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->_gen = this;
216
217 _gc_stats = new CMSGCStats();
218
219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
220 // offsets match. The ability to tell free chunks from objects
221 // depends on this property.
222 debug_only(
223 FreeChunk* junk = NULL;
224 assert(UseCompressedClassPointers ||
225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
226 "Offset of FreeChunk::_prev within FreeChunk must match"
227 " that of OopDesc::_klass within OopDesc");
228 )
229
230 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
231 for (uint i = 0; i < ParallelGCThreads; i++) {
232 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
233 }
234
235 _incremental_collection_failed = false;
236 // The "dilatation_factor" is the expansion that can occur on
237 // account of the fact that the minimum object size in the CMS
238 // generation may be larger than that in, say, a contiguous young
239 // generation.
240 // Ideally, in the calculation below, we'd compute the dilatation
241 // factor as: MinChunkSize/(promoting_gen's min object size)
242 // Since we do not have such a general query interface for the
243 // promoting generation, we'll instead just use the minimum
244 // object size (which today is a header's worth of space);
245 // note that all arithmetic is in units of HeapWords.
246 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
247 assert(_dilatation_factor >= 1.0, "from previous assert");
248 }
249
250
251 // The field "_initiating_occupancy" represents the occupancy percentage
252 // at which we trigger a new collection cycle. Unless explicitly specified
253 // via CMSInitiatingOccupancyFraction (argument "io" below), it
254 // is calculated by:
255 //
256 // Let "f" be MinHeapFreeRatio in
257 //
258 // _initiating_occupancy = 100-f +
259 // f * (CMSTriggerRatio/100)
260 // where CMSTriggerRatio is the argument "tr" below.
261 //
262 // That is, if we assume the heap is at its desired maximum occupancy at the
263 // end of a collection, we let CMSTriggerRatio of the (purported) free
264 // space be allocated before initiating a new collection cycle.
265 //
266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
267 assert(io <= 100 && tr <= 100, "Check the arguments");
268 if (io >= 0) {
269 _initiating_occupancy = (double)io / 100.0;
270 } else {
271 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
272 (double)(tr * MinHeapFreeRatio) / 100.0)
273 / 100.0;
274 }
275 }
276
277 void ConcurrentMarkSweepGeneration::ref_processor_init() {
278 assert(collector() != NULL, "no collector");
279 collector()->ref_processor_init();
280 }
281
282 void CMSCollector::ref_processor_init() {
283 if (_ref_processor == NULL) {
284 // Allocate and initialize a reference processor
285 _ref_processor =
286 new ReferenceProcessor(_span, // span
287 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
288 (int) ParallelGCThreads, // mt processing degree
289 _cmsGen->refs_discovery_is_mt(), // mt discovery
290 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
291 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
292 &_is_alive_closure); // closure for liveness info
293 // Initialize the _ref_processor field of CMSGen
294 _cmsGen->set_ref_processor(_ref_processor);
295
296 }
297 }
298
299 AdaptiveSizePolicy* CMSCollector::size_policy() {
300 GenCollectedHeap* gch = GenCollectedHeap::heap();
301 return gch->gen_policy()->size_policy();
302 }
303
304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
305
306 const char* gen_name = "old";
307 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy();
308
309 // Generation Counters - generation 1, 1 subspace
310 _gen_counters = new GenerationCounters(gen_name, 1, 1,
311 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
312
313 _space_counters = new GSpaceCounters(gen_name, 0,
314 _virtual_space.reserved_size(),
315 this, _gen_counters);
316 }
317
318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
319 _cms_gen(cms_gen)
320 {
321 assert(alpha <= 100, "bad value");
322 _saved_alpha = alpha;
323
324 // Initialize the alphas to the bootstrap value of 100.
325 _gc0_alpha = _cms_alpha = 100;
326
327 _cms_begin_time.update();
328 _cms_end_time.update();
329
330 _gc0_duration = 0.0;
331 _gc0_period = 0.0;
332 _gc0_promoted = 0;
333
334 _cms_duration = 0.0;
335 _cms_period = 0.0;
336 _cms_allocated = 0;
337
338 _cms_used_at_gc0_begin = 0;
339 _cms_used_at_gc0_end = 0;
340 _allow_duty_cycle_reduction = false;
341 _valid_bits = 0;
342 }
343
344 double CMSStats::cms_free_adjustment_factor(size_t free) const {
345 // TBD: CR 6909490
346 return 1.0;
347 }
348
349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
350 }
351
352 // If promotion failure handling is on use
353 // the padded average size of the promotion for each
354 // young generation collection.
355 double CMSStats::time_until_cms_gen_full() const {
356 size_t cms_free = _cms_gen->cmsSpace()->free();
357 GenCollectedHeap* gch = GenCollectedHeap::heap();
358 size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
359 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
360 if (cms_free > expected_promotion) {
361 // Start a cms collection if there isn't enough space to promote
362 // for the next minor collection. Use the padded average as
363 // a safety factor.
364 cms_free -= expected_promotion;
365
366 // Adjust by the safety factor.
367 double cms_free_dbl = (double)cms_free;
368 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
369 // Apply a further correction factor which tries to adjust
370 // for recent occurance of concurrent mode failures.
371 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
372 cms_free_dbl = cms_free_dbl * cms_adjustment;
373
374 if (PrintGCDetails && Verbose) {
375 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
376 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
377 cms_free, expected_promotion);
378 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
379 cms_free_dbl, cms_consumption_rate() + 1.0);
380 }
381 // Add 1 in case the consumption rate goes to zero.
382 return cms_free_dbl / (cms_consumption_rate() + 1.0);
383 }
384 return 0.0;
385 }
386
387 // Compare the duration of the cms collection to the
388 // time remaining before the cms generation is empty.
389 // Note that the time from the start of the cms collection
390 // to the start of the cms sweep (less than the total
391 // duration of the cms collection) can be used. This
392 // has been tried and some applications experienced
393 // promotion failures early in execution. This was
394 // possibly because the averages were not accurate
395 // enough at the beginning.
396 double CMSStats::time_until_cms_start() const {
397 // We add "gc0_period" to the "work" calculation
398 // below because this query is done (mostly) at the
399 // end of a scavenge, so we need to conservatively
400 // account for that much possible delay
401 // in the query so as to avoid concurrent mode failures
402 // due to starting the collection just a wee bit too
403 // late.
404 double work = cms_duration() + gc0_period();
405 double deadline = time_until_cms_gen_full();
406 // If a concurrent mode failure occurred recently, we want to be
407 // more conservative and halve our expected time_until_cms_gen_full()
408 if (work > deadline) {
409 if (Verbose && PrintGCDetails) {
410 gclog_or_tty->print(
411 " CMSCollector: collect because of anticipated promotion "
412 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
413 gc0_period(), time_until_cms_gen_full());
414 }
415 return 0.0;
416 }
417 return work - deadline;
418 }
419
420 #ifndef PRODUCT
421 void CMSStats::print_on(outputStream *st) const {
422 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
423 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
424 gc0_duration(), gc0_period(), gc0_promoted());
425 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
426 cms_duration(), cms_period(), cms_allocated());
427 st->print(",cms_since_beg=%g,cms_since_end=%g",
428 cms_time_since_begin(), cms_time_since_end());
429 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
430 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
431
432 if (valid()) {
433 st->print(",promo_rate=%g,cms_alloc_rate=%g",
434 promotion_rate(), cms_allocation_rate());
435 st->print(",cms_consumption_rate=%g,time_until_full=%g",
436 cms_consumption_rate(), time_until_cms_gen_full());
437 }
438 st->print(" ");
439 }
440 #endif // #ifndef PRODUCT
441
442 CMSCollector::CollectorState CMSCollector::_collectorState =
443 CMSCollector::Idling;
444 bool CMSCollector::_foregroundGCIsActive = false;
445 bool CMSCollector::_foregroundGCShouldWait = false;
446
447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
448 CardTableRS* ct,
449 ConcurrentMarkSweepPolicy* cp):
450 _cmsGen(cmsGen),
451 _ct(ct),
452 _ref_processor(NULL), // will be set later
453 _conc_workers(NULL), // may be set later
454 _abort_preclean(false),
455 _start_sampling(false),
456 _between_prologue_and_epilogue(false),
457 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
458 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
459 -1 /* lock-free */, "No_lock" /* dummy */),
460 _modUnionClosurePar(&_modUnionTable),
461 // Adjust my span to cover old (cms) gen
462 _span(cmsGen->reserved()),
463 // Construct the is_alive_closure with _span & markBitMap
464 _is_alive_closure(_span, &_markBitMap),
465 _restart_addr(NULL),
466 _overflow_list(NULL),
467 _stats(cmsGen),
468 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
469 //verify that this lock should be acquired with safepoint check.
470 Monitor::_safepoint_check_sometimes)),
471 _eden_chunk_array(NULL), // may be set in ctor body
472 _eden_chunk_capacity(0), // -- ditto --
473 _eden_chunk_index(0), // -- ditto --
474 _survivor_plab_array(NULL), // -- ditto --
475 _survivor_chunk_array(NULL), // -- ditto --
476 _survivor_chunk_capacity(0), // -- ditto --
477 _survivor_chunk_index(0), // -- ditto --
478 _ser_pmc_preclean_ovflw(0),
479 _ser_kac_preclean_ovflw(0),
480 _ser_pmc_remark_ovflw(0),
481 _par_pmc_remark_ovflw(0),
482 _ser_kac_ovflw(0),
483 _par_kac_ovflw(0),
484 #ifndef PRODUCT
485 _num_par_pushes(0),
486 #endif
487 _collection_count_start(0),
488 _verifying(false),
489 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
490 _completed_initialization(false),
491 _collector_policy(cp),
492 _should_unload_classes(CMSClassUnloadingEnabled),
493 _concurrent_cycles_since_last_unload(0),
494 _roots_scanning_options(GenCollectedHeap::SO_None),
495 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
496 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
497 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
498 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
499 _cms_start_registered(false)
500 {
501 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
502 ExplicitGCInvokesConcurrent = true;
503 }
504 // Now expand the span and allocate the collection support structures
505 // (MUT, marking bit map etc.) to cover both generations subject to
506 // collection.
507
508 // For use by dirty card to oop closures.
509 _cmsGen->cmsSpace()->set_collector(this);
510
511 // Allocate MUT and marking bit map
512 {
513 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
514 if (!_markBitMap.allocate(_span)) {
515 warning("Failed to allocate CMS Bit Map");
516 return;
517 }
518 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
519 }
520 {
521 _modUnionTable.allocate(_span);
522 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
523 }
524
525 if (!_markStack.allocate(MarkStackSize)) {
526 warning("Failed to allocate CMS Marking Stack");
527 return;
528 }
529
530 // Support for multi-threaded concurrent phases
531 if (CMSConcurrentMTEnabled) {
532 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
533 // just for now
534 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
535 }
536 if (ConcGCThreads > 1) {
537 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
538 ConcGCThreads, true);
539 if (_conc_workers == NULL) {
540 warning("GC/CMS: _conc_workers allocation failure: "
541 "forcing -CMSConcurrentMTEnabled");
542 CMSConcurrentMTEnabled = false;
543 } else {
544 _conc_workers->initialize_workers();
545 }
546 } else {
547 CMSConcurrentMTEnabled = false;
548 }
549 }
550 if (!CMSConcurrentMTEnabled) {
551 ConcGCThreads = 0;
552 } else {
553 // Turn off CMSCleanOnEnter optimization temporarily for
554 // the MT case where it's not fixed yet; see 6178663.
555 CMSCleanOnEnter = false;
556 }
557 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
558 "Inconsistency");
559
560 // Parallel task queues; these are shared for the
561 // concurrent and stop-world phases of CMS, but
562 // are not shared with parallel scavenge (ParNew).
563 {
564 uint i;
565 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
566
567 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
568 || ParallelRefProcEnabled)
569 && num_queues > 0) {
570 _task_queues = new OopTaskQueueSet(num_queues);
571 if (_task_queues == NULL) {
572 warning("task_queues allocation failure.");
573 return;
574 }
575 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
576 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
577 for (i = 0; i < num_queues; i++) {
578 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
579 if (q == NULL) {
580 warning("work_queue allocation failure.");
581 return;
582 }
583 _task_queues->register_queue(i, q);
584 }
585 for (i = 0; i < num_queues; i++) {
586 _task_queues->queue(i)->initialize();
587 _hash_seed[i] = 17; // copied from ParNew
588 }
589 }
590 }
591
592 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
593
594 // Clip CMSBootstrapOccupancy between 0 and 100.
595 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
596
597 // Now tell CMS generations the identity of their collector
598 ConcurrentMarkSweepGeneration::set_collector(this);
599
600 // Create & start a CMS thread for this CMS collector
601 _cmsThread = ConcurrentMarkSweepThread::start(this);
602 assert(cmsThread() != NULL, "CMS Thread should have been created");
603 assert(cmsThread()->collector() == this,
604 "CMS Thread should refer to this gen");
605 assert(CGC_lock != NULL, "Where's the CGC_lock?");
606
607 // Support for parallelizing young gen rescan
608 GenCollectedHeap* gch = GenCollectedHeap::heap();
609 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
610 _young_gen = (ParNewGeneration*)gch->young_gen();
611 if (gch->supports_inline_contig_alloc()) {
612 _top_addr = gch->top_addr();
613 _end_addr = gch->end_addr();
614 assert(_young_gen != NULL, "no _young_gen");
615 _eden_chunk_index = 0;
616 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
617 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
618 }
619
620 // Support for parallelizing survivor space rescan
621 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
622 const size_t max_plab_samples =
623 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size();
624
625 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
626 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
627 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
628 _survivor_chunk_capacity = 2*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 size_t CMSCollector::plab_sample_minimum_size() {
645 // The default value of MinTLABSize is 2k, but there is
646 // no way to get the default value if the flag has been overridden.
647 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K);
648 }
649
650 const char* ConcurrentMarkSweepGeneration::name() const {
651 return "concurrent mark-sweep generation";
652 }
653 void ConcurrentMarkSweepGeneration::update_counters() {
654 if (UsePerfData) {
655 _space_counters->update_all();
656 _gen_counters->update_all();
657 }
658 }
659
660 // this is an optimized version of update_counters(). it takes the
661 // used value as a parameter rather than computing it.
662 //
663 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
664 if (UsePerfData) {
665 _space_counters->update_used(used);
666 _space_counters->update_capacity();
667 _gen_counters->update_all();
668 }
669 }
670
671 void ConcurrentMarkSweepGeneration::print() const {
672 Generation::print();
673 cmsSpace()->print();
674 }
675
676 #ifndef PRODUCT
677 void ConcurrentMarkSweepGeneration::print_statistics() {
678 cmsSpace()->printFLCensus(0);
679 }
680 #endif
681
682 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
683 GenCollectedHeap* gch = GenCollectedHeap::heap();
684 if (PrintGCDetails) {
685 // I didn't want to change the logging when removing the level concept,
686 // but I guess this logging could say "old" or something instead of "1".
687 assert(gch->is_old_gen(this),
688 "The CMS generation should be the old generation");
689 uint level = 1;
690 if (Verbose) {
691 gclog_or_tty->print("[%u %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
692 level, short_name(), s, used(), capacity());
693 } else {
694 gclog_or_tty->print("[%u %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
695 level, short_name(), s, used() / K, capacity() / K);
696 }
697 }
698 if (Verbose) {
699 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
700 gch->used(), gch->capacity());
701 } else {
702 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
703 gch->used() / K, gch->capacity() / K);
704 }
705 }
706
707 size_t
708 ConcurrentMarkSweepGeneration::contiguous_available() const {
709 // dld proposes an improvement in precision here. If the committed
710 // part of the space ends in a free block we should add that to
711 // uncommitted size in the calculation below. Will make this
712 // change later, staying with the approximation below for the
713 // time being. -- ysr.
714 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
715 }
716
717 size_t
718 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
719 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
720 }
721
722 size_t ConcurrentMarkSweepGeneration::max_available() const {
723 return free() + _virtual_space.uncommitted_size();
724 }
725
726 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
727 size_t available = max_available();
728 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
729 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
730 if (Verbose && PrintGCDetails) {
731 gclog_or_tty->print_cr(
732 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
733 "max_promo("SIZE_FORMAT")",
734 res? "":" not", available, res? ">=":"<",
735 av_promo, max_promotion_in_bytes);
736 }
737 return res;
738 }
739
740 // At a promotion failure dump information on block layout in heap
741 // (cms old generation).
742 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
743 if (CMSDumpAtPromotionFailure) {
744 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
745 }
746 }
747
748 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
749 // Clear the promotion information. These pointers can be adjusted
750 // along with all the other pointers into the heap but
751 // compaction is expected to be a rare event with
752 // a heap using cms so don't do it without seeing the need.
753 for (uint i = 0; i < ParallelGCThreads; i++) {
754 _par_gc_thread_states[i]->promo.reset();
755 }
756 }
757
758 void ConcurrentMarkSweepGeneration::compute_new_size() {
759 assert_locked_or_safepoint(Heap_lock);
760
761 // If incremental collection failed, we just want to expand
762 // to the limit.
763 if (incremental_collection_failed()) {
764 clear_incremental_collection_failed();
765 grow_to_reserved();
766 return;
767 }
768
769 // The heap has been compacted but not reset yet.
770 // Any metric such as free() or used() will be incorrect.
771
772 CardGeneration::compute_new_size();
773
774 // Reset again after a possible resizing
775 if (did_compact()) {
776 cmsSpace()->reset_after_compaction();
777 }
778 }
779
780 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
781 assert_locked_or_safepoint(Heap_lock);
782
783 // If incremental collection failed, we just want to expand
784 // to the limit.
785 if (incremental_collection_failed()) {
786 clear_incremental_collection_failed();
787 grow_to_reserved();
788 return;
789 }
790
791 double free_percentage = ((double) free()) / capacity();
792 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
793 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
794
795 // compute expansion delta needed for reaching desired free percentage
796 if (free_percentage < desired_free_percentage) {
797 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
798 assert(desired_capacity >= capacity(), "invalid expansion size");
799 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
800 if (PrintGCDetails && Verbose) {
801 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
802 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
803 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
804 gclog_or_tty->print_cr(" Desired free fraction %f",
805 desired_free_percentage);
806 gclog_or_tty->print_cr(" Maximum free fraction %f",
807 maximum_free_percentage);
808 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000);
809 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
810 desired_capacity/1000);
811 GenCollectedHeap* gch = GenCollectedHeap::heap();
812 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
813 size_t young_size = gch->young_gen()->capacity();
814 gclog_or_tty->print_cr(" Young gen size " SIZE_FORMAT, young_size / 1000);
815 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
816 unsafe_max_alloc_nogc()/1000);
817 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
818 contiguous_available()/1000);
819 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
820 expand_bytes);
821 }
822 // safe if expansion fails
823 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
824 if (PrintGCDetails && Verbose) {
825 gclog_or_tty->print_cr(" Expanded free fraction %f",
826 ((double) free()) / capacity());
827 }
828 } else {
829 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
830 assert(desired_capacity <= capacity(), "invalid expansion size");
831 size_t shrink_bytes = capacity() - desired_capacity;
832 // Don't shrink unless the delta is greater than the minimum shrink we want
833 if (shrink_bytes >= MinHeapDeltaBytes) {
834 shrink_free_list_by(shrink_bytes);
835 }
836 }
837 }
838
839 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
840 return cmsSpace()->freelistLock();
841 }
842
843 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
844 bool tlab) {
845 CMSSynchronousYieldRequest yr;
846 MutexLockerEx x(freelistLock(),
847 Mutex::_no_safepoint_check_flag);
848 return have_lock_and_allocate(size, tlab);
849 }
850
851 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
852 bool tlab /* ignored */) {
853 assert_lock_strong(freelistLock());
854 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
855 HeapWord* res = cmsSpace()->allocate(adjustedSize);
856 // Allocate the object live (grey) if the background collector has
857 // started marking. This is necessary because the marker may
858 // have passed this address and consequently this object will
859 // not otherwise be greyed and would be incorrectly swept up.
860 // Note that if this object contains references, the writing
861 // of those references will dirty the card containing this object
862 // allowing the object to be blackened (and its references scanned)
863 // either during a preclean phase or at the final checkpoint.
864 if (res != NULL) {
865 // We may block here with an uninitialized object with
866 // its mark-bit or P-bits not yet set. Such objects need
867 // to be safely navigable by block_start().
868 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
869 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
870 collector()->direct_allocated(res, adjustedSize);
871 _direct_allocated_words += adjustedSize;
872 // allocation counters
873 NOT_PRODUCT(
874 _numObjectsAllocated++;
875 _numWordsAllocated += (int)adjustedSize;
876 )
877 }
878 return res;
879 }
880
881 // In the case of direct allocation by mutators in a generation that
882 // is being concurrently collected, the object must be allocated
883 // live (grey) if the background collector has started marking.
884 // This is necessary because the marker may
885 // have passed this address and consequently this object will
886 // not otherwise be greyed and would be incorrectly swept up.
887 // Note that if this object contains references, the writing
888 // of those references will dirty the card containing this object
889 // allowing the object to be blackened (and its references scanned)
890 // either during a preclean phase or at the final checkpoint.
891 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
892 assert(_markBitMap.covers(start, size), "Out of bounds");
893 if (_collectorState >= Marking) {
894 MutexLockerEx y(_markBitMap.lock(),
895 Mutex::_no_safepoint_check_flag);
896 // [see comments preceding SweepClosure::do_blk() below for details]
897 //
898 // Can the P-bits be deleted now? JJJ
899 //
900 // 1. need to mark the object as live so it isn't collected
901 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
902 // 3. need to mark the end of the object so marking, precleaning or sweeping
903 // can skip over uninitialized or unparsable objects. An allocated
904 // object is considered uninitialized for our purposes as long as
905 // its klass word is NULL. All old gen objects are parsable
906 // as soon as they are initialized.)
907 _markBitMap.mark(start); // object is live
908 _markBitMap.mark(start + 1); // object is potentially uninitialized?
909 _markBitMap.mark(start + size - 1);
910 // mark end of object
911 }
912 // check that oop looks uninitialized
913 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
914 }
915
916 void CMSCollector::promoted(bool par, HeapWord* start,
917 bool is_obj_array, size_t obj_size) {
918 assert(_markBitMap.covers(start), "Out of bounds");
919 // See comment in direct_allocated() about when objects should
920 // be allocated live.
921 if (_collectorState >= Marking) {
922 // we already hold the marking bit map lock, taken in
923 // the prologue
924 if (par) {
925 _markBitMap.par_mark(start);
926 } else {
927 _markBitMap.mark(start);
928 }
929 // We don't need to mark the object as uninitialized (as
930 // in direct_allocated above) because this is being done with the
931 // world stopped and the object will be initialized by the
932 // time the marking, precleaning or sweeping get to look at it.
933 // But see the code for copying objects into the CMS generation,
934 // where we need to ensure that concurrent readers of the
935 // block offset table are able to safely navigate a block that
936 // is in flux from being free to being allocated (and in
937 // transition while being copied into) and subsequently
938 // becoming a bona-fide object when the copy/promotion is complete.
939 assert(SafepointSynchronize::is_at_safepoint(),
940 "expect promotion only at safepoints");
941
942 if (_collectorState < Sweeping) {
943 // Mark the appropriate cards in the modUnionTable, so that
944 // this object gets scanned before the sweep. If this is
945 // not done, CMS generation references in the object might
946 // not get marked.
947 // For the case of arrays, which are otherwise precisely
948 // marked, we need to dirty the entire array, not just its head.
949 if (is_obj_array) {
950 // The [par_]mark_range() method expects mr.end() below to
951 // be aligned to the granularity of a bit's representation
952 // in the heap. In the case of the MUT below, that's a
953 // card size.
954 MemRegion mr(start,
955 (HeapWord*)round_to((intptr_t)(start + obj_size),
956 CardTableModRefBS::card_size /* bytes */));
957 if (par) {
958 _modUnionTable.par_mark_range(mr);
959 } else {
960 _modUnionTable.mark_range(mr);
961 }
962 } else { // not an obj array; we can just mark the head
963 if (par) {
964 _modUnionTable.par_mark(start);
965 } else {
966 _modUnionTable.mark(start);
967 }
968 }
969 }
970 }
971 }
972
973 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
974 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
975 // allocate, copy and if necessary update promoinfo --
976 // delegate to underlying space.
977 assert_lock_strong(freelistLock());
978
979 #ifndef PRODUCT
980 if (GenCollectedHeap::heap()->promotion_should_fail()) {
981 return NULL;
982 }
983 #endif // #ifndef PRODUCT
984
985 oop res = _cmsSpace->promote(obj, obj_size);
986 if (res == NULL) {
987 // expand and retry
988 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
989 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
990 // Since this is the old generation, we don't try to promote
991 // into a more senior generation.
992 res = _cmsSpace->promote(obj, obj_size);
993 }
994 if (res != NULL) {
995 // See comment in allocate() about when objects should
996 // be allocated live.
997 assert(obj->is_oop(), "Will dereference klass pointer below");
998 collector()->promoted(false, // Not parallel
999 (HeapWord*)res, obj->is_objArray(), obj_size);
1000 // promotion counters
1001 NOT_PRODUCT(
1002 _numObjectsPromoted++;
1003 _numWordsPromoted +=
1004 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1005 )
1006 }
1007 return res;
1008 }
1009
1010
1011 // IMPORTANT: Notes on object size recognition in CMS.
1012 // ---------------------------------------------------
1013 // A block of storage in the CMS generation is always in
1014 // one of three states. A free block (FREE), an allocated
1015 // object (OBJECT) whose size() method reports the correct size,
1016 // and an intermediate state (TRANSIENT) in which its size cannot
1017 // be accurately determined.
1018 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1019 // -----------------------------------------------------
1020 // FREE: klass_word & 1 == 1; mark_word holds block size
1021 //
1022 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1023 // obj->size() computes correct size
1024 //
1025 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1026 //
1027 // STATE IDENTIFICATION: (64 bit+COOPS)
1028 // ------------------------------------
1029 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1030 //
1031 // OBJECT: klass_word installed; klass_word != 0;
1032 // obj->size() computes correct size
1033 //
1034 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1035 //
1036 //
1037 // STATE TRANSITION DIAGRAM
1038 //
1039 // mut / parnew mut / parnew
1040 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1041 // ^ |
1042 // |------------------------ DEAD <------------------------------------|
1043 // sweep mut
1044 //
1045 // While a block is in TRANSIENT state its size cannot be determined
1046 // so readers will either need to come back later or stall until
1047 // the size can be determined. Note that for the case of direct
1048 // allocation, P-bits, when available, may be used to determine the
1049 // size of an object that may not yet have been initialized.
1050
1051 // Things to support parallel young-gen collection.
1052 oop
1053 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1054 oop old, markOop m,
1055 size_t word_sz) {
1056 #ifndef PRODUCT
1057 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1058 return NULL;
1059 }
1060 #endif // #ifndef PRODUCT
1061
1062 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1063 PromotionInfo* promoInfo = &ps->promo;
1064 // if we are tracking promotions, then first ensure space for
1065 // promotion (including spooling space for saving header if necessary).
1066 // then allocate and copy, then track promoted info if needed.
1067 // When tracking (see PromotionInfo::track()), the mark word may
1068 // be displaced and in this case restoration of the mark word
1069 // occurs in the (oop_since_save_marks_)iterate phase.
1070 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1071 // Out of space for allocating spooling buffers;
1072 // try expanding and allocating spooling buffers.
1073 if (!expand_and_ensure_spooling_space(promoInfo)) {
1074 return NULL;
1075 }
1076 }
1077 assert(promoInfo->has_spooling_space(), "Control point invariant");
1078 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1079 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1080 if (obj_ptr == NULL) {
1081 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1082 if (obj_ptr == NULL) {
1083 return NULL;
1084 }
1085 }
1086 oop obj = oop(obj_ptr);
1087 OrderAccess::storestore();
1088 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1089 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1090 // IMPORTANT: See note on object initialization for CMS above.
1091 // Otherwise, copy the object. Here we must be careful to insert the
1092 // klass pointer last, since this marks the block as an allocated object.
1093 // Except with compressed oops it's the mark word.
1094 HeapWord* old_ptr = (HeapWord*)old;
1095 // Restore the mark word copied above.
1096 obj->set_mark(m);
1097 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1098 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1099 OrderAccess::storestore();
1100
1101 if (UseCompressedClassPointers) {
1102 // Copy gap missed by (aligned) header size calculation below
1103 obj->set_klass_gap(old->klass_gap());
1104 }
1105 if (word_sz > (size_t)oopDesc::header_size()) {
1106 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1107 obj_ptr + oopDesc::header_size(),
1108 word_sz - oopDesc::header_size());
1109 }
1110
1111 // Now we can track the promoted object, if necessary. We take care
1112 // to delay the transition from uninitialized to full object
1113 // (i.e., insertion of klass pointer) until after, so that it
1114 // atomically becomes a promoted object.
1115 if (promoInfo->tracking()) {
1116 promoInfo->track((PromotedObject*)obj, old->klass());
1117 }
1118 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1119 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1120 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1121
1122 // Finally, install the klass pointer (this should be volatile).
1123 OrderAccess::storestore();
1124 obj->set_klass(old->klass());
1125 // We should now be able to calculate the right size for this object
1126 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1127
1128 collector()->promoted(true, // parallel
1129 obj_ptr, old->is_objArray(), word_sz);
1130
1131 NOT_PRODUCT(
1132 Atomic::inc_ptr(&_numObjectsPromoted);
1133 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1134 )
1135
1136 return obj;
1137 }
1138
1139 void
1140 ConcurrentMarkSweepGeneration::
1141 par_promote_alloc_done(int thread_num) {
1142 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1143 ps->lab.retire(thread_num);
1144 }
1145
1146 void
1147 ConcurrentMarkSweepGeneration::
1148 par_oop_since_save_marks_iterate_done(int thread_num) {
1149 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1150 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1151 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1152 }
1153
1154 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1155 size_t size,
1156 bool tlab)
1157 {
1158 // We allow a STW collection only if a full
1159 // collection was requested.
1160 return full || should_allocate(size, tlab); // FIX ME !!!
1161 // This and promotion failure handling are connected at the
1162 // hip and should be fixed by untying them.
1163 }
1164
1165 bool CMSCollector::shouldConcurrentCollect() {
1166 if (_full_gc_requested) {
1167 if (Verbose && PrintGCDetails) {
1168 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1169 " gc request (or gc_locker)");
1170 }
1171 return true;
1172 }
1173
1174 FreelistLocker x(this);
1175 // ------------------------------------------------------------------
1176 // Print out lots of information which affects the initiation of
1177 // a collection.
1178 if (PrintCMSInitiationStatistics && stats().valid()) {
1179 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1180 gclog_or_tty->stamp();
1181 gclog_or_tty->cr();
1182 stats().print_on(gclog_or_tty);
1183 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1184 stats().time_until_cms_gen_full());
1185 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1186 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1187 _cmsGen->contiguous_available());
1188 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1189 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1190 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1191 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1192 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1193 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1194 gclog_or_tty->print_cr("metadata initialized %d",
1195 MetaspaceGC::should_concurrent_collect());
1196 }
1197 // ------------------------------------------------------------------
1198
1199 // If the estimated time to complete a cms collection (cms_duration())
1200 // is less than the estimated time remaining until the cms generation
1201 // is full, start a collection.
1202 if (!UseCMSInitiatingOccupancyOnly) {
1203 if (stats().valid()) {
1204 if (stats().time_until_cms_start() == 0.0) {
1205 return true;
1206 }
1207 } else {
1208 // We want to conservatively collect somewhat early in order
1209 // to try and "bootstrap" our CMS/promotion statistics;
1210 // this branch will not fire after the first successful CMS
1211 // collection because the stats should then be valid.
1212 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1213 if (Verbose && PrintGCDetails) {
1214 gclog_or_tty->print_cr(
1215 " CMSCollector: collect for bootstrapping statistics:"
1216 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1217 _bootstrap_occupancy);
1218 }
1219 return true;
1220 }
1221 }
1222 }
1223
1224 // Otherwise, we start a collection cycle if
1225 // old gen want a collection cycle started. Each may use
1226 // an appropriate criterion for making this decision.
1227 // XXX We need to make sure that the gen expansion
1228 // criterion dovetails well with this. XXX NEED TO FIX THIS
1229 if (_cmsGen->should_concurrent_collect()) {
1230 if (Verbose && PrintGCDetails) {
1231 gclog_or_tty->print_cr("CMS old gen initiated");
1232 }
1233 return true;
1234 }
1235
1236 // We start a collection if we believe an incremental collection may fail;
1237 // this is not likely to be productive in practice because it's probably too
1238 // late anyway.
1239 GenCollectedHeap* gch = GenCollectedHeap::heap();
1240 assert(gch->collector_policy()->is_generation_policy(),
1241 "You may want to check the correctness of the following");
1242 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1243 if (Verbose && PrintGCDetails) {
1244 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1245 }
1246 return true;
1247 }
1248
1249 if (MetaspaceGC::should_concurrent_collect()) {
1250 if (Verbose && PrintGCDetails) {
1251 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1252 }
1253 return true;
1254 }
1255
1256 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1257 if (CMSTriggerInterval >= 0) {
1258 if (CMSTriggerInterval == 0) {
1259 // Trigger always
1260 return true;
1261 }
1262
1263 // Check the CMS time since begin (we do not check the stats validity
1264 // as we want to be able to trigger the first CMS cycle as well)
1265 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1266 if (Verbose && PrintGCDetails) {
1267 if (stats().valid()) {
1268 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1269 stats().cms_time_since_begin());
1270 } else {
1271 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1272 }
1273 }
1274 return true;
1275 }
1276 }
1277
1278 return false;
1279 }
1280
1281 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1282
1283 // Clear _expansion_cause fields of constituent generations
1284 void CMSCollector::clear_expansion_cause() {
1285 _cmsGen->clear_expansion_cause();
1286 }
1287
1288 // We should be conservative in starting a collection cycle. To
1289 // start too eagerly runs the risk of collecting too often in the
1290 // extreme. To collect too rarely falls back on full collections,
1291 // which works, even if not optimum in terms of concurrent work.
1292 // As a work around for too eagerly collecting, use the flag
1293 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1294 // giving the user an easily understandable way of controlling the
1295 // collections.
1296 // We want to start a new collection cycle if any of the following
1297 // conditions hold:
1298 // . our current occupancy exceeds the configured initiating occupancy
1299 // for this generation, or
1300 // . we recently needed to expand this space and have not, since that
1301 // expansion, done a collection of this generation, or
1302 // . the underlying space believes that it may be a good idea to initiate
1303 // a concurrent collection (this may be based on criteria such as the
1304 // following: the space uses linear allocation and linear allocation is
1305 // going to fail, or there is believed to be excessive fragmentation in
1306 // the generation, etc... or ...
1307 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1308 // the case of the old generation; see CR 6543076):
1309 // we may be approaching a point at which allocation requests may fail because
1310 // we will be out of sufficient free space given allocation rate estimates.]
1311 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1312
1313 assert_lock_strong(freelistLock());
1314 if (occupancy() > initiating_occupancy()) {
1315 if (PrintGCDetails && Verbose) {
1316 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1317 short_name(), occupancy(), initiating_occupancy());
1318 }
1319 return true;
1320 }
1321 if (UseCMSInitiatingOccupancyOnly) {
1322 return false;
1323 }
1324 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1325 if (PrintGCDetails && Verbose) {
1326 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1327 short_name());
1328 }
1329 return true;
1330 }
1331 if (_cmsSpace->should_concurrent_collect()) {
1332 if (PrintGCDetails && Verbose) {
1333 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1334 short_name());
1335 }
1336 return true;
1337 }
1338 return false;
1339 }
1340
1341 void ConcurrentMarkSweepGeneration::collect(bool full,
1342 bool clear_all_soft_refs,
1343 size_t size,
1344 bool tlab)
1345 {
1346 collector()->collect(full, clear_all_soft_refs, size, tlab);
1347 }
1348
1349 void CMSCollector::collect(bool full,
1350 bool clear_all_soft_refs,
1351 size_t size,
1352 bool tlab)
1353 {
1354 // The following "if" branch is present for defensive reasons.
1355 // In the current uses of this interface, it can be replaced with:
1356 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1357 // But I am not placing that assert here to allow future
1358 // generality in invoking this interface.
1359 if (GC_locker::is_active()) {
1360 // A consistency test for GC_locker
1361 assert(GC_locker::needs_gc(), "Should have been set already");
1362 // Skip this foreground collection, instead
1363 // expanding the heap if necessary.
1364 // Need the free list locks for the call to free() in compute_new_size()
1365 compute_new_size();
1366 return;
1367 }
1368 acquire_control_and_collect(full, clear_all_soft_refs);
1369 }
1370
1371 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1372 GenCollectedHeap* gch = GenCollectedHeap::heap();
1373 unsigned int gc_count = gch->total_full_collections();
1374 if (gc_count == full_gc_count) {
1375 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1376 _full_gc_requested = true;
1377 _full_gc_cause = cause;
1378 CGC_lock->notify(); // nudge CMS thread
1379 } else {
1380 assert(gc_count > full_gc_count, "Error: causal loop");
1381 }
1382 }
1383
1384 bool CMSCollector::is_external_interruption() {
1385 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1386 return GCCause::is_user_requested_gc(cause) ||
1387 GCCause::is_serviceability_requested_gc(cause);
1388 }
1389
1390 void CMSCollector::report_concurrent_mode_interruption() {
1391 if (is_external_interruption()) {
1392 if (PrintGCDetails) {
1393 gclog_or_tty->print(" (concurrent mode interrupted)");
1394 }
1395 } else {
1396 if (PrintGCDetails) {
1397 gclog_or_tty->print(" (concurrent mode failure)");
1398 }
1399 _gc_tracer_cm->report_concurrent_mode_failure();
1400 }
1401 }
1402
1403
1404 // The foreground and background collectors need to coordinate in order
1405 // to make sure that they do not mutually interfere with CMS collections.
1406 // When a background collection is active,
1407 // the foreground collector may need to take over (preempt) and
1408 // synchronously complete an ongoing collection. Depending on the
1409 // frequency of the background collections and the heap usage
1410 // of the application, this preemption can be seldom or frequent.
1411 // There are only certain
1412 // points in the background collection that the "collection-baton"
1413 // can be passed to the foreground collector.
1414 //
1415 // The foreground collector will wait for the baton before
1416 // starting any part of the collection. The foreground collector
1417 // will only wait at one location.
1418 //
1419 // The background collector will yield the baton before starting a new
1420 // phase of the collection (e.g., before initial marking, marking from roots,
1421 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1422 // of the loop which switches the phases. The background collector does some
1423 // of the phases (initial mark, final re-mark) with the world stopped.
1424 // Because of locking involved in stopping the world,
1425 // the foreground collector should not block waiting for the background
1426 // collector when it is doing a stop-the-world phase. The background
1427 // collector will yield the baton at an additional point just before
1428 // it enters a stop-the-world phase. Once the world is stopped, the
1429 // background collector checks the phase of the collection. If the
1430 // phase has not changed, it proceeds with the collection. If the
1431 // phase has changed, it skips that phase of the collection. See
1432 // the comments on the use of the Heap_lock in collect_in_background().
1433 //
1434 // Variable used in baton passing.
1435 // _foregroundGCIsActive - Set to true by the foreground collector when
1436 // it wants the baton. The foreground clears it when it has finished
1437 // the collection.
1438 // _foregroundGCShouldWait - Set to true by the background collector
1439 // when it is running. The foreground collector waits while
1440 // _foregroundGCShouldWait is true.
1441 // CGC_lock - monitor used to protect access to the above variables
1442 // and to notify the foreground and background collectors.
1443 // _collectorState - current state of the CMS collection.
1444 //
1445 // The foreground collector
1446 // acquires the CGC_lock
1447 // sets _foregroundGCIsActive
1448 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1449 // various locks acquired in preparation for the collection
1450 // are released so as not to block the background collector
1451 // that is in the midst of a collection
1452 // proceeds with the collection
1453 // clears _foregroundGCIsActive
1454 // returns
1455 //
1456 // The background collector in a loop iterating on the phases of the
1457 // collection
1458 // acquires the CGC_lock
1459 // sets _foregroundGCShouldWait
1460 // if _foregroundGCIsActive is set
1461 // clears _foregroundGCShouldWait, notifies _CGC_lock
1462 // waits on _CGC_lock for _foregroundGCIsActive to become false
1463 // and exits the loop.
1464 // otherwise
1465 // proceed with that phase of the collection
1466 // if the phase is a stop-the-world phase,
1467 // yield the baton once more just before enqueueing
1468 // the stop-world CMS operation (executed by the VM thread).
1469 // returns after all phases of the collection are done
1470 //
1471
1472 void CMSCollector::acquire_control_and_collect(bool full,
1473 bool clear_all_soft_refs) {
1474 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1475 assert(!Thread::current()->is_ConcurrentGC_thread(),
1476 "shouldn't try to acquire control from self!");
1477
1478 // Start the protocol for acquiring control of the
1479 // collection from the background collector (aka CMS thread).
1480 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1481 "VM thread should have CMS token");
1482 // Remember the possibly interrupted state of an ongoing
1483 // concurrent collection
1484 CollectorState first_state = _collectorState;
1485
1486 // Signal to a possibly ongoing concurrent collection that
1487 // we want to do a foreground collection.
1488 _foregroundGCIsActive = true;
1489
1490 // release locks and wait for a notify from the background collector
1491 // releasing the locks in only necessary for phases which
1492 // do yields to improve the granularity of the collection.
1493 assert_lock_strong(bitMapLock());
1494 // We need to lock the Free list lock for the space that we are
1495 // currently collecting.
1496 assert(haveFreelistLocks(), "Must be holding free list locks");
1497 bitMapLock()->unlock();
1498 releaseFreelistLocks();
1499 {
1500 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1501 if (_foregroundGCShouldWait) {
1502 // We are going to be waiting for action for the CMS thread;
1503 // it had better not be gone (for instance at shutdown)!
1504 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1505 "CMS thread must be running");
1506 // Wait here until the background collector gives us the go-ahead
1507 ConcurrentMarkSweepThread::clear_CMS_flag(
1508 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1509 // Get a possibly blocked CMS thread going:
1510 // Note that we set _foregroundGCIsActive true above,
1511 // without protection of the CGC_lock.
1512 CGC_lock->notify();
1513 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1514 "Possible deadlock");
1515 while (_foregroundGCShouldWait) {
1516 // wait for notification
1517 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1518 // Possibility of delay/starvation here, since CMS token does
1519 // not know to give priority to VM thread? Actually, i think
1520 // there wouldn't be any delay/starvation, but the proof of
1521 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1522 }
1523 ConcurrentMarkSweepThread::set_CMS_flag(
1524 ConcurrentMarkSweepThread::CMS_vm_has_token);
1525 }
1526 }
1527 // The CMS_token is already held. Get back the other locks.
1528 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1529 "VM thread should have CMS token");
1530 getFreelistLocks();
1531 bitMapLock()->lock_without_safepoint_check();
1532 if (TraceCMSState) {
1533 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1534 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1535 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1536 }
1537
1538 // Inform cms gen if this was due to partial collection failing.
1539 // The CMS gen may use this fact to determine its expansion policy.
1540 GenCollectedHeap* gch = GenCollectedHeap::heap();
1541 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1542 assert(!_cmsGen->incremental_collection_failed(),
1543 "Should have been noticed, reacted to and cleared");
1544 _cmsGen->set_incremental_collection_failed();
1545 }
1546
1547 if (first_state > Idling) {
1548 report_concurrent_mode_interruption();
1549 }
1550
1551 set_did_compact(true);
1552
1553 // If the collection is being acquired from the background
1554 // collector, there may be references on the discovered
1555 // references lists. Abandon those references, since some
1556 // of them may have become unreachable after concurrent
1557 // discovery; the STW compacting collector will redo discovery
1558 // more precisely, without being subject to floating garbage.
1559 // Leaving otherwise unreachable references in the discovered
1560 // lists would require special handling.
1561 ref_processor()->disable_discovery();
1562 ref_processor()->abandon_partial_discovery();
1563 ref_processor()->verify_no_references_recorded();
1564
1565 if (first_state > Idling) {
1566 save_heap_summary();
1567 }
1568
1569 do_compaction_work(clear_all_soft_refs);
1570
1571 // Has the GC time limit been exceeded?
1572 size_t max_eden_size = _young_gen->max_capacity() -
1573 _young_gen->to()->capacity() -
1574 _young_gen->from()->capacity();
1575 GCCause::Cause gc_cause = gch->gc_cause();
1576 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1577 _young_gen->eden()->used(),
1578 _cmsGen->max_capacity(),
1579 max_eden_size,
1580 full,
1581 gc_cause,
1582 gch->collector_policy());
1583
1584 // Reset the expansion cause, now that we just completed
1585 // a collection cycle.
1586 clear_expansion_cause();
1587 _foregroundGCIsActive = false;
1588 return;
1589 }
1590
1591 // Resize the tenured generation
1592 // after obtaining the free list locks for the
1593 // two generations.
1594 void CMSCollector::compute_new_size() {
1595 assert_locked_or_safepoint(Heap_lock);
1596 FreelistLocker z(this);
1597 MetaspaceGC::compute_new_size();
1598 _cmsGen->compute_new_size_free_list();
1599 }
1600
1601 // A work method used by the foreground collector to do
1602 // a mark-sweep-compact.
1603 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1604 GenCollectedHeap* gch = GenCollectedHeap::heap();
1605
1606 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1607 gc_timer->register_gc_start();
1608
1609 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1610 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1611
1612 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1613
1614 // Temporarily widen the span of the weak reference processing to
1615 // the entire heap.
1616 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1617 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1618 // Temporarily, clear the "is_alive_non_header" field of the
1619 // reference processor.
1620 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1621 // Temporarily make reference _processing_ single threaded (non-MT).
1622 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1623 // Temporarily make refs discovery atomic
1624 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1625 // Temporarily make reference _discovery_ single threaded (non-MT)
1626 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1627
1628 ref_processor()->set_enqueuing_is_done(false);
1629 ref_processor()->enable_discovery();
1630 ref_processor()->setup_policy(clear_all_soft_refs);
1631 // If an asynchronous collection finishes, the _modUnionTable is
1632 // all clear. If we are assuming the collection from an asynchronous
1633 // collection, clear the _modUnionTable.
1634 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1635 "_modUnionTable should be clear if the baton was not passed");
1636 _modUnionTable.clear_all();
1637 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1638 "mod union for klasses should be clear if the baton was passed");
1639 _ct->klass_rem_set()->clear_mod_union();
1640
1641 // We must adjust the allocation statistics being maintained
1642 // in the free list space. We do so by reading and clearing
1643 // the sweep timer and updating the block flux rate estimates below.
1644 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1645 if (_inter_sweep_timer.is_active()) {
1646 _inter_sweep_timer.stop();
1647 // Note that we do not use this sample to update the _inter_sweep_estimate.
1648 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1649 _inter_sweep_estimate.padded_average(),
1650 _intra_sweep_estimate.padded_average());
1651 }
1652
1653 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1654 #ifdef ASSERT
1655 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1656 size_t free_size = cms_space->free();
1657 assert(free_size ==
1658 pointer_delta(cms_space->end(), cms_space->compaction_top())
1659 * HeapWordSize,
1660 "All the free space should be compacted into one chunk at top");
1661 assert(cms_space->dictionary()->total_chunk_size(
1662 debug_only(cms_space->freelistLock())) == 0 ||
1663 cms_space->totalSizeInIndexedFreeLists() == 0,
1664 "All the free space should be in a single chunk");
1665 size_t num = cms_space->totalCount();
1666 assert((free_size == 0 && num == 0) ||
1667 (free_size > 0 && (num == 1 || num == 2)),
1668 "There should be at most 2 free chunks after compaction");
1669 #endif // ASSERT
1670 _collectorState = Resetting;
1671 assert(_restart_addr == NULL,
1672 "Should have been NULL'd before baton was passed");
1673 reset(false /* == !concurrent */);
1674 _cmsGen->reset_after_compaction();
1675 _concurrent_cycles_since_last_unload = 0;
1676
1677 // Clear any data recorded in the PLAB chunk arrays.
1678 if (_survivor_plab_array != NULL) {
1679 reset_survivor_plab_arrays();
1680 }
1681
1682 // Adjust the per-size allocation stats for the next epoch.
1683 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1684 // Restart the "inter sweep timer" for the next epoch.
1685 _inter_sweep_timer.reset();
1686 _inter_sweep_timer.start();
1687
1688 gc_timer->register_gc_end();
1689
1690 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1691
1692 // For a mark-sweep-compact, compute_new_size() will be called
1693 // in the heap's do_collection() method.
1694 }
1695
1696 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1697 ContiguousSpace* eden_space = _young_gen->eden();
1698 ContiguousSpace* from_space = _young_gen->from();
1699 ContiguousSpace* to_space = _young_gen->to();
1700 // Eden
1701 if (_eden_chunk_array != NULL) {
1702 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1703 p2i(eden_space->bottom()), p2i(eden_space->top()),
1704 p2i(eden_space->end()), eden_space->capacity());
1705 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1706 "_eden_chunk_capacity=" SIZE_FORMAT,
1707 _eden_chunk_index, _eden_chunk_capacity);
1708 for (size_t i = 0; i < _eden_chunk_index; i++) {
1709 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1710 i, p2i(_eden_chunk_array[i]));
1711 }
1712 }
1713 // Survivor
1714 if (_survivor_chunk_array != NULL) {
1715 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1716 p2i(from_space->bottom()), p2i(from_space->top()),
1717 p2i(from_space->end()), from_space->capacity());
1718 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1719 "_survivor_chunk_capacity=" SIZE_FORMAT,
1720 _survivor_chunk_index, _survivor_chunk_capacity);
1721 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1722 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1723 i, p2i(_survivor_chunk_array[i]));
1724 }
1725 }
1726 }
1727
1728 void CMSCollector::getFreelistLocks() const {
1729 // Get locks for all free lists in all generations that this
1730 // collector is responsible for
1731 _cmsGen->freelistLock()->lock_without_safepoint_check();
1732 }
1733
1734 void CMSCollector::releaseFreelistLocks() const {
1735 // Release locks for all free lists in all generations that this
1736 // collector is responsible for
1737 _cmsGen->freelistLock()->unlock();
1738 }
1739
1740 bool CMSCollector::haveFreelistLocks() const {
1741 // Check locks for all free lists in all generations that this
1742 // collector is responsible for
1743 assert_lock_strong(_cmsGen->freelistLock());
1744 PRODUCT_ONLY(ShouldNotReachHere());
1745 return true;
1746 }
1747
1748 // A utility class that is used by the CMS collector to
1749 // temporarily "release" the foreground collector from its
1750 // usual obligation to wait for the background collector to
1751 // complete an ongoing phase before proceeding.
1752 class ReleaseForegroundGC: public StackObj {
1753 private:
1754 CMSCollector* _c;
1755 public:
1756 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1757 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1758 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1759 // allow a potentially blocked foreground collector to proceed
1760 _c->_foregroundGCShouldWait = false;
1761 if (_c->_foregroundGCIsActive) {
1762 CGC_lock->notify();
1763 }
1764 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1765 "Possible deadlock");
1766 }
1767
1768 ~ReleaseForegroundGC() {
1769 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1770 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1771 _c->_foregroundGCShouldWait = true;
1772 }
1773 };
1774
1775 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1776 assert(Thread::current()->is_ConcurrentGC_thread(),
1777 "A CMS asynchronous collection is only allowed on a CMS thread.");
1778
1779 GenCollectedHeap* gch = GenCollectedHeap::heap();
1780 {
1781 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1782 MutexLockerEx hl(Heap_lock, safepoint_check);
1783 FreelistLocker fll(this);
1784 MutexLockerEx x(CGC_lock, safepoint_check);
1785 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1786 // The foreground collector is active or we're
1787 // not using asynchronous collections. Skip this
1788 // background collection.
1789 assert(!_foregroundGCShouldWait, "Should be clear");
1790 return;
1791 } else {
1792 assert(_collectorState == Idling, "Should be idling before start.");
1793 _collectorState = InitialMarking;
1794 register_gc_start(cause);
1795 // Reset the expansion cause, now that we are about to begin
1796 // a new cycle.
1797 clear_expansion_cause();
1798
1799 // Clear the MetaspaceGC flag since a concurrent collection
1800 // is starting but also clear it after the collection.
1801 MetaspaceGC::set_should_concurrent_collect(false);
1802 }
1803 // Decide if we want to enable class unloading as part of the
1804 // ensuing concurrent GC cycle.
1805 update_should_unload_classes();
1806 _full_gc_requested = false; // acks all outstanding full gc requests
1807 _full_gc_cause = GCCause::_no_gc;
1808 // Signal that we are about to start a collection
1809 gch->increment_total_full_collections(); // ... starting a collection cycle
1810 _collection_count_start = gch->total_full_collections();
1811 }
1812
1813 // Used for PrintGC
1814 size_t prev_used;
1815 if (PrintGC && Verbose) {
1816 prev_used = _cmsGen->used();
1817 }
1818
1819 // The change of the collection state is normally done at this level;
1820 // the exceptions are phases that are executed while the world is
1821 // stopped. For those phases the change of state is done while the
1822 // world is stopped. For baton passing purposes this allows the
1823 // background collector to finish the phase and change state atomically.
1824 // The foreground collector cannot wait on a phase that is done
1825 // while the world is stopped because the foreground collector already
1826 // has the world stopped and would deadlock.
1827 while (_collectorState != Idling) {
1828 if (TraceCMSState) {
1829 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1830 p2i(Thread::current()), _collectorState);
1831 }
1832 // The foreground collector
1833 // holds the Heap_lock throughout its collection.
1834 // holds the CMS token (but not the lock)
1835 // except while it is waiting for the background collector to yield.
1836 //
1837 // The foreground collector should be blocked (not for long)
1838 // if the background collector is about to start a phase
1839 // executed with world stopped. If the background
1840 // collector has already started such a phase, the
1841 // foreground collector is blocked waiting for the
1842 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1843 // are executed in the VM thread.
1844 //
1845 // The locking order is
1846 // PendingListLock (PLL) -- if applicable (FinalMarking)
1847 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1848 // CMS token (claimed in
1849 // stop_world_and_do() -->
1850 // safepoint_synchronize() -->
1851 // CMSThread::synchronize())
1852
1853 {
1854 // Check if the FG collector wants us to yield.
1855 CMSTokenSync x(true); // is cms thread
1856 if (waitForForegroundGC()) {
1857 // We yielded to a foreground GC, nothing more to be
1858 // done this round.
1859 assert(_foregroundGCShouldWait == false, "We set it to false in "
1860 "waitForForegroundGC()");
1861 if (TraceCMSState) {
1862 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1863 " exiting collection CMS state %d",
1864 p2i(Thread::current()), _collectorState);
1865 }
1866 return;
1867 } else {
1868 // The background collector can run but check to see if the
1869 // foreground collector has done a collection while the
1870 // background collector was waiting to get the CGC_lock
1871 // above. If yes, break so that _foregroundGCShouldWait
1872 // is cleared before returning.
1873 if (_collectorState == Idling) {
1874 break;
1875 }
1876 }
1877 }
1878
1879 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1880 "should be waiting");
1881
1882 switch (_collectorState) {
1883 case InitialMarking:
1884 {
1885 ReleaseForegroundGC x(this);
1886 stats().record_cms_begin();
1887 VM_CMS_Initial_Mark initial_mark_op(this);
1888 VMThread::execute(&initial_mark_op);
1889 }
1890 // The collector state may be any legal state at this point
1891 // since the background collector may have yielded to the
1892 // foreground collector.
1893 break;
1894 case Marking:
1895 // initial marking in checkpointRootsInitialWork has been completed
1896 if (markFromRoots()) { // we were successful
1897 assert(_collectorState == Precleaning, "Collector state should "
1898 "have changed");
1899 } else {
1900 assert(_foregroundGCIsActive, "Internal state inconsistency");
1901 }
1902 break;
1903 case Precleaning:
1904 // marking from roots in markFromRoots has been completed
1905 preclean();
1906 assert(_collectorState == AbortablePreclean ||
1907 _collectorState == FinalMarking,
1908 "Collector state should have changed");
1909 break;
1910 case AbortablePreclean:
1911 abortable_preclean();
1912 assert(_collectorState == FinalMarking, "Collector state should "
1913 "have changed");
1914 break;
1915 case FinalMarking:
1916 {
1917 ReleaseForegroundGC x(this);
1918
1919 VM_CMS_Final_Remark final_remark_op(this);
1920 VMThread::execute(&final_remark_op);
1921 }
1922 assert(_foregroundGCShouldWait, "block post-condition");
1923 break;
1924 case Sweeping:
1925 // final marking in checkpointRootsFinal has been completed
1926 sweep();
1927 assert(_collectorState == Resizing, "Collector state change "
1928 "to Resizing must be done under the free_list_lock");
1929
1930 case Resizing: {
1931 // Sweeping has been completed...
1932 // At this point the background collection has completed.
1933 // Don't move the call to compute_new_size() down
1934 // into code that might be executed if the background
1935 // collection was preempted.
1936 {
1937 ReleaseForegroundGC x(this); // unblock FG collection
1938 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1939 CMSTokenSync z(true); // not strictly needed.
1940 if (_collectorState == Resizing) {
1941 compute_new_size();
1942 save_heap_summary();
1943 _collectorState = Resetting;
1944 } else {
1945 assert(_collectorState == Idling, "The state should only change"
1946 " because the foreground collector has finished the collection");
1947 }
1948 }
1949 break;
1950 }
1951 case Resetting:
1952 // CMS heap resizing has been completed
1953 reset(true);
1954 assert(_collectorState == Idling, "Collector state should "
1955 "have changed");
1956
1957 MetaspaceGC::set_should_concurrent_collect(false);
1958
1959 stats().record_cms_end();
1960 // Don't move the concurrent_phases_end() and compute_new_size()
1961 // calls to here because a preempted background collection
1962 // has it's state set to "Resetting".
1963 break;
1964 case Idling:
1965 default:
1966 ShouldNotReachHere();
1967 break;
1968 }
1969 if (TraceCMSState) {
1970 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1971 p2i(Thread::current()), _collectorState);
1972 }
1973 assert(_foregroundGCShouldWait, "block post-condition");
1974 }
1975
1976 // Should this be in gc_epilogue?
1977 collector_policy()->counters()->update_counters();
1978
1979 {
1980 // Clear _foregroundGCShouldWait and, in the event that the
1981 // foreground collector is waiting, notify it, before
1982 // returning.
1983 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1984 _foregroundGCShouldWait = false;
1985 if (_foregroundGCIsActive) {
1986 CGC_lock->notify();
1987 }
1988 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1989 "Possible deadlock");
1990 }
1991 if (TraceCMSState) {
1992 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1993 " exiting collection CMS state %d",
1994 p2i(Thread::current()), _collectorState);
1995 }
1996 if (PrintGC && Verbose) {
1997 _cmsGen->print_heap_change(prev_used);
1998 }
1999 }
2000
2001 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2002 _cms_start_registered = true;
2003 _gc_timer_cm->register_gc_start();
2004 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2005 }
2006
2007 void CMSCollector::register_gc_end() {
2008 if (_cms_start_registered) {
2009 report_heap_summary(GCWhen::AfterGC);
2010
2011 _gc_timer_cm->register_gc_end();
2012 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2013 _cms_start_registered = false;
2014 }
2015 }
2016
2017 void CMSCollector::save_heap_summary() {
2018 GenCollectedHeap* gch = GenCollectedHeap::heap();
2019 _last_heap_summary = gch->create_heap_summary();
2020 _last_metaspace_summary = gch->create_metaspace_summary();
2021 }
2022
2023 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2024 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2025 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2026 }
2027
2028 bool CMSCollector::waitForForegroundGC() {
2029 bool res = false;
2030 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2031 "CMS thread should have CMS token");
2032 // Block the foreground collector until the
2033 // background collectors decides whether to
2034 // yield.
2035 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2036 _foregroundGCShouldWait = true;
2037 if (_foregroundGCIsActive) {
2038 // The background collector yields to the
2039 // foreground collector and returns a value
2040 // indicating that it has yielded. The foreground
2041 // collector can proceed.
2042 res = true;
2043 _foregroundGCShouldWait = false;
2044 ConcurrentMarkSweepThread::clear_CMS_flag(
2045 ConcurrentMarkSweepThread::CMS_cms_has_token);
2046 ConcurrentMarkSweepThread::set_CMS_flag(
2047 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2048 // Get a possibly blocked foreground thread going
2049 CGC_lock->notify();
2050 if (TraceCMSState) {
2051 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2052 p2i(Thread::current()), _collectorState);
2053 }
2054 while (_foregroundGCIsActive) {
2055 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2056 }
2057 ConcurrentMarkSweepThread::set_CMS_flag(
2058 ConcurrentMarkSweepThread::CMS_cms_has_token);
2059 ConcurrentMarkSweepThread::clear_CMS_flag(
2060 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2061 }
2062 if (TraceCMSState) {
2063 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2064 p2i(Thread::current()), _collectorState);
2065 }
2066 return res;
2067 }
2068
2069 // Because of the need to lock the free lists and other structures in
2070 // the collector, common to all the generations that the collector is
2071 // collecting, we need the gc_prologues of individual CMS generations
2072 // delegate to their collector. It may have been simpler had the
2073 // current infrastructure allowed one to call a prologue on a
2074 // collector. In the absence of that we have the generation's
2075 // prologue delegate to the collector, which delegates back
2076 // some "local" work to a worker method in the individual generations
2077 // that it's responsible for collecting, while itself doing any
2078 // work common to all generations it's responsible for. A similar
2079 // comment applies to the gc_epilogue()'s.
2080 // The role of the variable _between_prologue_and_epilogue is to
2081 // enforce the invocation protocol.
2082 void CMSCollector::gc_prologue(bool full) {
2083 // Call gc_prologue_work() for the CMSGen
2084 // we are responsible for.
2085
2086 // The following locking discipline assumes that we are only called
2087 // when the world is stopped.
2088 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2089
2090 // The CMSCollector prologue must call the gc_prologues for the
2091 // "generations" that it's responsible
2092 // for.
2093
2094 assert( Thread::current()->is_VM_thread()
2095 || ( CMSScavengeBeforeRemark
2096 && Thread::current()->is_ConcurrentGC_thread()),
2097 "Incorrect thread type for prologue execution");
2098
2099 if (_between_prologue_and_epilogue) {
2100 // We have already been invoked; this is a gc_prologue delegation
2101 // from yet another CMS generation that we are responsible for, just
2102 // ignore it since all relevant work has already been done.
2103 return;
2104 }
2105
2106 // set a bit saying prologue has been called; cleared in epilogue
2107 _between_prologue_and_epilogue = true;
2108 // Claim locks for common data structures, then call gc_prologue_work()
2109 // for each CMSGen.
2110
2111 getFreelistLocks(); // gets free list locks on constituent spaces
2112 bitMapLock()->lock_without_safepoint_check();
2113
2114 // Should call gc_prologue_work() for all cms gens we are responsible for
2115 bool duringMarking = _collectorState >= Marking
2116 && _collectorState < Sweeping;
2117
2118 // The young collections clear the modified oops state, which tells if
2119 // there are any modified oops in the class. The remark phase also needs
2120 // that information. Tell the young collection to save the union of all
2121 // modified klasses.
2122 if (duringMarking) {
2123 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2124 }
2125
2126 bool registerClosure = duringMarking;
2127
2128 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2129
2130 if (!full) {
2131 stats().record_gc0_begin();
2132 }
2133 }
2134
2135 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2136
2137 _capacity_at_prologue = capacity();
2138 _used_at_prologue = used();
2139
2140 // Delegate to CMScollector which knows how to coordinate between
2141 // this and any other CMS generations that it is responsible for
2142 // collecting.
2143 collector()->gc_prologue(full);
2144 }
2145
2146 // This is a "private" interface for use by this generation's CMSCollector.
2147 // Not to be called directly by any other entity (for instance,
2148 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2149 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2150 bool registerClosure, ModUnionClosure* modUnionClosure) {
2151 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2152 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2153 "Should be NULL");
2154 if (registerClosure) {
2155 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2156 }
2157 cmsSpace()->gc_prologue();
2158 // Clear stat counters
2159 NOT_PRODUCT(
2160 assert(_numObjectsPromoted == 0, "check");
2161 assert(_numWordsPromoted == 0, "check");
2162 if (Verbose && PrintGC) {
2163 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2164 SIZE_FORMAT" bytes concurrently",
2165 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2166 }
2167 _numObjectsAllocated = 0;
2168 _numWordsAllocated = 0;
2169 )
2170 }
2171
2172 void CMSCollector::gc_epilogue(bool full) {
2173 // The following locking discipline assumes that we are only called
2174 // when the world is stopped.
2175 assert(SafepointSynchronize::is_at_safepoint(),
2176 "world is stopped assumption");
2177
2178 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2179 // if linear allocation blocks need to be appropriately marked to allow the
2180 // the blocks to be parsable. We also check here whether we need to nudge the
2181 // CMS collector thread to start a new cycle (if it's not already active).
2182 assert( Thread::current()->is_VM_thread()
2183 || ( CMSScavengeBeforeRemark
2184 && Thread::current()->is_ConcurrentGC_thread()),
2185 "Incorrect thread type for epilogue execution");
2186
2187 if (!_between_prologue_and_epilogue) {
2188 // We have already been invoked; this is a gc_epilogue delegation
2189 // from yet another CMS generation that we are responsible for, just
2190 // ignore it since all relevant work has already been done.
2191 return;
2192 }
2193 assert(haveFreelistLocks(), "must have freelist locks");
2194 assert_lock_strong(bitMapLock());
2195
2196 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2197
2198 _cmsGen->gc_epilogue_work(full);
2199
2200 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2201 // in case sampling was not already enabled, enable it
2202 _start_sampling = true;
2203 }
2204 // reset _eden_chunk_array so sampling starts afresh
2205 _eden_chunk_index = 0;
2206
2207 size_t cms_used = _cmsGen->cmsSpace()->used();
2208
2209 // update performance counters - this uses a special version of
2210 // update_counters() that allows the utilization to be passed as a
2211 // parameter, avoiding multiple calls to used().
2212 //
2213 _cmsGen->update_counters(cms_used);
2214
2215 bitMapLock()->unlock();
2216 releaseFreelistLocks();
2217
2218 if (!CleanChunkPoolAsync) {
2219 Chunk::clean_chunk_pool();
2220 }
2221
2222 set_did_compact(false);
2223 _between_prologue_and_epilogue = false; // ready for next cycle
2224 }
2225
2226 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2227 collector()->gc_epilogue(full);
2228
2229 // Also reset promotion tracking in par gc thread states.
2230 for (uint i = 0; i < ParallelGCThreads; i++) {
2231 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2232 }
2233 }
2234
2235 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2236 assert(!incremental_collection_failed(), "Should have been cleared");
2237 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2238 cmsSpace()->gc_epilogue();
2239 // Print stat counters
2240 NOT_PRODUCT(
2241 assert(_numObjectsAllocated == 0, "check");
2242 assert(_numWordsAllocated == 0, "check");
2243 if (Verbose && PrintGC) {
2244 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2245 SIZE_FORMAT" bytes",
2246 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2247 }
2248 _numObjectsPromoted = 0;
2249 _numWordsPromoted = 0;
2250 )
2251
2252 if (PrintGC && Verbose) {
2253 // Call down the chain in contiguous_available needs the freelistLock
2254 // so print this out before releasing the freeListLock.
2255 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2256 contiguous_available());
2257 }
2258 }
2259
2260 #ifndef PRODUCT
2261 bool CMSCollector::have_cms_token() {
2262 Thread* thr = Thread::current();
2263 if (thr->is_VM_thread()) {
2264 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2265 } else if (thr->is_ConcurrentGC_thread()) {
2266 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2267 } else if (thr->is_GC_task_thread()) {
2268 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2269 ParGCRareEvent_lock->owned_by_self();
2270 }
2271 return false;
2272 }
2273 #endif
2274
2275 // Check reachability of the given heap address in CMS generation,
2276 // treating all other generations as roots.
2277 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2278 // We could "guarantee" below, rather than assert, but I'll
2279 // leave these as "asserts" so that an adventurous debugger
2280 // could try this in the product build provided some subset of
2281 // the conditions were met, provided they were interested in the
2282 // results and knew that the computation below wouldn't interfere
2283 // with other concurrent computations mutating the structures
2284 // being read or written.
2285 assert(SafepointSynchronize::is_at_safepoint(),
2286 "Else mutations in object graph will make answer suspect");
2287 assert(have_cms_token(), "Should hold cms token");
2288 assert(haveFreelistLocks(), "must hold free list locks");
2289 assert_lock_strong(bitMapLock());
2290
2291 // Clear the marking bit map array before starting, but, just
2292 // for kicks, first report if the given address is already marked
2293 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2294 _markBitMap.isMarked(addr) ? "" : " not");
2295
2296 if (verify_after_remark()) {
2297 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2298 bool result = verification_mark_bm()->isMarked(addr);
2299 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2300 result ? "IS" : "is NOT");
2301 return result;
2302 } else {
2303 gclog_or_tty->print_cr("Could not compute result");
2304 return false;
2305 }
2306 }
2307
2308
2309 void
2310 CMSCollector::print_on_error(outputStream* st) {
2311 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2312 if (collector != NULL) {
2313 CMSBitMap* bitmap = &collector->_markBitMap;
2314 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2315 bitmap->print_on_error(st, " Bits: ");
2316
2317 st->cr();
2318
2319 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2320 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2321 mut_bitmap->print_on_error(st, " Bits: ");
2322 }
2323 }
2324
2325 ////////////////////////////////////////////////////////
2326 // CMS Verification Support
2327 ////////////////////////////////////////////////////////
2328 // Following the remark phase, the following invariant
2329 // should hold -- each object in the CMS heap which is
2330 // marked in markBitMap() should be marked in the verification_mark_bm().
2331
2332 class VerifyMarkedClosure: public BitMapClosure {
2333 CMSBitMap* _marks;
2334 bool _failed;
2335
2336 public:
2337 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2338
2339 bool do_bit(size_t offset) {
2340 HeapWord* addr = _marks->offsetToHeapWord(offset);
2341 if (!_marks->isMarked(addr)) {
2342 oop(addr)->print_on(gclog_or_tty);
2343 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", p2i(addr));
2344 _failed = true;
2345 }
2346 return true;
2347 }
2348
2349 bool failed() { return _failed; }
2350 };
2351
2352 bool CMSCollector::verify_after_remark(bool silent) {
2353 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2354 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2355 static bool init = false;
2356
2357 assert(SafepointSynchronize::is_at_safepoint(),
2358 "Else mutations in object graph will make answer suspect");
2359 assert(have_cms_token(),
2360 "Else there may be mutual interference in use of "
2361 " verification data structures");
2362 assert(_collectorState > Marking && _collectorState <= Sweeping,
2363 "Else marking info checked here may be obsolete");
2364 assert(haveFreelistLocks(), "must hold free list locks");
2365 assert_lock_strong(bitMapLock());
2366
2367
2368 // Allocate marking bit map if not already allocated
2369 if (!init) { // first time
2370 if (!verification_mark_bm()->allocate(_span)) {
2371 return false;
2372 }
2373 init = true;
2374 }
2375
2376 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2377
2378 // Turn off refs discovery -- so we will be tracing through refs.
2379 // This is as intended, because by this time
2380 // GC must already have cleared any refs that need to be cleared,
2381 // and traced those that need to be marked; moreover,
2382 // the marking done here is not going to interfere in any
2383 // way with the marking information used by GC.
2384 NoRefDiscovery no_discovery(ref_processor());
2385
2386 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2387
2388 // Clear any marks from a previous round
2389 verification_mark_bm()->clear_all();
2390 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2391 verify_work_stacks_empty();
2392
2393 GenCollectedHeap* gch = GenCollectedHeap::heap();
2394 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2395 // Update the saved marks which may affect the root scans.
2396 gch->save_marks();
2397
2398 if (CMSRemarkVerifyVariant == 1) {
2399 // In this first variant of verification, we complete
2400 // all marking, then check if the new marks-vector is
2401 // a subset of the CMS marks-vector.
2402 verify_after_remark_work_1();
2403 } else if (CMSRemarkVerifyVariant == 2) {
2404 // In this second variant of verification, we flag an error
2405 // (i.e. an object reachable in the new marks-vector not reachable
2406 // in the CMS marks-vector) immediately, also indicating the
2407 // identify of an object (A) that references the unmarked object (B) --
2408 // presumably, a mutation to A failed to be picked up by preclean/remark?
2409 verify_after_remark_work_2();
2410 } else {
2411 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2412 CMSRemarkVerifyVariant);
2413 }
2414 if (!silent) gclog_or_tty->print(" done] ");
2415 return true;
2416 }
2417
2418 void CMSCollector::verify_after_remark_work_1() {
2419 ResourceMark rm;
2420 HandleMark hm;
2421 GenCollectedHeap* gch = GenCollectedHeap::heap();
2422
2423 // Get a clear set of claim bits for the roots processing to work with.
2424 ClassLoaderDataGraph::clear_claimed_marks();
2425
2426 // Mark from roots one level into CMS
2427 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2428 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2429
2430 {
2431 StrongRootsScope srs(1);
2432
2433 gch->gen_process_roots(&srs,
2434 Generation::Old,
2435 true, // younger gens are roots
2436 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2437 should_unload_classes(),
2438 ¬Older,
2439 NULL,
2440 NULL);
2441 }
2442
2443 // Now mark from the roots
2444 MarkFromRootsClosure markFromRootsClosure(this, _span,
2445 verification_mark_bm(), verification_mark_stack(),
2446 false /* don't yield */, true /* verifying */);
2447 assert(_restart_addr == NULL, "Expected pre-condition");
2448 verification_mark_bm()->iterate(&markFromRootsClosure);
2449 while (_restart_addr != NULL) {
2450 // Deal with stack overflow: by restarting at the indicated
2451 // address.
2452 HeapWord* ra = _restart_addr;
2453 markFromRootsClosure.reset(ra);
2454 _restart_addr = NULL;
2455 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2456 }
2457 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2458 verify_work_stacks_empty();
2459
2460 // Marking completed -- now verify that each bit marked in
2461 // verification_mark_bm() is also marked in markBitMap(); flag all
2462 // errors by printing corresponding objects.
2463 VerifyMarkedClosure vcl(markBitMap());
2464 verification_mark_bm()->iterate(&vcl);
2465 if (vcl.failed()) {
2466 gclog_or_tty->print("Verification failed");
2467 gch->print_on(gclog_or_tty);
2468 fatal("CMS: failed marking verification after remark");
2469 }
2470 }
2471
2472 class VerifyKlassOopsKlassClosure : public KlassClosure {
2473 class VerifyKlassOopsClosure : public OopClosure {
2474 CMSBitMap* _bitmap;
2475 public:
2476 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2477 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2478 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2479 } _oop_closure;
2480 public:
2481 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2482 void do_klass(Klass* k) {
2483 k->oops_do(&_oop_closure);
2484 }
2485 };
2486
2487 void CMSCollector::verify_after_remark_work_2() {
2488 ResourceMark rm;
2489 HandleMark hm;
2490 GenCollectedHeap* gch = GenCollectedHeap::heap();
2491
2492 // Get a clear set of claim bits for the roots processing to work with.
2493 ClassLoaderDataGraph::clear_claimed_marks();
2494
2495 // Mark from roots one level into CMS
2496 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2497 markBitMap());
2498 CLDToOopClosure cld_closure(¬Older, true);
2499
2500 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2501
2502 {
2503 StrongRootsScope srs(1);
2504
2505 gch->gen_process_roots(&srs,
2506 Generation::Old,
2507 true, // younger gens are roots
2508 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2509 should_unload_classes(),
2510 ¬Older,
2511 NULL,
2512 &cld_closure);
2513 }
2514
2515 // Now mark from the roots
2516 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2517 verification_mark_bm(), markBitMap(), verification_mark_stack());
2518 assert(_restart_addr == NULL, "Expected pre-condition");
2519 verification_mark_bm()->iterate(&markFromRootsClosure);
2520 while (_restart_addr != NULL) {
2521 // Deal with stack overflow: by restarting at the indicated
2522 // address.
2523 HeapWord* ra = _restart_addr;
2524 markFromRootsClosure.reset(ra);
2525 _restart_addr = NULL;
2526 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2527 }
2528 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2529 verify_work_stacks_empty();
2530
2531 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2532 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2533
2534 // Marking completed -- now verify that each bit marked in
2535 // verification_mark_bm() is also marked in markBitMap(); flag all
2536 // errors by printing corresponding objects.
2537 VerifyMarkedClosure vcl(markBitMap());
2538 verification_mark_bm()->iterate(&vcl);
2539 assert(!vcl.failed(), "Else verification above should not have succeeded");
2540 }
2541
2542 void ConcurrentMarkSweepGeneration::save_marks() {
2543 // delegate to CMS space
2544 cmsSpace()->save_marks();
2545 for (uint i = 0; i < ParallelGCThreads; i++) {
2546 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2547 }
2548 }
2549
2550 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2551 return cmsSpace()->no_allocs_since_save_marks();
2552 }
2553
2554 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2555 \
2556 void ConcurrentMarkSweepGeneration:: \
2557 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2558 cl->set_generation(this); \
2559 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2560 cl->reset_generation(); \
2561 save_marks(); \
2562 }
2563
2564 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2565
2566 void
2567 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2568 if (freelistLock()->owned_by_self()) {
2569 Generation::oop_iterate(cl);
2570 } else {
2571 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2572 Generation::oop_iterate(cl);
2573 }
2574 }
2575
2576 void
2577 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2578 if (freelistLock()->owned_by_self()) {
2579 Generation::object_iterate(cl);
2580 } else {
2581 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2582 Generation::object_iterate(cl);
2583 }
2584 }
2585
2586 void
2587 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2588 if (freelistLock()->owned_by_self()) {
2589 Generation::safe_object_iterate(cl);
2590 } else {
2591 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2592 Generation::safe_object_iterate(cl);
2593 }
2594 }
2595
2596 void
2597 ConcurrentMarkSweepGeneration::post_compact() {
2598 }
2599
2600 void
2601 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2602 // Fix the linear allocation blocks to look like free blocks.
2603
2604 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2605 // are not called when the heap is verified during universe initialization and
2606 // at vm shutdown.
2607 if (freelistLock()->owned_by_self()) {
2608 cmsSpace()->prepare_for_verify();
2609 } else {
2610 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2611 cmsSpace()->prepare_for_verify();
2612 }
2613 }
2614
2615 void
2616 ConcurrentMarkSweepGeneration::verify() {
2617 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2618 // are not called when the heap is verified during universe initialization and
2619 // at vm shutdown.
2620 if (freelistLock()->owned_by_self()) {
2621 cmsSpace()->verify();
2622 } else {
2623 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2624 cmsSpace()->verify();
2625 }
2626 }
2627
2628 void CMSCollector::verify() {
2629 _cmsGen->verify();
2630 }
2631
2632 #ifndef PRODUCT
2633 bool CMSCollector::overflow_list_is_empty() const {
2634 assert(_num_par_pushes >= 0, "Inconsistency");
2635 if (_overflow_list == NULL) {
2636 assert(_num_par_pushes == 0, "Inconsistency");
2637 }
2638 return _overflow_list == NULL;
2639 }
2640
2641 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2642 // merely consolidate assertion checks that appear to occur together frequently.
2643 void CMSCollector::verify_work_stacks_empty() const {
2644 assert(_markStack.isEmpty(), "Marking stack should be empty");
2645 assert(overflow_list_is_empty(), "Overflow list should be empty");
2646 }
2647
2648 void CMSCollector::verify_overflow_empty() const {
2649 assert(overflow_list_is_empty(), "Overflow list should be empty");
2650 assert(no_preserved_marks(), "No preserved marks");
2651 }
2652 #endif // PRODUCT
2653
2654 // Decide if we want to enable class unloading as part of the
2655 // ensuing concurrent GC cycle. We will collect and
2656 // unload classes if it's the case that:
2657 // (1) an explicit gc request has been made and the flag
2658 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2659 // (2) (a) class unloading is enabled at the command line, and
2660 // (b) old gen is getting really full
2661 // NOTE: Provided there is no change in the state of the heap between
2662 // calls to this method, it should have idempotent results. Moreover,
2663 // its results should be monotonically increasing (i.e. going from 0 to 1,
2664 // but not 1 to 0) between successive calls between which the heap was
2665 // not collected. For the implementation below, it must thus rely on
2666 // the property that concurrent_cycles_since_last_unload()
2667 // will not decrease unless a collection cycle happened and that
2668 // _cmsGen->is_too_full() are
2669 // themselves also monotonic in that sense. See check_monotonicity()
2670 // below.
2671 void CMSCollector::update_should_unload_classes() {
2672 _should_unload_classes = false;
2673 // Condition 1 above
2674 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2675 _should_unload_classes = true;
2676 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2677 // Disjuncts 2.b.(i,ii,iii) above
2678 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2679 CMSClassUnloadingMaxInterval)
2680 || _cmsGen->is_too_full();
2681 }
2682 }
2683
2684 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2685 bool res = should_concurrent_collect();
2686 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2687 return res;
2688 }
2689
2690 void CMSCollector::setup_cms_unloading_and_verification_state() {
2691 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2692 || VerifyBeforeExit;
2693 const int rso = GenCollectedHeap::SO_AllCodeCache;
2694
2695 // We set the proper root for this CMS cycle here.
2696 if (should_unload_classes()) { // Should unload classes this cycle
2697 remove_root_scanning_option(rso); // Shrink the root set appropriately
2698 set_verifying(should_verify); // Set verification state for this cycle
2699 return; // Nothing else needs to be done at this time
2700 }
2701
2702 // Not unloading classes this cycle
2703 assert(!should_unload_classes(), "Inconsistency!");
2704
2705 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2706 // Include symbols, strings and code cache elements to prevent their resurrection.
2707 add_root_scanning_option(rso);
2708 set_verifying(true);
2709 } else if (verifying() && !should_verify) {
2710 // We were verifying, but some verification flags got disabled.
2711 set_verifying(false);
2712 // Exclude symbols, strings and code cache elements from root scanning to
2713 // reduce IM and RM pauses.
2714 remove_root_scanning_option(rso);
2715 }
2716 }
2717
2718
2719 #ifndef PRODUCT
2720 HeapWord* CMSCollector::block_start(const void* p) const {
2721 const HeapWord* addr = (HeapWord*)p;
2722 if (_span.contains(p)) {
2723 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2724 return _cmsGen->cmsSpace()->block_start(p);
2725 }
2726 }
2727 return NULL;
2728 }
2729 #endif
2730
2731 HeapWord*
2732 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2733 bool tlab,
2734 bool parallel) {
2735 CMSSynchronousYieldRequest yr;
2736 assert(!tlab, "Can't deal with TLAB allocation");
2737 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2738 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2739 if (GCExpandToAllocateDelayMillis > 0) {
2740 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2741 }
2742 return have_lock_and_allocate(word_size, tlab);
2743 }
2744
2745 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2746 size_t bytes,
2747 size_t expand_bytes,
2748 CMSExpansionCause::Cause cause)
2749 {
2750
2751 bool success = expand(bytes, expand_bytes);
2752
2753 // remember why we expanded; this information is used
2754 // by shouldConcurrentCollect() when making decisions on whether to start
2755 // a new CMS cycle.
2756 if (success) {
2757 set_expansion_cause(cause);
2758 if (PrintGCDetails && Verbose) {
2759 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2760 CMSExpansionCause::to_string(cause));
2761 }
2762 }
2763 }
2764
2765 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2766 HeapWord* res = NULL;
2767 MutexLocker x(ParGCRareEvent_lock);
2768 while (true) {
2769 // Expansion by some other thread might make alloc OK now:
2770 res = ps->lab.alloc(word_sz);
2771 if (res != NULL) return res;
2772 // If there's not enough expansion space available, give up.
2773 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2774 return NULL;
2775 }
2776 // Otherwise, we try expansion.
2777 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2778 // Now go around the loop and try alloc again;
2779 // A competing par_promote might beat us to the expansion space,
2780 // so we may go around the loop again if promotion fails again.
2781 if (GCExpandToAllocateDelayMillis > 0) {
2782 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2783 }
2784 }
2785 }
2786
2787
2788 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2789 PromotionInfo* promo) {
2790 MutexLocker x(ParGCRareEvent_lock);
2791 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2792 while (true) {
2793 // Expansion by some other thread might make alloc OK now:
2794 if (promo->ensure_spooling_space()) {
2795 assert(promo->has_spooling_space(),
2796 "Post-condition of successful ensure_spooling_space()");
2797 return true;
2798 }
2799 // If there's not enough expansion space available, give up.
2800 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2801 return false;
2802 }
2803 // Otherwise, we try expansion.
2804 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2805 // Now go around the loop and try alloc again;
2806 // A competing allocation might beat us to the expansion space,
2807 // so we may go around the loop again if allocation fails again.
2808 if (GCExpandToAllocateDelayMillis > 0) {
2809 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2810 }
2811 }
2812 }
2813
2814 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2815 // Only shrink if a compaction was done so that all the free space
2816 // in the generation is in a contiguous block at the end.
2817 if (did_compact()) {
2818 CardGeneration::shrink(bytes);
2819 }
2820 }
2821
2822 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2823 assert_locked_or_safepoint(Heap_lock);
2824 }
2825
2826 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2827 assert_locked_or_safepoint(Heap_lock);
2828 assert_lock_strong(freelistLock());
2829 if (PrintGCDetails && Verbose) {
2830 warning("Shrinking of CMS not yet implemented");
2831 }
2832 return;
2833 }
2834
2835
2836 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2837 // phases.
2838 class CMSPhaseAccounting: public StackObj {
2839 public:
2840 CMSPhaseAccounting(CMSCollector *collector,
2841 const char *phase,
2842 const GCId gc_id,
2843 bool print_cr = true);
2844 ~CMSPhaseAccounting();
2845
2846 private:
2847 CMSCollector *_collector;
2848 const char *_phase;
2849 elapsedTimer _wallclock;
2850 bool _print_cr;
2851 const GCId _gc_id;
2852
2853 public:
2854 // Not MT-safe; so do not pass around these StackObj's
2855 // where they may be accessed by other threads.
2856 jlong wallclock_millis() {
2857 assert(_wallclock.is_active(), "Wall clock should not stop");
2858 _wallclock.stop(); // to record time
2859 jlong ret = _wallclock.milliseconds();
2860 _wallclock.start(); // restart
2861 return ret;
2862 }
2863 };
2864
2865 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2866 const char *phase,
2867 const GCId gc_id,
2868 bool print_cr) :
2869 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2870
2871 if (PrintCMSStatistics != 0) {
2872 _collector->resetYields();
2873 }
2874 if (PrintGCDetails) {
2875 gclog_or_tty->gclog_stamp(_gc_id);
2876 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2877 _collector->cmsGen()->short_name(), _phase);
2878 }
2879 _collector->resetTimer();
2880 _wallclock.start();
2881 _collector->startTimer();
2882 }
2883
2884 CMSPhaseAccounting::~CMSPhaseAccounting() {
2885 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2886 _collector->stopTimer();
2887 _wallclock.stop();
2888 if (PrintGCDetails) {
2889 gclog_or_tty->gclog_stamp(_gc_id);
2890 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2891 _collector->cmsGen()->short_name(),
2892 _phase, _collector->timerValue(), _wallclock.seconds());
2893 if (_print_cr) {
2894 gclog_or_tty->cr();
2895 }
2896 if (PrintCMSStatistics != 0) {
2897 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2898 _collector->yields());
2899 }
2900 }
2901 }
2902
2903 // CMS work
2904
2905 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2906 class CMSParMarkTask : public AbstractGangTask {
2907 protected:
2908 CMSCollector* _collector;
2909 uint _n_workers;
2910 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2911 AbstractGangTask(name),
2912 _collector(collector),
2913 _n_workers(n_workers) {}
2914 // Work method in support of parallel rescan ... of young gen spaces
2915 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2916 ContiguousSpace* space,
2917 HeapWord** chunk_array, size_t chunk_top);
2918 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2919 };
2920
2921 // Parallel initial mark task
2922 class CMSParInitialMarkTask: public CMSParMarkTask {
2923 StrongRootsScope* _strong_roots_scope;
2924 public:
2925 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2926 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2927 _strong_roots_scope(strong_roots_scope) {}
2928 void work(uint worker_id);
2929 };
2930
2931 // Checkpoint the roots into this generation from outside
2932 // this generation. [Note this initial checkpoint need only
2933 // be approximate -- we'll do a catch up phase subsequently.]
2934 void CMSCollector::checkpointRootsInitial() {
2935 assert(_collectorState == InitialMarking, "Wrong collector state");
2936 check_correct_thread_executing();
2937 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2938
2939 save_heap_summary();
2940 report_heap_summary(GCWhen::BeforeGC);
2941
2942 ReferenceProcessor* rp = ref_processor();
2943 assert(_restart_addr == NULL, "Control point invariant");
2944 {
2945 // acquire locks for subsequent manipulations
2946 MutexLockerEx x(bitMapLock(),
2947 Mutex::_no_safepoint_check_flag);
2948 checkpointRootsInitialWork();
2949 // enable ("weak") refs discovery
2950 rp->enable_discovery();
2951 _collectorState = Marking;
2952 }
2953 }
2954
2955 void CMSCollector::checkpointRootsInitialWork() {
2956 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2957 assert(_collectorState == InitialMarking, "just checking");
2958
2959 // If there has not been a GC[n-1] since last GC[n] cycle completed,
2960 // precede our marking with a collection of all
2961 // younger generations to keep floating garbage to a minimum.
2962 // XXX: we won't do this for now -- it's an optimization to be done later.
2963
2964 // already have locks
2965 assert_lock_strong(bitMapLock());
2966 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2967
2968 // Setup the verification and class unloading state for this
2969 // CMS collection cycle.
2970 setup_cms_unloading_and_verification_state();
2971
2972 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2973 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
2974
2975 // Reset all the PLAB chunk arrays if necessary.
2976 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2977 reset_survivor_plab_arrays();
2978 }
2979
2980 ResourceMark rm;
2981 HandleMark hm;
2982
2983 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2984 GenCollectedHeap* gch = GenCollectedHeap::heap();
2985
2986 verify_work_stacks_empty();
2987 verify_overflow_empty();
2988
2989 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2990 // Update the saved marks which may affect the root scans.
2991 gch->save_marks();
2992
2993 // weak reference processing has not started yet.
2994 ref_processor()->set_enqueuing_is_done(false);
2995
2996 // Need to remember all newly created CLDs,
2997 // so that we can guarantee that the remark finds them.
2998 ClassLoaderDataGraph::remember_new_clds(true);
2999
3000 // Whenever a CLD is found, it will be claimed before proceeding to mark
3001 // the klasses. The claimed marks need to be cleared before marking starts.
3002 ClassLoaderDataGraph::clear_claimed_marks();
3003
3004 if (CMSPrintEdenSurvivorChunks) {
3005 print_eden_and_survivor_chunk_arrays();
3006 }
3007
3008 {
3009 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3010 if (CMSParallelInitialMarkEnabled) {
3011 // The parallel version.
3012 FlexibleWorkGang* workers = gch->workers();
3013 assert(workers != NULL, "Need parallel worker threads.");
3014 uint n_workers = workers->active_workers();
3015
3016 StrongRootsScope srs(n_workers);
3017
3018 CMSParInitialMarkTask tsk(this, &srs, n_workers);
3019 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3020 if (n_workers > 1) {
3021 workers->run_task(&tsk);
3022 } else {
3023 tsk.work(0);
3024 }
3025 } else {
3026 // The serial version.
3027 CLDToOopClosure cld_closure(¬Older, true);
3028 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3029
3030 StrongRootsScope srs(1);
3031
3032 gch->gen_process_roots(&srs,
3033 Generation::Old,
3034 true, // younger gens are roots
3035 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3036 should_unload_classes(),
3037 ¬Older,
3038 NULL,
3039 &cld_closure);
3040 }
3041 }
3042
3043 // Clear mod-union table; it will be dirtied in the prologue of
3044 // CMS generation per each younger generation collection.
3045
3046 assert(_modUnionTable.isAllClear(),
3047 "Was cleared in most recent final checkpoint phase"
3048 " or no bits are set in the gc_prologue before the start of the next "
3049 "subsequent marking phase.");
3050
3051 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3052
3053 // Save the end of the used_region of the constituent generations
3054 // to be used to limit the extent of sweep in each generation.
3055 save_sweep_limits();
3056 verify_overflow_empty();
3057 }
3058
3059 bool CMSCollector::markFromRoots() {
3060 // we might be tempted to assert that:
3061 // assert(!SafepointSynchronize::is_at_safepoint(),
3062 // "inconsistent argument?");
3063 // However that wouldn't be right, because it's possible that
3064 // a safepoint is indeed in progress as a younger generation
3065 // stop-the-world GC happens even as we mark in this generation.
3066 assert(_collectorState == Marking, "inconsistent state?");
3067 check_correct_thread_executing();
3068 verify_overflow_empty();
3069
3070 // Weak ref discovery note: We may be discovering weak
3071 // refs in this generation concurrent (but interleaved) with
3072 // weak ref discovery by a younger generation collector.
3073
3074 CMSTokenSyncWithLocks ts(true, bitMapLock());
3075 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3076 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3077 bool res = markFromRootsWork();
3078 if (res) {
3079 _collectorState = Precleaning;
3080 } else { // We failed and a foreground collection wants to take over
3081 assert(_foregroundGCIsActive, "internal state inconsistency");
3082 assert(_restart_addr == NULL, "foreground will restart from scratch");
3083 if (PrintGCDetails) {
3084 gclog_or_tty->print_cr("bailing out to foreground collection");
3085 }
3086 }
3087 verify_overflow_empty();
3088 return res;
3089 }
3090
3091 bool CMSCollector::markFromRootsWork() {
3092 // iterate over marked bits in bit map, doing a full scan and mark
3093 // from these roots using the following algorithm:
3094 // . if oop is to the right of the current scan pointer,
3095 // mark corresponding bit (we'll process it later)
3096 // . else (oop is to left of current scan pointer)
3097 // push oop on marking stack
3098 // . drain the marking stack
3099
3100 // Note that when we do a marking step we need to hold the
3101 // bit map lock -- recall that direct allocation (by mutators)
3102 // and promotion (by younger generation collectors) is also
3103 // marking the bit map. [the so-called allocate live policy.]
3104 // Because the implementation of bit map marking is not
3105 // robust wrt simultaneous marking of bits in the same word,
3106 // we need to make sure that there is no such interference
3107 // between concurrent such updates.
3108
3109 // already have locks
3110 assert_lock_strong(bitMapLock());
3111
3112 verify_work_stacks_empty();
3113 verify_overflow_empty();
3114 bool result = false;
3115 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3116 result = do_marking_mt();
3117 } else {
3118 result = do_marking_st();
3119 }
3120 return result;
3121 }
3122
3123 // Forward decl
3124 class CMSConcMarkingTask;
3125
3126 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3127 CMSCollector* _collector;
3128 CMSConcMarkingTask* _task;
3129 public:
3130 virtual void yield();
3131
3132 // "n_threads" is the number of threads to be terminated.
3133 // "queue_set" is a set of work queues of other threads.
3134 // "collector" is the CMS collector associated with this task terminator.
3135 // "yield" indicates whether we need the gang as a whole to yield.
3136 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3137 ParallelTaskTerminator(n_threads, queue_set),
3138 _collector(collector) { }
3139
3140 void set_task(CMSConcMarkingTask* task) {
3141 _task = task;
3142 }
3143 };
3144
3145 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3146 CMSConcMarkingTask* _task;
3147 public:
3148 bool should_exit_termination();
3149 void set_task(CMSConcMarkingTask* task) {
3150 _task = task;
3151 }
3152 };
3153
3154 // MT Concurrent Marking Task
3155 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3156 CMSCollector* _collector;
3157 uint _n_workers; // requested/desired # workers
3158 bool _result;
3159 CompactibleFreeListSpace* _cms_space;
3160 char _pad_front[64]; // padding to ...
3161 HeapWord* _global_finger; // ... avoid sharing cache line
3162 char _pad_back[64];
3163 HeapWord* _restart_addr;
3164
3165 // Exposed here for yielding support
3166 Mutex* const _bit_map_lock;
3167
3168 // The per thread work queues, available here for stealing
3169 OopTaskQueueSet* _task_queues;
3170
3171 // Termination (and yielding) support
3172 CMSConcMarkingTerminator _term;
3173 CMSConcMarkingTerminatorTerminator _term_term;
3174
3175 public:
3176 CMSConcMarkingTask(CMSCollector* collector,
3177 CompactibleFreeListSpace* cms_space,
3178 YieldingFlexibleWorkGang* workers,
3179 OopTaskQueueSet* task_queues):
3180 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3181 _collector(collector),
3182 _cms_space(cms_space),
3183 _n_workers(0), _result(true),
3184 _task_queues(task_queues),
3185 _term(_n_workers, task_queues, _collector),
3186 _bit_map_lock(collector->bitMapLock())
3187 {
3188 _requested_size = _n_workers;
3189 _term.set_task(this);
3190 _term_term.set_task(this);
3191 _restart_addr = _global_finger = _cms_space->bottom();
3192 }
3193
3194
3195 OopTaskQueueSet* task_queues() { return _task_queues; }
3196
3197 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3198
3199 HeapWord** global_finger_addr() { return &_global_finger; }
3200
3201 CMSConcMarkingTerminator* terminator() { return &_term; }
3202
3203 virtual void set_for_termination(uint active_workers) {
3204 terminator()->reset_for_reuse(active_workers);
3205 }
3206
3207 void work(uint worker_id);
3208 bool should_yield() {
3209 return ConcurrentMarkSweepThread::should_yield()
3210 && !_collector->foregroundGCIsActive();
3211 }
3212
3213 virtual void coordinator_yield(); // stuff done by coordinator
3214 bool result() { return _result; }
3215
3216 void reset(HeapWord* ra) {
3217 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3218 _restart_addr = _global_finger = ra;
3219 _term.reset_for_reuse();
3220 }
3221
3222 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3223 OopTaskQueue* work_q);
3224
3225 private:
3226 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3227 void do_work_steal(int i);
3228 void bump_global_finger(HeapWord* f);
3229 };
3230
3231 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3232 assert(_task != NULL, "Error");
3233 return _task->yielding();
3234 // Note that we do not need the disjunct || _task->should_yield() above
3235 // because we want terminating threads to yield only if the task
3236 // is already in the midst of yielding, which happens only after at least one
3237 // thread has yielded.
3238 }
3239
3240 void CMSConcMarkingTerminator::yield() {
3241 if (_task->should_yield()) {
3242 _task->yield();
3243 } else {
3244 ParallelTaskTerminator::yield();
3245 }
3246 }
3247
3248 ////////////////////////////////////////////////////////////////
3249 // Concurrent Marking Algorithm Sketch
3250 ////////////////////////////////////////////////////////////////
3251 // Until all tasks exhausted (both spaces):
3252 // -- claim next available chunk
3253 // -- bump global finger via CAS
3254 // -- find first object that starts in this chunk
3255 // and start scanning bitmap from that position
3256 // -- scan marked objects for oops
3257 // -- CAS-mark target, and if successful:
3258 // . if target oop is above global finger (volatile read)
3259 // nothing to do
3260 // . if target oop is in chunk and above local finger
3261 // then nothing to do
3262 // . else push on work-queue
3263 // -- Deal with possible overflow issues:
3264 // . local work-queue overflow causes stuff to be pushed on
3265 // global (common) overflow queue
3266 // . always first empty local work queue
3267 // . then get a batch of oops from global work queue if any
3268 // . then do work stealing
3269 // -- When all tasks claimed (both spaces)
3270 // and local work queue empty,
3271 // then in a loop do:
3272 // . check global overflow stack; steal a batch of oops and trace
3273 // . try to steal from other threads oif GOS is empty
3274 // . if neither is available, offer termination
3275 // -- Terminate and return result
3276 //
3277 void CMSConcMarkingTask::work(uint worker_id) {
3278 elapsedTimer _timer;
3279 ResourceMark rm;
3280 HandleMark hm;
3281
3282 DEBUG_ONLY(_collector->verify_overflow_empty();)
3283
3284 // Before we begin work, our work queue should be empty
3285 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3286 // Scan the bitmap covering _cms_space, tracing through grey objects.
3287 _timer.start();
3288 do_scan_and_mark(worker_id, _cms_space);
3289 _timer.stop();
3290 if (PrintCMSStatistics != 0) {
3291 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3292 worker_id, _timer.seconds());
3293 // XXX: need xxx/xxx type of notation, two timers
3294 }
3295
3296 // ... do work stealing
3297 _timer.reset();
3298 _timer.start();
3299 do_work_steal(worker_id);
3300 _timer.stop();
3301 if (PrintCMSStatistics != 0) {
3302 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3303 worker_id, _timer.seconds());
3304 // XXX: need xxx/xxx type of notation, two timers
3305 }
3306 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3307 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3308 // Note that under the current task protocol, the
3309 // following assertion is true even of the spaces
3310 // expanded since the completion of the concurrent
3311 // marking. XXX This will likely change under a strict
3312 // ABORT semantics.
3313 // After perm removal the comparison was changed to
3314 // greater than or equal to from strictly greater than.
3315 // Before perm removal the highest address sweep would
3316 // have been at the end of perm gen but now is at the
3317 // end of the tenured gen.
3318 assert(_global_finger >= _cms_space->end(),
3319 "All tasks have been completed");
3320 DEBUG_ONLY(_collector->verify_overflow_empty();)
3321 }
3322
3323 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3324 HeapWord* read = _global_finger;
3325 HeapWord* cur = read;
3326 while (f > read) {
3327 cur = read;
3328 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3329 if (cur == read) {
3330 // our cas succeeded
3331 assert(_global_finger >= f, "protocol consistency");
3332 break;
3333 }
3334 }
3335 }
3336
3337 // This is really inefficient, and should be redone by
3338 // using (not yet available) block-read and -write interfaces to the
3339 // stack and the work_queue. XXX FIX ME !!!
3340 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3341 OopTaskQueue* work_q) {
3342 // Fast lock-free check
3343 if (ovflw_stk->length() == 0) {
3344 return false;
3345 }
3346 assert(work_q->size() == 0, "Shouldn't steal");
3347 MutexLockerEx ml(ovflw_stk->par_lock(),
3348 Mutex::_no_safepoint_check_flag);
3349 // Grab up to 1/4 the size of the work queue
3350 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3351 (size_t)ParGCDesiredObjsFromOverflowList);
3352 num = MIN2(num, ovflw_stk->length());
3353 for (int i = (int) num; i > 0; i--) {
3354 oop cur = ovflw_stk->pop();
3355 assert(cur != NULL, "Counted wrong?");
3356 work_q->push(cur);
3357 }
3358 return num > 0;
3359 }
3360
3361 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3362 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3363 int n_tasks = pst->n_tasks();
3364 // We allow that there may be no tasks to do here because
3365 // we are restarting after a stack overflow.
3366 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3367 uint nth_task = 0;
3368
3369 HeapWord* aligned_start = sp->bottom();
3370 if (sp->used_region().contains(_restart_addr)) {
3371 // Align down to a card boundary for the start of 0th task
3372 // for this space.
3373 aligned_start =
3374 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3375 CardTableModRefBS::card_size);
3376 }
3377
3378 size_t chunk_size = sp->marking_task_size();
3379 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3380 // Having claimed the nth task in this space,
3381 // compute the chunk that it corresponds to:
3382 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3383 aligned_start + (nth_task+1)*chunk_size);
3384 // Try and bump the global finger via a CAS;
3385 // note that we need to do the global finger bump
3386 // _before_ taking the intersection below, because
3387 // the task corresponding to that region will be
3388 // deemed done even if the used_region() expands
3389 // because of allocation -- as it almost certainly will
3390 // during start-up while the threads yield in the
3391 // closure below.
3392 HeapWord* finger = span.end();
3393 bump_global_finger(finger); // atomically
3394 // There are null tasks here corresponding to chunks
3395 // beyond the "top" address of the space.
3396 span = span.intersection(sp->used_region());
3397 if (!span.is_empty()) { // Non-null task
3398 HeapWord* prev_obj;
3399 assert(!span.contains(_restart_addr) || nth_task == 0,
3400 "Inconsistency");
3401 if (nth_task == 0) {
3402 // For the 0th task, we'll not need to compute a block_start.
3403 if (span.contains(_restart_addr)) {
3404 // In the case of a restart because of stack overflow,
3405 // we might additionally skip a chunk prefix.
3406 prev_obj = _restart_addr;
3407 } else {
3408 prev_obj = span.start();
3409 }
3410 } else {
3411 // We want to skip the first object because
3412 // the protocol is to scan any object in its entirety
3413 // that _starts_ in this span; a fortiori, any
3414 // object starting in an earlier span is scanned
3415 // as part of an earlier claimed task.
3416 // Below we use the "careful" version of block_start
3417 // so we do not try to navigate uninitialized objects.
3418 prev_obj = sp->block_start_careful(span.start());
3419 // Below we use a variant of block_size that uses the
3420 // Printezis bits to avoid waiting for allocated
3421 // objects to become initialized/parsable.
3422 while (prev_obj < span.start()) {
3423 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3424 if (sz > 0) {
3425 prev_obj += sz;
3426 } else {
3427 // In this case we may end up doing a bit of redundant
3428 // scanning, but that appears unavoidable, short of
3429 // locking the free list locks; see bug 6324141.
3430 break;
3431 }
3432 }
3433 }
3434 if (prev_obj < span.end()) {
3435 MemRegion my_span = MemRegion(prev_obj, span.end());
3436 // Do the marking work within a non-empty span --
3437 // the last argument to the constructor indicates whether the
3438 // iteration should be incremental with periodic yields.
3439 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3440 &_collector->_markBitMap,
3441 work_queue(i),
3442 &_collector->_markStack);
3443 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3444 } // else nothing to do for this task
3445 } // else nothing to do for this task
3446 }
3447 // We'd be tempted to assert here that since there are no
3448 // more tasks left to claim in this space, the global_finger
3449 // must exceed space->top() and a fortiori space->end(). However,
3450 // that would not quite be correct because the bumping of
3451 // global_finger occurs strictly after the claiming of a task,
3452 // so by the time we reach here the global finger may not yet
3453 // have been bumped up by the thread that claimed the last
3454 // task.
3455 pst->all_tasks_completed();
3456 }
3457
3458 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3459 private:
3460 CMSCollector* _collector;
3461 CMSConcMarkingTask* _task;
3462 MemRegion _span;
3463 CMSBitMap* _bit_map;
3464 CMSMarkStack* _overflow_stack;
3465 OopTaskQueue* _work_queue;
3466 protected:
3467 DO_OOP_WORK_DEFN
3468 public:
3469 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3470 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3471 MetadataAwareOopClosure(collector->ref_processor()),
3472 _collector(collector),
3473 _task(task),
3474 _span(collector->_span),
3475 _work_queue(work_queue),
3476 _bit_map(bit_map),
3477 _overflow_stack(overflow_stack)
3478 { }
3479 virtual void do_oop(oop* p);
3480 virtual void do_oop(narrowOop* p);
3481
3482 void trim_queue(size_t max);
3483 void handle_stack_overflow(HeapWord* lost);
3484 void do_yield_check() {
3485 if (_task->should_yield()) {
3486 _task->yield();
3487 }
3488 }
3489 };
3490
3491 // Grey object scanning during work stealing phase --
3492 // the salient assumption here is that any references
3493 // that are in these stolen objects being scanned must
3494 // already have been initialized (else they would not have
3495 // been published), so we do not need to check for
3496 // uninitialized objects before pushing here.
3497 void Par_ConcMarkingClosure::do_oop(oop obj) {
3498 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3499 HeapWord* addr = (HeapWord*)obj;
3500 // Check if oop points into the CMS generation
3501 // and is not marked
3502 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3503 // a white object ...
3504 // If we manage to "claim" the object, by being the
3505 // first thread to mark it, then we push it on our
3506 // marking stack
3507 if (_bit_map->par_mark(addr)) { // ... now grey
3508 // push on work queue (grey set)
3509 bool simulate_overflow = false;
3510 NOT_PRODUCT(
3511 if (CMSMarkStackOverflowALot &&
3512 _collector->simulate_overflow()) {
3513 // simulate a stack overflow
3514 simulate_overflow = true;
3515 }
3516 )
3517 if (simulate_overflow ||
3518 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3519 // stack overflow
3520 if (PrintCMSStatistics != 0) {
3521 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3522 SIZE_FORMAT, _overflow_stack->capacity());
3523 }
3524 // We cannot assert that the overflow stack is full because
3525 // it may have been emptied since.
3526 assert(simulate_overflow ||
3527 _work_queue->size() == _work_queue->max_elems(),
3528 "Else push should have succeeded");
3529 handle_stack_overflow(addr);
3530 }
3531 } // Else, some other thread got there first
3532 do_yield_check();
3533 }
3534 }
3535
3536 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3537 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3538
3539 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3540 while (_work_queue->size() > max) {
3541 oop new_oop;
3542 if (_work_queue->pop_local(new_oop)) {
3543 assert(new_oop->is_oop(), "Should be an oop");
3544 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3545 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3546 new_oop->oop_iterate(this); // do_oop() above
3547 do_yield_check();
3548 }
3549 }
3550 }
3551
3552 // Upon stack overflow, we discard (part of) the stack,
3553 // remembering the least address amongst those discarded
3554 // in CMSCollector's _restart_address.
3555 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3556 // We need to do this under a mutex to prevent other
3557 // workers from interfering with the work done below.
3558 MutexLockerEx ml(_overflow_stack->par_lock(),
3559 Mutex::_no_safepoint_check_flag);
3560 // Remember the least grey address discarded
3561 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3562 _collector->lower_restart_addr(ra);
3563 _overflow_stack->reset(); // discard stack contents
3564 _overflow_stack->expand(); // expand the stack if possible
3565 }
3566
3567
3568 void CMSConcMarkingTask::do_work_steal(int i) {
3569 OopTaskQueue* work_q = work_queue(i);
3570 oop obj_to_scan;
3571 CMSBitMap* bm = &(_collector->_markBitMap);
3572 CMSMarkStack* ovflw = &(_collector->_markStack);
3573 int* seed = _collector->hash_seed(i);
3574 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3575 while (true) {
3576 cl.trim_queue(0);
3577 assert(work_q->size() == 0, "Should have been emptied above");
3578 if (get_work_from_overflow_stack(ovflw, work_q)) {
3579 // Can't assert below because the work obtained from the
3580 // overflow stack may already have been stolen from us.
3581 // assert(work_q->size() > 0, "Work from overflow stack");
3582 continue;
3583 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3584 assert(obj_to_scan->is_oop(), "Should be an oop");
3585 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3586 obj_to_scan->oop_iterate(&cl);
3587 } else if (terminator()->offer_termination(&_term_term)) {
3588 assert(work_q->size() == 0, "Impossible!");
3589 break;
3590 } else if (yielding() || should_yield()) {
3591 yield();
3592 }
3593 }
3594 }
3595
3596 // This is run by the CMS (coordinator) thread.
3597 void CMSConcMarkingTask::coordinator_yield() {
3598 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3599 "CMS thread should hold CMS token");
3600 // First give up the locks, then yield, then re-lock
3601 // We should probably use a constructor/destructor idiom to
3602 // do this unlock/lock or modify the MutexUnlocker class to
3603 // serve our purpose. XXX
3604 assert_lock_strong(_bit_map_lock);
3605 _bit_map_lock->unlock();
3606 ConcurrentMarkSweepThread::desynchronize(true);
3607 _collector->stopTimer();
3608 if (PrintCMSStatistics != 0) {
3609 _collector->incrementYields();
3610 }
3611
3612 // It is possible for whichever thread initiated the yield request
3613 // not to get a chance to wake up and take the bitmap lock between
3614 // this thread releasing it and reacquiring it. So, while the
3615 // should_yield() flag is on, let's sleep for a bit to give the
3616 // other thread a chance to wake up. The limit imposed on the number
3617 // of iterations is defensive, to avoid any unforseen circumstances
3618 // putting us into an infinite loop. Since it's always been this
3619 // (coordinator_yield()) method that was observed to cause the
3620 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3621 // which is by default non-zero. For the other seven methods that
3622 // also perform the yield operation, as are using a different
3623 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3624 // can enable the sleeping for those methods too, if necessary.
3625 // See 6442774.
3626 //
3627 // We really need to reconsider the synchronization between the GC
3628 // thread and the yield-requesting threads in the future and we
3629 // should really use wait/notify, which is the recommended
3630 // way of doing this type of interaction. Additionally, we should
3631 // consolidate the eight methods that do the yield operation and they
3632 // are almost identical into one for better maintainability and
3633 // readability. See 6445193.
3634 //
3635 // Tony 2006.06.29
3636 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3637 ConcurrentMarkSweepThread::should_yield() &&
3638 !CMSCollector::foregroundGCIsActive(); ++i) {
3639 os::sleep(Thread::current(), 1, false);
3640 }
3641
3642 ConcurrentMarkSweepThread::synchronize(true);
3643 _bit_map_lock->lock_without_safepoint_check();
3644 _collector->startTimer();
3645 }
3646
3647 bool CMSCollector::do_marking_mt() {
3648 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3649 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3650 conc_workers()->active_workers(),
3651 Threads::number_of_non_daemon_threads());
3652 conc_workers()->set_active_workers(num_workers);
3653
3654 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3655
3656 CMSConcMarkingTask tsk(this,
3657 cms_space,
3658 conc_workers(),
3659 task_queues());
3660
3661 // Since the actual number of workers we get may be different
3662 // from the number we requested above, do we need to do anything different
3663 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3664 // class?? XXX
3665 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3666
3667 // Refs discovery is already non-atomic.
3668 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3669 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3670 conc_workers()->start_task(&tsk);
3671 while (tsk.yielded()) {
3672 tsk.coordinator_yield();
3673 conc_workers()->continue_task(&tsk);
3674 }
3675 // If the task was aborted, _restart_addr will be non-NULL
3676 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3677 while (_restart_addr != NULL) {
3678 // XXX For now we do not make use of ABORTED state and have not
3679 // yet implemented the right abort semantics (even in the original
3680 // single-threaded CMS case). That needs some more investigation
3681 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3682 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3683 // If _restart_addr is non-NULL, a marking stack overflow
3684 // occurred; we need to do a fresh marking iteration from the
3685 // indicated restart address.
3686 if (_foregroundGCIsActive) {
3687 // We may be running into repeated stack overflows, having
3688 // reached the limit of the stack size, while making very
3689 // slow forward progress. It may be best to bail out and
3690 // let the foreground collector do its job.
3691 // Clear _restart_addr, so that foreground GC
3692 // works from scratch. This avoids the headache of
3693 // a "rescan" which would otherwise be needed because
3694 // of the dirty mod union table & card table.
3695 _restart_addr = NULL;
3696 return false;
3697 }
3698 // Adjust the task to restart from _restart_addr
3699 tsk.reset(_restart_addr);
3700 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3701 _restart_addr);
3702 _restart_addr = NULL;
3703 // Get the workers going again
3704 conc_workers()->start_task(&tsk);
3705 while (tsk.yielded()) {
3706 tsk.coordinator_yield();
3707 conc_workers()->continue_task(&tsk);
3708 }
3709 }
3710 assert(tsk.completed(), "Inconsistency");
3711 assert(tsk.result() == true, "Inconsistency");
3712 return true;
3713 }
3714
3715 bool CMSCollector::do_marking_st() {
3716 ResourceMark rm;
3717 HandleMark hm;
3718
3719 // Temporarily make refs discovery single threaded (non-MT)
3720 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3721 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3722 &_markStack, CMSYield);
3723 // the last argument to iterate indicates whether the iteration
3724 // should be incremental with periodic yields.
3725 _markBitMap.iterate(&markFromRootsClosure);
3726 // If _restart_addr is non-NULL, a marking stack overflow
3727 // occurred; we need to do a fresh iteration from the
3728 // indicated restart address.
3729 while (_restart_addr != NULL) {
3730 if (_foregroundGCIsActive) {
3731 // We may be running into repeated stack overflows, having
3732 // reached the limit of the stack size, while making very
3733 // slow forward progress. It may be best to bail out and
3734 // let the foreground collector do its job.
3735 // Clear _restart_addr, so that foreground GC
3736 // works from scratch. This avoids the headache of
3737 // a "rescan" which would otherwise be needed because
3738 // of the dirty mod union table & card table.
3739 _restart_addr = NULL;
3740 return false; // indicating failure to complete marking
3741 }
3742 // Deal with stack overflow:
3743 // we restart marking from _restart_addr
3744 HeapWord* ra = _restart_addr;
3745 markFromRootsClosure.reset(ra);
3746 _restart_addr = NULL;
3747 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3748 }
3749 return true;
3750 }
3751
3752 void CMSCollector::preclean() {
3753 check_correct_thread_executing();
3754 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3755 verify_work_stacks_empty();
3756 verify_overflow_empty();
3757 _abort_preclean = false;
3758 if (CMSPrecleaningEnabled) {
3759 if (!CMSEdenChunksRecordAlways) {
3760 _eden_chunk_index = 0;
3761 }
3762 size_t used = get_eden_used();
3763 size_t capacity = get_eden_capacity();
3764 // Don't start sampling unless we will get sufficiently
3765 // many samples.
3766 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3767 * CMSScheduleRemarkEdenPenetration)) {
3768 _start_sampling = true;
3769 } else {
3770 _start_sampling = false;
3771 }
3772 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3773 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3774 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3775 }
3776 CMSTokenSync x(true); // is cms thread
3777 if (CMSPrecleaningEnabled) {
3778 sample_eden();
3779 _collectorState = AbortablePreclean;
3780 } else {
3781 _collectorState = FinalMarking;
3782 }
3783 verify_work_stacks_empty();
3784 verify_overflow_empty();
3785 }
3786
3787 // Try and schedule the remark such that young gen
3788 // occupancy is CMSScheduleRemarkEdenPenetration %.
3789 void CMSCollector::abortable_preclean() {
3790 check_correct_thread_executing();
3791 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3792 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3793
3794 // If Eden's current occupancy is below this threshold,
3795 // immediately schedule the remark; else preclean
3796 // past the next scavenge in an effort to
3797 // schedule the pause as described above. By choosing
3798 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3799 // we will never do an actual abortable preclean cycle.
3800 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3801 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3802 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3803 // We need more smarts in the abortable preclean
3804 // loop below to deal with cases where allocation
3805 // in young gen is very very slow, and our precleaning
3806 // is running a losing race against a horde of
3807 // mutators intent on flooding us with CMS updates
3808 // (dirty cards).
3809 // One, admittedly dumb, strategy is to give up
3810 // after a certain number of abortable precleaning loops
3811 // or after a certain maximum time. We want to make
3812 // this smarter in the next iteration.
3813 // XXX FIX ME!!! YSR
3814 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3815 while (!(should_abort_preclean() ||
3816 ConcurrentMarkSweepThread::should_terminate())) {
3817 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3818 cumworkdone += workdone;
3819 loops++;
3820 // Voluntarily terminate abortable preclean phase if we have
3821 // been at it for too long.
3822 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3823 loops >= CMSMaxAbortablePrecleanLoops) {
3824 if (PrintGCDetails) {
3825 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3826 }
3827 break;
3828 }
3829 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3830 if (PrintGCDetails) {
3831 gclog_or_tty->print(" CMS: abort preclean due to time ");
3832 }
3833 break;
3834 }
3835 // If we are doing little work each iteration, we should
3836 // take a short break.
3837 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3838 // Sleep for some time, waiting for work to accumulate
3839 stopTimer();
3840 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3841 startTimer();
3842 waited++;
3843 }
3844 }
3845 if (PrintCMSStatistics > 0) {
3846 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3847 loops, waited, cumworkdone);
3848 }
3849 }
3850 CMSTokenSync x(true); // is cms thread
3851 if (_collectorState != Idling) {
3852 assert(_collectorState == AbortablePreclean,
3853 "Spontaneous state transition?");
3854 _collectorState = FinalMarking;
3855 } // Else, a foreground collection completed this CMS cycle.
3856 return;
3857 }
3858
3859 // Respond to an Eden sampling opportunity
3860 void CMSCollector::sample_eden() {
3861 // Make sure a young gc cannot sneak in between our
3862 // reading and recording of a sample.
3863 assert(Thread::current()->is_ConcurrentGC_thread(),
3864 "Only the cms thread may collect Eden samples");
3865 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3866 "Should collect samples while holding CMS token");
3867 if (!_start_sampling) {
3868 return;
3869 }
3870 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3871 // is populated by the young generation.
3872 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3873 if (_eden_chunk_index < _eden_chunk_capacity) {
3874 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3875 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3876 "Unexpected state of Eden");
3877 // We'd like to check that what we just sampled is an oop-start address;
3878 // however, we cannot do that here since the object may not yet have been
3879 // initialized. So we'll instead do the check when we _use_ this sample
3880 // later.
3881 if (_eden_chunk_index == 0 ||
3882 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3883 _eden_chunk_array[_eden_chunk_index-1])
3884 >= CMSSamplingGrain)) {
3885 _eden_chunk_index++; // commit sample
3886 }
3887 }
3888 }
3889 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3890 size_t used = get_eden_used();
3891 size_t capacity = get_eden_capacity();
3892 assert(used <= capacity, "Unexpected state of Eden");
3893 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3894 _abort_preclean = true;
3895 }
3896 }
3897 }
3898
3899
3900 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3901 assert(_collectorState == Precleaning ||
3902 _collectorState == AbortablePreclean, "incorrect state");
3903 ResourceMark rm;
3904 HandleMark hm;
3905
3906 // Precleaning is currently not MT but the reference processor
3907 // may be set for MT. Disable it temporarily here.
3908 ReferenceProcessor* rp = ref_processor();
3909 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3910
3911 // Do one pass of scrubbing the discovered reference lists
3912 // to remove any reference objects with strongly-reachable
3913 // referents.
3914 if (clean_refs) {
3915 CMSPrecleanRefsYieldClosure yield_cl(this);
3916 assert(rp->span().equals(_span), "Spans should be equal");
3917 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3918 &_markStack, true /* preclean */);
3919 CMSDrainMarkingStackClosure complete_trace(this,
3920 _span, &_markBitMap, &_markStack,
3921 &keep_alive, true /* preclean */);
3922
3923 // We don't want this step to interfere with a young
3924 // collection because we don't want to take CPU
3925 // or memory bandwidth away from the young GC threads
3926 // (which may be as many as there are CPUs).
3927 // Note that we don't need to protect ourselves from
3928 // interference with mutators because they can't
3929 // manipulate the discovered reference lists nor affect
3930 // the computed reachability of the referents, the
3931 // only properties manipulated by the precleaning
3932 // of these reference lists.
3933 stopTimer();
3934 CMSTokenSyncWithLocks x(true /* is cms thread */,
3935 bitMapLock());
3936 startTimer();
3937 sample_eden();
3938
3939 // The following will yield to allow foreground
3940 // collection to proceed promptly. XXX YSR:
3941 // The code in this method may need further
3942 // tweaking for better performance and some restructuring
3943 // for cleaner interfaces.
3944 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3945 rp->preclean_discovered_references(
3946 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3947 gc_timer, _gc_tracer_cm->gc_id());
3948 }
3949
3950 if (clean_survivor) { // preclean the active survivor space(s)
3951 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3952 &_markBitMap, &_modUnionTable,
3953 &_markStack, true /* precleaning phase */);
3954 stopTimer();
3955 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3956 bitMapLock());
3957 startTimer();
3958 unsigned int before_count =
3959 GenCollectedHeap::heap()->total_collections();
3960 SurvivorSpacePrecleanClosure
3961 sss_cl(this, _span, &_markBitMap, &_markStack,
3962 &pam_cl, before_count, CMSYield);
3963 _young_gen->from()->object_iterate_careful(&sss_cl);
3964 _young_gen->to()->object_iterate_careful(&sss_cl);
3965 }
3966 MarkRefsIntoAndScanClosure
3967 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3968 &_markStack, this, CMSYield,
3969 true /* precleaning phase */);
3970 // CAUTION: The following closure has persistent state that may need to
3971 // be reset upon a decrease in the sequence of addresses it
3972 // processes.
3973 ScanMarkedObjectsAgainCarefullyClosure
3974 smoac_cl(this, _span,
3975 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3976
3977 // Preclean dirty cards in ModUnionTable and CardTable using
3978 // appropriate convergence criterion;
3979 // repeat CMSPrecleanIter times unless we find that
3980 // we are losing.
3981 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3982 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3983 "Bad convergence multiplier");
3984 assert(CMSPrecleanThreshold >= 100,
3985 "Unreasonably low CMSPrecleanThreshold");
3986
3987 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3988 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3989 numIter < CMSPrecleanIter;
3990 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3991 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3992 if (Verbose && PrintGCDetails) {
3993 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3994 }
3995 // Either there are very few dirty cards, so re-mark
3996 // pause will be small anyway, or our pre-cleaning isn't
3997 // that much faster than the rate at which cards are being
3998 // dirtied, so we might as well stop and re-mark since
3999 // precleaning won't improve our re-mark time by much.
4000 if (curNumCards <= CMSPrecleanThreshold ||
4001 (numIter > 0 &&
4002 (curNumCards * CMSPrecleanDenominator >
4003 lastNumCards * CMSPrecleanNumerator))) {
4004 numIter++;
4005 cumNumCards += curNumCards;
4006 break;
4007 }
4008 }
4009
4010 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4011
4012 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4013 cumNumCards += curNumCards;
4014 if (PrintGCDetails && PrintCMSStatistics != 0) {
4015 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4016 curNumCards, cumNumCards, numIter);
4017 }
4018 return cumNumCards; // as a measure of useful work done
4019 }
4020
4021 // PRECLEANING NOTES:
4022 // Precleaning involves:
4023 // . reading the bits of the modUnionTable and clearing the set bits.
4024 // . For the cards corresponding to the set bits, we scan the
4025 // objects on those cards. This means we need the free_list_lock
4026 // so that we can safely iterate over the CMS space when scanning
4027 // for oops.
4028 // . When we scan the objects, we'll be both reading and setting
4029 // marks in the marking bit map, so we'll need the marking bit map.
4030 // . For protecting _collector_state transitions, we take the CGC_lock.
4031 // Note that any races in the reading of of card table entries by the
4032 // CMS thread on the one hand and the clearing of those entries by the
4033 // VM thread or the setting of those entries by the mutator threads on the
4034 // other are quite benign. However, for efficiency it makes sense to keep
4035 // the VM thread from racing with the CMS thread while the latter is
4036 // dirty card info to the modUnionTable. We therefore also use the
4037 // CGC_lock to protect the reading of the card table and the mod union
4038 // table by the CM thread.
4039 // . We run concurrently with mutator updates, so scanning
4040 // needs to be done carefully -- we should not try to scan
4041 // potentially uninitialized objects.
4042 //
4043 // Locking strategy: While holding the CGC_lock, we scan over and
4044 // reset a maximal dirty range of the mod union / card tables, then lock
4045 // the free_list_lock and bitmap lock to do a full marking, then
4046 // release these locks; and repeat the cycle. This allows for a
4047 // certain amount of fairness in the sharing of these locks between
4048 // the CMS collector on the one hand, and the VM thread and the
4049 // mutators on the other.
4050
4051 // NOTE: preclean_mod_union_table() and preclean_card_table()
4052 // further below are largely identical; if you need to modify
4053 // one of these methods, please check the other method too.
4054
4055 size_t CMSCollector::preclean_mod_union_table(
4056 ConcurrentMarkSweepGeneration* gen,
4057 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4058 verify_work_stacks_empty();
4059 verify_overflow_empty();
4060
4061 // strategy: starting with the first card, accumulate contiguous
4062 // ranges of dirty cards; clear these cards, then scan the region
4063 // covered by these cards.
4064
4065 // Since all of the MUT is committed ahead, we can just use
4066 // that, in case the generations expand while we are precleaning.
4067 // It might also be fine to just use the committed part of the
4068 // generation, but we might potentially miss cards when the
4069 // generation is rapidly expanding while we are in the midst
4070 // of precleaning.
4071 HeapWord* startAddr = gen->reserved().start();
4072 HeapWord* endAddr = gen->reserved().end();
4073
4074 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4075
4076 size_t numDirtyCards, cumNumDirtyCards;
4077 HeapWord *nextAddr, *lastAddr;
4078 for (cumNumDirtyCards = numDirtyCards = 0,
4079 nextAddr = lastAddr = startAddr;
4080 nextAddr < endAddr;
4081 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4082
4083 ResourceMark rm;
4084 HandleMark hm;
4085
4086 MemRegion dirtyRegion;
4087 {
4088 stopTimer();
4089 // Potential yield point
4090 CMSTokenSync ts(true);
4091 startTimer();
4092 sample_eden();
4093 // Get dirty region starting at nextOffset (inclusive),
4094 // simultaneously clearing it.
4095 dirtyRegion =
4096 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4097 assert(dirtyRegion.start() >= nextAddr,
4098 "returned region inconsistent?");
4099 }
4100 // Remember where the next search should begin.
4101 // The returned region (if non-empty) is a right open interval,
4102 // so lastOffset is obtained from the right end of that
4103 // interval.
4104 lastAddr = dirtyRegion.end();
4105 // Should do something more transparent and less hacky XXX
4106 numDirtyCards =
4107 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4108
4109 // We'll scan the cards in the dirty region (with periodic
4110 // yields for foreground GC as needed).
4111 if (!dirtyRegion.is_empty()) {
4112 assert(numDirtyCards > 0, "consistency check");
4113 HeapWord* stop_point = NULL;
4114 stopTimer();
4115 // Potential yield point
4116 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4117 bitMapLock());
4118 startTimer();
4119 {
4120 verify_work_stacks_empty();
4121 verify_overflow_empty();
4122 sample_eden();
4123 stop_point =
4124 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4125 }
4126 if (stop_point != NULL) {
4127 // The careful iteration stopped early either because it found an
4128 // uninitialized object, or because we were in the midst of an
4129 // "abortable preclean", which should now be aborted. Redirty
4130 // the bits corresponding to the partially-scanned or unscanned
4131 // cards. We'll either restart at the next block boundary or
4132 // abort the preclean.
4133 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4134 "Should only be AbortablePreclean.");
4135 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4136 if (should_abort_preclean()) {
4137 break; // out of preclean loop
4138 } else {
4139 // Compute the next address at which preclean should pick up;
4140 // might need bitMapLock in order to read P-bits.
4141 lastAddr = next_card_start_after_block(stop_point);
4142 }
4143 }
4144 } else {
4145 assert(lastAddr == endAddr, "consistency check");
4146 assert(numDirtyCards == 0, "consistency check");
4147 break;
4148 }
4149 }
4150 verify_work_stacks_empty();
4151 verify_overflow_empty();
4152 return cumNumDirtyCards;
4153 }
4154
4155 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4156 // below are largely identical; if you need to modify
4157 // one of these methods, please check the other method too.
4158
4159 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4160 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4161 // strategy: it's similar to precleamModUnionTable above, in that
4162 // we accumulate contiguous ranges of dirty cards, mark these cards
4163 // precleaned, then scan the region covered by these cards.
4164 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4165 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4166
4167 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4168
4169 size_t numDirtyCards, cumNumDirtyCards;
4170 HeapWord *lastAddr, *nextAddr;
4171
4172 for (cumNumDirtyCards = numDirtyCards = 0,
4173 nextAddr = lastAddr = startAddr;
4174 nextAddr < endAddr;
4175 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4176
4177 ResourceMark rm;
4178 HandleMark hm;
4179
4180 MemRegion dirtyRegion;
4181 {
4182 // See comments in "Precleaning notes" above on why we
4183 // do this locking. XXX Could the locking overheads be
4184 // too high when dirty cards are sparse? [I don't think so.]
4185 stopTimer();
4186 CMSTokenSync x(true); // is cms thread
4187 startTimer();
4188 sample_eden();
4189 // Get and clear dirty region from card table
4190 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4191 MemRegion(nextAddr, endAddr),
4192 true,
4193 CardTableModRefBS::precleaned_card_val());
4194
4195 assert(dirtyRegion.start() >= nextAddr,
4196 "returned region inconsistent?");
4197 }
4198 lastAddr = dirtyRegion.end();
4199 numDirtyCards =
4200 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4201
4202 if (!dirtyRegion.is_empty()) {
4203 stopTimer();
4204 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4205 startTimer();
4206 sample_eden();
4207 verify_work_stacks_empty();
4208 verify_overflow_empty();
4209 HeapWord* stop_point =
4210 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4211 if (stop_point != NULL) {
4212 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4213 "Should only be AbortablePreclean.");
4214 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4215 if (should_abort_preclean()) {
4216 break; // out of preclean loop
4217 } else {
4218 // Compute the next address at which preclean should pick up.
4219 lastAddr = next_card_start_after_block(stop_point);
4220 }
4221 }
4222 } else {
4223 break;
4224 }
4225 }
4226 verify_work_stacks_empty();
4227 verify_overflow_empty();
4228 return cumNumDirtyCards;
4229 }
4230
4231 class PrecleanKlassClosure : public KlassClosure {
4232 KlassToOopClosure _cm_klass_closure;
4233 public:
4234 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4235 void do_klass(Klass* k) {
4236 if (k->has_accumulated_modified_oops()) {
4237 k->clear_accumulated_modified_oops();
4238
4239 _cm_klass_closure.do_klass(k);
4240 }
4241 }
4242 };
4243
4244 // The freelist lock is needed to prevent asserts, is it really needed?
4245 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4246
4247 cl->set_freelistLock(freelistLock);
4248
4249 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4250
4251 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4252 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4253 PrecleanKlassClosure preclean_klass_closure(cl);
4254 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4255
4256 verify_work_stacks_empty();
4257 verify_overflow_empty();
4258 }
4259
4260 void CMSCollector::checkpointRootsFinal() {
4261 assert(_collectorState == FinalMarking, "incorrect state transition?");
4262 check_correct_thread_executing();
4263 // world is stopped at this checkpoint
4264 assert(SafepointSynchronize::is_at_safepoint(),
4265 "world should be stopped");
4266 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4267
4268 verify_work_stacks_empty();
4269 verify_overflow_empty();
4270
4271 if (PrintGCDetails) {
4272 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4273 _young_gen->used() / K,
4274 _young_gen->capacity() / K);
4275 }
4276 {
4277 if (CMSScavengeBeforeRemark) {
4278 GenCollectedHeap* gch = GenCollectedHeap::heap();
4279 // Temporarily set flag to false, GCH->do_collection will
4280 // expect it to be false and set to true
4281 FlagSetting fl(gch->_is_gc_active, false);
4282 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4283 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4284 gch->do_collection(true, // full (i.e. force, see below)
4285 false, // !clear_all_soft_refs
4286 0, // size
4287 false, // is_tlab
4288 Generation::Young // type
4289 );
4290 }
4291 FreelistLocker x(this);
4292 MutexLockerEx y(bitMapLock(),
4293 Mutex::_no_safepoint_check_flag);
4294 checkpointRootsFinalWork();
4295 }
4296 verify_work_stacks_empty();
4297 verify_overflow_empty();
4298 }
4299
4300 void CMSCollector::checkpointRootsFinalWork() {
4301 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4302
4303 assert(haveFreelistLocks(), "must have free list locks");
4304 assert_lock_strong(bitMapLock());
4305
4306 ResourceMark rm;
4307 HandleMark hm;
4308
4309 GenCollectedHeap* gch = GenCollectedHeap::heap();
4310
4311 if (should_unload_classes()) {
4312 CodeCache::gc_prologue();
4313 }
4314 assert(haveFreelistLocks(), "must have free list locks");
4315 assert_lock_strong(bitMapLock());
4316
4317 // We might assume that we need not fill TLAB's when
4318 // CMSScavengeBeforeRemark is set, because we may have just done
4319 // a scavenge which would have filled all TLAB's -- and besides
4320 // Eden would be empty. This however may not always be the case --
4321 // for instance although we asked for a scavenge, it may not have
4322 // happened because of a JNI critical section. We probably need
4323 // a policy for deciding whether we can in that case wait until
4324 // the critical section releases and then do the remark following
4325 // the scavenge, and skip it here. In the absence of that policy,
4326 // or of an indication of whether the scavenge did indeed occur,
4327 // we cannot rely on TLAB's having been filled and must do
4328 // so here just in case a scavenge did not happen.
4329 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4330 // Update the saved marks which may affect the root scans.
4331 gch->save_marks();
4332
4333 if (CMSPrintEdenSurvivorChunks) {
4334 print_eden_and_survivor_chunk_arrays();
4335 }
4336
4337 {
4338 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4339
4340 // Note on the role of the mod union table:
4341 // Since the marker in "markFromRoots" marks concurrently with
4342 // mutators, it is possible for some reachable objects not to have been
4343 // scanned. For instance, an only reference to an object A was
4344 // placed in object B after the marker scanned B. Unless B is rescanned,
4345 // A would be collected. Such updates to references in marked objects
4346 // are detected via the mod union table which is the set of all cards
4347 // dirtied since the first checkpoint in this GC cycle and prior to
4348 // the most recent young generation GC, minus those cleaned up by the
4349 // concurrent precleaning.
4350 if (CMSParallelRemarkEnabled) {
4351 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4352 do_remark_parallel();
4353 } else {
4354 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4355 _gc_timer_cm, _gc_tracer_cm->gc_id());
4356 do_remark_non_parallel();
4357 }
4358 }
4359 verify_work_stacks_empty();
4360 verify_overflow_empty();
4361
4362 {
4363 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4364 refProcessingWork();
4365 }
4366 verify_work_stacks_empty();
4367 verify_overflow_empty();
4368
4369 if (should_unload_classes()) {
4370 CodeCache::gc_epilogue();
4371 }
4372 JvmtiExport::gc_epilogue();
4373
4374 // If we encountered any (marking stack / work queue) overflow
4375 // events during the current CMS cycle, take appropriate
4376 // remedial measures, where possible, so as to try and avoid
4377 // recurrence of that condition.
4378 assert(_markStack.isEmpty(), "No grey objects");
4379 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4380 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4381 if (ser_ovflw > 0) {
4382 if (PrintCMSStatistics != 0) {
4383 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4384 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4385 ", kac_preclean="SIZE_FORMAT")",
4386 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4387 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4388 }
4389 _markStack.expand();
4390 _ser_pmc_remark_ovflw = 0;
4391 _ser_pmc_preclean_ovflw = 0;
4392 _ser_kac_preclean_ovflw = 0;
4393 _ser_kac_ovflw = 0;
4394 }
4395 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4396 if (PrintCMSStatistics != 0) {
4397 gclog_or_tty->print_cr("Work queue overflow (benign) "
4398 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4399 _par_pmc_remark_ovflw, _par_kac_ovflw);
4400 }
4401 _par_pmc_remark_ovflw = 0;
4402 _par_kac_ovflw = 0;
4403 }
4404 if (PrintCMSStatistics != 0) {
4405 if (_markStack._hit_limit > 0) {
4406 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4407 _markStack._hit_limit);
4408 }
4409 if (_markStack._failed_double > 0) {
4410 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4411 " current capacity "SIZE_FORMAT,
4412 _markStack._failed_double,
4413 _markStack.capacity());
4414 }
4415 }
4416 _markStack._hit_limit = 0;
4417 _markStack._failed_double = 0;
4418
4419 if ((VerifyAfterGC || VerifyDuringGC) &&
4420 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4421 verify_after_remark();
4422 }
4423
4424 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4425
4426 // Change under the freelistLocks.
4427 _collectorState = Sweeping;
4428 // Call isAllClear() under bitMapLock
4429 assert(_modUnionTable.isAllClear(),
4430 "Should be clear by end of the final marking");
4431 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4432 "Should be clear by end of the final marking");
4433 }
4434
4435 void CMSParInitialMarkTask::work(uint worker_id) {
4436 elapsedTimer _timer;
4437 ResourceMark rm;
4438 HandleMark hm;
4439
4440 // ---------- scan from roots --------------
4441 _timer.start();
4442 GenCollectedHeap* gch = GenCollectedHeap::heap();
4443 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4444
4445 // ---------- young gen roots --------------
4446 {
4447 work_on_young_gen_roots(worker_id, &par_mri_cl);
4448 _timer.stop();
4449 if (PrintCMSStatistics != 0) {
4450 gclog_or_tty->print_cr(
4451 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4452 worker_id, _timer.seconds());
4453 }
4454 }
4455
4456 // ---------- remaining roots --------------
4457 _timer.reset();
4458 _timer.start();
4459
4460 CLDToOopClosure cld_closure(&par_mri_cl, true);
4461
4462 gch->gen_process_roots(_strong_roots_scope,
4463 Generation::Old,
4464 false, // yg was scanned above
4465 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4466 _collector->should_unload_classes(),
4467 &par_mri_cl,
4468 NULL,
4469 &cld_closure);
4470 assert(_collector->should_unload_classes()
4471 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4472 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4473 _timer.stop();
4474 if (PrintCMSStatistics != 0) {
4475 gclog_or_tty->print_cr(
4476 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4477 worker_id, _timer.seconds());
4478 }
4479 }
4480
4481 // Parallel remark task
4482 class CMSParRemarkTask: public CMSParMarkTask {
4483 CompactibleFreeListSpace* _cms_space;
4484
4485 // The per-thread work queues, available here for stealing.
4486 OopTaskQueueSet* _task_queues;
4487 ParallelTaskTerminator _term;
4488 StrongRootsScope* _strong_roots_scope;
4489
4490 public:
4491 // A value of 0 passed to n_workers will cause the number of
4492 // workers to be taken from the active workers in the work gang.
4493 CMSParRemarkTask(CMSCollector* collector,
4494 CompactibleFreeListSpace* cms_space,
4495 uint n_workers, FlexibleWorkGang* workers,
4496 OopTaskQueueSet* task_queues,
4497 StrongRootsScope* strong_roots_scope):
4498 CMSParMarkTask("Rescan roots and grey objects in parallel",
4499 collector, n_workers),
4500 _cms_space(cms_space),
4501 _task_queues(task_queues),
4502 _term(n_workers, task_queues),
4503 _strong_roots_scope(strong_roots_scope) { }
4504
4505 OopTaskQueueSet* task_queues() { return _task_queues; }
4506
4507 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4508
4509 ParallelTaskTerminator* terminator() { return &_term; }
4510 uint n_workers() { return _n_workers; }
4511
4512 void work(uint worker_id);
4513
4514 private:
4515 // ... of dirty cards in old space
4516 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4517 Par_MarkRefsIntoAndScanClosure* cl);
4518
4519 // ... work stealing for the above
4520 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4521 };
4522
4523 class RemarkKlassClosure : public KlassClosure {
4524 KlassToOopClosure _cm_klass_closure;
4525 public:
4526 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4527 void do_klass(Klass* k) {
4528 // Check if we have modified any oops in the Klass during the concurrent marking.
4529 if (k->has_accumulated_modified_oops()) {
4530 k->clear_accumulated_modified_oops();
4531
4532 // We could have transfered the current modified marks to the accumulated marks,
4533 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4534 } else if (k->has_modified_oops()) {
4535 // Don't clear anything, this info is needed by the next young collection.
4536 } else {
4537 // No modified oops in the Klass.
4538 return;
4539 }
4540
4541 // The klass has modified fields, need to scan the klass.
4542 _cm_klass_closure.do_klass(k);
4543 }
4544 };
4545
4546 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4547 ParNewGeneration* young_gen = _collector->_young_gen;
4548 ContiguousSpace* eden_space = young_gen->eden();
4549 ContiguousSpace* from_space = young_gen->from();
4550 ContiguousSpace* to_space = young_gen->to();
4551
4552 HeapWord** eca = _collector->_eden_chunk_array;
4553 size_t ect = _collector->_eden_chunk_index;
4554 HeapWord** sca = _collector->_survivor_chunk_array;
4555 size_t sct = _collector->_survivor_chunk_index;
4556
4557 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4558 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4559
4560 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4561 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4562 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4563 }
4564
4565 // work_queue(i) is passed to the closure
4566 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4567 // also is passed to do_dirty_card_rescan_tasks() and to
4568 // do_work_steal() to select the i-th task_queue.
4569
4570 void CMSParRemarkTask::work(uint worker_id) {
4571 elapsedTimer _timer;
4572 ResourceMark rm;
4573 HandleMark hm;
4574
4575 // ---------- rescan from roots --------------
4576 _timer.start();
4577 GenCollectedHeap* gch = GenCollectedHeap::heap();
4578 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4579 _collector->_span, _collector->ref_processor(),
4580 &(_collector->_markBitMap),
4581 work_queue(worker_id));
4582
4583 // Rescan young gen roots first since these are likely
4584 // coarsely partitioned and may, on that account, constitute
4585 // the critical path; thus, it's best to start off that
4586 // work first.
4587 // ---------- young gen roots --------------
4588 {
4589 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4590 _timer.stop();
4591 if (PrintCMSStatistics != 0) {
4592 gclog_or_tty->print_cr(
4593 "Finished young gen rescan work in %dth thread: %3.3f sec",
4594 worker_id, _timer.seconds());
4595 }
4596 }
4597
4598 // ---------- remaining roots --------------
4599 _timer.reset();
4600 _timer.start();
4601 gch->gen_process_roots(_strong_roots_scope,
4602 Generation::Old,
4603 false, // yg was scanned above
4604 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4605 _collector->should_unload_classes(),
4606 &par_mrias_cl,
4607 NULL,
4608 NULL); // The dirty klasses will be handled below
4609
4610 assert(_collector->should_unload_classes()
4611 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4612 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4613 _timer.stop();
4614 if (PrintCMSStatistics != 0) {
4615 gclog_or_tty->print_cr(
4616 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4617 worker_id, _timer.seconds());
4618 }
4619
4620 // ---------- unhandled CLD scanning ----------
4621 if (worker_id == 0) { // Single threaded at the moment.
4622 _timer.reset();
4623 _timer.start();
4624
4625 // Scan all new class loader data objects and new dependencies that were
4626 // introduced during concurrent marking.
4627 ResourceMark rm;
4628 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4629 for (int i = 0; i < array->length(); i++) {
4630 par_mrias_cl.do_class_loader_data(array->at(i));
4631 }
4632
4633 // We don't need to keep track of new CLDs anymore.
4634 ClassLoaderDataGraph::remember_new_clds(false);
4635
4636 _timer.stop();
4637 if (PrintCMSStatistics != 0) {
4638 gclog_or_tty->print_cr(
4639 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4640 worker_id, _timer.seconds());
4641 }
4642 }
4643
4644 // ---------- dirty klass scanning ----------
4645 if (worker_id == 0) { // Single threaded at the moment.
4646 _timer.reset();
4647 _timer.start();
4648
4649 // Scan all classes that was dirtied during the concurrent marking phase.
4650 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4651 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4652
4653 _timer.stop();
4654 if (PrintCMSStatistics != 0) {
4655 gclog_or_tty->print_cr(
4656 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4657 worker_id, _timer.seconds());
4658 }
4659 }
4660
4661 // We might have added oops to ClassLoaderData::_handles during the
4662 // concurrent marking phase. These oops point to newly allocated objects
4663 // that are guaranteed to be kept alive. Either by the direct allocation
4664 // code, or when the young collector processes the roots. Hence,
4665 // we don't have to revisit the _handles block during the remark phase.
4666
4667 // ---------- rescan dirty cards ------------
4668 _timer.reset();
4669 _timer.start();
4670
4671 // Do the rescan tasks for each of the two spaces
4672 // (cms_space) in turn.
4673 // "worker_id" is passed to select the task_queue for "worker_id"
4674 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4675 _timer.stop();
4676 if (PrintCMSStatistics != 0) {
4677 gclog_or_tty->print_cr(
4678 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4679 worker_id, _timer.seconds());
4680 }
4681
4682 // ---------- steal work from other threads ...
4683 // ---------- ... and drain overflow list.
4684 _timer.reset();
4685 _timer.start();
4686 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4687 _timer.stop();
4688 if (PrintCMSStatistics != 0) {
4689 gclog_or_tty->print_cr(
4690 "Finished work stealing in %dth thread: %3.3f sec",
4691 worker_id, _timer.seconds());
4692 }
4693 }
4694
4695 // Note that parameter "i" is not used.
4696 void
4697 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4698 OopsInGenClosure* cl, ContiguousSpace* space,
4699 HeapWord** chunk_array, size_t chunk_top) {
4700 // Until all tasks completed:
4701 // . claim an unclaimed task
4702 // . compute region boundaries corresponding to task claimed
4703 // using chunk_array
4704 // . par_oop_iterate(cl) over that region
4705
4706 ResourceMark rm;
4707 HandleMark hm;
4708
4709 SequentialSubTasksDone* pst = space->par_seq_tasks();
4710
4711 uint nth_task = 0;
4712 uint n_tasks = pst->n_tasks();
4713
4714 if (n_tasks > 0) {
4715 assert(pst->valid(), "Uninitialized use?");
4716 HeapWord *start, *end;
4717 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4718 // We claimed task # nth_task; compute its boundaries.
4719 if (chunk_top == 0) { // no samples were taken
4720 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4721 start = space->bottom();
4722 end = space->top();
4723 } else if (nth_task == 0) {
4724 start = space->bottom();
4725 end = chunk_array[nth_task];
4726 } else if (nth_task < (uint)chunk_top) {
4727 assert(nth_task >= 1, "Control point invariant");
4728 start = chunk_array[nth_task - 1];
4729 end = chunk_array[nth_task];
4730 } else {
4731 assert(nth_task == (uint)chunk_top, "Control point invariant");
4732 start = chunk_array[chunk_top - 1];
4733 end = space->top();
4734 }
4735 MemRegion mr(start, end);
4736 // Verify that mr is in space
4737 assert(mr.is_empty() || space->used_region().contains(mr),
4738 "Should be in space");
4739 // Verify that "start" is an object boundary
4740 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4741 "Should be an oop");
4742 space->par_oop_iterate(mr, cl);
4743 }
4744 pst->all_tasks_completed();
4745 }
4746 }
4747
4748 void
4749 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4750 CompactibleFreeListSpace* sp, int i,
4751 Par_MarkRefsIntoAndScanClosure* cl) {
4752 // Until all tasks completed:
4753 // . claim an unclaimed task
4754 // . compute region boundaries corresponding to task claimed
4755 // . transfer dirty bits ct->mut for that region
4756 // . apply rescanclosure to dirty mut bits for that region
4757
4758 ResourceMark rm;
4759 HandleMark hm;
4760
4761 OopTaskQueue* work_q = work_queue(i);
4762 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4763 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4764 // CAUTION: This closure has state that persists across calls to
4765 // the work method dirty_range_iterate_clear() in that it has
4766 // embedded in it a (subtype of) UpwardsObjectClosure. The
4767 // use of that state in the embedded UpwardsObjectClosure instance
4768 // assumes that the cards are always iterated (even if in parallel
4769 // by several threads) in monotonically increasing order per each
4770 // thread. This is true of the implementation below which picks
4771 // card ranges (chunks) in monotonically increasing order globally
4772 // and, a-fortiori, in monotonically increasing order per thread
4773 // (the latter order being a subsequence of the former).
4774 // If the work code below is ever reorganized into a more chaotic
4775 // work-partitioning form than the current "sequential tasks"
4776 // paradigm, the use of that persistent state will have to be
4777 // revisited and modified appropriately. See also related
4778 // bug 4756801 work on which should examine this code to make
4779 // sure that the changes there do not run counter to the
4780 // assumptions made here and necessary for correctness and
4781 // efficiency. Note also that this code might yield inefficient
4782 // behavior in the case of very large objects that span one or
4783 // more work chunks. Such objects would potentially be scanned
4784 // several times redundantly. Work on 4756801 should try and
4785 // address that performance anomaly if at all possible. XXX
4786 MemRegion full_span = _collector->_span;
4787 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4788 MarkFromDirtyCardsClosure
4789 greyRescanClosure(_collector, full_span, // entire span of interest
4790 sp, bm, work_q, cl);
4791
4792 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4793 assert(pst->valid(), "Uninitialized use?");
4794 uint nth_task = 0;
4795 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4796 MemRegion span = sp->used_region();
4797 HeapWord* start_addr = span.start();
4798 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4799 alignment);
4800 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4801 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4802 start_addr, "Check alignment");
4803 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4804 chunk_size, "Check alignment");
4805
4806 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4807 // Having claimed the nth_task, compute corresponding mem-region,
4808 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4809 // The alignment restriction ensures that we do not need any
4810 // synchronization with other gang-workers while setting or
4811 // clearing bits in thus chunk of the MUT.
4812 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4813 start_addr + (nth_task+1)*chunk_size);
4814 // The last chunk's end might be way beyond end of the
4815 // used region. In that case pull back appropriately.
4816 if (this_span.end() > end_addr) {
4817 this_span.set_end(end_addr);
4818 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4819 }
4820 // Iterate over the dirty cards covering this chunk, marking them
4821 // precleaned, and setting the corresponding bits in the mod union
4822 // table. Since we have been careful to partition at Card and MUT-word
4823 // boundaries no synchronization is needed between parallel threads.
4824 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4825 &modUnionClosure);
4826
4827 // Having transferred these marks into the modUnionTable,
4828 // rescan the marked objects on the dirty cards in the modUnionTable.
4829 // Even if this is at a synchronous collection, the initial marking
4830 // may have been done during an asynchronous collection so there
4831 // may be dirty bits in the mod-union table.
4832 _collector->_modUnionTable.dirty_range_iterate_clear(
4833 this_span, &greyRescanClosure);
4834 _collector->_modUnionTable.verifyNoOneBitsInRange(
4835 this_span.start(),
4836 this_span.end());
4837 }
4838 pst->all_tasks_completed(); // declare that i am done
4839 }
4840
4841 // . see if we can share work_queues with ParNew? XXX
4842 void
4843 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4844 int* seed) {
4845 OopTaskQueue* work_q = work_queue(i);
4846 NOT_PRODUCT(int num_steals = 0;)
4847 oop obj_to_scan;
4848 CMSBitMap* bm = &(_collector->_markBitMap);
4849
4850 while (true) {
4851 // Completely finish any left over work from (an) earlier round(s)
4852 cl->trim_queue(0);
4853 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4854 (size_t)ParGCDesiredObjsFromOverflowList);
4855 // Now check if there's any work in the overflow list
4856 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4857 // only affects the number of attempts made to get work from the
4858 // overflow list and does not affect the number of workers. Just
4859 // pass ParallelGCThreads so this behavior is unchanged.
4860 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4861 work_q,
4862 ParallelGCThreads)) {
4863 // found something in global overflow list;
4864 // not yet ready to go stealing work from others.
4865 // We'd like to assert(work_q->size() != 0, ...)
4866 // because we just took work from the overflow list,
4867 // but of course we can't since all of that could have
4868 // been already stolen from us.
4869 // "He giveth and He taketh away."
4870 continue;
4871 }
4872 // Verify that we have no work before we resort to stealing
4873 assert(work_q->size() == 0, "Have work, shouldn't steal");
4874 // Try to steal from other queues that have work
4875 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4876 NOT_PRODUCT(num_steals++;)
4877 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4878 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4879 // Do scanning work
4880 obj_to_scan->oop_iterate(cl);
4881 // Loop around, finish this work, and try to steal some more
4882 } else if (terminator()->offer_termination()) {
4883 break; // nirvana from the infinite cycle
4884 }
4885 }
4886 NOT_PRODUCT(
4887 if (PrintCMSStatistics != 0) {
4888 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4889 }
4890 )
4891 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4892 "Else our work is not yet done");
4893 }
4894
4895 // Record object boundaries in _eden_chunk_array by sampling the eden
4896 // top in the slow-path eden object allocation code path and record
4897 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4898 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4899 // sampling in sample_eden() that activates during the part of the
4900 // preclean phase.
4901 void CMSCollector::sample_eden_chunk() {
4902 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4903 if (_eden_chunk_lock->try_lock()) {
4904 // Record a sample. This is the critical section. The contents
4905 // of the _eden_chunk_array have to be non-decreasing in the
4906 // address order.
4907 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4908 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4909 "Unexpected state of Eden");
4910 if (_eden_chunk_index == 0 ||
4911 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4912 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4913 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4914 _eden_chunk_index++; // commit sample
4915 }
4916 _eden_chunk_lock->unlock();
4917 }
4918 }
4919 }
4920
4921 // Return a thread-local PLAB recording array, as appropriate.
4922 void* CMSCollector::get_data_recorder(int thr_num) {
4923 if (_survivor_plab_array != NULL &&
4924 (CMSPLABRecordAlways ||
4925 (_collectorState > Marking && _collectorState < FinalMarking))) {
4926 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4927 ChunkArray* ca = &_survivor_plab_array[thr_num];
4928 ca->reset(); // clear it so that fresh data is recorded
4929 return (void*) ca;
4930 } else {
4931 return NULL;
4932 }
4933 }
4934
4935 // Reset all the thread-local PLAB recording arrays
4936 void CMSCollector::reset_survivor_plab_arrays() {
4937 for (uint i = 0; i < ParallelGCThreads; i++) {
4938 _survivor_plab_array[i].reset();
4939 }
4940 }
4941
4942 // Merge the per-thread plab arrays into the global survivor chunk
4943 // array which will provide the partitioning of the survivor space
4944 // for CMS initial scan and rescan.
4945 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4946 int no_of_gc_threads) {
4947 assert(_survivor_plab_array != NULL, "Error");
4948 assert(_survivor_chunk_array != NULL, "Error");
4949 assert(_collectorState == FinalMarking ||
4950 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4951 for (int j = 0; j < no_of_gc_threads; j++) {
4952 _cursor[j] = 0;
4953 }
4954 HeapWord* top = surv->top();
4955 size_t i;
4956 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4957 HeapWord* min_val = top; // Higher than any PLAB address
4958 uint min_tid = 0; // position of min_val this round
4959 for (int j = 0; j < no_of_gc_threads; j++) {
4960 ChunkArray* cur_sca = &_survivor_plab_array[j];
4961 if (_cursor[j] == cur_sca->end()) {
4962 continue;
4963 }
4964 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4965 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4966 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4967 if (cur_val < min_val) {
4968 min_tid = j;
4969 min_val = cur_val;
4970 } else {
4971 assert(cur_val < top, "All recorded addresses should be less");
4972 }
4973 }
4974 // At this point min_val and min_tid are respectively
4975 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4976 // and the thread (j) that witnesses that address.
4977 // We record this address in the _survivor_chunk_array[i]
4978 // and increment _cursor[min_tid] prior to the next round i.
4979 if (min_val == top) {
4980 break;
4981 }
4982 _survivor_chunk_array[i] = min_val;
4983 _cursor[min_tid]++;
4984 }
4985 // We are all done; record the size of the _survivor_chunk_array
4986 _survivor_chunk_index = i; // exclusive: [0, i)
4987 if (PrintCMSStatistics > 0) {
4988 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4989 }
4990 // Verify that we used up all the recorded entries
4991 #ifdef ASSERT
4992 size_t total = 0;
4993 for (int j = 0; j < no_of_gc_threads; j++) {
4994 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4995 total += _cursor[j];
4996 }
4997 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4998 // Check that the merged array is in sorted order
4999 if (total > 0) {
5000 for (size_t i = 0; i < total - 1; i++) {
5001 if (PrintCMSStatistics > 0) {
5002 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5003 i, p2i(_survivor_chunk_array[i]));
5004 }
5005 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5006 "Not sorted");
5007 }
5008 }
5009 #endif // ASSERT
5010 }
5011
5012 // Set up the space's par_seq_tasks structure for work claiming
5013 // for parallel initial scan and rescan of young gen.
5014 // See ParRescanTask where this is currently used.
5015 void
5016 CMSCollector::
5017 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5018 assert(n_threads > 0, "Unexpected n_threads argument");
5019
5020 // Eden space
5021 if (!_young_gen->eden()->is_empty()) {
5022 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5023 assert(!pst->valid(), "Clobbering existing data?");
5024 // Each valid entry in [0, _eden_chunk_index) represents a task.
5025 size_t n_tasks = _eden_chunk_index + 1;
5026 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5027 // Sets the condition for completion of the subtask (how many threads
5028 // need to finish in order to be done).
5029 pst->set_n_threads(n_threads);
5030 pst->set_n_tasks((int)n_tasks);
5031 }
5032
5033 // Merge the survivor plab arrays into _survivor_chunk_array
5034 if (_survivor_plab_array != NULL) {
5035 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5036 } else {
5037 assert(_survivor_chunk_index == 0, "Error");
5038 }
5039
5040 // To space
5041 {
5042 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5043 assert(!pst->valid(), "Clobbering existing data?");
5044 // Sets the condition for completion of the subtask (how many threads
5045 // need to finish in order to be done).
5046 pst->set_n_threads(n_threads);
5047 pst->set_n_tasks(1);
5048 assert(pst->valid(), "Error");
5049 }
5050
5051 // From space
5052 {
5053 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5054 assert(!pst->valid(), "Clobbering existing data?");
5055 size_t n_tasks = _survivor_chunk_index + 1;
5056 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5057 // Sets the condition for completion of the subtask (how many threads
5058 // need to finish in order to be done).
5059 pst->set_n_threads(n_threads);
5060 pst->set_n_tasks((int)n_tasks);
5061 assert(pst->valid(), "Error");
5062 }
5063 }
5064
5065 // Parallel version of remark
5066 void CMSCollector::do_remark_parallel() {
5067 GenCollectedHeap* gch = GenCollectedHeap::heap();
5068 FlexibleWorkGang* workers = gch->workers();
5069 assert(workers != NULL, "Need parallel worker threads.");
5070 // Choose to use the number of GC workers most recently set
5071 // into "active_workers".
5072 uint n_workers = workers->active_workers();
5073
5074 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5075
5076 StrongRootsScope srs(n_workers);
5077
5078 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
5079
5080 // We won't be iterating over the cards in the card table updating
5081 // the younger_gen cards, so we shouldn't call the following else
5082 // the verification code as well as subsequent younger_refs_iterate
5083 // code would get confused. XXX
5084 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5085
5086 // The young gen rescan work will not be done as part of
5087 // process_roots (which currently doesn't know how to
5088 // parallelize such a scan), but rather will be broken up into
5089 // a set of parallel tasks (via the sampling that the [abortable]
5090 // preclean phase did of eden, plus the [two] tasks of
5091 // scanning the [two] survivor spaces. Further fine-grain
5092 // parallelization of the scanning of the survivor spaces
5093 // themselves, and of precleaning of the younger gen itself
5094 // is deferred to the future.
5095 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5096
5097 // The dirty card rescan work is broken up into a "sequence"
5098 // of parallel tasks (per constituent space) that are dynamically
5099 // claimed by the parallel threads.
5100 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5101
5102 // It turns out that even when we're using 1 thread, doing the work in a
5103 // separate thread causes wide variance in run times. We can't help this
5104 // in the multi-threaded case, but we special-case n=1 here to get
5105 // repeatable measurements of the 1-thread overhead of the parallel code.
5106 if (n_workers > 1) {
5107 // Make refs discovery MT-safe, if it isn't already: it may not
5108 // necessarily be so, since it's possible that we are doing
5109 // ST marking.
5110 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5111 workers->run_task(&tsk);
5112 } else {
5113 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5114 tsk.work(0);
5115 }
5116
5117 // restore, single-threaded for now, any preserved marks
5118 // as a result of work_q overflow
5119 restore_preserved_marks_if_any();
5120 }
5121
5122 // Non-parallel version of remark
5123 void CMSCollector::do_remark_non_parallel() {
5124 ResourceMark rm;
5125 HandleMark hm;
5126 GenCollectedHeap* gch = GenCollectedHeap::heap();
5127 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5128
5129 MarkRefsIntoAndScanClosure
5130 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5131 &_markStack, this,
5132 false /* should_yield */, false /* not precleaning */);
5133 MarkFromDirtyCardsClosure
5134 markFromDirtyCardsClosure(this, _span,
5135 NULL, // space is set further below
5136 &_markBitMap, &_markStack, &mrias_cl);
5137 {
5138 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5139 // Iterate over the dirty cards, setting the corresponding bits in the
5140 // mod union table.
5141 {
5142 ModUnionClosure modUnionClosure(&_modUnionTable);
5143 _ct->ct_bs()->dirty_card_iterate(
5144 _cmsGen->used_region(),
5145 &modUnionClosure);
5146 }
5147 // Having transferred these marks into the modUnionTable, we just need
5148 // to rescan the marked objects on the dirty cards in the modUnionTable.
5149 // The initial marking may have been done during an asynchronous
5150 // collection so there may be dirty bits in the mod-union table.
5151 const int alignment =
5152 CardTableModRefBS::card_size * BitsPerWord;
5153 {
5154 // ... First handle dirty cards in CMS gen
5155 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5156 MemRegion ur = _cmsGen->used_region();
5157 HeapWord* lb = ur.start();
5158 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5159 MemRegion cms_span(lb, ub);
5160 _modUnionTable.dirty_range_iterate_clear(cms_span,
5161 &markFromDirtyCardsClosure);
5162 verify_work_stacks_empty();
5163 if (PrintCMSStatistics != 0) {
5164 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5165 markFromDirtyCardsClosure.num_dirty_cards());
5166 }
5167 }
5168 }
5169 if (VerifyDuringGC &&
5170 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5171 HandleMark hm; // Discard invalid handles created during verification
5172 Universe::verify();
5173 }
5174 {
5175 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5176
5177 verify_work_stacks_empty();
5178
5179 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5180 StrongRootsScope srs(1);
5181
5182 gch->gen_process_roots(&srs,
5183 Generation::Old,
5184 true, // younger gens as roots
5185 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5186 should_unload_classes(),
5187 &mrias_cl,
5188 NULL,
5189 NULL); // The dirty klasses will be handled below
5190
5191 assert(should_unload_classes()
5192 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5193 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5194 }
5195
5196 {
5197 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5198
5199 verify_work_stacks_empty();
5200
5201 // Scan all class loader data objects that might have been introduced
5202 // during concurrent marking.
5203 ResourceMark rm;
5204 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5205 for (int i = 0; i < array->length(); i++) {
5206 mrias_cl.do_class_loader_data(array->at(i));
5207 }
5208
5209 // We don't need to keep track of new CLDs anymore.
5210 ClassLoaderDataGraph::remember_new_clds(false);
5211
5212 verify_work_stacks_empty();
5213 }
5214
5215 {
5216 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5217
5218 verify_work_stacks_empty();
5219
5220 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5221 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5222
5223 verify_work_stacks_empty();
5224 }
5225
5226 // We might have added oops to ClassLoaderData::_handles during the
5227 // concurrent marking phase. These oops point to newly allocated objects
5228 // that are guaranteed to be kept alive. Either by the direct allocation
5229 // code, or when the young collector processes the roots. Hence,
5230 // we don't have to revisit the _handles block during the remark phase.
5231
5232 verify_work_stacks_empty();
5233 // Restore evacuated mark words, if any, used for overflow list links
5234 if (!CMSOverflowEarlyRestoration) {
5235 restore_preserved_marks_if_any();
5236 }
5237 verify_overflow_empty();
5238 }
5239
5240 ////////////////////////////////////////////////////////
5241 // Parallel Reference Processing Task Proxy Class
5242 ////////////////////////////////////////////////////////
5243 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5244 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5245 CMSCollector* _collector;
5246 CMSBitMap* _mark_bit_map;
5247 const MemRegion _span;
5248 ProcessTask& _task;
5249
5250 public:
5251 CMSRefProcTaskProxy(ProcessTask& task,
5252 CMSCollector* collector,
5253 const MemRegion& span,
5254 CMSBitMap* mark_bit_map,
5255 AbstractWorkGang* workers,
5256 OopTaskQueueSet* task_queues):
5257 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5258 task_queues,
5259 workers->active_workers()),
5260 _task(task),
5261 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5262 {
5263 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5264 "Inconsistency in _span");
5265 }
5266
5267 OopTaskQueueSet* task_queues() { return queues(); }
5268
5269 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5270
5271 void do_work_steal(int i,
5272 CMSParDrainMarkingStackClosure* drain,
5273 CMSParKeepAliveClosure* keep_alive,
5274 int* seed);
5275
5276 virtual void work(uint worker_id);
5277 };
5278
5279 void CMSRefProcTaskProxy::work(uint worker_id) {
5280 ResourceMark rm;
5281 HandleMark hm;
5282 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5283 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5284 _mark_bit_map,
5285 work_queue(worker_id));
5286 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5287 _mark_bit_map,
5288 work_queue(worker_id));
5289 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5290 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5291 if (_task.marks_oops_alive()) {
5292 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5293 _collector->hash_seed(worker_id));
5294 }
5295 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5296 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5297 }
5298
5299 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5300 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5301 EnqueueTask& _task;
5302
5303 public:
5304 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5305 : AbstractGangTask("Enqueue reference objects in parallel"),
5306 _task(task)
5307 { }
5308
5309 virtual void work(uint worker_id)
5310 {
5311 _task.work(worker_id);
5312 }
5313 };
5314
5315 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5316 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5317 _span(span),
5318 _bit_map(bit_map),
5319 _work_queue(work_queue),
5320 _mark_and_push(collector, span, bit_map, work_queue),
5321 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5322 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5323 { }
5324
5325 // . see if we can share work_queues with ParNew? XXX
5326 void CMSRefProcTaskProxy::do_work_steal(int i,
5327 CMSParDrainMarkingStackClosure* drain,
5328 CMSParKeepAliveClosure* keep_alive,
5329 int* seed) {
5330 OopTaskQueue* work_q = work_queue(i);
5331 NOT_PRODUCT(int num_steals = 0;)
5332 oop obj_to_scan;
5333
5334 while (true) {
5335 // Completely finish any left over work from (an) earlier round(s)
5336 drain->trim_queue(0);
5337 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5338 (size_t)ParGCDesiredObjsFromOverflowList);
5339 // Now check if there's any work in the overflow list
5340 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5341 // only affects the number of attempts made to get work from the
5342 // overflow list and does not affect the number of workers. Just
5343 // pass ParallelGCThreads so this behavior is unchanged.
5344 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5345 work_q,
5346 ParallelGCThreads)) {
5347 // Found something in global overflow list;
5348 // not yet ready to go stealing work from others.
5349 // We'd like to assert(work_q->size() != 0, ...)
5350 // because we just took work from the overflow list,
5351 // but of course we can't, since all of that might have
5352 // been already stolen from us.
5353 continue;
5354 }
5355 // Verify that we have no work before we resort to stealing
5356 assert(work_q->size() == 0, "Have work, shouldn't steal");
5357 // Try to steal from other queues that have work
5358 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5359 NOT_PRODUCT(num_steals++;)
5360 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5361 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5362 // Do scanning work
5363 obj_to_scan->oop_iterate(keep_alive);
5364 // Loop around, finish this work, and try to steal some more
5365 } else if (terminator()->offer_termination()) {
5366 break; // nirvana from the infinite cycle
5367 }
5368 }
5369 NOT_PRODUCT(
5370 if (PrintCMSStatistics != 0) {
5371 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5372 }
5373 )
5374 }
5375
5376 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5377 {
5378 GenCollectedHeap* gch = GenCollectedHeap::heap();
5379 FlexibleWorkGang* workers = gch->workers();
5380 assert(workers != NULL, "Need parallel worker threads.");
5381 CMSRefProcTaskProxy rp_task(task, &_collector,
5382 _collector.ref_processor()->span(),
5383 _collector.markBitMap(),
5384 workers, _collector.task_queues());
5385 workers->run_task(&rp_task);
5386 }
5387
5388 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5389 {
5390
5391 GenCollectedHeap* gch = GenCollectedHeap::heap();
5392 FlexibleWorkGang* workers = gch->workers();
5393 assert(workers != NULL, "Need parallel worker threads.");
5394 CMSRefEnqueueTaskProxy enq_task(task);
5395 workers->run_task(&enq_task);
5396 }
5397
5398 void CMSCollector::refProcessingWork() {
5399 ResourceMark rm;
5400 HandleMark hm;
5401
5402 ReferenceProcessor* rp = ref_processor();
5403 assert(rp->span().equals(_span), "Spans should be equal");
5404 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5405 // Process weak references.
5406 rp->setup_policy(false);
5407 verify_work_stacks_empty();
5408
5409 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5410 &_markStack, false /* !preclean */);
5411 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5412 _span, &_markBitMap, &_markStack,
5413 &cmsKeepAliveClosure, false /* !preclean */);
5414 {
5415 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5416
5417 ReferenceProcessorStats stats;
5418 if (rp->processing_is_mt()) {
5419 // Set the degree of MT here. If the discovery is done MT, there
5420 // may have been a different number of threads doing the discovery
5421 // and a different number of discovered lists may have Ref objects.
5422 // That is OK as long as the Reference lists are balanced (see
5423 // balance_all_queues() and balance_queues()).
5424 GenCollectedHeap* gch = GenCollectedHeap::heap();
5425 uint active_workers = ParallelGCThreads;
5426 FlexibleWorkGang* workers = gch->workers();
5427 if (workers != NULL) {
5428 active_workers = workers->active_workers();
5429 // The expectation is that active_workers will have already
5430 // been set to a reasonable value. If it has not been set,
5431 // investigate.
5432 assert(active_workers > 0, "Should have been set during scavenge");
5433 }
5434 rp->set_active_mt_degree(active_workers);
5435 CMSRefProcTaskExecutor task_executor(*this);
5436 stats = rp->process_discovered_references(&_is_alive_closure,
5437 &cmsKeepAliveClosure,
5438 &cmsDrainMarkingStackClosure,
5439 &task_executor,
5440 _gc_timer_cm,
5441 _gc_tracer_cm->gc_id());
5442 } else {
5443 stats = rp->process_discovered_references(&_is_alive_closure,
5444 &cmsKeepAliveClosure,
5445 &cmsDrainMarkingStackClosure,
5446 NULL,
5447 _gc_timer_cm,
5448 _gc_tracer_cm->gc_id());
5449 }
5450 _gc_tracer_cm->report_gc_reference_stats(stats);
5451
5452 }
5453
5454 // This is the point where the entire marking should have completed.
5455 verify_work_stacks_empty();
5456
5457 if (should_unload_classes()) {
5458 {
5459 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5460
5461 // Unload classes and purge the SystemDictionary.
5462 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5463
5464 // Unload nmethods.
5465 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5466
5467 // Prune dead klasses from subklass/sibling/implementor lists.
5468 Klass::clean_weak_klass_links(&_is_alive_closure);
5469 }
5470
5471 {
5472 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5473 // Clean up unreferenced symbols in symbol table.
5474 SymbolTable::unlink();
5475 }
5476
5477 {
5478 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5479 // Delete entries for dead interned strings.
5480 StringTable::unlink(&_is_alive_closure);
5481 }
5482 }
5483
5484
5485 // Restore any preserved marks as a result of mark stack or
5486 // work queue overflow
5487 restore_preserved_marks_if_any(); // done single-threaded for now
5488
5489 rp->set_enqueuing_is_done(true);
5490 if (rp->processing_is_mt()) {
5491 rp->balance_all_queues();
5492 CMSRefProcTaskExecutor task_executor(*this);
5493 rp->enqueue_discovered_references(&task_executor);
5494 } else {
5495 rp->enqueue_discovered_references(NULL);
5496 }
5497 rp->verify_no_references_recorded();
5498 assert(!rp->discovery_enabled(), "should have been disabled");
5499 }
5500
5501 #ifndef PRODUCT
5502 void CMSCollector::check_correct_thread_executing() {
5503 Thread* t = Thread::current();
5504 // Only the VM thread or the CMS thread should be here.
5505 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5506 "Unexpected thread type");
5507 // If this is the vm thread, the foreground process
5508 // should not be waiting. Note that _foregroundGCIsActive is
5509 // true while the foreground collector is waiting.
5510 if (_foregroundGCShouldWait) {
5511 // We cannot be the VM thread
5512 assert(t->is_ConcurrentGC_thread(),
5513 "Should be CMS thread");
5514 } else {
5515 // We can be the CMS thread only if we are in a stop-world
5516 // phase of CMS collection.
5517 if (t->is_ConcurrentGC_thread()) {
5518 assert(_collectorState == InitialMarking ||
5519 _collectorState == FinalMarking,
5520 "Should be a stop-world phase");
5521 // The CMS thread should be holding the CMS_token.
5522 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5523 "Potential interference with concurrently "
5524 "executing VM thread");
5525 }
5526 }
5527 }
5528 #endif
5529
5530 void CMSCollector::sweep() {
5531 assert(_collectorState == Sweeping, "just checking");
5532 check_correct_thread_executing();
5533 verify_work_stacks_empty();
5534 verify_overflow_empty();
5535 increment_sweep_count();
5536 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5537
5538 _inter_sweep_timer.stop();
5539 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5540
5541 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5542 _intra_sweep_timer.reset();
5543 _intra_sweep_timer.start();
5544 {
5545 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5546 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5547 // First sweep the old gen
5548 {
5549 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5550 bitMapLock());
5551 sweepWork(_cmsGen);
5552 }
5553
5554 // Update Universe::_heap_*_at_gc figures.
5555 // We need all the free list locks to make the abstract state
5556 // transition from Sweeping to Resetting. See detailed note
5557 // further below.
5558 {
5559 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5560 // Update heap occupancy information which is used as
5561 // input to soft ref clearing policy at the next gc.
5562 Universe::update_heap_info_at_gc();
5563 _collectorState = Resizing;
5564 }
5565 }
5566 verify_work_stacks_empty();
5567 verify_overflow_empty();
5568
5569 if (should_unload_classes()) {
5570 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5571 // requires that the virtual spaces are stable and not deleted.
5572 ClassLoaderDataGraph::set_should_purge(true);
5573 }
5574
5575 _intra_sweep_timer.stop();
5576 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5577
5578 _inter_sweep_timer.reset();
5579 _inter_sweep_timer.start();
5580
5581 // We need to use a monotonically non-decreasing time in ms
5582 // or we will see time-warp warnings and os::javaTimeMillis()
5583 // does not guarantee monotonicity.
5584 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5585 update_time_of_last_gc(now);
5586
5587 // NOTE on abstract state transitions:
5588 // Mutators allocate-live and/or mark the mod-union table dirty
5589 // based on the state of the collection. The former is done in
5590 // the interval [Marking, Sweeping] and the latter in the interval
5591 // [Marking, Sweeping). Thus the transitions into the Marking state
5592 // and out of the Sweeping state must be synchronously visible
5593 // globally to the mutators.
5594 // The transition into the Marking state happens with the world
5595 // stopped so the mutators will globally see it. Sweeping is
5596 // done asynchronously by the background collector so the transition
5597 // from the Sweeping state to the Resizing state must be done
5598 // under the freelistLock (as is the check for whether to
5599 // allocate-live and whether to dirty the mod-union table).
5600 assert(_collectorState == Resizing, "Change of collector state to"
5601 " Resizing must be done under the freelistLocks (plural)");
5602
5603 // Now that sweeping has been completed, we clear
5604 // the incremental_collection_failed flag,
5605 // thus inviting a younger gen collection to promote into
5606 // this generation. If such a promotion may still fail,
5607 // the flag will be set again when a young collection is
5608 // attempted.
5609 GenCollectedHeap* gch = GenCollectedHeap::heap();
5610 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5611 gch->update_full_collections_completed(_collection_count_start);
5612 }
5613
5614 // FIX ME!!! Looks like this belongs in CFLSpace, with
5615 // CMSGen merely delegating to it.
5616 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5617 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5618 HeapWord* minAddr = _cmsSpace->bottom();
5619 HeapWord* largestAddr =
5620 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5621 if (largestAddr == NULL) {
5622 // The dictionary appears to be empty. In this case
5623 // try to coalesce at the end of the heap.
5624 largestAddr = _cmsSpace->end();
5625 }
5626 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5627 size_t nearLargestOffset =
5628 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5629 if (PrintFLSStatistics != 0) {
5630 gclog_or_tty->print_cr(
5631 "CMS: Large Block: " PTR_FORMAT ";"
5632 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5633 p2i(largestAddr),
5634 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5635 }
5636 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5637 }
5638
5639 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5640 return addr >= _cmsSpace->nearLargestChunk();
5641 }
5642
5643 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5644 return _cmsSpace->find_chunk_at_end();
5645 }
5646
5647 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5648 bool full) {
5649 // If the young generation has been collected, gather any statistics
5650 // that are of interest at this point.
5651 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5652 if (!full && current_is_young) {
5653 // Gather statistics on the young generation collection.
5654 collector()->stats().record_gc0_end(used());
5655 }
5656 }
5657
5658 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) {
5659 // We iterate over the space(s) underlying this generation,
5660 // checking the mark bit map to see if the bits corresponding
5661 // to specific blocks are marked or not. Blocks that are
5662 // marked are live and are not swept up. All remaining blocks
5663 // are swept up, with coalescing on-the-fly as we sweep up
5664 // contiguous free and/or garbage blocks:
5665 // We need to ensure that the sweeper synchronizes with allocators
5666 // and stop-the-world collectors. In particular, the following
5667 // locks are used:
5668 // . CMS token: if this is held, a stop the world collection cannot occur
5669 // . freelistLock: if this is held no allocation can occur from this
5670 // generation by another thread
5671 // . bitMapLock: if this is held, no other thread can access or update
5672 //
5673
5674 // Note that we need to hold the freelistLock if we use
5675 // block iterate below; else the iterator might go awry if
5676 // a mutator (or promotion) causes block contents to change
5677 // (for instance if the allocator divvies up a block).
5678 // If we hold the free list lock, for all practical purposes
5679 // young generation GC's can't occur (they'll usually need to
5680 // promote), so we might as well prevent all young generation
5681 // GC's while we do a sweeping step. For the same reason, we might
5682 // as well take the bit map lock for the entire duration
5683
5684 // check that we hold the requisite locks
5685 assert(have_cms_token(), "Should hold cms token");
5686 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5687 assert_lock_strong(gen->freelistLock());
5688 assert_lock_strong(bitMapLock());
5689
5690 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5691 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5692 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5693 _inter_sweep_estimate.padded_average(),
5694 _intra_sweep_estimate.padded_average());
5695 gen->setNearLargestChunk();
5696
5697 {
5698 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield);
5699 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5700 // We need to free-up/coalesce garbage/blocks from a
5701 // co-terminal free run. This is done in the SweepClosure
5702 // destructor; so, do not remove this scope, else the
5703 // end-of-sweep-census below will be off by a little bit.
5704 }
5705 gen->cmsSpace()->sweep_completed();
5706 gen->cmsSpace()->endSweepFLCensus(sweep_count());
5707 if (should_unload_classes()) { // unloaded classes this cycle,
5708 _concurrent_cycles_since_last_unload = 0; // ... reset count
5709 } else { // did not unload classes,
5710 _concurrent_cycles_since_last_unload++; // ... increment count
5711 }
5712 }
5713
5714 // Reset CMS data structures (for now just the marking bit map)
5715 // preparatory for the next cycle.
5716 void CMSCollector::reset(bool concurrent) {
5717 if (concurrent) {
5718 CMSTokenSyncWithLocks ts(true, bitMapLock());
5719
5720 // If the state is not "Resetting", the foreground thread
5721 // has done a collection and the resetting.
5722 if (_collectorState != Resetting) {
5723 assert(_collectorState == Idling, "The state should only change"
5724 " because the foreground collector has finished the collection");
5725 return;
5726 }
5727
5728 // Clear the mark bitmap (no grey objects to start with)
5729 // for the next cycle.
5730 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5731 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5732
5733 HeapWord* curAddr = _markBitMap.startWord();
5734 while (curAddr < _markBitMap.endWord()) {
5735 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5736 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5737 _markBitMap.clear_large_range(chunk);
5738 if (ConcurrentMarkSweepThread::should_yield() &&
5739 !foregroundGCIsActive() &&
5740 CMSYield) {
5741 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5742 "CMS thread should hold CMS token");
5743 assert_lock_strong(bitMapLock());
5744 bitMapLock()->unlock();
5745 ConcurrentMarkSweepThread::desynchronize(true);
5746 stopTimer();
5747 if (PrintCMSStatistics != 0) {
5748 incrementYields();
5749 }
5750
5751 // See the comment in coordinator_yield()
5752 for (unsigned i = 0; i < CMSYieldSleepCount &&
5753 ConcurrentMarkSweepThread::should_yield() &&
5754 !CMSCollector::foregroundGCIsActive(); ++i) {
5755 os::sleep(Thread::current(), 1, false);
5756 }
5757
5758 ConcurrentMarkSweepThread::synchronize(true);
5759 bitMapLock()->lock_without_safepoint_check();
5760 startTimer();
5761 }
5762 curAddr = chunk.end();
5763 }
5764 // A successful mostly concurrent collection has been done.
5765 // Because only the full (i.e., concurrent mode failure) collections
5766 // are being measured for gc overhead limits, clean the "near" flag
5767 // and count.
5768 size_policy()->reset_gc_overhead_limit_count();
5769 _collectorState = Idling;
5770 } else {
5771 // already have the lock
5772 assert(_collectorState == Resetting, "just checking");
5773 assert_lock_strong(bitMapLock());
5774 _markBitMap.clear_all();
5775 _collectorState = Idling;
5776 }
5777
5778 register_gc_end();
5779 }
5780
5781 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5782 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5783 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5784 TraceCollectorStats tcs(counters());
5785
5786 switch (op) {
5787 case CMS_op_checkpointRootsInitial: {
5788 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5789 checkpointRootsInitial();
5790 if (PrintGC) {
5791 _cmsGen->printOccupancy("initial-mark");
5792 }
5793 break;
5794 }
5795 case CMS_op_checkpointRootsFinal: {
5796 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5797 checkpointRootsFinal();
5798 if (PrintGC) {
5799 _cmsGen->printOccupancy("remark");
5800 }
5801 break;
5802 }
5803 default:
5804 fatal("No such CMS_op");
5805 }
5806 }
5807
5808 #ifndef PRODUCT
5809 size_t const CMSCollector::skip_header_HeapWords() {
5810 return FreeChunk::header_size();
5811 }
5812
5813 // Try and collect here conditions that should hold when
5814 // CMS thread is exiting. The idea is that the foreground GC
5815 // thread should not be blocked if it wants to terminate
5816 // the CMS thread and yet continue to run the VM for a while
5817 // after that.
5818 void CMSCollector::verify_ok_to_terminate() const {
5819 assert(Thread::current()->is_ConcurrentGC_thread(),
5820 "should be called by CMS thread");
5821 assert(!_foregroundGCShouldWait, "should be false");
5822 // We could check here that all the various low-level locks
5823 // are not held by the CMS thread, but that is overkill; see
5824 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5825 // is checked.
5826 }
5827 #endif
5828
5829 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5830 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5831 "missing Printezis mark?");
5832 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5833 size_t size = pointer_delta(nextOneAddr + 1, addr);
5834 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5835 "alignment problem");
5836 assert(size >= 3, "Necessary for Printezis marks to work");
5837 return size;
5838 }
5839
5840 // A variant of the above (block_size_using_printezis_bits()) except
5841 // that we return 0 if the P-bits are not yet set.
5842 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5843 if (_markBitMap.isMarked(addr + 1)) {
5844 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5845 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5846 size_t size = pointer_delta(nextOneAddr + 1, addr);
5847 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5848 "alignment problem");
5849 assert(size >= 3, "Necessary for Printezis marks to work");
5850 return size;
5851 }
5852 return 0;
5853 }
5854
5855 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5856 size_t sz = 0;
5857 oop p = (oop)addr;
5858 if (p->klass_or_null() != NULL) {
5859 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5860 } else {
5861 sz = block_size_using_printezis_bits(addr);
5862 }
5863 assert(sz > 0, "size must be nonzero");
5864 HeapWord* next_block = addr + sz;
5865 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5866 CardTableModRefBS::card_size);
5867 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5868 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5869 "must be different cards");
5870 return next_card;
5871 }
5872
5873
5874 // CMS Bit Map Wrapper /////////////////////////////////////////
5875
5876 // Construct a CMS bit map infrastructure, but don't create the
5877 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5878 // further below.
5879 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5880 _bm(),
5881 _shifter(shifter),
5882 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5883 Monitor::_safepoint_check_sometimes) : NULL)
5884 {
5885 _bmStartWord = 0;
5886 _bmWordSize = 0;
5887 }
5888
5889 bool CMSBitMap::allocate(MemRegion mr) {
5890 _bmStartWord = mr.start();
5891 _bmWordSize = mr.word_size();
5892 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5893 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5894 if (!brs.is_reserved()) {
5895 warning("CMS bit map allocation failure");
5896 return false;
5897 }
5898 // For now we'll just commit all of the bit map up front.
5899 // Later on we'll try to be more parsimonious with swap.
5900 if (!_virtual_space.initialize(brs, brs.size())) {
5901 warning("CMS bit map backing store failure");
5902 return false;
5903 }
5904 assert(_virtual_space.committed_size() == brs.size(),
5905 "didn't reserve backing store for all of CMS bit map?");
5906 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5907 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5908 _bmWordSize, "inconsistency in bit map sizing");
5909 _bm.set_size(_bmWordSize >> _shifter);
5910
5911 // bm.clear(); // can we rely on getting zero'd memory? verify below
5912 assert(isAllClear(),
5913 "Expected zero'd memory from ReservedSpace constructor");
5914 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5915 "consistency check");
5916 return true;
5917 }
5918
5919 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5920 HeapWord *next_addr, *end_addr, *last_addr;
5921 assert_locked();
5922 assert(covers(mr), "out-of-range error");
5923 // XXX assert that start and end are appropriately aligned
5924 for (next_addr = mr.start(), end_addr = mr.end();
5925 next_addr < end_addr; next_addr = last_addr) {
5926 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5927 last_addr = dirty_region.end();
5928 if (!dirty_region.is_empty()) {
5929 cl->do_MemRegion(dirty_region);
5930 } else {
5931 assert(last_addr == end_addr, "program logic");
5932 return;
5933 }
5934 }
5935 }
5936
5937 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5938 _bm.print_on_error(st, prefix);
5939 }
5940
5941 #ifndef PRODUCT
5942 void CMSBitMap::assert_locked() const {
5943 CMSLockVerifier::assert_locked(lock());
5944 }
5945
5946 bool CMSBitMap::covers(MemRegion mr) const {
5947 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5948 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5949 "size inconsistency");
5950 return (mr.start() >= _bmStartWord) &&
5951 (mr.end() <= endWord());
5952 }
5953
5954 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5955 return (start >= _bmStartWord && (start + size) <= endWord());
5956 }
5957
5958 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5959 // verify that there are no 1 bits in the interval [left, right)
5960 FalseBitMapClosure falseBitMapClosure;
5961 iterate(&falseBitMapClosure, left, right);
5962 }
5963
5964 void CMSBitMap::region_invariant(MemRegion mr)
5965 {
5966 assert_locked();
5967 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5968 assert(!mr.is_empty(), "unexpected empty region");
5969 assert(covers(mr), "mr should be covered by bit map");
5970 // convert address range into offset range
5971 size_t start_ofs = heapWordToOffset(mr.start());
5972 // Make sure that end() is appropriately aligned
5973 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5974 (1 << (_shifter+LogHeapWordSize))),
5975 "Misaligned mr.end()");
5976 size_t end_ofs = heapWordToOffset(mr.end());
5977 assert(end_ofs > start_ofs, "Should mark at least one bit");
5978 }
5979
5980 #endif
5981
5982 bool CMSMarkStack::allocate(size_t size) {
5983 // allocate a stack of the requisite depth
5984 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5985 size * sizeof(oop)));
5986 if (!rs.is_reserved()) {
5987 warning("CMSMarkStack allocation failure");
5988 return false;
5989 }
5990 if (!_virtual_space.initialize(rs, rs.size())) {
5991 warning("CMSMarkStack backing store failure");
5992 return false;
5993 }
5994 assert(_virtual_space.committed_size() == rs.size(),
5995 "didn't reserve backing store for all of CMS stack?");
5996 _base = (oop*)(_virtual_space.low());
5997 _index = 0;
5998 _capacity = size;
5999 NOT_PRODUCT(_max_depth = 0);
6000 return true;
6001 }
6002
6003 // XXX FIX ME !!! In the MT case we come in here holding a
6004 // leaf lock. For printing we need to take a further lock
6005 // which has lower rank. We need to recalibrate the two
6006 // lock-ranks involved in order to be able to print the
6007 // messages below. (Or defer the printing to the caller.
6008 // For now we take the expedient path of just disabling the
6009 // messages for the problematic case.)
6010 void CMSMarkStack::expand() {
6011 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6012 if (_capacity == MarkStackSizeMax) {
6013 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6014 // We print a warning message only once per CMS cycle.
6015 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6016 }
6017 return;
6018 }
6019 // Double capacity if possible
6020 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6021 // Do not give up existing stack until we have managed to
6022 // get the double capacity that we desired.
6023 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6024 new_capacity * sizeof(oop)));
6025 if (rs.is_reserved()) {
6026 // Release the backing store associated with old stack
6027 _virtual_space.release();
6028 // Reinitialize virtual space for new stack
6029 if (!_virtual_space.initialize(rs, rs.size())) {
6030 fatal("Not enough swap for expanded marking stack");
6031 }
6032 _base = (oop*)(_virtual_space.low());
6033 _index = 0;
6034 _capacity = new_capacity;
6035 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6036 // Failed to double capacity, continue;
6037 // we print a detail message only once per CMS cycle.
6038 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6039 SIZE_FORMAT"K",
6040 _capacity / K, new_capacity / K);
6041 }
6042 }
6043
6044
6045 // Closures
6046 // XXX: there seems to be a lot of code duplication here;
6047 // should refactor and consolidate common code.
6048
6049 // This closure is used to mark refs into the CMS generation in
6050 // the CMS bit map. Called at the first checkpoint. This closure
6051 // assumes that we do not need to re-mark dirty cards; if the CMS
6052 // generation on which this is used is not an oldest
6053 // generation then this will lose younger_gen cards!
6054
6055 MarkRefsIntoClosure::MarkRefsIntoClosure(
6056 MemRegion span, CMSBitMap* bitMap):
6057 _span(span),
6058 _bitMap(bitMap)
6059 {
6060 assert(_ref_processor == NULL, "deliberately left NULL");
6061 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6062 }
6063
6064 void MarkRefsIntoClosure::do_oop(oop obj) {
6065 // if p points into _span, then mark corresponding bit in _markBitMap
6066 assert(obj->is_oop(), "expected an oop");
6067 HeapWord* addr = (HeapWord*)obj;
6068 if (_span.contains(addr)) {
6069 // this should be made more efficient
6070 _bitMap->mark(addr);
6071 }
6072 }
6073
6074 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6075 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6076
6077 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6078 MemRegion span, CMSBitMap* bitMap):
6079 _span(span),
6080 _bitMap(bitMap)
6081 {
6082 assert(_ref_processor == NULL, "deliberately left NULL");
6083 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6084 }
6085
6086 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6087 // if p points into _span, then mark corresponding bit in _markBitMap
6088 assert(obj->is_oop(), "expected an oop");
6089 HeapWord* addr = (HeapWord*)obj;
6090 if (_span.contains(addr)) {
6091 // this should be made more efficient
6092 _bitMap->par_mark(addr);
6093 }
6094 }
6095
6096 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6097 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6098
6099 // A variant of the above, used for CMS marking verification.
6100 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6101 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6102 _span(span),
6103 _verification_bm(verification_bm),
6104 _cms_bm(cms_bm)
6105 {
6106 assert(_ref_processor == NULL, "deliberately left NULL");
6107 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6108 }
6109
6110 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6111 // if p points into _span, then mark corresponding bit in _markBitMap
6112 assert(obj->is_oop(), "expected an oop");
6113 HeapWord* addr = (HeapWord*)obj;
6114 if (_span.contains(addr)) {
6115 _verification_bm->mark(addr);
6116 if (!_cms_bm->isMarked(addr)) {
6117 oop(addr)->print();
6118 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6119 fatal("... aborting");
6120 }
6121 }
6122 }
6123
6124 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6125 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6126
6127 //////////////////////////////////////////////////
6128 // MarkRefsIntoAndScanClosure
6129 //////////////////////////////////////////////////
6130
6131 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6132 ReferenceProcessor* rp,
6133 CMSBitMap* bit_map,
6134 CMSBitMap* mod_union_table,
6135 CMSMarkStack* mark_stack,
6136 CMSCollector* collector,
6137 bool should_yield,
6138 bool concurrent_precleaning):
6139 _collector(collector),
6140 _span(span),
6141 _bit_map(bit_map),
6142 _mark_stack(mark_stack),
6143 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6144 mark_stack, concurrent_precleaning),
6145 _yield(should_yield),
6146 _concurrent_precleaning(concurrent_precleaning),
6147 _freelistLock(NULL)
6148 {
6149 _ref_processor = rp;
6150 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6151 }
6152
6153 // This closure is used to mark refs into the CMS generation at the
6154 // second (final) checkpoint, and to scan and transitively follow
6155 // the unmarked oops. It is also used during the concurrent precleaning
6156 // phase while scanning objects on dirty cards in the CMS generation.
6157 // The marks are made in the marking bit map and the marking stack is
6158 // used for keeping the (newly) grey objects during the scan.
6159 // The parallel version (Par_...) appears further below.
6160 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6161 if (obj != NULL) {
6162 assert(obj->is_oop(), "expected an oop");
6163 HeapWord* addr = (HeapWord*)obj;
6164 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6165 assert(_collector->overflow_list_is_empty(),
6166 "overflow list should be empty");
6167 if (_span.contains(addr) &&
6168 !_bit_map->isMarked(addr)) {
6169 // mark bit map (object is now grey)
6170 _bit_map->mark(addr);
6171 // push on marking stack (stack should be empty), and drain the
6172 // stack by applying this closure to the oops in the oops popped
6173 // from the stack (i.e. blacken the grey objects)
6174 bool res = _mark_stack->push(obj);
6175 assert(res, "Should have space to push on empty stack");
6176 do {
6177 oop new_oop = _mark_stack->pop();
6178 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6179 assert(_bit_map->isMarked((HeapWord*)new_oop),
6180 "only grey objects on this stack");
6181 // iterate over the oops in this oop, marking and pushing
6182 // the ones in CMS heap (i.e. in _span).
6183 new_oop->oop_iterate(&_pushAndMarkClosure);
6184 // check if it's time to yield
6185 do_yield_check();
6186 } while (!_mark_stack->isEmpty() ||
6187 (!_concurrent_precleaning && take_from_overflow_list()));
6188 // if marking stack is empty, and we are not doing this
6189 // during precleaning, then check the overflow list
6190 }
6191 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6192 assert(_collector->overflow_list_is_empty(),
6193 "overflow list was drained above");
6194 // We could restore evacuated mark words, if any, used for
6195 // overflow list links here because the overflow list is
6196 // provably empty here. That would reduce the maximum
6197 // size requirements for preserved_{oop,mark}_stack.
6198 // But we'll just postpone it until we are all done
6199 // so we can just stream through.
6200 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6201 _collector->restore_preserved_marks_if_any();
6202 assert(_collector->no_preserved_marks(), "No preserved marks");
6203 }
6204 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6205 "All preserved marks should have been restored above");
6206 }
6207 }
6208
6209 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6210 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6211
6212 void MarkRefsIntoAndScanClosure::do_yield_work() {
6213 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6214 "CMS thread should hold CMS token");
6215 assert_lock_strong(_freelistLock);
6216 assert_lock_strong(_bit_map->lock());
6217 // relinquish the free_list_lock and bitMaplock()
6218 _bit_map->lock()->unlock();
6219 _freelistLock->unlock();
6220 ConcurrentMarkSweepThread::desynchronize(true);
6221 _collector->stopTimer();
6222 if (PrintCMSStatistics != 0) {
6223 _collector->incrementYields();
6224 }
6225
6226 // See the comment in coordinator_yield()
6227 for (unsigned i = 0;
6228 i < CMSYieldSleepCount &&
6229 ConcurrentMarkSweepThread::should_yield() &&
6230 !CMSCollector::foregroundGCIsActive();
6231 ++i) {
6232 os::sleep(Thread::current(), 1, false);
6233 }
6234
6235 ConcurrentMarkSweepThread::synchronize(true);
6236 _freelistLock->lock_without_safepoint_check();
6237 _bit_map->lock()->lock_without_safepoint_check();
6238 _collector->startTimer();
6239 }
6240
6241 ///////////////////////////////////////////////////////////
6242 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6243 // MarkRefsIntoAndScanClosure
6244 ///////////////////////////////////////////////////////////
6245 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6246 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6247 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6248 _span(span),
6249 _bit_map(bit_map),
6250 _work_queue(work_queue),
6251 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6252 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6253 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6254 {
6255 _ref_processor = rp;
6256 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6257 }
6258
6259 // This closure is used to mark refs into the CMS generation at the
6260 // second (final) checkpoint, and to scan and transitively follow
6261 // the unmarked oops. The marks are made in the marking bit map and
6262 // the work_queue is used for keeping the (newly) grey objects during
6263 // the scan phase whence they are also available for stealing by parallel
6264 // threads. Since the marking bit map is shared, updates are
6265 // synchronized (via CAS).
6266 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6267 if (obj != NULL) {
6268 // Ignore mark word because this could be an already marked oop
6269 // that may be chained at the end of the overflow list.
6270 assert(obj->is_oop(true), "expected an oop");
6271 HeapWord* addr = (HeapWord*)obj;
6272 if (_span.contains(addr) &&
6273 !_bit_map->isMarked(addr)) {
6274 // mark bit map (object will become grey):
6275 // It is possible for several threads to be
6276 // trying to "claim" this object concurrently;
6277 // the unique thread that succeeds in marking the
6278 // object first will do the subsequent push on
6279 // to the work queue (or overflow list).
6280 if (_bit_map->par_mark(addr)) {
6281 // push on work_queue (which may not be empty), and trim the
6282 // queue to an appropriate length by applying this closure to
6283 // the oops in the oops popped from the stack (i.e. blacken the
6284 // grey objects)
6285 bool res = _work_queue->push(obj);
6286 assert(res, "Low water mark should be less than capacity?");
6287 trim_queue(_low_water_mark);
6288 } // Else, another thread claimed the object
6289 }
6290 }
6291 }
6292
6293 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6294 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6295
6296 // This closure is used to rescan the marked objects on the dirty cards
6297 // in the mod union table and the card table proper.
6298 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6299 oop p, MemRegion mr) {
6300
6301 size_t size = 0;
6302 HeapWord* addr = (HeapWord*)p;
6303 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6304 assert(_span.contains(addr), "we are scanning the CMS generation");
6305 // check if it's time to yield
6306 if (do_yield_check()) {
6307 // We yielded for some foreground stop-world work,
6308 // and we have been asked to abort this ongoing preclean cycle.
6309 return 0;
6310 }
6311 if (_bitMap->isMarked(addr)) {
6312 // it's marked; is it potentially uninitialized?
6313 if (p->klass_or_null() != NULL) {
6314 // an initialized object; ignore mark word in verification below
6315 // since we are running concurrent with mutators
6316 assert(p->is_oop(true), "should be an oop");
6317 if (p->is_objArray()) {
6318 // objArrays are precisely marked; restrict scanning
6319 // to dirty cards only.
6320 size = CompactibleFreeListSpace::adjustObjectSize(
6321 p->oop_iterate(_scanningClosure, mr));
6322 } else {
6323 // A non-array may have been imprecisely marked; we need
6324 // to scan object in its entirety.
6325 size = CompactibleFreeListSpace::adjustObjectSize(
6326 p->oop_iterate(_scanningClosure));
6327 }
6328 #ifdef ASSERT
6329 size_t direct_size =
6330 CompactibleFreeListSpace::adjustObjectSize(p->size());
6331 assert(size == direct_size, "Inconsistency in size");
6332 assert(size >= 3, "Necessary for Printezis marks to work");
6333 if (!_bitMap->isMarked(addr+1)) {
6334 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6335 } else {
6336 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6337 assert(_bitMap->isMarked(addr+size-1),
6338 "inconsistent Printezis mark");
6339 }
6340 #endif // ASSERT
6341 } else {
6342 // An uninitialized object.
6343 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6344 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6345 size = pointer_delta(nextOneAddr + 1, addr);
6346 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6347 "alignment problem");
6348 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6349 // will dirty the card when the klass pointer is installed in the
6350 // object (signaling the completion of initialization).
6351 }
6352 } else {
6353 // Either a not yet marked object or an uninitialized object
6354 if (p->klass_or_null() == NULL) {
6355 // An uninitialized object, skip to the next card, since
6356 // we may not be able to read its P-bits yet.
6357 assert(size == 0, "Initial value");
6358 } else {
6359 // An object not (yet) reached by marking: we merely need to
6360 // compute its size so as to go look at the next block.
6361 assert(p->is_oop(true), "should be an oop");
6362 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6363 }
6364 }
6365 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6366 return size;
6367 }
6368
6369 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6370 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6371 "CMS thread should hold CMS token");
6372 assert_lock_strong(_freelistLock);
6373 assert_lock_strong(_bitMap->lock());
6374 // relinquish the free_list_lock and bitMaplock()
6375 _bitMap->lock()->unlock();
6376 _freelistLock->unlock();
6377 ConcurrentMarkSweepThread::desynchronize(true);
6378 _collector->stopTimer();
6379 if (PrintCMSStatistics != 0) {
6380 _collector->incrementYields();
6381 }
6382
6383 // See the comment in coordinator_yield()
6384 for (unsigned i = 0; i < CMSYieldSleepCount &&
6385 ConcurrentMarkSweepThread::should_yield() &&
6386 !CMSCollector::foregroundGCIsActive(); ++i) {
6387 os::sleep(Thread::current(), 1, false);
6388 }
6389
6390 ConcurrentMarkSweepThread::synchronize(true);
6391 _freelistLock->lock_without_safepoint_check();
6392 _bitMap->lock()->lock_without_safepoint_check();
6393 _collector->startTimer();
6394 }
6395
6396
6397 //////////////////////////////////////////////////////////////////
6398 // SurvivorSpacePrecleanClosure
6399 //////////////////////////////////////////////////////////////////
6400 // This (single-threaded) closure is used to preclean the oops in
6401 // the survivor spaces.
6402 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6403
6404 HeapWord* addr = (HeapWord*)p;
6405 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6406 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6407 assert(p->klass_or_null() != NULL, "object should be initialized");
6408 // an initialized object; ignore mark word in verification below
6409 // since we are running concurrent with mutators
6410 assert(p->is_oop(true), "should be an oop");
6411 // Note that we do not yield while we iterate over
6412 // the interior oops of p, pushing the relevant ones
6413 // on our marking stack.
6414 size_t size = p->oop_iterate(_scanning_closure);
6415 do_yield_check();
6416 // Observe that below, we do not abandon the preclean
6417 // phase as soon as we should; rather we empty the
6418 // marking stack before returning. This is to satisfy
6419 // some existing assertions. In general, it may be a
6420 // good idea to abort immediately and complete the marking
6421 // from the grey objects at a later time.
6422 while (!_mark_stack->isEmpty()) {
6423 oop new_oop = _mark_stack->pop();
6424 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6425 assert(_bit_map->isMarked((HeapWord*)new_oop),
6426 "only grey objects on this stack");
6427 // iterate over the oops in this oop, marking and pushing
6428 // the ones in CMS heap (i.e. in _span).
6429 new_oop->oop_iterate(_scanning_closure);
6430 // check if it's time to yield
6431 do_yield_check();
6432 }
6433 unsigned int after_count =
6434 GenCollectedHeap::heap()->total_collections();
6435 bool abort = (_before_count != after_count) ||
6436 _collector->should_abort_preclean();
6437 return abort ? 0 : size;
6438 }
6439
6440 void SurvivorSpacePrecleanClosure::do_yield_work() {
6441 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6442 "CMS thread should hold CMS token");
6443 assert_lock_strong(_bit_map->lock());
6444 // Relinquish the bit map lock
6445 _bit_map->lock()->unlock();
6446 ConcurrentMarkSweepThread::desynchronize(true);
6447 _collector->stopTimer();
6448 if (PrintCMSStatistics != 0) {
6449 _collector->incrementYields();
6450 }
6451
6452 // See the comment in coordinator_yield()
6453 for (unsigned i = 0; i < CMSYieldSleepCount &&
6454 ConcurrentMarkSweepThread::should_yield() &&
6455 !CMSCollector::foregroundGCIsActive(); ++i) {
6456 os::sleep(Thread::current(), 1, false);
6457 }
6458
6459 ConcurrentMarkSweepThread::synchronize(true);
6460 _bit_map->lock()->lock_without_safepoint_check();
6461 _collector->startTimer();
6462 }
6463
6464 // This closure is used to rescan the marked objects on the dirty cards
6465 // in the mod union table and the card table proper. In the parallel
6466 // case, although the bitMap is shared, we do a single read so the
6467 // isMarked() query is "safe".
6468 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6469 // Ignore mark word because we are running concurrent with mutators
6470 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6471 HeapWord* addr = (HeapWord*)p;
6472 assert(_span.contains(addr), "we are scanning the CMS generation");
6473 bool is_obj_array = false;
6474 #ifdef ASSERT
6475 if (!_parallel) {
6476 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6477 assert(_collector->overflow_list_is_empty(),
6478 "overflow list should be empty");
6479
6480 }
6481 #endif // ASSERT
6482 if (_bit_map->isMarked(addr)) {
6483 // Obj arrays are precisely marked, non-arrays are not;
6484 // so we scan objArrays precisely and non-arrays in their
6485 // entirety.
6486 if (p->is_objArray()) {
6487 is_obj_array = true;
6488 if (_parallel) {
6489 p->oop_iterate(_par_scan_closure, mr);
6490 } else {
6491 p->oop_iterate(_scan_closure, mr);
6492 }
6493 } else {
6494 if (_parallel) {
6495 p->oop_iterate(_par_scan_closure);
6496 } else {
6497 p->oop_iterate(_scan_closure);
6498 }
6499 }
6500 }
6501 #ifdef ASSERT
6502 if (!_parallel) {
6503 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6504 assert(_collector->overflow_list_is_empty(),
6505 "overflow list should be empty");
6506
6507 }
6508 #endif // ASSERT
6509 return is_obj_array;
6510 }
6511
6512 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6513 MemRegion span,
6514 CMSBitMap* bitMap, CMSMarkStack* markStack,
6515 bool should_yield, bool verifying):
6516 _collector(collector),
6517 _span(span),
6518 _bitMap(bitMap),
6519 _mut(&collector->_modUnionTable),
6520 _markStack(markStack),
6521 _yield(should_yield),
6522 _skipBits(0)
6523 {
6524 assert(_markStack->isEmpty(), "stack should be empty");
6525 _finger = _bitMap->startWord();
6526 _threshold = _finger;
6527 assert(_collector->_restart_addr == NULL, "Sanity check");
6528 assert(_span.contains(_finger), "Out of bounds _finger?");
6529 DEBUG_ONLY(_verifying = verifying;)
6530 }
6531
6532 void MarkFromRootsClosure::reset(HeapWord* addr) {
6533 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6534 assert(_span.contains(addr), "Out of bounds _finger?");
6535 _finger = addr;
6536 _threshold = (HeapWord*)round_to(
6537 (intptr_t)_finger, CardTableModRefBS::card_size);
6538 }
6539
6540 // Should revisit to see if this should be restructured for
6541 // greater efficiency.
6542 bool MarkFromRootsClosure::do_bit(size_t offset) {
6543 if (_skipBits > 0) {
6544 _skipBits--;
6545 return true;
6546 }
6547 // convert offset into a HeapWord*
6548 HeapWord* addr = _bitMap->startWord() + offset;
6549 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6550 "address out of range");
6551 assert(_bitMap->isMarked(addr), "tautology");
6552 if (_bitMap->isMarked(addr+1)) {
6553 // this is an allocated but not yet initialized object
6554 assert(_skipBits == 0, "tautology");
6555 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6556 oop p = oop(addr);
6557 if (p->klass_or_null() == NULL) {
6558 DEBUG_ONLY(if (!_verifying) {)
6559 // We re-dirty the cards on which this object lies and increase
6560 // the _threshold so that we'll come back to scan this object
6561 // during the preclean or remark phase. (CMSCleanOnEnter)
6562 if (CMSCleanOnEnter) {
6563 size_t sz = _collector->block_size_using_printezis_bits(addr);
6564 HeapWord* end_card_addr = (HeapWord*)round_to(
6565 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6566 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6567 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6568 // Bump _threshold to end_card_addr; note that
6569 // _threshold cannot possibly exceed end_card_addr, anyhow.
6570 // This prevents future clearing of the card as the scan proceeds
6571 // to the right.
6572 assert(_threshold <= end_card_addr,
6573 "Because we are just scanning into this object");
6574 if (_threshold < end_card_addr) {
6575 _threshold = end_card_addr;
6576 }
6577 if (p->klass_or_null() != NULL) {
6578 // Redirty the range of cards...
6579 _mut->mark_range(redirty_range);
6580 } // ...else the setting of klass will dirty the card anyway.
6581 }
6582 DEBUG_ONLY(})
6583 return true;
6584 }
6585 }
6586 scanOopsInOop(addr);
6587 return true;
6588 }
6589
6590 // We take a break if we've been at this for a while,
6591 // so as to avoid monopolizing the locks involved.
6592 void MarkFromRootsClosure::do_yield_work() {
6593 // First give up the locks, then yield, then re-lock
6594 // We should probably use a constructor/destructor idiom to
6595 // do this unlock/lock or modify the MutexUnlocker class to
6596 // serve our purpose. XXX
6597 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6598 "CMS thread should hold CMS token");
6599 assert_lock_strong(_bitMap->lock());
6600 _bitMap->lock()->unlock();
6601 ConcurrentMarkSweepThread::desynchronize(true);
6602 _collector->stopTimer();
6603 if (PrintCMSStatistics != 0) {
6604 _collector->incrementYields();
6605 }
6606
6607 // See the comment in coordinator_yield()
6608 for (unsigned i = 0; i < CMSYieldSleepCount &&
6609 ConcurrentMarkSweepThread::should_yield() &&
6610 !CMSCollector::foregroundGCIsActive(); ++i) {
6611 os::sleep(Thread::current(), 1, false);
6612 }
6613
6614 ConcurrentMarkSweepThread::synchronize(true);
6615 _bitMap->lock()->lock_without_safepoint_check();
6616 _collector->startTimer();
6617 }
6618
6619 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6620 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6621 assert(_markStack->isEmpty(),
6622 "should drain stack to limit stack usage");
6623 // convert ptr to an oop preparatory to scanning
6624 oop obj = oop(ptr);
6625 // Ignore mark word in verification below, since we
6626 // may be running concurrent with mutators.
6627 assert(obj->is_oop(true), "should be an oop");
6628 assert(_finger <= ptr, "_finger runneth ahead");
6629 // advance the finger to right end of this object
6630 _finger = ptr + obj->size();
6631 assert(_finger > ptr, "we just incremented it above");
6632 // On large heaps, it may take us some time to get through
6633 // the marking phase. During
6634 // this time it's possible that a lot of mutations have
6635 // accumulated in the card table and the mod union table --
6636 // these mutation records are redundant until we have
6637 // actually traced into the corresponding card.
6638 // Here, we check whether advancing the finger would make
6639 // us cross into a new card, and if so clear corresponding
6640 // cards in the MUT (preclean them in the card-table in the
6641 // future).
6642
6643 DEBUG_ONLY(if (!_verifying) {)
6644 // The clean-on-enter optimization is disabled by default,
6645 // until we fix 6178663.
6646 if (CMSCleanOnEnter && (_finger > _threshold)) {
6647 // [_threshold, _finger) represents the interval
6648 // of cards to be cleared in MUT (or precleaned in card table).
6649 // The set of cards to be cleared is all those that overlap
6650 // with the interval [_threshold, _finger); note that
6651 // _threshold is always kept card-aligned but _finger isn't
6652 // always card-aligned.
6653 HeapWord* old_threshold = _threshold;
6654 assert(old_threshold == (HeapWord*)round_to(
6655 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6656 "_threshold should always be card-aligned");
6657 _threshold = (HeapWord*)round_to(
6658 (intptr_t)_finger, CardTableModRefBS::card_size);
6659 MemRegion mr(old_threshold, _threshold);
6660 assert(!mr.is_empty(), "Control point invariant");
6661 assert(_span.contains(mr), "Should clear within span");
6662 _mut->clear_range(mr);
6663 }
6664 DEBUG_ONLY(})
6665 // Note: the finger doesn't advance while we drain
6666 // the stack below.
6667 PushOrMarkClosure pushOrMarkClosure(_collector,
6668 _span, _bitMap, _markStack,
6669 _finger, this);
6670 bool res = _markStack->push(obj);
6671 assert(res, "Empty non-zero size stack should have space for single push");
6672 while (!_markStack->isEmpty()) {
6673 oop new_oop = _markStack->pop();
6674 // Skip verifying header mark word below because we are
6675 // running concurrent with mutators.
6676 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6677 // now scan this oop's oops
6678 new_oop->oop_iterate(&pushOrMarkClosure);
6679 do_yield_check();
6680 }
6681 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6682 }
6683
6684 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6685 CMSCollector* collector, MemRegion span,
6686 CMSBitMap* bit_map,
6687 OopTaskQueue* work_queue,
6688 CMSMarkStack* overflow_stack):
6689 _collector(collector),
6690 _whole_span(collector->_span),
6691 _span(span),
6692 _bit_map(bit_map),
6693 _mut(&collector->_modUnionTable),
6694 _work_queue(work_queue),
6695 _overflow_stack(overflow_stack),
6696 _skip_bits(0),
6697 _task(task)
6698 {
6699 assert(_work_queue->size() == 0, "work_queue should be empty");
6700 _finger = span.start();
6701 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6702 assert(_span.contains(_finger), "Out of bounds _finger?");
6703 }
6704
6705 // Should revisit to see if this should be restructured for
6706 // greater efficiency.
6707 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6708 if (_skip_bits > 0) {
6709 _skip_bits--;
6710 return true;
6711 }
6712 // convert offset into a HeapWord*
6713 HeapWord* addr = _bit_map->startWord() + offset;
6714 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6715 "address out of range");
6716 assert(_bit_map->isMarked(addr), "tautology");
6717 if (_bit_map->isMarked(addr+1)) {
6718 // this is an allocated object that might not yet be initialized
6719 assert(_skip_bits == 0, "tautology");
6720 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6721 oop p = oop(addr);
6722 if (p->klass_or_null() == NULL) {
6723 // in the case of Clean-on-Enter optimization, redirty card
6724 // and avoid clearing card by increasing the threshold.
6725 return true;
6726 }
6727 }
6728 scan_oops_in_oop(addr);
6729 return true;
6730 }
6731
6732 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6733 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6734 // Should we assert that our work queue is empty or
6735 // below some drain limit?
6736 assert(_work_queue->size() == 0,
6737 "should drain stack to limit stack usage");
6738 // convert ptr to an oop preparatory to scanning
6739 oop obj = oop(ptr);
6740 // Ignore mark word in verification below, since we
6741 // may be running concurrent with mutators.
6742 assert(obj->is_oop(true), "should be an oop");
6743 assert(_finger <= ptr, "_finger runneth ahead");
6744 // advance the finger to right end of this object
6745 _finger = ptr + obj->size();
6746 assert(_finger > ptr, "we just incremented it above");
6747 // On large heaps, it may take us some time to get through
6748 // the marking phase. During
6749 // this time it's possible that a lot of mutations have
6750 // accumulated in the card table and the mod union table --
6751 // these mutation records are redundant until we have
6752 // actually traced into the corresponding card.
6753 // Here, we check whether advancing the finger would make
6754 // us cross into a new card, and if so clear corresponding
6755 // cards in the MUT (preclean them in the card-table in the
6756 // future).
6757
6758 // The clean-on-enter optimization is disabled by default,
6759 // until we fix 6178663.
6760 if (CMSCleanOnEnter && (_finger > _threshold)) {
6761 // [_threshold, _finger) represents the interval
6762 // of cards to be cleared in MUT (or precleaned in card table).
6763 // The set of cards to be cleared is all those that overlap
6764 // with the interval [_threshold, _finger); note that
6765 // _threshold is always kept card-aligned but _finger isn't
6766 // always card-aligned.
6767 HeapWord* old_threshold = _threshold;
6768 assert(old_threshold == (HeapWord*)round_to(
6769 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6770 "_threshold should always be card-aligned");
6771 _threshold = (HeapWord*)round_to(
6772 (intptr_t)_finger, CardTableModRefBS::card_size);
6773 MemRegion mr(old_threshold, _threshold);
6774 assert(!mr.is_empty(), "Control point invariant");
6775 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6776 _mut->clear_range(mr);
6777 }
6778
6779 // Note: the local finger doesn't advance while we drain
6780 // the stack below, but the global finger sure can and will.
6781 HeapWord** gfa = _task->global_finger_addr();
6782 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6783 _span, _bit_map,
6784 _work_queue,
6785 _overflow_stack,
6786 _finger,
6787 gfa, this);
6788 bool res = _work_queue->push(obj); // overflow could occur here
6789 assert(res, "Will hold once we use workqueues");
6790 while (true) {
6791 oop new_oop;
6792 if (!_work_queue->pop_local(new_oop)) {
6793 // We emptied our work_queue; check if there's stuff that can
6794 // be gotten from the overflow stack.
6795 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6796 _overflow_stack, _work_queue)) {
6797 do_yield_check();
6798 continue;
6799 } else { // done
6800 break;
6801 }
6802 }
6803 // Skip verifying header mark word below because we are
6804 // running concurrent with mutators.
6805 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6806 // now scan this oop's oops
6807 new_oop->oop_iterate(&pushOrMarkClosure);
6808 do_yield_check();
6809 }
6810 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6811 }
6812
6813 // Yield in response to a request from VM Thread or
6814 // from mutators.
6815 void Par_MarkFromRootsClosure::do_yield_work() {
6816 assert(_task != NULL, "sanity");
6817 _task->yield();
6818 }
6819
6820 // A variant of the above used for verifying CMS marking work.
6821 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6822 MemRegion span,
6823 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6824 CMSMarkStack* mark_stack):
6825 _collector(collector),
6826 _span(span),
6827 _verification_bm(verification_bm),
6828 _cms_bm(cms_bm),
6829 _mark_stack(mark_stack),
6830 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6831 mark_stack)
6832 {
6833 assert(_mark_stack->isEmpty(), "stack should be empty");
6834 _finger = _verification_bm->startWord();
6835 assert(_collector->_restart_addr == NULL, "Sanity check");
6836 assert(_span.contains(_finger), "Out of bounds _finger?");
6837 }
6838
6839 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6840 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6841 assert(_span.contains(addr), "Out of bounds _finger?");
6842 _finger = addr;
6843 }
6844
6845 // Should revisit to see if this should be restructured for
6846 // greater efficiency.
6847 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6848 // convert offset into a HeapWord*
6849 HeapWord* addr = _verification_bm->startWord() + offset;
6850 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6851 "address out of range");
6852 assert(_verification_bm->isMarked(addr), "tautology");
6853 assert(_cms_bm->isMarked(addr), "tautology");
6854
6855 assert(_mark_stack->isEmpty(),
6856 "should drain stack to limit stack usage");
6857 // convert addr to an oop preparatory to scanning
6858 oop obj = oop(addr);
6859 assert(obj->is_oop(), "should be an oop");
6860 assert(_finger <= addr, "_finger runneth ahead");
6861 // advance the finger to right end of this object
6862 _finger = addr + obj->size();
6863 assert(_finger > addr, "we just incremented it above");
6864 // Note: the finger doesn't advance while we drain
6865 // the stack below.
6866 bool res = _mark_stack->push(obj);
6867 assert(res, "Empty non-zero size stack should have space for single push");
6868 while (!_mark_stack->isEmpty()) {
6869 oop new_oop = _mark_stack->pop();
6870 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6871 // now scan this oop's oops
6872 new_oop->oop_iterate(&_pam_verify_closure);
6873 }
6874 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6875 return true;
6876 }
6877
6878 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6879 CMSCollector* collector, MemRegion span,
6880 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6881 CMSMarkStack* mark_stack):
6882 MetadataAwareOopClosure(collector->ref_processor()),
6883 _collector(collector),
6884 _span(span),
6885 _verification_bm(verification_bm),
6886 _cms_bm(cms_bm),
6887 _mark_stack(mark_stack)
6888 { }
6889
6890 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6891 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6892
6893 // Upon stack overflow, we discard (part of) the stack,
6894 // remembering the least address amongst those discarded
6895 // in CMSCollector's _restart_address.
6896 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6897 // Remember the least grey address discarded
6898 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6899 _collector->lower_restart_addr(ra);
6900 _mark_stack->reset(); // discard stack contents
6901 _mark_stack->expand(); // expand the stack if possible
6902 }
6903
6904 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6905 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6906 HeapWord* addr = (HeapWord*)obj;
6907 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6908 // Oop lies in _span and isn't yet grey or black
6909 _verification_bm->mark(addr); // now grey
6910 if (!_cms_bm->isMarked(addr)) {
6911 oop(addr)->print();
6912 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6913 p2i(addr));
6914 fatal("... aborting");
6915 }
6916
6917 if (!_mark_stack->push(obj)) { // stack overflow
6918 if (PrintCMSStatistics != 0) {
6919 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6920 SIZE_FORMAT, _mark_stack->capacity());
6921 }
6922 assert(_mark_stack->isFull(), "Else push should have succeeded");
6923 handle_stack_overflow(addr);
6924 }
6925 // anything including and to the right of _finger
6926 // will be scanned as we iterate over the remainder of the
6927 // bit map
6928 }
6929 }
6930
6931 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6932 MemRegion span,
6933 CMSBitMap* bitMap, CMSMarkStack* markStack,
6934 HeapWord* finger, MarkFromRootsClosure* parent) :
6935 MetadataAwareOopClosure(collector->ref_processor()),
6936 _collector(collector),
6937 _span(span),
6938 _bitMap(bitMap),
6939 _markStack(markStack),
6940 _finger(finger),
6941 _parent(parent)
6942 { }
6943
6944 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6945 MemRegion span,
6946 CMSBitMap* bit_map,
6947 OopTaskQueue* work_queue,
6948 CMSMarkStack* overflow_stack,
6949 HeapWord* finger,
6950 HeapWord** global_finger_addr,
6951 Par_MarkFromRootsClosure* parent) :
6952 MetadataAwareOopClosure(collector->ref_processor()),
6953 _collector(collector),
6954 _whole_span(collector->_span),
6955 _span(span),
6956 _bit_map(bit_map),
6957 _work_queue(work_queue),
6958 _overflow_stack(overflow_stack),
6959 _finger(finger),
6960 _global_finger_addr(global_finger_addr),
6961 _parent(parent)
6962 { }
6963
6964 // Assumes thread-safe access by callers, who are
6965 // responsible for mutual exclusion.
6966 void CMSCollector::lower_restart_addr(HeapWord* low) {
6967 assert(_span.contains(low), "Out of bounds addr");
6968 if (_restart_addr == NULL) {
6969 _restart_addr = low;
6970 } else {
6971 _restart_addr = MIN2(_restart_addr, low);
6972 }
6973 }
6974
6975 // Upon stack overflow, we discard (part of) the stack,
6976 // remembering the least address amongst those discarded
6977 // in CMSCollector's _restart_address.
6978 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6979 // Remember the least grey address discarded
6980 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6981 _collector->lower_restart_addr(ra);
6982 _markStack->reset(); // discard stack contents
6983 _markStack->expand(); // expand the stack if possible
6984 }
6985
6986 // Upon stack overflow, we discard (part of) the stack,
6987 // remembering the least address amongst those discarded
6988 // in CMSCollector's _restart_address.
6989 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6990 // We need to do this under a mutex to prevent other
6991 // workers from interfering with the work done below.
6992 MutexLockerEx ml(_overflow_stack->par_lock(),
6993 Mutex::_no_safepoint_check_flag);
6994 // Remember the least grey address discarded
6995 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6996 _collector->lower_restart_addr(ra);
6997 _overflow_stack->reset(); // discard stack contents
6998 _overflow_stack->expand(); // expand the stack if possible
6999 }
7000
7001 void PushOrMarkClosure::do_oop(oop obj) {
7002 // Ignore mark word because we are running concurrent with mutators.
7003 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7004 HeapWord* addr = (HeapWord*)obj;
7005 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7006 // Oop lies in _span and isn't yet grey or black
7007 _bitMap->mark(addr); // now grey
7008 if (addr < _finger) {
7009 // the bit map iteration has already either passed, or
7010 // sampled, this bit in the bit map; we'll need to
7011 // use the marking stack to scan this oop's oops.
7012 bool simulate_overflow = false;
7013 NOT_PRODUCT(
7014 if (CMSMarkStackOverflowALot &&
7015 _collector->simulate_overflow()) {
7016 // simulate a stack overflow
7017 simulate_overflow = true;
7018 }
7019 )
7020 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7021 if (PrintCMSStatistics != 0) {
7022 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7023 SIZE_FORMAT, _markStack->capacity());
7024 }
7025 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7026 handle_stack_overflow(addr);
7027 }
7028 }
7029 // anything including and to the right of _finger
7030 // will be scanned as we iterate over the remainder of the
7031 // bit map
7032 do_yield_check();
7033 }
7034 }
7035
7036 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7037 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7038
7039 void Par_PushOrMarkClosure::do_oop(oop obj) {
7040 // Ignore mark word because we are running concurrent with mutators.
7041 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7042 HeapWord* addr = (HeapWord*)obj;
7043 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7044 // Oop lies in _span and isn't yet grey or black
7045 // We read the global_finger (volatile read) strictly after marking oop
7046 bool res = _bit_map->par_mark(addr); // now grey
7047 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7048 // Should we push this marked oop on our stack?
7049 // -- if someone else marked it, nothing to do
7050 // -- if target oop is above global finger nothing to do
7051 // -- if target oop is in chunk and above local finger
7052 // then nothing to do
7053 // -- else push on work queue
7054 if ( !res // someone else marked it, they will deal with it
7055 || (addr >= *gfa) // will be scanned in a later task
7056 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7057 return;
7058 }
7059 // the bit map iteration has already either passed, or
7060 // sampled, this bit in the bit map; we'll need to
7061 // use the marking stack to scan this oop's oops.
7062 bool simulate_overflow = false;
7063 NOT_PRODUCT(
7064 if (CMSMarkStackOverflowALot &&
7065 _collector->simulate_overflow()) {
7066 // simulate a stack overflow
7067 simulate_overflow = true;
7068 }
7069 )
7070 if (simulate_overflow ||
7071 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7072 // stack overflow
7073 if (PrintCMSStatistics != 0) {
7074 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7075 SIZE_FORMAT, _overflow_stack->capacity());
7076 }
7077 // We cannot assert that the overflow stack is full because
7078 // it may have been emptied since.
7079 assert(simulate_overflow ||
7080 _work_queue->size() == _work_queue->max_elems(),
7081 "Else push should have succeeded");
7082 handle_stack_overflow(addr);
7083 }
7084 do_yield_check();
7085 }
7086 }
7087
7088 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7089 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7090
7091 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7092 MemRegion span,
7093 ReferenceProcessor* rp,
7094 CMSBitMap* bit_map,
7095 CMSBitMap* mod_union_table,
7096 CMSMarkStack* mark_stack,
7097 bool concurrent_precleaning):
7098 MetadataAwareOopClosure(rp),
7099 _collector(collector),
7100 _span(span),
7101 _bit_map(bit_map),
7102 _mod_union_table(mod_union_table),
7103 _mark_stack(mark_stack),
7104 _concurrent_precleaning(concurrent_precleaning)
7105 {
7106 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7107 }
7108
7109 // Grey object rescan during pre-cleaning and second checkpoint phases --
7110 // the non-parallel version (the parallel version appears further below.)
7111 void PushAndMarkClosure::do_oop(oop obj) {
7112 // Ignore mark word verification. If during concurrent precleaning,
7113 // the object monitor may be locked. If during the checkpoint
7114 // phases, the object may already have been reached by a different
7115 // path and may be at the end of the global overflow list (so
7116 // the mark word may be NULL).
7117 assert(obj->is_oop_or_null(true /* ignore mark word */),
7118 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7119 HeapWord* addr = (HeapWord*)obj;
7120 // Check if oop points into the CMS generation
7121 // and is not marked
7122 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7123 // a white object ...
7124 _bit_map->mark(addr); // ... now grey
7125 // push on the marking stack (grey set)
7126 bool simulate_overflow = false;
7127 NOT_PRODUCT(
7128 if (CMSMarkStackOverflowALot &&
7129 _collector->simulate_overflow()) {
7130 // simulate a stack overflow
7131 simulate_overflow = true;
7132 }
7133 )
7134 if (simulate_overflow || !_mark_stack->push(obj)) {
7135 if (_concurrent_precleaning) {
7136 // During precleaning we can just dirty the appropriate card(s)
7137 // in the mod union table, thus ensuring that the object remains
7138 // in the grey set and continue. In the case of object arrays
7139 // we need to dirty all of the cards that the object spans,
7140 // since the rescan of object arrays will be limited to the
7141 // dirty cards.
7142 // Note that no one can be interfering with us in this action
7143 // of dirtying the mod union table, so no locking or atomics
7144 // are required.
7145 if (obj->is_objArray()) {
7146 size_t sz = obj->size();
7147 HeapWord* end_card_addr = (HeapWord*)round_to(
7148 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7149 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7150 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7151 _mod_union_table->mark_range(redirty_range);
7152 } else {
7153 _mod_union_table->mark(addr);
7154 }
7155 _collector->_ser_pmc_preclean_ovflw++;
7156 } else {
7157 // During the remark phase, we need to remember this oop
7158 // in the overflow list.
7159 _collector->push_on_overflow_list(obj);
7160 _collector->_ser_pmc_remark_ovflw++;
7161 }
7162 }
7163 }
7164 }
7165
7166 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7167 MemRegion span,
7168 ReferenceProcessor* rp,
7169 CMSBitMap* bit_map,
7170 OopTaskQueue* work_queue):
7171 MetadataAwareOopClosure(rp),
7172 _collector(collector),
7173 _span(span),
7174 _bit_map(bit_map),
7175 _work_queue(work_queue)
7176 {
7177 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7178 }
7179
7180 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7181 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7182
7183 // Grey object rescan during second checkpoint phase --
7184 // the parallel version.
7185 void Par_PushAndMarkClosure::do_oop(oop obj) {
7186 // In the assert below, we ignore the mark word because
7187 // this oop may point to an already visited object that is
7188 // on the overflow stack (in which case the mark word has
7189 // been hijacked for chaining into the overflow stack --
7190 // if this is the last object in the overflow stack then
7191 // its mark word will be NULL). Because this object may
7192 // have been subsequently popped off the global overflow
7193 // stack, and the mark word possibly restored to the prototypical
7194 // value, by the time we get to examined this failing assert in
7195 // the debugger, is_oop_or_null(false) may subsequently start
7196 // to hold.
7197 assert(obj->is_oop_or_null(true),
7198 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7199 HeapWord* addr = (HeapWord*)obj;
7200 // Check if oop points into the CMS generation
7201 // and is not marked
7202 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7203 // a white object ...
7204 // If we manage to "claim" the object, by being the
7205 // first thread to mark it, then we push it on our
7206 // marking stack
7207 if (_bit_map->par_mark(addr)) { // ... now grey
7208 // push on work queue (grey set)
7209 bool simulate_overflow = false;
7210 NOT_PRODUCT(
7211 if (CMSMarkStackOverflowALot &&
7212 _collector->par_simulate_overflow()) {
7213 // simulate a stack overflow
7214 simulate_overflow = true;
7215 }
7216 )
7217 if (simulate_overflow || !_work_queue->push(obj)) {
7218 _collector->par_push_on_overflow_list(obj);
7219 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7220 }
7221 } // Else, some other thread got there first
7222 }
7223 }
7224
7225 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7226 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7227
7228 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7229 Mutex* bml = _collector->bitMapLock();
7230 assert_lock_strong(bml);
7231 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7232 "CMS thread should hold CMS token");
7233
7234 bml->unlock();
7235 ConcurrentMarkSweepThread::desynchronize(true);
7236
7237 _collector->stopTimer();
7238 if (PrintCMSStatistics != 0) {
7239 _collector->incrementYields();
7240 }
7241
7242 // See the comment in coordinator_yield()
7243 for (unsigned i = 0; i < CMSYieldSleepCount &&
7244 ConcurrentMarkSweepThread::should_yield() &&
7245 !CMSCollector::foregroundGCIsActive(); ++i) {
7246 os::sleep(Thread::current(), 1, false);
7247 }
7248
7249 ConcurrentMarkSweepThread::synchronize(true);
7250 bml->lock();
7251
7252 _collector->startTimer();
7253 }
7254
7255 bool CMSPrecleanRefsYieldClosure::should_return() {
7256 if (ConcurrentMarkSweepThread::should_yield()) {
7257 do_yield_work();
7258 }
7259 return _collector->foregroundGCIsActive();
7260 }
7261
7262 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7263 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7264 "mr should be aligned to start at a card boundary");
7265 // We'd like to assert:
7266 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7267 // "mr should be a range of cards");
7268 // However, that would be too strong in one case -- the last
7269 // partition ends at _unallocated_block which, in general, can be
7270 // an arbitrary boundary, not necessarily card aligned.
7271 if (PrintCMSStatistics != 0) {
7272 _num_dirty_cards +=
7273 mr.word_size()/CardTableModRefBS::card_size_in_words;
7274 }
7275 _space->object_iterate_mem(mr, &_scan_cl);
7276 }
7277
7278 SweepClosure::SweepClosure(CMSCollector* collector,
7279 ConcurrentMarkSweepGeneration* g,
7280 CMSBitMap* bitMap, bool should_yield) :
7281 _collector(collector),
7282 _g(g),
7283 _sp(g->cmsSpace()),
7284 _limit(_sp->sweep_limit()),
7285 _freelistLock(_sp->freelistLock()),
7286 _bitMap(bitMap),
7287 _yield(should_yield),
7288 _inFreeRange(false), // No free range at beginning of sweep
7289 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7290 _lastFreeRangeCoalesced(false),
7291 _freeFinger(g->used_region().start())
7292 {
7293 NOT_PRODUCT(
7294 _numObjectsFreed = 0;
7295 _numWordsFreed = 0;
7296 _numObjectsLive = 0;
7297 _numWordsLive = 0;
7298 _numObjectsAlreadyFree = 0;
7299 _numWordsAlreadyFree = 0;
7300 _last_fc = NULL;
7301
7302 _sp->initializeIndexedFreeListArrayReturnedBytes();
7303 _sp->dictionary()->initialize_dict_returned_bytes();
7304 )
7305 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7306 "sweep _limit out of bounds");
7307 if (CMSTraceSweeper) {
7308 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7309 p2i(_limit));
7310 }
7311 }
7312
7313 void SweepClosure::print_on(outputStream* st) const {
7314 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7315 p2i(_sp->bottom()), p2i(_sp->end()));
7316 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7317 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7318 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7319 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7320 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7321 }
7322
7323 #ifndef PRODUCT
7324 // Assertion checking only: no useful work in product mode --
7325 // however, if any of the flags below become product flags,
7326 // you may need to review this code to see if it needs to be
7327 // enabled in product mode.
7328 SweepClosure::~SweepClosure() {
7329 assert_lock_strong(_freelistLock);
7330 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7331 "sweep _limit out of bounds");
7332 if (inFreeRange()) {
7333 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7334 print();
7335 ShouldNotReachHere();
7336 }
7337 if (Verbose && PrintGC) {
7338 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7339 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7340 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7341 SIZE_FORMAT" bytes "
7342 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7343 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7344 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7345 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7346 * sizeof(HeapWord);
7347 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7348
7349 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7350 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7351 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7352 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7353 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
7354 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7355 indexListReturnedBytes);
7356 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7357 dict_returned_bytes);
7358 }
7359 }
7360 if (CMSTraceSweeper) {
7361 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7362 p2i(_limit));
7363 }
7364 }
7365 #endif // PRODUCT
7366
7367 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7368 bool freeRangeInFreeLists) {
7369 if (CMSTraceSweeper) {
7370 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7371 p2i(freeFinger), freeRangeInFreeLists);
7372 }
7373 assert(!inFreeRange(), "Trampling existing free range");
7374 set_inFreeRange(true);
7375 set_lastFreeRangeCoalesced(false);
7376
7377 set_freeFinger(freeFinger);
7378 set_freeRangeInFreeLists(freeRangeInFreeLists);
7379 if (CMSTestInFreeList) {
7380 if (freeRangeInFreeLists) {
7381 FreeChunk* fc = (FreeChunk*) freeFinger;
7382 assert(fc->is_free(), "A chunk on the free list should be free.");
7383 assert(fc->size() > 0, "Free range should have a size");
7384 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7385 }
7386 }
7387 }
7388
7389 // Note that the sweeper runs concurrently with mutators. Thus,
7390 // it is possible for direct allocation in this generation to happen
7391 // in the middle of the sweep. Note that the sweeper also coalesces
7392 // contiguous free blocks. Thus, unless the sweeper and the allocator
7393 // synchronize appropriately freshly allocated blocks may get swept up.
7394 // This is accomplished by the sweeper locking the free lists while
7395 // it is sweeping. Thus blocks that are determined to be free are
7396 // indeed free. There is however one additional complication:
7397 // blocks that have been allocated since the final checkpoint and
7398 // mark, will not have been marked and so would be treated as
7399 // unreachable and swept up. To prevent this, the allocator marks
7400 // the bit map when allocating during the sweep phase. This leads,
7401 // however, to a further complication -- objects may have been allocated
7402 // but not yet initialized -- in the sense that the header isn't yet
7403 // installed. The sweeper can not then determine the size of the block
7404 // in order to skip over it. To deal with this case, we use a technique
7405 // (due to Printezis) to encode such uninitialized block sizes in the
7406 // bit map. Since the bit map uses a bit per every HeapWord, but the
7407 // CMS generation has a minimum object size of 3 HeapWords, it follows
7408 // that "normal marks" won't be adjacent in the bit map (there will
7409 // always be at least two 0 bits between successive 1 bits). We make use
7410 // of these "unused" bits to represent uninitialized blocks -- the bit
7411 // corresponding to the start of the uninitialized object and the next
7412 // bit are both set. Finally, a 1 bit marks the end of the object that
7413 // started with the two consecutive 1 bits to indicate its potentially
7414 // uninitialized state.
7415
7416 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7417 FreeChunk* fc = (FreeChunk*)addr;
7418 size_t res;
7419
7420 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7421 // than "addr == _limit" because although _limit was a block boundary when
7422 // we started the sweep, it may no longer be one because heap expansion
7423 // may have caused us to coalesce the block ending at the address _limit
7424 // with a newly expanded chunk (this happens when _limit was set to the
7425 // previous _end of the space), so we may have stepped past _limit:
7426 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7427 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7428 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7429 "sweep _limit out of bounds");
7430 assert(addr < _sp->end(), "addr out of bounds");
7431 // Flush any free range we might be holding as a single
7432 // coalesced chunk to the appropriate free list.
7433 if (inFreeRange()) {
7434 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7435 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger())));
7436 flush_cur_free_chunk(freeFinger(),
7437 pointer_delta(addr, freeFinger()));
7438 if (CMSTraceSweeper) {
7439 gclog_or_tty->print("Sweep: last chunk: ");
7440 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") "
7441 "[coalesced:%d]\n",
7442 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7443 lastFreeRangeCoalesced() ? 1 : 0);
7444 }
7445 }
7446
7447 // help the iterator loop finish
7448 return pointer_delta(_sp->end(), addr);
7449 }
7450
7451 assert(addr < _limit, "sweep invariant");
7452 // check if we should yield
7453 do_yield_check(addr);
7454 if (fc->is_free()) {
7455 // Chunk that is already free
7456 res = fc->size();
7457 do_already_free_chunk(fc);
7458 debug_only(_sp->verifyFreeLists());
7459 // If we flush the chunk at hand in lookahead_and_flush()
7460 // and it's coalesced with a preceding chunk, then the
7461 // process of "mangling" the payload of the coalesced block
7462 // will cause erasure of the size information from the
7463 // (erstwhile) header of all the coalesced blocks but the
7464 // first, so the first disjunct in the assert will not hold
7465 // in that specific case (in which case the second disjunct
7466 // will hold).
7467 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7468 "Otherwise the size info doesn't change at this step");
7469 NOT_PRODUCT(
7470 _numObjectsAlreadyFree++;
7471 _numWordsAlreadyFree += res;
7472 )
7473 NOT_PRODUCT(_last_fc = fc;)
7474 } else if (!_bitMap->isMarked(addr)) {
7475 // Chunk is fresh garbage
7476 res = do_garbage_chunk(fc);
7477 debug_only(_sp->verifyFreeLists());
7478 NOT_PRODUCT(
7479 _numObjectsFreed++;
7480 _numWordsFreed += res;
7481 )
7482 } else {
7483 // Chunk that is alive.
7484 res = do_live_chunk(fc);
7485 debug_only(_sp->verifyFreeLists());
7486 NOT_PRODUCT(
7487 _numObjectsLive++;
7488 _numWordsLive += res;
7489 )
7490 }
7491 return res;
7492 }
7493
7494 // For the smart allocation, record following
7495 // split deaths - a free chunk is removed from its free list because
7496 // it is being split into two or more chunks.
7497 // split birth - a free chunk is being added to its free list because
7498 // a larger free chunk has been split and resulted in this free chunk.
7499 // coal death - a free chunk is being removed from its free list because
7500 // it is being coalesced into a large free chunk.
7501 // coal birth - a free chunk is being added to its free list because
7502 // it was created when two or more free chunks where coalesced into
7503 // this free chunk.
7504 //
7505 // These statistics are used to determine the desired number of free
7506 // chunks of a given size. The desired number is chosen to be relative
7507 // to the end of a CMS sweep. The desired number at the end of a sweep
7508 // is the
7509 // count-at-end-of-previous-sweep (an amount that was enough)
7510 // - count-at-beginning-of-current-sweep (the excess)
7511 // + split-births (gains in this size during interval)
7512 // - split-deaths (demands on this size during interval)
7513 // where the interval is from the end of one sweep to the end of the
7514 // next.
7515 //
7516 // When sweeping the sweeper maintains an accumulated chunk which is
7517 // the chunk that is made up of chunks that have been coalesced. That
7518 // will be termed the left-hand chunk. A new chunk of garbage that
7519 // is being considered for coalescing will be referred to as the
7520 // right-hand chunk.
7521 //
7522 // When making a decision on whether to coalesce a right-hand chunk with
7523 // the current left-hand chunk, the current count vs. the desired count
7524 // of the left-hand chunk is considered. Also if the right-hand chunk
7525 // is near the large chunk at the end of the heap (see
7526 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7527 // left-hand chunk is coalesced.
7528 //
7529 // When making a decision about whether to split a chunk, the desired count
7530 // vs. the current count of the candidate to be split is also considered.
7531 // If the candidate is underpopulated (currently fewer chunks than desired)
7532 // a chunk of an overpopulated (currently more chunks than desired) size may
7533 // be chosen. The "hint" associated with a free list, if non-null, points
7534 // to a free list which may be overpopulated.
7535 //
7536
7537 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7538 const size_t size = fc->size();
7539 // Chunks that cannot be coalesced are not in the
7540 // free lists.
7541 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7542 assert(_sp->verify_chunk_in_free_list(fc),
7543 "free chunk should be in free lists");
7544 }
7545 // a chunk that is already free, should not have been
7546 // marked in the bit map
7547 HeapWord* const addr = (HeapWord*) fc;
7548 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7549 // Verify that the bit map has no bits marked between
7550 // addr and purported end of this block.
7551 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7552
7553 // Some chunks cannot be coalesced under any circumstances.
7554 // See the definition of cantCoalesce().
7555 if (!fc->cantCoalesce()) {
7556 // This chunk can potentially be coalesced.
7557 if (_sp->adaptive_freelists()) {
7558 // All the work is done in
7559 do_post_free_or_garbage_chunk(fc, size);
7560 } else { // Not adaptive free lists
7561 // this is a free chunk that can potentially be coalesced by the sweeper;
7562 if (!inFreeRange()) {
7563 // if the next chunk is a free block that can't be coalesced
7564 // it doesn't make sense to remove this chunk from the free lists
7565 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7566 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7567 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
7568 nextChunk->is_free() && // ... which is free...
7569 nextChunk->cantCoalesce()) { // ... but can't be coalesced
7570 // nothing to do
7571 } else {
7572 // Potentially the start of a new free range:
7573 // Don't eagerly remove it from the free lists.
7574 // No need to remove it if it will just be put
7575 // back again. (Also from a pragmatic point of view
7576 // if it is a free block in a region that is beyond
7577 // any allocated blocks, an assertion will fail)
7578 // Remember the start of a free run.
7579 initialize_free_range(addr, true);
7580 // end - can coalesce with next chunk
7581 }
7582 } else {
7583 // the midst of a free range, we are coalescing
7584 print_free_block_coalesced(fc);
7585 if (CMSTraceSweeper) {
7586 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7587 }
7588 // remove it from the free lists
7589 _sp->removeFreeChunkFromFreeLists(fc);
7590 set_lastFreeRangeCoalesced(true);
7591 // If the chunk is being coalesced and the current free range is
7592 // in the free lists, remove the current free range so that it
7593 // will be returned to the free lists in its entirety - all
7594 // the coalesced pieces included.
7595 if (freeRangeInFreeLists()) {
7596 FreeChunk* ffc = (FreeChunk*) freeFinger();
7597 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7598 "Size of free range is inconsistent with chunk size.");
7599 if (CMSTestInFreeList) {
7600 assert(_sp->verify_chunk_in_free_list(ffc),
7601 "free range is not in free lists");
7602 }
7603 _sp->removeFreeChunkFromFreeLists(ffc);
7604 set_freeRangeInFreeLists(false);
7605 }
7606 }
7607 }
7608 // Note that if the chunk is not coalescable (the else arm
7609 // below), we unconditionally flush, without needing to do
7610 // a "lookahead," as we do below.
7611 if (inFreeRange()) lookahead_and_flush(fc, size);
7612 } else {
7613 // Code path common to both original and adaptive free lists.
7614
7615 // cant coalesce with previous block; this should be treated
7616 // as the end of a free run if any
7617 if (inFreeRange()) {
7618 // we kicked some butt; time to pick up the garbage
7619 assert(freeFinger() < addr, "freeFinger points too high");
7620 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7621 }
7622 // else, nothing to do, just continue
7623 }
7624 }
7625
7626 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7627 // This is a chunk of garbage. It is not in any free list.
7628 // Add it to a free list or let it possibly be coalesced into
7629 // a larger chunk.
7630 HeapWord* const addr = (HeapWord*) fc;
7631 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7632
7633 if (_sp->adaptive_freelists()) {
7634 // Verify that the bit map has no bits marked between
7635 // addr and purported end of just dead object.
7636 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7637
7638 do_post_free_or_garbage_chunk(fc, size);
7639 } else {
7640 if (!inFreeRange()) {
7641 // start of a new free range
7642 assert(size > 0, "A free range should have a size");
7643 initialize_free_range(addr, false);
7644 } else {
7645 // this will be swept up when we hit the end of the
7646 // free range
7647 if (CMSTraceSweeper) {
7648 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7649 }
7650 // If the chunk is being coalesced and the current free range is
7651 // in the free lists, remove the current free range so that it
7652 // will be returned to the free lists in its entirety - all
7653 // the coalesced pieces included.
7654 if (freeRangeInFreeLists()) {
7655 FreeChunk* ffc = (FreeChunk*)freeFinger();
7656 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7657 "Size of free range is inconsistent with chunk size.");
7658 if (CMSTestInFreeList) {
7659 assert(_sp->verify_chunk_in_free_list(ffc),
7660 "free range is not in free lists");
7661 }
7662 _sp->removeFreeChunkFromFreeLists(ffc);
7663 set_freeRangeInFreeLists(false);
7664 }
7665 set_lastFreeRangeCoalesced(true);
7666 }
7667 // this will be swept up when we hit the end of the free range
7668
7669 // Verify that the bit map has no bits marked between
7670 // addr and purported end of just dead object.
7671 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7672 }
7673 assert(_limit >= addr + size,
7674 "A freshly garbage chunk can't possibly straddle over _limit");
7675 if (inFreeRange()) lookahead_and_flush(fc, size);
7676 return size;
7677 }
7678
7679 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7680 HeapWord* addr = (HeapWord*) fc;
7681 // The sweeper has just found a live object. Return any accumulated
7682 // left hand chunk to the free lists.
7683 if (inFreeRange()) {
7684 assert(freeFinger() < addr, "freeFinger points too high");
7685 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7686 }
7687
7688 // This object is live: we'd normally expect this to be
7689 // an oop, and like to assert the following:
7690 // assert(oop(addr)->is_oop(), "live block should be an oop");
7691 // However, as we commented above, this may be an object whose
7692 // header hasn't yet been initialized.
7693 size_t size;
7694 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7695 if (_bitMap->isMarked(addr + 1)) {
7696 // Determine the size from the bit map, rather than trying to
7697 // compute it from the object header.
7698 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7699 size = pointer_delta(nextOneAddr + 1, addr);
7700 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7701 "alignment problem");
7702
7703 #ifdef ASSERT
7704 if (oop(addr)->klass_or_null() != NULL) {
7705 // Ignore mark word because we are running concurrent with mutators
7706 assert(oop(addr)->is_oop(true), "live block should be an oop");
7707 assert(size ==
7708 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7709 "P-mark and computed size do not agree");
7710 }
7711 #endif
7712
7713 } else {
7714 // This should be an initialized object that's alive.
7715 assert(oop(addr)->klass_or_null() != NULL,
7716 "Should be an initialized object");
7717 // Ignore mark word because we are running concurrent with mutators
7718 assert(oop(addr)->is_oop(true), "live block should be an oop");
7719 // Verify that the bit map has no bits marked between
7720 // addr and purported end of this block.
7721 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7722 assert(size >= 3, "Necessary for Printezis marks to work");
7723 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7724 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7725 }
7726 return size;
7727 }
7728
7729 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7730 size_t chunkSize) {
7731 // do_post_free_or_garbage_chunk() should only be called in the case
7732 // of the adaptive free list allocator.
7733 const bool fcInFreeLists = fc->is_free();
7734 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7735 assert((HeapWord*)fc <= _limit, "sweep invariant");
7736 if (CMSTestInFreeList && fcInFreeLists) {
7737 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7738 }
7739
7740 if (CMSTraceSweeper) {
7741 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7742 }
7743
7744 HeapWord* const fc_addr = (HeapWord*) fc;
7745
7746 bool coalesce;
7747 const size_t left = pointer_delta(fc_addr, freeFinger());
7748 const size_t right = chunkSize;
7749 switch (FLSCoalescePolicy) {
7750 // numeric value forms a coalition aggressiveness metric
7751 case 0: { // never coalesce
7752 coalesce = false;
7753 break;
7754 }
7755 case 1: { // coalesce if left & right chunks on overpopulated lists
7756 coalesce = _sp->coalOverPopulated(left) &&
7757 _sp->coalOverPopulated(right);
7758 break;
7759 }
7760 case 2: { // coalesce if left chunk on overpopulated list (default)
7761 coalesce = _sp->coalOverPopulated(left);
7762 break;
7763 }
7764 case 3: { // coalesce if left OR right chunk on overpopulated list
7765 coalesce = _sp->coalOverPopulated(left) ||
7766 _sp->coalOverPopulated(right);
7767 break;
7768 }
7769 case 4: { // always coalesce
7770 coalesce = true;
7771 break;
7772 }
7773 default:
7774 ShouldNotReachHere();
7775 }
7776
7777 // Should the current free range be coalesced?
7778 // If the chunk is in a free range and either we decided to coalesce above
7779 // or the chunk is near the large block at the end of the heap
7780 // (isNearLargestChunk() returns true), then coalesce this chunk.
7781 const bool doCoalesce = inFreeRange()
7782 && (coalesce || _g->isNearLargestChunk(fc_addr));
7783 if (doCoalesce) {
7784 // Coalesce the current free range on the left with the new
7785 // chunk on the right. If either is on a free list,
7786 // it must be removed from the list and stashed in the closure.
7787 if (freeRangeInFreeLists()) {
7788 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7789 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7790 "Size of free range is inconsistent with chunk size.");
7791 if (CMSTestInFreeList) {
7792 assert(_sp->verify_chunk_in_free_list(ffc),
7793 "Chunk is not in free lists");
7794 }
7795 _sp->coalDeath(ffc->size());
7796 _sp->removeFreeChunkFromFreeLists(ffc);
7797 set_freeRangeInFreeLists(false);
7798 }
7799 if (fcInFreeLists) {
7800 _sp->coalDeath(chunkSize);
7801 assert(fc->size() == chunkSize,
7802 "The chunk has the wrong size or is not in the free lists");
7803 _sp->removeFreeChunkFromFreeLists(fc);
7804 }
7805 set_lastFreeRangeCoalesced(true);
7806 print_free_block_coalesced(fc);
7807 } else { // not in a free range and/or should not coalesce
7808 // Return the current free range and start a new one.
7809 if (inFreeRange()) {
7810 // In a free range but cannot coalesce with the right hand chunk.
7811 // Put the current free range into the free lists.
7812 flush_cur_free_chunk(freeFinger(),
7813 pointer_delta(fc_addr, freeFinger()));
7814 }
7815 // Set up for new free range. Pass along whether the right hand
7816 // chunk is in the free lists.
7817 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7818 }
7819 }
7820
7821 // Lookahead flush:
7822 // If we are tracking a free range, and this is the last chunk that
7823 // we'll look at because its end crosses past _limit, we'll preemptively
7824 // flush it along with any free range we may be holding on to. Note that
7825 // this can be the case only for an already free or freshly garbage
7826 // chunk. If this block is an object, it can never straddle
7827 // over _limit. The "straddling" occurs when _limit is set at
7828 // the previous end of the space when this cycle started, and
7829 // a subsequent heap expansion caused the previously co-terminal
7830 // free block to be coalesced with the newly expanded portion,
7831 // thus rendering _limit a non-block-boundary making it dangerous
7832 // for the sweeper to step over and examine.
7833 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7834 assert(inFreeRange(), "Should only be called if currently in a free range.");
7835 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7836 assert(_sp->used_region().contains(eob - 1),
7837 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7838 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7839 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7840 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size));
7841 if (eob >= _limit) {
7842 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7843 if (CMSTraceSweeper) {
7844 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7845 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7846 "[" PTR_FORMAT "," PTR_FORMAT ")",
7847 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7848 }
7849 // Return the storage we are tracking back into the free lists.
7850 if (CMSTraceSweeper) {
7851 gclog_or_tty->print_cr("Flushing ... ");
7852 }
7853 assert(freeFinger() < eob, "Error");
7854 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7855 }
7856 }
7857
7858 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7859 assert(inFreeRange(), "Should only be called if currently in a free range.");
7860 assert(size > 0,
7861 "A zero sized chunk cannot be added to the free lists.");
7862 if (!freeRangeInFreeLists()) {
7863 if (CMSTestInFreeList) {
7864 FreeChunk* fc = (FreeChunk*) chunk;
7865 fc->set_size(size);
7866 assert(!_sp->verify_chunk_in_free_list(fc),
7867 "chunk should not be in free lists yet");
7868 }
7869 if (CMSTraceSweeper) {
7870 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7871 p2i(chunk), size);
7872 }
7873 // A new free range is going to be starting. The current
7874 // free range has not been added to the free lists yet or
7875 // was removed so add it back.
7876 // If the current free range was coalesced, then the death
7877 // of the free range was recorded. Record a birth now.
7878 if (lastFreeRangeCoalesced()) {
7879 _sp->coalBirth(size);
7880 }
7881 _sp->addChunkAndRepairOffsetTable(chunk, size,
7882 lastFreeRangeCoalesced());
7883 } else if (CMSTraceSweeper) {
7884 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7885 }
7886 set_inFreeRange(false);
7887 set_freeRangeInFreeLists(false);
7888 }
7889
7890 // We take a break if we've been at this for a while,
7891 // so as to avoid monopolizing the locks involved.
7892 void SweepClosure::do_yield_work(HeapWord* addr) {
7893 // Return current free chunk being used for coalescing (if any)
7894 // to the appropriate freelist. After yielding, the next
7895 // free block encountered will start a coalescing range of
7896 // free blocks. If the next free block is adjacent to the
7897 // chunk just flushed, they will need to wait for the next
7898 // sweep to be coalesced.
7899 if (inFreeRange()) {
7900 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7901 }
7902
7903 // First give up the locks, then yield, then re-lock.
7904 // We should probably use a constructor/destructor idiom to
7905 // do this unlock/lock or modify the MutexUnlocker class to
7906 // serve our purpose. XXX
7907 assert_lock_strong(_bitMap->lock());
7908 assert_lock_strong(_freelistLock);
7909 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7910 "CMS thread should hold CMS token");
7911 _bitMap->lock()->unlock();
7912 _freelistLock->unlock();
7913 ConcurrentMarkSweepThread::desynchronize(true);
7914 _collector->stopTimer();
7915 if (PrintCMSStatistics != 0) {
7916 _collector->incrementYields();
7917 }
7918
7919 // See the comment in coordinator_yield()
7920 for (unsigned i = 0; i < CMSYieldSleepCount &&
7921 ConcurrentMarkSweepThread::should_yield() &&
7922 !CMSCollector::foregroundGCIsActive(); ++i) {
7923 os::sleep(Thread::current(), 1, false);
7924 }
7925
7926 ConcurrentMarkSweepThread::synchronize(true);
7927 _freelistLock->lock();
7928 _bitMap->lock()->lock_without_safepoint_check();
7929 _collector->startTimer();
7930 }
7931
7932 #ifndef PRODUCT
7933 // This is actually very useful in a product build if it can
7934 // be called from the debugger. Compile it into the product
7935 // as needed.
7936 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7937 return debug_cms_space->verify_chunk_in_free_list(fc);
7938 }
7939 #endif
7940
7941 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7942 if (CMSTraceSweeper) {
7943 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7944 p2i(fc), fc->size());
7945 }
7946 }
7947
7948 // CMSIsAliveClosure
7949 bool CMSIsAliveClosure::do_object_b(oop obj) {
7950 HeapWord* addr = (HeapWord*)obj;
7951 return addr != NULL &&
7952 (!_span.contains(addr) || _bit_map->isMarked(addr));
7953 }
7954
7955
7956 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7957 MemRegion span,
7958 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7959 bool cpc):
7960 _collector(collector),
7961 _span(span),
7962 _bit_map(bit_map),
7963 _mark_stack(mark_stack),
7964 _concurrent_precleaning(cpc) {
7965 assert(!_span.is_empty(), "Empty span could spell trouble");
7966 }
7967
7968
7969 // CMSKeepAliveClosure: the serial version
7970 void CMSKeepAliveClosure::do_oop(oop obj) {
7971 HeapWord* addr = (HeapWord*)obj;
7972 if (_span.contains(addr) &&
7973 !_bit_map->isMarked(addr)) {
7974 _bit_map->mark(addr);
7975 bool simulate_overflow = false;
7976 NOT_PRODUCT(
7977 if (CMSMarkStackOverflowALot &&
7978 _collector->simulate_overflow()) {
7979 // simulate a stack overflow
7980 simulate_overflow = true;
7981 }
7982 )
7983 if (simulate_overflow || !_mark_stack->push(obj)) {
7984 if (_concurrent_precleaning) {
7985 // We dirty the overflown object and let the remark
7986 // phase deal with it.
7987 assert(_collector->overflow_list_is_empty(), "Error");
7988 // In the case of object arrays, we need to dirty all of
7989 // the cards that the object spans. No locking or atomics
7990 // are needed since no one else can be mutating the mod union
7991 // table.
7992 if (obj->is_objArray()) {
7993 size_t sz = obj->size();
7994 HeapWord* end_card_addr =
7995 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7996 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7997 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7998 _collector->_modUnionTable.mark_range(redirty_range);
7999 } else {
8000 _collector->_modUnionTable.mark(addr);
8001 }
8002 _collector->_ser_kac_preclean_ovflw++;
8003 } else {
8004 _collector->push_on_overflow_list(obj);
8005 _collector->_ser_kac_ovflw++;
8006 }
8007 }
8008 }
8009 }
8010
8011 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8012 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8013
8014 // CMSParKeepAliveClosure: a parallel version of the above.
8015 // The work queues are private to each closure (thread),
8016 // but (may be) available for stealing by other threads.
8017 void CMSParKeepAliveClosure::do_oop(oop obj) {
8018 HeapWord* addr = (HeapWord*)obj;
8019 if (_span.contains(addr) &&
8020 !_bit_map->isMarked(addr)) {
8021 // In general, during recursive tracing, several threads
8022 // may be concurrently getting here; the first one to
8023 // "tag" it, claims it.
8024 if (_bit_map->par_mark(addr)) {
8025 bool res = _work_queue->push(obj);
8026 assert(res, "Low water mark should be much less than capacity");
8027 // Do a recursive trim in the hope that this will keep
8028 // stack usage lower, but leave some oops for potential stealers
8029 trim_queue(_low_water_mark);
8030 } // Else, another thread got there first
8031 }
8032 }
8033
8034 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8035 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8036
8037 void CMSParKeepAliveClosure::trim_queue(uint max) {
8038 while (_work_queue->size() > max) {
8039 oop new_oop;
8040 if (_work_queue->pop_local(new_oop)) {
8041 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8042 assert(_bit_map->isMarked((HeapWord*)new_oop),
8043 "no white objects on this stack!");
8044 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8045 // iterate over the oops in this oop, marking and pushing
8046 // the ones in CMS heap (i.e. in _span).
8047 new_oop->oop_iterate(&_mark_and_push);
8048 }
8049 }
8050 }
8051
8052 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8053 CMSCollector* collector,
8054 MemRegion span, CMSBitMap* bit_map,
8055 OopTaskQueue* work_queue):
8056 _collector(collector),
8057 _span(span),
8058 _bit_map(bit_map),
8059 _work_queue(work_queue) { }
8060
8061 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8062 HeapWord* addr = (HeapWord*)obj;
8063 if (_span.contains(addr) &&
8064 !_bit_map->isMarked(addr)) {
8065 if (_bit_map->par_mark(addr)) {
8066 bool simulate_overflow = false;
8067 NOT_PRODUCT(
8068 if (CMSMarkStackOverflowALot &&
8069 _collector->par_simulate_overflow()) {
8070 // simulate a stack overflow
8071 simulate_overflow = true;
8072 }
8073 )
8074 if (simulate_overflow || !_work_queue->push(obj)) {
8075 _collector->par_push_on_overflow_list(obj);
8076 _collector->_par_kac_ovflw++;
8077 }
8078 } // Else another thread got there already
8079 }
8080 }
8081
8082 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8083 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8084
8085 //////////////////////////////////////////////////////////////////
8086 // CMSExpansionCause /////////////////////////////
8087 //////////////////////////////////////////////////////////////////
8088 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8089 switch (cause) {
8090 case _no_expansion:
8091 return "No expansion";
8092 case _satisfy_free_ratio:
8093 return "Free ratio";
8094 case _satisfy_promotion:
8095 return "Satisfy promotion";
8096 case _satisfy_allocation:
8097 return "allocation";
8098 case _allocate_par_lab:
8099 return "Par LAB";
8100 case _allocate_par_spooling_space:
8101 return "Par Spooling Space";
8102 case _adaptive_size_policy:
8103 return "Ergonomics";
8104 default:
8105 return "unknown";
8106 }
8107 }
8108
8109 void CMSDrainMarkingStackClosure::do_void() {
8110 // the max number to take from overflow list at a time
8111 const size_t num = _mark_stack->capacity()/4;
8112 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8113 "Overflow list should be NULL during concurrent phases");
8114 while (!_mark_stack->isEmpty() ||
8115 // if stack is empty, check the overflow list
8116 _collector->take_from_overflow_list(num, _mark_stack)) {
8117 oop obj = _mark_stack->pop();
8118 HeapWord* addr = (HeapWord*)obj;
8119 assert(_span.contains(addr), "Should be within span");
8120 assert(_bit_map->isMarked(addr), "Should be marked");
8121 assert(obj->is_oop(), "Should be an oop");
8122 obj->oop_iterate(_keep_alive);
8123 }
8124 }
8125
8126 void CMSParDrainMarkingStackClosure::do_void() {
8127 // drain queue
8128 trim_queue(0);
8129 }
8130
8131 // Trim our work_queue so its length is below max at return
8132 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8133 while (_work_queue->size() > max) {
8134 oop new_oop;
8135 if (_work_queue->pop_local(new_oop)) {
8136 assert(new_oop->is_oop(), "Expected an oop");
8137 assert(_bit_map->isMarked((HeapWord*)new_oop),
8138 "no white objects on this stack!");
8139 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8140 // iterate over the oops in this oop, marking and pushing
8141 // the ones in CMS heap (i.e. in _span).
8142 new_oop->oop_iterate(&_mark_and_push);
8143 }
8144 }
8145 }
8146
8147 ////////////////////////////////////////////////////////////////////
8148 // Support for Marking Stack Overflow list handling and related code
8149 ////////////////////////////////////////////////////////////////////
8150 // Much of the following code is similar in shape and spirit to the
8151 // code used in ParNewGC. We should try and share that code
8152 // as much as possible in the future.
8153
8154 #ifndef PRODUCT
8155 // Debugging support for CMSStackOverflowALot
8156
8157 // It's OK to call this multi-threaded; the worst thing
8158 // that can happen is that we'll get a bunch of closely
8159 // spaced simulated overflows, but that's OK, in fact
8160 // probably good as it would exercise the overflow code
8161 // under contention.
8162 bool CMSCollector::simulate_overflow() {
8163 if (_overflow_counter-- <= 0) { // just being defensive
8164 _overflow_counter = CMSMarkStackOverflowInterval;
8165 return true;
8166 } else {
8167 return false;
8168 }
8169 }
8170
8171 bool CMSCollector::par_simulate_overflow() {
8172 return simulate_overflow();
8173 }
8174 #endif
8175
8176 // Single-threaded
8177 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8178 assert(stack->isEmpty(), "Expected precondition");
8179 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8180 size_t i = num;
8181 oop cur = _overflow_list;
8182 const markOop proto = markOopDesc::prototype();
8183 NOT_PRODUCT(ssize_t n = 0;)
8184 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8185 next = oop(cur->mark());
8186 cur->set_mark(proto); // until proven otherwise
8187 assert(cur->is_oop(), "Should be an oop");
8188 bool res = stack->push(cur);
8189 assert(res, "Bit off more than can chew?");
8190 NOT_PRODUCT(n++;)
8191 }
8192 _overflow_list = cur;
8193 #ifndef PRODUCT
8194 assert(_num_par_pushes >= n, "Too many pops?");
8195 _num_par_pushes -=n;
8196 #endif
8197 return !stack->isEmpty();
8198 }
8199
8200 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8201 // (MT-safe) Get a prefix of at most "num" from the list.
8202 // The overflow list is chained through the mark word of
8203 // each object in the list. We fetch the entire list,
8204 // break off a prefix of the right size and return the
8205 // remainder. If other threads try to take objects from
8206 // the overflow list at that time, they will wait for
8207 // some time to see if data becomes available. If (and
8208 // only if) another thread places one or more object(s)
8209 // on the global list before we have returned the suffix
8210 // to the global list, we will walk down our local list
8211 // to find its end and append the global list to
8212 // our suffix before returning it. This suffix walk can
8213 // prove to be expensive (quadratic in the amount of traffic)
8214 // when there are many objects in the overflow list and
8215 // there is much producer-consumer contention on the list.
8216 // *NOTE*: The overflow list manipulation code here and
8217 // in ParNewGeneration:: are very similar in shape,
8218 // except that in the ParNew case we use the old (from/eden)
8219 // copy of the object to thread the list via its klass word.
8220 // Because of the common code, if you make any changes in
8221 // the code below, please check the ParNew version to see if
8222 // similar changes might be needed.
8223 // CR 6797058 has been filed to consolidate the common code.
8224 bool CMSCollector::par_take_from_overflow_list(size_t num,
8225 OopTaskQueue* work_q,
8226 int no_of_gc_threads) {
8227 assert(work_q->size() == 0, "First empty local work queue");
8228 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8229 if (_overflow_list == NULL) {
8230 return false;
8231 }
8232 // Grab the entire list; we'll put back a suffix
8233 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8234 Thread* tid = Thread::current();
8235 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8236 // set to ParallelGCThreads.
8237 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8238 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8239 // If the list is busy, we spin for a short while,
8240 // sleeping between attempts to get the list.
8241 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8242 os::sleep(tid, sleep_time_millis, false);
8243 if (_overflow_list == NULL) {
8244 // Nothing left to take
8245 return false;
8246 } else if (_overflow_list != BUSY) {
8247 // Try and grab the prefix
8248 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8249 }
8250 }
8251 // If the list was found to be empty, or we spun long
8252 // enough, we give up and return empty-handed. If we leave
8253 // the list in the BUSY state below, it must be the case that
8254 // some other thread holds the overflow list and will set it
8255 // to a non-BUSY state in the future.
8256 if (prefix == NULL || prefix == BUSY) {
8257 // Nothing to take or waited long enough
8258 if (prefix == NULL) {
8259 // Write back the NULL in case we overwrote it with BUSY above
8260 // and it is still the same value.
8261 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8262 }
8263 return false;
8264 }
8265 assert(prefix != NULL && prefix != BUSY, "Error");
8266 size_t i = num;
8267 oop cur = prefix;
8268 // Walk down the first "num" objects, unless we reach the end.
8269 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8270 if (cur->mark() == NULL) {
8271 // We have "num" or fewer elements in the list, so there
8272 // is nothing to return to the global list.
8273 // Write back the NULL in lieu of the BUSY we wrote
8274 // above, if it is still the same value.
8275 if (_overflow_list == BUSY) {
8276 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8277 }
8278 } else {
8279 // Chop off the suffix and return it to the global list.
8280 assert(cur->mark() != BUSY, "Error");
8281 oop suffix_head = cur->mark(); // suffix will be put back on global list
8282 cur->set_mark(NULL); // break off suffix
8283 // It's possible that the list is still in the empty(busy) state
8284 // we left it in a short while ago; in that case we may be
8285 // able to place back the suffix without incurring the cost
8286 // of a walk down the list.
8287 oop observed_overflow_list = _overflow_list;
8288 oop cur_overflow_list = observed_overflow_list;
8289 bool attached = false;
8290 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8291 observed_overflow_list =
8292 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8293 if (cur_overflow_list == observed_overflow_list) {
8294 attached = true;
8295 break;
8296 } else cur_overflow_list = observed_overflow_list;
8297 }
8298 if (!attached) {
8299 // Too bad, someone else sneaked in (at least) an element; we'll need
8300 // to do a splice. Find tail of suffix so we can prepend suffix to global
8301 // list.
8302 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8303 oop suffix_tail = cur;
8304 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8305 "Tautology");
8306 observed_overflow_list = _overflow_list;
8307 do {
8308 cur_overflow_list = observed_overflow_list;
8309 if (cur_overflow_list != BUSY) {
8310 // Do the splice ...
8311 suffix_tail->set_mark(markOop(cur_overflow_list));
8312 } else { // cur_overflow_list == BUSY
8313 suffix_tail->set_mark(NULL);
8314 }
8315 // ... and try to place spliced list back on overflow_list ...
8316 observed_overflow_list =
8317 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8318 } while (cur_overflow_list != observed_overflow_list);
8319 // ... until we have succeeded in doing so.
8320 }
8321 }
8322
8323 // Push the prefix elements on work_q
8324 assert(prefix != NULL, "control point invariant");
8325 const markOop proto = markOopDesc::prototype();
8326 oop next;
8327 NOT_PRODUCT(ssize_t n = 0;)
8328 for (cur = prefix; cur != NULL; cur = next) {
8329 next = oop(cur->mark());
8330 cur->set_mark(proto); // until proven otherwise
8331 assert(cur->is_oop(), "Should be an oop");
8332 bool res = work_q->push(cur);
8333 assert(res, "Bit off more than we can chew?");
8334 NOT_PRODUCT(n++;)
8335 }
8336 #ifndef PRODUCT
8337 assert(_num_par_pushes >= n, "Too many pops?");
8338 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8339 #endif
8340 return true;
8341 }
8342
8343 // Single-threaded
8344 void CMSCollector::push_on_overflow_list(oop p) {
8345 NOT_PRODUCT(_num_par_pushes++;)
8346 assert(p->is_oop(), "Not an oop");
8347 preserve_mark_if_necessary(p);
8348 p->set_mark((markOop)_overflow_list);
8349 _overflow_list = p;
8350 }
8351
8352 // Multi-threaded; use CAS to prepend to overflow list
8353 void CMSCollector::par_push_on_overflow_list(oop p) {
8354 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8355 assert(p->is_oop(), "Not an oop");
8356 par_preserve_mark_if_necessary(p);
8357 oop observed_overflow_list = _overflow_list;
8358 oop cur_overflow_list;
8359 do {
8360 cur_overflow_list = observed_overflow_list;
8361 if (cur_overflow_list != BUSY) {
8362 p->set_mark(markOop(cur_overflow_list));
8363 } else {
8364 p->set_mark(NULL);
8365 }
8366 observed_overflow_list =
8367 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8368 } while (cur_overflow_list != observed_overflow_list);
8369 }
8370 #undef BUSY
8371
8372 // Single threaded
8373 // General Note on GrowableArray: pushes may silently fail
8374 // because we are (temporarily) out of C-heap for expanding
8375 // the stack. The problem is quite ubiquitous and affects
8376 // a lot of code in the JVM. The prudent thing for GrowableArray
8377 // to do (for now) is to exit with an error. However, that may
8378 // be too draconian in some cases because the caller may be
8379 // able to recover without much harm. For such cases, we
8380 // should probably introduce a "soft_push" method which returns
8381 // an indication of success or failure with the assumption that
8382 // the caller may be able to recover from a failure; code in
8383 // the VM can then be changed, incrementally, to deal with such
8384 // failures where possible, thus, incrementally hardening the VM
8385 // in such low resource situations.
8386 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8387 _preserved_oop_stack.push(p);
8388 _preserved_mark_stack.push(m);
8389 assert(m == p->mark(), "Mark word changed");
8390 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8391 "bijection");
8392 }
8393
8394 // Single threaded
8395 void CMSCollector::preserve_mark_if_necessary(oop p) {
8396 markOop m = p->mark();
8397 if (m->must_be_preserved(p)) {
8398 preserve_mark_work(p, m);
8399 }
8400 }
8401
8402 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8403 markOop m = p->mark();
8404 if (m->must_be_preserved(p)) {
8405 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8406 // Even though we read the mark word without holding
8407 // the lock, we are assured that it will not change
8408 // because we "own" this oop, so no other thread can
8409 // be trying to push it on the overflow list; see
8410 // the assertion in preserve_mark_work() that checks
8411 // that m == p->mark().
8412 preserve_mark_work(p, m);
8413 }
8414 }
8415
8416 // We should be able to do this multi-threaded,
8417 // a chunk of stack being a task (this is
8418 // correct because each oop only ever appears
8419 // once in the overflow list. However, it's
8420 // not very easy to completely overlap this with
8421 // other operations, so will generally not be done
8422 // until all work's been completed. Because we
8423 // expect the preserved oop stack (set) to be small,
8424 // it's probably fine to do this single-threaded.
8425 // We can explore cleverer concurrent/overlapped/parallel
8426 // processing of preserved marks if we feel the
8427 // need for this in the future. Stack overflow should
8428 // be so rare in practice and, when it happens, its
8429 // effect on performance so great that this will
8430 // likely just be in the noise anyway.
8431 void CMSCollector::restore_preserved_marks_if_any() {
8432 assert(SafepointSynchronize::is_at_safepoint(),
8433 "world should be stopped");
8434 assert(Thread::current()->is_ConcurrentGC_thread() ||
8435 Thread::current()->is_VM_thread(),
8436 "should be single-threaded");
8437 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8438 "bijection");
8439
8440 while (!_preserved_oop_stack.is_empty()) {
8441 oop p = _preserved_oop_stack.pop();
8442 assert(p->is_oop(), "Should be an oop");
8443 assert(_span.contains(p), "oop should be in _span");
8444 assert(p->mark() == markOopDesc::prototype(),
8445 "Set when taken from overflow list");
8446 markOop m = _preserved_mark_stack.pop();
8447 p->set_mark(m);
8448 }
8449 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8450 "stacks were cleared above");
8451 }
8452
8453 #ifndef PRODUCT
8454 bool CMSCollector::no_preserved_marks() const {
8455 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8456 }
8457 #endif
8458
8459 // Transfer some number of overflown objects to usual marking
8460 // stack. Return true if some objects were transferred.
8461 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8462 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8463 (size_t)ParGCDesiredObjsFromOverflowList);
8464
8465 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8466 assert(_collector->overflow_list_is_empty() || res,
8467 "If list is not empty, we should have taken something");
8468 assert(!res || !_mark_stack->isEmpty(),
8469 "If we took something, it should now be on our stack");
8470 return res;
8471 }
8472
8473 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8474 size_t res = _sp->block_size_no_stall(addr, _collector);
8475 if (_sp->block_is_obj(addr)) {
8476 if (_live_bit_map->isMarked(addr)) {
8477 // It can't have been dead in a previous cycle
8478 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8479 } else {
8480 _dead_bit_map->mark(addr); // mark the dead object
8481 }
8482 }
8483 // Could be 0, if the block size could not be computed without stalling.
8484 return res;
8485 }
8486
8487 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8488
8489 switch (phase) {
8490 case CMSCollector::InitialMarking:
8491 initialize(true /* fullGC */ ,
8492 cause /* cause of the GC */,
8493 true /* recordGCBeginTime */,
8494 true /* recordPreGCUsage */,
8495 false /* recordPeakUsage */,
8496 false /* recordPostGCusage */,
8497 true /* recordAccumulatedGCTime */,
8498 false /* recordGCEndTime */,
8499 false /* countCollection */ );
8500 break;
8501
8502 case CMSCollector::FinalMarking:
8503 initialize(true /* fullGC */ ,
8504 cause /* cause of the GC */,
8505 false /* recordGCBeginTime */,
8506 false /* recordPreGCUsage */,
8507 false /* recordPeakUsage */,
8508 false /* recordPostGCusage */,
8509 true /* recordAccumulatedGCTime */,
8510 false /* recordGCEndTime */,
8511 false /* countCollection */ );
8512 break;
8513
8514 case CMSCollector::Sweeping:
8515 initialize(true /* fullGC */ ,
8516 cause /* cause of the GC */,
8517 false /* recordGCBeginTime */,
8518 false /* recordPreGCUsage */,
8519 true /* recordPeakUsage */,
8520 true /* recordPostGCusage */,
8521 false /* recordAccumulatedGCTime */,
8522 true /* recordGCEndTime */,
8523 true /* countCollection */ );
8524 break;
8525
8526 default:
8527 ShouldNotReachHere();
8528 }
8529 }