non-arrays in their
6483 // entirety.
6484 if (p->is_objArray()) {
6485 is_obj_array = true;
6486 if (_parallel) {
6487 p->oop_iterate(_par_scan_closure, mr);
6488 } else {
6489 p->oop_iterate(_scan_closure, mr);
6490 }
6491 } else {
6492 if (_parallel) {
6493 p->oop_iterate(_par_scan_closure);
6494 } else {
6495 p->oop_iterate(_scan_closure);
6496 }
6497 }
6498 }
6499 #ifdef ASSERT
6500 if (!_parallel) {
6501 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6502 assert(_collector->overflow_list_is_empty(),
6503 "overflow list should be empty");
6504
6505 }
6506 #endif // ASSERT
6507 return is_obj_array;
6508 }
6509
6510 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6511 MemRegion span,
6512 CMSBitMap* bitMap, CMSMarkStack* markStack,
6513 bool should_yield, bool verifying):
6514 _collector(collector),
6515 _span(span),
6516 _bitMap(bitMap),
6517 _mut(&collector->_modUnionTable),
6518 _markStack(markStack),
6519 _yield(should_yield),
6520 _skipBits(0)
6521 {
6522 assert(_markStack->isEmpty(), "stack should be empty");
6523 _finger = _bitMap->startWord();
6524 _threshold = _finger;
6525 assert(_collector->_restart_addr == NULL, "Sanity check");
6526 assert(_span.contains(_finger), "Out of bounds _finger?");
6527 DEBUG_ONLY(_verifying = verifying;)
6528 }
6529
6530 void MarkFromRootsClosure::reset(HeapWord* addr) {
6531 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6532 assert(_span.contains(addr), "Out of bounds _finger?");
6533 _finger = addr;
6534 _threshold = (HeapWord*)round_to(
6535 (intptr_t)_finger, CardTableModRefBS::card_size);
6536 }
6537
6538 // Should revisit to see if this should be restructured for
6539 // greater efficiency.
6540 bool MarkFromRootsClosure::do_bit(size_t offset) {
6541 if (_skipBits > 0) {
6542 _skipBits--;
6543 return true;
6544 }
6545 // convert offset into a HeapWord*
6546 HeapWord* addr = _bitMap->startWord() + offset;
6547 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6548 "address out of range");
6549 assert(_bitMap->isMarked(addr), "tautology");
6550 if (_bitMap->isMarked(addr+1)) {
6551 // this is an allocated but not yet initialized object
6552 assert(_skipBits == 0, "tautology");
6553 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6554 oop p = oop(addr);
6555 if (p->klass_or_null() == NULL) {
6556 DEBUG_ONLY(if (!_verifying) {)
6557 // We re-dirty the cards on which this object lies and increase
6558 // the _threshold so that we'll come back to scan this object
6559 // during the preclean or remark phase. (CMSCleanOnEnter)
6560 if (CMSCleanOnEnter) {
6561 size_t sz = _collector->block_size_using_printezis_bits(addr);
6562 HeapWord* end_card_addr = (HeapWord*)round_to(
6563 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6564 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6565 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6566 // Bump _threshold to end_card_addr; note that
6567 // _threshold cannot possibly exceed end_card_addr, anyhow.
6568 // This prevents future clearing of the card as the scan proceeds
6569 // to the right.
6570 assert(_threshold <= end_card_addr,
6571 "Because we are just scanning into this object");
6572 if (_threshold < end_card_addr) {
6573 _threshold = end_card_addr;
6574 }
6575 if (p->klass_or_null() != NULL) {
6576 // Redirty the range of cards...
6577 _mut->mark_range(redirty_range);
6578 } // ...else the setting of klass will dirty the card anyway.
6579 }
6580 DEBUG_ONLY(})
6581 return true;
6582 }
6583 }
6584 scanOopsInOop(addr);
6585 return true;
6586 }
6587
6588 // We take a break if we've been at this for a while,
6589 // so as to avoid monopolizing the locks involved.
6590 void MarkFromRootsClosure::do_yield_work() {
6591 // First give up the locks, then yield, then re-lock
6592 // We should probably use a constructor/destructor idiom to
6593 // do this unlock/lock or modify the MutexUnlocker class to
6594 // serve our purpose. XXX
6595 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6596 "CMS thread should hold CMS token");
6597 assert_lock_strong(_bitMap->lock());
6598 _bitMap->lock()->unlock();
6599 ConcurrentMarkSweepThread::desynchronize(true);
6600 _collector->stopTimer();
6601 if (PrintCMSStatistics != 0) {
6602 _collector->incrementYields();
6603 }
6604
6605 // See the comment in coordinator_yield()
6606 for (unsigned i = 0; i < CMSYieldSleepCount &&
6607 ConcurrentMarkSweepThread::should_yield() &&
6608 !CMSCollector::foregroundGCIsActive(); ++i) {
6609 os::sleep(Thread::current(), 1, false);
6610 }
6611
6612 ConcurrentMarkSweepThread::synchronize(true);
6613 _bitMap->lock()->lock_without_safepoint_check();
6614 _collector->startTimer();
6615 }
6616
6617 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6618 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6619 assert(_markStack->isEmpty(),
6620 "should drain stack to limit stack usage");
6621 // convert ptr to an oop preparatory to scanning
6622 oop obj = oop(ptr);
6623 // Ignore mark word in verification below, since we
6624 // may be running concurrent with mutators.
6625 assert(obj->is_oop(true), "should be an oop");
6626 assert(_finger <= ptr, "_finger runneth ahead");
6627 // advance the finger to right end of this object
6628 _finger = ptr + obj->size();
6629 assert(_finger > ptr, "we just incremented it above");
6630 // On large heaps, it may take us some time to get through
6631 // the marking phase. During
6632 // this time it's possible that a lot of mutations have
6633 // accumulated in the card table and the mod union table --
6634 // these mutation records are redundant until we have
6635 // actually traced into the corresponding card.
6636 // Here, we check whether advancing the finger would make
6637 // us cross into a new card, and if so clear corresponding
6638 // cards in the MUT (preclean them in the card-table in the
6639 // future).
6640
6641 DEBUG_ONLY(if (!_verifying) {)
6642 // The clean-on-enter optimization is disabled by default,
6643 // until we fix 6178663.
6644 if (CMSCleanOnEnter && (_finger > _threshold)) {
6645 // [_threshold, _finger) represents the interval
6646 // of cards to be cleared in MUT (or precleaned in card table).
6647 // The set of cards to be cleared is all those that overlap
6648 // with the interval [_threshold, _finger); note that
6649 // _threshold is always kept card-aligned but _finger isn't
6650 // always card-aligned.
6651 HeapWord* old_threshold = _threshold;
6652 assert(old_threshold == (HeapWord*)round_to(
6653 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6654 "_threshold should always be card-aligned");
6655 _threshold = (HeapWord*)round_to(
6656 (intptr_t)_finger, CardTableModRefBS::card_size);
6657 MemRegion mr(old_threshold, _threshold);
6658 assert(!mr.is_empty(), "Control point invariant");
6659 assert(_span.contains(mr), "Should clear within span");
6660 _mut->clear_range(mr);
6661 }
6662 DEBUG_ONLY(})
6663 // Note: the finger doesn't advance while we drain
6664 // the stack below.
6665 PushOrMarkClosure pushOrMarkClosure(_collector,
6666 _span, _bitMap, _markStack,
6667 _finger, this);
6668 bool res = _markStack->push(obj);
6669 assert(res, "Empty non-zero size stack should have space for single push");
6670 while (!_markStack->isEmpty()) {
6671 oop new_oop = _markStack->pop();
6672 // Skip verifying header mark word below because we are
6673 // running concurrent with mutators.
6674 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6675 // now scan this oop's oops
6676 new_oop->oop_iterate(&pushOrMarkClosure);
6677 do_yield_check();
6678 }
6679 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6680 }
6681
6682 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6683 CMSCollector* collector, MemRegion span,
6684 CMSBitMap* bit_map,
6685 OopTaskQueue* work_queue,
6686 CMSMarkStack* overflow_stack):
6687 _collector(collector),
6688 _whole_span(collector->_span),
6689 _span(span),
6690 _bit_map(bit_map),
6691 _mut(&collector->_modUnionTable),
6692 _work_queue(work_queue),
6693 _overflow_stack(overflow_stack),
6694 _skip_bits(0),
6695 _task(task)
6696 {
6697 assert(_work_queue->size() == 0, "work_queue should be empty");
6698 _finger = span.start();
6699 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6700 assert(_span.contains(_finger), "Out of bounds _finger?");
6701 }
6702
6703 // Should revisit to see if this should be restructured for
6704 // greater efficiency.
6705 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6706 if (_skip_bits > 0) {
6707 _skip_bits--;
6708 return true;
6709 }
6710 // convert offset into a HeapWord*
6711 HeapWord* addr = _bit_map->startWord() + offset;
6712 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6713 "address out of range");
6714 assert(_bit_map->isMarked(addr), "tautology");
6715 if (_bit_map->isMarked(addr+1)) {
6716 // this is an allocated object that might not yet be initialized
6717 assert(_skip_bits == 0, "tautology");
6718 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6719 oop p = oop(addr);
6720 if (p->klass_or_null() == NULL) {
6721 // in the case of Clean-on-Enter optimization, redirty card
6722 // and avoid clearing card by increasing the threshold.
6723 return true;
6724 }
6725 }
6726 scan_oops_in_oop(addr);
6727 return true;
6728 }
6729
6730 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6731 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6732 // Should we assert that our work queue is empty or
6733 // below some drain limit?
6734 assert(_work_queue->size() == 0,
6735 "should drain stack to limit stack usage");
6736 // convert ptr to an oop preparatory to scanning
6737 oop obj = oop(ptr);
6738 // Ignore mark word in verification below, since we
6739 // may be running concurrent with mutators.
6740 assert(obj->is_oop(true), "should be an oop");
6741 assert(_finger <= ptr, "_finger runneth ahead");
6742 // advance the finger to right end of this object
6743 _finger = ptr + obj->size();
6744 assert(_finger > ptr, "we just incremented it above");
6745 // On large heaps, it may take us some time to get through
6746 // the marking phase. During
6747 // this time it's possible that a lot of mutations have
6748 // accumulated in the card table and the mod union table --
6749 // these mutation records are redundant until we have
6750 // actually traced into the corresponding card.
6751 // Here, we check whether advancing the finger would make
6752 // us cross into a new card, and if so clear corresponding
6753 // cards in the MUT (preclean them in the card-table in the
6754 // future).
6755
6756 // The clean-on-enter optimization is disabled by default,
6757 // until we fix 6178663.
6758 if (CMSCleanOnEnter && (_finger > _threshold)) {
6759 // [_threshold, _finger) represents the interval
6760 // of cards to be cleared in MUT (or precleaned in card table).
6761 // The set of cards to be cleared is all those that overlap
6762 // with the interval [_threshold, _finger); note that
6763 // _threshold is always kept card-aligned but _finger isn't
6764 // always card-aligned.
6765 HeapWord* old_threshold = _threshold;
6766 assert(old_threshold == (HeapWord*)round_to(
6767 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6768 "_threshold should always be card-aligned");
6769 _threshold = (HeapWord*)round_to(
6770 (intptr_t)_finger, CardTableModRefBS::card_size);
6771 MemRegion mr(old_threshold, _threshold);
6772 assert(!mr.is_empty(), "Control point invariant");
6773 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6774 _mut->clear_range(mr);
6775 }
6776
6777 // Note: the local finger doesn't advance while we drain
6778 // the stack below, but the global finger sure can and will.
6779 HeapWord** gfa = _task->global_finger_addr();
6780 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6781 _span, _bit_map,
6782 _work_queue,
6783 _overflow_stack,
6784 _finger,
6785 gfa, this);
6786 bool res = _work_queue->push(obj); // overflow could occur here
6787 assert(res, "Will hold once we use workqueues");
6788 while (true) {
6789 oop new_oop;
6790 if (!_work_queue->pop_local(new_oop)) {
6791 // We emptied our work_queue; check if there's stuff that can
6792 // be gotten from the overflow stack.
6793 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6794 _overflow_stack, _work_queue)) {
6795 do_yield_check();
6796 continue;
6797 } else { // done
6798 break;
6799 }
6800 }
6801 // Skip verifying header mark word below because we are
6802 // running concurrent with mutators.
6803 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6804 // now scan this oop's oops
6805 new_oop->oop_iterate(&pushOrMarkClosure);
6806 do_yield_check();
6807 }
6808 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6809 }
6810
6811 // Yield in response to a request from VM Thread or
6812 // from mutators.
6813 void Par_MarkFromRootsClosure::do_yield_work() {
6814 assert(_task != NULL, "sanity");
6815 _task->yield();
6816 }
6817
6818 // A variant of the above used for verifying CMS marking work.
6819 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6820 MemRegion span,
6821 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6822 CMSMarkStack* mark_stack):
6823 _collector(collector),
6824 _span(span),
6825 _verification_bm(verification_bm),
6826 _cms_bm(cms_bm),
6827 _mark_stack(mark_stack),
6828 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6829 mark_stack)
6830 {
6831 assert(_mark_stack->isEmpty(), "stack should be empty");
6832 _finger = _verification_bm->startWord();
6833 assert(_collector->_restart_addr == NULL, "Sanity check");
6834 assert(_span.contains(_finger), "Out of bounds _finger?");
6835 }
6836
6837 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6838 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6839 assert(_span.contains(addr), "Out of bounds _finger?");
6840 _finger = addr;
6841 }
6842
6843 // Should revisit to see if this should be restructured for
6844 // greater efficiency.
6845 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6846 // convert offset into a HeapWord*
6847 HeapWord* addr = _verification_bm->startWord() + offset;
6848 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6849 "address out of range");
6850 assert(_verification_bm->isMarked(addr), "tautology");
6851 assert(_cms_bm->isMarked(addr), "tautology");
6852
6853 assert(_mark_stack->isEmpty(),
6854 "should drain stack to limit stack usage");
6855 // convert addr to an oop preparatory to scanning
6856 oop obj = oop(addr);
6857 assert(obj->is_oop(), "should be an oop");
6858 assert(_finger <= addr, "_finger runneth ahead");
6859 // advance the finger to right end of this object
6860 _finger = addr + obj->size();
6861 assert(_finger > addr, "we just incremented it above");
6862 // Note: the finger doesn't advance while we drain
6863 // the stack below.
6864 bool res = _mark_stack->push(obj);
6865 assert(res, "Empty non-zero size stack should have space for single push");
6866 while (!_mark_stack->isEmpty()) {
6867 oop new_oop = _mark_stack->pop();
6868 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6869 // now scan this oop's oops
6870 new_oop->oop_iterate(&_pam_verify_closure);
6871 }
6872 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6873 return true;
6874 }
6875
6876 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6877 CMSCollector* collector, MemRegion span,
6878 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6879 CMSMarkStack* mark_stack):
6880 MetadataAwareOopClosure(collector->ref_processor()),
6881 _collector(collector),
6882 _span(span),
6883 _verification_bm(verification_bm),
6884 _cms_bm(cms_bm),
6885 _mark_stack(mark_stack)
6886 { }
6887
6888 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6889 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6890
6891 // Upon stack overflow, we discard (part of) the stack,
6892 // remembering the least address amongst those discarded
6893 // in CMSCollector's _restart_address.
6894 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6895 // Remember the least grey address discarded
6896 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6897 _collector->lower_restart_addr(ra);
6898 _mark_stack->reset(); // discard stack contents
6899 _mark_stack->expand(); // expand the stack if possible
6900 }
6901
6902 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6903 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6904 HeapWord* addr = (HeapWord*)obj;
6905 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6906 // Oop lies in _span and isn't yet grey or black
6907 _verification_bm->mark(addr); // now grey
6908 if (!_cms_bm->isMarked(addr)) {
6909 oop(addr)->print();
6910 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6911 p2i(addr));
6912 fatal("... aborting");
6913 }
6914
6915 if (!_mark_stack->push(obj)) { // stack overflow
6916 if (PrintCMSStatistics != 0) {
6917 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6918 SIZE_FORMAT, _mark_stack->capacity());
6919 }
6920 assert(_mark_stack->isFull(), "Else push should have succeeded");
6921 handle_stack_overflow(addr);
6922 }
6923 // anything including and to the right of _finger
6924 // will be scanned as we iterate over the remainder of the
6925 // bit map
6926 }
6927 }
6928
6929 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6930 MemRegion span,
6931 CMSBitMap* bitMap, CMSMarkStack* markStack,
6932 HeapWord* finger, MarkFromRootsClosure* parent) :
6933 MetadataAwareOopClosure(collector->ref_processor()),
6934 _collector(collector),
6935 _span(span),
6936 _bitMap(bitMap),
6937 _markStack(markStack),
6938 _finger(finger),
6939 _parent(parent)
6940 { }
6941
6942 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6943 MemRegion span,
6944 CMSBitMap* bit_map,
6945 OopTaskQueue* work_queue,
6946 CMSMarkStack* overflow_stack,
6947 HeapWord* finger,
6948 HeapWord** global_finger_addr,
6949 Par_MarkFromRootsClosure* parent) :
6950 MetadataAwareOopClosure(collector->ref_processor()),
6951 _collector(collector),
6952 _whole_span(collector->_span),
6953 _span(span),
6954 _bit_map(bit_map),
6955 _work_queue(work_queue),
6956 _overflow_stack(overflow_stack),
6957 _finger(finger),
6958 _global_finger_addr(global_finger_addr),
6959 _parent(parent)
6960 { }
6961
6962 // Assumes thread-safe access by callers, who are
6963 // responsible for mutual exclusion.
6964 void CMSCollector::lower_restart_addr(HeapWord* low) {
6965 assert(_span.contains(low), "Out of bounds addr");
6966 if (_restart_addr == NULL) {
6967 _restart_addr = low;
6968 } else {
6969 _restart_addr = MIN2(_restart_addr, low);
6970 }
6971 }
6972
6973 // Upon stack overflow, we discard (part of) the stack,
6974 // remembering the least address amongst those discarded
6975 // in CMSCollector's _restart_address.
6976 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6977 // Remember the least grey address discarded
6978 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6979 _collector->lower_restart_addr(ra);
6980 _markStack->reset(); // discard stack contents
6981 _markStack->expand(); // expand the stack if possible
6982 }
6983
6984 // Upon stack overflow, we discard (part of) the stack,
6985 // remembering the least address amongst those discarded
6986 // in CMSCollector's _restart_address.
6987 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6988 // We need to do this under a mutex to prevent other
6989 // workers from interfering with the work done below.
6990 MutexLockerEx ml(_overflow_stack->par_lock(),
6991 Mutex::_no_safepoint_check_flag);
6992 // Remember the least grey address discarded
6993 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6994 _collector->lower_restart_addr(ra);
6995 _overflow_stack->reset(); // discard stack contents
6996 _overflow_stack->expand(); // expand the stack if possible
6997 }
6998
6999 void PushOrMarkClosure::do_oop(oop obj) {
7000 // Ignore mark word because we are running concurrent with mutators.
7001 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7002 HeapWord* addr = (HeapWord*)obj;
7003 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7004 // Oop lies in _span and isn't yet grey or black
7005 _bitMap->mark(addr); // now grey
7006 if (addr < _finger) {
7007 // the bit map iteration has already either passed, or
7008 // sampled, this bit in the bit map; we'll need to
7009 // use the marking stack to scan this oop's oops.
7010 bool simulate_overflow = false;
7011 NOT_PRODUCT(
7012 if (CMSMarkStackOverflowALot &&
7013 _collector->simulate_overflow()) {
7014 // simulate a stack overflow
7015 simulate_overflow = true;
7016 }
7017 )
7018 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7019 if (PrintCMSStatistics != 0) {
7020 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7021 SIZE_FORMAT, _markStack->capacity());
7022 }
7023 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7024 handle_stack_overflow(addr);
7025 }
7026 }
7027 // anything including and to the right of _finger
7028 // will be scanned as we iterate over the remainder of the
7029 // bit map
7030 do_yield_check();
7031 }
7032 }
7033
7034 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7035 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7036
7037 void Par_PushOrMarkClosure::do_oop(oop obj) {
7038 // Ignore mark word because we are running concurrent with mutators.
7039 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7040 HeapWord* addr = (HeapWord*)obj;
7041 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7042 // Oop lies in _span and isn't yet grey or black
7043 // We read the global_finger (volatile read) strictly after marking oop
7044 bool res = _bit_map->par_mark(addr); // now grey
7045 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7046 // Should we push this marked oop on our stack?
7047 // -- if someone else marked it, nothing to do
7048 // -- if target oop is above global finger nothing to do
7049 // -- if target oop is in chunk and above local finger
7050 // then nothing to do
7051 // -- else push on work queue
7052 if ( !res // someone else marked it, they will deal with it
7053 || (addr >= *gfa) // will be scanned in a later task
7054 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7055 return;
7056 }
7057 // the bit map iteration has already either passed, or
7058 // sampled, this bit in the bit map; we'll need to
7059 // use the marking stack to scan this oop's oops.
7060 bool simulate_overflow = false;
7061 NOT_PRODUCT(
7062 if (CMSMarkStackOverflowALot &&
7063 _collector->simulate_overflow()) {
7064 // simulate a stack overflow
7065 simulate_overflow = true;
7066 }
7067 )
7068 if (simulate_overflow ||
7069 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7070 // stack overflow
7071 if (PrintCMSStatistics != 0) {
7072 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7073 SIZE_FORMAT, _overflow_stack->capacity());
7074 }
7075 // We cannot assert that the overflow stack is full because
7076 // it may have been emptied since.
7077 assert(simulate_overflow ||
7078 _work_queue->size() == _work_queue->max_elems(),
7079 "Else push should have succeeded");
7080 handle_stack_overflow(addr);
7081 }
7082 do_yield_check();
7083 }
7084 }
7085
7086 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7087 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7088
7089 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7090 MemRegion span,
7091 ReferenceProcessor* rp,
7092 CMSBitMap* bit_map,
7093 CMSBitMap* mod_union_table,
7094 CMSMarkStack* mark_stack,
7095 bool concurrent_precleaning):
7096 MetadataAwareOopClosure(rp),
7097 _collector(collector),
7098 _span(span),
7099 _bit_map(bit_map),
7100 _mod_union_table(mod_union_table),
7101 _mark_stack(mark_stack),
7102 _concurrent_precleaning(concurrent_precleaning)
7103 {
7104 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7105 }
7106
7107 // Grey object rescan during pre-cleaning and second checkpoint phases --
7108 // the non-parallel version (the parallel version appears further below.)
7109 void PushAndMarkClosure::do_oop(oop obj) {
7110 // Ignore mark word verification. If during concurrent precleaning,
7111 // the object monitor may be locked. If during the checkpoint
7112 // phases, the object may already have been reached by a different
7113 // path and may be at the end of the global overflow list (so
7114 // the mark word may be NULL).
7115 assert(obj->is_oop_or_null(true /* ignore mark word */),
7116 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7117 HeapWord* addr = (HeapWord*)obj;
7118 // Check if oop points into the CMS generation
7119 // and is not marked
7120 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7121 // a white object ...
7122 _bit_map->mark(addr); // ... now grey
7123 // push on the marking stack (grey set)
7124 bool simulate_overflow = false;
7125 NOT_PRODUCT(
7126 if (CMSMarkStackOverflowALot &&
7127 _collector->simulate_overflow()) {
7128 // simulate a stack overflow
7129 simulate_overflow = true;
7130 }
7131 )
7132 if (simulate_overflow || !_mark_stack->push(obj)) {
7133 if (_concurrent_precleaning) {
7134 // During precleaning we can just dirty the appropriate card(s)
7135 // in the mod union table, thus ensuring that the object remains
7136 // in the grey set and continue. In the case of object arrays
7137 // we need to dirty all of the cards that the object spans,
7138 // since the rescan of object arrays will be limited to the
7139 // dirty cards.
7140 // Note that no one can be interfering with us in this action
7141 // of dirtying the mod union table, so no locking or atomics
7142 // are required.
7143 if (obj->is_objArray()) {
7144 size_t sz = obj->size();
7145 HeapWord* end_card_addr = (HeapWord*)round_to(
7146 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7147 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7148 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7149 _mod_union_table->mark_range(redirty_range);
7150 } else {
7151 _mod_union_table->mark(addr);
7152 }
7153 _collector->_ser_pmc_preclean_ovflw++;
7154 } else {
7155 // During the remark phase, we need to remember this oop
7156 // in the overflow list.
7157 _collector->push_on_overflow_list(obj);
7158 _collector->_ser_pmc_remark_ovflw++;
7159 }
7160 }
7161 }
7162 }
7163
7164 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7165 MemRegion span,
7166 ReferenceProcessor* rp,
7167 CMSBitMap* bit_map,
7168 OopTaskQueue* work_queue):
7169 MetadataAwareOopClosure(rp),
7170 _collector(collector),
7171 _span(span),
7172 _bit_map(bit_map),
7173 _work_queue(work_queue)
7174 {
7175 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7176 }
7177
7178 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7179 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7180
7181 // Grey object rescan during second checkpoint phase --
7182 // the parallel version.
7183 void Par_PushAndMarkClosure::do_oop(oop obj) {
7184 // In the assert below, we ignore the mark word because
7185 // this oop may point to an already visited object that is
7186 // on the overflow stack (in which case the mark word has
7187 // been hijacked for chaining into the overflow stack --
7188 // if this is the last object in the overflow stack then
7189 // its mark word will be NULL). Because this object may
7190 // have been subsequently popped off the global overflow
7191 // stack, and the mark word possibly restored to the prototypical
7192 // value, by the time we get to examined this failing assert in
7193 // the debugger, is_oop_or_null(false) may subsequently start
7194 // to hold.
7195 assert(obj->is_oop_or_null(true),
7196 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7197 HeapWord* addr = (HeapWord*)obj;
7198 // Check if oop points into the CMS generation
7199 // and is not marked
7200 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7201 // a white object ...
7202 // If we manage to "claim" the object, by being the
7203 // first thread to mark it, then we push it on our
7204 // marking stack
7205 if (_bit_map->par_mark(addr)) { // ... now grey
7206 // push on work queue (grey set)
7207 bool simulate_overflow = false;
7208 NOT_PRODUCT(
7209 if (CMSMarkStackOverflowALot &&
7210 _collector->par_simulate_overflow()) {
7211 // simulate a stack overflow
7212 simulate_overflow = true;
7213 }
7214 )
7215 if (simulate_overflow || !_work_queue->push(obj)) {
7216 _collector->par_push_on_overflow_list(obj);
7217 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7218 }
7219 } // Else, some other thread got there first
7220 }
7221 }
7222
7223 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7224 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7225
7226 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7227 Mutex* bml = _collector->bitMapLock();
7228 assert_lock_strong(bml);
7229 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7230 "CMS thread should hold CMS token");
7231
7232 bml->unlock();
7233 ConcurrentMarkSweepThread::desynchronize(true);
7234
7235 _collector->stopTimer();
7236 if (PrintCMSStatistics != 0) {
7237 _collector->incrementYields();
7238 }
7239
7240 // See the comment in coordinator_yield()
7241 for (unsigned i = 0; i < CMSYieldSleepCount &&
7242 ConcurrentMarkSweepThread::should_yield() &&
7243 !CMSCollector::foregroundGCIsActive(); ++i) {
7244 os::sleep(Thread::current(), 1, false);
7245 }
7246
7247 ConcurrentMarkSweepThread::synchronize(true);
7248 bml->lock();
7249
7250 _collector->startTimer();
7251 }
7252
7253 bool CMSPrecleanRefsYieldClosure::should_return() {
7254 if (ConcurrentMarkSweepThread::should_yield()) {
7255 do_yield_work();
7256 }
7257 return _collector->foregroundGCIsActive();
7258 }
7259
7260 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7261 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7262 "mr should be aligned to start at a card boundary");
7263 // We'd like to assert:
7264 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7265 // "mr should be a range of cards");
7266 // However, that would be too strong in one case -- the last
7267 // partition ends at _unallocated_block which, in general, can be
7268 // an arbitrary boundary, not necessarily card aligned.
7269 if (PrintCMSStatistics != 0) {
7270 _num_dirty_cards +=
7271 mr.word_size()/CardTableModRefBS::card_size_in_words;
7272 }
7273 _space->object_iterate_mem(mr, &_scan_cl);
7274 }
7275
7276 SweepClosure::SweepClosure(CMSCollector* collector,
7277 ConcurrentMarkSweepGeneration* g,
7278 CMSBitMap* bitMap, bool should_yield) :
7279 _collector(collector),
7280 _g(g),
7281 _sp(g->cmsSpace()),
7282 _limit(_sp->sweep_limit()),
7283 _freelistLock(_sp->freelistLock()),
7284 _bitMap(bitMap),
7285 _yield(should_yield),
7286 _inFreeRange(false), // No free range at beginning of sweep
7287 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7288 _lastFreeRangeCoalesced(false),
7289 _freeFinger(g->used_region().start())
7290 {
7291 NOT_PRODUCT(
7292 _numObjectsFreed = 0;
7293 _numWordsFreed = 0;
7294 _numObjectsLive = 0;
7295 _numWordsLive = 0;
7296 _numObjectsAlreadyFree = 0;
7297 _numWordsAlreadyFree = 0;
7298 _last_fc = NULL;
7299
7300 _sp->initializeIndexedFreeListArrayReturnedBytes();
7301 _sp->dictionary()->initialize_dict_returned_bytes();
7302 )
7303 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7304 "sweep _limit out of bounds");
7305 if (CMSTraceSweeper) {
7306 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7307 p2i(_limit));
7308 }
7309 }
7310
7311 void SweepClosure::print_on(outputStream* st) const {
7312 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7313 p2i(_sp->bottom()), p2i(_sp->end()));
7314 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7315 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7316 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7317 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7318 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7319 }
7320
7321 #ifndef PRODUCT
7322 // Assertion checking only: no useful work in product mode --
7323 // however, if any of the flags below become product flags,
7324 // you may need to review this code to see if it needs to be
7325 // enabled in product mode.
7326 SweepClosure::~SweepClosure() {
7327 assert_lock_strong(_freelistLock);
7328 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7329 "sweep _limit out of bounds");
7330 if (inFreeRange()) {
7331 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7332 print();
7333 ShouldNotReachHere();
7334 }
7335 if (Verbose && PrintGC) {
7336 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7337 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7338 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7339 SIZE_FORMAT" bytes "
7340 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7341 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7342 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7343 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7344 * sizeof(HeapWord);
7345 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7346
7347 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7348 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7349 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7350 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7351 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
7352 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7353 indexListReturnedBytes);
7354 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7355 dict_returned_bytes);
7356 }
7357 }
7358 if (CMSTraceSweeper) {
7359 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7360 p2i(_limit));
7361 }
7362 }
7363 #endif // PRODUCT
7364
7365 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7366 bool freeRangeInFreeLists) {
7367 if (CMSTraceSweeper) {
7368 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7369 p2i(freeFinger), freeRangeInFreeLists);
7370 }
7371 assert(!inFreeRange(), "Trampling existing free range");
7372 set_inFreeRange(true);
7373 set_lastFreeRangeCoalesced(false);
7374
7375 set_freeFinger(freeFinger);
7376 set_freeRangeInFreeLists(freeRangeInFreeLists);
7377 if (CMSTestInFreeList) {
7378 if (freeRangeInFreeLists) {
7379 FreeChunk* fc = (FreeChunk*) freeFinger;
7380 assert(fc->is_free(), "A chunk on the free list should be free.");
7381 assert(fc->size() > 0, "Free range should have a size");
7382 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7383 }
7384 }
7385 }
7386
7387 // Note that the sweeper runs concurrently with mutators. Thus,
7388 // it is possible for direct allocation in this generation to happen
7389 // in the middle of the sweep. Note that the sweeper also coalesces
7390 // contiguous free blocks. Thus, unless the sweeper and the allocator
7391 // synchronize appropriately freshly allocated blocks may get swept up.
7392 // This is accomplished by the sweeper locking the free lists while
7393 // it is sweeping. Thus blocks that are determined to be free are
7394 // indeed free. There is however one additional complication:
7395 // blocks that have been allocated since the final checkpoint and
7396 // mark, will not have been marked and so would be treated as
7397 // unreachable and swept up. To prevent this, the allocator marks
7398 // the bit map when allocating during the sweep phase. This leads,
7399 // however, to a further complication -- objects may have been allocated
7400 // but not yet initialized -- in the sense that the header isn't yet
7401 // installed. The sweeper can not then determine the size of the block
7402 // in order to skip over it. To deal with this case, we use a technique
7403 // (due to Printezis) to encode such uninitialized block sizes in the
7404 // bit map. Since the bit map uses a bit per every HeapWord, but the
7405 // CMS generation has a minimum object size of 3 HeapWords, it follows
7406 // that "normal marks" won't be adjacent in the bit map (there will
7407 // always be at least two 0 bits between successive 1 bits). We make use
7408 // of these "unused" bits to represent uninitialized blocks -- the bit
7409 // corresponding to the start of the uninitialized object and the next
7410 // bit are both set. Finally, a 1 bit marks the end of the object that
7411 // started with the two consecutive 1 bits to indicate its potentially
7412 // uninitialized state.
7413
7414 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7415 FreeChunk* fc = (FreeChunk*)addr;
7416 size_t res;
7417
7418 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7419 // than "addr == _limit" because although _limit was a block boundary when
7420 // we started the sweep, it may no longer be one because heap expansion
7421 // may have caused us to coalesce the block ending at the address _limit
7422 // with a newly expanded chunk (this happens when _limit was set to the
7423 // previous _end of the space), so we may have stepped past _limit:
7424 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7425 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7426 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7427 "sweep _limit out of bounds");
7428 assert(addr < _sp->end(), "addr out of bounds");
7429 // Flush any free range we might be holding as a single
7430 // coalesced chunk to the appropriate free list.
7431 if (inFreeRange()) {
7432 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7433 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger())));
7434 flush_cur_free_chunk(freeFinger(),
7435 pointer_delta(addr, freeFinger()));
7436 if (CMSTraceSweeper) {
7437 gclog_or_tty->print("Sweep: last chunk: ");
7438 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") "
7439 "[coalesced:%d]\n",
7440 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7441 lastFreeRangeCoalesced() ? 1 : 0);
7442 }
7443 }
7444
7445 // help the iterator loop finish
7446 return pointer_delta(_sp->end(), addr);
7447 }
7448
7449 assert(addr < _limit, "sweep invariant");
7450 // check if we should yield
7451 do_yield_check(addr);
7452 if (fc->is_free()) {
7453 // Chunk that is already free
7454 res = fc->size();
7455 do_already_free_chunk(fc);
7456 debug_only(_sp->verifyFreeLists());
7457 // If we flush the chunk at hand in lookahead_and_flush()
7458 // and it's coalesced with a preceding chunk, then the
7459 // process of "mangling" the payload of the coalesced block
7460 // will cause erasure of the size information from the
7461 // (erstwhile) header of all the coalesced blocks but the
7462 // first, so the first disjunct in the assert will not hold
7463 // in that specific case (in which case the second disjunct
7464 // will hold).
7465 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7466 "Otherwise the size info doesn't change at this step");
7467 NOT_PRODUCT(
7468 _numObjectsAlreadyFree++;
7469 _numWordsAlreadyFree += res;
7470 )
7471 NOT_PRODUCT(_last_fc = fc;)
7472 } else if (!_bitMap->isMarked(addr)) {
7473 // Chunk is fresh garbage
7474 res = do_garbage_chunk(fc);
7475 debug_only(_sp->verifyFreeLists());
7476 NOT_PRODUCT(
7477 _numObjectsFreed++;
7478 _numWordsFreed += res;
7479 )
7480 } else {
7481 // Chunk that is alive.
7482 res = do_live_chunk(fc);
7483 debug_only(_sp->verifyFreeLists());
7484 NOT_PRODUCT(
7485 _numObjectsLive++;
7486 _numWordsLive += res;
7487 )
7488 }
7489 return res;
7490 }
7491
7492 // For the smart allocation, record following
7493 // split deaths - a free chunk is removed from its free list because
7494 // it is being split into two or more chunks.
7495 // split birth - a free chunk is being added to its free list because
7496 // a larger free chunk has been split and resulted in this free chunk.
7497 // coal death - a free chunk is being removed from its free list because
7498 // it is being coalesced into a large free chunk.
7499 // coal birth - a free chunk is being added to its free list because
7500 // it was created when two or more free chunks where coalesced into
7501 // this free chunk.
7502 //
7503 // These statistics are used to determine the desired number of free
7504 // chunks of a given size. The desired number is chosen to be relative
7505 // to the end of a CMS sweep. The desired number at the end of a sweep
7506 // is the
7507 // count-at-end-of-previous-sweep (an amount that was enough)
7508 // - count-at-beginning-of-current-sweep (the excess)
7509 // + split-births (gains in this size during interval)
7510 // - split-deaths (demands on this size during interval)
7511 // where the interval is from the end of one sweep to the end of the
7512 // next.
7513 //
7514 // When sweeping the sweeper maintains an accumulated chunk which is
7515 // the chunk that is made up of chunks that have been coalesced. That
7516 // will be termed the left-hand chunk. A new chunk of garbage that
7517 // is being considered for coalescing will be referred to as the
7518 // right-hand chunk.
7519 //
7520 // When making a decision on whether to coalesce a right-hand chunk with
7521 // the current left-hand chunk, the current count vs. the desired count
7522 // of the left-hand chunk is considered. Also if the right-hand chunk
7523 // is near the large chunk at the end of the heap (see
7524 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7525 // left-hand chunk is coalesced.
7526 //
7527 // When making a decision about whether to split a chunk, the desired count
7528 // vs. the current count of the candidate to be split is also considered.
7529 // If the candidate is underpopulated (currently fewer chunks than desired)
7530 // a chunk of an overpopulated (currently more chunks than desired) size may
7531 // be chosen. The "hint" associated with a free list, if non-null, points
7532 // to a free list which may be overpopulated.
7533 //
7534
7535 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7536 const size_t size = fc->size();
7537 // Chunks that cannot be coalesced are not in the
7538 // free lists.
7539 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7540 assert(_sp->verify_chunk_in_free_list(fc),
7541 "free chunk should be in free lists");
7542 }
7543 // a chunk that is already free, should not have been
7544 // marked in the bit map
7545 HeapWord* const addr = (HeapWord*) fc;
7546 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7547 // Verify that the bit map has no bits marked between
7548 // addr and purported end of this block.
7549 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7550
7551 // Some chunks cannot be coalesced under any circumstances.
7552 // See the definition of cantCoalesce().
7553 if (!fc->cantCoalesce()) {
7554 // This chunk can potentially be coalesced.
7555 if (_sp->adaptive_freelists()) {
7556 // All the work is done in
7557 do_post_free_or_garbage_chunk(fc, size);
7558 } else { // Not adaptive free lists
7559 // this is a free chunk that can potentially be coalesced by the sweeper;
7560 if (!inFreeRange()) {
7561 // if the next chunk is a free block that can't be coalesced
7562 // it doesn't make sense to remove this chunk from the free lists
7563 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7564 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7565 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
7566 nextChunk->is_free() && // ... which is free...
7567 nextChunk->cantCoalesce()) { // ... but can't be coalesced
7568 // nothing to do
7569 } else {
7570 // Potentially the start of a new free range:
7571 // Don't eagerly remove it from the free lists.
7572 // No need to remove it if it will just be put
7573 // back again. (Also from a pragmatic point of view
7574 // if it is a free block in a region that is beyond
7575 // any allocated blocks, an assertion will fail)
7576 // Remember the start of a free run.
7577 initialize_free_range(addr, true);
7578 // end - can coalesce with next chunk
7579 }
7580 } else {
7581 // the midst of a free range, we are coalescing
7582 print_free_block_coalesced(fc);
7583 if (CMSTraceSweeper) {
7584 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7585 }
7586 // remove it from the free lists
7587 _sp->removeFreeChunkFromFreeLists(fc);
7588 set_lastFreeRangeCoalesced(true);
7589 // If the chunk is being coalesced and the current free range is
7590 // in the free lists, remove the current free range so that it
7591 // will be returned to the free lists in its entirety - all
7592 // the coalesced pieces included.
7593 if (freeRangeInFreeLists()) {
7594 FreeChunk* ffc = (FreeChunk*) freeFinger();
7595 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7596 "Size of free range is inconsistent with chunk size.");
7597 if (CMSTestInFreeList) {
7598 assert(_sp->verify_chunk_in_free_list(ffc),
7599 "free range is not in free lists");
7600 }
7601 _sp->removeFreeChunkFromFreeLists(ffc);
7602 set_freeRangeInFreeLists(false);
7603 }
7604 }
7605 }
7606 // Note that if the chunk is not coalescable (the else arm
7607 // below), we unconditionally flush, without needing to do
7608 // a "lookahead," as we do below.
7609 if (inFreeRange()) lookahead_and_flush(fc, size);
7610 } else {
7611 // Code path common to both original and adaptive free lists.
7612
7613 // cant coalesce with previous block; this should be treated
7614 // as the end of a free run if any
7615 if (inFreeRange()) {
7616 // we kicked some butt; time to pick up the garbage
7617 assert(freeFinger() < addr, "freeFinger points too high");
7618 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7619 }
7620 // else, nothing to do, just continue
7621 }
7622 }
7623
7624 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7625 // This is a chunk of garbage. It is not in any free list.
7626 // Add it to a free list or let it possibly be coalesced into
7627 // a larger chunk.
7628 HeapWord* const addr = (HeapWord*) fc;
7629 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7630
7631 if (_sp->adaptive_freelists()) {
7632 // Verify that the bit map has no bits marked between
7633 // addr and purported end of just dead object.
7634 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7635
7636 do_post_free_or_garbage_chunk(fc, size);
7637 } else {
7638 if (!inFreeRange()) {
7639 // start of a new free range
7640 assert(size > 0, "A free range should have a size");
7641 initialize_free_range(addr, false);
7642 } else {
7643 // this will be swept up when we hit the end of the
7644 // free range
7645 if (CMSTraceSweeper) {
7646 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7647 }
7648 // If the chunk is being coalesced and the current free range is
7649 // in the free lists, remove the current free range so that it
7650 // will be returned to the free lists in its entirety - all
7651 // the coalesced pieces included.
7652 if (freeRangeInFreeLists()) {
7653 FreeChunk* ffc = (FreeChunk*)freeFinger();
7654 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7655 "Size of free range is inconsistent with chunk size.");
7656 if (CMSTestInFreeList) {
7657 assert(_sp->verify_chunk_in_free_list(ffc),
7658 "free range is not in free lists");
7659 }
7660 _sp->removeFreeChunkFromFreeLists(ffc);
7661 set_freeRangeInFreeLists(false);
7662 }
7663 set_lastFreeRangeCoalesced(true);
7664 }
7665 // this will be swept up when we hit the end of the free range
7666
7667 // Verify that the bit map has no bits marked between
7668 // addr and purported end of just dead object.
7669 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7670 }
7671 assert(_limit >= addr + size,
7672 "A freshly garbage chunk can't possibly straddle over _limit");
7673 if (inFreeRange()) lookahead_and_flush(fc, size);
7674 return size;
7675 }
7676
7677 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7678 HeapWord* addr = (HeapWord*) fc;
7679 // The sweeper has just found a live object. Return any accumulated
7680 // left hand chunk to the free lists.
7681 if (inFreeRange()) {
7682 assert(freeFinger() < addr, "freeFinger points too high");
7683 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7684 }
7685
7686 // This object is live: we'd normally expect this to be
7687 // an oop, and like to assert the following:
7688 // assert(oop(addr)->is_oop(), "live block should be an oop");
7689 // However, as we commented above, this may be an object whose
7690 // header hasn't yet been initialized.
7691 size_t size;
7692 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7693 if (_bitMap->isMarked(addr + 1)) {
7694 // Determine the size from the bit map, rather than trying to
7695 // compute it from the object header.
7696 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7697 size = pointer_delta(nextOneAddr + 1, addr);
7698 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7699 "alignment problem");
7700
7701 #ifdef ASSERT
7702 if (oop(addr)->klass_or_null() != NULL) {
7703 // Ignore mark word because we are running concurrent with mutators
7704 assert(oop(addr)->is_oop(true), "live block should be an oop");
7705 assert(size ==
7706 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7707 "P-mark and computed size do not agree");
7708 }
7709 #endif
7710
7711 } else {
7712 // This should be an initialized object that's alive.
7713 assert(oop(addr)->klass_or_null() != NULL,
7714 "Should be an initialized object");
7715 // Ignore mark word because we are running concurrent with mutators
7716 assert(oop(addr)->is_oop(true), "live block should be an oop");
7717 // Verify that the bit map has no bits marked between
7718 // addr and purported end of this block.
7719 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7720 assert(size >= 3, "Necessary for Printezis marks to work");
7721 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7722 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7723 }
7724 return size;
7725 }
7726
7727 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7728 size_t chunkSize) {
7729 // do_post_free_or_garbage_chunk() should only be called in the case
7730 // of the adaptive free list allocator.
7731 const bool fcInFreeLists = fc->is_free();
7732 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7733 assert((HeapWord*)fc <= _limit, "sweep invariant");
7734 if (CMSTestInFreeList && fcInFreeLists) {
7735 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7736 }
7737
7738 if (CMSTraceSweeper) {
7739 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7740 }
7741
7742 HeapWord* const fc_addr = (HeapWord*) fc;
7743
7744 bool coalesce;
7745 const size_t left = pointer_delta(fc_addr, freeFinger());
7746 const size_t right = chunkSize;
7747 switch (FLSCoalescePolicy) {
7748 // numeric value forms a coalition aggressiveness metric
7749 case 0: { // never coalesce
7750 coalesce = false;
7751 break;
7752 }
7753 case 1: { // coalesce if left & right chunks on overpopulated lists
7754 coalesce = _sp->coalOverPopulated(left) &&
7755 _sp->coalOverPopulated(right);
7756 break;
7757 }
7758 case 2: { // coalesce if left chunk on overpopulated list (default)
7759 coalesce = _sp->coalOverPopulated(left);
7760 break;
7761 }
7762 case 3: { // coalesce if left OR right chunk on overpopulated list
7763 coalesce = _sp->coalOverPopulated(left) ||
7764 _sp->coalOverPopulated(right);
7765 break;
7766 }
7767 case 4: { // always coalesce
7768 coalesce = true;
7769 break;
7770 }
7771 default:
7772 ShouldNotReachHere();
7773 }
7774
7775 // Should the current free range be coalesced?
7776 // If the chunk is in a free range and either we decided to coalesce above
7777 // or the chunk is near the large block at the end of the heap
7778 // (isNearLargestChunk() returns true), then coalesce this chunk.
7779 const bool doCoalesce = inFreeRange()
7780 && (coalesce || _g->isNearLargestChunk(fc_addr));
7781 if (doCoalesce) {
7782 // Coalesce the current free range on the left with the new
7783 // chunk on the right. If either is on a free list,
7784 // it must be removed from the list and stashed in the closure.
7785 if (freeRangeInFreeLists()) {
7786 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7787 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7788 "Size of free range is inconsistent with chunk size.");
7789 if (CMSTestInFreeList) {
7790 assert(_sp->verify_chunk_in_free_list(ffc),
7791 "Chunk is not in free lists");
7792 }
7793 _sp->coalDeath(ffc->size());
7794 _sp->removeFreeChunkFromFreeLists(ffc);
7795 set_freeRangeInFreeLists(false);
7796 }
7797 if (fcInFreeLists) {
7798 _sp->coalDeath(chunkSize);
7799 assert(fc->size() == chunkSize,
7800 "The chunk has the wrong size or is not in the free lists");
7801 _sp->removeFreeChunkFromFreeLists(fc);
7802 }
7803 set_lastFreeRangeCoalesced(true);
7804 print_free_block_coalesced(fc);
7805 } else { // not in a free range and/or should not coalesce
7806 // Return the current free range and start a new one.
7807 if (inFreeRange()) {
7808 // In a free range but cannot coalesce with the right hand chunk.
7809 // Put the current free range into the free lists.
7810 flush_cur_free_chunk(freeFinger(),
7811 pointer_delta(fc_addr, freeFinger()));
7812 }
7813 // Set up for new free range. Pass along whether the right hand
7814 // chunk is in the free lists.
7815 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7816 }
7817 }
7818
7819 // Lookahead flush:
7820 // If we are tracking a free range, and this is the last chunk that
7821 // we'll look at because its end crosses past _limit, we'll preemptively
7822 // flush it along with any free range we may be holding on to. Note that
7823 // this can be the case only for an already free or freshly garbage
7824 // chunk. If this block is an object, it can never straddle
7825 // over _limit. The "straddling" occurs when _limit is set at
7826 // the previous end of the space when this cycle started, and
7827 // a subsequent heap expansion caused the previously co-terminal
7828 // free block to be coalesced with the newly expanded portion,
7829 // thus rendering _limit a non-block-boundary making it dangerous
7830 // for the sweeper to step over and examine.
7831 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7832 assert(inFreeRange(), "Should only be called if currently in a free range.");
7833 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7834 assert(_sp->used_region().contains(eob - 1),
7835 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7836 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7837 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7838 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size));
7839 if (eob >= _limit) {
7840 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7841 if (CMSTraceSweeper) {
7842 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7843 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7844 "[" PTR_FORMAT "," PTR_FORMAT ")",
7845 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7846 }
7847 // Return the storage we are tracking back into the free lists.
7848 if (CMSTraceSweeper) {
7849 gclog_or_tty->print_cr("Flushing ... ");
7850 }
7851 assert(freeFinger() < eob, "Error");
7852 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7853 }
7854 }
7855
7856 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7857 assert(inFreeRange(), "Should only be called if currently in a free range.");
7858 assert(size > 0,
7859 "A zero sized chunk cannot be added to the free lists.");
7860 if (!freeRangeInFreeLists()) {
7861 if (CMSTestInFreeList) {
7862 FreeChunk* fc = (FreeChunk*) chunk;
7863 fc->set_size(size);
7864 assert(!_sp->verify_chunk_in_free_list(fc),
7865 "chunk should not be in free lists yet");
7866 }
7867 if (CMSTraceSweeper) {
7868 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7869 p2i(chunk), size);
7870 }
7871 // A new free range is going to be starting. The current
7872 // free range has not been added to the free lists yet or
7873 // was removed so add it back.
7874 // If the current free range was coalesced, then the death
7875 // of the free range was recorded. Record a birth now.
7876 if (lastFreeRangeCoalesced()) {
7877 _sp->coalBirth(size);
7878 }
7879 _sp->addChunkAndRepairOffsetTable(chunk, size,
7880 lastFreeRangeCoalesced());
7881 } else if (CMSTraceSweeper) {
7882 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7883 }
7884 set_inFreeRange(false);
7885 set_freeRangeInFreeLists(false);
7886 }
7887
7888 // We take a break if we've been at this for a while,
7889 // so as to avoid monopolizing the locks involved.
7890 void SweepClosure::do_yield_work(HeapWord* addr) {
7891 // Return current free chunk being used for coalescing (if any)
7892 // to the appropriate freelist. After yielding, the next
7893 // free block encountered will start a coalescing range of
7894 // free blocks. If the next free block is adjacent to the
7895 // chunk just flushed, they will need to wait for the next
7896 // sweep to be coalesced.
7897 if (inFreeRange()) {
7898 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7899 }
7900
7901 // First give up the locks, then yield, then re-lock.
7902 // We should probably use a constructor/destructor idiom to
7903 // do this unlock/lock or modify the MutexUnlocker class to
7904 // serve our purpose. XXX
7905 assert_lock_strong(_bitMap->lock());
7906 assert_lock_strong(_freelistLock);
7907 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7908 "CMS thread should hold CMS token");
7909 _bitMap->lock()->unlock();
7910 _freelistLock->unlock();
7911 ConcurrentMarkSweepThread::desynchronize(true);
7912 _collector->stopTimer();
7913 if (PrintCMSStatistics != 0) {
7914 _collector->incrementYields();
7915 }
7916
7917 // See the comment in coordinator_yield()
7918 for (unsigned i = 0; i < CMSYieldSleepCount &&
7919 ConcurrentMarkSweepThread::should_yield() &&
7920 !CMSCollector::foregroundGCIsActive(); ++i) {
7921 os::sleep(Thread::current(), 1, false);
7922 }
7923
7924 ConcurrentMarkSweepThread::synchronize(true);
7925 _freelistLock->lock();
7926 _bitMap->lock()->lock_without_safepoint_check();
7927 _collector->startTimer();
7928 }
7929
7930 #ifndef PRODUCT
7931 // This is actually very useful in a product build if it can
7932 // be called from the debugger. Compile it into the product
7933 // as needed.
7934 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7935 return debug_cms_space->verify_chunk_in_free_list(fc);
7936 }
7937 #endif
7938
7939 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7940 if (CMSTraceSweeper) {
7941 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7942 p2i(fc), fc->size());
7943 }
7944 }
7945
7946 // CMSIsAliveClosure
7947 bool CMSIsAliveClosure::do_object_b(oop obj) {
7948 HeapWord* addr = (HeapWord*)obj;
7949 return addr != NULL &&
7950 (!_span.contains(addr) || _bit_map->isMarked(addr));
7951 }
7952
7953
7954 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7955 MemRegion span,
7956 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7957 bool cpc):
7958 _collector(collector),
7959 _span(span),
7960 _bit_map(bit_map),
7961 _mark_stack(mark_stack),
7962 _concurrent_precleaning(cpc) {
7963 assert(!_span.is_empty(), "Empty span could spell trouble");
7964 }
7965
7966
7967 // CMSKeepAliveClosure: the serial version
7968 void CMSKeepAliveClosure::do_oop(oop obj) {
7969 HeapWord* addr = (HeapWord*)obj;
7970 if (_span.contains(addr) &&
7971 !_bit_map->isMarked(addr)) {
7972 _bit_map->mark(addr);
7973 bool simulate_overflow = false;
7974 NOT_PRODUCT(
7975 if (CMSMarkStackOverflowALot &&
7976 _collector->simulate_overflow()) {
7977 // simulate a stack overflow
7978 simulate_overflow = true;
7979 }
7980 )
7981 if (simulate_overflow || !_mark_stack->push(obj)) {
7982 if (_concurrent_precleaning) {
7983 // We dirty the overflown object and let the remark
7984 // phase deal with it.
7985 assert(_collector->overflow_list_is_empty(), "Error");
7986 // In the case of object arrays, we need to dirty all of
7987 // the cards that the object spans. No locking or atomics
7988 // are needed since no one else can be mutating the mod union
7989 // table.
7990 if (obj->is_objArray()) {
7991 size_t sz = obj->size();
7992 HeapWord* end_card_addr =
7993 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7994 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7995 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7996 _collector->_modUnionTable.mark_range(redirty_range);
7997 } else {
7998 _collector->_modUnionTable.mark(addr);
7999 }
8000 _collector->_ser_kac_preclean_ovflw++;
8001 } else {
8002 _collector->push_on_overflow_list(obj);
8003 _collector->_ser_kac_ovflw++;
8004 }
8005 }
8006 }
8007 }
8008
8009 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8010 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8011
8012 // CMSParKeepAliveClosure: a parallel version of the above.
8013 // The work queues are private to each closure (thread),
8014 // but (may be) available for stealing by other threads.
8015 void CMSParKeepAliveClosure::do_oop(oop obj) {
8016 HeapWord* addr = (HeapWord*)obj;
8017 if (_span.contains(addr) &&
8018 !_bit_map->isMarked(addr)) {
8019 // In general, during recursive tracing, several threads
8020 // may be concurrently getting here; the first one to
8021 // "tag" it, claims it.
8022 if (_bit_map->par_mark(addr)) {
8023 bool res = _work_queue->push(obj);
8024 assert(res, "Low water mark should be much less than capacity");
8025 // Do a recursive trim in the hope that this will keep
8026 // stack usage lower, but leave some oops for potential stealers
8027 trim_queue(_low_water_mark);
8028 } // Else, another thread got there first
8029 }
8030 }
8031
8032 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8033 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8034
8035 void CMSParKeepAliveClosure::trim_queue(uint max) {
8036 while (_work_queue->size() > max) {
8037 oop new_oop;
8038 if (_work_queue->pop_local(new_oop)) {
8039 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8040 assert(_bit_map->isMarked((HeapWord*)new_oop),
8041 "no white objects on this stack!");
8042 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8043 // iterate over the oops in this oop, marking and pushing
8044 // the ones in CMS heap (i.e. in _span).
8045 new_oop->oop_iterate(&_mark_and_push);
8046 }
8047 }
8048 }
8049
8050 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8051 CMSCollector* collector,
8052 MemRegion span, CMSBitMap* bit_map,
8053 OopTaskQueue* work_queue):
8054 _collector(collector),
8055 _span(span),
8056 _bit_map(bit_map),
8057 _work_queue(work_queue) { }
8058
8059 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8060 HeapWord* addr = (HeapWord*)obj;
8061 if (_span.contains(addr) &&
8062 !_bit_map->isMarked(addr)) {
8063 if (_bit_map->par_mark(addr)) {
8064 bool simulate_overflow = false;
8065 NOT_PRODUCT(
8066 if (CMSMarkStackOverflowALot &&
8067 _collector->par_simulate_overflow()) {
8068 // simulate a stack overflow
8069 simulate_overflow = true;
8070 }
8071 )
8072 if (simulate_overflow || !_work_queue->push(obj)) {
8073 _collector->par_push_on_overflow_list(obj);
8074 _collector->_par_kac_ovflw++;
8075 }
8076 } // Else another thread got there already
8077 }
8078 }
8079
8080 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8081 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8082
8083 //////////////////////////////////////////////////////////////////
8084 // CMSExpansionCause /////////////////////////////
8085 //////////////////////////////////////////////////////////////////
8086 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8087 switch (cause) {
8088 case _no_expansion:
8089 return "No expansion";
8090 case _satisfy_free_ratio:
8091 return "Free ratio";
8092 case _satisfy_promotion:
8093 return "Satisfy promotion";
8094 case _satisfy_allocation:
8095 return "allocation";
8096 case _allocate_par_lab:
8097 return "Par LAB";
8098 case _allocate_par_spooling_space:
8099 return "Par Spooling Space";
8100 case _adaptive_size_policy:
8101 return "Ergonomics";
8102 default:
8103 return "unknown";
8104 }
8105 }
8106
8107 void CMSDrainMarkingStackClosure::do_void() {
8108 // the max number to take from overflow list at a time
8109 const size_t num = _mark_stack->capacity()/4;
8110 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8111 "Overflow list should be NULL during concurrent phases");
8112 while (!_mark_stack->isEmpty() ||
8113 // if stack is empty, check the overflow list
8114 _collector->take_from_overflow_list(num, _mark_stack)) {
8115 oop obj = _mark_stack->pop();
8116 HeapWord* addr = (HeapWord*)obj;
8117 assert(_span.contains(addr), "Should be within span");
8118 assert(_bit_map->isMarked(addr), "Should be marked");
8119 assert(obj->is_oop(), "Should be an oop");
8120 obj->oop_iterate(_keep_alive);
8121 }
8122 }
8123
8124 void CMSParDrainMarkingStackClosure::do_void() {
8125 // drain queue
8126 trim_queue(0);
8127 }
8128
8129 // Trim our work_queue so its length is below max at return
8130 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8131 while (_work_queue->size() > max) {
8132 oop new_oop;
8133 if (_work_queue->pop_local(new_oop)) {
8134 assert(new_oop->is_oop(), "Expected an oop");
8135 assert(_bit_map->isMarked((HeapWord*)new_oop),
8136 "no white objects on this stack!");
8137 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8138 // iterate over the oops in this oop, marking and pushing
8139 // the ones in CMS heap (i.e. in _span).
8140 new_oop->oop_iterate(&_mark_and_push);
8141 }
8142 }
8143 }
8144
8145 ////////////////////////////////////////////////////////////////////
8146 // Support for Marking Stack Overflow list handling and related code
8147 ////////////////////////////////////////////////////////////////////
8148 // Much of the following code is similar in shape and spirit to the
8149 // code used in ParNewGC. We should try and share that code
8150 // as much as possible in the future.
8151
8152 #ifndef PRODUCT
8153 // Debugging support for CMSStackOverflowALot
8154
8155 // It's OK to call this multi-threaded; the worst thing
8156 // that can happen is that we'll get a bunch of closely
8157 // spaced simulated overflows, but that's OK, in fact
8158 // probably good as it would exercise the overflow code
8159 // under contention.
8160 bool CMSCollector::simulate_overflow() {
8161 if (_overflow_counter-- <= 0) { // just being defensive
8162 _overflow_counter = CMSMarkStackOverflowInterval;
8163 return true;
8164 } else {
8165 return false;
8166 }
8167 }
8168
8169 bool CMSCollector::par_simulate_overflow() {
8170 return simulate_overflow();
8171 }
8172 #endif
8173
8174 // Single-threaded
8175 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8176 assert(stack->isEmpty(), "Expected precondition");
8177 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8178 size_t i = num;
8179 oop cur = _overflow_list;
8180 const markOop proto = markOopDesc::prototype();
8181 NOT_PRODUCT(ssize_t n = 0;)
8182 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8183 next = oop(cur->mark());
8184 cur->set_mark(proto); // until proven otherwise
8185 assert(cur->is_oop(), "Should be an oop");
8186 bool res = stack->push(cur);
8187 assert(res, "Bit off more than can chew?");
8188 NOT_PRODUCT(n++;)
8189 }
8190 _overflow_list = cur;
8191 #ifndef PRODUCT
8192 assert(_num_par_pushes >= n, "Too many pops?");
8193 _num_par_pushes -=n;
8194 #endif
8195 return !stack->isEmpty();
8196 }
8197
8198 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8199 // (MT-safe) Get a prefix of at most "num" from the list.
8200 // The overflow list is chained through the mark word of
8201 // each object in the list. We fetch the entire list,
8202 // break off a prefix of the right size and return the
8203 // remainder. If other threads try to take objects from
8204 // the overflow list at that time, they will wait for
8205 // some time to see if data becomes available. If (and
8206 // only if) another thread places one or more object(s)
8207 // on the global list before we have returned the suffix
8208 // to the global list, we will walk down our local list
8209 // to find its end and append the global list to
8210 // our suffix before returning it. This suffix walk can
8211 // prove to be expensive (quadratic in the amount of traffic)
8212 // when there are many objects in the overflow list and
8213 // there is much producer-consumer contention on the list.
8214 // *NOTE*: The overflow list manipulation code here and
8215 // in ParNewGeneration:: are very similar in shape,
8216 // except that in the ParNew case we use the old (from/eden)
8217 // copy of the object to thread the list via its klass word.
8218 // Because of the common code, if you make any changes in
8219 // the code below, please check the ParNew version to see if
8220 // similar changes might be needed.
8221 // CR 6797058 has been filed to consolidate the common code.
8222 bool CMSCollector::par_take_from_overflow_list(size_t num,
8223 OopTaskQueue* work_q,
8224 int no_of_gc_threads) {
8225 assert(work_q->size() == 0, "First empty local work queue");
8226 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8227 if (_overflow_list == NULL) {
8228 return false;
8229 }
8230 // Grab the entire list; we'll put back a suffix
8231 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8232 Thread* tid = Thread::current();
8233 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8234 // set to ParallelGCThreads.
8235 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8236 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8237 // If the list is busy, we spin for a short while,
8238 // sleeping between attempts to get the list.
8239 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8240 os::sleep(tid, sleep_time_millis, false);
8241 if (_overflow_list == NULL) {
8242 // Nothing left to take
8243 return false;
8244 } else if (_overflow_list != BUSY) {
8245 // Try and grab the prefix
8246 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8247 }
8248 }
8249 // If the list was found to be empty, or we spun long
8250 // enough, we give up and return empty-handed. If we leave
8251 // the list in the BUSY state below, it must be the case that
8252 // some other thread holds the overflow list and will set it
8253 // to a non-BUSY state in the future.
8254 if (prefix == NULL || prefix == BUSY) {
8255 // Nothing to take or waited long enough
8256 if (prefix == NULL) {
8257 // Write back the NULL in case we overwrote it with BUSY above
8258 // and it is still the same value.
8259 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8260 }
8261 return false;
8262 }
8263 assert(prefix != NULL && prefix != BUSY, "Error");
8264 size_t i = num;
8265 oop cur = prefix;
8266 // Walk down the first "num" objects, unless we reach the end.
8267 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8268 if (cur->mark() == NULL) {
8269 // We have "num" or fewer elements in the list, so there
8270 // is nothing to return to the global list.
8271 // Write back the NULL in lieu of the BUSY we wrote
8272 // above, if it is still the same value.
8273 if (_overflow_list == BUSY) {
8274 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8275 }
8276 } else {
8277 // Chop off the suffix and return it to the global list.
8278 assert(cur->mark() != BUSY, "Error");
8279 oop suffix_head = cur->mark(); // suffix will be put back on global list
8280 cur->set_mark(NULL); // break off suffix
8281 // It's possible that the list is still in the empty(busy) state
8282 // we left it in a short while ago; in that case we may be
8283 // able to place back the suffix without incurring the cost
8284 // of a walk down the list.
8285 oop observed_overflow_list = _overflow_list;
8286 oop cur_overflow_list = observed_overflow_list;
8287 bool attached = false;
8288 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8289 observed_overflow_list =
8290 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8291 if (cur_overflow_list == observed_overflow_list) {
8292 attached = true;
8293 break;
8294 } else cur_overflow_list = observed_overflow_list;
8295 }
8296 if (!attached) {
8297 // Too bad, someone else sneaked in (at least) an element; we'll need
8298 // to do a splice. Find tail of suffix so we can prepend suffix to global
8299 // list.
8300 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8301 oop suffix_tail = cur;
8302 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8303 "Tautology");
8304 observed_overflow_list = _overflow_list;
8305 do {
8306 cur_overflow_list = observed_overflow_list;
8307 if (cur_overflow_list != BUSY) {
8308 // Do the splice ...
8309 suffix_tail->set_mark(markOop(cur_overflow_list));
8310 } else { // cur_overflow_list == BUSY
8311 suffix_tail->set_mark(NULL);
8312 }
8313 // ... and try to place spliced list back on overflow_list ...
8314 observed_overflow_list =
8315 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8316 } while (cur_overflow_list != observed_overflow_list);
8317 // ... until we have succeeded in doing so.
8318 }
8319 }
8320
8321 // Push the prefix elements on work_q
8322 assert(prefix != NULL, "control point invariant");
8323 const markOop proto = markOopDesc::prototype();
8324 oop next;
8325 NOT_PRODUCT(ssize_t n = 0;)
8326 for (cur = prefix; cur != NULL; cur = next) {
8327 next = oop(cur->mark());
8328 cur->set_mark(proto); // until proven otherwise
8329 assert(cur->is_oop(), "Should be an oop");
8330 bool res = work_q->push(cur);
8331 assert(res, "Bit off more than we can chew?");
8332 NOT_PRODUCT(n++;)
8333 }
8334 #ifndef PRODUCT
8335 assert(_num_par_pushes >= n, "Too many pops?");
8336 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8337 #endif
8338 return true;
8339 }
8340
8341 // Single-threaded
8342 void CMSCollector::push_on_overflow_list(oop p) {
8343 NOT_PRODUCT(_num_par_pushes++;)
8344 assert(p->is_oop(), "Not an oop");
8345 preserve_mark_if_necessary(p);
8346 p->set_mark((markOop)_overflow_list);
8347 _overflow_list = p;
8348 }
8349
8350 // Multi-threaded; use CAS to prepend to overflow list
8351 void CMSCollector::par_push_on_overflow_list(oop p) {
8352 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8353 assert(p->is_oop(), "Not an oop");
8354 par_preserve_mark_if_necessary(p);
8355 oop observed_overflow_list = _overflow_list;
8356 oop cur_overflow_list;
8357 do {
8358 cur_overflow_list = observed_overflow_list;
8359 if (cur_overflow_list != BUSY) {
8360 p->set_mark(markOop(cur_overflow_list));
8361 } else {
8362 p->set_mark(NULL);
8363 }
8364 observed_overflow_list =
8365 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8366 } while (cur_overflow_list != observed_overflow_list);
8367 }
8368 #undef BUSY
8369
8370 // Single threaded
8371 // General Note on GrowableArray: pushes may silently fail
8372 // because we are (temporarily) out of C-heap for expanding
8373 // the stack. The problem is quite ubiquitous and affects
8374 // a lot of code in the JVM. The prudent thing for GrowableArray
8375 // to do (for now) is to exit with an error. However, that may
8376 // be too draconian in some cases because the caller may be
8377 // able to recover without much harm. For such cases, we
8378 // should probably introduce a "soft_push" method which returns
8379 // an indication of success or failure with the assumption that
8380 // the caller may be able to recover from a failure; code in
8381 // the VM can then be changed, incrementally, to deal with such
8382 // failures where possible, thus, incrementally hardening the VM
8383 // in such low resource situations.
8384 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8385 _preserved_oop_stack.push(p);
8386 _preserved_mark_stack.push(m);
8387 assert(m == p->mark(), "Mark word changed");
8388 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8389 "bijection");
8390 }
8391
8392 // Single threaded
8393 void CMSCollector::preserve_mark_if_necessary(oop p) {
8394 markOop m = p->mark();
8395 if (m->must_be_preserved(p)) {
8396 preserve_mark_work(p, m);
8397 }
8398 }
8399
8400 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8401 markOop m = p->mark();
8402 if (m->must_be_preserved(p)) {
8403 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8404 // Even though we read the mark word without holding
8405 // the lock, we are assured that it will not change
8406 // because we "own" this oop, so no other thread can
8407 // be trying to push it on the overflow list; see
8408 // the assertion in preserve_mark_work() that checks
8409 // that m == p->mark().
8410 preserve_mark_work(p, m);
8411 }
8412 }
8413
8414 // We should be able to do this multi-threaded,
8415 // a chunk of stack being a task (this is
8416 // correct because each oop only ever appears
8417 // once in the overflow list. However, it's
8418 // not very easy to completely overlap this with
8419 // other operations, so will generally not be done
8420 // until all work's been completed. Because we
8421 // expect the preserved oop stack (set) to be small,
8422 // it's probably fine to do this single-threaded.
8423 // We can explore cleverer concurrent/overlapped/parallel
8424 // processing of preserved marks if we feel the
8425 // need for this in the future. Stack overflow should
8426 // be so rare in practice and, when it happens, its
8427 // effect on performance so great that this will
8428 // likely just be in the noise anyway.
8429 void CMSCollector::restore_preserved_marks_if_any() {
8430 assert(SafepointSynchronize::is_at_safepoint(),
8431 "world should be stopped");
8432 assert(Thread::current()->is_ConcurrentGC_thread() ||
8433 Thread::current()->is_VM_thread(),
8434 "should be single-threaded");
8435 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8436 "bijection");
8437
8438 while (!_preserved_oop_stack.is_empty()) {
8439 oop p = _preserved_oop_stack.pop();
8440 assert(p->is_oop(), "Should be an oop");
8441 assert(_span.contains(p), "oop should be in _span");
8442 assert(p->mark() == markOopDesc::prototype(),
8443 "Set when taken from overflow list");
8444 markOop m = _preserved_mark_stack.pop();
8445 p->set_mark(m);
8446 }
8447 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8448 "stacks were cleared above");
8449 }
8450
8451 #ifndef PRODUCT
8452 bool CMSCollector::no_preserved_marks() const {
8453 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8454 }
8455 #endif
8456
8457 // Transfer some number of overflown objects to usual marking
8458 // stack. Return true if some objects were transferred.
8459 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8460 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8461 (size_t)ParGCDesiredObjsFromOverflowList);
8462
8463 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8464 assert(_collector->overflow_list_is_empty() || res,
8465 "If list is not empty, we should have taken something");
8466 assert(!res || !_mark_stack->isEmpty(),
8467 "If we took something, it should now be on our stack");
8468 return res;
8469 }
8470
8471 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8472 size_t res = _sp->block_size_no_stall(addr, _collector);
8473 if (_sp->block_is_obj(addr)) {
8474 if (_live_bit_map->isMarked(addr)) {
8475 // It can't have been dead in a previous cycle
8476 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8477 } else {
8478 _dead_bit_map->mark(addr); // mark the dead object
8479 }
8480 }
8481 // Could be 0, if the block size could not be computed without stalling.
8482 return res;
8483 }
8484
8485 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8486
8487 switch (phase) {
8488 case CMSCollector::InitialMarking:
8489 initialize(true /* fullGC */ ,
8490 cause /* cause of the GC */,
8491 true /* recordGCBeginTime */,
8492 true /* recordPreGCUsage */,
8493 false /* recordPeakUsage */,
8494 false /* recordPostGCusage */,
8495 true /* recordAccumulatedGCTime */,
8496 false /* recordGCEndTime */,
8497 false /* countCollection */ );
8498 break;
8499
8500 case CMSCollector::FinalMarking:
8501 initialize(true /* fullGC */ ,
8502 cause /* cause of the GC */,
8503 false /* recordGCBeginTime */,
8504 false /* recordPreGCUsage */,
8505 false /* recordPeakUsage */,
8506 false /* recordPostGCusage */,
8507 true /* recordAccumulatedGCTime */,
8508 false /* recordGCEndTime */,
8509 false /* countCollection */ );
8510 break;
8511
8512 case CMSCollector::Sweeping:
8513 initialize(true /* fullGC */ ,
8514 cause /* cause of the GC */,
8515 false /* recordGCBeginTime */,
8516 false /* recordPreGCUsage */,
8517 true /* recordPeakUsage */,
8518 true /* recordPostGCusage */,
8519 false /* recordAccumulatedGCTime */,
8520 true /* recordGCEndTime */,
8521 true /* countCollection */ );
8522 break;
8523
8524 default:
8525 ShouldNotReachHere();
8526 }
8527 }