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,
194 CardTableRS* ct, bool use_adaptive_freelists,
195 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
196 CardGeneration(rs, initial_byte_size, 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->_old_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 ParallelGCThreads, // mt processing degree
289 _cmsGen->refs_discovery_is_mt(), // mt discovery
290 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 young 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 = 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 = CMSBootstrapOccupancy / 100.0;
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 const size_t max_plab_samples =
623 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
624
625 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
626 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
627 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
628 _survivor_chunk_capacity = max_plab_samples;
629 for (uint i = 0; i < ParallelGCThreads; i++) {
630 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
631 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
632 assert(cur->end() == 0, "Should be 0");
633 assert(cur->array() == vec, "Should be vec");
634 assert(cur->capacity() == max_plab_samples, "Error");
635 }
636 }
637
638 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
639 _gc_counters = new CollectorCounters("CMS", 1);
640 _completed_initialization = true;
641 _inter_sweep_timer.start(); // start of time
642 }
643
644 const char* ConcurrentMarkSweepGeneration::name() const {
645 return "concurrent mark-sweep generation";
646 }
647 void ConcurrentMarkSweepGeneration::update_counters() {
648 if (UsePerfData) {
649 _space_counters->update_all();
650 _gen_counters->update_all();
651 }
652 }
653
654 // this is an optimized version of update_counters(). it takes the
655 // used value as a parameter rather than computing it.
656 //
657 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
658 if (UsePerfData) {
659 _space_counters->update_used(used);
660 _space_counters->update_capacity();
661 _gen_counters->update_all();
662 }
663 }
664
665 void ConcurrentMarkSweepGeneration::print() const {
666 Generation::print();
667 cmsSpace()->print();
668 }
669
670 #ifndef PRODUCT
671 void ConcurrentMarkSweepGeneration::print_statistics() {
672 cmsSpace()->printFLCensus(0);
673 }
674 #endif
675
676 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
677 GenCollectedHeap* gch = GenCollectedHeap::heap();
678 if (PrintGCDetails) {
679 // I didn't want to change the logging when removing the level concept,
680 // but I guess this logging could say "old" or something instead of "1".
681 assert(gch->is_old_gen(this),
682 "The CMS generation should be the old generation");
683 uint level = 1;
684 if (Verbose) {
685 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]",
686 level, short_name(), s, used(), capacity());
687 } else {
688 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]",
689 level, short_name(), s, used() / K, capacity() / K);
690 }
691 }
692 if (Verbose) {
693 gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")",
694 gch->used(), gch->capacity());
695 } else {
696 gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)",
697 gch->used() / K, gch->capacity() / K);
698 }
699 }
700
701 size_t
702 ConcurrentMarkSweepGeneration::contiguous_available() const {
703 // dld proposes an improvement in precision here. If the committed
704 // part of the space ends in a free block we should add that to
705 // uncommitted size in the calculation below. Will make this
706 // change later, staying with the approximation below for the
707 // time being. -- ysr.
708 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
709 }
710
711 size_t
712 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
713 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
714 }
715
716 size_t ConcurrentMarkSweepGeneration::max_available() const {
717 return free() + _virtual_space.uncommitted_size();
718 }
719
720 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
721 size_t available = max_available();
722 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
723 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
724 if (Verbose && PrintGCDetails) {
725 gclog_or_tty->print_cr(
726 "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "),"
727 "max_promo(" SIZE_FORMAT ")",
728 res? "":" not", available, res? ">=":"<",
729 av_promo, max_promotion_in_bytes);
730 }
731 return res;
732 }
733
734 // At a promotion failure dump information on block layout in heap
735 // (cms old generation).
736 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
737 if (CMSDumpAtPromotionFailure) {
738 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
739 }
740 }
741
742 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
743 // Clear the promotion information. These pointers can be adjusted
744 // along with all the other pointers into the heap but
745 // compaction is expected to be a rare event with
746 // a heap using cms so don't do it without seeing the need.
747 for (uint i = 0; i < ParallelGCThreads; i++) {
748 _par_gc_thread_states[i]->promo.reset();
749 }
750 }
751
752 void ConcurrentMarkSweepGeneration::compute_new_size() {
753 assert_locked_or_safepoint(Heap_lock);
754
755 // If incremental collection failed, we just want to expand
756 // to the limit.
757 if (incremental_collection_failed()) {
758 clear_incremental_collection_failed();
759 grow_to_reserved();
760 return;
761 }
762
763 // The heap has been compacted but not reset yet.
764 // Any metric such as free() or used() will be incorrect.
765
766 CardGeneration::compute_new_size();
767
768 // Reset again after a possible resizing
769 if (did_compact()) {
770 cmsSpace()->reset_after_compaction();
771 }
772 }
773
774 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
775 assert_locked_or_safepoint(Heap_lock);
776
777 // If incremental collection failed, we just want to expand
778 // to the limit.
779 if (incremental_collection_failed()) {
780 clear_incremental_collection_failed();
781 grow_to_reserved();
782 return;
783 }
784
785 double free_percentage = ((double) free()) / capacity();
786 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
787 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
788
789 // compute expansion delta needed for reaching desired free percentage
790 if (free_percentage < desired_free_percentage) {
791 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
792 assert(desired_capacity >= capacity(), "invalid expansion size");
793 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
794 if (PrintGCDetails && Verbose) {
795 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
796 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
797 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
798 gclog_or_tty->print_cr(" Desired free fraction %f", desired_free_percentage);
799 gclog_or_tty->print_cr(" Maximum free fraction %f", maximum_free_percentage);
800 gclog_or_tty->print_cr(" Capacity " SIZE_FORMAT, capacity() / 1000);
801 gclog_or_tty->print_cr(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
802 GenCollectedHeap* gch = GenCollectedHeap::heap();
803 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
804 size_t young_size = gch->young_gen()->capacity();
805 gclog_or_tty->print_cr(" Young gen size " SIZE_FORMAT, young_size / 1000);
806 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
807 gclog_or_tty->print_cr(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
808 gclog_or_tty->print_cr(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
809 }
810 // safe if expansion fails
811 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
812 if (PrintGCDetails && Verbose) {
813 gclog_or_tty->print_cr(" Expanded free fraction %f", ((double) free()) / capacity());
814 }
815 } else {
816 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
817 assert(desired_capacity <= capacity(), "invalid expansion size");
818 size_t shrink_bytes = capacity() - desired_capacity;
819 // Don't shrink unless the delta is greater than the minimum shrink we want
820 if (shrink_bytes >= MinHeapDeltaBytes) {
821 shrink_free_list_by(shrink_bytes);
822 }
823 }
824 }
825
826 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
827 return cmsSpace()->freelistLock();
828 }
829
830 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
831 CMSSynchronousYieldRequest yr;
832 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
833 return have_lock_and_allocate(size, tlab);
834 }
835
836 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
837 bool tlab /* ignored */) {
838 assert_lock_strong(freelistLock());
839 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
840 HeapWord* res = cmsSpace()->allocate(adjustedSize);
841 // Allocate the object live (grey) if the background collector has
842 // started marking. This is necessary because the marker may
843 // have passed this address and consequently this object will
844 // not otherwise be greyed and would be incorrectly swept up.
845 // Note that if this object contains references, the writing
846 // of those references will dirty the card containing this object
847 // allowing the object to be blackened (and its references scanned)
848 // either during a preclean phase or at the final checkpoint.
849 if (res != NULL) {
850 // We may block here with an uninitialized object with
851 // its mark-bit or P-bits not yet set. Such objects need
852 // to be safely navigable by block_start().
853 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
854 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
855 collector()->direct_allocated(res, adjustedSize);
856 _direct_allocated_words += adjustedSize;
857 // allocation counters
858 NOT_PRODUCT(
859 _numObjectsAllocated++;
860 _numWordsAllocated += (int)adjustedSize;
861 )
862 }
863 return res;
864 }
865
866 // In the case of direct allocation by mutators in a generation that
867 // is being concurrently collected, the object must be allocated
868 // live (grey) if the background collector has started marking.
869 // This is necessary because the marker may
870 // have passed this address and consequently this object will
871 // not otherwise be greyed and would be incorrectly swept up.
872 // Note that if this object contains references, the writing
873 // of those references will dirty the card containing this object
874 // allowing the object to be blackened (and its references scanned)
875 // either during a preclean phase or at the final checkpoint.
876 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
877 assert(_markBitMap.covers(start, size), "Out of bounds");
878 if (_collectorState >= Marking) {
879 MutexLockerEx y(_markBitMap.lock(),
880 Mutex::_no_safepoint_check_flag);
881 // [see comments preceding SweepClosure::do_blk() below for details]
882 //
883 // Can the P-bits be deleted now? JJJ
884 //
885 // 1. need to mark the object as live so it isn't collected
886 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
887 // 3. need to mark the end of the object so marking, precleaning or sweeping
888 // can skip over uninitialized or unparsable objects. An allocated
889 // object is considered uninitialized for our purposes as long as
890 // its klass word is NULL. All old gen objects are parsable
891 // as soon as they are initialized.)
892 _markBitMap.mark(start); // object is live
893 _markBitMap.mark(start + 1); // object is potentially uninitialized?
894 _markBitMap.mark(start + size - 1);
895 // mark end of object
896 }
897 // check that oop looks uninitialized
898 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
899 }
900
901 void CMSCollector::promoted(bool par, HeapWord* start,
902 bool is_obj_array, size_t obj_size) {
903 assert(_markBitMap.covers(start), "Out of bounds");
904 // See comment in direct_allocated() about when objects should
905 // be allocated live.
906 if (_collectorState >= Marking) {
907 // we already hold the marking bit map lock, taken in
908 // the prologue
909 if (par) {
910 _markBitMap.par_mark(start);
911 } else {
912 _markBitMap.mark(start);
913 }
914 // We don't need to mark the object as uninitialized (as
915 // in direct_allocated above) because this is being done with the
916 // world stopped and the object will be initialized by the
917 // time the marking, precleaning or sweeping get to look at it.
918 // But see the code for copying objects into the CMS generation,
919 // where we need to ensure that concurrent readers of the
920 // block offset table are able to safely navigate a block that
921 // is in flux from being free to being allocated (and in
922 // transition while being copied into) and subsequently
923 // becoming a bona-fide object when the copy/promotion is complete.
924 assert(SafepointSynchronize::is_at_safepoint(),
925 "expect promotion only at safepoints");
926
927 if (_collectorState < Sweeping) {
928 // Mark the appropriate cards in the modUnionTable, so that
929 // this object gets scanned before the sweep. If this is
930 // not done, CMS generation references in the object might
931 // not get marked.
932 // For the case of arrays, which are otherwise precisely
933 // marked, we need to dirty the entire array, not just its head.
934 if (is_obj_array) {
935 // The [par_]mark_range() method expects mr.end() below to
936 // be aligned to the granularity of a bit's representation
937 // in the heap. In the case of the MUT below, that's a
938 // card size.
939 MemRegion mr(start,
940 (HeapWord*)round_to((intptr_t)(start + obj_size),
941 CardTableModRefBS::card_size /* bytes */));
942 if (par) {
943 _modUnionTable.par_mark_range(mr);
944 } else {
945 _modUnionTable.mark_range(mr);
946 }
947 } else { // not an obj array; we can just mark the head
948 if (par) {
949 _modUnionTable.par_mark(start);
950 } else {
951 _modUnionTable.mark(start);
952 }
953 }
954 }
955 }
956 }
957
958 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
959 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
960 // allocate, copy and if necessary update promoinfo --
961 // delegate to underlying space.
962 assert_lock_strong(freelistLock());
963
964 #ifndef PRODUCT
965 if (GenCollectedHeap::heap()->promotion_should_fail()) {
966 return NULL;
967 }
968 #endif // #ifndef PRODUCT
969
970 oop res = _cmsSpace->promote(obj, obj_size);
971 if (res == NULL) {
972 // expand and retry
973 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
974 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
975 // Since this is the old generation, we don't try to promote
976 // into a more senior generation.
977 res = _cmsSpace->promote(obj, obj_size);
978 }
979 if (res != NULL) {
980 // See comment in allocate() about when objects should
981 // be allocated live.
982 assert(obj->is_oop(), "Will dereference klass pointer below");
983 collector()->promoted(false, // Not parallel
984 (HeapWord*)res, obj->is_objArray(), obj_size);
985 // promotion counters
986 NOT_PRODUCT(
987 _numObjectsPromoted++;
988 _numWordsPromoted +=
989 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
990 )
991 }
992 return res;
993 }
994
995
996 // IMPORTANT: Notes on object size recognition in CMS.
997 // ---------------------------------------------------
998 // A block of storage in the CMS generation is always in
999 // one of three states. A free block (FREE), an allocated
1000 // object (OBJECT) whose size() method reports the correct size,
1001 // and an intermediate state (TRANSIENT) in which its size cannot
1002 // be accurately determined.
1003 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1004 // -----------------------------------------------------
1005 // FREE: klass_word & 1 == 1; mark_word holds block size
1006 //
1007 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1008 // obj->size() computes correct size
1009 //
1010 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1011 //
1012 // STATE IDENTIFICATION: (64 bit+COOPS)
1013 // ------------------------------------
1014 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1015 //
1016 // OBJECT: klass_word installed; klass_word != 0;
1017 // obj->size() computes correct size
1018 //
1019 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1020 //
1021 //
1022 // STATE TRANSITION DIAGRAM
1023 //
1024 // mut / parnew mut / parnew
1025 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1026 // ^ |
1027 // |------------------------ DEAD <------------------------------------|
1028 // sweep mut
1029 //
1030 // While a block is in TRANSIENT state its size cannot be determined
1031 // so readers will either need to come back later or stall until
1032 // the size can be determined. Note that for the case of direct
1033 // allocation, P-bits, when available, may be used to determine the
1034 // size of an object that may not yet have been initialized.
1035
1036 // Things to support parallel young-gen collection.
1037 oop
1038 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1039 oop old, markOop m,
1040 size_t word_sz) {
1041 #ifndef PRODUCT
1042 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1043 return NULL;
1044 }
1045 #endif // #ifndef PRODUCT
1046
1047 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1048 PromotionInfo* promoInfo = &ps->promo;
1049 // if we are tracking promotions, then first ensure space for
1050 // promotion (including spooling space for saving header if necessary).
1051 // then allocate and copy, then track promoted info if needed.
1052 // When tracking (see PromotionInfo::track()), the mark word may
1053 // be displaced and in this case restoration of the mark word
1054 // occurs in the (oop_since_save_marks_)iterate phase.
1055 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1056 // Out of space for allocating spooling buffers;
1057 // try expanding and allocating spooling buffers.
1058 if (!expand_and_ensure_spooling_space(promoInfo)) {
1059 return NULL;
1060 }
1061 }
1062 assert(promoInfo->has_spooling_space(), "Control point invariant");
1063 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1064 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1065 if (obj_ptr == NULL) {
1066 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1067 if (obj_ptr == NULL) {
1068 return NULL;
1069 }
1070 }
1071 oop obj = oop(obj_ptr);
1072 OrderAccess::storestore();
1073 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1074 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1075 // IMPORTANT: See note on object initialization for CMS above.
1076 // Otherwise, copy the object. Here we must be careful to insert the
1077 // klass pointer last, since this marks the block as an allocated object.
1078 // Except with compressed oops it's the mark word.
1079 HeapWord* old_ptr = (HeapWord*)old;
1080 // Restore the mark word copied above.
1081 obj->set_mark(m);
1082 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1083 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1084 OrderAccess::storestore();
1085
1086 if (UseCompressedClassPointers) {
1087 // Copy gap missed by (aligned) header size calculation below
1088 obj->set_klass_gap(old->klass_gap());
1089 }
1090 if (word_sz > (size_t)oopDesc::header_size()) {
1091 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1092 obj_ptr + oopDesc::header_size(),
1093 word_sz - oopDesc::header_size());
1094 }
1095
1096 // Now we can track the promoted object, if necessary. We take care
1097 // to delay the transition from uninitialized to full object
1098 // (i.e., insertion of klass pointer) until after, so that it
1099 // atomically becomes a promoted object.
1100 if (promoInfo->tracking()) {
1101 promoInfo->track((PromotedObject*)obj, old->klass());
1102 }
1103 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1104 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1105 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1106
1107 // Finally, install the klass pointer (this should be volatile).
1108 OrderAccess::storestore();
1109 obj->set_klass(old->klass());
1110 // We should now be able to calculate the right size for this object
1111 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1112
1113 collector()->promoted(true, // parallel
1114 obj_ptr, old->is_objArray(), word_sz);
1115
1116 NOT_PRODUCT(
1117 Atomic::inc_ptr(&_numObjectsPromoted);
1118 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1119 )
1120
1121 return obj;
1122 }
1123
1124 void
1125 ConcurrentMarkSweepGeneration::
1126 par_promote_alloc_done(int thread_num) {
1127 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1128 ps->lab.retire(thread_num);
1129 }
1130
1131 void
1132 ConcurrentMarkSweepGeneration::
1133 par_oop_since_save_marks_iterate_done(int thread_num) {
1134 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1135 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1136 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1137 }
1138
1139 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1140 size_t size,
1141 bool tlab)
1142 {
1143 // We allow a STW collection only if a full
1144 // collection was requested.
1145 return full || should_allocate(size, tlab); // FIX ME !!!
1146 // This and promotion failure handling are connected at the
1147 // hip and should be fixed by untying them.
1148 }
1149
1150 bool CMSCollector::shouldConcurrentCollect() {
1151 if (_full_gc_requested) {
1152 if (Verbose && PrintGCDetails) {
1153 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1154 " gc request (or gc_locker)");
1155 }
1156 return true;
1157 }
1158
1159 FreelistLocker x(this);
1160 // ------------------------------------------------------------------
1161 // Print out lots of information which affects the initiation of
1162 // a collection.
1163 if (PrintCMSInitiationStatistics && stats().valid()) {
1164 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1165 gclog_or_tty->stamp();
1166 gclog_or_tty->cr();
1167 stats().print_on(gclog_or_tty);
1168 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1169 stats().time_until_cms_gen_full());
1170 gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free());
1171 gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT,
1172 _cmsGen->contiguous_available());
1173 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1174 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1175 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1176 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1177 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1178 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1179 gclog_or_tty->print_cr("metadata initialized %d",
1180 MetaspaceGC::should_concurrent_collect());
1181 }
1182 // ------------------------------------------------------------------
1183
1184 // If the estimated time to complete a cms collection (cms_duration())
1185 // is less than the estimated time remaining until the cms generation
1186 // is full, start a collection.
1187 if (!UseCMSInitiatingOccupancyOnly) {
1188 if (stats().valid()) {
1189 if (stats().time_until_cms_start() == 0.0) {
1190 return true;
1191 }
1192 } else {
1193 // We want to conservatively collect somewhat early in order
1194 // to try and "bootstrap" our CMS/promotion statistics;
1195 // this branch will not fire after the first successful CMS
1196 // collection because the stats should then be valid.
1197 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1198 if (Verbose && PrintGCDetails) {
1199 gclog_or_tty->print_cr(
1200 " CMSCollector: collect for bootstrapping statistics:"
1201 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1202 _bootstrap_occupancy);
1203 }
1204 return true;
1205 }
1206 }
1207 }
1208
1209 // Otherwise, we start a collection cycle if
1210 // old gen want a collection cycle started. Each may use
1211 // an appropriate criterion for making this decision.
1212 // XXX We need to make sure that the gen expansion
1213 // criterion dovetails well with this. XXX NEED TO FIX THIS
1214 if (_cmsGen->should_concurrent_collect()) {
1215 if (Verbose && PrintGCDetails) {
1216 gclog_or_tty->print_cr("CMS old gen initiated");
1217 }
1218 return true;
1219 }
1220
1221 // We start a collection if we believe an incremental collection may fail;
1222 // this is not likely to be productive in practice because it's probably too
1223 // late anyway.
1224 GenCollectedHeap* gch = GenCollectedHeap::heap();
1225 assert(gch->collector_policy()->is_generation_policy(),
1226 "You may want to check the correctness of the following");
1227 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1228 if (Verbose && PrintGCDetails) {
1229 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1230 }
1231 return true;
1232 }
1233
1234 if (MetaspaceGC::should_concurrent_collect()) {
1235 if (Verbose && PrintGCDetails) {
1236 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1237 }
1238 return true;
1239 }
1240
1241 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1242 if (CMSTriggerInterval >= 0) {
1243 if (CMSTriggerInterval == 0) {
1244 // Trigger always
1245 return true;
1246 }
1247
1248 // Check the CMS time since begin (we do not check the stats validity
1249 // as we want to be able to trigger the first CMS cycle as well)
1250 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1251 if (Verbose && PrintGCDetails) {
1252 if (stats().valid()) {
1253 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1254 stats().cms_time_since_begin());
1255 } else {
1256 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1257 }
1258 }
1259 return true;
1260 }
1261 }
1262
1263 return false;
1264 }
1265
1266 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1267
1268 // Clear _expansion_cause fields of constituent generations
1269 void CMSCollector::clear_expansion_cause() {
1270 _cmsGen->clear_expansion_cause();
1271 }
1272
1273 // We should be conservative in starting a collection cycle. To
1274 // start too eagerly runs the risk of collecting too often in the
1275 // extreme. To collect too rarely falls back on full collections,
1276 // which works, even if not optimum in terms of concurrent work.
1277 // As a work around for too eagerly collecting, use the flag
1278 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1279 // giving the user an easily understandable way of controlling the
1280 // collections.
1281 // We want to start a new collection cycle if any of the following
1282 // conditions hold:
1283 // . our current occupancy exceeds the configured initiating occupancy
1284 // for this generation, or
1285 // . we recently needed to expand this space and have not, since that
1286 // expansion, done a collection of this generation, or
1287 // . the underlying space believes that it may be a good idea to initiate
1288 // a concurrent collection (this may be based on criteria such as the
1289 // following: the space uses linear allocation and linear allocation is
1290 // going to fail, or there is believed to be excessive fragmentation in
1291 // the generation, etc... or ...
1292 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1293 // the case of the old generation; see CR 6543076):
1294 // we may be approaching a point at which allocation requests may fail because
1295 // we will be out of sufficient free space given allocation rate estimates.]
1296 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1297
1298 assert_lock_strong(freelistLock());
1299 if (occupancy() > initiating_occupancy()) {
1300 if (PrintGCDetails && Verbose) {
1301 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1302 short_name(), occupancy(), initiating_occupancy());
1303 }
1304 return true;
1305 }
1306 if (UseCMSInitiatingOccupancyOnly) {
1307 return false;
1308 }
1309 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1310 if (PrintGCDetails && Verbose) {
1311 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1312 short_name());
1313 }
1314 return true;
1315 }
1316 if (_cmsSpace->should_concurrent_collect()) {
1317 if (PrintGCDetails && Verbose) {
1318 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1319 short_name());
1320 }
1321 return true;
1322 }
1323 return false;
1324 }
1325
1326 void ConcurrentMarkSweepGeneration::collect(bool full,
1327 bool clear_all_soft_refs,
1328 size_t size,
1329 bool tlab)
1330 {
1331 collector()->collect(full, clear_all_soft_refs, size, tlab);
1332 }
1333
1334 void CMSCollector::collect(bool full,
1335 bool clear_all_soft_refs,
1336 size_t size,
1337 bool tlab)
1338 {
1339 // The following "if" branch is present for defensive reasons.
1340 // In the current uses of this interface, it can be replaced with:
1341 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1342 // But I am not placing that assert here to allow future
1343 // generality in invoking this interface.
1344 if (GC_locker::is_active()) {
1345 // A consistency test for GC_locker
1346 assert(GC_locker::needs_gc(), "Should have been set already");
1347 // Skip this foreground collection, instead
1348 // expanding the heap if necessary.
1349 // Need the free list locks for the call to free() in compute_new_size()
1350 compute_new_size();
1351 return;
1352 }
1353 acquire_control_and_collect(full, clear_all_soft_refs);
1354 }
1355
1356 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1357 GenCollectedHeap* gch = GenCollectedHeap::heap();
1358 unsigned int gc_count = gch->total_full_collections();
1359 if (gc_count == full_gc_count) {
1360 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1361 _full_gc_requested = true;
1362 _full_gc_cause = cause;
1363 CGC_lock->notify(); // nudge CMS thread
1364 } else {
1365 assert(gc_count > full_gc_count, "Error: causal loop");
1366 }
1367 }
1368
1369 bool CMSCollector::is_external_interruption() {
1370 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1371 return GCCause::is_user_requested_gc(cause) ||
1372 GCCause::is_serviceability_requested_gc(cause);
1373 }
1374
1375 void CMSCollector::report_concurrent_mode_interruption() {
1376 if (is_external_interruption()) {
1377 if (PrintGCDetails) {
1378 gclog_or_tty->print(" (concurrent mode interrupted)");
1379 }
1380 } else {
1381 if (PrintGCDetails) {
1382 gclog_or_tty->print(" (concurrent mode failure)");
1383 }
1384 _gc_tracer_cm->report_concurrent_mode_failure();
1385 }
1386 }
1387
1388
1389 // The foreground and background collectors need to coordinate in order
1390 // to make sure that they do not mutually interfere with CMS collections.
1391 // When a background collection is active,
1392 // the foreground collector may need to take over (preempt) and
1393 // synchronously complete an ongoing collection. Depending on the
1394 // frequency of the background collections and the heap usage
1395 // of the application, this preemption can be seldom or frequent.
1396 // There are only certain
1397 // points in the background collection that the "collection-baton"
1398 // can be passed to the foreground collector.
1399 //
1400 // The foreground collector will wait for the baton before
1401 // starting any part of the collection. The foreground collector
1402 // will only wait at one location.
1403 //
1404 // The background collector will yield the baton before starting a new
1405 // phase of the collection (e.g., before initial marking, marking from roots,
1406 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1407 // of the loop which switches the phases. The background collector does some
1408 // of the phases (initial mark, final re-mark) with the world stopped.
1409 // Because of locking involved in stopping the world,
1410 // the foreground collector should not block waiting for the background
1411 // collector when it is doing a stop-the-world phase. The background
1412 // collector will yield the baton at an additional point just before
1413 // it enters a stop-the-world phase. Once the world is stopped, the
1414 // background collector checks the phase of the collection. If the
1415 // phase has not changed, it proceeds with the collection. If the
1416 // phase has changed, it skips that phase of the collection. See
1417 // the comments on the use of the Heap_lock in collect_in_background().
1418 //
1419 // Variable used in baton passing.
1420 // _foregroundGCIsActive - Set to true by the foreground collector when
1421 // it wants the baton. The foreground clears it when it has finished
1422 // the collection.
1423 // _foregroundGCShouldWait - Set to true by the background collector
1424 // when it is running. The foreground collector waits while
1425 // _foregroundGCShouldWait is true.
1426 // CGC_lock - monitor used to protect access to the above variables
1427 // and to notify the foreground and background collectors.
1428 // _collectorState - current state of the CMS collection.
1429 //
1430 // The foreground collector
1431 // acquires the CGC_lock
1432 // sets _foregroundGCIsActive
1433 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1434 // various locks acquired in preparation for the collection
1435 // are released so as not to block the background collector
1436 // that is in the midst of a collection
1437 // proceeds with the collection
1438 // clears _foregroundGCIsActive
1439 // returns
1440 //
1441 // The background collector in a loop iterating on the phases of the
1442 // collection
1443 // acquires the CGC_lock
1444 // sets _foregroundGCShouldWait
1445 // if _foregroundGCIsActive is set
1446 // clears _foregroundGCShouldWait, notifies _CGC_lock
1447 // waits on _CGC_lock for _foregroundGCIsActive to become false
1448 // and exits the loop.
1449 // otherwise
1450 // proceed with that phase of the collection
1451 // if the phase is a stop-the-world phase,
1452 // yield the baton once more just before enqueueing
1453 // the stop-world CMS operation (executed by the VM thread).
1454 // returns after all phases of the collection are done
1455 //
1456
1457 void CMSCollector::acquire_control_and_collect(bool full,
1458 bool clear_all_soft_refs) {
1459 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1460 assert(!Thread::current()->is_ConcurrentGC_thread(),
1461 "shouldn't try to acquire control from self!");
1462
1463 // Start the protocol for acquiring control of the
1464 // collection from the background collector (aka CMS thread).
1465 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1466 "VM thread should have CMS token");
1467 // Remember the possibly interrupted state of an ongoing
1468 // concurrent collection
1469 CollectorState first_state = _collectorState;
1470
1471 // Signal to a possibly ongoing concurrent collection that
1472 // we want to do a foreground collection.
1473 _foregroundGCIsActive = true;
1474
1475 // release locks and wait for a notify from the background collector
1476 // releasing the locks in only necessary for phases which
1477 // do yields to improve the granularity of the collection.
1478 assert_lock_strong(bitMapLock());
1479 // We need to lock the Free list lock for the space that we are
1480 // currently collecting.
1481 assert(haveFreelistLocks(), "Must be holding free list locks");
1482 bitMapLock()->unlock();
1483 releaseFreelistLocks();
1484 {
1485 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1486 if (_foregroundGCShouldWait) {
1487 // We are going to be waiting for action for the CMS thread;
1488 // it had better not be gone (for instance at shutdown)!
1489 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1490 "CMS thread must be running");
1491 // Wait here until the background collector gives us the go-ahead
1492 ConcurrentMarkSweepThread::clear_CMS_flag(
1493 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1494 // Get a possibly blocked CMS thread going:
1495 // Note that we set _foregroundGCIsActive true above,
1496 // without protection of the CGC_lock.
1497 CGC_lock->notify();
1498 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1499 "Possible deadlock");
1500 while (_foregroundGCShouldWait) {
1501 // wait for notification
1502 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1503 // Possibility of delay/starvation here, since CMS token does
1504 // not know to give priority to VM thread? Actually, i think
1505 // there wouldn't be any delay/starvation, but the proof of
1506 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1507 }
1508 ConcurrentMarkSweepThread::set_CMS_flag(
1509 ConcurrentMarkSweepThread::CMS_vm_has_token);
1510 }
1511 }
1512 // The CMS_token is already held. Get back the other locks.
1513 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1514 "VM thread should have CMS token");
1515 getFreelistLocks();
1516 bitMapLock()->lock_without_safepoint_check();
1517 if (TraceCMSState) {
1518 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1519 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1520 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1521 }
1522
1523 // Inform cms gen if this was due to partial collection failing.
1524 // The CMS gen may use this fact to determine its expansion policy.
1525 GenCollectedHeap* gch = GenCollectedHeap::heap();
1526 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1527 assert(!_cmsGen->incremental_collection_failed(),
1528 "Should have been noticed, reacted to and cleared");
1529 _cmsGen->set_incremental_collection_failed();
1530 }
1531
1532 if (first_state > Idling) {
1533 report_concurrent_mode_interruption();
1534 }
1535
1536 set_did_compact(true);
1537
1538 // If the collection is being acquired from the background
1539 // collector, there may be references on the discovered
1540 // references lists. Abandon those references, since some
1541 // of them may have become unreachable after concurrent
1542 // discovery; the STW compacting collector will redo discovery
1543 // more precisely, without being subject to floating garbage.
1544 // Leaving otherwise unreachable references in the discovered
1545 // lists would require special handling.
1546 ref_processor()->disable_discovery();
1547 ref_processor()->abandon_partial_discovery();
1548 ref_processor()->verify_no_references_recorded();
1549
1550 if (first_state > Idling) {
1551 save_heap_summary();
1552 }
1553
1554 do_compaction_work(clear_all_soft_refs);
1555
1556 // Has the GC time limit been exceeded?
1557 size_t max_eden_size = _young_gen->max_capacity() -
1558 _young_gen->to()->capacity() -
1559 _young_gen->from()->capacity();
1560 GCCause::Cause gc_cause = gch->gc_cause();
1561 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1562 _young_gen->eden()->used(),
1563 _cmsGen->max_capacity(),
1564 max_eden_size,
1565 full,
1566 gc_cause,
1567 gch->collector_policy());
1568
1569 // Reset the expansion cause, now that we just completed
1570 // a collection cycle.
1571 clear_expansion_cause();
1572 _foregroundGCIsActive = false;
1573 return;
1574 }
1575
1576 // Resize the tenured generation
1577 // after obtaining the free list locks for the
1578 // two generations.
1579 void CMSCollector::compute_new_size() {
1580 assert_locked_or_safepoint(Heap_lock);
1581 FreelistLocker z(this);
1582 MetaspaceGC::compute_new_size();
1583 _cmsGen->compute_new_size_free_list();
1584 }
1585
1586 // A work method used by the foreground collector to do
1587 // a mark-sweep-compact.
1588 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1589 GenCollectedHeap* gch = GenCollectedHeap::heap();
1590
1591 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1592 gc_timer->register_gc_start();
1593
1594 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1595 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1596
1597 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1598
1599 // Temporarily widen the span of the weak reference processing to
1600 // the entire heap.
1601 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1602 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1603 // Temporarily, clear the "is_alive_non_header" field of the
1604 // reference processor.
1605 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1606 // Temporarily make reference _processing_ single threaded (non-MT).
1607 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1608 // Temporarily make refs discovery atomic
1609 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1610 // Temporarily make reference _discovery_ single threaded (non-MT)
1611 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1612
1613 ref_processor()->set_enqueuing_is_done(false);
1614 ref_processor()->enable_discovery();
1615 ref_processor()->setup_policy(clear_all_soft_refs);
1616 // If an asynchronous collection finishes, the _modUnionTable is
1617 // all clear. If we are assuming the collection from an asynchronous
1618 // collection, clear the _modUnionTable.
1619 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1620 "_modUnionTable should be clear if the baton was not passed");
1621 _modUnionTable.clear_all();
1622 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1623 "mod union for klasses should be clear if the baton was passed");
1624 _ct->klass_rem_set()->clear_mod_union();
1625
1626 // We must adjust the allocation statistics being maintained
1627 // in the free list space. We do so by reading and clearing
1628 // the sweep timer and updating the block flux rate estimates below.
1629 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1630 if (_inter_sweep_timer.is_active()) {
1631 _inter_sweep_timer.stop();
1632 // Note that we do not use this sample to update the _inter_sweep_estimate.
1633 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1634 _inter_sweep_estimate.padded_average(),
1635 _intra_sweep_estimate.padded_average());
1636 }
1637
1638 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1639 #ifdef ASSERT
1640 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1641 size_t free_size = cms_space->free();
1642 assert(free_size ==
1643 pointer_delta(cms_space->end(), cms_space->compaction_top())
1644 * HeapWordSize,
1645 "All the free space should be compacted into one chunk at top");
1646 assert(cms_space->dictionary()->total_chunk_size(
1647 debug_only(cms_space->freelistLock())) == 0 ||
1648 cms_space->totalSizeInIndexedFreeLists() == 0,
1649 "All the free space should be in a single chunk");
1650 size_t num = cms_space->totalCount();
1651 assert((free_size == 0 && num == 0) ||
1652 (free_size > 0 && (num == 1 || num == 2)),
1653 "There should be at most 2 free chunks after compaction");
1654 #endif // ASSERT
1655 _collectorState = Resetting;
1656 assert(_restart_addr == NULL,
1657 "Should have been NULL'd before baton was passed");
1658 reset(false /* == !concurrent */);
1659 _cmsGen->reset_after_compaction();
1660 _concurrent_cycles_since_last_unload = 0;
1661
1662 // Clear any data recorded in the PLAB chunk arrays.
1663 if (_survivor_plab_array != NULL) {
1664 reset_survivor_plab_arrays();
1665 }
1666
1667 // Adjust the per-size allocation stats for the next epoch.
1668 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1669 // Restart the "inter sweep timer" for the next epoch.
1670 _inter_sweep_timer.reset();
1671 _inter_sweep_timer.start();
1672
1673 gc_timer->register_gc_end();
1674
1675 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1676
1677 // For a mark-sweep-compact, compute_new_size() will be called
1678 // in the heap's do_collection() method.
1679 }
1680
1681 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1682 ContiguousSpace* eden_space = _young_gen->eden();
1683 ContiguousSpace* from_space = _young_gen->from();
1684 ContiguousSpace* to_space = _young_gen->to();
1685 // Eden
1686 if (_eden_chunk_array != NULL) {
1687 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1688 p2i(eden_space->bottom()), p2i(eden_space->top()),
1689 p2i(eden_space->end()), eden_space->capacity());
1690 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1691 "_eden_chunk_capacity=" SIZE_FORMAT,
1692 _eden_chunk_index, _eden_chunk_capacity);
1693 for (size_t i = 0; i < _eden_chunk_index; i++) {
1694 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1695 i, p2i(_eden_chunk_array[i]));
1696 }
1697 }
1698 // Survivor
1699 if (_survivor_chunk_array != NULL) {
1700 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1701 p2i(from_space->bottom()), p2i(from_space->top()),
1702 p2i(from_space->end()), from_space->capacity());
1703 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1704 "_survivor_chunk_capacity=" SIZE_FORMAT,
1705 _survivor_chunk_index, _survivor_chunk_capacity);
1706 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1707 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1708 i, p2i(_survivor_chunk_array[i]));
1709 }
1710 }
1711 }
1712
1713 void CMSCollector::getFreelistLocks() const {
1714 // Get locks for all free lists in all generations that this
1715 // collector is responsible for
1716 _cmsGen->freelistLock()->lock_without_safepoint_check();
1717 }
1718
1719 void CMSCollector::releaseFreelistLocks() const {
1720 // Release locks for all free lists in all generations that this
1721 // collector is responsible for
1722 _cmsGen->freelistLock()->unlock();
1723 }
1724
1725 bool CMSCollector::haveFreelistLocks() const {
1726 // Check locks for all free lists in all generations that this
1727 // collector is responsible for
1728 assert_lock_strong(_cmsGen->freelistLock());
1729 PRODUCT_ONLY(ShouldNotReachHere());
1730 return true;
1731 }
1732
1733 // A utility class that is used by the CMS collector to
1734 // temporarily "release" the foreground collector from its
1735 // usual obligation to wait for the background collector to
1736 // complete an ongoing phase before proceeding.
1737 class ReleaseForegroundGC: public StackObj {
1738 private:
1739 CMSCollector* _c;
1740 public:
1741 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1742 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1743 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1744 // allow a potentially blocked foreground collector to proceed
1745 _c->_foregroundGCShouldWait = false;
1746 if (_c->_foregroundGCIsActive) {
1747 CGC_lock->notify();
1748 }
1749 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1750 "Possible deadlock");
1751 }
1752
1753 ~ReleaseForegroundGC() {
1754 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1755 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1756 _c->_foregroundGCShouldWait = true;
1757 }
1758 };
1759
1760 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1761 assert(Thread::current()->is_ConcurrentGC_thread(),
1762 "A CMS asynchronous collection is only allowed on a CMS thread.");
1763
1764 GenCollectedHeap* gch = GenCollectedHeap::heap();
1765 {
1766 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1767 MutexLockerEx hl(Heap_lock, safepoint_check);
1768 FreelistLocker fll(this);
1769 MutexLockerEx x(CGC_lock, safepoint_check);
1770 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1771 // The foreground collector is active or we're
1772 // not using asynchronous collections. Skip this
1773 // background collection.
1774 assert(!_foregroundGCShouldWait, "Should be clear");
1775 return;
1776 } else {
1777 assert(_collectorState == Idling, "Should be idling before start.");
1778 _collectorState = InitialMarking;
1779 register_gc_start(cause);
1780 // Reset the expansion cause, now that we are about to begin
1781 // a new cycle.
1782 clear_expansion_cause();
1783
1784 // Clear the MetaspaceGC flag since a concurrent collection
1785 // is starting but also clear it after the collection.
1786 MetaspaceGC::set_should_concurrent_collect(false);
1787 }
1788 // Decide if we want to enable class unloading as part of the
1789 // ensuing concurrent GC cycle.
1790 update_should_unload_classes();
1791 _full_gc_requested = false; // acks all outstanding full gc requests
1792 _full_gc_cause = GCCause::_no_gc;
1793 // Signal that we are about to start a collection
1794 gch->increment_total_full_collections(); // ... starting a collection cycle
1795 _collection_count_start = gch->total_full_collections();
1796 }
1797
1798 // Used for PrintGC
1799 size_t prev_used;
1800 if (PrintGC && Verbose) {
1801 prev_used = _cmsGen->used();
1802 }
1803
1804 // The change of the collection state is normally done at this level;
1805 // the exceptions are phases that are executed while the world is
1806 // stopped. For those phases the change of state is done while the
1807 // world is stopped. For baton passing purposes this allows the
1808 // background collector to finish the phase and change state atomically.
1809 // The foreground collector cannot wait on a phase that is done
1810 // while the world is stopped because the foreground collector already
1811 // has the world stopped and would deadlock.
1812 while (_collectorState != Idling) {
1813 if (TraceCMSState) {
1814 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1815 p2i(Thread::current()), _collectorState);
1816 }
1817 // The foreground collector
1818 // holds the Heap_lock throughout its collection.
1819 // holds the CMS token (but not the lock)
1820 // except while it is waiting for the background collector to yield.
1821 //
1822 // The foreground collector should be blocked (not for long)
1823 // if the background collector is about to start a phase
1824 // executed with world stopped. If the background
1825 // collector has already started such a phase, the
1826 // foreground collector is blocked waiting for the
1827 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1828 // are executed in the VM thread.
1829 //
1830 // The locking order is
1831 // PendingListLock (PLL) -- if applicable (FinalMarking)
1832 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1833 // CMS token (claimed in
1834 // stop_world_and_do() -->
1835 // safepoint_synchronize() -->
1836 // CMSThread::synchronize())
1837
1838 {
1839 // Check if the FG collector wants us to yield.
1840 CMSTokenSync x(true); // is cms thread
1841 if (waitForForegroundGC()) {
1842 // We yielded to a foreground GC, nothing more to be
1843 // done this round.
1844 assert(_foregroundGCShouldWait == false, "We set it to false in "
1845 "waitForForegroundGC()");
1846 if (TraceCMSState) {
1847 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1848 " exiting collection CMS state %d",
1849 p2i(Thread::current()), _collectorState);
1850 }
1851 return;
1852 } else {
1853 // The background collector can run but check to see if the
1854 // foreground collector has done a collection while the
1855 // background collector was waiting to get the CGC_lock
1856 // above. If yes, break so that _foregroundGCShouldWait
1857 // is cleared before returning.
1858 if (_collectorState == Idling) {
1859 break;
1860 }
1861 }
1862 }
1863
1864 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1865 "should be waiting");
1866
1867 switch (_collectorState) {
1868 case InitialMarking:
1869 {
1870 ReleaseForegroundGC x(this);
1871 stats().record_cms_begin();
1872 VM_CMS_Initial_Mark initial_mark_op(this);
1873 VMThread::execute(&initial_mark_op);
1874 }
1875 // The collector state may be any legal state at this point
1876 // since the background collector may have yielded to the
1877 // foreground collector.
1878 break;
1879 case Marking:
1880 // initial marking in checkpointRootsInitialWork has been completed
1881 if (markFromRoots()) { // we were successful
1882 assert(_collectorState == Precleaning, "Collector state should "
1883 "have changed");
1884 } else {
1885 assert(_foregroundGCIsActive, "Internal state inconsistency");
1886 }
1887 break;
1888 case Precleaning:
1889 // marking from roots in markFromRoots has been completed
1890 preclean();
1891 assert(_collectorState == AbortablePreclean ||
1892 _collectorState == FinalMarking,
1893 "Collector state should have changed");
1894 break;
1895 case AbortablePreclean:
1896 abortable_preclean();
1897 assert(_collectorState == FinalMarking, "Collector state should "
1898 "have changed");
1899 break;
1900 case FinalMarking:
1901 {
1902 ReleaseForegroundGC x(this);
1903
1904 VM_CMS_Final_Remark final_remark_op(this);
1905 VMThread::execute(&final_remark_op);
1906 }
1907 assert(_foregroundGCShouldWait, "block post-condition");
1908 break;
1909 case Sweeping:
1910 // final marking in checkpointRootsFinal has been completed
1911 sweep();
1912 assert(_collectorState == Resizing, "Collector state change "
1913 "to Resizing must be done under the free_list_lock");
1914
1915 case Resizing: {
1916 // Sweeping has been completed...
1917 // At this point the background collection has completed.
1918 // Don't move the call to compute_new_size() down
1919 // into code that might be executed if the background
1920 // collection was preempted.
1921 {
1922 ReleaseForegroundGC x(this); // unblock FG collection
1923 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1924 CMSTokenSync z(true); // not strictly needed.
1925 if (_collectorState == Resizing) {
1926 compute_new_size();
1927 save_heap_summary();
1928 _collectorState = Resetting;
1929 } else {
1930 assert(_collectorState == Idling, "The state should only change"
1931 " because the foreground collector has finished the collection");
1932 }
1933 }
1934 break;
1935 }
1936 case Resetting:
1937 // CMS heap resizing has been completed
1938 reset(true);
1939 assert(_collectorState == Idling, "Collector state should "
1940 "have changed");
1941
1942 MetaspaceGC::set_should_concurrent_collect(false);
1943
1944 stats().record_cms_end();
1945 // Don't move the concurrent_phases_end() and compute_new_size()
1946 // calls to here because a preempted background collection
1947 // has it's state set to "Resetting".
1948 break;
1949 case Idling:
1950 default:
1951 ShouldNotReachHere();
1952 break;
1953 }
1954 if (TraceCMSState) {
1955 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1956 p2i(Thread::current()), _collectorState);
1957 }
1958 assert(_foregroundGCShouldWait, "block post-condition");
1959 }
1960
1961 // Should this be in gc_epilogue?
1962 collector_policy()->counters()->update_counters();
1963
1964 {
1965 // Clear _foregroundGCShouldWait and, in the event that the
1966 // foreground collector is waiting, notify it, before
1967 // returning.
1968 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1969 _foregroundGCShouldWait = false;
1970 if (_foregroundGCIsActive) {
1971 CGC_lock->notify();
1972 }
1973 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1974 "Possible deadlock");
1975 }
1976 if (TraceCMSState) {
1977 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1978 " exiting collection CMS state %d",
1979 p2i(Thread::current()), _collectorState);
1980 }
1981 if (PrintGC && Verbose) {
1982 _cmsGen->print_heap_change(prev_used);
1983 }
1984 }
1985
1986 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1987 _cms_start_registered = true;
1988 _gc_timer_cm->register_gc_start();
1989 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1990 }
1991
1992 void CMSCollector::register_gc_end() {
1993 if (_cms_start_registered) {
1994 report_heap_summary(GCWhen::AfterGC);
1995
1996 _gc_timer_cm->register_gc_end();
1997 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1998 _cms_start_registered = false;
1999 }
2000 }
2001
2002 void CMSCollector::save_heap_summary() {
2003 GenCollectedHeap* gch = GenCollectedHeap::heap();
2004 _last_heap_summary = gch->create_heap_summary();
2005 _last_metaspace_summary = gch->create_metaspace_summary();
2006 }
2007
2008 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2009 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2010 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2011 }
2012
2013 bool CMSCollector::waitForForegroundGC() {
2014 bool res = false;
2015 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2016 "CMS thread should have CMS token");
2017 // Block the foreground collector until the
2018 // background collectors decides whether to
2019 // yield.
2020 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2021 _foregroundGCShouldWait = true;
2022 if (_foregroundGCIsActive) {
2023 // The background collector yields to the
2024 // foreground collector and returns a value
2025 // indicating that it has yielded. The foreground
2026 // collector can proceed.
2027 res = true;
2028 _foregroundGCShouldWait = false;
2029 ConcurrentMarkSweepThread::clear_CMS_flag(
2030 ConcurrentMarkSweepThread::CMS_cms_has_token);
2031 ConcurrentMarkSweepThread::set_CMS_flag(
2032 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2033 // Get a possibly blocked foreground thread going
2034 CGC_lock->notify();
2035 if (TraceCMSState) {
2036 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2037 p2i(Thread::current()), _collectorState);
2038 }
2039 while (_foregroundGCIsActive) {
2040 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2041 }
2042 ConcurrentMarkSweepThread::set_CMS_flag(
2043 ConcurrentMarkSweepThread::CMS_cms_has_token);
2044 ConcurrentMarkSweepThread::clear_CMS_flag(
2045 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2046 }
2047 if (TraceCMSState) {
2048 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2049 p2i(Thread::current()), _collectorState);
2050 }
2051 return res;
2052 }
2053
2054 // Because of the need to lock the free lists and other structures in
2055 // the collector, common to all the generations that the collector is
2056 // collecting, we need the gc_prologues of individual CMS generations
2057 // delegate to their collector. It may have been simpler had the
2058 // current infrastructure allowed one to call a prologue on a
2059 // collector. In the absence of that we have the generation's
2060 // prologue delegate to the collector, which delegates back
2061 // some "local" work to a worker method in the individual generations
2062 // that it's responsible for collecting, while itself doing any
2063 // work common to all generations it's responsible for. A similar
2064 // comment applies to the gc_epilogue()'s.
2065 // The role of the variable _between_prologue_and_epilogue is to
2066 // enforce the invocation protocol.
2067 void CMSCollector::gc_prologue(bool full) {
2068 // Call gc_prologue_work() for the CMSGen
2069 // we are responsible for.
2070
2071 // The following locking discipline assumes that we are only called
2072 // when the world is stopped.
2073 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2074
2075 // The CMSCollector prologue must call the gc_prologues for the
2076 // "generations" that it's responsible
2077 // for.
2078
2079 assert( Thread::current()->is_VM_thread()
2080 || ( CMSScavengeBeforeRemark
2081 && Thread::current()->is_ConcurrentGC_thread()),
2082 "Incorrect thread type for prologue execution");
2083
2084 if (_between_prologue_and_epilogue) {
2085 // We have already been invoked; this is a gc_prologue delegation
2086 // from yet another CMS generation that we are responsible for, just
2087 // ignore it since all relevant work has already been done.
2088 return;
2089 }
2090
2091 // set a bit saying prologue has been called; cleared in epilogue
2092 _between_prologue_and_epilogue = true;
2093 // Claim locks for common data structures, then call gc_prologue_work()
2094 // for each CMSGen.
2095
2096 getFreelistLocks(); // gets free list locks on constituent spaces
2097 bitMapLock()->lock_without_safepoint_check();
2098
2099 // Should call gc_prologue_work() for all cms gens we are responsible for
2100 bool duringMarking = _collectorState >= Marking
2101 && _collectorState < Sweeping;
2102
2103 // The young collections clear the modified oops state, which tells if
2104 // there are any modified oops in the class. The remark phase also needs
2105 // that information. Tell the young collection to save the union of all
2106 // modified klasses.
2107 if (duringMarking) {
2108 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2109 }
2110
2111 bool registerClosure = duringMarking;
2112
2113 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2114
2115 if (!full) {
2116 stats().record_gc0_begin();
2117 }
2118 }
2119
2120 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2121
2122 _capacity_at_prologue = capacity();
2123 _used_at_prologue = used();
2124
2125 // Delegate to CMScollector which knows how to coordinate between
2126 // this and any other CMS generations that it is responsible for
2127 // collecting.
2128 collector()->gc_prologue(full);
2129 }
2130
2131 // This is a "private" interface for use by this generation's CMSCollector.
2132 // Not to be called directly by any other entity (for instance,
2133 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2134 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2135 bool registerClosure, ModUnionClosure* modUnionClosure) {
2136 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2137 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2138 "Should be NULL");
2139 if (registerClosure) {
2140 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2141 }
2142 cmsSpace()->gc_prologue();
2143 // Clear stat counters
2144 NOT_PRODUCT(
2145 assert(_numObjectsPromoted == 0, "check");
2146 assert(_numWordsPromoted == 0, "check");
2147 if (Verbose && PrintGC) {
2148 gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, "
2149 SIZE_FORMAT " bytes concurrently",
2150 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2151 }
2152 _numObjectsAllocated = 0;
2153 _numWordsAllocated = 0;
2154 )
2155 }
2156
2157 void CMSCollector::gc_epilogue(bool full) {
2158 // The following locking discipline assumes that we are only called
2159 // when the world is stopped.
2160 assert(SafepointSynchronize::is_at_safepoint(),
2161 "world is stopped assumption");
2162
2163 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2164 // if linear allocation blocks need to be appropriately marked to allow the
2165 // the blocks to be parsable. We also check here whether we need to nudge the
2166 // CMS collector thread to start a new cycle (if it's not already active).
2167 assert( Thread::current()->is_VM_thread()
2168 || ( CMSScavengeBeforeRemark
2169 && Thread::current()->is_ConcurrentGC_thread()),
2170 "Incorrect thread type for epilogue execution");
2171
2172 if (!_between_prologue_and_epilogue) {
2173 // We have already been invoked; this is a gc_epilogue delegation
2174 // from yet another CMS generation that we are responsible for, just
2175 // ignore it since all relevant work has already been done.
2176 return;
2177 }
2178 assert(haveFreelistLocks(), "must have freelist locks");
2179 assert_lock_strong(bitMapLock());
2180
2181 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2182
2183 _cmsGen->gc_epilogue_work(full);
2184
2185 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2186 // in case sampling was not already enabled, enable it
2187 _start_sampling = true;
2188 }
2189 // reset _eden_chunk_array so sampling starts afresh
2190 _eden_chunk_index = 0;
2191
2192 size_t cms_used = _cmsGen->cmsSpace()->used();
2193
2194 // update performance counters - this uses a special version of
2195 // update_counters() that allows the utilization to be passed as a
2196 // parameter, avoiding multiple calls to used().
2197 //
2198 _cmsGen->update_counters(cms_used);
2199
2200 bitMapLock()->unlock();
2201 releaseFreelistLocks();
2202
2203 if (!CleanChunkPoolAsync) {
2204 Chunk::clean_chunk_pool();
2205 }
2206
2207 set_did_compact(false);
2208 _between_prologue_and_epilogue = false; // ready for next cycle
2209 }
2210
2211 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2212 collector()->gc_epilogue(full);
2213
2214 // Also reset promotion tracking in par gc thread states.
2215 for (uint i = 0; i < ParallelGCThreads; i++) {
2216 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2217 }
2218 }
2219
2220 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2221 assert(!incremental_collection_failed(), "Should have been cleared");
2222 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2223 cmsSpace()->gc_epilogue();
2224 // Print stat counters
2225 NOT_PRODUCT(
2226 assert(_numObjectsAllocated == 0, "check");
2227 assert(_numWordsAllocated == 0, "check");
2228 if (Verbose && PrintGC) {
2229 gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, "
2230 SIZE_FORMAT " bytes",
2231 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2232 }
2233 _numObjectsPromoted = 0;
2234 _numWordsPromoted = 0;
2235 )
2236
2237 if (PrintGC && Verbose) {
2238 // Call down the chain in contiguous_available needs the freelistLock
2239 // so print this out before releasing the freeListLock.
2240 gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ",
2241 contiguous_available());
2242 }
2243 }
2244
2245 #ifndef PRODUCT
2246 bool CMSCollector::have_cms_token() {
2247 Thread* thr = Thread::current();
2248 if (thr->is_VM_thread()) {
2249 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2250 } else if (thr->is_ConcurrentGC_thread()) {
2251 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2252 } else if (thr->is_GC_task_thread()) {
2253 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2254 ParGCRareEvent_lock->owned_by_self();
2255 }
2256 return false;
2257 }
2258 #endif
2259
2260 // Check reachability of the given heap address in CMS generation,
2261 // treating all other generations as roots.
2262 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2263 // We could "guarantee" below, rather than assert, but I'll
2264 // leave these as "asserts" so that an adventurous debugger
2265 // could try this in the product build provided some subset of
2266 // the conditions were met, provided they were interested in the
2267 // results and knew that the computation below wouldn't interfere
2268 // with other concurrent computations mutating the structures
2269 // being read or written.
2270 assert(SafepointSynchronize::is_at_safepoint(),
2271 "Else mutations in object graph will make answer suspect");
2272 assert(have_cms_token(), "Should hold cms token");
2273 assert(haveFreelistLocks(), "must hold free list locks");
2274 assert_lock_strong(bitMapLock());
2275
2276 // Clear the marking bit map array before starting, but, just
2277 // for kicks, first report if the given address is already marked
2278 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2279 _markBitMap.isMarked(addr) ? "" : " not");
2280
2281 if (verify_after_remark()) {
2282 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2283 bool result = verification_mark_bm()->isMarked(addr);
2284 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2285 result ? "IS" : "is NOT");
2286 return result;
2287 } else {
2288 gclog_or_tty->print_cr("Could not compute result");
2289 return false;
2290 }
2291 }
2292
2293
2294 void
2295 CMSCollector::print_on_error(outputStream* st) {
2296 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2297 if (collector != NULL) {
2298 CMSBitMap* bitmap = &collector->_markBitMap;
2299 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2300 bitmap->print_on_error(st, " Bits: ");
2301
2302 st->cr();
2303
2304 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2305 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2306 mut_bitmap->print_on_error(st, " Bits: ");
2307 }
2308 }
2309
2310 ////////////////////////////////////////////////////////
2311 // CMS Verification Support
2312 ////////////////////////////////////////////////////////
2313 // Following the remark phase, the following invariant
2314 // should hold -- each object in the CMS heap which is
2315 // marked in markBitMap() should be marked in the verification_mark_bm().
2316
2317 class VerifyMarkedClosure: public BitMapClosure {
2318 CMSBitMap* _marks;
2319 bool _failed;
2320
2321 public:
2322 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2323
2324 bool do_bit(size_t offset) {
2325 HeapWord* addr = _marks->offsetToHeapWord(offset);
2326 if (!_marks->isMarked(addr)) {
2327 oop(addr)->print_on(gclog_or_tty);
2328 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2329 _failed = true;
2330 }
2331 return true;
2332 }
2333
2334 bool failed() { return _failed; }
2335 };
2336
2337 bool CMSCollector::verify_after_remark(bool silent) {
2338 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2339 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2340 static bool init = false;
2341
2342 assert(SafepointSynchronize::is_at_safepoint(),
2343 "Else mutations in object graph will make answer suspect");
2344 assert(have_cms_token(),
2345 "Else there may be mutual interference in use of "
2346 " verification data structures");
2347 assert(_collectorState > Marking && _collectorState <= Sweeping,
2348 "Else marking info checked here may be obsolete");
2349 assert(haveFreelistLocks(), "must hold free list locks");
2350 assert_lock_strong(bitMapLock());
2351
2352
2353 // Allocate marking bit map if not already allocated
2354 if (!init) { // first time
2355 if (!verification_mark_bm()->allocate(_span)) {
2356 return false;
2357 }
2358 init = true;
2359 }
2360
2361 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2362
2363 // Turn off refs discovery -- so we will be tracing through refs.
2364 // This is as intended, because by this time
2365 // GC must already have cleared any refs that need to be cleared,
2366 // and traced those that need to be marked; moreover,
2367 // the marking done here is not going to interfere in any
2368 // way with the marking information used by GC.
2369 NoRefDiscovery no_discovery(ref_processor());
2370
2371 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2372
2373 // Clear any marks from a previous round
2374 verification_mark_bm()->clear_all();
2375 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2376 verify_work_stacks_empty();
2377
2378 GenCollectedHeap* gch = GenCollectedHeap::heap();
2379 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2380 // Update the saved marks which may affect the root scans.
2381 gch->save_marks();
2382
2383 if (CMSRemarkVerifyVariant == 1) {
2384 // In this first variant of verification, we complete
2385 // all marking, then check if the new marks-vector is
2386 // a subset of the CMS marks-vector.
2387 verify_after_remark_work_1();
2388 } else if (CMSRemarkVerifyVariant == 2) {
2389 // In this second variant of verification, we flag an error
2390 // (i.e. an object reachable in the new marks-vector not reachable
2391 // in the CMS marks-vector) immediately, also indicating the
2392 // identify of an object (A) that references the unmarked object (B) --
2393 // presumably, a mutation to A failed to be picked up by preclean/remark?
2394 verify_after_remark_work_2();
2395 } else {
2396 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2397 CMSRemarkVerifyVariant);
2398 }
2399 if (!silent) gclog_or_tty->print(" done] ");
2400 return true;
2401 }
2402
2403 void CMSCollector::verify_after_remark_work_1() {
2404 ResourceMark rm;
2405 HandleMark hm;
2406 GenCollectedHeap* gch = GenCollectedHeap::heap();
2407
2408 // Get a clear set of claim bits for the roots processing to work with.
2409 ClassLoaderDataGraph::clear_claimed_marks();
2410
2411 // Mark from roots one level into CMS
2412 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2413 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2414
2415 {
2416 StrongRootsScope srs(1);
2417
2418 gch->gen_process_roots(&srs,
2419 GenCollectedHeap::OldGen,
2420 true, // young gen as roots
2421 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2422 should_unload_classes(),
2423 ¬Older,
2424 NULL,
2425 NULL);
2426 }
2427
2428 // Now mark from the roots
2429 MarkFromRootsClosure markFromRootsClosure(this, _span,
2430 verification_mark_bm(), verification_mark_stack(),
2431 false /* don't yield */, true /* verifying */);
2432 assert(_restart_addr == NULL, "Expected pre-condition");
2433 verification_mark_bm()->iterate(&markFromRootsClosure);
2434 while (_restart_addr != NULL) {
2435 // Deal with stack overflow: by restarting at the indicated
2436 // address.
2437 HeapWord* ra = _restart_addr;
2438 markFromRootsClosure.reset(ra);
2439 _restart_addr = NULL;
2440 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2441 }
2442 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2443 verify_work_stacks_empty();
2444
2445 // Marking completed -- now verify that each bit marked in
2446 // verification_mark_bm() is also marked in markBitMap(); flag all
2447 // errors by printing corresponding objects.
2448 VerifyMarkedClosure vcl(markBitMap());
2449 verification_mark_bm()->iterate(&vcl);
2450 if (vcl.failed()) {
2451 gclog_or_tty->print("Verification failed");
2452 gch->print_on(gclog_or_tty);
2453 fatal("CMS: failed marking verification after remark");
2454 }
2455 }
2456
2457 class VerifyKlassOopsKlassClosure : public KlassClosure {
2458 class VerifyKlassOopsClosure : public OopClosure {
2459 CMSBitMap* _bitmap;
2460 public:
2461 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2462 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2463 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2464 } _oop_closure;
2465 public:
2466 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2467 void do_klass(Klass* k) {
2468 k->oops_do(&_oop_closure);
2469 }
2470 };
2471
2472 void CMSCollector::verify_after_remark_work_2() {
2473 ResourceMark rm;
2474 HandleMark hm;
2475 GenCollectedHeap* gch = GenCollectedHeap::heap();
2476
2477 // Get a clear set of claim bits for the roots processing to work with.
2478 ClassLoaderDataGraph::clear_claimed_marks();
2479
2480 // Mark from roots one level into CMS
2481 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2482 markBitMap());
2483 CLDToOopClosure cld_closure(¬Older, true);
2484
2485 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2486
2487 {
2488 StrongRootsScope srs(1);
2489
2490 gch->gen_process_roots(&srs,
2491 GenCollectedHeap::OldGen,
2492 true, // young gen as roots
2493 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2494 should_unload_classes(),
2495 ¬Older,
2496 NULL,
2497 &cld_closure);
2498 }
2499
2500 // Now mark from the roots
2501 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2502 verification_mark_bm(), markBitMap(), verification_mark_stack());
2503 assert(_restart_addr == NULL, "Expected pre-condition");
2504 verification_mark_bm()->iterate(&markFromRootsClosure);
2505 while (_restart_addr != NULL) {
2506 // Deal with stack overflow: by restarting at the indicated
2507 // address.
2508 HeapWord* ra = _restart_addr;
2509 markFromRootsClosure.reset(ra);
2510 _restart_addr = NULL;
2511 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2512 }
2513 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2514 verify_work_stacks_empty();
2515
2516 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2517 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2518
2519 // Marking completed -- now verify that each bit marked in
2520 // verification_mark_bm() is also marked in markBitMap(); flag all
2521 // errors by printing corresponding objects.
2522 VerifyMarkedClosure vcl(markBitMap());
2523 verification_mark_bm()->iterate(&vcl);
2524 assert(!vcl.failed(), "Else verification above should not have succeeded");
2525 }
2526
2527 void ConcurrentMarkSweepGeneration::save_marks() {
2528 // delegate to CMS space
2529 cmsSpace()->save_marks();
2530 for (uint i = 0; i < ParallelGCThreads; i++) {
2531 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2532 }
2533 }
2534
2535 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2536 return cmsSpace()->no_allocs_since_save_marks();
2537 }
2538
2539 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2540 \
2541 void ConcurrentMarkSweepGeneration:: \
2542 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2543 cl->set_generation(this); \
2544 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2545 cl->reset_generation(); \
2546 save_marks(); \
2547 }
2548
2549 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2550
2551 void
2552 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2553 if (freelistLock()->owned_by_self()) {
2554 Generation::oop_iterate(cl);
2555 } else {
2556 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2557 Generation::oop_iterate(cl);
2558 }
2559 }
2560
2561 void
2562 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2563 if (freelistLock()->owned_by_self()) {
2564 Generation::object_iterate(cl);
2565 } else {
2566 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2567 Generation::object_iterate(cl);
2568 }
2569 }
2570
2571 void
2572 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2573 if (freelistLock()->owned_by_self()) {
2574 Generation::safe_object_iterate(cl);
2575 } else {
2576 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2577 Generation::safe_object_iterate(cl);
2578 }
2579 }
2580
2581 void
2582 ConcurrentMarkSweepGeneration::post_compact() {
2583 }
2584
2585 void
2586 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2587 // Fix the linear allocation blocks to look like free blocks.
2588
2589 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2590 // are not called when the heap is verified during universe initialization and
2591 // at vm shutdown.
2592 if (freelistLock()->owned_by_self()) {
2593 cmsSpace()->prepare_for_verify();
2594 } else {
2595 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2596 cmsSpace()->prepare_for_verify();
2597 }
2598 }
2599
2600 void
2601 ConcurrentMarkSweepGeneration::verify() {
2602 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2603 // are not called when the heap is verified during universe initialization and
2604 // at vm shutdown.
2605 if (freelistLock()->owned_by_self()) {
2606 cmsSpace()->verify();
2607 } else {
2608 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2609 cmsSpace()->verify();
2610 }
2611 }
2612
2613 void CMSCollector::verify() {
2614 _cmsGen->verify();
2615 }
2616
2617 #ifndef PRODUCT
2618 bool CMSCollector::overflow_list_is_empty() const {
2619 assert(_num_par_pushes >= 0, "Inconsistency");
2620 if (_overflow_list == NULL) {
2621 assert(_num_par_pushes == 0, "Inconsistency");
2622 }
2623 return _overflow_list == NULL;
2624 }
2625
2626 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2627 // merely consolidate assertion checks that appear to occur together frequently.
2628 void CMSCollector::verify_work_stacks_empty() const {
2629 assert(_markStack.isEmpty(), "Marking stack should be empty");
2630 assert(overflow_list_is_empty(), "Overflow list should be empty");
2631 }
2632
2633 void CMSCollector::verify_overflow_empty() const {
2634 assert(overflow_list_is_empty(), "Overflow list should be empty");
2635 assert(no_preserved_marks(), "No preserved marks");
2636 }
2637 #endif // PRODUCT
2638
2639 // Decide if we want to enable class unloading as part of the
2640 // ensuing concurrent GC cycle. We will collect and
2641 // unload classes if it's the case that:
2642 // (1) an explicit gc request has been made and the flag
2643 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2644 // (2) (a) class unloading is enabled at the command line, and
2645 // (b) old gen is getting really full
2646 // NOTE: Provided there is no change in the state of the heap between
2647 // calls to this method, it should have idempotent results. Moreover,
2648 // its results should be monotonically increasing (i.e. going from 0 to 1,
2649 // but not 1 to 0) between successive calls between which the heap was
2650 // not collected. For the implementation below, it must thus rely on
2651 // the property that concurrent_cycles_since_last_unload()
2652 // will not decrease unless a collection cycle happened and that
2653 // _cmsGen->is_too_full() are
2654 // themselves also monotonic in that sense. See check_monotonicity()
2655 // below.
2656 void CMSCollector::update_should_unload_classes() {
2657 _should_unload_classes = false;
2658 // Condition 1 above
2659 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2660 _should_unload_classes = true;
2661 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2662 // Disjuncts 2.b.(i,ii,iii) above
2663 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2664 CMSClassUnloadingMaxInterval)
2665 || _cmsGen->is_too_full();
2666 }
2667 }
2668
2669 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2670 bool res = should_concurrent_collect();
2671 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2672 return res;
2673 }
2674
2675 void CMSCollector::setup_cms_unloading_and_verification_state() {
2676 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2677 || VerifyBeforeExit;
2678 const int rso = GenCollectedHeap::SO_AllCodeCache;
2679
2680 // We set the proper root for this CMS cycle here.
2681 if (should_unload_classes()) { // Should unload classes this cycle
2682 remove_root_scanning_option(rso); // Shrink the root set appropriately
2683 set_verifying(should_verify); // Set verification state for this cycle
2684 return; // Nothing else needs to be done at this time
2685 }
2686
2687 // Not unloading classes this cycle
2688 assert(!should_unload_classes(), "Inconsistency!");
2689
2690 // If we are not unloading classes then add SO_AllCodeCache to root
2691 // scanning options.
2692 add_root_scanning_option(rso);
2693
2694 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2695 set_verifying(true);
2696 } else if (verifying() && !should_verify) {
2697 // We were verifying, but some verification flags got disabled.
2698 set_verifying(false);
2699 // Exclude symbols, strings and code cache elements from root scanning to
2700 // reduce IM and RM pauses.
2701 remove_root_scanning_option(rso);
2702 }
2703 }
2704
2705
2706 #ifndef PRODUCT
2707 HeapWord* CMSCollector::block_start(const void* p) const {
2708 const HeapWord* addr = (HeapWord*)p;
2709 if (_span.contains(p)) {
2710 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2711 return _cmsGen->cmsSpace()->block_start(p);
2712 }
2713 }
2714 return NULL;
2715 }
2716 #endif
2717
2718 HeapWord*
2719 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2720 bool tlab,
2721 bool parallel) {
2722 CMSSynchronousYieldRequest yr;
2723 assert(!tlab, "Can't deal with TLAB allocation");
2724 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2725 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2726 if (GCExpandToAllocateDelayMillis > 0) {
2727 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2728 }
2729 return have_lock_and_allocate(word_size, tlab);
2730 }
2731
2732 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2733 size_t bytes,
2734 size_t expand_bytes,
2735 CMSExpansionCause::Cause cause)
2736 {
2737
2738 bool success = expand(bytes, expand_bytes);
2739
2740 // remember why we expanded; this information is used
2741 // by shouldConcurrentCollect() when making decisions on whether to start
2742 // a new CMS cycle.
2743 if (success) {
2744 set_expansion_cause(cause);
2745 if (PrintGCDetails && Verbose) {
2746 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2747 CMSExpansionCause::to_string(cause));
2748 }
2749 }
2750 }
2751
2752 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2753 HeapWord* res = NULL;
2754 MutexLocker x(ParGCRareEvent_lock);
2755 while (true) {
2756 // Expansion by some other thread might make alloc OK now:
2757 res = ps->lab.alloc(word_sz);
2758 if (res != NULL) return res;
2759 // If there's not enough expansion space available, give up.
2760 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2761 return NULL;
2762 }
2763 // Otherwise, we try expansion.
2764 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2765 // Now go around the loop and try alloc again;
2766 // A competing par_promote might beat us to the expansion space,
2767 // so we may go around the loop again if promotion fails again.
2768 if (GCExpandToAllocateDelayMillis > 0) {
2769 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2770 }
2771 }
2772 }
2773
2774
2775 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2776 PromotionInfo* promo) {
2777 MutexLocker x(ParGCRareEvent_lock);
2778 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2779 while (true) {
2780 // Expansion by some other thread might make alloc OK now:
2781 if (promo->ensure_spooling_space()) {
2782 assert(promo->has_spooling_space(),
2783 "Post-condition of successful ensure_spooling_space()");
2784 return true;
2785 }
2786 // If there's not enough expansion space available, give up.
2787 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2788 return false;
2789 }
2790 // Otherwise, we try expansion.
2791 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2792 // Now go around the loop and try alloc again;
2793 // A competing allocation might beat us to the expansion space,
2794 // so we may go around the loop again if allocation fails again.
2795 if (GCExpandToAllocateDelayMillis > 0) {
2796 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2797 }
2798 }
2799 }
2800
2801 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2802 // Only shrink if a compaction was done so that all the free space
2803 // in the generation is in a contiguous block at the end.
2804 if (did_compact()) {
2805 CardGeneration::shrink(bytes);
2806 }
2807 }
2808
2809 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2810 assert_locked_or_safepoint(Heap_lock);
2811 }
2812
2813 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2814 assert_locked_or_safepoint(Heap_lock);
2815 assert_lock_strong(freelistLock());
2816 if (PrintGCDetails && Verbose) {
2817 warning("Shrinking of CMS not yet implemented");
2818 }
2819 return;
2820 }
2821
2822
2823 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2824 // phases.
2825 class CMSPhaseAccounting: public StackObj {
2826 public:
2827 CMSPhaseAccounting(CMSCollector *collector,
2828 const char *phase,
2829 const GCId gc_id,
2830 bool print_cr = true);
2831 ~CMSPhaseAccounting();
2832
2833 private:
2834 CMSCollector *_collector;
2835 const char *_phase;
2836 elapsedTimer _wallclock;
2837 bool _print_cr;
2838 const GCId _gc_id;
2839
2840 public:
2841 // Not MT-safe; so do not pass around these StackObj's
2842 // where they may be accessed by other threads.
2843 jlong wallclock_millis() {
2844 assert(_wallclock.is_active(), "Wall clock should not stop");
2845 _wallclock.stop(); // to record time
2846 jlong ret = _wallclock.milliseconds();
2847 _wallclock.start(); // restart
2848 return ret;
2849 }
2850 };
2851
2852 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2853 const char *phase,
2854 const GCId gc_id,
2855 bool print_cr) :
2856 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2857
2858 if (PrintCMSStatistics != 0) {
2859 _collector->resetYields();
2860 }
2861 if (PrintGCDetails) {
2862 gclog_or_tty->gclog_stamp(_gc_id);
2863 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2864 _collector->cmsGen()->short_name(), _phase);
2865 }
2866 _collector->resetTimer();
2867 _wallclock.start();
2868 _collector->startTimer();
2869 }
2870
2871 CMSPhaseAccounting::~CMSPhaseAccounting() {
2872 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2873 _collector->stopTimer();
2874 _wallclock.stop();
2875 if (PrintGCDetails) {
2876 gclog_or_tty->gclog_stamp(_gc_id);
2877 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2878 _collector->cmsGen()->short_name(),
2879 _phase, _collector->timerValue(), _wallclock.seconds());
2880 if (_print_cr) {
2881 gclog_or_tty->cr();
2882 }
2883 if (PrintCMSStatistics != 0) {
2884 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2885 _collector->yields());
2886 }
2887 }
2888 }
2889
2890 // CMS work
2891
2892 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2893 class CMSParMarkTask : public AbstractGangTask {
2894 protected:
2895 CMSCollector* _collector;
2896 uint _n_workers;
2897 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2898 AbstractGangTask(name),
2899 _collector(collector),
2900 _n_workers(n_workers) {}
2901 // Work method in support of parallel rescan ... of young gen spaces
2902 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2903 ContiguousSpace* space,
2904 HeapWord** chunk_array, size_t chunk_top);
2905 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2906 };
2907
2908 // Parallel initial mark task
2909 class CMSParInitialMarkTask: public CMSParMarkTask {
2910 StrongRootsScope* _strong_roots_scope;
2911 public:
2912 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2913 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2914 _strong_roots_scope(strong_roots_scope) {}
2915 void work(uint worker_id);
2916 };
2917
2918 // Checkpoint the roots into this generation from outside
2919 // this generation. [Note this initial checkpoint need only
2920 // be approximate -- we'll do a catch up phase subsequently.]
2921 void CMSCollector::checkpointRootsInitial() {
2922 assert(_collectorState == InitialMarking, "Wrong collector state");
2923 check_correct_thread_executing();
2924 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2925
2926 save_heap_summary();
2927 report_heap_summary(GCWhen::BeforeGC);
2928
2929 ReferenceProcessor* rp = ref_processor();
2930 assert(_restart_addr == NULL, "Control point invariant");
2931 {
2932 // acquire locks for subsequent manipulations
2933 MutexLockerEx x(bitMapLock(),
2934 Mutex::_no_safepoint_check_flag);
2935 checkpointRootsInitialWork();
2936 // enable ("weak") refs discovery
2937 rp->enable_discovery();
2938 _collectorState = Marking;
2939 }
2940 }
2941
2942 void CMSCollector::checkpointRootsInitialWork() {
2943 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2944 assert(_collectorState == InitialMarking, "just checking");
2945
2946 // Already have locks.
2947 assert_lock_strong(bitMapLock());
2948 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2949
2950 // Setup the verification and class unloading state for this
2951 // CMS collection cycle.
2952 setup_cms_unloading_and_verification_state();
2953
2954 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2955 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
2956
2957 // Reset all the PLAB chunk arrays if necessary.
2958 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2959 reset_survivor_plab_arrays();
2960 }
2961
2962 ResourceMark rm;
2963 HandleMark hm;
2964
2965 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2966 GenCollectedHeap* gch = GenCollectedHeap::heap();
2967
2968 verify_work_stacks_empty();
2969 verify_overflow_empty();
2970
2971 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2972 // Update the saved marks which may affect the root scans.
2973 gch->save_marks();
2974
2975 // weak reference processing has not started yet.
2976 ref_processor()->set_enqueuing_is_done(false);
2977
2978 // Need to remember all newly created CLDs,
2979 // so that we can guarantee that the remark finds them.
2980 ClassLoaderDataGraph::remember_new_clds(true);
2981
2982 // Whenever a CLD is found, it will be claimed before proceeding to mark
2983 // the klasses. The claimed marks need to be cleared before marking starts.
2984 ClassLoaderDataGraph::clear_claimed_marks();
2985
2986 if (CMSPrintEdenSurvivorChunks) {
2987 print_eden_and_survivor_chunk_arrays();
2988 }
2989
2990 {
2991 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2992 if (CMSParallelInitialMarkEnabled) {
2993 // The parallel version.
2994 WorkGang* workers = gch->workers();
2995 assert(workers != NULL, "Need parallel worker threads.");
2996 uint n_workers = workers->active_workers();
2997
2998 StrongRootsScope srs(n_workers);
2999
3000 CMSParInitialMarkTask tsk(this, &srs, n_workers);
3001 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3002 if (n_workers > 1) {
3003 workers->run_task(&tsk);
3004 } else {
3005 tsk.work(0);
3006 }
3007 } else {
3008 // The serial version.
3009 CLDToOopClosure cld_closure(¬Older, true);
3010 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3011
3012 StrongRootsScope srs(1);
3013
3014 gch->gen_process_roots(&srs,
3015 GenCollectedHeap::OldGen,
3016 true, // young gen as roots
3017 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3018 should_unload_classes(),
3019 ¬Older,
3020 NULL,
3021 &cld_closure);
3022 }
3023 }
3024
3025 // Clear mod-union table; it will be dirtied in the prologue of
3026 // CMS generation per each young generation collection.
3027
3028 assert(_modUnionTable.isAllClear(),
3029 "Was cleared in most recent final checkpoint phase"
3030 " or no bits are set in the gc_prologue before the start of the next "
3031 "subsequent marking phase.");
3032
3033 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3034
3035 // Save the end of the used_region of the constituent generations
3036 // to be used to limit the extent of sweep in each generation.
3037 save_sweep_limits();
3038 verify_overflow_empty();
3039 }
3040
3041 bool CMSCollector::markFromRoots() {
3042 // we might be tempted to assert that:
3043 // assert(!SafepointSynchronize::is_at_safepoint(),
3044 // "inconsistent argument?");
3045 // However that wouldn't be right, because it's possible that
3046 // a safepoint is indeed in progress as a young generation
3047 // stop-the-world GC happens even as we mark in this generation.
3048 assert(_collectorState == Marking, "inconsistent state?");
3049 check_correct_thread_executing();
3050 verify_overflow_empty();
3051
3052 // Weak ref discovery note: We may be discovering weak
3053 // refs in this generation concurrent (but interleaved) with
3054 // weak ref discovery by the young generation collector.
3055
3056 CMSTokenSyncWithLocks ts(true, bitMapLock());
3057 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3058 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3059 bool res = markFromRootsWork();
3060 if (res) {
3061 _collectorState = Precleaning;
3062 } else { // We failed and a foreground collection wants to take over
3063 assert(_foregroundGCIsActive, "internal state inconsistency");
3064 assert(_restart_addr == NULL, "foreground will restart from scratch");
3065 if (PrintGCDetails) {
3066 gclog_or_tty->print_cr("bailing out to foreground collection");
3067 }
3068 }
3069 verify_overflow_empty();
3070 return res;
3071 }
3072
3073 bool CMSCollector::markFromRootsWork() {
3074 // iterate over marked bits in bit map, doing a full scan and mark
3075 // from these roots using the following algorithm:
3076 // . if oop is to the right of the current scan pointer,
3077 // mark corresponding bit (we'll process it later)
3078 // . else (oop is to left of current scan pointer)
3079 // push oop on marking stack
3080 // . drain the marking stack
3081
3082 // Note that when we do a marking step we need to hold the
3083 // bit map lock -- recall that direct allocation (by mutators)
3084 // and promotion (by the young generation collector) is also
3085 // marking the bit map. [the so-called allocate live policy.]
3086 // Because the implementation of bit map marking is not
3087 // robust wrt simultaneous marking of bits in the same word,
3088 // we need to make sure that there is no such interference
3089 // between concurrent such updates.
3090
3091 // already have locks
3092 assert_lock_strong(bitMapLock());
3093
3094 verify_work_stacks_empty();
3095 verify_overflow_empty();
3096 bool result = false;
3097 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3098 result = do_marking_mt();
3099 } else {
3100 result = do_marking_st();
3101 }
3102 return result;
3103 }
3104
3105 // Forward decl
3106 class CMSConcMarkingTask;
3107
3108 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3109 CMSCollector* _collector;
3110 CMSConcMarkingTask* _task;
3111 public:
3112 virtual void yield();
3113
3114 // "n_threads" is the number of threads to be terminated.
3115 // "queue_set" is a set of work queues of other threads.
3116 // "collector" is the CMS collector associated with this task terminator.
3117 // "yield" indicates whether we need the gang as a whole to yield.
3118 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3119 ParallelTaskTerminator(n_threads, queue_set),
3120 _collector(collector) { }
3121
3122 void set_task(CMSConcMarkingTask* task) {
3123 _task = task;
3124 }
3125 };
3126
3127 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3128 CMSConcMarkingTask* _task;
3129 public:
3130 bool should_exit_termination();
3131 void set_task(CMSConcMarkingTask* task) {
3132 _task = task;
3133 }
3134 };
3135
3136 // MT Concurrent Marking Task
3137 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3138 CMSCollector* _collector;
3139 uint _n_workers; // requested/desired # workers
3140 bool _result;
3141 CompactibleFreeListSpace* _cms_space;
3142 char _pad_front[64]; // padding to ...
3143 HeapWord* _global_finger; // ... avoid sharing cache line
3144 char _pad_back[64];
3145 HeapWord* _restart_addr;
3146
3147 // Exposed here for yielding support
3148 Mutex* const _bit_map_lock;
3149
3150 // The per thread work queues, available here for stealing
3151 OopTaskQueueSet* _task_queues;
3152
3153 // Termination (and yielding) support
3154 CMSConcMarkingTerminator _term;
3155 CMSConcMarkingTerminatorTerminator _term_term;
3156
3157 public:
3158 CMSConcMarkingTask(CMSCollector* collector,
3159 CompactibleFreeListSpace* cms_space,
3160 YieldingFlexibleWorkGang* workers,
3161 OopTaskQueueSet* task_queues):
3162 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3163 _collector(collector),
3164 _cms_space(cms_space),
3165 _n_workers(0), _result(true),
3166 _task_queues(task_queues),
3167 _term(_n_workers, task_queues, _collector),
3168 _bit_map_lock(collector->bitMapLock())
3169 {
3170 _requested_size = _n_workers;
3171 _term.set_task(this);
3172 _term_term.set_task(this);
3173 _restart_addr = _global_finger = _cms_space->bottom();
3174 }
3175
3176
3177 OopTaskQueueSet* task_queues() { return _task_queues; }
3178
3179 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3180
3181 HeapWord** global_finger_addr() { return &_global_finger; }
3182
3183 CMSConcMarkingTerminator* terminator() { return &_term; }
3184
3185 virtual void set_for_termination(uint active_workers) {
3186 terminator()->reset_for_reuse(active_workers);
3187 }
3188
3189 void work(uint worker_id);
3190 bool should_yield() {
3191 return ConcurrentMarkSweepThread::should_yield()
3192 && !_collector->foregroundGCIsActive();
3193 }
3194
3195 virtual void coordinator_yield(); // stuff done by coordinator
3196 bool result() { return _result; }
3197
3198 void reset(HeapWord* ra) {
3199 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3200 _restart_addr = _global_finger = ra;
3201 _term.reset_for_reuse();
3202 }
3203
3204 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3205 OopTaskQueue* work_q);
3206
3207 private:
3208 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3209 void do_work_steal(int i);
3210 void bump_global_finger(HeapWord* f);
3211 };
3212
3213 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3214 assert(_task != NULL, "Error");
3215 return _task->yielding();
3216 // Note that we do not need the disjunct || _task->should_yield() above
3217 // because we want terminating threads to yield only if the task
3218 // is already in the midst of yielding, which happens only after at least one
3219 // thread has yielded.
3220 }
3221
3222 void CMSConcMarkingTerminator::yield() {
3223 if (_task->should_yield()) {
3224 _task->yield();
3225 } else {
3226 ParallelTaskTerminator::yield();
3227 }
3228 }
3229
3230 ////////////////////////////////////////////////////////////////
3231 // Concurrent Marking Algorithm Sketch
3232 ////////////////////////////////////////////////////////////////
3233 // Until all tasks exhausted (both spaces):
3234 // -- claim next available chunk
3235 // -- bump global finger via CAS
3236 // -- find first object that starts in this chunk
3237 // and start scanning bitmap from that position
3238 // -- scan marked objects for oops
3239 // -- CAS-mark target, and if successful:
3240 // . if target oop is above global finger (volatile read)
3241 // nothing to do
3242 // . if target oop is in chunk and above local finger
3243 // then nothing to do
3244 // . else push on work-queue
3245 // -- Deal with possible overflow issues:
3246 // . local work-queue overflow causes stuff to be pushed on
3247 // global (common) overflow queue
3248 // . always first empty local work queue
3249 // . then get a batch of oops from global work queue if any
3250 // . then do work stealing
3251 // -- When all tasks claimed (both spaces)
3252 // and local work queue empty,
3253 // then in a loop do:
3254 // . check global overflow stack; steal a batch of oops and trace
3255 // . try to steal from other threads oif GOS is empty
3256 // . if neither is available, offer termination
3257 // -- Terminate and return result
3258 //
3259 void CMSConcMarkingTask::work(uint worker_id) {
3260 elapsedTimer _timer;
3261 ResourceMark rm;
3262 HandleMark hm;
3263
3264 DEBUG_ONLY(_collector->verify_overflow_empty();)
3265
3266 // Before we begin work, our work queue should be empty
3267 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3268 // Scan the bitmap covering _cms_space, tracing through grey objects.
3269 _timer.start();
3270 do_scan_and_mark(worker_id, _cms_space);
3271 _timer.stop();
3272 if (PrintCMSStatistics != 0) {
3273 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3274 worker_id, _timer.seconds());
3275 // XXX: need xxx/xxx type of notation, two timers
3276 }
3277
3278 // ... do work stealing
3279 _timer.reset();
3280 _timer.start();
3281 do_work_steal(worker_id);
3282 _timer.stop();
3283 if (PrintCMSStatistics != 0) {
3284 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3285 worker_id, _timer.seconds());
3286 // XXX: need xxx/xxx type of notation, two timers
3287 }
3288 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3289 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3290 // Note that under the current task protocol, the
3291 // following assertion is true even of the spaces
3292 // expanded since the completion of the concurrent
3293 // marking. XXX This will likely change under a strict
3294 // ABORT semantics.
3295 // After perm removal the comparison was changed to
3296 // greater than or equal to from strictly greater than.
3297 // Before perm removal the highest address sweep would
3298 // have been at the end of perm gen but now is at the
3299 // end of the tenured gen.
3300 assert(_global_finger >= _cms_space->end(),
3301 "All tasks have been completed");
3302 DEBUG_ONLY(_collector->verify_overflow_empty();)
3303 }
3304
3305 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3306 HeapWord* read = _global_finger;
3307 HeapWord* cur = read;
3308 while (f > read) {
3309 cur = read;
3310 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3311 if (cur == read) {
3312 // our cas succeeded
3313 assert(_global_finger >= f, "protocol consistency");
3314 break;
3315 }
3316 }
3317 }
3318
3319 // This is really inefficient, and should be redone by
3320 // using (not yet available) block-read and -write interfaces to the
3321 // stack and the work_queue. XXX FIX ME !!!
3322 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3323 OopTaskQueue* work_q) {
3324 // Fast lock-free check
3325 if (ovflw_stk->length() == 0) {
3326 return false;
3327 }
3328 assert(work_q->size() == 0, "Shouldn't steal");
3329 MutexLockerEx ml(ovflw_stk->par_lock(),
3330 Mutex::_no_safepoint_check_flag);
3331 // Grab up to 1/4 the size of the work queue
3332 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3333 (size_t)ParGCDesiredObjsFromOverflowList);
3334 num = MIN2(num, ovflw_stk->length());
3335 for (int i = (int) num; i > 0; i--) {
3336 oop cur = ovflw_stk->pop();
3337 assert(cur != NULL, "Counted wrong?");
3338 work_q->push(cur);
3339 }
3340 return num > 0;
3341 }
3342
3343 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3344 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3345 int n_tasks = pst->n_tasks();
3346 // We allow that there may be no tasks to do here because
3347 // we are restarting after a stack overflow.
3348 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3349 uint nth_task = 0;
3350
3351 HeapWord* aligned_start = sp->bottom();
3352 if (sp->used_region().contains(_restart_addr)) {
3353 // Align down to a card boundary for the start of 0th task
3354 // for this space.
3355 aligned_start =
3356 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3357 CardTableModRefBS::card_size);
3358 }
3359
3360 size_t chunk_size = sp->marking_task_size();
3361 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3362 // Having claimed the nth task in this space,
3363 // compute the chunk that it corresponds to:
3364 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3365 aligned_start + (nth_task+1)*chunk_size);
3366 // Try and bump the global finger via a CAS;
3367 // note that we need to do the global finger bump
3368 // _before_ taking the intersection below, because
3369 // the task corresponding to that region will be
3370 // deemed done even if the used_region() expands
3371 // because of allocation -- as it almost certainly will
3372 // during start-up while the threads yield in the
3373 // closure below.
3374 HeapWord* finger = span.end();
3375 bump_global_finger(finger); // atomically
3376 // There are null tasks here corresponding to chunks
3377 // beyond the "top" address of the space.
3378 span = span.intersection(sp->used_region());
3379 if (!span.is_empty()) { // Non-null task
3380 HeapWord* prev_obj;
3381 assert(!span.contains(_restart_addr) || nth_task == 0,
3382 "Inconsistency");
3383 if (nth_task == 0) {
3384 // For the 0th task, we'll not need to compute a block_start.
3385 if (span.contains(_restart_addr)) {
3386 // In the case of a restart because of stack overflow,
3387 // we might additionally skip a chunk prefix.
3388 prev_obj = _restart_addr;
3389 } else {
3390 prev_obj = span.start();
3391 }
3392 } else {
3393 // We want to skip the first object because
3394 // the protocol is to scan any object in its entirety
3395 // that _starts_ in this span; a fortiori, any
3396 // object starting in an earlier span is scanned
3397 // as part of an earlier claimed task.
3398 // Below we use the "careful" version of block_start
3399 // so we do not try to navigate uninitialized objects.
3400 prev_obj = sp->block_start_careful(span.start());
3401 // Below we use a variant of block_size that uses the
3402 // Printezis bits to avoid waiting for allocated
3403 // objects to become initialized/parsable.
3404 while (prev_obj < span.start()) {
3405 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3406 if (sz > 0) {
3407 prev_obj += sz;
3408 } else {
3409 // In this case we may end up doing a bit of redundant
3410 // scanning, but that appears unavoidable, short of
3411 // locking the free list locks; see bug 6324141.
3412 break;
3413 }
3414 }
3415 }
3416 if (prev_obj < span.end()) {
3417 MemRegion my_span = MemRegion(prev_obj, span.end());
3418 // Do the marking work within a non-empty span --
3419 // the last argument to the constructor indicates whether the
3420 // iteration should be incremental with periodic yields.
3421 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3422 &_collector->_markBitMap,
3423 work_queue(i),
3424 &_collector->_markStack);
3425 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3426 } // else nothing to do for this task
3427 } // else nothing to do for this task
3428 }
3429 // We'd be tempted to assert here that since there are no
3430 // more tasks left to claim in this space, the global_finger
3431 // must exceed space->top() and a fortiori space->end(). However,
3432 // that would not quite be correct because the bumping of
3433 // global_finger occurs strictly after the claiming of a task,
3434 // so by the time we reach here the global finger may not yet
3435 // have been bumped up by the thread that claimed the last
3436 // task.
3437 pst->all_tasks_completed();
3438 }
3439
3440 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3441 private:
3442 CMSCollector* _collector;
3443 CMSConcMarkingTask* _task;
3444 MemRegion _span;
3445 CMSBitMap* _bit_map;
3446 CMSMarkStack* _overflow_stack;
3447 OopTaskQueue* _work_queue;
3448 protected:
3449 DO_OOP_WORK_DEFN
3450 public:
3451 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3452 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3453 MetadataAwareOopClosure(collector->ref_processor()),
3454 _collector(collector),
3455 _task(task),
3456 _span(collector->_span),
3457 _work_queue(work_queue),
3458 _bit_map(bit_map),
3459 _overflow_stack(overflow_stack)
3460 { }
3461 virtual void do_oop(oop* p);
3462 virtual void do_oop(narrowOop* p);
3463
3464 void trim_queue(size_t max);
3465 void handle_stack_overflow(HeapWord* lost);
3466 void do_yield_check() {
3467 if (_task->should_yield()) {
3468 _task->yield();
3469 }
3470 }
3471 };
3472
3473 // Grey object scanning during work stealing phase --
3474 // the salient assumption here is that any references
3475 // that are in these stolen objects being scanned must
3476 // already have been initialized (else they would not have
3477 // been published), so we do not need to check for
3478 // uninitialized objects before pushing here.
3479 void Par_ConcMarkingClosure::do_oop(oop obj) {
3480 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3481 HeapWord* addr = (HeapWord*)obj;
3482 // Check if oop points into the CMS generation
3483 // and is not marked
3484 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3485 // a white object ...
3486 // If we manage to "claim" the object, by being the
3487 // first thread to mark it, then we push it on our
3488 // marking stack
3489 if (_bit_map->par_mark(addr)) { // ... now grey
3490 // push on work queue (grey set)
3491 bool simulate_overflow = false;
3492 NOT_PRODUCT(
3493 if (CMSMarkStackOverflowALot &&
3494 _collector->simulate_overflow()) {
3495 // simulate a stack overflow
3496 simulate_overflow = true;
3497 }
3498 )
3499 if (simulate_overflow ||
3500 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3501 // stack overflow
3502 if (PrintCMSStatistics != 0) {
3503 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3504 SIZE_FORMAT, _overflow_stack->capacity());
3505 }
3506 // We cannot assert that the overflow stack is full because
3507 // it may have been emptied since.
3508 assert(simulate_overflow ||
3509 _work_queue->size() == _work_queue->max_elems(),
3510 "Else push should have succeeded");
3511 handle_stack_overflow(addr);
3512 }
3513 } // Else, some other thread got there first
3514 do_yield_check();
3515 }
3516 }
3517
3518 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3519 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3520
3521 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3522 while (_work_queue->size() > max) {
3523 oop new_oop;
3524 if (_work_queue->pop_local(new_oop)) {
3525 assert(new_oop->is_oop(), "Should be an oop");
3526 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3527 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3528 new_oop->oop_iterate(this); // do_oop() above
3529 do_yield_check();
3530 }
3531 }
3532 }
3533
3534 // Upon stack overflow, we discard (part of) the stack,
3535 // remembering the least address amongst those discarded
3536 // in CMSCollector's _restart_address.
3537 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3538 // We need to do this under a mutex to prevent other
3539 // workers from interfering with the work done below.
3540 MutexLockerEx ml(_overflow_stack->par_lock(),
3541 Mutex::_no_safepoint_check_flag);
3542 // Remember the least grey address discarded
3543 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3544 _collector->lower_restart_addr(ra);
3545 _overflow_stack->reset(); // discard stack contents
3546 _overflow_stack->expand(); // expand the stack if possible
3547 }
3548
3549
3550 void CMSConcMarkingTask::do_work_steal(int i) {
3551 OopTaskQueue* work_q = work_queue(i);
3552 oop obj_to_scan;
3553 CMSBitMap* bm = &(_collector->_markBitMap);
3554 CMSMarkStack* ovflw = &(_collector->_markStack);
3555 int* seed = _collector->hash_seed(i);
3556 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3557 while (true) {
3558 cl.trim_queue(0);
3559 assert(work_q->size() == 0, "Should have been emptied above");
3560 if (get_work_from_overflow_stack(ovflw, work_q)) {
3561 // Can't assert below because the work obtained from the
3562 // overflow stack may already have been stolen from us.
3563 // assert(work_q->size() > 0, "Work from overflow stack");
3564 continue;
3565 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3566 assert(obj_to_scan->is_oop(), "Should be an oop");
3567 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3568 obj_to_scan->oop_iterate(&cl);
3569 } else if (terminator()->offer_termination(&_term_term)) {
3570 assert(work_q->size() == 0, "Impossible!");
3571 break;
3572 } else if (yielding() || should_yield()) {
3573 yield();
3574 }
3575 }
3576 }
3577
3578 // This is run by the CMS (coordinator) thread.
3579 void CMSConcMarkingTask::coordinator_yield() {
3580 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3581 "CMS thread should hold CMS token");
3582 // First give up the locks, then yield, then re-lock
3583 // We should probably use a constructor/destructor idiom to
3584 // do this unlock/lock or modify the MutexUnlocker class to
3585 // serve our purpose. XXX
3586 assert_lock_strong(_bit_map_lock);
3587 _bit_map_lock->unlock();
3588 ConcurrentMarkSweepThread::desynchronize(true);
3589 _collector->stopTimer();
3590 if (PrintCMSStatistics != 0) {
3591 _collector->incrementYields();
3592 }
3593
3594 // It is possible for whichever thread initiated the yield request
3595 // not to get a chance to wake up and take the bitmap lock between
3596 // this thread releasing it and reacquiring it. So, while the
3597 // should_yield() flag is on, let's sleep for a bit to give the
3598 // other thread a chance to wake up. The limit imposed on the number
3599 // of iterations is defensive, to avoid any unforseen circumstances
3600 // putting us into an infinite loop. Since it's always been this
3601 // (coordinator_yield()) method that was observed to cause the
3602 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3603 // which is by default non-zero. For the other seven methods that
3604 // also perform the yield operation, as are using a different
3605 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3606 // can enable the sleeping for those methods too, if necessary.
3607 // See 6442774.
3608 //
3609 // We really need to reconsider the synchronization between the GC
3610 // thread and the yield-requesting threads in the future and we
3611 // should really use wait/notify, which is the recommended
3612 // way of doing this type of interaction. Additionally, we should
3613 // consolidate the eight methods that do the yield operation and they
3614 // are almost identical into one for better maintainability and
3615 // readability. See 6445193.
3616 //
3617 // Tony 2006.06.29
3618 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3619 ConcurrentMarkSweepThread::should_yield() &&
3620 !CMSCollector::foregroundGCIsActive(); ++i) {
3621 os::sleep(Thread::current(), 1, false);
3622 }
3623
3624 ConcurrentMarkSweepThread::synchronize(true);
3625 _bit_map_lock->lock_without_safepoint_check();
3626 _collector->startTimer();
3627 }
3628
3629 bool CMSCollector::do_marking_mt() {
3630 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3631 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3632 conc_workers()->active_workers(),
3633 Threads::number_of_non_daemon_threads());
3634 conc_workers()->set_active_workers(num_workers);
3635
3636 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3637
3638 CMSConcMarkingTask tsk(this,
3639 cms_space,
3640 conc_workers(),
3641 task_queues());
3642
3643 // Since the actual number of workers we get may be different
3644 // from the number we requested above, do we need to do anything different
3645 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3646 // class?? XXX
3647 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3648
3649 // Refs discovery is already non-atomic.
3650 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3651 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3652 conc_workers()->start_task(&tsk);
3653 while (tsk.yielded()) {
3654 tsk.coordinator_yield();
3655 conc_workers()->continue_task(&tsk);
3656 }
3657 // If the task was aborted, _restart_addr will be non-NULL
3658 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3659 while (_restart_addr != NULL) {
3660 // XXX For now we do not make use of ABORTED state and have not
3661 // yet implemented the right abort semantics (even in the original
3662 // single-threaded CMS case). That needs some more investigation
3663 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3664 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3665 // If _restart_addr is non-NULL, a marking stack overflow
3666 // occurred; we need to do a fresh marking iteration from the
3667 // indicated restart address.
3668 if (_foregroundGCIsActive) {
3669 // We may be running into repeated stack overflows, having
3670 // reached the limit of the stack size, while making very
3671 // slow forward progress. It may be best to bail out and
3672 // let the foreground collector do its job.
3673 // Clear _restart_addr, so that foreground GC
3674 // works from scratch. This avoids the headache of
3675 // a "rescan" which would otherwise be needed because
3676 // of the dirty mod union table & card table.
3677 _restart_addr = NULL;
3678 return false;
3679 }
3680 // Adjust the task to restart from _restart_addr
3681 tsk.reset(_restart_addr);
3682 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3683 _restart_addr);
3684 _restart_addr = NULL;
3685 // Get the workers going again
3686 conc_workers()->start_task(&tsk);
3687 while (tsk.yielded()) {
3688 tsk.coordinator_yield();
3689 conc_workers()->continue_task(&tsk);
3690 }
3691 }
3692 assert(tsk.completed(), "Inconsistency");
3693 assert(tsk.result() == true, "Inconsistency");
3694 return true;
3695 }
3696
3697 bool CMSCollector::do_marking_st() {
3698 ResourceMark rm;
3699 HandleMark hm;
3700
3701 // Temporarily make refs discovery single threaded (non-MT)
3702 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3703 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3704 &_markStack, CMSYield);
3705 // the last argument to iterate indicates whether the iteration
3706 // should be incremental with periodic yields.
3707 _markBitMap.iterate(&markFromRootsClosure);
3708 // If _restart_addr is non-NULL, a marking stack overflow
3709 // occurred; we need to do a fresh iteration from the
3710 // indicated restart address.
3711 while (_restart_addr != NULL) {
3712 if (_foregroundGCIsActive) {
3713 // We may be running into repeated stack overflows, having
3714 // reached the limit of the stack size, while making very
3715 // slow forward progress. It may be best to bail out and
3716 // let the foreground collector do its job.
3717 // Clear _restart_addr, so that foreground GC
3718 // works from scratch. This avoids the headache of
3719 // a "rescan" which would otherwise be needed because
3720 // of the dirty mod union table & card table.
3721 _restart_addr = NULL;
3722 return false; // indicating failure to complete marking
3723 }
3724 // Deal with stack overflow:
3725 // we restart marking from _restart_addr
3726 HeapWord* ra = _restart_addr;
3727 markFromRootsClosure.reset(ra);
3728 _restart_addr = NULL;
3729 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3730 }
3731 return true;
3732 }
3733
3734 void CMSCollector::preclean() {
3735 check_correct_thread_executing();
3736 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3737 verify_work_stacks_empty();
3738 verify_overflow_empty();
3739 _abort_preclean = false;
3740 if (CMSPrecleaningEnabled) {
3741 if (!CMSEdenChunksRecordAlways) {
3742 _eden_chunk_index = 0;
3743 }
3744 size_t used = get_eden_used();
3745 size_t capacity = get_eden_capacity();
3746 // Don't start sampling unless we will get sufficiently
3747 // many samples.
3748 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3749 * CMSScheduleRemarkEdenPenetration)) {
3750 _start_sampling = true;
3751 } else {
3752 _start_sampling = false;
3753 }
3754 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3755 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3756 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3757 }
3758 CMSTokenSync x(true); // is cms thread
3759 if (CMSPrecleaningEnabled) {
3760 sample_eden();
3761 _collectorState = AbortablePreclean;
3762 } else {
3763 _collectorState = FinalMarking;
3764 }
3765 verify_work_stacks_empty();
3766 verify_overflow_empty();
3767 }
3768
3769 // Try and schedule the remark such that young gen
3770 // occupancy is CMSScheduleRemarkEdenPenetration %.
3771 void CMSCollector::abortable_preclean() {
3772 check_correct_thread_executing();
3773 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3774 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3775
3776 // If Eden's current occupancy is below this threshold,
3777 // immediately schedule the remark; else preclean
3778 // past the next scavenge in an effort to
3779 // schedule the pause as described above. By choosing
3780 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3781 // we will never do an actual abortable preclean cycle.
3782 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3783 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3784 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3785 // We need more smarts in the abortable preclean
3786 // loop below to deal with cases where allocation
3787 // in young gen is very very slow, and our precleaning
3788 // is running a losing race against a horde of
3789 // mutators intent on flooding us with CMS updates
3790 // (dirty cards).
3791 // One, admittedly dumb, strategy is to give up
3792 // after a certain number of abortable precleaning loops
3793 // or after a certain maximum time. We want to make
3794 // this smarter in the next iteration.
3795 // XXX FIX ME!!! YSR
3796 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3797 while (!(should_abort_preclean() ||
3798 ConcurrentMarkSweepThread::should_terminate())) {
3799 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3800 cumworkdone += workdone;
3801 loops++;
3802 // Voluntarily terminate abortable preclean phase if we have
3803 // been at it for too long.
3804 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3805 loops >= CMSMaxAbortablePrecleanLoops) {
3806 if (PrintGCDetails) {
3807 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3808 }
3809 break;
3810 }
3811 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3812 if (PrintGCDetails) {
3813 gclog_or_tty->print(" CMS: abort preclean due to time ");
3814 }
3815 break;
3816 }
3817 // If we are doing little work each iteration, we should
3818 // take a short break.
3819 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3820 // Sleep for some time, waiting for work to accumulate
3821 stopTimer();
3822 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3823 startTimer();
3824 waited++;
3825 }
3826 }
3827 if (PrintCMSStatistics > 0) {
3828 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3829 loops, waited, cumworkdone);
3830 }
3831 }
3832 CMSTokenSync x(true); // is cms thread
3833 if (_collectorState != Idling) {
3834 assert(_collectorState == AbortablePreclean,
3835 "Spontaneous state transition?");
3836 _collectorState = FinalMarking;
3837 } // Else, a foreground collection completed this CMS cycle.
3838 return;
3839 }
3840
3841 // Respond to an Eden sampling opportunity
3842 void CMSCollector::sample_eden() {
3843 // Make sure a young gc cannot sneak in between our
3844 // reading and recording of a sample.
3845 assert(Thread::current()->is_ConcurrentGC_thread(),
3846 "Only the cms thread may collect Eden samples");
3847 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3848 "Should collect samples while holding CMS token");
3849 if (!_start_sampling) {
3850 return;
3851 }
3852 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3853 // is populated by the young generation.
3854 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3855 if (_eden_chunk_index < _eden_chunk_capacity) {
3856 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3857 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3858 "Unexpected state of Eden");
3859 // We'd like to check that what we just sampled is an oop-start address;
3860 // however, we cannot do that here since the object may not yet have been
3861 // initialized. So we'll instead do the check when we _use_ this sample
3862 // later.
3863 if (_eden_chunk_index == 0 ||
3864 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3865 _eden_chunk_array[_eden_chunk_index-1])
3866 >= CMSSamplingGrain)) {
3867 _eden_chunk_index++; // commit sample
3868 }
3869 }
3870 }
3871 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3872 size_t used = get_eden_used();
3873 size_t capacity = get_eden_capacity();
3874 assert(used <= capacity, "Unexpected state of Eden");
3875 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3876 _abort_preclean = true;
3877 }
3878 }
3879 }
3880
3881
3882 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3883 assert(_collectorState == Precleaning ||
3884 _collectorState == AbortablePreclean, "incorrect state");
3885 ResourceMark rm;
3886 HandleMark hm;
3887
3888 // Precleaning is currently not MT but the reference processor
3889 // may be set for MT. Disable it temporarily here.
3890 ReferenceProcessor* rp = ref_processor();
3891 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3892
3893 // Do one pass of scrubbing the discovered reference lists
3894 // to remove any reference objects with strongly-reachable
3895 // referents.
3896 if (clean_refs) {
3897 CMSPrecleanRefsYieldClosure yield_cl(this);
3898 assert(rp->span().equals(_span), "Spans should be equal");
3899 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3900 &_markStack, true /* preclean */);
3901 CMSDrainMarkingStackClosure complete_trace(this,
3902 _span, &_markBitMap, &_markStack,
3903 &keep_alive, true /* preclean */);
3904
3905 // We don't want this step to interfere with a young
3906 // collection because we don't want to take CPU
3907 // or memory bandwidth away from the young GC threads
3908 // (which may be as many as there are CPUs).
3909 // Note that we don't need to protect ourselves from
3910 // interference with mutators because they can't
3911 // manipulate the discovered reference lists nor affect
3912 // the computed reachability of the referents, the
3913 // only properties manipulated by the precleaning
3914 // of these reference lists.
3915 stopTimer();
3916 CMSTokenSyncWithLocks x(true /* is cms thread */,
3917 bitMapLock());
3918 startTimer();
3919 sample_eden();
3920
3921 // The following will yield to allow foreground
3922 // collection to proceed promptly. XXX YSR:
3923 // The code in this method may need further
3924 // tweaking for better performance and some restructuring
3925 // for cleaner interfaces.
3926 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3927 rp->preclean_discovered_references(
3928 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3929 gc_timer, _gc_tracer_cm->gc_id());
3930 }
3931
3932 if (clean_survivor) { // preclean the active survivor space(s)
3933 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3934 &_markBitMap, &_modUnionTable,
3935 &_markStack, true /* precleaning phase */);
3936 stopTimer();
3937 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3938 bitMapLock());
3939 startTimer();
3940 unsigned int before_count =
3941 GenCollectedHeap::heap()->total_collections();
3942 SurvivorSpacePrecleanClosure
3943 sss_cl(this, _span, &_markBitMap, &_markStack,
3944 &pam_cl, before_count, CMSYield);
3945 _young_gen->from()->object_iterate_careful(&sss_cl);
3946 _young_gen->to()->object_iterate_careful(&sss_cl);
3947 }
3948 MarkRefsIntoAndScanClosure
3949 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3950 &_markStack, this, CMSYield,
3951 true /* precleaning phase */);
3952 // CAUTION: The following closure has persistent state that may need to
3953 // be reset upon a decrease in the sequence of addresses it
3954 // processes.
3955 ScanMarkedObjectsAgainCarefullyClosure
3956 smoac_cl(this, _span,
3957 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3958
3959 // Preclean dirty cards in ModUnionTable and CardTable using
3960 // appropriate convergence criterion;
3961 // repeat CMSPrecleanIter times unless we find that
3962 // we are losing.
3963 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3964 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3965 "Bad convergence multiplier");
3966 assert(CMSPrecleanThreshold >= 100,
3967 "Unreasonably low CMSPrecleanThreshold");
3968
3969 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3970 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3971 numIter < CMSPrecleanIter;
3972 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3973 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3974 if (Verbose && PrintGCDetails) {
3975 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3976 }
3977 // Either there are very few dirty cards, so re-mark
3978 // pause will be small anyway, or our pre-cleaning isn't
3979 // that much faster than the rate at which cards are being
3980 // dirtied, so we might as well stop and re-mark since
3981 // precleaning won't improve our re-mark time by much.
3982 if (curNumCards <= CMSPrecleanThreshold ||
3983 (numIter > 0 &&
3984 (curNumCards * CMSPrecleanDenominator >
3985 lastNumCards * CMSPrecleanNumerator))) {
3986 numIter++;
3987 cumNumCards += curNumCards;
3988 break;
3989 }
3990 }
3991
3992 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
3993
3994 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3995 cumNumCards += curNumCards;
3996 if (PrintGCDetails && PrintCMSStatistics != 0) {
3997 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3998 curNumCards, cumNumCards, numIter);
3999 }
4000 return cumNumCards; // as a measure of useful work done
4001 }
4002
4003 // PRECLEANING NOTES:
4004 // Precleaning involves:
4005 // . reading the bits of the modUnionTable and clearing the set bits.
4006 // . For the cards corresponding to the set bits, we scan the
4007 // objects on those cards. This means we need the free_list_lock
4008 // so that we can safely iterate over the CMS space when scanning
4009 // for oops.
4010 // . When we scan the objects, we'll be both reading and setting
4011 // marks in the marking bit map, so we'll need the marking bit map.
4012 // . For protecting _collector_state transitions, we take the CGC_lock.
4013 // Note that any races in the reading of of card table entries by the
4014 // CMS thread on the one hand and the clearing of those entries by the
4015 // VM thread or the setting of those entries by the mutator threads on the
4016 // other are quite benign. However, for efficiency it makes sense to keep
4017 // the VM thread from racing with the CMS thread while the latter is
4018 // dirty card info to the modUnionTable. We therefore also use the
4019 // CGC_lock to protect the reading of the card table and the mod union
4020 // table by the CM thread.
4021 // . We run concurrently with mutator updates, so scanning
4022 // needs to be done carefully -- we should not try to scan
4023 // potentially uninitialized objects.
4024 //
4025 // Locking strategy: While holding the CGC_lock, we scan over and
4026 // reset a maximal dirty range of the mod union / card tables, then lock
4027 // the free_list_lock and bitmap lock to do a full marking, then
4028 // release these locks; and repeat the cycle. This allows for a
4029 // certain amount of fairness in the sharing of these locks between
4030 // the CMS collector on the one hand, and the VM thread and the
4031 // mutators on the other.
4032
4033 // NOTE: preclean_mod_union_table() and preclean_card_table()
4034 // further below are largely identical; if you need to modify
4035 // one of these methods, please check the other method too.
4036
4037 size_t CMSCollector::preclean_mod_union_table(
4038 ConcurrentMarkSweepGeneration* old_gen,
4039 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4040 verify_work_stacks_empty();
4041 verify_overflow_empty();
4042
4043 // strategy: starting with the first card, accumulate contiguous
4044 // ranges of dirty cards; clear these cards, then scan the region
4045 // covered by these cards.
4046
4047 // Since all of the MUT is committed ahead, we can just use
4048 // that, in case the generations expand while we are precleaning.
4049 // It might also be fine to just use the committed part of the
4050 // generation, but we might potentially miss cards when the
4051 // generation is rapidly expanding while we are in the midst
4052 // of precleaning.
4053 HeapWord* startAddr = old_gen->reserved().start();
4054 HeapWord* endAddr = old_gen->reserved().end();
4055
4056 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4057
4058 size_t numDirtyCards, cumNumDirtyCards;
4059 HeapWord *nextAddr, *lastAddr;
4060 for (cumNumDirtyCards = numDirtyCards = 0,
4061 nextAddr = lastAddr = startAddr;
4062 nextAddr < endAddr;
4063 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4064
4065 ResourceMark rm;
4066 HandleMark hm;
4067
4068 MemRegion dirtyRegion;
4069 {
4070 stopTimer();
4071 // Potential yield point
4072 CMSTokenSync ts(true);
4073 startTimer();
4074 sample_eden();
4075 // Get dirty region starting at nextOffset (inclusive),
4076 // simultaneously clearing it.
4077 dirtyRegion =
4078 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4079 assert(dirtyRegion.start() >= nextAddr,
4080 "returned region inconsistent?");
4081 }
4082 // Remember where the next search should begin.
4083 // The returned region (if non-empty) is a right open interval,
4084 // so lastOffset is obtained from the right end of that
4085 // interval.
4086 lastAddr = dirtyRegion.end();
4087 // Should do something more transparent and less hacky XXX
4088 numDirtyCards =
4089 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4090
4091 // We'll scan the cards in the dirty region (with periodic
4092 // yields for foreground GC as needed).
4093 if (!dirtyRegion.is_empty()) {
4094 assert(numDirtyCards > 0, "consistency check");
4095 HeapWord* stop_point = NULL;
4096 stopTimer();
4097 // Potential yield point
4098 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
4099 bitMapLock());
4100 startTimer();
4101 {
4102 verify_work_stacks_empty();
4103 verify_overflow_empty();
4104 sample_eden();
4105 stop_point =
4106 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4107 }
4108 if (stop_point != NULL) {
4109 // The careful iteration stopped early either because it found an
4110 // uninitialized object, or because we were in the midst of an
4111 // "abortable preclean", which should now be aborted. Redirty
4112 // the bits corresponding to the partially-scanned or unscanned
4113 // cards. We'll either restart at the next block boundary or
4114 // abort the preclean.
4115 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4116 "Should only be AbortablePreclean.");
4117 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4118 if (should_abort_preclean()) {
4119 break; // out of preclean loop
4120 } else {
4121 // Compute the next address at which preclean should pick up;
4122 // might need bitMapLock in order to read P-bits.
4123 lastAddr = next_card_start_after_block(stop_point);
4124 }
4125 }
4126 } else {
4127 assert(lastAddr == endAddr, "consistency check");
4128 assert(numDirtyCards == 0, "consistency check");
4129 break;
4130 }
4131 }
4132 verify_work_stacks_empty();
4133 verify_overflow_empty();
4134 return cumNumDirtyCards;
4135 }
4136
4137 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4138 // below are largely identical; if you need to modify
4139 // one of these methods, please check the other method too.
4140
4141 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4142 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4143 // strategy: it's similar to precleamModUnionTable above, in that
4144 // we accumulate contiguous ranges of dirty cards, mark these cards
4145 // precleaned, then scan the region covered by these cards.
4146 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4147 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4148
4149 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4150
4151 size_t numDirtyCards, cumNumDirtyCards;
4152 HeapWord *lastAddr, *nextAddr;
4153
4154 for (cumNumDirtyCards = numDirtyCards = 0,
4155 nextAddr = lastAddr = startAddr;
4156 nextAddr < endAddr;
4157 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4158
4159 ResourceMark rm;
4160 HandleMark hm;
4161
4162 MemRegion dirtyRegion;
4163 {
4164 // See comments in "Precleaning notes" above on why we
4165 // do this locking. XXX Could the locking overheads be
4166 // too high when dirty cards are sparse? [I don't think so.]
4167 stopTimer();
4168 CMSTokenSync x(true); // is cms thread
4169 startTimer();
4170 sample_eden();
4171 // Get and clear dirty region from card table
4172 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4173 MemRegion(nextAddr, endAddr),
4174 true,
4175 CardTableModRefBS::precleaned_card_val());
4176
4177 assert(dirtyRegion.start() >= nextAddr,
4178 "returned region inconsistent?");
4179 }
4180 lastAddr = dirtyRegion.end();
4181 numDirtyCards =
4182 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4183
4184 if (!dirtyRegion.is_empty()) {
4185 stopTimer();
4186 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4187 startTimer();
4188 sample_eden();
4189 verify_work_stacks_empty();
4190 verify_overflow_empty();
4191 HeapWord* stop_point =
4192 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4193 if (stop_point != NULL) {
4194 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4195 "Should only be AbortablePreclean.");
4196 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4197 if (should_abort_preclean()) {
4198 break; // out of preclean loop
4199 } else {
4200 // Compute the next address at which preclean should pick up.
4201 lastAddr = next_card_start_after_block(stop_point);
4202 }
4203 }
4204 } else {
4205 break;
4206 }
4207 }
4208 verify_work_stacks_empty();
4209 verify_overflow_empty();
4210 return cumNumDirtyCards;
4211 }
4212
4213 class PrecleanKlassClosure : public KlassClosure {
4214 KlassToOopClosure _cm_klass_closure;
4215 public:
4216 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4217 void do_klass(Klass* k) {
4218 if (k->has_accumulated_modified_oops()) {
4219 k->clear_accumulated_modified_oops();
4220
4221 _cm_klass_closure.do_klass(k);
4222 }
4223 }
4224 };
4225
4226 // The freelist lock is needed to prevent asserts, is it really needed?
4227 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4228
4229 cl->set_freelistLock(freelistLock);
4230
4231 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4232
4233 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4234 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4235 PrecleanKlassClosure preclean_klass_closure(cl);
4236 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4237
4238 verify_work_stacks_empty();
4239 verify_overflow_empty();
4240 }
4241
4242 void CMSCollector::checkpointRootsFinal() {
4243 assert(_collectorState == FinalMarking, "incorrect state transition?");
4244 check_correct_thread_executing();
4245 // world is stopped at this checkpoint
4246 assert(SafepointSynchronize::is_at_safepoint(),
4247 "world should be stopped");
4248 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4249
4250 verify_work_stacks_empty();
4251 verify_overflow_empty();
4252
4253 if (PrintGCDetails) {
4254 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]",
4255 _young_gen->used() / K,
4256 _young_gen->capacity() / K);
4257 }
4258 {
4259 if (CMSScavengeBeforeRemark) {
4260 GenCollectedHeap* gch = GenCollectedHeap::heap();
4261 // Temporarily set flag to false, GCH->do_collection will
4262 // expect it to be false and set to true
4263 FlagSetting fl(gch->_is_gc_active, false);
4264 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4265 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4266 gch->do_collection(true, // full (i.e. force, see below)
4267 false, // !clear_all_soft_refs
4268 0, // size
4269 false, // is_tlab
4270 GenCollectedHeap::YoungGen // type
4271 );
4272 }
4273 FreelistLocker x(this);
4274 MutexLockerEx y(bitMapLock(),
4275 Mutex::_no_safepoint_check_flag);
4276 checkpointRootsFinalWork();
4277 }
4278 verify_work_stacks_empty();
4279 verify_overflow_empty();
4280 }
4281
4282 void CMSCollector::checkpointRootsFinalWork() {
4283 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4284
4285 assert(haveFreelistLocks(), "must have free list locks");
4286 assert_lock_strong(bitMapLock());
4287
4288 ResourceMark rm;
4289 HandleMark hm;
4290
4291 GenCollectedHeap* gch = GenCollectedHeap::heap();
4292
4293 if (should_unload_classes()) {
4294 CodeCache::gc_prologue();
4295 }
4296 assert(haveFreelistLocks(), "must have free list locks");
4297 assert_lock_strong(bitMapLock());
4298
4299 // We might assume that we need not fill TLAB's when
4300 // CMSScavengeBeforeRemark is set, because we may have just done
4301 // a scavenge which would have filled all TLAB's -- and besides
4302 // Eden would be empty. This however may not always be the case --
4303 // for instance although we asked for a scavenge, it may not have
4304 // happened because of a JNI critical section. We probably need
4305 // a policy for deciding whether we can in that case wait until
4306 // the critical section releases and then do the remark following
4307 // the scavenge, and skip it here. In the absence of that policy,
4308 // or of an indication of whether the scavenge did indeed occur,
4309 // we cannot rely on TLAB's having been filled and must do
4310 // so here just in case a scavenge did not happen.
4311 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4312 // Update the saved marks which may affect the root scans.
4313 gch->save_marks();
4314
4315 if (CMSPrintEdenSurvivorChunks) {
4316 print_eden_and_survivor_chunk_arrays();
4317 }
4318
4319 {
4320 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4321
4322 // Note on the role of the mod union table:
4323 // Since the marker in "markFromRoots" marks concurrently with
4324 // mutators, it is possible for some reachable objects not to have been
4325 // scanned. For instance, an only reference to an object A was
4326 // placed in object B after the marker scanned B. Unless B is rescanned,
4327 // A would be collected. Such updates to references in marked objects
4328 // are detected via the mod union table which is the set of all cards
4329 // dirtied since the first checkpoint in this GC cycle and prior to
4330 // the most recent young generation GC, minus those cleaned up by the
4331 // concurrent precleaning.
4332 if (CMSParallelRemarkEnabled) {
4333 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4334 do_remark_parallel();
4335 } else {
4336 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4337 _gc_timer_cm, _gc_tracer_cm->gc_id());
4338 do_remark_non_parallel();
4339 }
4340 }
4341 verify_work_stacks_empty();
4342 verify_overflow_empty();
4343
4344 {
4345 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4346 refProcessingWork();
4347 }
4348 verify_work_stacks_empty();
4349 verify_overflow_empty();
4350
4351 if (should_unload_classes()) {
4352 CodeCache::gc_epilogue();
4353 }
4354 JvmtiExport::gc_epilogue();
4355
4356 // If we encountered any (marking stack / work queue) overflow
4357 // events during the current CMS cycle, take appropriate
4358 // remedial measures, where possible, so as to try and avoid
4359 // recurrence of that condition.
4360 assert(_markStack.isEmpty(), "No grey objects");
4361 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4362 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4363 if (ser_ovflw > 0) {
4364 if (PrintCMSStatistics != 0) {
4365 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4366 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT
4367 ", kac_preclean=" SIZE_FORMAT ")",
4368 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4369 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4370 }
4371 _markStack.expand();
4372 _ser_pmc_remark_ovflw = 0;
4373 _ser_pmc_preclean_ovflw = 0;
4374 _ser_kac_preclean_ovflw = 0;
4375 _ser_kac_ovflw = 0;
4376 }
4377 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4378 if (PrintCMSStatistics != 0) {
4379 gclog_or_tty->print_cr("Work queue overflow (benign) "
4380 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4381 _par_pmc_remark_ovflw, _par_kac_ovflw);
4382 }
4383 _par_pmc_remark_ovflw = 0;
4384 _par_kac_ovflw = 0;
4385 }
4386 if (PrintCMSStatistics != 0) {
4387 if (_markStack._hit_limit > 0) {
4388 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4389 _markStack._hit_limit);
4390 }
4391 if (_markStack._failed_double > 0) {
4392 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT "),"
4393 " current capacity " SIZE_FORMAT,
4394 _markStack._failed_double,
4395 _markStack.capacity());
4396 }
4397 }
4398 _markStack._hit_limit = 0;
4399 _markStack._failed_double = 0;
4400
4401 if ((VerifyAfterGC || VerifyDuringGC) &&
4402 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4403 verify_after_remark();
4404 }
4405
4406 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4407
4408 // Change under the freelistLocks.
4409 _collectorState = Sweeping;
4410 // Call isAllClear() under bitMapLock
4411 assert(_modUnionTable.isAllClear(),
4412 "Should be clear by end of the final marking");
4413 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4414 "Should be clear by end of the final marking");
4415 }
4416
4417 void CMSParInitialMarkTask::work(uint worker_id) {
4418 elapsedTimer _timer;
4419 ResourceMark rm;
4420 HandleMark hm;
4421
4422 // ---------- scan from roots --------------
4423 _timer.start();
4424 GenCollectedHeap* gch = GenCollectedHeap::heap();
4425 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4426
4427 // ---------- young gen roots --------------
4428 {
4429 work_on_young_gen_roots(worker_id, &par_mri_cl);
4430 _timer.stop();
4431 if (PrintCMSStatistics != 0) {
4432 gclog_or_tty->print_cr(
4433 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4434 worker_id, _timer.seconds());
4435 }
4436 }
4437
4438 // ---------- remaining roots --------------
4439 _timer.reset();
4440 _timer.start();
4441
4442 CLDToOopClosure cld_closure(&par_mri_cl, true);
4443
4444 gch->gen_process_roots(_strong_roots_scope,
4445 GenCollectedHeap::OldGen,
4446 false, // yg was scanned above
4447 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4448 _collector->should_unload_classes(),
4449 &par_mri_cl,
4450 NULL,
4451 &cld_closure);
4452 assert(_collector->should_unload_classes()
4453 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4454 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4455 _timer.stop();
4456 if (PrintCMSStatistics != 0) {
4457 gclog_or_tty->print_cr(
4458 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4459 worker_id, _timer.seconds());
4460 }
4461 }
4462
4463 // Parallel remark task
4464 class CMSParRemarkTask: public CMSParMarkTask {
4465 CompactibleFreeListSpace* _cms_space;
4466
4467 // The per-thread work queues, available here for stealing.
4468 OopTaskQueueSet* _task_queues;
4469 ParallelTaskTerminator _term;
4470 StrongRootsScope* _strong_roots_scope;
4471
4472 public:
4473 // A value of 0 passed to n_workers will cause the number of
4474 // workers to be taken from the active workers in the work gang.
4475 CMSParRemarkTask(CMSCollector* collector,
4476 CompactibleFreeListSpace* cms_space,
4477 uint n_workers, WorkGang* workers,
4478 OopTaskQueueSet* task_queues,
4479 StrongRootsScope* strong_roots_scope):
4480 CMSParMarkTask("Rescan roots and grey objects in parallel",
4481 collector, n_workers),
4482 _cms_space(cms_space),
4483 _task_queues(task_queues),
4484 _term(n_workers, task_queues),
4485 _strong_roots_scope(strong_roots_scope) { }
4486
4487 OopTaskQueueSet* task_queues() { return _task_queues; }
4488
4489 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4490
4491 ParallelTaskTerminator* terminator() { return &_term; }
4492 uint n_workers() { return _n_workers; }
4493
4494 void work(uint worker_id);
4495
4496 private:
4497 // ... of dirty cards in old space
4498 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4499 Par_MarkRefsIntoAndScanClosure* cl);
4500
4501 // ... work stealing for the above
4502 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4503 };
4504
4505 class RemarkKlassClosure : public KlassClosure {
4506 KlassToOopClosure _cm_klass_closure;
4507 public:
4508 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4509 void do_klass(Klass* k) {
4510 // Check if we have modified any oops in the Klass during the concurrent marking.
4511 if (k->has_accumulated_modified_oops()) {
4512 k->clear_accumulated_modified_oops();
4513
4514 // We could have transfered the current modified marks to the accumulated marks,
4515 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4516 } else if (k->has_modified_oops()) {
4517 // Don't clear anything, this info is needed by the next young collection.
4518 } else {
4519 // No modified oops in the Klass.
4520 return;
4521 }
4522
4523 // The klass has modified fields, need to scan the klass.
4524 _cm_klass_closure.do_klass(k);
4525 }
4526 };
4527
4528 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4529 ParNewGeneration* young_gen = _collector->_young_gen;
4530 ContiguousSpace* eden_space = young_gen->eden();
4531 ContiguousSpace* from_space = young_gen->from();
4532 ContiguousSpace* to_space = young_gen->to();
4533
4534 HeapWord** eca = _collector->_eden_chunk_array;
4535 size_t ect = _collector->_eden_chunk_index;
4536 HeapWord** sca = _collector->_survivor_chunk_array;
4537 size_t sct = _collector->_survivor_chunk_index;
4538
4539 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4540 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4541
4542 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4543 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4544 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4545 }
4546
4547 // work_queue(i) is passed to the closure
4548 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4549 // also is passed to do_dirty_card_rescan_tasks() and to
4550 // do_work_steal() to select the i-th task_queue.
4551
4552 void CMSParRemarkTask::work(uint worker_id) {
4553 elapsedTimer _timer;
4554 ResourceMark rm;
4555 HandleMark hm;
4556
4557 // ---------- rescan from roots --------------
4558 _timer.start();
4559 GenCollectedHeap* gch = GenCollectedHeap::heap();
4560 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4561 _collector->_span, _collector->ref_processor(),
4562 &(_collector->_markBitMap),
4563 work_queue(worker_id));
4564
4565 // Rescan young gen roots first since these are likely
4566 // coarsely partitioned and may, on that account, constitute
4567 // the critical path; thus, it's best to start off that
4568 // work first.
4569 // ---------- young gen roots --------------
4570 {
4571 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4572 _timer.stop();
4573 if (PrintCMSStatistics != 0) {
4574 gclog_or_tty->print_cr(
4575 "Finished young gen rescan work in %dth thread: %3.3f sec",
4576 worker_id, _timer.seconds());
4577 }
4578 }
4579
4580 // ---------- remaining roots --------------
4581 _timer.reset();
4582 _timer.start();
4583 gch->gen_process_roots(_strong_roots_scope,
4584 GenCollectedHeap::OldGen,
4585 false, // yg was scanned above
4586 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4587 _collector->should_unload_classes(),
4588 &par_mrias_cl,
4589 NULL,
4590 NULL); // The dirty klasses will be handled below
4591
4592 assert(_collector->should_unload_classes()
4593 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4594 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4595 _timer.stop();
4596 if (PrintCMSStatistics != 0) {
4597 gclog_or_tty->print_cr(
4598 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4599 worker_id, _timer.seconds());
4600 }
4601
4602 // ---------- unhandled CLD scanning ----------
4603 if (worker_id == 0) { // Single threaded at the moment.
4604 _timer.reset();
4605 _timer.start();
4606
4607 // Scan all new class loader data objects and new dependencies that were
4608 // introduced during concurrent marking.
4609 ResourceMark rm;
4610 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4611 for (int i = 0; i < array->length(); i++) {
4612 par_mrias_cl.do_cld_nv(array->at(i));
4613 }
4614
4615 // We don't need to keep track of new CLDs anymore.
4616 ClassLoaderDataGraph::remember_new_clds(false);
4617
4618 _timer.stop();
4619 if (PrintCMSStatistics != 0) {
4620 gclog_or_tty->print_cr(
4621 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4622 worker_id, _timer.seconds());
4623 }
4624 }
4625
4626 // ---------- dirty klass scanning ----------
4627 if (worker_id == 0) { // Single threaded at the moment.
4628 _timer.reset();
4629 _timer.start();
4630
4631 // Scan all classes that was dirtied during the concurrent marking phase.
4632 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4633 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4634
4635 _timer.stop();
4636 if (PrintCMSStatistics != 0) {
4637 gclog_or_tty->print_cr(
4638 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4639 worker_id, _timer.seconds());
4640 }
4641 }
4642
4643 // We might have added oops to ClassLoaderData::_handles during the
4644 // concurrent marking phase. These oops point to newly allocated objects
4645 // that are guaranteed to be kept alive. Either by the direct allocation
4646 // code, or when the young collector processes the roots. Hence,
4647 // we don't have to revisit the _handles block during the remark phase.
4648
4649 // ---------- rescan dirty cards ------------
4650 _timer.reset();
4651 _timer.start();
4652
4653 // Do the rescan tasks for each of the two spaces
4654 // (cms_space) in turn.
4655 // "worker_id" is passed to select the task_queue for "worker_id"
4656 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4657 _timer.stop();
4658 if (PrintCMSStatistics != 0) {
4659 gclog_or_tty->print_cr(
4660 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4661 worker_id, _timer.seconds());
4662 }
4663
4664 // ---------- steal work from other threads ...
4665 // ---------- ... and drain overflow list.
4666 _timer.reset();
4667 _timer.start();
4668 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4669 _timer.stop();
4670 if (PrintCMSStatistics != 0) {
4671 gclog_or_tty->print_cr(
4672 "Finished work stealing in %dth thread: %3.3f sec",
4673 worker_id, _timer.seconds());
4674 }
4675 }
4676
4677 // Note that parameter "i" is not used.
4678 void
4679 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4680 OopsInGenClosure* cl, ContiguousSpace* space,
4681 HeapWord** chunk_array, size_t chunk_top) {
4682 // Until all tasks completed:
4683 // . claim an unclaimed task
4684 // . compute region boundaries corresponding to task claimed
4685 // using chunk_array
4686 // . par_oop_iterate(cl) over that region
4687
4688 ResourceMark rm;
4689 HandleMark hm;
4690
4691 SequentialSubTasksDone* pst = space->par_seq_tasks();
4692
4693 uint nth_task = 0;
4694 uint n_tasks = pst->n_tasks();
4695
4696 if (n_tasks > 0) {
4697 assert(pst->valid(), "Uninitialized use?");
4698 HeapWord *start, *end;
4699 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4700 // We claimed task # nth_task; compute its boundaries.
4701 if (chunk_top == 0) { // no samples were taken
4702 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4703 start = space->bottom();
4704 end = space->top();
4705 } else if (nth_task == 0) {
4706 start = space->bottom();
4707 end = chunk_array[nth_task];
4708 } else if (nth_task < (uint)chunk_top) {
4709 assert(nth_task >= 1, "Control point invariant");
4710 start = chunk_array[nth_task - 1];
4711 end = chunk_array[nth_task];
4712 } else {
4713 assert(nth_task == (uint)chunk_top, "Control point invariant");
4714 start = chunk_array[chunk_top - 1];
4715 end = space->top();
4716 }
4717 MemRegion mr(start, end);
4718 // Verify that mr is in space
4719 assert(mr.is_empty() || space->used_region().contains(mr),
4720 "Should be in space");
4721 // Verify that "start" is an object boundary
4722 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4723 "Should be an oop");
4724 space->par_oop_iterate(mr, cl);
4725 }
4726 pst->all_tasks_completed();
4727 }
4728 }
4729
4730 void
4731 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4732 CompactibleFreeListSpace* sp, int i,
4733 Par_MarkRefsIntoAndScanClosure* cl) {
4734 // Until all tasks completed:
4735 // . claim an unclaimed task
4736 // . compute region boundaries corresponding to task claimed
4737 // . transfer dirty bits ct->mut for that region
4738 // . apply rescanclosure to dirty mut bits for that region
4739
4740 ResourceMark rm;
4741 HandleMark hm;
4742
4743 OopTaskQueue* work_q = work_queue(i);
4744 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4745 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4746 // CAUTION: This closure has state that persists across calls to
4747 // the work method dirty_range_iterate_clear() in that it has
4748 // embedded in it a (subtype of) UpwardsObjectClosure. The
4749 // use of that state in the embedded UpwardsObjectClosure instance
4750 // assumes that the cards are always iterated (even if in parallel
4751 // by several threads) in monotonically increasing order per each
4752 // thread. This is true of the implementation below which picks
4753 // card ranges (chunks) in monotonically increasing order globally
4754 // and, a-fortiori, in monotonically increasing order per thread
4755 // (the latter order being a subsequence of the former).
4756 // If the work code below is ever reorganized into a more chaotic
4757 // work-partitioning form than the current "sequential tasks"
4758 // paradigm, the use of that persistent state will have to be
4759 // revisited and modified appropriately. See also related
4760 // bug 4756801 work on which should examine this code to make
4761 // sure that the changes there do not run counter to the
4762 // assumptions made here and necessary for correctness and
4763 // efficiency. Note also that this code might yield inefficient
4764 // behavior in the case of very large objects that span one or
4765 // more work chunks. Such objects would potentially be scanned
4766 // several times redundantly. Work on 4756801 should try and
4767 // address that performance anomaly if at all possible. XXX
4768 MemRegion full_span = _collector->_span;
4769 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4770 MarkFromDirtyCardsClosure
4771 greyRescanClosure(_collector, full_span, // entire span of interest
4772 sp, bm, work_q, cl);
4773
4774 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4775 assert(pst->valid(), "Uninitialized use?");
4776 uint nth_task = 0;
4777 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4778 MemRegion span = sp->used_region();
4779 HeapWord* start_addr = span.start();
4780 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4781 alignment);
4782 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4783 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4784 start_addr, "Check alignment");
4785 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4786 chunk_size, "Check alignment");
4787
4788 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4789 // Having claimed the nth_task, compute corresponding mem-region,
4790 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4791 // The alignment restriction ensures that we do not need any
4792 // synchronization with other gang-workers while setting or
4793 // clearing bits in thus chunk of the MUT.
4794 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4795 start_addr + (nth_task+1)*chunk_size);
4796 // The last chunk's end might be way beyond end of the
4797 // used region. In that case pull back appropriately.
4798 if (this_span.end() > end_addr) {
4799 this_span.set_end(end_addr);
4800 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4801 }
4802 // Iterate over the dirty cards covering this chunk, marking them
4803 // precleaned, and setting the corresponding bits in the mod union
4804 // table. Since we have been careful to partition at Card and MUT-word
4805 // boundaries no synchronization is needed between parallel threads.
4806 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4807 &modUnionClosure);
4808
4809 // Having transferred these marks into the modUnionTable,
4810 // rescan the marked objects on the dirty cards in the modUnionTable.
4811 // Even if this is at a synchronous collection, the initial marking
4812 // may have been done during an asynchronous collection so there
4813 // may be dirty bits in the mod-union table.
4814 _collector->_modUnionTable.dirty_range_iterate_clear(
4815 this_span, &greyRescanClosure);
4816 _collector->_modUnionTable.verifyNoOneBitsInRange(
4817 this_span.start(),
4818 this_span.end());
4819 }
4820 pst->all_tasks_completed(); // declare that i am done
4821 }
4822
4823 // . see if we can share work_queues with ParNew? XXX
4824 void
4825 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4826 int* seed) {
4827 OopTaskQueue* work_q = work_queue(i);
4828 NOT_PRODUCT(int num_steals = 0;)
4829 oop obj_to_scan;
4830 CMSBitMap* bm = &(_collector->_markBitMap);
4831
4832 while (true) {
4833 // Completely finish any left over work from (an) earlier round(s)
4834 cl->trim_queue(0);
4835 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4836 (size_t)ParGCDesiredObjsFromOverflowList);
4837 // Now check if there's any work in the overflow list
4838 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4839 // only affects the number of attempts made to get work from the
4840 // overflow list and does not affect the number of workers. Just
4841 // pass ParallelGCThreads so this behavior is unchanged.
4842 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4843 work_q,
4844 ParallelGCThreads)) {
4845 // found something in global overflow list;
4846 // not yet ready to go stealing work from others.
4847 // We'd like to assert(work_q->size() != 0, ...)
4848 // because we just took work from the overflow list,
4849 // but of course we can't since all of that could have
4850 // been already stolen from us.
4851 // "He giveth and He taketh away."
4852 continue;
4853 }
4854 // Verify that we have no work before we resort to stealing
4855 assert(work_q->size() == 0, "Have work, shouldn't steal");
4856 // Try to steal from other queues that have work
4857 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4858 NOT_PRODUCT(num_steals++;)
4859 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4860 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4861 // Do scanning work
4862 obj_to_scan->oop_iterate(cl);
4863 // Loop around, finish this work, and try to steal some more
4864 } else if (terminator()->offer_termination()) {
4865 break; // nirvana from the infinite cycle
4866 }
4867 }
4868 NOT_PRODUCT(
4869 if (PrintCMSStatistics != 0) {
4870 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4871 }
4872 )
4873 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4874 "Else our work is not yet done");
4875 }
4876
4877 // Record object boundaries in _eden_chunk_array by sampling the eden
4878 // top in the slow-path eden object allocation code path and record
4879 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4880 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4881 // sampling in sample_eden() that activates during the part of the
4882 // preclean phase.
4883 void CMSCollector::sample_eden_chunk() {
4884 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4885 if (_eden_chunk_lock->try_lock()) {
4886 // Record a sample. This is the critical section. The contents
4887 // of the _eden_chunk_array have to be non-decreasing in the
4888 // address order.
4889 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4890 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4891 "Unexpected state of Eden");
4892 if (_eden_chunk_index == 0 ||
4893 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4894 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4895 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4896 _eden_chunk_index++; // commit sample
4897 }
4898 _eden_chunk_lock->unlock();
4899 }
4900 }
4901 }
4902
4903 // Return a thread-local PLAB recording array, as appropriate.
4904 void* CMSCollector::get_data_recorder(int thr_num) {
4905 if (_survivor_plab_array != NULL &&
4906 (CMSPLABRecordAlways ||
4907 (_collectorState > Marking && _collectorState < FinalMarking))) {
4908 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4909 ChunkArray* ca = &_survivor_plab_array[thr_num];
4910 ca->reset(); // clear it so that fresh data is recorded
4911 return (void*) ca;
4912 } else {
4913 return NULL;
4914 }
4915 }
4916
4917 // Reset all the thread-local PLAB recording arrays
4918 void CMSCollector::reset_survivor_plab_arrays() {
4919 for (uint i = 0; i < ParallelGCThreads; i++) {
4920 _survivor_plab_array[i].reset();
4921 }
4922 }
4923
4924 // Merge the per-thread plab arrays into the global survivor chunk
4925 // array which will provide the partitioning of the survivor space
4926 // for CMS initial scan and rescan.
4927 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4928 int no_of_gc_threads) {
4929 assert(_survivor_plab_array != NULL, "Error");
4930 assert(_survivor_chunk_array != NULL, "Error");
4931 assert(_collectorState == FinalMarking ||
4932 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4933 for (int j = 0; j < no_of_gc_threads; j++) {
4934 _cursor[j] = 0;
4935 }
4936 HeapWord* top = surv->top();
4937 size_t i;
4938 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4939 HeapWord* min_val = top; // Higher than any PLAB address
4940 uint min_tid = 0; // position of min_val this round
4941 for (int j = 0; j < no_of_gc_threads; j++) {
4942 ChunkArray* cur_sca = &_survivor_plab_array[j];
4943 if (_cursor[j] == cur_sca->end()) {
4944 continue;
4945 }
4946 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4947 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4948 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4949 if (cur_val < min_val) {
4950 min_tid = j;
4951 min_val = cur_val;
4952 } else {
4953 assert(cur_val < top, "All recorded addresses should be less");
4954 }
4955 }
4956 // At this point min_val and min_tid are respectively
4957 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4958 // and the thread (j) that witnesses that address.
4959 // We record this address in the _survivor_chunk_array[i]
4960 // and increment _cursor[min_tid] prior to the next round i.
4961 if (min_val == top) {
4962 break;
4963 }
4964 _survivor_chunk_array[i] = min_val;
4965 _cursor[min_tid]++;
4966 }
4967 // We are all done; record the size of the _survivor_chunk_array
4968 _survivor_chunk_index = i; // exclusive: [0, i)
4969 if (PrintCMSStatistics > 0) {
4970 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4971 }
4972 // Verify that we used up all the recorded entries
4973 #ifdef ASSERT
4974 size_t total = 0;
4975 for (int j = 0; j < no_of_gc_threads; j++) {
4976 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4977 total += _cursor[j];
4978 }
4979 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4980 // Check that the merged array is in sorted order
4981 if (total > 0) {
4982 for (size_t i = 0; i < total - 1; i++) {
4983 if (PrintCMSStatistics > 0) {
4984 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4985 i, p2i(_survivor_chunk_array[i]));
4986 }
4987 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4988 "Not sorted");
4989 }
4990 }
4991 #endif // ASSERT
4992 }
4993
4994 // Set up the space's par_seq_tasks structure for work claiming
4995 // for parallel initial scan and rescan of young gen.
4996 // See ParRescanTask where this is currently used.
4997 void
4998 CMSCollector::
4999 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5000 assert(n_threads > 0, "Unexpected n_threads argument");
5001
5002 // Eden space
5003 if (!_young_gen->eden()->is_empty()) {
5004 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5005 assert(!pst->valid(), "Clobbering existing data?");
5006 // Each valid entry in [0, _eden_chunk_index) represents a task.
5007 size_t n_tasks = _eden_chunk_index + 1;
5008 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5009 // Sets the condition for completion of the subtask (how many threads
5010 // need to finish in order to be done).
5011 pst->set_n_threads(n_threads);
5012 pst->set_n_tasks((int)n_tasks);
5013 }
5014
5015 // Merge the survivor plab arrays into _survivor_chunk_array
5016 if (_survivor_plab_array != NULL) {
5017 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5018 } else {
5019 assert(_survivor_chunk_index == 0, "Error");
5020 }
5021
5022 // To space
5023 {
5024 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5025 assert(!pst->valid(), "Clobbering existing data?");
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(1);
5030 assert(pst->valid(), "Error");
5031 }
5032
5033 // From space
5034 {
5035 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5036 assert(!pst->valid(), "Clobbering existing data?");
5037 size_t n_tasks = _survivor_chunk_index + 1;
5038 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5039 // Sets the condition for completion of the subtask (how many threads
5040 // need to finish in order to be done).
5041 pst->set_n_threads(n_threads);
5042 pst->set_n_tasks((int)n_tasks);
5043 assert(pst->valid(), "Error");
5044 }
5045 }
5046
5047 // Parallel version of remark
5048 void CMSCollector::do_remark_parallel() {
5049 GenCollectedHeap* gch = GenCollectedHeap::heap();
5050 WorkGang* workers = gch->workers();
5051 assert(workers != NULL, "Need parallel worker threads.");
5052 // Choose to use the number of GC workers most recently set
5053 // into "active_workers".
5054 uint n_workers = workers->active_workers();
5055
5056 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5057
5058 StrongRootsScope srs(n_workers);
5059
5060 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
5061
5062 // We won't be iterating over the cards in the card table updating
5063 // the younger_gen cards, so we shouldn't call the following else
5064 // the verification code as well as subsequent younger_refs_iterate
5065 // code would get confused. XXX
5066 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5067
5068 // The young gen rescan work will not be done as part of
5069 // process_roots (which currently doesn't know how to
5070 // parallelize such a scan), but rather will be broken up into
5071 // a set of parallel tasks (via the sampling that the [abortable]
5072 // preclean phase did of eden, plus the [two] tasks of
5073 // scanning the [two] survivor spaces. Further fine-grain
5074 // parallelization of the scanning of the survivor spaces
5075 // themselves, and of precleaning of the young gen itself
5076 // is deferred to the future.
5077 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5078
5079 // The dirty card rescan work is broken up into a "sequence"
5080 // of parallel tasks (per constituent space) that are dynamically
5081 // claimed by the parallel threads.
5082 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5083
5084 // It turns out that even when we're using 1 thread, doing the work in a
5085 // separate thread causes wide variance in run times. We can't help this
5086 // in the multi-threaded case, but we special-case n=1 here to get
5087 // repeatable measurements of the 1-thread overhead of the parallel code.
5088 if (n_workers > 1) {
5089 // Make refs discovery MT-safe, if it isn't already: it may not
5090 // necessarily be so, since it's possible that we are doing
5091 // ST marking.
5092 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5093 workers->run_task(&tsk);
5094 } else {
5095 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5096 tsk.work(0);
5097 }
5098
5099 // restore, single-threaded for now, any preserved marks
5100 // as a result of work_q overflow
5101 restore_preserved_marks_if_any();
5102 }
5103
5104 // Non-parallel version of remark
5105 void CMSCollector::do_remark_non_parallel() {
5106 ResourceMark rm;
5107 HandleMark hm;
5108 GenCollectedHeap* gch = GenCollectedHeap::heap();
5109 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5110
5111 MarkRefsIntoAndScanClosure
5112 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5113 &_markStack, this,
5114 false /* should_yield */, false /* not precleaning */);
5115 MarkFromDirtyCardsClosure
5116 markFromDirtyCardsClosure(this, _span,
5117 NULL, // space is set further below
5118 &_markBitMap, &_markStack, &mrias_cl);
5119 {
5120 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5121 // Iterate over the dirty cards, setting the corresponding bits in the
5122 // mod union table.
5123 {
5124 ModUnionClosure modUnionClosure(&_modUnionTable);
5125 _ct->ct_bs()->dirty_card_iterate(
5126 _cmsGen->used_region(),
5127 &modUnionClosure);
5128 }
5129 // Having transferred these marks into the modUnionTable, we just need
5130 // to rescan the marked objects on the dirty cards in the modUnionTable.
5131 // The initial marking may have been done during an asynchronous
5132 // collection so there may be dirty bits in the mod-union table.
5133 const int alignment =
5134 CardTableModRefBS::card_size * BitsPerWord;
5135 {
5136 // ... First handle dirty cards in CMS gen
5137 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5138 MemRegion ur = _cmsGen->used_region();
5139 HeapWord* lb = ur.start();
5140 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5141 MemRegion cms_span(lb, ub);
5142 _modUnionTable.dirty_range_iterate_clear(cms_span,
5143 &markFromDirtyCardsClosure);
5144 verify_work_stacks_empty();
5145 if (PrintCMSStatistics != 0) {
5146 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ",
5147 markFromDirtyCardsClosure.num_dirty_cards());
5148 }
5149 }
5150 }
5151 if (VerifyDuringGC &&
5152 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5153 HandleMark hm; // Discard invalid handles created during verification
5154 Universe::verify();
5155 }
5156 {
5157 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5158
5159 verify_work_stacks_empty();
5160
5161 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5162 StrongRootsScope srs(1);
5163
5164 gch->gen_process_roots(&srs,
5165 GenCollectedHeap::OldGen,
5166 true, // young gen as roots
5167 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5168 should_unload_classes(),
5169 &mrias_cl,
5170 NULL,
5171 NULL); // The dirty klasses will be handled below
5172
5173 assert(should_unload_classes()
5174 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5175 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5176 }
5177
5178 {
5179 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5180
5181 verify_work_stacks_empty();
5182
5183 // Scan all class loader data objects that might have been introduced
5184 // during concurrent marking.
5185 ResourceMark rm;
5186 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5187 for (int i = 0; i < array->length(); i++) {
5188 mrias_cl.do_cld_nv(array->at(i));
5189 }
5190
5191 // We don't need to keep track of new CLDs anymore.
5192 ClassLoaderDataGraph::remember_new_clds(false);
5193
5194 verify_work_stacks_empty();
5195 }
5196
5197 {
5198 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5199
5200 verify_work_stacks_empty();
5201
5202 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5203 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5204
5205 verify_work_stacks_empty();
5206 }
5207
5208 // We might have added oops to ClassLoaderData::_handles during the
5209 // concurrent marking phase. These oops point to newly allocated objects
5210 // that are guaranteed to be kept alive. Either by the direct allocation
5211 // code, or when the young collector processes the roots. Hence,
5212 // we don't have to revisit the _handles block during the remark phase.
5213
5214 verify_work_stacks_empty();
5215 // Restore evacuated mark words, if any, used for overflow list links
5216 if (!CMSOverflowEarlyRestoration) {
5217 restore_preserved_marks_if_any();
5218 }
5219 verify_overflow_empty();
5220 }
5221
5222 ////////////////////////////////////////////////////////
5223 // Parallel Reference Processing Task Proxy Class
5224 ////////////////////////////////////////////////////////
5225 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5226 OopTaskQueueSet* _queues;
5227 ParallelTaskTerminator _terminator;
5228 public:
5229 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5230 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5231 ParallelTaskTerminator* terminator() { return &_terminator; }
5232 OopTaskQueueSet* queues() { return _queues; }
5233 };
5234
5235 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5236 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5237 CMSCollector* _collector;
5238 CMSBitMap* _mark_bit_map;
5239 const MemRegion _span;
5240 ProcessTask& _task;
5241
5242 public:
5243 CMSRefProcTaskProxy(ProcessTask& task,
5244 CMSCollector* collector,
5245 const MemRegion& span,
5246 CMSBitMap* mark_bit_map,
5247 AbstractWorkGang* workers,
5248 OopTaskQueueSet* task_queues):
5249 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5250 task_queues,
5251 workers->active_workers()),
5252 _task(task),
5253 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5254 {
5255 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5256 "Inconsistency in _span");
5257 }
5258
5259 OopTaskQueueSet* task_queues() { return queues(); }
5260
5261 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5262
5263 void do_work_steal(int i,
5264 CMSParDrainMarkingStackClosure* drain,
5265 CMSParKeepAliveClosure* keep_alive,
5266 int* seed);
5267
5268 virtual void work(uint worker_id);
5269 };
5270
5271 void CMSRefProcTaskProxy::work(uint worker_id) {
5272 ResourceMark rm;
5273 HandleMark hm;
5274 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5275 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5276 _mark_bit_map,
5277 work_queue(worker_id));
5278 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5279 _mark_bit_map,
5280 work_queue(worker_id));
5281 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5282 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5283 if (_task.marks_oops_alive()) {
5284 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5285 _collector->hash_seed(worker_id));
5286 }
5287 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5288 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5289 }
5290
5291 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5292 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5293 EnqueueTask& _task;
5294
5295 public:
5296 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5297 : AbstractGangTask("Enqueue reference objects in parallel"),
5298 _task(task)
5299 { }
5300
5301 virtual void work(uint worker_id)
5302 {
5303 _task.work(worker_id);
5304 }
5305 };
5306
5307 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5308 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5309 _span(span),
5310 _bit_map(bit_map),
5311 _work_queue(work_queue),
5312 _mark_and_push(collector, span, bit_map, work_queue),
5313 _low_water_mark(MIN2((work_queue->max_elems()/4),
5314 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5315 { }
5316
5317 // . see if we can share work_queues with ParNew? XXX
5318 void CMSRefProcTaskProxy::do_work_steal(int i,
5319 CMSParDrainMarkingStackClosure* drain,
5320 CMSParKeepAliveClosure* keep_alive,
5321 int* seed) {
5322 OopTaskQueue* work_q = work_queue(i);
5323 NOT_PRODUCT(int num_steals = 0;)
5324 oop obj_to_scan;
5325
5326 while (true) {
5327 // Completely finish any left over work from (an) earlier round(s)
5328 drain->trim_queue(0);
5329 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5330 (size_t)ParGCDesiredObjsFromOverflowList);
5331 // Now check if there's any work in the overflow list
5332 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5333 // only affects the number of attempts made to get work from the
5334 // overflow list and does not affect the number of workers. Just
5335 // pass ParallelGCThreads so this behavior is unchanged.
5336 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5337 work_q,
5338 ParallelGCThreads)) {
5339 // Found something in global overflow list;
5340 // not yet ready to go stealing work from others.
5341 // We'd like to assert(work_q->size() != 0, ...)
5342 // because we just took work from the overflow list,
5343 // but of course we can't, since all of that might have
5344 // been already stolen from us.
5345 continue;
5346 }
5347 // Verify that we have no work before we resort to stealing
5348 assert(work_q->size() == 0, "Have work, shouldn't steal");
5349 // Try to steal from other queues that have work
5350 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5351 NOT_PRODUCT(num_steals++;)
5352 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5353 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5354 // Do scanning work
5355 obj_to_scan->oop_iterate(keep_alive);
5356 // Loop around, finish this work, and try to steal some more
5357 } else if (terminator()->offer_termination()) {
5358 break; // nirvana from the infinite cycle
5359 }
5360 }
5361 NOT_PRODUCT(
5362 if (PrintCMSStatistics != 0) {
5363 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5364 }
5365 )
5366 }
5367
5368 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5369 {
5370 GenCollectedHeap* gch = GenCollectedHeap::heap();
5371 WorkGang* workers = gch->workers();
5372 assert(workers != NULL, "Need parallel worker threads.");
5373 CMSRefProcTaskProxy rp_task(task, &_collector,
5374 _collector.ref_processor()->span(),
5375 _collector.markBitMap(),
5376 workers, _collector.task_queues());
5377 workers->run_task(&rp_task);
5378 }
5379
5380 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5381 {
5382
5383 GenCollectedHeap* gch = GenCollectedHeap::heap();
5384 WorkGang* workers = gch->workers();
5385 assert(workers != NULL, "Need parallel worker threads.");
5386 CMSRefEnqueueTaskProxy enq_task(task);
5387 workers->run_task(&enq_task);
5388 }
5389
5390 void CMSCollector::refProcessingWork() {
5391 ResourceMark rm;
5392 HandleMark hm;
5393
5394 ReferenceProcessor* rp = ref_processor();
5395 assert(rp->span().equals(_span), "Spans should be equal");
5396 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5397 // Process weak references.
5398 rp->setup_policy(false);
5399 verify_work_stacks_empty();
5400
5401 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5402 &_markStack, false /* !preclean */);
5403 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5404 _span, &_markBitMap, &_markStack,
5405 &cmsKeepAliveClosure, false /* !preclean */);
5406 {
5407 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5408
5409 ReferenceProcessorStats stats;
5410 if (rp->processing_is_mt()) {
5411 // Set the degree of MT here. If the discovery is done MT, there
5412 // may have been a different number of threads doing the discovery
5413 // and a different number of discovered lists may have Ref objects.
5414 // That is OK as long as the Reference lists are balanced (see
5415 // balance_all_queues() and balance_queues()).
5416 GenCollectedHeap* gch = GenCollectedHeap::heap();
5417 uint active_workers = ParallelGCThreads;
5418 WorkGang* workers = gch->workers();
5419 if (workers != NULL) {
5420 active_workers = workers->active_workers();
5421 // The expectation is that active_workers will have already
5422 // been set to a reasonable value. If it has not been set,
5423 // investigate.
5424 assert(active_workers > 0, "Should have been set during scavenge");
5425 }
5426 rp->set_active_mt_degree(active_workers);
5427 CMSRefProcTaskExecutor task_executor(*this);
5428 stats = rp->process_discovered_references(&_is_alive_closure,
5429 &cmsKeepAliveClosure,
5430 &cmsDrainMarkingStackClosure,
5431 &task_executor,
5432 _gc_timer_cm,
5433 _gc_tracer_cm->gc_id());
5434 } else {
5435 stats = rp->process_discovered_references(&_is_alive_closure,
5436 &cmsKeepAliveClosure,
5437 &cmsDrainMarkingStackClosure,
5438 NULL,
5439 _gc_timer_cm,
5440 _gc_tracer_cm->gc_id());
5441 }
5442 _gc_tracer_cm->report_gc_reference_stats(stats);
5443
5444 }
5445
5446 // This is the point where the entire marking should have completed.
5447 verify_work_stacks_empty();
5448
5449 if (should_unload_classes()) {
5450 {
5451 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5452
5453 // Unload classes and purge the SystemDictionary.
5454 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5455
5456 // Unload nmethods.
5457 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5458
5459 // Prune dead klasses from subklass/sibling/implementor lists.
5460 Klass::clean_weak_klass_links(&_is_alive_closure);
5461 }
5462
5463 {
5464 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5465 // Clean up unreferenced symbols in symbol table.
5466 SymbolTable::unlink();
5467 }
5468
5469 {
5470 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5471 // Delete entries for dead interned strings.
5472 StringTable::unlink(&_is_alive_closure);
5473 }
5474 }
5475
5476
5477 // Restore any preserved marks as a result of mark stack or
5478 // work queue overflow
5479 restore_preserved_marks_if_any(); // done single-threaded for now
5480
5481 rp->set_enqueuing_is_done(true);
5482 if (rp->processing_is_mt()) {
5483 rp->balance_all_queues();
5484 CMSRefProcTaskExecutor task_executor(*this);
5485 rp->enqueue_discovered_references(&task_executor);
5486 } else {
5487 rp->enqueue_discovered_references(NULL);
5488 }
5489 rp->verify_no_references_recorded();
5490 assert(!rp->discovery_enabled(), "should have been disabled");
5491 }
5492
5493 #ifndef PRODUCT
5494 void CMSCollector::check_correct_thread_executing() {
5495 Thread* t = Thread::current();
5496 // Only the VM thread or the CMS thread should be here.
5497 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5498 "Unexpected thread type");
5499 // If this is the vm thread, the foreground process
5500 // should not be waiting. Note that _foregroundGCIsActive is
5501 // true while the foreground collector is waiting.
5502 if (_foregroundGCShouldWait) {
5503 // We cannot be the VM thread
5504 assert(t->is_ConcurrentGC_thread(),
5505 "Should be CMS thread");
5506 } else {
5507 // We can be the CMS thread only if we are in a stop-world
5508 // phase of CMS collection.
5509 if (t->is_ConcurrentGC_thread()) {
5510 assert(_collectorState == InitialMarking ||
5511 _collectorState == FinalMarking,
5512 "Should be a stop-world phase");
5513 // The CMS thread should be holding the CMS_token.
5514 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5515 "Potential interference with concurrently "
5516 "executing VM thread");
5517 }
5518 }
5519 }
5520 #endif
5521
5522 void CMSCollector::sweep() {
5523 assert(_collectorState == Sweeping, "just checking");
5524 check_correct_thread_executing();
5525 verify_work_stacks_empty();
5526 verify_overflow_empty();
5527 increment_sweep_count();
5528 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5529
5530 _inter_sweep_timer.stop();
5531 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5532
5533 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5534 _intra_sweep_timer.reset();
5535 _intra_sweep_timer.start();
5536 {
5537 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5538 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5539 // First sweep the old gen
5540 {
5541 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5542 bitMapLock());
5543 sweepWork(_cmsGen);
5544 }
5545
5546 // Update Universe::_heap_*_at_gc figures.
5547 // We need all the free list locks to make the abstract state
5548 // transition from Sweeping to Resetting. See detailed note
5549 // further below.
5550 {
5551 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5552 // Update heap occupancy information which is used as
5553 // input to soft ref clearing policy at the next gc.
5554 Universe::update_heap_info_at_gc();
5555 _collectorState = Resizing;
5556 }
5557 }
5558 verify_work_stacks_empty();
5559 verify_overflow_empty();
5560
5561 if (should_unload_classes()) {
5562 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5563 // requires that the virtual spaces are stable and not deleted.
5564 ClassLoaderDataGraph::set_should_purge(true);
5565 }
5566
5567 _intra_sweep_timer.stop();
5568 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5569
5570 _inter_sweep_timer.reset();
5571 _inter_sweep_timer.start();
5572
5573 // We need to use a monotonically non-decreasing time in ms
5574 // or we will see time-warp warnings and os::javaTimeMillis()
5575 // does not guarantee monotonicity.
5576 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5577 update_time_of_last_gc(now);
5578
5579 // NOTE on abstract state transitions:
5580 // Mutators allocate-live and/or mark the mod-union table dirty
5581 // based on the state of the collection. The former is done in
5582 // the interval [Marking, Sweeping] and the latter in the interval
5583 // [Marking, Sweeping). Thus the transitions into the Marking state
5584 // and out of the Sweeping state must be synchronously visible
5585 // globally to the mutators.
5586 // The transition into the Marking state happens with the world
5587 // stopped so the mutators will globally see it. Sweeping is
5588 // done asynchronously by the background collector so the transition
5589 // from the Sweeping state to the Resizing state must be done
5590 // under the freelistLock (as is the check for whether to
5591 // allocate-live and whether to dirty the mod-union table).
5592 assert(_collectorState == Resizing, "Change of collector state to"
5593 " Resizing must be done under the freelistLocks (plural)");
5594
5595 // Now that sweeping has been completed, we clear
5596 // the incremental_collection_failed flag,
5597 // thus inviting a younger gen collection to promote into
5598 // this generation. If such a promotion may still fail,
5599 // the flag will be set again when a young collection is
5600 // attempted.
5601 GenCollectedHeap* gch = GenCollectedHeap::heap();
5602 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5603 gch->update_full_collections_completed(_collection_count_start);
5604 }
5605
5606 // FIX ME!!! Looks like this belongs in CFLSpace, with
5607 // CMSGen merely delegating to it.
5608 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5609 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5610 HeapWord* minAddr = _cmsSpace->bottom();
5611 HeapWord* largestAddr =
5612 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5613 if (largestAddr == NULL) {
5614 // The dictionary appears to be empty. In this case
5615 // try to coalesce at the end of the heap.
5616 largestAddr = _cmsSpace->end();
5617 }
5618 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5619 size_t nearLargestOffset =
5620 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5621 if (PrintFLSStatistics != 0) {
5622 gclog_or_tty->print_cr(
5623 "CMS: Large Block: " PTR_FORMAT ";"
5624 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5625 p2i(largestAddr),
5626 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5627 }
5628 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5629 }
5630
5631 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5632 return addr >= _cmsSpace->nearLargestChunk();
5633 }
5634
5635 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5636 return _cmsSpace->find_chunk_at_end();
5637 }
5638
5639 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5640 bool full) {
5641 // If the young generation has been collected, gather any statistics
5642 // that are of interest at this point.
5643 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5644 if (!full && current_is_young) {
5645 // Gather statistics on the young generation collection.
5646 collector()->stats().record_gc0_end(used());
5647 }
5648 }
5649
5650 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5651 // We iterate over the space(s) underlying this generation,
5652 // checking the mark bit map to see if the bits corresponding
5653 // to specific blocks are marked or not. Blocks that are
5654 // marked are live and are not swept up. All remaining blocks
5655 // are swept up, with coalescing on-the-fly as we sweep up
5656 // contiguous free and/or garbage blocks:
5657 // We need to ensure that the sweeper synchronizes with allocators
5658 // and stop-the-world collectors. In particular, the following
5659 // locks are used:
5660 // . CMS token: if this is held, a stop the world collection cannot occur
5661 // . freelistLock: if this is held no allocation can occur from this
5662 // generation by another thread
5663 // . bitMapLock: if this is held, no other thread can access or update
5664 //
5665
5666 // Note that we need to hold the freelistLock if we use
5667 // block iterate below; else the iterator might go awry if
5668 // a mutator (or promotion) causes block contents to change
5669 // (for instance if the allocator divvies up a block).
5670 // If we hold the free list lock, for all practical purposes
5671 // young generation GC's can't occur (they'll usually need to
5672 // promote), so we might as well prevent all young generation
5673 // GC's while we do a sweeping step. For the same reason, we might
5674 // as well take the bit map lock for the entire duration
5675
5676 // check that we hold the requisite locks
5677 assert(have_cms_token(), "Should hold cms token");
5678 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5679 assert_lock_strong(old_gen->freelistLock());
5680 assert_lock_strong(bitMapLock());
5681
5682 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5683 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5684 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5685 _inter_sweep_estimate.padded_average(),
5686 _intra_sweep_estimate.padded_average());
5687 old_gen->setNearLargestChunk();
5688
5689 {
5690 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5691 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5692 // We need to free-up/coalesce garbage/blocks from a
5693 // co-terminal free run. This is done in the SweepClosure
5694 // destructor; so, do not remove this scope, else the
5695 // end-of-sweep-census below will be off by a little bit.
5696 }
5697 old_gen->cmsSpace()->sweep_completed();
5698 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5699 if (should_unload_classes()) { // unloaded classes this cycle,
5700 _concurrent_cycles_since_last_unload = 0; // ... reset count
5701 } else { // did not unload classes,
5702 _concurrent_cycles_since_last_unload++; // ... increment count
5703 }
5704 }
5705
5706 // Reset CMS data structures (for now just the marking bit map)
5707 // preparatory for the next cycle.
5708 void CMSCollector::reset(bool concurrent) {
5709 if (concurrent) {
5710 CMSTokenSyncWithLocks ts(true, bitMapLock());
5711
5712 // If the state is not "Resetting", the foreground thread
5713 // has done a collection and the resetting.
5714 if (_collectorState != Resetting) {
5715 assert(_collectorState == Idling, "The state should only change"
5716 " because the foreground collector has finished the collection");
5717 return;
5718 }
5719
5720 // Clear the mark bitmap (no grey objects to start with)
5721 // for the next cycle.
5722 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5723 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5724
5725 HeapWord* curAddr = _markBitMap.startWord();
5726 while (curAddr < _markBitMap.endWord()) {
5727 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5728 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5729 _markBitMap.clear_large_range(chunk);
5730 if (ConcurrentMarkSweepThread::should_yield() &&
5731 !foregroundGCIsActive() &&
5732 CMSYield) {
5733 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5734 "CMS thread should hold CMS token");
5735 assert_lock_strong(bitMapLock());
5736 bitMapLock()->unlock();
5737 ConcurrentMarkSweepThread::desynchronize(true);
5738 stopTimer();
5739 if (PrintCMSStatistics != 0) {
5740 incrementYields();
5741 }
5742
5743 // See the comment in coordinator_yield()
5744 for (unsigned i = 0; i < CMSYieldSleepCount &&
5745 ConcurrentMarkSweepThread::should_yield() &&
5746 !CMSCollector::foregroundGCIsActive(); ++i) {
5747 os::sleep(Thread::current(), 1, false);
5748 }
5749
5750 ConcurrentMarkSweepThread::synchronize(true);
5751 bitMapLock()->lock_without_safepoint_check();
5752 startTimer();
5753 }
5754 curAddr = chunk.end();
5755 }
5756 // A successful mostly concurrent collection has been done.
5757 // Because only the full (i.e., concurrent mode failure) collections
5758 // are being measured for gc overhead limits, clean the "near" flag
5759 // and count.
5760 size_policy()->reset_gc_overhead_limit_count();
5761 _collectorState = Idling;
5762 } else {
5763 // already have the lock
5764 assert(_collectorState == Resetting, "just checking");
5765 assert_lock_strong(bitMapLock());
5766 _markBitMap.clear_all();
5767 _collectorState = Idling;
5768 }
5769
5770 register_gc_end();
5771 }
5772
5773 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5774 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5775 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5776 TraceCollectorStats tcs(counters());
5777
5778 switch (op) {
5779 case CMS_op_checkpointRootsInitial: {
5780 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5781 checkpointRootsInitial();
5782 if (PrintGC) {
5783 _cmsGen->printOccupancy("initial-mark");
5784 }
5785 break;
5786 }
5787 case CMS_op_checkpointRootsFinal: {
5788 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5789 checkpointRootsFinal();
5790 if (PrintGC) {
5791 _cmsGen->printOccupancy("remark");
5792 }
5793 break;
5794 }
5795 default:
5796 fatal("No such CMS_op");
5797 }
5798 }
5799
5800 #ifndef PRODUCT
5801 size_t const CMSCollector::skip_header_HeapWords() {
5802 return FreeChunk::header_size();
5803 }
5804
5805 // Try and collect here conditions that should hold when
5806 // CMS thread is exiting. The idea is that the foreground GC
5807 // thread should not be blocked if it wants to terminate
5808 // the CMS thread and yet continue to run the VM for a while
5809 // after that.
5810 void CMSCollector::verify_ok_to_terminate() const {
5811 assert(Thread::current()->is_ConcurrentGC_thread(),
5812 "should be called by CMS thread");
5813 assert(!_foregroundGCShouldWait, "should be false");
5814 // We could check here that all the various low-level locks
5815 // are not held by the CMS thread, but that is overkill; see
5816 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5817 // is checked.
5818 }
5819 #endif
5820
5821 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5822 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5823 "missing Printezis mark?");
5824 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5825 size_t size = pointer_delta(nextOneAddr + 1, addr);
5826 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5827 "alignment problem");
5828 assert(size >= 3, "Necessary for Printezis marks to work");
5829 return size;
5830 }
5831
5832 // A variant of the above (block_size_using_printezis_bits()) except
5833 // that we return 0 if the P-bits are not yet set.
5834 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5835 if (_markBitMap.isMarked(addr + 1)) {
5836 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5837 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5838 size_t size = pointer_delta(nextOneAddr + 1, addr);
5839 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5840 "alignment problem");
5841 assert(size >= 3, "Necessary for Printezis marks to work");
5842 return size;
5843 }
5844 return 0;
5845 }
5846
5847 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5848 size_t sz = 0;
5849 oop p = (oop)addr;
5850 if (p->klass_or_null() != NULL) {
5851 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5852 } else {
5853 sz = block_size_using_printezis_bits(addr);
5854 }
5855 assert(sz > 0, "size must be nonzero");
5856 HeapWord* next_block = addr + sz;
5857 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5858 CardTableModRefBS::card_size);
5859 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5860 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5861 "must be different cards");
5862 return next_card;
5863 }
5864
5865
5866 // CMS Bit Map Wrapper /////////////////////////////////////////
5867
5868 // Construct a CMS bit map infrastructure, but don't create the
5869 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5870 // further below.
5871 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5872 _bm(),
5873 _shifter(shifter),
5874 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5875 Monitor::_safepoint_check_sometimes) : NULL)
5876 {
5877 _bmStartWord = 0;
5878 _bmWordSize = 0;
5879 }
5880
5881 bool CMSBitMap::allocate(MemRegion mr) {
5882 _bmStartWord = mr.start();
5883 _bmWordSize = mr.word_size();
5884 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5885 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5886 if (!brs.is_reserved()) {
5887 warning("CMS bit map allocation failure");
5888 return false;
5889 }
5890 // For now we'll just commit all of the bit map up front.
5891 // Later on we'll try to be more parsimonious with swap.
5892 if (!_virtual_space.initialize(brs, brs.size())) {
5893 warning("CMS bit map backing store failure");
5894 return false;
5895 }
5896 assert(_virtual_space.committed_size() == brs.size(),
5897 "didn't reserve backing store for all of CMS bit map?");
5898 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5899 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5900 _bmWordSize, "inconsistency in bit map sizing");
5901 _bm.set_size(_bmWordSize >> _shifter);
5902
5903 // bm.clear(); // can we rely on getting zero'd memory? verify below
5904 assert(isAllClear(),
5905 "Expected zero'd memory from ReservedSpace constructor");
5906 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5907 "consistency check");
5908 return true;
5909 }
5910
5911 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5912 HeapWord *next_addr, *end_addr, *last_addr;
5913 assert_locked();
5914 assert(covers(mr), "out-of-range error");
5915 // XXX assert that start and end are appropriately aligned
5916 for (next_addr = mr.start(), end_addr = mr.end();
5917 next_addr < end_addr; next_addr = last_addr) {
5918 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5919 last_addr = dirty_region.end();
5920 if (!dirty_region.is_empty()) {
5921 cl->do_MemRegion(dirty_region);
5922 } else {
5923 assert(last_addr == end_addr, "program logic");
5924 return;
5925 }
5926 }
5927 }
5928
5929 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5930 _bm.print_on_error(st, prefix);
5931 }
5932
5933 #ifndef PRODUCT
5934 void CMSBitMap::assert_locked() const {
5935 CMSLockVerifier::assert_locked(lock());
5936 }
5937
5938 bool CMSBitMap::covers(MemRegion mr) const {
5939 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5940 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5941 "size inconsistency");
5942 return (mr.start() >= _bmStartWord) &&
5943 (mr.end() <= endWord());
5944 }
5945
5946 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5947 return (start >= _bmStartWord && (start + size) <= endWord());
5948 }
5949
5950 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5951 // verify that there are no 1 bits in the interval [left, right)
5952 FalseBitMapClosure falseBitMapClosure;
5953 iterate(&falseBitMapClosure, left, right);
5954 }
5955
5956 void CMSBitMap::region_invariant(MemRegion mr)
5957 {
5958 assert_locked();
5959 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5960 assert(!mr.is_empty(), "unexpected empty region");
5961 assert(covers(mr), "mr should be covered by bit map");
5962 // convert address range into offset range
5963 size_t start_ofs = heapWordToOffset(mr.start());
5964 // Make sure that end() is appropriately aligned
5965 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5966 (1 << (_shifter+LogHeapWordSize))),
5967 "Misaligned mr.end()");
5968 size_t end_ofs = heapWordToOffset(mr.end());
5969 assert(end_ofs > start_ofs, "Should mark at least one bit");
5970 }
5971
5972 #endif
5973
5974 bool CMSMarkStack::allocate(size_t size) {
5975 // allocate a stack of the requisite depth
5976 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5977 size * sizeof(oop)));
5978 if (!rs.is_reserved()) {
5979 warning("CMSMarkStack allocation failure");
5980 return false;
5981 }
5982 if (!_virtual_space.initialize(rs, rs.size())) {
5983 warning("CMSMarkStack backing store failure");
5984 return false;
5985 }
5986 assert(_virtual_space.committed_size() == rs.size(),
5987 "didn't reserve backing store for all of CMS stack?");
5988 _base = (oop*)(_virtual_space.low());
5989 _index = 0;
5990 _capacity = size;
5991 NOT_PRODUCT(_max_depth = 0);
5992 return true;
5993 }
5994
5995 // XXX FIX ME !!! In the MT case we come in here holding a
5996 // leaf lock. For printing we need to take a further lock
5997 // which has lower rank. We need to recalibrate the two
5998 // lock-ranks involved in order to be able to print the
5999 // messages below. (Or defer the printing to the caller.
6000 // For now we take the expedient path of just disabling the
6001 // messages for the problematic case.)
6002 void CMSMarkStack::expand() {
6003 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6004 if (_capacity == MarkStackSizeMax) {
6005 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6006 // We print a warning message only once per CMS cycle.
6007 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6008 }
6009 return;
6010 }
6011 // Double capacity if possible
6012 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6013 // Do not give up existing stack until we have managed to
6014 // get the double capacity that we desired.
6015 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6016 new_capacity * sizeof(oop)));
6017 if (rs.is_reserved()) {
6018 // Release the backing store associated with old stack
6019 _virtual_space.release();
6020 // Reinitialize virtual space for new stack
6021 if (!_virtual_space.initialize(rs, rs.size())) {
6022 fatal("Not enough swap for expanded marking stack");
6023 }
6024 _base = (oop*)(_virtual_space.low());
6025 _index = 0;
6026 _capacity = new_capacity;
6027 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6028 // Failed to double capacity, continue;
6029 // we print a detail message only once per CMS cycle.
6030 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to "
6031 SIZE_FORMAT "K",
6032 _capacity / K, new_capacity / K);
6033 }
6034 }
6035
6036
6037 // Closures
6038 // XXX: there seems to be a lot of code duplication here;
6039 // should refactor and consolidate common code.
6040
6041 // This closure is used to mark refs into the CMS generation in
6042 // the CMS bit map. Called at the first checkpoint. This closure
6043 // assumes that we do not need to re-mark dirty cards; if the CMS
6044 // generation on which this is used is not an oldest
6045 // generation then this will lose younger_gen cards!
6046
6047 MarkRefsIntoClosure::MarkRefsIntoClosure(
6048 MemRegion span, CMSBitMap* bitMap):
6049 _span(span),
6050 _bitMap(bitMap)
6051 {
6052 assert(_ref_processor == NULL, "deliberately left NULL");
6053 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6054 }
6055
6056 void MarkRefsIntoClosure::do_oop(oop obj) {
6057 // if p points into _span, then mark corresponding bit in _markBitMap
6058 assert(obj->is_oop(), "expected an oop");
6059 HeapWord* addr = (HeapWord*)obj;
6060 if (_span.contains(addr)) {
6061 // this should be made more efficient
6062 _bitMap->mark(addr);
6063 }
6064 }
6065
6066 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6067 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6068
6069 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6070 MemRegion span, CMSBitMap* bitMap):
6071 _span(span),
6072 _bitMap(bitMap)
6073 {
6074 assert(_ref_processor == NULL, "deliberately left NULL");
6075 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6076 }
6077
6078 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6079 // if p points into _span, then mark corresponding bit in _markBitMap
6080 assert(obj->is_oop(), "expected an oop");
6081 HeapWord* addr = (HeapWord*)obj;
6082 if (_span.contains(addr)) {
6083 // this should be made more efficient
6084 _bitMap->par_mark(addr);
6085 }
6086 }
6087
6088 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6089 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6090
6091 // A variant of the above, used for CMS marking verification.
6092 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6093 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6094 _span(span),
6095 _verification_bm(verification_bm),
6096 _cms_bm(cms_bm)
6097 {
6098 assert(_ref_processor == NULL, "deliberately left NULL");
6099 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6100 }
6101
6102 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6103 // if p points into _span, then mark corresponding bit in _markBitMap
6104 assert(obj->is_oop(), "expected an oop");
6105 HeapWord* addr = (HeapWord*)obj;
6106 if (_span.contains(addr)) {
6107 _verification_bm->mark(addr);
6108 if (!_cms_bm->isMarked(addr)) {
6109 oop(addr)->print();
6110 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6111 fatal("... aborting");
6112 }
6113 }
6114 }
6115
6116 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6117 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6118
6119 //////////////////////////////////////////////////
6120 // MarkRefsIntoAndScanClosure
6121 //////////////////////////////////////////////////
6122
6123 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6124 ReferenceProcessor* rp,
6125 CMSBitMap* bit_map,
6126 CMSBitMap* mod_union_table,
6127 CMSMarkStack* mark_stack,
6128 CMSCollector* collector,
6129 bool should_yield,
6130 bool concurrent_precleaning):
6131 _collector(collector),
6132 _span(span),
6133 _bit_map(bit_map),
6134 _mark_stack(mark_stack),
6135 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6136 mark_stack, concurrent_precleaning),
6137 _yield(should_yield),
6138 _concurrent_precleaning(concurrent_precleaning),
6139 _freelistLock(NULL)
6140 {
6141 _ref_processor = rp;
6142 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6143 }
6144
6145 // This closure is used to mark refs into the CMS generation at the
6146 // second (final) checkpoint, and to scan and transitively follow
6147 // the unmarked oops. It is also used during the concurrent precleaning
6148 // phase while scanning objects on dirty cards in the CMS generation.
6149 // The marks are made in the marking bit map and the marking stack is
6150 // used for keeping the (newly) grey objects during the scan.
6151 // The parallel version (Par_...) appears further below.
6152 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6153 if (obj != NULL) {
6154 assert(obj->is_oop(), "expected an oop");
6155 HeapWord* addr = (HeapWord*)obj;
6156 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6157 assert(_collector->overflow_list_is_empty(),
6158 "overflow list should be empty");
6159 if (_span.contains(addr) &&
6160 !_bit_map->isMarked(addr)) {
6161 // mark bit map (object is now grey)
6162 _bit_map->mark(addr);
6163 // push on marking stack (stack should be empty), and drain the
6164 // stack by applying this closure to the oops in the oops popped
6165 // from the stack (i.e. blacken the grey objects)
6166 bool res = _mark_stack->push(obj);
6167 assert(res, "Should have space to push on empty stack");
6168 do {
6169 oop new_oop = _mark_stack->pop();
6170 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6171 assert(_bit_map->isMarked((HeapWord*)new_oop),
6172 "only grey objects on this stack");
6173 // iterate over the oops in this oop, marking and pushing
6174 // the ones in CMS heap (i.e. in _span).
6175 new_oop->oop_iterate(&_pushAndMarkClosure);
6176 // check if it's time to yield
6177 do_yield_check();
6178 } while (!_mark_stack->isEmpty() ||
6179 (!_concurrent_precleaning && take_from_overflow_list()));
6180 // if marking stack is empty, and we are not doing this
6181 // during precleaning, then check the overflow list
6182 }
6183 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6184 assert(_collector->overflow_list_is_empty(),
6185 "overflow list was drained above");
6186 // We could restore evacuated mark words, if any, used for
6187 // overflow list links here because the overflow list is
6188 // provably empty here. That would reduce the maximum
6189 // size requirements for preserved_{oop,mark}_stack.
6190 // But we'll just postpone it until we are all done
6191 // so we can just stream through.
6192 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6193 _collector->restore_preserved_marks_if_any();
6194 assert(_collector->no_preserved_marks(), "No preserved marks");
6195 }
6196 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6197 "All preserved marks should have been restored above");
6198 }
6199 }
6200
6201 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6202 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6203
6204 void MarkRefsIntoAndScanClosure::do_yield_work() {
6205 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6206 "CMS thread should hold CMS token");
6207 assert_lock_strong(_freelistLock);
6208 assert_lock_strong(_bit_map->lock());
6209 // relinquish the free_list_lock and bitMaplock()
6210 _bit_map->lock()->unlock();
6211 _freelistLock->unlock();
6212 ConcurrentMarkSweepThread::desynchronize(true);
6213 _collector->stopTimer();
6214 if (PrintCMSStatistics != 0) {
6215 _collector->incrementYields();
6216 }
6217
6218 // See the comment in coordinator_yield()
6219 for (unsigned i = 0;
6220 i < CMSYieldSleepCount &&
6221 ConcurrentMarkSweepThread::should_yield() &&
6222 !CMSCollector::foregroundGCIsActive();
6223 ++i) {
6224 os::sleep(Thread::current(), 1, false);
6225 }
6226
6227 ConcurrentMarkSweepThread::synchronize(true);
6228 _freelistLock->lock_without_safepoint_check();
6229 _bit_map->lock()->lock_without_safepoint_check();
6230 _collector->startTimer();
6231 }
6232
6233 ///////////////////////////////////////////////////////////
6234 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6235 // MarkRefsIntoAndScanClosure
6236 ///////////////////////////////////////////////////////////
6237 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6238 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6239 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6240 _span(span),
6241 _bit_map(bit_map),
6242 _work_queue(work_queue),
6243 _low_water_mark(MIN2((work_queue->max_elems()/4),
6244 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6245 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6246 {
6247 _ref_processor = rp;
6248 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6249 }
6250
6251 // This closure is used to mark refs into the CMS generation at the
6252 // second (final) checkpoint, and to scan and transitively follow
6253 // the unmarked oops. The marks are made in the marking bit map and
6254 // the work_queue is used for keeping the (newly) grey objects during
6255 // the scan phase whence they are also available for stealing by parallel
6256 // threads. Since the marking bit map is shared, updates are
6257 // synchronized (via CAS).
6258 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6259 if (obj != NULL) {
6260 // Ignore mark word because this could be an already marked oop
6261 // that may be chained at the end of the overflow list.
6262 assert(obj->is_oop(true), "expected an oop");
6263 HeapWord* addr = (HeapWord*)obj;
6264 if (_span.contains(addr) &&
6265 !_bit_map->isMarked(addr)) {
6266 // mark bit map (object will become grey):
6267 // It is possible for several threads to be
6268 // trying to "claim" this object concurrently;
6269 // the unique thread that succeeds in marking the
6270 // object first will do the subsequent push on
6271 // to the work queue (or overflow list).
6272 if (_bit_map->par_mark(addr)) {
6273 // push on work_queue (which may not be empty), and trim the
6274 // queue to an appropriate length by applying this closure to
6275 // the oops in the oops popped from the stack (i.e. blacken the
6276 // grey objects)
6277 bool res = _work_queue->push(obj);
6278 assert(res, "Low water mark should be less than capacity?");
6279 trim_queue(_low_water_mark);
6280 } // Else, another thread claimed the object
6281 }
6282 }
6283 }
6284
6285 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6286 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6287
6288 // This closure is used to rescan the marked objects on the dirty cards
6289 // in the mod union table and the card table proper.
6290 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6291 oop p, MemRegion mr) {
6292
6293 size_t size = 0;
6294 HeapWord* addr = (HeapWord*)p;
6295 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6296 assert(_span.contains(addr), "we are scanning the CMS generation");
6297 // check if it's time to yield
6298 if (do_yield_check()) {
6299 // We yielded for some foreground stop-world work,
6300 // and we have been asked to abort this ongoing preclean cycle.
6301 return 0;
6302 }
6303 if (_bitMap->isMarked(addr)) {
6304 // it's marked; is it potentially uninitialized?
6305 if (p->klass_or_null() != NULL) {
6306 // an initialized object; ignore mark word in verification below
6307 // since we are running concurrent with mutators
6308 assert(p->is_oop(true), "should be an oop");
6309 if (p->is_objArray()) {
6310 // objArrays are precisely marked; restrict scanning
6311 // to dirty cards only.
6312 size = CompactibleFreeListSpace::adjustObjectSize(
6313 p->oop_iterate_size(_scanningClosure, mr));
6314 } else {
6315 // A non-array may have been imprecisely marked; we need
6316 // to scan object in its entirety.
6317 size = CompactibleFreeListSpace::adjustObjectSize(
6318 p->oop_iterate_size(_scanningClosure));
6319 }
6320 #ifdef ASSERT
6321 size_t direct_size =
6322 CompactibleFreeListSpace::adjustObjectSize(p->size());
6323 assert(size == direct_size, "Inconsistency in size");
6324 assert(size >= 3, "Necessary for Printezis marks to work");
6325 if (!_bitMap->isMarked(addr+1)) {
6326 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6327 } else {
6328 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6329 assert(_bitMap->isMarked(addr+size-1),
6330 "inconsistent Printezis mark");
6331 }
6332 #endif // ASSERT
6333 } else {
6334 // An uninitialized object.
6335 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6336 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6337 size = pointer_delta(nextOneAddr + 1, addr);
6338 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6339 "alignment problem");
6340 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6341 // will dirty the card when the klass pointer is installed in the
6342 // object (signaling the completion of initialization).
6343 }
6344 } else {
6345 // Either a not yet marked object or an uninitialized object
6346 if (p->klass_or_null() == NULL) {
6347 // An uninitialized object, skip to the next card, since
6348 // we may not be able to read its P-bits yet.
6349 assert(size == 0, "Initial value");
6350 } else {
6351 // An object not (yet) reached by marking: we merely need to
6352 // compute its size so as to go look at the next block.
6353 assert(p->is_oop(true), "should be an oop");
6354 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6355 }
6356 }
6357 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6358 return size;
6359 }
6360
6361 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6362 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6363 "CMS thread should hold CMS token");
6364 assert_lock_strong(_freelistLock);
6365 assert_lock_strong(_bitMap->lock());
6366 // relinquish the free_list_lock and bitMaplock()
6367 _bitMap->lock()->unlock();
6368 _freelistLock->unlock();
6369 ConcurrentMarkSweepThread::desynchronize(true);
6370 _collector->stopTimer();
6371 if (PrintCMSStatistics != 0) {
6372 _collector->incrementYields();
6373 }
6374
6375 // See the comment in coordinator_yield()
6376 for (unsigned i = 0; i < CMSYieldSleepCount &&
6377 ConcurrentMarkSweepThread::should_yield() &&
6378 !CMSCollector::foregroundGCIsActive(); ++i) {
6379 os::sleep(Thread::current(), 1, false);
6380 }
6381
6382 ConcurrentMarkSweepThread::synchronize(true);
6383 _freelistLock->lock_without_safepoint_check();
6384 _bitMap->lock()->lock_without_safepoint_check();
6385 _collector->startTimer();
6386 }
6387
6388
6389 //////////////////////////////////////////////////////////////////
6390 // SurvivorSpacePrecleanClosure
6391 //////////////////////////////////////////////////////////////////
6392 // This (single-threaded) closure is used to preclean the oops in
6393 // the survivor spaces.
6394 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6395
6396 HeapWord* addr = (HeapWord*)p;
6397 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6398 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6399 assert(p->klass_or_null() != NULL, "object should be initialized");
6400 // an initialized object; ignore mark word in verification below
6401 // since we are running concurrent with mutators
6402 assert(p->is_oop(true), "should be an oop");
6403 // Note that we do not yield while we iterate over
6404 // the interior oops of p, pushing the relevant ones
6405 // on our marking stack.
6406 size_t size = p->oop_iterate_size(_scanning_closure);
6407 do_yield_check();
6408 // Observe that below, we do not abandon the preclean
6409 // phase as soon as we should; rather we empty the
6410 // marking stack before returning. This is to satisfy
6411 // some existing assertions. In general, it may be a
6412 // good idea to abort immediately and complete the marking
6413 // from the grey objects at a later time.
6414 while (!_mark_stack->isEmpty()) {
6415 oop new_oop = _mark_stack->pop();
6416 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6417 assert(_bit_map->isMarked((HeapWord*)new_oop),
6418 "only grey objects on this stack");
6419 // iterate over the oops in this oop, marking and pushing
6420 // the ones in CMS heap (i.e. in _span).
6421 new_oop->oop_iterate(_scanning_closure);
6422 // check if it's time to yield
6423 do_yield_check();
6424 }
6425 unsigned int after_count =
6426 GenCollectedHeap::heap()->total_collections();
6427 bool abort = (_before_count != after_count) ||
6428 _collector->should_abort_preclean();
6429 return abort ? 0 : size;
6430 }
6431
6432 void SurvivorSpacePrecleanClosure::do_yield_work() {
6433 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6434 "CMS thread should hold CMS token");
6435 assert_lock_strong(_bit_map->lock());
6436 // Relinquish the bit map lock
6437 _bit_map->lock()->unlock();
6438 ConcurrentMarkSweepThread::desynchronize(true);
6439 _collector->stopTimer();
6440 if (PrintCMSStatistics != 0) {
6441 _collector->incrementYields();
6442 }
6443
6444 // See the comment in coordinator_yield()
6445 for (unsigned i = 0; i < CMSYieldSleepCount &&
6446 ConcurrentMarkSweepThread::should_yield() &&
6447 !CMSCollector::foregroundGCIsActive(); ++i) {
6448 os::sleep(Thread::current(), 1, false);
6449 }
6450
6451 ConcurrentMarkSweepThread::synchronize(true);
6452 _bit_map->lock()->lock_without_safepoint_check();
6453 _collector->startTimer();
6454 }
6455
6456 // This closure is used to rescan the marked objects on the dirty cards
6457 // in the mod union table and the card table proper. In the parallel
6458 // case, although the bitMap is shared, we do a single read so the
6459 // isMarked() query is "safe".
6460 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6461 // Ignore mark word because we are running concurrent with mutators
6462 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6463 HeapWord* addr = (HeapWord*)p;
6464 assert(_span.contains(addr), "we are scanning the CMS generation");
6465 bool is_obj_array = false;
6466 #ifdef ASSERT
6467 if (!_parallel) {
6468 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6469 assert(_collector->overflow_list_is_empty(),
6470 "overflow list should be empty");
6471
6472 }
6473 #endif // ASSERT
6474 if (_bit_map->isMarked(addr)) {
6475 // Obj arrays are precisely marked, non-arrays are not;
6476 // so we scan objArrays precisely and non-arrays in their
6477 // entirety.
6478 if (p->is_objArray()) {
6479 is_obj_array = true;
6480 if (_parallel) {
6481 p->oop_iterate(_par_scan_closure, mr);
6482 } else {
6483 p->oop_iterate(_scan_closure, mr);
6484 }
6485 } else {
6486 if (_parallel) {
6487 p->oop_iterate(_par_scan_closure);
6488 } else {
6489 p->oop_iterate(_scan_closure);
6490 }
6491 }
6492 }
6493 #ifdef ASSERT
6494 if (!_parallel) {
6495 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6496 assert(_collector->overflow_list_is_empty(),
6497 "overflow list should be empty");
6498
6499 }
6500 #endif // ASSERT
6501 return is_obj_array;
6502 }
6503
6504 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6505 MemRegion span,
6506 CMSBitMap* bitMap, CMSMarkStack* markStack,
6507 bool should_yield, bool verifying):
6508 _collector(collector),
6509 _span(span),
6510 _bitMap(bitMap),
6511 _mut(&collector->_modUnionTable),
6512 _markStack(markStack),
6513 _yield(should_yield),
6514 _skipBits(0)
6515 {
6516 assert(_markStack->isEmpty(), "stack should be empty");
6517 _finger = _bitMap->startWord();
6518 _threshold = _finger;
6519 assert(_collector->_restart_addr == NULL, "Sanity check");
6520 assert(_span.contains(_finger), "Out of bounds _finger?");
6521 DEBUG_ONLY(_verifying = verifying;)
6522 }
6523
6524 void MarkFromRootsClosure::reset(HeapWord* addr) {
6525 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6526 assert(_span.contains(addr), "Out of bounds _finger?");
6527 _finger = addr;
6528 _threshold = (HeapWord*)round_to(
6529 (intptr_t)_finger, CardTableModRefBS::card_size);
6530 }
6531
6532 // Should revisit to see if this should be restructured for
6533 // greater efficiency.
6534 bool MarkFromRootsClosure::do_bit(size_t offset) {
6535 if (_skipBits > 0) {
6536 _skipBits--;
6537 return true;
6538 }
6539 // convert offset into a HeapWord*
6540 HeapWord* addr = _bitMap->startWord() + offset;
6541 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6542 "address out of range");
6543 assert(_bitMap->isMarked(addr), "tautology");
6544 if (_bitMap->isMarked(addr+1)) {
6545 // this is an allocated but not yet initialized object
6546 assert(_skipBits == 0, "tautology");
6547 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6548 oop p = oop(addr);
6549 if (p->klass_or_null() == NULL) {
6550 DEBUG_ONLY(if (!_verifying) {)
6551 // We re-dirty the cards on which this object lies and increase
6552 // the _threshold so that we'll come back to scan this object
6553 // during the preclean or remark phase. (CMSCleanOnEnter)
6554 if (CMSCleanOnEnter) {
6555 size_t sz = _collector->block_size_using_printezis_bits(addr);
6556 HeapWord* end_card_addr = (HeapWord*)round_to(
6557 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6558 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6559 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6560 // Bump _threshold to end_card_addr; note that
6561 // _threshold cannot possibly exceed end_card_addr, anyhow.
6562 // This prevents future clearing of the card as the scan proceeds
6563 // to the right.
6564 assert(_threshold <= end_card_addr,
6565 "Because we are just scanning into this object");
6566 if (_threshold < end_card_addr) {
6567 _threshold = end_card_addr;
6568 }
6569 if (p->klass_or_null() != NULL) {
6570 // Redirty the range of cards...
6571 _mut->mark_range(redirty_range);
6572 } // ...else the setting of klass will dirty the card anyway.
6573 }
6574 DEBUG_ONLY(})
6575 return true;
6576 }
6577 }
6578 scanOopsInOop(addr);
6579 return true;
6580 }
6581
6582 // We take a break if we've been at this for a while,
6583 // so as to avoid monopolizing the locks involved.
6584 void MarkFromRootsClosure::do_yield_work() {
6585 // First give up the locks, then yield, then re-lock
6586 // We should probably use a constructor/destructor idiom to
6587 // do this unlock/lock or modify the MutexUnlocker class to
6588 // serve our purpose. XXX
6589 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6590 "CMS thread should hold CMS token");
6591 assert_lock_strong(_bitMap->lock());
6592 _bitMap->lock()->unlock();
6593 ConcurrentMarkSweepThread::desynchronize(true);
6594 _collector->stopTimer();
6595 if (PrintCMSStatistics != 0) {
6596 _collector->incrementYields();
6597 }
6598
6599 // See the comment in coordinator_yield()
6600 for (unsigned i = 0; i < CMSYieldSleepCount &&
6601 ConcurrentMarkSweepThread::should_yield() &&
6602 !CMSCollector::foregroundGCIsActive(); ++i) {
6603 os::sleep(Thread::current(), 1, false);
6604 }
6605
6606 ConcurrentMarkSweepThread::synchronize(true);
6607 _bitMap->lock()->lock_without_safepoint_check();
6608 _collector->startTimer();
6609 }
6610
6611 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6612 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6613 assert(_markStack->isEmpty(),
6614 "should drain stack to limit stack usage");
6615 // convert ptr to an oop preparatory to scanning
6616 oop obj = oop(ptr);
6617 // Ignore mark word in verification below, since we
6618 // may be running concurrent with mutators.
6619 assert(obj->is_oop(true), "should be an oop");
6620 assert(_finger <= ptr, "_finger runneth ahead");
6621 // advance the finger to right end of this object
6622 _finger = ptr + obj->size();
6623 assert(_finger > ptr, "we just incremented it above");
6624 // On large heaps, it may take us some time to get through
6625 // the marking phase. During
6626 // this time it's possible that a lot of mutations have
6627 // accumulated in the card table and the mod union table --
6628 // these mutation records are redundant until we have
6629 // actually traced into the corresponding card.
6630 // Here, we check whether advancing the finger would make
6631 // us cross into a new card, and if so clear corresponding
6632 // cards in the MUT (preclean them in the card-table in the
6633 // future).
6634
6635 DEBUG_ONLY(if (!_verifying) {)
6636 // The clean-on-enter optimization is disabled by default,
6637 // until we fix 6178663.
6638 if (CMSCleanOnEnter && (_finger > _threshold)) {
6639 // [_threshold, _finger) represents the interval
6640 // of cards to be cleared in MUT (or precleaned in card table).
6641 // The set of cards to be cleared is all those that overlap
6642 // with the interval [_threshold, _finger); note that
6643 // _threshold is always kept card-aligned but _finger isn't
6644 // always card-aligned.
6645 HeapWord* old_threshold = _threshold;
6646 assert(old_threshold == (HeapWord*)round_to(
6647 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6648 "_threshold should always be card-aligned");
6649 _threshold = (HeapWord*)round_to(
6650 (intptr_t)_finger, CardTableModRefBS::card_size);
6651 MemRegion mr(old_threshold, _threshold);
6652 assert(!mr.is_empty(), "Control point invariant");
6653 assert(_span.contains(mr), "Should clear within span");
6654 _mut->clear_range(mr);
6655 }
6656 DEBUG_ONLY(})
6657 // Note: the finger doesn't advance while we drain
6658 // the stack below.
6659 PushOrMarkClosure pushOrMarkClosure(_collector,
6660 _span, _bitMap, _markStack,
6661 _finger, this);
6662 bool res = _markStack->push(obj);
6663 assert(res, "Empty non-zero size stack should have space for single push");
6664 while (!_markStack->isEmpty()) {
6665 oop new_oop = _markStack->pop();
6666 // Skip verifying header mark word below because we are
6667 // running concurrent with mutators.
6668 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6669 // now scan this oop's oops
6670 new_oop->oop_iterate(&pushOrMarkClosure);
6671 do_yield_check();
6672 }
6673 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6674 }
6675
6676 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6677 CMSCollector* collector, MemRegion span,
6678 CMSBitMap* bit_map,
6679 OopTaskQueue* work_queue,
6680 CMSMarkStack* overflow_stack):
6681 _collector(collector),
6682 _whole_span(collector->_span),
6683 _span(span),
6684 _bit_map(bit_map),
6685 _mut(&collector->_modUnionTable),
6686 _work_queue(work_queue),
6687 _overflow_stack(overflow_stack),
6688 _skip_bits(0),
6689 _task(task)
6690 {
6691 assert(_work_queue->size() == 0, "work_queue should be empty");
6692 _finger = span.start();
6693 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6694 assert(_span.contains(_finger), "Out of bounds _finger?");
6695 }
6696
6697 // Should revisit to see if this should be restructured for
6698 // greater efficiency.
6699 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6700 if (_skip_bits > 0) {
6701 _skip_bits--;
6702 return true;
6703 }
6704 // convert offset into a HeapWord*
6705 HeapWord* addr = _bit_map->startWord() + offset;
6706 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6707 "address out of range");
6708 assert(_bit_map->isMarked(addr), "tautology");
6709 if (_bit_map->isMarked(addr+1)) {
6710 // this is an allocated object that might not yet be initialized
6711 assert(_skip_bits == 0, "tautology");
6712 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6713 oop p = oop(addr);
6714 if (p->klass_or_null() == NULL) {
6715 // in the case of Clean-on-Enter optimization, redirty card
6716 // and avoid clearing card by increasing the threshold.
6717 return true;
6718 }
6719 }
6720 scan_oops_in_oop(addr);
6721 return true;
6722 }
6723
6724 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6725 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6726 // Should we assert that our work queue is empty or
6727 // below some drain limit?
6728 assert(_work_queue->size() == 0,
6729 "should drain stack to limit stack usage");
6730 // convert ptr to an oop preparatory to scanning
6731 oop obj = oop(ptr);
6732 // Ignore mark word in verification below, since we
6733 // may be running concurrent with mutators.
6734 assert(obj->is_oop(true), "should be an oop");
6735 assert(_finger <= ptr, "_finger runneth ahead");
6736 // advance the finger to right end of this object
6737 _finger = ptr + obj->size();
6738 assert(_finger > ptr, "we just incremented it above");
6739 // On large heaps, it may take us some time to get through
6740 // the marking phase. During
6741 // this time it's possible that a lot of mutations have
6742 // accumulated in the card table and the mod union table --
6743 // these mutation records are redundant until we have
6744 // actually traced into the corresponding card.
6745 // Here, we check whether advancing the finger would make
6746 // us cross into a new card, and if so clear corresponding
6747 // cards in the MUT (preclean them in the card-table in the
6748 // future).
6749
6750 // The clean-on-enter optimization is disabled by default,
6751 // until we fix 6178663.
6752 if (CMSCleanOnEnter && (_finger > _threshold)) {
6753 // [_threshold, _finger) represents the interval
6754 // of cards to be cleared in MUT (or precleaned in card table).
6755 // The set of cards to be cleared is all those that overlap
6756 // with the interval [_threshold, _finger); note that
6757 // _threshold is always kept card-aligned but _finger isn't
6758 // always card-aligned.
6759 HeapWord* old_threshold = _threshold;
6760 assert(old_threshold == (HeapWord*)round_to(
6761 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6762 "_threshold should always be card-aligned");
6763 _threshold = (HeapWord*)round_to(
6764 (intptr_t)_finger, CardTableModRefBS::card_size);
6765 MemRegion mr(old_threshold, _threshold);
6766 assert(!mr.is_empty(), "Control point invariant");
6767 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6768 _mut->clear_range(mr);
6769 }
6770
6771 // Note: the local finger doesn't advance while we drain
6772 // the stack below, but the global finger sure can and will.
6773 HeapWord** gfa = _task->global_finger_addr();
6774 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6775 _span, _bit_map,
6776 _work_queue,
6777 _overflow_stack,
6778 _finger,
6779 gfa, this);
6780 bool res = _work_queue->push(obj); // overflow could occur here
6781 assert(res, "Will hold once we use workqueues");
6782 while (true) {
6783 oop new_oop;
6784 if (!_work_queue->pop_local(new_oop)) {
6785 // We emptied our work_queue; check if there's stuff that can
6786 // be gotten from the overflow stack.
6787 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6788 _overflow_stack, _work_queue)) {
6789 do_yield_check();
6790 continue;
6791 } else { // done
6792 break;
6793 }
6794 }
6795 // Skip verifying header mark word below because we are
6796 // running concurrent with mutators.
6797 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6798 // now scan this oop's oops
6799 new_oop->oop_iterate(&pushOrMarkClosure);
6800 do_yield_check();
6801 }
6802 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6803 }
6804
6805 // Yield in response to a request from VM Thread or
6806 // from mutators.
6807 void Par_MarkFromRootsClosure::do_yield_work() {
6808 assert(_task != NULL, "sanity");
6809 _task->yield();
6810 }
6811
6812 // A variant of the above used for verifying CMS marking work.
6813 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6814 MemRegion span,
6815 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6816 CMSMarkStack* mark_stack):
6817 _collector(collector),
6818 _span(span),
6819 _verification_bm(verification_bm),
6820 _cms_bm(cms_bm),
6821 _mark_stack(mark_stack),
6822 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6823 mark_stack)
6824 {
6825 assert(_mark_stack->isEmpty(), "stack should be empty");
6826 _finger = _verification_bm->startWord();
6827 assert(_collector->_restart_addr == NULL, "Sanity check");
6828 assert(_span.contains(_finger), "Out of bounds _finger?");
6829 }
6830
6831 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6832 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6833 assert(_span.contains(addr), "Out of bounds _finger?");
6834 _finger = addr;
6835 }
6836
6837 // Should revisit to see if this should be restructured for
6838 // greater efficiency.
6839 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6840 // convert offset into a HeapWord*
6841 HeapWord* addr = _verification_bm->startWord() + offset;
6842 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6843 "address out of range");
6844 assert(_verification_bm->isMarked(addr), "tautology");
6845 assert(_cms_bm->isMarked(addr), "tautology");
6846
6847 assert(_mark_stack->isEmpty(),
6848 "should drain stack to limit stack usage");
6849 // convert addr to an oop preparatory to scanning
6850 oop obj = oop(addr);
6851 assert(obj->is_oop(), "should be an oop");
6852 assert(_finger <= addr, "_finger runneth ahead");
6853 // advance the finger to right end of this object
6854 _finger = addr + obj->size();
6855 assert(_finger > addr, "we just incremented it above");
6856 // Note: the finger doesn't advance while we drain
6857 // the stack below.
6858 bool res = _mark_stack->push(obj);
6859 assert(res, "Empty non-zero size stack should have space for single push");
6860 while (!_mark_stack->isEmpty()) {
6861 oop new_oop = _mark_stack->pop();
6862 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6863 // now scan this oop's oops
6864 new_oop->oop_iterate(&_pam_verify_closure);
6865 }
6866 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6867 return true;
6868 }
6869
6870 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6871 CMSCollector* collector, MemRegion span,
6872 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6873 CMSMarkStack* mark_stack):
6874 MetadataAwareOopClosure(collector->ref_processor()),
6875 _collector(collector),
6876 _span(span),
6877 _verification_bm(verification_bm),
6878 _cms_bm(cms_bm),
6879 _mark_stack(mark_stack)
6880 { }
6881
6882 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6883 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6884
6885 // Upon stack overflow, we discard (part of) the stack,
6886 // remembering the least address amongst those discarded
6887 // in CMSCollector's _restart_address.
6888 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6889 // Remember the least grey address discarded
6890 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6891 _collector->lower_restart_addr(ra);
6892 _mark_stack->reset(); // discard stack contents
6893 _mark_stack->expand(); // expand the stack if possible
6894 }
6895
6896 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6897 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6898 HeapWord* addr = (HeapWord*)obj;
6899 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6900 // Oop lies in _span and isn't yet grey or black
6901 _verification_bm->mark(addr); // now grey
6902 if (!_cms_bm->isMarked(addr)) {
6903 oop(addr)->print();
6904 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6905 p2i(addr));
6906 fatal("... aborting");
6907 }
6908
6909 if (!_mark_stack->push(obj)) { // stack overflow
6910 if (PrintCMSStatistics != 0) {
6911 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6912 SIZE_FORMAT, _mark_stack->capacity());
6913 }
6914 assert(_mark_stack->isFull(), "Else push should have succeeded");
6915 handle_stack_overflow(addr);
6916 }
6917 // anything including and to the right of _finger
6918 // will be scanned as we iterate over the remainder of the
6919 // bit map
6920 }
6921 }
6922
6923 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6924 MemRegion span,
6925 CMSBitMap* bitMap, CMSMarkStack* markStack,
6926 HeapWord* finger, MarkFromRootsClosure* parent) :
6927 MetadataAwareOopClosure(collector->ref_processor()),
6928 _collector(collector),
6929 _span(span),
6930 _bitMap(bitMap),
6931 _markStack(markStack),
6932 _finger(finger),
6933 _parent(parent)
6934 { }
6935
6936 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6937 MemRegion span,
6938 CMSBitMap* bit_map,
6939 OopTaskQueue* work_queue,
6940 CMSMarkStack* overflow_stack,
6941 HeapWord* finger,
6942 HeapWord** global_finger_addr,
6943 Par_MarkFromRootsClosure* parent) :
6944 MetadataAwareOopClosure(collector->ref_processor()),
6945 _collector(collector),
6946 _whole_span(collector->_span),
6947 _span(span),
6948 _bit_map(bit_map),
6949 _work_queue(work_queue),
6950 _overflow_stack(overflow_stack),
6951 _finger(finger),
6952 _global_finger_addr(global_finger_addr),
6953 _parent(parent)
6954 { }
6955
6956 // Assumes thread-safe access by callers, who are
6957 // responsible for mutual exclusion.
6958 void CMSCollector::lower_restart_addr(HeapWord* low) {
6959 assert(_span.contains(low), "Out of bounds addr");
6960 if (_restart_addr == NULL) {
6961 _restart_addr = low;
6962 } else {
6963 _restart_addr = MIN2(_restart_addr, low);
6964 }
6965 }
6966
6967 // Upon stack overflow, we discard (part of) the stack,
6968 // remembering the least address amongst those discarded
6969 // in CMSCollector's _restart_address.
6970 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6971 // Remember the least grey address discarded
6972 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6973 _collector->lower_restart_addr(ra);
6974 _markStack->reset(); // discard stack contents
6975 _markStack->expand(); // expand the stack if possible
6976 }
6977
6978 // Upon stack overflow, we discard (part of) the stack,
6979 // remembering the least address amongst those discarded
6980 // in CMSCollector's _restart_address.
6981 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6982 // We need to do this under a mutex to prevent other
6983 // workers from interfering with the work done below.
6984 MutexLockerEx ml(_overflow_stack->par_lock(),
6985 Mutex::_no_safepoint_check_flag);
6986 // Remember the least grey address discarded
6987 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6988 _collector->lower_restart_addr(ra);
6989 _overflow_stack->reset(); // discard stack contents
6990 _overflow_stack->expand(); // expand the stack if possible
6991 }
6992
6993 void PushOrMarkClosure::do_oop(oop obj) {
6994 // Ignore mark word because we are running concurrent with mutators.
6995 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6996 HeapWord* addr = (HeapWord*)obj;
6997 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6998 // Oop lies in _span and isn't yet grey or black
6999 _bitMap->mark(addr); // now grey
7000 if (addr < _finger) {
7001 // the bit map iteration has already either passed, or
7002 // sampled, this bit in the bit map; we'll need to
7003 // use the marking stack to scan this oop's oops.
7004 bool simulate_overflow = false;
7005 NOT_PRODUCT(
7006 if (CMSMarkStackOverflowALot &&
7007 _collector->simulate_overflow()) {
7008 // simulate a stack overflow
7009 simulate_overflow = true;
7010 }
7011 )
7012 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7013 if (PrintCMSStatistics != 0) {
7014 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7015 SIZE_FORMAT, _markStack->capacity());
7016 }
7017 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7018 handle_stack_overflow(addr);
7019 }
7020 }
7021 // anything including and to the right of _finger
7022 // will be scanned as we iterate over the remainder of the
7023 // bit map
7024 do_yield_check();
7025 }
7026 }
7027
7028 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7029 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7030
7031 void Par_PushOrMarkClosure::do_oop(oop obj) {
7032 // Ignore mark word because we are running concurrent with mutators.
7033 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7034 HeapWord* addr = (HeapWord*)obj;
7035 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7036 // Oop lies in _span and isn't yet grey or black
7037 // We read the global_finger (volatile read) strictly after marking oop
7038 bool res = _bit_map->par_mark(addr); // now grey
7039 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7040 // Should we push this marked oop on our stack?
7041 // -- if someone else marked it, nothing to do
7042 // -- if target oop is above global finger nothing to do
7043 // -- if target oop is in chunk and above local finger
7044 // then nothing to do
7045 // -- else push on work queue
7046 if ( !res // someone else marked it, they will deal with it
7047 || (addr >= *gfa) // will be scanned in a later task
7048 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7049 return;
7050 }
7051 // the bit map iteration has already either passed, or
7052 // sampled, this bit in the bit map; we'll need to
7053 // use the marking stack to scan this oop's oops.
7054 bool simulate_overflow = false;
7055 NOT_PRODUCT(
7056 if (CMSMarkStackOverflowALot &&
7057 _collector->simulate_overflow()) {
7058 // simulate a stack overflow
7059 simulate_overflow = true;
7060 }
7061 )
7062 if (simulate_overflow ||
7063 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7064 // stack overflow
7065 if (PrintCMSStatistics != 0) {
7066 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7067 SIZE_FORMAT, _overflow_stack->capacity());
7068 }
7069 // We cannot assert that the overflow stack is full because
7070 // it may have been emptied since.
7071 assert(simulate_overflow ||
7072 _work_queue->size() == _work_queue->max_elems(),
7073 "Else push should have succeeded");
7074 handle_stack_overflow(addr);
7075 }
7076 do_yield_check();
7077 }
7078 }
7079
7080 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7081 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7082
7083 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7084 MemRegion span,
7085 ReferenceProcessor* rp,
7086 CMSBitMap* bit_map,
7087 CMSBitMap* mod_union_table,
7088 CMSMarkStack* mark_stack,
7089 bool concurrent_precleaning):
7090 MetadataAwareOopClosure(rp),
7091 _collector(collector),
7092 _span(span),
7093 _bit_map(bit_map),
7094 _mod_union_table(mod_union_table),
7095 _mark_stack(mark_stack),
7096 _concurrent_precleaning(concurrent_precleaning)
7097 {
7098 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7099 }
7100
7101 // Grey object rescan during pre-cleaning and second checkpoint phases --
7102 // the non-parallel version (the parallel version appears further below.)
7103 void PushAndMarkClosure::do_oop(oop obj) {
7104 // Ignore mark word verification. If during concurrent precleaning,
7105 // the object monitor may be locked. If during the checkpoint
7106 // phases, the object may already have been reached by a different
7107 // path and may be at the end of the global overflow list (so
7108 // the mark word may be NULL).
7109 assert(obj->is_oop_or_null(true /* ignore mark word */),
7110 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7111 HeapWord* addr = (HeapWord*)obj;
7112 // Check if oop points into the CMS generation
7113 // and is not marked
7114 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7115 // a white object ...
7116 _bit_map->mark(addr); // ... now grey
7117 // push on the marking stack (grey set)
7118 bool simulate_overflow = false;
7119 NOT_PRODUCT(
7120 if (CMSMarkStackOverflowALot &&
7121 _collector->simulate_overflow()) {
7122 // simulate a stack overflow
7123 simulate_overflow = true;
7124 }
7125 )
7126 if (simulate_overflow || !_mark_stack->push(obj)) {
7127 if (_concurrent_precleaning) {
7128 // During precleaning we can just dirty the appropriate card(s)
7129 // in the mod union table, thus ensuring that the object remains
7130 // in the grey set and continue. In the case of object arrays
7131 // we need to dirty all of the cards that the object spans,
7132 // since the rescan of object arrays will be limited to the
7133 // dirty cards.
7134 // Note that no one can be interfering with us in this action
7135 // of dirtying the mod union table, so no locking or atomics
7136 // are required.
7137 if (obj->is_objArray()) {
7138 size_t sz = obj->size();
7139 HeapWord* end_card_addr = (HeapWord*)round_to(
7140 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7141 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7142 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7143 _mod_union_table->mark_range(redirty_range);
7144 } else {
7145 _mod_union_table->mark(addr);
7146 }
7147 _collector->_ser_pmc_preclean_ovflw++;
7148 } else {
7149 // During the remark phase, we need to remember this oop
7150 // in the overflow list.
7151 _collector->push_on_overflow_list(obj);
7152 _collector->_ser_pmc_remark_ovflw++;
7153 }
7154 }
7155 }
7156 }
7157
7158 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7159 MemRegion span,
7160 ReferenceProcessor* rp,
7161 CMSBitMap* bit_map,
7162 OopTaskQueue* work_queue):
7163 MetadataAwareOopClosure(rp),
7164 _collector(collector),
7165 _span(span),
7166 _bit_map(bit_map),
7167 _work_queue(work_queue)
7168 {
7169 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7170 }
7171
7172 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7173 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7174
7175 // Grey object rescan during second checkpoint phase --
7176 // the parallel version.
7177 void Par_PushAndMarkClosure::do_oop(oop obj) {
7178 // In the assert below, we ignore the mark word because
7179 // this oop may point to an already visited object that is
7180 // on the overflow stack (in which case the mark word has
7181 // been hijacked for chaining into the overflow stack --
7182 // if this is the last object in the overflow stack then
7183 // its mark word will be NULL). Because this object may
7184 // have been subsequently popped off the global overflow
7185 // stack, and the mark word possibly restored to the prototypical
7186 // value, by the time we get to examined this failing assert in
7187 // the debugger, is_oop_or_null(false) may subsequently start
7188 // to hold.
7189 assert(obj->is_oop_or_null(true),
7190 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7191 HeapWord* addr = (HeapWord*)obj;
7192 // Check if oop points into the CMS generation
7193 // and is not marked
7194 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7195 // a white object ...
7196 // If we manage to "claim" the object, by being the
7197 // first thread to mark it, then we push it on our
7198 // marking stack
7199 if (_bit_map->par_mark(addr)) { // ... now grey
7200 // push on work queue (grey set)
7201 bool simulate_overflow = false;
7202 NOT_PRODUCT(
7203 if (CMSMarkStackOverflowALot &&
7204 _collector->par_simulate_overflow()) {
7205 // simulate a stack overflow
7206 simulate_overflow = true;
7207 }
7208 )
7209 if (simulate_overflow || !_work_queue->push(obj)) {
7210 _collector->par_push_on_overflow_list(obj);
7211 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7212 }
7213 } // Else, some other thread got there first
7214 }
7215 }
7216
7217 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7218 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7219
7220 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7221 Mutex* bml = _collector->bitMapLock();
7222 assert_lock_strong(bml);
7223 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7224 "CMS thread should hold CMS token");
7225
7226 bml->unlock();
7227 ConcurrentMarkSweepThread::desynchronize(true);
7228
7229 _collector->stopTimer();
7230 if (PrintCMSStatistics != 0) {
7231 _collector->incrementYields();
7232 }
7233
7234 // See the comment in coordinator_yield()
7235 for (unsigned i = 0; i < CMSYieldSleepCount &&
7236 ConcurrentMarkSweepThread::should_yield() &&
7237 !CMSCollector::foregroundGCIsActive(); ++i) {
7238 os::sleep(Thread::current(), 1, false);
7239 }
7240
7241 ConcurrentMarkSweepThread::synchronize(true);
7242 bml->lock();
7243
7244 _collector->startTimer();
7245 }
7246
7247 bool CMSPrecleanRefsYieldClosure::should_return() {
7248 if (ConcurrentMarkSweepThread::should_yield()) {
7249 do_yield_work();
7250 }
7251 return _collector->foregroundGCIsActive();
7252 }
7253
7254 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7255 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7256 "mr should be aligned to start at a card boundary");
7257 // We'd like to assert:
7258 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7259 // "mr should be a range of cards");
7260 // However, that would be too strong in one case -- the last
7261 // partition ends at _unallocated_block which, in general, can be
7262 // an arbitrary boundary, not necessarily card aligned.
7263 if (PrintCMSStatistics != 0) {
7264 _num_dirty_cards +=
7265 mr.word_size()/CardTableModRefBS::card_size_in_words;
7266 }
7267 _space->object_iterate_mem(mr, &_scan_cl);
7268 }
7269
7270 SweepClosure::SweepClosure(CMSCollector* collector,
7271 ConcurrentMarkSweepGeneration* g,
7272 CMSBitMap* bitMap, bool should_yield) :
7273 _collector(collector),
7274 _g(g),
7275 _sp(g->cmsSpace()),
7276 _limit(_sp->sweep_limit()),
7277 _freelistLock(_sp->freelistLock()),
7278 _bitMap(bitMap),
7279 _yield(should_yield),
7280 _inFreeRange(false), // No free range at beginning of sweep
7281 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7282 _lastFreeRangeCoalesced(false),
7283 _freeFinger(g->used_region().start())
7284 {
7285 NOT_PRODUCT(
7286 _numObjectsFreed = 0;
7287 _numWordsFreed = 0;
7288 _numObjectsLive = 0;
7289 _numWordsLive = 0;
7290 _numObjectsAlreadyFree = 0;
7291 _numWordsAlreadyFree = 0;
7292 _last_fc = NULL;
7293
7294 _sp->initializeIndexedFreeListArrayReturnedBytes();
7295 _sp->dictionary()->initialize_dict_returned_bytes();
7296 )
7297 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7298 "sweep _limit out of bounds");
7299 if (CMSTraceSweeper) {
7300 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7301 p2i(_limit));
7302 }
7303 }
7304
7305 void SweepClosure::print_on(outputStream* st) const {
7306 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7307 p2i(_sp->bottom()), p2i(_sp->end()));
7308 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7309 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7310 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7311 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7312 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7313 }
7314
7315 #ifndef PRODUCT
7316 // Assertion checking only: no useful work in product mode --
7317 // however, if any of the flags below become product flags,
7318 // you may need to review this code to see if it needs to be
7319 // enabled in product mode.
7320 SweepClosure::~SweepClosure() {
7321 assert_lock_strong(_freelistLock);
7322 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7323 "sweep _limit out of bounds");
7324 if (inFreeRange()) {
7325 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7326 print();
7327 ShouldNotReachHere();
7328 }
7329 if (Verbose && PrintGC) {
7330 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7331 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7332 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, "
7333 SIZE_FORMAT " bytes "
7334 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7335 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7336 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7337 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7338 * sizeof(HeapWord);
7339 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7340
7341 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7342 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7343 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7344 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7345 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes);
7346 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes",
7347 indexListReturnedBytes);
7348 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes",
7349 dict_returned_bytes);
7350 }
7351 }
7352 if (CMSTraceSweeper) {
7353 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7354 p2i(_limit));
7355 }
7356 }
7357 #endif // PRODUCT
7358
7359 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7360 bool freeRangeInFreeLists) {
7361 if (CMSTraceSweeper) {
7362 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7363 p2i(freeFinger), freeRangeInFreeLists);
7364 }
7365 assert(!inFreeRange(), "Trampling existing free range");
7366 set_inFreeRange(true);
7367 set_lastFreeRangeCoalesced(false);
7368
7369 set_freeFinger(freeFinger);
7370 set_freeRangeInFreeLists(freeRangeInFreeLists);
7371 if (CMSTestInFreeList) {
7372 if (freeRangeInFreeLists) {
7373 FreeChunk* fc = (FreeChunk*) freeFinger;
7374 assert(fc->is_free(), "A chunk on the free list should be free.");
7375 assert(fc->size() > 0, "Free range should have a size");
7376 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7377 }
7378 }
7379 }
7380
7381 // Note that the sweeper runs concurrently with mutators. Thus,
7382 // it is possible for direct allocation in this generation to happen
7383 // in the middle of the sweep. Note that the sweeper also coalesces
7384 // contiguous free blocks. Thus, unless the sweeper and the allocator
7385 // synchronize appropriately freshly allocated blocks may get swept up.
7386 // This is accomplished by the sweeper locking the free lists while
7387 // it is sweeping. Thus blocks that are determined to be free are
7388 // indeed free. There is however one additional complication:
7389 // blocks that have been allocated since the final checkpoint and
7390 // mark, will not have been marked and so would be treated as
7391 // unreachable and swept up. To prevent this, the allocator marks
7392 // the bit map when allocating during the sweep phase. This leads,
7393 // however, to a further complication -- objects may have been allocated
7394 // but not yet initialized -- in the sense that the header isn't yet
7395 // installed. The sweeper can not then determine the size of the block
7396 // in order to skip over it. To deal with this case, we use a technique
7397 // (due to Printezis) to encode such uninitialized block sizes in the
7398 // bit map. Since the bit map uses a bit per every HeapWord, but the
7399 // CMS generation has a minimum object size of 3 HeapWords, it follows
7400 // that "normal marks" won't be adjacent in the bit map (there will
7401 // always be at least two 0 bits between successive 1 bits). We make use
7402 // of these "unused" bits to represent uninitialized blocks -- the bit
7403 // corresponding to the start of the uninitialized object and the next
7404 // bit are both set. Finally, a 1 bit marks the end of the object that
7405 // started with the two consecutive 1 bits to indicate its potentially
7406 // uninitialized state.
7407
7408 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7409 FreeChunk* fc = (FreeChunk*)addr;
7410 size_t res;
7411
7412 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7413 // than "addr == _limit" because although _limit was a block boundary when
7414 // we started the sweep, it may no longer be one because heap expansion
7415 // may have caused us to coalesce the block ending at the address _limit
7416 // with a newly expanded chunk (this happens when _limit was set to the
7417 // previous _end of the space), so we may have stepped past _limit:
7418 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7419 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7420 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7421 "sweep _limit out of bounds");
7422 assert(addr < _sp->end(), "addr out of bounds");
7423 // Flush any free range we might be holding as a single
7424 // coalesced chunk to the appropriate free list.
7425 if (inFreeRange()) {
7426 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7427 err_msg("freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())));
7428 flush_cur_free_chunk(freeFinger(),
7429 pointer_delta(addr, freeFinger()));
7430 if (CMSTraceSweeper) {
7431 gclog_or_tty->print("Sweep: last chunk: ");
7432 gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") "
7433 "[coalesced:%d]\n",
7434 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7435 lastFreeRangeCoalesced() ? 1 : 0);
7436 }
7437 }
7438
7439 // help the iterator loop finish
7440 return pointer_delta(_sp->end(), addr);
7441 }
7442
7443 assert(addr < _limit, "sweep invariant");
7444 // check if we should yield
7445 do_yield_check(addr);
7446 if (fc->is_free()) {
7447 // Chunk that is already free
7448 res = fc->size();
7449 do_already_free_chunk(fc);
7450 debug_only(_sp->verifyFreeLists());
7451 // If we flush the chunk at hand in lookahead_and_flush()
7452 // and it's coalesced with a preceding chunk, then the
7453 // process of "mangling" the payload of the coalesced block
7454 // will cause erasure of the size information from the
7455 // (erstwhile) header of all the coalesced blocks but the
7456 // first, so the first disjunct in the assert will not hold
7457 // in that specific case (in which case the second disjunct
7458 // will hold).
7459 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7460 "Otherwise the size info doesn't change at this step");
7461 NOT_PRODUCT(
7462 _numObjectsAlreadyFree++;
7463 _numWordsAlreadyFree += res;
7464 )
7465 NOT_PRODUCT(_last_fc = fc;)
7466 } else if (!_bitMap->isMarked(addr)) {
7467 // Chunk is fresh garbage
7468 res = do_garbage_chunk(fc);
7469 debug_only(_sp->verifyFreeLists());
7470 NOT_PRODUCT(
7471 _numObjectsFreed++;
7472 _numWordsFreed += res;
7473 )
7474 } else {
7475 // Chunk that is alive.
7476 res = do_live_chunk(fc);
7477 debug_only(_sp->verifyFreeLists());
7478 NOT_PRODUCT(
7479 _numObjectsLive++;
7480 _numWordsLive += res;
7481 )
7482 }
7483 return res;
7484 }
7485
7486 // For the smart allocation, record following
7487 // split deaths - a free chunk is removed from its free list because
7488 // it is being split into two or more chunks.
7489 // split birth - a free chunk is being added to its free list because
7490 // a larger free chunk has been split and resulted in this free chunk.
7491 // coal death - a free chunk is being removed from its free list because
7492 // it is being coalesced into a large free chunk.
7493 // coal birth - a free chunk is being added to its free list because
7494 // it was created when two or more free chunks where coalesced into
7495 // this free chunk.
7496 //
7497 // These statistics are used to determine the desired number of free
7498 // chunks of a given size. The desired number is chosen to be relative
7499 // to the end of a CMS sweep. The desired number at the end of a sweep
7500 // is the
7501 // count-at-end-of-previous-sweep (an amount that was enough)
7502 // - count-at-beginning-of-current-sweep (the excess)
7503 // + split-births (gains in this size during interval)
7504 // - split-deaths (demands on this size during interval)
7505 // where the interval is from the end of one sweep to the end of the
7506 // next.
7507 //
7508 // When sweeping the sweeper maintains an accumulated chunk which is
7509 // the chunk that is made up of chunks that have been coalesced. That
7510 // will be termed the left-hand chunk. A new chunk of garbage that
7511 // is being considered for coalescing will be referred to as the
7512 // right-hand chunk.
7513 //
7514 // When making a decision on whether to coalesce a right-hand chunk with
7515 // the current left-hand chunk, the current count vs. the desired count
7516 // of the left-hand chunk is considered. Also if the right-hand chunk
7517 // is near the large chunk at the end of the heap (see
7518 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7519 // left-hand chunk is coalesced.
7520 //
7521 // When making a decision about whether to split a chunk, the desired count
7522 // vs. the current count of the candidate to be split is also considered.
7523 // If the candidate is underpopulated (currently fewer chunks than desired)
7524 // a chunk of an overpopulated (currently more chunks than desired) size may
7525 // be chosen. The "hint" associated with a free list, if non-null, points
7526 // to a free list which may be overpopulated.
7527 //
7528
7529 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7530 const size_t size = fc->size();
7531 // Chunks that cannot be coalesced are not in the
7532 // free lists.
7533 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7534 assert(_sp->verify_chunk_in_free_list(fc),
7535 "free chunk should be in free lists");
7536 }
7537 // a chunk that is already free, should not have been
7538 // marked in the bit map
7539 HeapWord* const addr = (HeapWord*) fc;
7540 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7541 // Verify that the bit map has no bits marked between
7542 // addr and purported end of this block.
7543 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7544
7545 // Some chunks cannot be coalesced under any circumstances.
7546 // See the definition of cantCoalesce().
7547 if (!fc->cantCoalesce()) {
7548 // This chunk can potentially be coalesced.
7549 if (_sp->adaptive_freelists()) {
7550 // All the work is done in
7551 do_post_free_or_garbage_chunk(fc, size);
7552 } else { // Not adaptive free lists
7553 // this is a free chunk that can potentially be coalesced by the sweeper;
7554 if (!inFreeRange()) {
7555 // if the next chunk is a free block that can't be coalesced
7556 // it doesn't make sense to remove this chunk from the free lists
7557 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7558 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7559 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
7560 nextChunk->is_free() && // ... which is free...
7561 nextChunk->cantCoalesce()) { // ... but can't be coalesced
7562 // nothing to do
7563 } else {
7564 // Potentially the start of a new free range:
7565 // Don't eagerly remove it from the free lists.
7566 // No need to remove it if it will just be put
7567 // back again. (Also from a pragmatic point of view
7568 // if it is a free block in a region that is beyond
7569 // any allocated blocks, an assertion will fail)
7570 // Remember the start of a free run.
7571 initialize_free_range(addr, true);
7572 // end - can coalesce with next chunk
7573 }
7574 } else {
7575 // the midst of a free range, we are coalescing
7576 print_free_block_coalesced(fc);
7577 if (CMSTraceSweeper) {
7578 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7579 }
7580 // remove it from the free lists
7581 _sp->removeFreeChunkFromFreeLists(fc);
7582 set_lastFreeRangeCoalesced(true);
7583 // If the chunk is being coalesced and the current free range is
7584 // in the free lists, remove the current free range so that it
7585 // will be returned to the free lists in its entirety - all
7586 // the coalesced pieces included.
7587 if (freeRangeInFreeLists()) {
7588 FreeChunk* ffc = (FreeChunk*) freeFinger();
7589 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7590 "Size of free range is inconsistent with chunk size.");
7591 if (CMSTestInFreeList) {
7592 assert(_sp->verify_chunk_in_free_list(ffc),
7593 "free range is not in free lists");
7594 }
7595 _sp->removeFreeChunkFromFreeLists(ffc);
7596 set_freeRangeInFreeLists(false);
7597 }
7598 }
7599 }
7600 // Note that if the chunk is not coalescable (the else arm
7601 // below), we unconditionally flush, without needing to do
7602 // a "lookahead," as we do below.
7603 if (inFreeRange()) lookahead_and_flush(fc, size);
7604 } else {
7605 // Code path common to both original and adaptive free lists.
7606
7607 // cant coalesce with previous block; this should be treated
7608 // as the end of a free run if any
7609 if (inFreeRange()) {
7610 // we kicked some butt; time to pick up the garbage
7611 assert(freeFinger() < addr, "freeFinger points too high");
7612 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7613 }
7614 // else, nothing to do, just continue
7615 }
7616 }
7617
7618 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7619 // This is a chunk of garbage. It is not in any free list.
7620 // Add it to a free list or let it possibly be coalesced into
7621 // a larger chunk.
7622 HeapWord* const addr = (HeapWord*) fc;
7623 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7624
7625 if (_sp->adaptive_freelists()) {
7626 // Verify that the bit map has no bits marked between
7627 // addr and purported end of just dead object.
7628 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7629
7630 do_post_free_or_garbage_chunk(fc, size);
7631 } else {
7632 if (!inFreeRange()) {
7633 // start of a new free range
7634 assert(size > 0, "A free range should have a size");
7635 initialize_free_range(addr, false);
7636 } else {
7637 // this will be swept up when we hit the end of the
7638 // free range
7639 if (CMSTraceSweeper) {
7640 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7641 }
7642 // If the chunk is being coalesced and the current free range is
7643 // in the free lists, remove the current free range so that it
7644 // will be returned to the free lists in its entirety - all
7645 // the coalesced pieces included.
7646 if (freeRangeInFreeLists()) {
7647 FreeChunk* ffc = (FreeChunk*)freeFinger();
7648 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7649 "Size of free range is inconsistent with chunk size.");
7650 if (CMSTestInFreeList) {
7651 assert(_sp->verify_chunk_in_free_list(ffc),
7652 "free range is not in free lists");
7653 }
7654 _sp->removeFreeChunkFromFreeLists(ffc);
7655 set_freeRangeInFreeLists(false);
7656 }
7657 set_lastFreeRangeCoalesced(true);
7658 }
7659 // this will be swept up when we hit the end of the free range
7660
7661 // Verify that the bit map has no bits marked between
7662 // addr and purported end of just dead object.
7663 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7664 }
7665 assert(_limit >= addr + size,
7666 "A freshly garbage chunk can't possibly straddle over _limit");
7667 if (inFreeRange()) lookahead_and_flush(fc, size);
7668 return size;
7669 }
7670
7671 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7672 HeapWord* addr = (HeapWord*) fc;
7673 // The sweeper has just found a live object. Return any accumulated
7674 // left hand chunk to the free lists.
7675 if (inFreeRange()) {
7676 assert(freeFinger() < addr, "freeFinger points too high");
7677 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7678 }
7679
7680 // This object is live: we'd normally expect this to be
7681 // an oop, and like to assert the following:
7682 // assert(oop(addr)->is_oop(), "live block should be an oop");
7683 // However, as we commented above, this may be an object whose
7684 // header hasn't yet been initialized.
7685 size_t size;
7686 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7687 if (_bitMap->isMarked(addr + 1)) {
7688 // Determine the size from the bit map, rather than trying to
7689 // compute it from the object header.
7690 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7691 size = pointer_delta(nextOneAddr + 1, addr);
7692 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7693 "alignment problem");
7694
7695 #ifdef ASSERT
7696 if (oop(addr)->klass_or_null() != NULL) {
7697 // Ignore mark word because we are running concurrent with mutators
7698 assert(oop(addr)->is_oop(true), "live block should be an oop");
7699 assert(size ==
7700 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7701 "P-mark and computed size do not agree");
7702 }
7703 #endif
7704
7705 } else {
7706 // This should be an initialized object that's alive.
7707 assert(oop(addr)->klass_or_null() != NULL,
7708 "Should be an initialized object");
7709 // Ignore mark word because we are running concurrent with mutators
7710 assert(oop(addr)->is_oop(true), "live block should be an oop");
7711 // Verify that the bit map has no bits marked between
7712 // addr and purported end of this block.
7713 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7714 assert(size >= 3, "Necessary for Printezis marks to work");
7715 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7716 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7717 }
7718 return size;
7719 }
7720
7721 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7722 size_t chunkSize) {
7723 // do_post_free_or_garbage_chunk() should only be called in the case
7724 // of the adaptive free list allocator.
7725 const bool fcInFreeLists = fc->is_free();
7726 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7727 assert((HeapWord*)fc <= _limit, "sweep invariant");
7728 if (CMSTestInFreeList && fcInFreeLists) {
7729 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7730 }
7731
7732 if (CMSTraceSweeper) {
7733 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7734 }
7735
7736 HeapWord* const fc_addr = (HeapWord*) fc;
7737
7738 bool coalesce;
7739 const size_t left = pointer_delta(fc_addr, freeFinger());
7740 const size_t right = chunkSize;
7741 switch (FLSCoalescePolicy) {
7742 // numeric value forms a coalition aggressiveness metric
7743 case 0: { // never coalesce
7744 coalesce = false;
7745 break;
7746 }
7747 case 1: { // coalesce if left & right chunks on overpopulated lists
7748 coalesce = _sp->coalOverPopulated(left) &&
7749 _sp->coalOverPopulated(right);
7750 break;
7751 }
7752 case 2: { // coalesce if left chunk on overpopulated list (default)
7753 coalesce = _sp->coalOverPopulated(left);
7754 break;
7755 }
7756 case 3: { // coalesce if left OR right chunk on overpopulated list
7757 coalesce = _sp->coalOverPopulated(left) ||
7758 _sp->coalOverPopulated(right);
7759 break;
7760 }
7761 case 4: { // always coalesce
7762 coalesce = true;
7763 break;
7764 }
7765 default:
7766 ShouldNotReachHere();
7767 }
7768
7769 // Should the current free range be coalesced?
7770 // If the chunk is in a free range and either we decided to coalesce above
7771 // or the chunk is near the large block at the end of the heap
7772 // (isNearLargestChunk() returns true), then coalesce this chunk.
7773 const bool doCoalesce = inFreeRange()
7774 && (coalesce || _g->isNearLargestChunk(fc_addr));
7775 if (doCoalesce) {
7776 // Coalesce the current free range on the left with the new
7777 // chunk on the right. If either is on a free list,
7778 // it must be removed from the list and stashed in the closure.
7779 if (freeRangeInFreeLists()) {
7780 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7781 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7782 "Size of free range is inconsistent with chunk size.");
7783 if (CMSTestInFreeList) {
7784 assert(_sp->verify_chunk_in_free_list(ffc),
7785 "Chunk is not in free lists");
7786 }
7787 _sp->coalDeath(ffc->size());
7788 _sp->removeFreeChunkFromFreeLists(ffc);
7789 set_freeRangeInFreeLists(false);
7790 }
7791 if (fcInFreeLists) {
7792 _sp->coalDeath(chunkSize);
7793 assert(fc->size() == chunkSize,
7794 "The chunk has the wrong size or is not in the free lists");
7795 _sp->removeFreeChunkFromFreeLists(fc);
7796 }
7797 set_lastFreeRangeCoalesced(true);
7798 print_free_block_coalesced(fc);
7799 } else { // not in a free range and/or should not coalesce
7800 // Return the current free range and start a new one.
7801 if (inFreeRange()) {
7802 // In a free range but cannot coalesce with the right hand chunk.
7803 // Put the current free range into the free lists.
7804 flush_cur_free_chunk(freeFinger(),
7805 pointer_delta(fc_addr, freeFinger()));
7806 }
7807 // Set up for new free range. Pass along whether the right hand
7808 // chunk is in the free lists.
7809 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7810 }
7811 }
7812
7813 // Lookahead flush:
7814 // If we are tracking a free range, and this is the last chunk that
7815 // we'll look at because its end crosses past _limit, we'll preemptively
7816 // flush it along with any free range we may be holding on to. Note that
7817 // this can be the case only for an already free or freshly garbage
7818 // chunk. If this block is an object, it can never straddle
7819 // over _limit. The "straddling" occurs when _limit is set at
7820 // the previous end of the space when this cycle started, and
7821 // a subsequent heap expansion caused the previously co-terminal
7822 // free block to be coalesced with the newly expanded portion,
7823 // thus rendering _limit a non-block-boundary making it dangerous
7824 // for the sweeper to step over and examine.
7825 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7826 assert(inFreeRange(), "Should only be called if currently in a free range.");
7827 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7828 assert(_sp->used_region().contains(eob - 1),
7829 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7830 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7831 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7832 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size));
7833 if (eob >= _limit) {
7834 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7835 if (CMSTraceSweeper) {
7836 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7837 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7838 "[" PTR_FORMAT "," PTR_FORMAT ")",
7839 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7840 }
7841 // Return the storage we are tracking back into the free lists.
7842 if (CMSTraceSweeper) {
7843 gclog_or_tty->print_cr("Flushing ... ");
7844 }
7845 assert(freeFinger() < eob, "Error");
7846 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7847 }
7848 }
7849
7850 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7851 assert(inFreeRange(), "Should only be called if currently in a free range.");
7852 assert(size > 0,
7853 "A zero sized chunk cannot be added to the free lists.");
7854 if (!freeRangeInFreeLists()) {
7855 if (CMSTestInFreeList) {
7856 FreeChunk* fc = (FreeChunk*) chunk;
7857 fc->set_size(size);
7858 assert(!_sp->verify_chunk_in_free_list(fc),
7859 "chunk should not be in free lists yet");
7860 }
7861 if (CMSTraceSweeper) {
7862 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7863 p2i(chunk), size);
7864 }
7865 // A new free range is going to be starting. The current
7866 // free range has not been added to the free lists yet or
7867 // was removed so add it back.
7868 // If the current free range was coalesced, then the death
7869 // of the free range was recorded. Record a birth now.
7870 if (lastFreeRangeCoalesced()) {
7871 _sp->coalBirth(size);
7872 }
7873 _sp->addChunkAndRepairOffsetTable(chunk, size,
7874 lastFreeRangeCoalesced());
7875 } else if (CMSTraceSweeper) {
7876 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7877 }
7878 set_inFreeRange(false);
7879 set_freeRangeInFreeLists(false);
7880 }
7881
7882 // We take a break if we've been at this for a while,
7883 // so as to avoid monopolizing the locks involved.
7884 void SweepClosure::do_yield_work(HeapWord* addr) {
7885 // Return current free chunk being used for coalescing (if any)
7886 // to the appropriate freelist. After yielding, the next
7887 // free block encountered will start a coalescing range of
7888 // free blocks. If the next free block is adjacent to the
7889 // chunk just flushed, they will need to wait for the next
7890 // sweep to be coalesced.
7891 if (inFreeRange()) {
7892 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7893 }
7894
7895 // First give up the locks, then yield, then re-lock.
7896 // We should probably use a constructor/destructor idiom to
7897 // do this unlock/lock or modify the MutexUnlocker class to
7898 // serve our purpose. XXX
7899 assert_lock_strong(_bitMap->lock());
7900 assert_lock_strong(_freelistLock);
7901 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7902 "CMS thread should hold CMS token");
7903 _bitMap->lock()->unlock();
7904 _freelistLock->unlock();
7905 ConcurrentMarkSweepThread::desynchronize(true);
7906 _collector->stopTimer();
7907 if (PrintCMSStatistics != 0) {
7908 _collector->incrementYields();
7909 }
7910
7911 // See the comment in coordinator_yield()
7912 for (unsigned i = 0; i < CMSYieldSleepCount &&
7913 ConcurrentMarkSweepThread::should_yield() &&
7914 !CMSCollector::foregroundGCIsActive(); ++i) {
7915 os::sleep(Thread::current(), 1, false);
7916 }
7917
7918 ConcurrentMarkSweepThread::synchronize(true);
7919 _freelistLock->lock();
7920 _bitMap->lock()->lock_without_safepoint_check();
7921 _collector->startTimer();
7922 }
7923
7924 #ifndef PRODUCT
7925 // This is actually very useful in a product build if it can
7926 // be called from the debugger. Compile it into the product
7927 // as needed.
7928 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7929 return debug_cms_space->verify_chunk_in_free_list(fc);
7930 }
7931 #endif
7932
7933 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7934 if (CMSTraceSweeper) {
7935 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7936 p2i(fc), fc->size());
7937 }
7938 }
7939
7940 // CMSIsAliveClosure
7941 bool CMSIsAliveClosure::do_object_b(oop obj) {
7942 HeapWord* addr = (HeapWord*)obj;
7943 return addr != NULL &&
7944 (!_span.contains(addr) || _bit_map->isMarked(addr));
7945 }
7946
7947
7948 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7949 MemRegion span,
7950 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7951 bool cpc):
7952 _collector(collector),
7953 _span(span),
7954 _bit_map(bit_map),
7955 _mark_stack(mark_stack),
7956 _concurrent_precleaning(cpc) {
7957 assert(!_span.is_empty(), "Empty span could spell trouble");
7958 }
7959
7960
7961 // CMSKeepAliveClosure: the serial version
7962 void CMSKeepAliveClosure::do_oop(oop obj) {
7963 HeapWord* addr = (HeapWord*)obj;
7964 if (_span.contains(addr) &&
7965 !_bit_map->isMarked(addr)) {
7966 _bit_map->mark(addr);
7967 bool simulate_overflow = false;
7968 NOT_PRODUCT(
7969 if (CMSMarkStackOverflowALot &&
7970 _collector->simulate_overflow()) {
7971 // simulate a stack overflow
7972 simulate_overflow = true;
7973 }
7974 )
7975 if (simulate_overflow || !_mark_stack->push(obj)) {
7976 if (_concurrent_precleaning) {
7977 // We dirty the overflown object and let the remark
7978 // phase deal with it.
7979 assert(_collector->overflow_list_is_empty(), "Error");
7980 // In the case of object arrays, we need to dirty all of
7981 // the cards that the object spans. No locking or atomics
7982 // are needed since no one else can be mutating the mod union
7983 // table.
7984 if (obj->is_objArray()) {
7985 size_t sz = obj->size();
7986 HeapWord* end_card_addr =
7987 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7988 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7989 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7990 _collector->_modUnionTable.mark_range(redirty_range);
7991 } else {
7992 _collector->_modUnionTable.mark(addr);
7993 }
7994 _collector->_ser_kac_preclean_ovflw++;
7995 } else {
7996 _collector->push_on_overflow_list(obj);
7997 _collector->_ser_kac_ovflw++;
7998 }
7999 }
8000 }
8001 }
8002
8003 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8004 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8005
8006 // CMSParKeepAliveClosure: a parallel version of the above.
8007 // The work queues are private to each closure (thread),
8008 // but (may be) available for stealing by other threads.
8009 void CMSParKeepAliveClosure::do_oop(oop obj) {
8010 HeapWord* addr = (HeapWord*)obj;
8011 if (_span.contains(addr) &&
8012 !_bit_map->isMarked(addr)) {
8013 // In general, during recursive tracing, several threads
8014 // may be concurrently getting here; the first one to
8015 // "tag" it, claims it.
8016 if (_bit_map->par_mark(addr)) {
8017 bool res = _work_queue->push(obj);
8018 assert(res, "Low water mark should be much less than capacity");
8019 // Do a recursive trim in the hope that this will keep
8020 // stack usage lower, but leave some oops for potential stealers
8021 trim_queue(_low_water_mark);
8022 } // Else, another thread got there first
8023 }
8024 }
8025
8026 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8027 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8028
8029 void CMSParKeepAliveClosure::trim_queue(uint max) {
8030 while (_work_queue->size() > max) {
8031 oop new_oop;
8032 if (_work_queue->pop_local(new_oop)) {
8033 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8034 assert(_bit_map->isMarked((HeapWord*)new_oop),
8035 "no white objects on this stack!");
8036 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8037 // iterate over the oops in this oop, marking and pushing
8038 // the ones in CMS heap (i.e. in _span).
8039 new_oop->oop_iterate(&_mark_and_push);
8040 }
8041 }
8042 }
8043
8044 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8045 CMSCollector* collector,
8046 MemRegion span, CMSBitMap* bit_map,
8047 OopTaskQueue* work_queue):
8048 _collector(collector),
8049 _span(span),
8050 _bit_map(bit_map),
8051 _work_queue(work_queue) { }
8052
8053 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8054 HeapWord* addr = (HeapWord*)obj;
8055 if (_span.contains(addr) &&
8056 !_bit_map->isMarked(addr)) {
8057 if (_bit_map->par_mark(addr)) {
8058 bool simulate_overflow = false;
8059 NOT_PRODUCT(
8060 if (CMSMarkStackOverflowALot &&
8061 _collector->par_simulate_overflow()) {
8062 // simulate a stack overflow
8063 simulate_overflow = true;
8064 }
8065 )
8066 if (simulate_overflow || !_work_queue->push(obj)) {
8067 _collector->par_push_on_overflow_list(obj);
8068 _collector->_par_kac_ovflw++;
8069 }
8070 } // Else another thread got there already
8071 }
8072 }
8073
8074 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8075 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8076
8077 //////////////////////////////////////////////////////////////////
8078 // CMSExpansionCause /////////////////////////////
8079 //////////////////////////////////////////////////////////////////
8080 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8081 switch (cause) {
8082 case _no_expansion:
8083 return "No expansion";
8084 case _satisfy_free_ratio:
8085 return "Free ratio";
8086 case _satisfy_promotion:
8087 return "Satisfy promotion";
8088 case _satisfy_allocation:
8089 return "allocation";
8090 case _allocate_par_lab:
8091 return "Par LAB";
8092 case _allocate_par_spooling_space:
8093 return "Par Spooling Space";
8094 case _adaptive_size_policy:
8095 return "Ergonomics";
8096 default:
8097 return "unknown";
8098 }
8099 }
8100
8101 void CMSDrainMarkingStackClosure::do_void() {
8102 // the max number to take from overflow list at a time
8103 const size_t num = _mark_stack->capacity()/4;
8104 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8105 "Overflow list should be NULL during concurrent phases");
8106 while (!_mark_stack->isEmpty() ||
8107 // if stack is empty, check the overflow list
8108 _collector->take_from_overflow_list(num, _mark_stack)) {
8109 oop obj = _mark_stack->pop();
8110 HeapWord* addr = (HeapWord*)obj;
8111 assert(_span.contains(addr), "Should be within span");
8112 assert(_bit_map->isMarked(addr), "Should be marked");
8113 assert(obj->is_oop(), "Should be an oop");
8114 obj->oop_iterate(_keep_alive);
8115 }
8116 }
8117
8118 void CMSParDrainMarkingStackClosure::do_void() {
8119 // drain queue
8120 trim_queue(0);
8121 }
8122
8123 // Trim our work_queue so its length is below max at return
8124 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8125 while (_work_queue->size() > max) {
8126 oop new_oop;
8127 if (_work_queue->pop_local(new_oop)) {
8128 assert(new_oop->is_oop(), "Expected an oop");
8129 assert(_bit_map->isMarked((HeapWord*)new_oop),
8130 "no white objects on this stack!");
8131 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8132 // iterate over the oops in this oop, marking and pushing
8133 // the ones in CMS heap (i.e. in _span).
8134 new_oop->oop_iterate(&_mark_and_push);
8135 }
8136 }
8137 }
8138
8139 ////////////////////////////////////////////////////////////////////
8140 // Support for Marking Stack Overflow list handling and related code
8141 ////////////////////////////////////////////////////////////////////
8142 // Much of the following code is similar in shape and spirit to the
8143 // code used in ParNewGC. We should try and share that code
8144 // as much as possible in the future.
8145
8146 #ifndef PRODUCT
8147 // Debugging support for CMSStackOverflowALot
8148
8149 // It's OK to call this multi-threaded; the worst thing
8150 // that can happen is that we'll get a bunch of closely
8151 // spaced simulated overflows, but that's OK, in fact
8152 // probably good as it would exercise the overflow code
8153 // under contention.
8154 bool CMSCollector::simulate_overflow() {
8155 if (_overflow_counter-- <= 0) { // just being defensive
8156 _overflow_counter = CMSMarkStackOverflowInterval;
8157 return true;
8158 } else {
8159 return false;
8160 }
8161 }
8162
8163 bool CMSCollector::par_simulate_overflow() {
8164 return simulate_overflow();
8165 }
8166 #endif
8167
8168 // Single-threaded
8169 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8170 assert(stack->isEmpty(), "Expected precondition");
8171 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8172 size_t i = num;
8173 oop cur = _overflow_list;
8174 const markOop proto = markOopDesc::prototype();
8175 NOT_PRODUCT(ssize_t n = 0;)
8176 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8177 next = oop(cur->mark());
8178 cur->set_mark(proto); // until proven otherwise
8179 assert(cur->is_oop(), "Should be an oop");
8180 bool res = stack->push(cur);
8181 assert(res, "Bit off more than can chew?");
8182 NOT_PRODUCT(n++;)
8183 }
8184 _overflow_list = cur;
8185 #ifndef PRODUCT
8186 assert(_num_par_pushes >= n, "Too many pops?");
8187 _num_par_pushes -=n;
8188 #endif
8189 return !stack->isEmpty();
8190 }
8191
8192 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8193 // (MT-safe) Get a prefix of at most "num" from the list.
8194 // The overflow list is chained through the mark word of
8195 // each object in the list. We fetch the entire list,
8196 // break off a prefix of the right size and return the
8197 // remainder. If other threads try to take objects from
8198 // the overflow list at that time, they will wait for
8199 // some time to see if data becomes available. If (and
8200 // only if) another thread places one or more object(s)
8201 // on the global list before we have returned the suffix
8202 // to the global list, we will walk down our local list
8203 // to find its end and append the global list to
8204 // our suffix before returning it. This suffix walk can
8205 // prove to be expensive (quadratic in the amount of traffic)
8206 // when there are many objects in the overflow list and
8207 // there is much producer-consumer contention on the list.
8208 // *NOTE*: The overflow list manipulation code here and
8209 // in ParNewGeneration:: are very similar in shape,
8210 // except that in the ParNew case we use the old (from/eden)
8211 // copy of the object to thread the list via its klass word.
8212 // Because of the common code, if you make any changes in
8213 // the code below, please check the ParNew version to see if
8214 // similar changes might be needed.
8215 // CR 6797058 has been filed to consolidate the common code.
8216 bool CMSCollector::par_take_from_overflow_list(size_t num,
8217 OopTaskQueue* work_q,
8218 int no_of_gc_threads) {
8219 assert(work_q->size() == 0, "First empty local work queue");
8220 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8221 if (_overflow_list == NULL) {
8222 return false;
8223 }
8224 // Grab the entire list; we'll put back a suffix
8225 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8226 Thread* tid = Thread::current();
8227 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8228 // set to ParallelGCThreads.
8229 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8230 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8231 // If the list is busy, we spin for a short while,
8232 // sleeping between attempts to get the list.
8233 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8234 os::sleep(tid, sleep_time_millis, false);
8235 if (_overflow_list == NULL) {
8236 // Nothing left to take
8237 return false;
8238 } else if (_overflow_list != BUSY) {
8239 // Try and grab the prefix
8240 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8241 }
8242 }
8243 // If the list was found to be empty, or we spun long
8244 // enough, we give up and return empty-handed. If we leave
8245 // the list in the BUSY state below, it must be the case that
8246 // some other thread holds the overflow list and will set it
8247 // to a non-BUSY state in the future.
8248 if (prefix == NULL || prefix == BUSY) {
8249 // Nothing to take or waited long enough
8250 if (prefix == NULL) {
8251 // Write back the NULL in case we overwrote it with BUSY above
8252 // and it is still the same value.
8253 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8254 }
8255 return false;
8256 }
8257 assert(prefix != NULL && prefix != BUSY, "Error");
8258 size_t i = num;
8259 oop cur = prefix;
8260 // Walk down the first "num" objects, unless we reach the end.
8261 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8262 if (cur->mark() == NULL) {
8263 // We have "num" or fewer elements in the list, so there
8264 // is nothing to return to the global list.
8265 // Write back the NULL in lieu of the BUSY we wrote
8266 // above, if it is still the same value.
8267 if (_overflow_list == BUSY) {
8268 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8269 }
8270 } else {
8271 // Chop off the suffix and return it to the global list.
8272 assert(cur->mark() != BUSY, "Error");
8273 oop suffix_head = cur->mark(); // suffix will be put back on global list
8274 cur->set_mark(NULL); // break off suffix
8275 // It's possible that the list is still in the empty(busy) state
8276 // we left it in a short while ago; in that case we may be
8277 // able to place back the suffix without incurring the cost
8278 // of a walk down the list.
8279 oop observed_overflow_list = _overflow_list;
8280 oop cur_overflow_list = observed_overflow_list;
8281 bool attached = false;
8282 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8283 observed_overflow_list =
8284 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8285 if (cur_overflow_list == observed_overflow_list) {
8286 attached = true;
8287 break;
8288 } else cur_overflow_list = observed_overflow_list;
8289 }
8290 if (!attached) {
8291 // Too bad, someone else sneaked in (at least) an element; we'll need
8292 // to do a splice. Find tail of suffix so we can prepend suffix to global
8293 // list.
8294 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8295 oop suffix_tail = cur;
8296 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8297 "Tautology");
8298 observed_overflow_list = _overflow_list;
8299 do {
8300 cur_overflow_list = observed_overflow_list;
8301 if (cur_overflow_list != BUSY) {
8302 // Do the splice ...
8303 suffix_tail->set_mark(markOop(cur_overflow_list));
8304 } else { // cur_overflow_list == BUSY
8305 suffix_tail->set_mark(NULL);
8306 }
8307 // ... and try to place spliced list back on overflow_list ...
8308 observed_overflow_list =
8309 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8310 } while (cur_overflow_list != observed_overflow_list);
8311 // ... until we have succeeded in doing so.
8312 }
8313 }
8314
8315 // Push the prefix elements on work_q
8316 assert(prefix != NULL, "control point invariant");
8317 const markOop proto = markOopDesc::prototype();
8318 oop next;
8319 NOT_PRODUCT(ssize_t n = 0;)
8320 for (cur = prefix; cur != NULL; cur = next) {
8321 next = oop(cur->mark());
8322 cur->set_mark(proto); // until proven otherwise
8323 assert(cur->is_oop(), "Should be an oop");
8324 bool res = work_q->push(cur);
8325 assert(res, "Bit off more than we can chew?");
8326 NOT_PRODUCT(n++;)
8327 }
8328 #ifndef PRODUCT
8329 assert(_num_par_pushes >= n, "Too many pops?");
8330 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8331 #endif
8332 return true;
8333 }
8334
8335 // Single-threaded
8336 void CMSCollector::push_on_overflow_list(oop p) {
8337 NOT_PRODUCT(_num_par_pushes++;)
8338 assert(p->is_oop(), "Not an oop");
8339 preserve_mark_if_necessary(p);
8340 p->set_mark((markOop)_overflow_list);
8341 _overflow_list = p;
8342 }
8343
8344 // Multi-threaded; use CAS to prepend to overflow list
8345 void CMSCollector::par_push_on_overflow_list(oop p) {
8346 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8347 assert(p->is_oop(), "Not an oop");
8348 par_preserve_mark_if_necessary(p);
8349 oop observed_overflow_list = _overflow_list;
8350 oop cur_overflow_list;
8351 do {
8352 cur_overflow_list = observed_overflow_list;
8353 if (cur_overflow_list != BUSY) {
8354 p->set_mark(markOop(cur_overflow_list));
8355 } else {
8356 p->set_mark(NULL);
8357 }
8358 observed_overflow_list =
8359 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8360 } while (cur_overflow_list != observed_overflow_list);
8361 }
8362 #undef BUSY
8363
8364 // Single threaded
8365 // General Note on GrowableArray: pushes may silently fail
8366 // because we are (temporarily) out of C-heap for expanding
8367 // the stack. The problem is quite ubiquitous and affects
8368 // a lot of code in the JVM. The prudent thing for GrowableArray
8369 // to do (for now) is to exit with an error. However, that may
8370 // be too draconian in some cases because the caller may be
8371 // able to recover without much harm. For such cases, we
8372 // should probably introduce a "soft_push" method which returns
8373 // an indication of success or failure with the assumption that
8374 // the caller may be able to recover from a failure; code in
8375 // the VM can then be changed, incrementally, to deal with such
8376 // failures where possible, thus, incrementally hardening the VM
8377 // in such low resource situations.
8378 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8379 _preserved_oop_stack.push(p);
8380 _preserved_mark_stack.push(m);
8381 assert(m == p->mark(), "Mark word changed");
8382 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8383 "bijection");
8384 }
8385
8386 // Single threaded
8387 void CMSCollector::preserve_mark_if_necessary(oop p) {
8388 markOop m = p->mark();
8389 if (m->must_be_preserved(p)) {
8390 preserve_mark_work(p, m);
8391 }
8392 }
8393
8394 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8395 markOop m = p->mark();
8396 if (m->must_be_preserved(p)) {
8397 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8398 // Even though we read the mark word without holding
8399 // the lock, we are assured that it will not change
8400 // because we "own" this oop, so no other thread can
8401 // be trying to push it on the overflow list; see
8402 // the assertion in preserve_mark_work() that checks
8403 // that m == p->mark().
8404 preserve_mark_work(p, m);
8405 }
8406 }
8407
8408 // We should be able to do this multi-threaded,
8409 // a chunk of stack being a task (this is
8410 // correct because each oop only ever appears
8411 // once in the overflow list. However, it's
8412 // not very easy to completely overlap this with
8413 // other operations, so will generally not be done
8414 // until all work's been completed. Because we
8415 // expect the preserved oop stack (set) to be small,
8416 // it's probably fine to do this single-threaded.
8417 // We can explore cleverer concurrent/overlapped/parallel
8418 // processing of preserved marks if we feel the
8419 // need for this in the future. Stack overflow should
8420 // be so rare in practice and, when it happens, its
8421 // effect on performance so great that this will
8422 // likely just be in the noise anyway.
8423 void CMSCollector::restore_preserved_marks_if_any() {
8424 assert(SafepointSynchronize::is_at_safepoint(),
8425 "world should be stopped");
8426 assert(Thread::current()->is_ConcurrentGC_thread() ||
8427 Thread::current()->is_VM_thread(),
8428 "should be single-threaded");
8429 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8430 "bijection");
8431
8432 while (!_preserved_oop_stack.is_empty()) {
8433 oop p = _preserved_oop_stack.pop();
8434 assert(p->is_oop(), "Should be an oop");
8435 assert(_span.contains(p), "oop should be in _span");
8436 assert(p->mark() == markOopDesc::prototype(),
8437 "Set when taken from overflow list");
8438 markOop m = _preserved_mark_stack.pop();
8439 p->set_mark(m);
8440 }
8441 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8442 "stacks were cleared above");
8443 }
8444
8445 #ifndef PRODUCT
8446 bool CMSCollector::no_preserved_marks() const {
8447 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8448 }
8449 #endif
8450
8451 // Transfer some number of overflown objects to usual marking
8452 // stack. Return true if some objects were transferred.
8453 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8454 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8455 (size_t)ParGCDesiredObjsFromOverflowList);
8456
8457 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8458 assert(_collector->overflow_list_is_empty() || res,
8459 "If list is not empty, we should have taken something");
8460 assert(!res || !_mark_stack->isEmpty(),
8461 "If we took something, it should now be on our stack");
8462 return res;
8463 }
8464
8465 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8466 size_t res = _sp->block_size_no_stall(addr, _collector);
8467 if (_sp->block_is_obj(addr)) {
8468 if (_live_bit_map->isMarked(addr)) {
8469 // It can't have been dead in a previous cycle
8470 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8471 } else {
8472 _dead_bit_map->mark(addr); // mark the dead object
8473 }
8474 }
8475 // Could be 0, if the block size could not be computed without stalling.
8476 return res;
8477 }
8478
8479 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8480
8481 switch (phase) {
8482 case CMSCollector::InitialMarking:
8483 initialize(true /* fullGC */ ,
8484 cause /* cause of the GC */,
8485 true /* recordGCBeginTime */,
8486 true /* recordPreGCUsage */,
8487 false /* recordPeakUsage */,
8488 false /* recordPostGCusage */,
8489 true /* recordAccumulatedGCTime */,
8490 false /* recordGCEndTime */,
8491 false /* countCollection */ );
8492 break;
8493
8494 case CMSCollector::FinalMarking:
8495 initialize(true /* fullGC */ ,
8496 cause /* cause of the GC */,
8497 false /* recordGCBeginTime */,
8498 false /* recordPreGCUsage */,
8499 false /* recordPeakUsage */,
8500 false /* recordPostGCusage */,
8501 true /* recordAccumulatedGCTime */,
8502 false /* recordGCEndTime */,
8503 false /* countCollection */ );
8504 break;
8505
8506 case CMSCollector::Sweeping:
8507 initialize(true /* fullGC */ ,
8508 cause /* cause of the GC */,
8509 false /* recordGCBeginTime */,
8510 false /* recordPreGCUsage */,
8511 true /* recordPeakUsage */,
8512 true /* recordPostGCusage */,
8513 false /* recordAccumulatedGCTime */,
8514 true /* recordGCEndTime */,
8515 true /* countCollection */ );
8516 break;
8517
8518 default:
8519 ShouldNotReachHere();
8520 }
8521 }