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_implementation/shared/adaptiveSizePolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
37 #include "gc_implementation/parNew/parNewGeneration.hpp"
38 #include "gc_implementation/shared/collectorCounters.hpp"
39 #include "gc_implementation/shared/gcTimer.hpp"
40 #include "gc_implementation/shared/gcTrace.hpp"
41 #include "gc_implementation/shared/gcTraceTime.hpp"
42 #include "gc_implementation/shared/isGCActiveMark.hpp"
43 #include "gc_interface/collectedHeap.inline.hpp"
44 #include "memory/allocation.hpp"
45 #include "memory/cardGeneration.inline.hpp"
46 #include "memory/cardTableRS.hpp"
47 #include "memory/collectorPolicy.hpp"
48 #include "memory/gcLocker.inline.hpp"
49 #include "memory/genCollectedHeap.hpp"
50 #include "memory/genMarkSweep.hpp"
51 #include "memory/genOopClosures.inline.hpp"
52 #include "memory/iterator.inline.hpp"
53 #include "memory/padded.hpp"
54 #include "memory/referencePolicy.hpp"
55 #include "memory/resourceArea.hpp"
56 #include "memory/strongRootsScope.hpp"
57 #include "memory/tenuredGeneration.hpp"
58 #include "oops/oop.inline.hpp"
59 #include "prims/jvmtiExport.hpp"
60 #include "runtime/atomic.inline.hpp"
61 #include "runtime/globals_extension.hpp"
62 #include "runtime/handles.inline.hpp"
63 #include "runtime/java.hpp"
64 #include "runtime/orderAccess.inline.hpp"
65 #include "runtime/vmThread.hpp"
66 #include "services/memoryService.hpp"
67 #include "services/runtimeService.hpp"
68 #include "utilities/stack.inline.hpp"
69
70 // statics
71 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
72 bool CMSCollector::_full_gc_requested = false;
73 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
74
75 //////////////////////////////////////////////////////////////////
76 // In support of CMS/VM thread synchronization
77 //////////////////////////////////////////////////////////////////
78 // We split use of the CGC_lock into 2 "levels".
79 // The low-level locking is of the usual CGC_lock monitor. We introduce
80 // a higher level "token" (hereafter "CMS token") built on top of the
81 // low level monitor (hereafter "CGC lock").
82 // The token-passing protocol gives priority to the VM thread. The
83 // CMS-lock doesn't provide any fairness guarantees, but clients
84 // should ensure that it is only held for very short, bounded
85 // durations.
86 //
87 // When either of the CMS thread or the VM thread is involved in
88 // collection operations during which it does not want the other
89 // thread to interfere, it obtains the CMS token.
90 //
91 // If either thread tries to get the token while the other has
92 // it, that thread waits. However, if the VM thread and CMS thread
93 // both want the token, then the VM thread gets priority while the
94 // CMS thread waits. This ensures, for instance, that the "concurrent"
95 // phases of the CMS thread's work do not block out the VM thread
96 // for long periods of time as the CMS thread continues to hog
97 // the token. (See bug 4616232).
98 //
99 // The baton-passing functions are, however, controlled by the
100 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
101 // and here the low-level CMS lock, not the high level token,
102 // ensures mutual exclusion.
103 //
104 // Two important conditions that we have to satisfy:
105 // 1. if a thread does a low-level wait on the CMS lock, then it
106 // relinquishes the CMS token if it were holding that token
107 // when it acquired the low-level CMS lock.
108 // 2. any low-level notifications on the low-level lock
109 // should only be sent when a thread has relinquished the token.
110 //
111 // In the absence of either property, we'd have potential deadlock.
112 //
113 // We protect each of the CMS (concurrent and sequential) phases
114 // with the CMS _token_, not the CMS _lock_.
115 //
116 // The only code protected by CMS lock is the token acquisition code
117 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
118 // baton-passing code.
119 //
120 // Unfortunately, i couldn't come up with a good abstraction to factor and
121 // hide the naked CGC_lock manipulation in the baton-passing code
122 // further below. That's something we should try to do. Also, the proof
123 // of correctness of this 2-level locking scheme is far from obvious,
124 // and potentially quite slippery. We have an uneasy suspicion, for instance,
125 // that there may be a theoretical possibility of delay/starvation in the
126 // low-level lock/wait/notify scheme used for the baton-passing because of
127 // potential interference with the priority scheme embodied in the
128 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
129 // invocation further below and marked with "XXX 20011219YSR".
130 // Indeed, as we note elsewhere, this may become yet more slippery
131 // in the presence of multiple CMS and/or multiple VM threads. XXX
132
133 class CMSTokenSync: public StackObj {
134 private:
135 bool _is_cms_thread;
136 public:
137 CMSTokenSync(bool is_cms_thread):
138 _is_cms_thread(is_cms_thread) {
139 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
140 "Incorrect argument to constructor");
141 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
142 }
143
144 ~CMSTokenSync() {
145 assert(_is_cms_thread ?
146 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
147 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
148 "Incorrect state");
149 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
150 }
151 };
152
153 // Convenience class that does a CMSTokenSync, and then acquires
154 // upto three locks.
155 class CMSTokenSyncWithLocks: public CMSTokenSync {
156 private:
157 // Note: locks are acquired in textual declaration order
158 // and released in the opposite order
159 MutexLockerEx _locker1, _locker2, _locker3;
160 public:
161 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
162 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
163 CMSTokenSync(is_cms_thread),
164 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
165 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
166 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
167 { }
168 };
169
170
171 //////////////////////////////////////////////////////////////////
172 // Concurrent Mark-Sweep Generation /////////////////////////////
173 //////////////////////////////////////////////////////////////////
174
175 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
176
177 // This struct contains per-thread things necessary to support parallel
178 // young-gen collection.
179 class CMSParGCThreadState: public CHeapObj<mtGC> {
180 public:
181 CFLS_LAB lab;
182 PromotionInfo promo;
183
184 // Constructor.
185 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
186 promo.setSpace(cfls);
187 }
188 };
189
190 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
191 ReservedSpace rs, size_t initial_byte_size,
192 CardTableRS* ct, bool use_adaptive_freelists,
193 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
194 CardGeneration(rs, initial_byte_size, ct),
195 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
196 _did_compact(false)
197 {
198 HeapWord* bottom = (HeapWord*) _virtual_space.low();
199 HeapWord* end = (HeapWord*) _virtual_space.high();
200
201 _direct_allocated_words = 0;
202 NOT_PRODUCT(
203 _numObjectsPromoted = 0;
204 _numWordsPromoted = 0;
205 _numObjectsAllocated = 0;
206 _numWordsAllocated = 0;
207 )
208
209 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
210 use_adaptive_freelists,
211 dictionaryChoice);
212 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
213 _cmsSpace->_gen = this;
214
215 _gc_stats = new CMSGCStats();
216
217 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
218 // offsets match. The ability to tell free chunks from objects
219 // depends on this property.
220 debug_only(
221 FreeChunk* junk = NULL;
222 assert(UseCompressedClassPointers ||
223 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
224 "Offset of FreeChunk::_prev within FreeChunk must match"
225 " that of OopDesc::_klass within OopDesc");
226 )
227 if (CollectedHeap::use_parallel_gc_threads()) {
228 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
229 _par_gc_thread_states =
230 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
231 for (uint i = 0; i < ParallelGCThreads; i++) {
232 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
233 }
234 } else {
235 _par_gc_thread_states = NULL;
236 }
237 _incremental_collection_failed = false;
238 // The "dilatation_factor" is the expansion that can occur on
239 // account of the fact that the minimum object size in the CMS
240 // generation may be larger than that in, say, a contiguous young
241 // generation.
242 // Ideally, in the calculation below, we'd compute the dilatation
243 // factor as: MinChunkSize/(promoting_gen's min object size)
244 // Since we do not have such a general query interface for the
245 // promoting generation, we'll instead just use the minimum
246 // object size (which today is a header's worth of space);
247 // note that all arithmetic is in units of HeapWords.
248 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
249 assert(_dilatation_factor >= 1.0, "from previous assert");
250 }
251
252
253 // The field "_initiating_occupancy" represents the occupancy percentage
254 // at which we trigger a new collection cycle. Unless explicitly specified
255 // via CMSInitiatingOccupancyFraction (argument "io" below), it
256 // is calculated by:
257 //
258 // Let "f" be MinHeapFreeRatio in
259 //
260 // _initiating_occupancy = 100-f +
261 // f * (CMSTriggerRatio/100)
262 // where CMSTriggerRatio is the argument "tr" below.
263 //
264 // That is, if we assume the heap is at its desired maximum occupancy at the
265 // end of a collection, we let CMSTriggerRatio of the (purported) free
266 // space be allocated before initiating a new collection cycle.
267 //
268 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
269 assert(io <= 100 && tr <= 100, "Check the arguments");
270 if (io >= 0) {
271 _initiating_occupancy = (double)io / 100.0;
272 } else {
273 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
274 (double)(tr * MinHeapFreeRatio) / 100.0)
275 / 100.0;
276 }
277 }
278
279 void ConcurrentMarkSweepGeneration::ref_processor_init() {
280 assert(collector() != NULL, "no collector");
281 collector()->ref_processor_init();
282 }
283
284 void CMSCollector::ref_processor_init() {
285 if (_ref_processor == NULL) {
286 // Allocate and initialize a reference processor
287 _ref_processor =
288 new ReferenceProcessor(_span, // span
289 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
290 (int) ParallelGCThreads, // mt processing degree
291 _cmsGen->refs_discovery_is_mt(), // mt discovery
292 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
293 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
294 &_is_alive_closure); // closure for liveness info
295 // Initialize the _ref_processor field of CMSGen
296 _cmsGen->set_ref_processor(_ref_processor);
297
298 }
299 }
300
301 AdaptiveSizePolicy* CMSCollector::size_policy() {
302 GenCollectedHeap* gch = GenCollectedHeap::heap();
303 return gch->gen_policy()->size_policy();
304 }
305
306 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
307
308 const char* gen_name = "old";
309 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy();
310
311 // Generation Counters - generation 1, 1 subspace
312 _gen_counters = new GenerationCounters(gen_name, 1, 1,
313 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
314
315 _space_counters = new GSpaceCounters(gen_name, 0,
316 _virtual_space.reserved_size(),
317 this, _gen_counters);
318 }
319
320 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
321 _cms_gen(cms_gen)
322 {
323 assert(alpha <= 100, "bad value");
324 _saved_alpha = alpha;
325
326 // Initialize the alphas to the bootstrap value of 100.
327 _gc0_alpha = _cms_alpha = 100;
328
329 _cms_begin_time.update();
330 _cms_end_time.update();
331
332 _gc0_duration = 0.0;
333 _gc0_period = 0.0;
334 _gc0_promoted = 0;
335
336 _cms_duration = 0.0;
337 _cms_period = 0.0;
338 _cms_allocated = 0;
339
340 _cms_used_at_gc0_begin = 0;
341 _cms_used_at_gc0_end = 0;
342 _allow_duty_cycle_reduction = false;
343 _valid_bits = 0;
344 }
345
346 double CMSStats::cms_free_adjustment_factor(size_t free) const {
347 // TBD: CR 6909490
348 return 1.0;
349 }
350
351 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
352 }
353
354 // If promotion failure handling is on use
355 // the padded average size of the promotion for each
356 // young generation collection.
357 double CMSStats::time_until_cms_gen_full() const {
358 size_t cms_free = _cms_gen->cmsSpace()->free();
359 GenCollectedHeap* gch = GenCollectedHeap::heap();
360 size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
361 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
362 if (cms_free > expected_promotion) {
363 // Start a cms collection if there isn't enough space to promote
364 // for the next minor collection. Use the padded average as
365 // a safety factor.
366 cms_free -= expected_promotion;
367
368 // Adjust by the safety factor.
369 double cms_free_dbl = (double)cms_free;
370 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
371 // Apply a further correction factor which tries to adjust
372 // for recent occurance of concurrent mode failures.
373 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
374 cms_free_dbl = cms_free_dbl * cms_adjustment;
375
376 if (PrintGCDetails && Verbose) {
377 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
378 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
379 cms_free, expected_promotion);
380 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
381 cms_free_dbl, cms_consumption_rate() + 1.0);
382 }
383 // Add 1 in case the consumption rate goes to zero.
384 return cms_free_dbl / (cms_consumption_rate() + 1.0);
385 }
386 return 0.0;
387 }
388
389 // Compare the duration of the cms collection to the
390 // time remaining before the cms generation is empty.
391 // Note that the time from the start of the cms collection
392 // to the start of the cms sweep (less than the total
393 // duration of the cms collection) can be used. This
394 // has been tried and some applications experienced
395 // promotion failures early in execution. This was
396 // possibly because the averages were not accurate
397 // enough at the beginning.
398 double CMSStats::time_until_cms_start() const {
399 // We add "gc0_period" to the "work" calculation
400 // below because this query is done (mostly) at the
401 // end of a scavenge, so we need to conservatively
402 // account for that much possible delay
403 // in the query so as to avoid concurrent mode failures
404 // due to starting the collection just a wee bit too
405 // late.
406 double work = cms_duration() + gc0_period();
407 double deadline = time_until_cms_gen_full();
408 // If a concurrent mode failure occurred recently, we want to be
409 // more conservative and halve our expected time_until_cms_gen_full()
410 if (work > deadline) {
411 if (Verbose && PrintGCDetails) {
412 gclog_or_tty->print(
413 " CMSCollector: collect because of anticipated promotion "
414 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
415 gc0_period(), time_until_cms_gen_full());
416 }
417 return 0.0;
418 }
419 return work - deadline;
420 }
421
422 #ifndef PRODUCT
423 void CMSStats::print_on(outputStream *st) const {
424 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
425 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
426 gc0_duration(), gc0_period(), gc0_promoted());
427 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
428 cms_duration(), cms_period(), cms_allocated());
429 st->print(",cms_since_beg=%g,cms_since_end=%g",
430 cms_time_since_begin(), cms_time_since_end());
431 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
432 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
433
434 if (valid()) {
435 st->print(",promo_rate=%g,cms_alloc_rate=%g",
436 promotion_rate(), cms_allocation_rate());
437 st->print(",cms_consumption_rate=%g,time_until_full=%g",
438 cms_consumption_rate(), time_until_cms_gen_full());
439 }
440 st->print(" ");
441 }
442 #endif // #ifndef PRODUCT
443
444 CMSCollector::CollectorState CMSCollector::_collectorState =
445 CMSCollector::Idling;
446 bool CMSCollector::_foregroundGCIsActive = false;
447 bool CMSCollector::_foregroundGCShouldWait = false;
448
449 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
450 CardTableRS* ct,
451 ConcurrentMarkSweepPolicy* cp):
452 _cmsGen(cmsGen),
453 _ct(ct),
454 _ref_processor(NULL), // will be set later
455 _conc_workers(NULL), // may be set later
456 _abort_preclean(false),
457 _start_sampling(false),
458 _between_prologue_and_epilogue(false),
459 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
460 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
461 -1 /* lock-free */, "No_lock" /* dummy */),
462 _modUnionClosure(&_modUnionTable),
463 _modUnionClosurePar(&_modUnionTable),
464 // Adjust my span to cover old (cms) gen
465 _span(cmsGen->reserved()),
466 // Construct the is_alive_closure with _span & markBitMap
467 _is_alive_closure(_span, &_markBitMap),
468 _restart_addr(NULL),
469 _overflow_list(NULL),
470 _stats(cmsGen),
471 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
472 //verify that this lock should be acquired with safepoint check.
473 Monitor::_safepoint_check_sometimes)),
474 _eden_chunk_array(NULL), // may be set in ctor body
475 _eden_chunk_capacity(0), // -- ditto --
476 _eden_chunk_index(0), // -- ditto --
477 _survivor_plab_array(NULL), // -- ditto --
478 _survivor_chunk_array(NULL), // -- ditto --
479 _survivor_chunk_capacity(0), // -- ditto --
480 _survivor_chunk_index(0), // -- ditto --
481 _ser_pmc_preclean_ovflw(0),
482 _ser_kac_preclean_ovflw(0),
483 _ser_pmc_remark_ovflw(0),
484 _par_pmc_remark_ovflw(0),
485 _ser_kac_ovflw(0),
486 _par_kac_ovflw(0),
487 #ifndef PRODUCT
488 _num_par_pushes(0),
489 #endif
490 _collection_count_start(0),
491 _verifying(false),
492 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
493 _completed_initialization(false),
494 _collector_policy(cp),
495 _should_unload_classes(CMSClassUnloadingEnabled),
496 _concurrent_cycles_since_last_unload(0),
497 _roots_scanning_options(GenCollectedHeap::SO_None),
498 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
499 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
500 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
501 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
502 _cms_start_registered(false)
503 {
504 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
505 ExplicitGCInvokesConcurrent = true;
506 }
507 // Now expand the span and allocate the collection support structures
508 // (MUT, marking bit map etc.) to cover both generations subject to
509 // collection.
510
511 // For use by dirty card to oop closures.
512 _cmsGen->cmsSpace()->set_collector(this);
513
514 // Allocate MUT and marking bit map
515 {
516 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
517 if (!_markBitMap.allocate(_span)) {
518 warning("Failed to allocate CMS Bit Map");
519 return;
520 }
521 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
522 }
523 {
524 _modUnionTable.allocate(_span);
525 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
526 }
527
528 if (!_markStack.allocate(MarkStackSize)) {
529 warning("Failed to allocate CMS Marking Stack");
530 return;
531 }
532
533 // Support for multi-threaded concurrent phases
534 if (CMSConcurrentMTEnabled) {
535 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
536 // just for now
537 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
538 }
539 if (ConcGCThreads > 1) {
540 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
541 ConcGCThreads, true);
542 if (_conc_workers == NULL) {
543 warning("GC/CMS: _conc_workers allocation failure: "
544 "forcing -CMSConcurrentMTEnabled");
545 CMSConcurrentMTEnabled = false;
546 } else {
547 _conc_workers->initialize_workers();
548 }
549 } else {
550 CMSConcurrentMTEnabled = false;
551 }
552 }
553 if (!CMSConcurrentMTEnabled) {
554 ConcGCThreads = 0;
555 } else {
556 // Turn off CMSCleanOnEnter optimization temporarily for
557 // the MT case where it's not fixed yet; see 6178663.
558 CMSCleanOnEnter = false;
559 }
560 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
561 "Inconsistency");
562
563 // Parallel task queues; these are shared for the
564 // concurrent and stop-world phases of CMS, but
565 // are not shared with parallel scavenge (ParNew).
566 {
567 uint i;
568 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
569
570 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
571 || ParallelRefProcEnabled)
572 && num_queues > 0) {
573 _task_queues = new OopTaskQueueSet(num_queues);
574 if (_task_queues == NULL) {
575 warning("task_queues allocation failure.");
576 return;
577 }
578 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
579 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
580 for (i = 0; i < num_queues; i++) {
581 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
582 if (q == NULL) {
583 warning("work_queue allocation failure.");
584 return;
585 }
586 _task_queues->register_queue(i, q);
587 }
588 for (i = 0; i < num_queues; i++) {
589 _task_queues->queue(i)->initialize();
590 _hash_seed[i] = 17; // copied from ParNew
591 }
592 }
593 }
594
595 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
596
597 // Clip CMSBootstrapOccupancy between 0 and 100.
598 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
599
600 // Now tell CMS generations the identity of their collector
601 ConcurrentMarkSweepGeneration::set_collector(this);
602
603 // Create & start a CMS thread for this CMS collector
604 _cmsThread = ConcurrentMarkSweepThread::start(this);
605 assert(cmsThread() != NULL, "CMS Thread should have been created");
606 assert(cmsThread()->collector() == this,
607 "CMS Thread should refer to this gen");
608 assert(CGC_lock != NULL, "Where's the CGC_lock?");
609
610 // Support for parallelizing young gen rescan
611 GenCollectedHeap* gch = GenCollectedHeap::heap();
612 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
613 _young_gen = (ParNewGeneration*)gch->young_gen();
614 if (gch->supports_inline_contig_alloc()) {
615 _top_addr = gch->top_addr();
616 _end_addr = gch->end_addr();
617 assert(_young_gen != NULL, "no _young_gen");
618 _eden_chunk_index = 0;
619 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
620 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
621 }
622
623 // Support for parallelizing survivor space rescan
624 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
625 const size_t max_plab_samples =
626 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size();
627
628 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
629 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
630 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
631 _survivor_chunk_capacity = 2*max_plab_samples;
632 for (uint i = 0; i < ParallelGCThreads; i++) {
633 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
634 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
635 assert(cur->end() == 0, "Should be 0");
636 assert(cur->array() == vec, "Should be vec");
637 assert(cur->capacity() == max_plab_samples, "Error");
638 }
639 }
640
641 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
642 _gc_counters = new CollectorCounters("CMS", 1);
643 _completed_initialization = true;
644 _inter_sweep_timer.start(); // start of time
645 }
646
647 size_t CMSCollector::plab_sample_minimum_size() {
648 // The default value of MinTLABSize is 2k, but there is
649 // no way to get the default value if the flag has been overridden.
650 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K);
651 }
652
653 const char* ConcurrentMarkSweepGeneration::name() const {
654 return "concurrent mark-sweep generation";
655 }
656 void ConcurrentMarkSweepGeneration::update_counters() {
657 if (UsePerfData) {
658 _space_counters->update_all();
659 _gen_counters->update_all();
660 }
661 }
662
663 // this is an optimized version of update_counters(). it takes the
664 // used value as a parameter rather than computing it.
665 //
666 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
667 if (UsePerfData) {
668 _space_counters->update_used(used);
669 _space_counters->update_capacity();
670 _gen_counters->update_all();
671 }
672 }
673
674 void ConcurrentMarkSweepGeneration::print() const {
675 Generation::print();
676 cmsSpace()->print();
677 }
678
679 #ifndef PRODUCT
680 void ConcurrentMarkSweepGeneration::print_statistics() {
681 cmsSpace()->printFLCensus(0);
682 }
683 #endif
684
685 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
686 GenCollectedHeap* gch = GenCollectedHeap::heap();
687 if (PrintGCDetails) {
688 // I didn't want to change the logging when removing the level concept,
689 // but I guess this logging could say "old" or something instead of "1".
690 assert(this == gch->old_gen(),
691 "The CMS generation should be the old generation");
692 uint level = 1;
693 if (Verbose) {
694 gclog_or_tty->print("[%u %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
695 level, short_name(), s, used(), capacity());
696 } else {
697 gclog_or_tty->print("[%u %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
698 level, short_name(), s, used() / K, capacity() / K);
699 }
700 }
701 if (Verbose) {
702 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
703 gch->used(), gch->capacity());
704 } else {
705 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
706 gch->used() / K, gch->capacity() / K);
707 }
708 }
709
710 size_t
711 ConcurrentMarkSweepGeneration::contiguous_available() const {
712 // dld proposes an improvement in precision here. If the committed
713 // part of the space ends in a free block we should add that to
714 // uncommitted size in the calculation below. Will make this
715 // change later, staying with the approximation below for the
716 // time being. -- ysr.
717 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
718 }
719
720 size_t
721 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
722 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
723 }
724
725 size_t ConcurrentMarkSweepGeneration::max_available() const {
726 return free() + _virtual_space.uncommitted_size();
727 }
728
729 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
730 size_t available = max_available();
731 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
732 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
733 if (Verbose && PrintGCDetails) {
734 gclog_or_tty->print_cr(
735 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
736 "max_promo("SIZE_FORMAT")",
737 res? "":" not", available, res? ">=":"<",
738 av_promo, max_promotion_in_bytes);
739 }
740 return res;
741 }
742
743 // At a promotion failure dump information on block layout in heap
744 // (cms old generation).
745 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
746 if (CMSDumpAtPromotionFailure) {
747 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
748 }
749 }
750
751 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
752 // Clear the promotion information. These pointers can be adjusted
753 // along with all the other pointers into the heap but
754 // compaction is expected to be a rare event with
755 // a heap using cms so don't do it without seeing the need.
756 for (uint i = 0; i < ParallelGCThreads; i++) {
757 _par_gc_thread_states[i]->promo.reset();
758 }
759 }
760
761 void ConcurrentMarkSweepGeneration::compute_new_size() {
762 assert_locked_or_safepoint(Heap_lock);
763
764 // If incremental collection failed, we just want to expand
765 // to the limit.
766 if (incremental_collection_failed()) {
767 clear_incremental_collection_failed();
768 grow_to_reserved();
769 return;
770 }
771
772 // The heap has been compacted but not reset yet.
773 // Any metric such as free() or used() will be incorrect.
774
775 CardGeneration::compute_new_size();
776
777 // Reset again after a possible resizing
778 if (did_compact()) {
779 cmsSpace()->reset_after_compaction();
780 }
781 }
782
783 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
784 assert_locked_or_safepoint(Heap_lock);
785
786 // If incremental collection failed, we just want to expand
787 // to the limit.
788 if (incremental_collection_failed()) {
789 clear_incremental_collection_failed();
790 grow_to_reserved();
791 return;
792 }
793
794 double free_percentage = ((double) free()) / capacity();
795 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
796 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
797
798 // compute expansion delta needed for reaching desired free percentage
799 if (free_percentage < desired_free_percentage) {
800 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
801 assert(desired_capacity >= capacity(), "invalid expansion size");
802 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
803 if (PrintGCDetails && Verbose) {
804 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
805 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
806 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
807 gclog_or_tty->print_cr(" Desired free fraction %f", desired_free_percentage);
808 gclog_or_tty->print_cr(" Maximum free fraction %f", maximum_free_percentage);
809 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity() / 1000);
810 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, desired_capacity / 1000);
811 GenCollectedHeap* gch = GenCollectedHeap::heap();
812 assert(this == gch->_old_gen, "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, unsafe_max_alloc_nogc() / 1000);
816 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, contiguous_available() / 1000);
817 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", expand_bytes);
818 }
819 // safe if expansion fails
820 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
821 if (PrintGCDetails && Verbose) {
822 gclog_or_tty->print_cr(" Expanded free fraction %f",
823 ((double) free()) / capacity());
824 }
825 } else {
826 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
827 assert(desired_capacity <= capacity(), "invalid expansion size");
828 size_t shrink_bytes = capacity() - desired_capacity;
829 // Don't shrink unless the delta is greater than the minimum shrink we want
830 if (shrink_bytes >= MinHeapDeltaBytes) {
831 shrink_free_list_by(shrink_bytes);
832 }
833 }
834 }
835
836 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
837 return cmsSpace()->freelistLock();
838 }
839
840 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
841 bool tlab) {
842 CMSSynchronousYieldRequest yr;
843 MutexLockerEx x(freelistLock(),
844 Mutex::_no_safepoint_check_flag);
845 return have_lock_and_allocate(size, tlab);
846 }
847
848 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
849 bool tlab /* ignored */) {
850 assert_lock_strong(freelistLock());
851 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
852 HeapWord* res = cmsSpace()->allocate(adjustedSize);
853 // Allocate the object live (grey) if the background collector has
854 // started marking. This is necessary because the marker may
855 // have passed this address and consequently this object will
856 // not otherwise be greyed and would be incorrectly swept up.
857 // Note that if this object contains references, the writing
858 // of those references will dirty the card containing this object
859 // allowing the object to be blackened (and its references scanned)
860 // either during a preclean phase or at the final checkpoint.
861 if (res != NULL) {
862 // We may block here with an uninitialized object with
863 // its mark-bit or P-bits not yet set. Such objects need
864 // to be safely navigable by block_start().
865 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
866 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
867 collector()->direct_allocated(res, adjustedSize);
868 _direct_allocated_words += adjustedSize;
869 // allocation counters
870 NOT_PRODUCT(
871 _numObjectsAllocated++;
872 _numWordsAllocated += (int)adjustedSize;
873 )
874 }
875 return res;
876 }
877
878 // In the case of direct allocation by mutators in a generation that
879 // is being concurrently collected, the object must be allocated
880 // live (grey) if the background collector has started marking.
881 // This is necessary because the marker may
882 // have passed this address and consequently this object will
883 // not otherwise be greyed and would be incorrectly swept up.
884 // Note that if this object contains references, the writing
885 // of those references will dirty the card containing this object
886 // allowing the object to be blackened (and its references scanned)
887 // either during a preclean phase or at the final checkpoint.
888 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
889 assert(_markBitMap.covers(start, size), "Out of bounds");
890 if (_collectorState >= Marking) {
891 MutexLockerEx y(_markBitMap.lock(),
892 Mutex::_no_safepoint_check_flag);
893 // [see comments preceding SweepClosure::do_blk() below for details]
894 //
895 // Can the P-bits be deleted now? JJJ
896 //
897 // 1. need to mark the object as live so it isn't collected
898 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
899 // 3. need to mark the end of the object so marking, precleaning or sweeping
900 // can skip over uninitialized or unparsable objects. An allocated
901 // object is considered uninitialized for our purposes as long as
902 // its klass word is NULL. All old gen objects are parsable
903 // as soon as they are initialized.)
904 _markBitMap.mark(start); // object is live
905 _markBitMap.mark(start + 1); // object is potentially uninitialized?
906 _markBitMap.mark(start + size - 1);
907 // mark end of object
908 }
909 // check that oop looks uninitialized
910 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
911 }
912
913 void CMSCollector::promoted(bool par, HeapWord* start,
914 bool is_obj_array, size_t obj_size) {
915 assert(_markBitMap.covers(start), "Out of bounds");
916 // See comment in direct_allocated() about when objects should
917 // be allocated live.
918 if (_collectorState >= Marking) {
919 // we already hold the marking bit map lock, taken in
920 // the prologue
921 if (par) {
922 _markBitMap.par_mark(start);
923 } else {
924 _markBitMap.mark(start);
925 }
926 // We don't need to mark the object as uninitialized (as
927 // in direct_allocated above) because this is being done with the
928 // world stopped and the object will be initialized by the
929 // time the marking, precleaning or sweeping get to look at it.
930 // But see the code for copying objects into the CMS generation,
931 // where we need to ensure that concurrent readers of the
932 // block offset table are able to safely navigate a block that
933 // is in flux from being free to being allocated (and in
934 // transition while being copied into) and subsequently
935 // becoming a bona-fide object when the copy/promotion is complete.
936 assert(SafepointSynchronize::is_at_safepoint(),
937 "expect promotion only at safepoints");
938
939 if (_collectorState < Sweeping) {
940 // Mark the appropriate cards in the modUnionTable, so that
941 // this object gets scanned before the sweep. If this is
942 // not done, CMS generation references in the object might
943 // not get marked.
944 // For the case of arrays, which are otherwise precisely
945 // marked, we need to dirty the entire array, not just its head.
946 if (is_obj_array) {
947 // The [par_]mark_range() method expects mr.end() below to
948 // be aligned to the granularity of a bit's representation
949 // in the heap. In the case of the MUT below, that's a
950 // card size.
951 MemRegion mr(start,
952 (HeapWord*)round_to((intptr_t)(start + obj_size),
953 CardTableModRefBS::card_size /* bytes */));
954 if (par) {
955 _modUnionTable.par_mark_range(mr);
956 } else {
957 _modUnionTable.mark_range(mr);
958 }
959 } else { // not an obj array; we can just mark the head
960 if (par) {
961 _modUnionTable.par_mark(start);
962 } else {
963 _modUnionTable.mark(start);
964 }
965 }
966 }
967 }
968 }
969
970 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
971 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
972 // allocate, copy and if necessary update promoinfo --
973 // delegate to underlying space.
974 assert_lock_strong(freelistLock());
975
976 #ifndef PRODUCT
977 if (GenCollectedHeap::heap()->promotion_should_fail()) {
978 return NULL;
979 }
980 #endif // #ifndef PRODUCT
981
982 oop res = _cmsSpace->promote(obj, obj_size);
983 if (res == NULL) {
984 // expand and retry
985 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
986 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
987 // Since this is the old generation, we don't try to promote
988 // into a more senior generation.
989 res = _cmsSpace->promote(obj, obj_size);
990 }
991 if (res != NULL) {
992 // See comment in allocate() about when objects should
993 // be allocated live.
994 assert(obj->is_oop(), "Will dereference klass pointer below");
995 collector()->promoted(false, // Not parallel
996 (HeapWord*)res, obj->is_objArray(), obj_size);
997 // promotion counters
998 NOT_PRODUCT(
999 _numObjectsPromoted++;
1000 _numWordsPromoted +=
1001 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1002 )
1003 }
1004 return res;
1005 }
1006
1007
1008 // IMPORTANT: Notes on object size recognition in CMS.
1009 // ---------------------------------------------------
1010 // A block of storage in the CMS generation is always in
1011 // one of three states. A free block (FREE), an allocated
1012 // object (OBJECT) whose size() method reports the correct size,
1013 // and an intermediate state (TRANSIENT) in which its size cannot
1014 // be accurately determined.
1015 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1016 // -----------------------------------------------------
1017 // FREE: klass_word & 1 == 1; mark_word holds block size
1018 //
1019 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1020 // obj->size() computes correct size
1021 //
1022 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1023 //
1024 // STATE IDENTIFICATION: (64 bit+COOPS)
1025 // ------------------------------------
1026 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1027 //
1028 // OBJECT: klass_word installed; klass_word != 0;
1029 // obj->size() computes correct size
1030 //
1031 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1032 //
1033 //
1034 // STATE TRANSITION DIAGRAM
1035 //
1036 // mut / parnew mut / parnew
1037 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1038 // ^ |
1039 // |------------------------ DEAD <------------------------------------|
1040 // sweep mut
1041 //
1042 // While a block is in TRANSIENT state its size cannot be determined
1043 // so readers will either need to come back later or stall until
1044 // the size can be determined. Note that for the case of direct
1045 // allocation, P-bits, when available, may be used to determine the
1046 // size of an object that may not yet have been initialized.
1047
1048 // Things to support parallel young-gen collection.
1049 oop
1050 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1051 oop old, markOop m,
1052 size_t word_sz) {
1053 #ifndef PRODUCT
1054 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1055 return NULL;
1056 }
1057 #endif // #ifndef PRODUCT
1058
1059 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1060 PromotionInfo* promoInfo = &ps->promo;
1061 // if we are tracking promotions, then first ensure space for
1062 // promotion (including spooling space for saving header if necessary).
1063 // then allocate and copy, then track promoted info if needed.
1064 // When tracking (see PromotionInfo::track()), the mark word may
1065 // be displaced and in this case restoration of the mark word
1066 // occurs in the (oop_since_save_marks_)iterate phase.
1067 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1068 // Out of space for allocating spooling buffers;
1069 // try expanding and allocating spooling buffers.
1070 if (!expand_and_ensure_spooling_space(promoInfo)) {
1071 return NULL;
1072 }
1073 }
1074 assert(promoInfo->has_spooling_space(), "Control point invariant");
1075 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1076 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1077 if (obj_ptr == NULL) {
1078 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1079 if (obj_ptr == NULL) {
1080 return NULL;
1081 }
1082 }
1083 oop obj = oop(obj_ptr);
1084 OrderAccess::storestore();
1085 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1086 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1087 // IMPORTANT: See note on object initialization for CMS above.
1088 // Otherwise, copy the object. Here we must be careful to insert the
1089 // klass pointer last, since this marks the block as an allocated object.
1090 // Except with compressed oops it's the mark word.
1091 HeapWord* old_ptr = (HeapWord*)old;
1092 // Restore the mark word copied above.
1093 obj->set_mark(m);
1094 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1095 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1096 OrderAccess::storestore();
1097
1098 if (UseCompressedClassPointers) {
1099 // Copy gap missed by (aligned) header size calculation below
1100 obj->set_klass_gap(old->klass_gap());
1101 }
1102 if (word_sz > (size_t)oopDesc::header_size()) {
1103 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1104 obj_ptr + oopDesc::header_size(),
1105 word_sz - oopDesc::header_size());
1106 }
1107
1108 // Now we can track the promoted object, if necessary. We take care
1109 // to delay the transition from uninitialized to full object
1110 // (i.e., insertion of klass pointer) until after, so that it
1111 // atomically becomes a promoted object.
1112 if (promoInfo->tracking()) {
1113 promoInfo->track((PromotedObject*)obj, old->klass());
1114 }
1115 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1116 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1117 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1118
1119 // Finally, install the klass pointer (this should be volatile).
1120 OrderAccess::storestore();
1121 obj->set_klass(old->klass());
1122 // We should now be able to calculate the right size for this object
1123 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1124
1125 collector()->promoted(true, // parallel
1126 obj_ptr, old->is_objArray(), word_sz);
1127
1128 NOT_PRODUCT(
1129 Atomic::inc_ptr(&_numObjectsPromoted);
1130 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1131 )
1132
1133 return obj;
1134 }
1135
1136 void
1137 ConcurrentMarkSweepGeneration::
1138 par_promote_alloc_done(int thread_num) {
1139 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1140 ps->lab.retire(thread_num);
1141 }
1142
1143 void
1144 ConcurrentMarkSweepGeneration::
1145 par_oop_since_save_marks_iterate_done(int thread_num) {
1146 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1147 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1148 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1149 }
1150
1151 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1152 size_t size,
1153 bool tlab)
1154 {
1155 // We allow a STW collection only if a full
1156 // collection was requested.
1157 return full || should_allocate(size, tlab); // FIX ME !!!
1158 // This and promotion failure handling are connected at the
1159 // hip and should be fixed by untying them.
1160 }
1161
1162 bool CMSCollector::shouldConcurrentCollect() {
1163 if (_full_gc_requested) {
1164 if (Verbose && PrintGCDetails) {
1165 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1166 " gc request (or gc_locker)");
1167 }
1168 return true;
1169 }
1170
1171 FreelistLocker x(this);
1172 // ------------------------------------------------------------------
1173 // Print out lots of information which affects the initiation of
1174 // a collection.
1175 if (PrintCMSInitiationStatistics && stats().valid()) {
1176 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1177 gclog_or_tty->stamp();
1178 gclog_or_tty->cr();
1179 stats().print_on(gclog_or_tty);
1180 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1181 stats().time_until_cms_gen_full());
1182 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1183 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1184 _cmsGen->contiguous_available());
1185 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1186 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1187 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1188 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1189 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1190 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1191 gclog_or_tty->print_cr("metadata initialized %d",
1192 MetaspaceGC::should_concurrent_collect());
1193 }
1194 // ------------------------------------------------------------------
1195
1196 // If the estimated time to complete a cms collection (cms_duration())
1197 // is less than the estimated time remaining until the cms generation
1198 // is full, start a collection.
1199 if (!UseCMSInitiatingOccupancyOnly) {
1200 if (stats().valid()) {
1201 if (stats().time_until_cms_start() == 0.0) {
1202 return true;
1203 }
1204 } else {
1205 // We want to conservatively collect somewhat early in order
1206 // to try and "bootstrap" our CMS/promotion statistics;
1207 // this branch will not fire after the first successful CMS
1208 // collection because the stats should then be valid.
1209 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1210 if (Verbose && PrintGCDetails) {
1211 gclog_or_tty->print_cr(
1212 " CMSCollector: collect for bootstrapping statistics:"
1213 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1214 _bootstrap_occupancy);
1215 }
1216 return true;
1217 }
1218 }
1219 }
1220
1221 // Otherwise, we start a collection cycle if
1222 // old gen want a collection cycle started. Each may use
1223 // an appropriate criterion for making this decision.
1224 // XXX We need to make sure that the gen expansion
1225 // criterion dovetails well with this. XXX NEED TO FIX THIS
1226 if (_cmsGen->should_concurrent_collect()) {
1227 if (Verbose && PrintGCDetails) {
1228 gclog_or_tty->print_cr("CMS old gen initiated");
1229 }
1230 return true;
1231 }
1232
1233 // We start a collection if we believe an incremental collection may fail;
1234 // this is not likely to be productive in practice because it's probably too
1235 // late anyway.
1236 GenCollectedHeap* gch = GenCollectedHeap::heap();
1237 assert(gch->collector_policy()->is_generation_policy(),
1238 "You may want to check the correctness of the following");
1239 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1240 if (Verbose && PrintGCDetails) {
1241 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1242 }
1243 return true;
1244 }
1245
1246 if (MetaspaceGC::should_concurrent_collect()) {
1247 if (Verbose && PrintGCDetails) {
1248 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1249 }
1250 return true;
1251 }
1252
1253 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1254 if (CMSTriggerInterval >= 0) {
1255 if (CMSTriggerInterval == 0) {
1256 // Trigger always
1257 return true;
1258 }
1259
1260 // Check the CMS time since begin (we do not check the stats validity
1261 // as we want to be able to trigger the first CMS cycle as well)
1262 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1263 if (Verbose && PrintGCDetails) {
1264 if (stats().valid()) {
1265 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1266 stats().cms_time_since_begin());
1267 } else {
1268 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1269 }
1270 }
1271 return true;
1272 }
1273 }
1274
1275 return false;
1276 }
1277
1278 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1279
1280 // Clear _expansion_cause fields of constituent generations
1281 void CMSCollector::clear_expansion_cause() {
1282 _cmsGen->clear_expansion_cause();
1283 }
1284
1285 // We should be conservative in starting a collection cycle. To
1286 // start too eagerly runs the risk of collecting too often in the
1287 // extreme. To collect too rarely falls back on full collections,
1288 // which works, even if not optimum in terms of concurrent work.
1289 // As a work around for too eagerly collecting, use the flag
1290 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1291 // giving the user an easily understandable way of controlling the
1292 // collections.
1293 // We want to start a new collection cycle if any of the following
1294 // conditions hold:
1295 // . our current occupancy exceeds the configured initiating occupancy
1296 // for this generation, or
1297 // . we recently needed to expand this space and have not, since that
1298 // expansion, done a collection of this generation, or
1299 // . the underlying space believes that it may be a good idea to initiate
1300 // a concurrent collection (this may be based on criteria such as the
1301 // following: the space uses linear allocation and linear allocation is
1302 // going to fail, or there is believed to be excessive fragmentation in
1303 // the generation, etc... or ...
1304 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1305 // the case of the old generation; see CR 6543076):
1306 // we may be approaching a point at which allocation requests may fail because
1307 // we will be out of sufficient free space given allocation rate estimates.]
1308 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1309
1310 assert_lock_strong(freelistLock());
1311 if (occupancy() > initiating_occupancy()) {
1312 if (PrintGCDetails && Verbose) {
1313 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1314 short_name(), occupancy(), initiating_occupancy());
1315 }
1316 return true;
1317 }
1318 if (UseCMSInitiatingOccupancyOnly) {
1319 return false;
1320 }
1321 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1322 if (PrintGCDetails && Verbose) {
1323 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1324 short_name());
1325 }
1326 return true;
1327 }
1328 if (_cmsSpace->should_concurrent_collect()) {
1329 if (PrintGCDetails && Verbose) {
1330 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1331 short_name());
1332 }
1333 return true;
1334 }
1335 return false;
1336 }
1337
1338 void ConcurrentMarkSweepGeneration::collect(bool full,
1339 bool clear_all_soft_refs,
1340 size_t size,
1341 bool tlab)
1342 {
1343 collector()->collect(full, clear_all_soft_refs, size, tlab);
1344 }
1345
1346 void CMSCollector::collect(bool full,
1347 bool clear_all_soft_refs,
1348 size_t size,
1349 bool tlab)
1350 {
1351 // The following "if" branch is present for defensive reasons.
1352 // In the current uses of this interface, it can be replaced with:
1353 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1354 // But I am not placing that assert here to allow future
1355 // generality in invoking this interface.
1356 if (GC_locker::is_active()) {
1357 // A consistency test for GC_locker
1358 assert(GC_locker::needs_gc(), "Should have been set already");
1359 // Skip this foreground collection, instead
1360 // expanding the heap if necessary.
1361 // Need the free list locks for the call to free() in compute_new_size()
1362 compute_new_size();
1363 return;
1364 }
1365 acquire_control_and_collect(full, clear_all_soft_refs);
1366 }
1367
1368 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1369 GenCollectedHeap* gch = GenCollectedHeap::heap();
1370 unsigned int gc_count = gch->total_full_collections();
1371 if (gc_count == full_gc_count) {
1372 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1373 _full_gc_requested = true;
1374 _full_gc_cause = cause;
1375 CGC_lock->notify(); // nudge CMS thread
1376 } else {
1377 assert(gc_count > full_gc_count, "Error: causal loop");
1378 }
1379 }
1380
1381 bool CMSCollector::is_external_interruption() {
1382 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1383 return GCCause::is_user_requested_gc(cause) ||
1384 GCCause::is_serviceability_requested_gc(cause);
1385 }
1386
1387 void CMSCollector::report_concurrent_mode_interruption() {
1388 if (is_external_interruption()) {
1389 if (PrintGCDetails) {
1390 gclog_or_tty->print(" (concurrent mode interrupted)");
1391 }
1392 } else {
1393 if (PrintGCDetails) {
1394 gclog_or_tty->print(" (concurrent mode failure)");
1395 }
1396 _gc_tracer_cm->report_concurrent_mode_failure();
1397 }
1398 }
1399
1400
1401 // The foreground and background collectors need to coordinate in order
1402 // to make sure that they do not mutually interfere with CMS collections.
1403 // When a background collection is active,
1404 // the foreground collector may need to take over (preempt) and
1405 // synchronously complete an ongoing collection. Depending on the
1406 // frequency of the background collections and the heap usage
1407 // of the application, this preemption can be seldom or frequent.
1408 // There are only certain
1409 // points in the background collection that the "collection-baton"
1410 // can be passed to the foreground collector.
1411 //
1412 // The foreground collector will wait for the baton before
1413 // starting any part of the collection. The foreground collector
1414 // will only wait at one location.
1415 //
1416 // The background collector will yield the baton before starting a new
1417 // phase of the collection (e.g., before initial marking, marking from roots,
1418 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1419 // of the loop which switches the phases. The background collector does some
1420 // of the phases (initial mark, final re-mark) with the world stopped.
1421 // Because of locking involved in stopping the world,
1422 // the foreground collector should not block waiting for the background
1423 // collector when it is doing a stop-the-world phase. The background
1424 // collector will yield the baton at an additional point just before
1425 // it enters a stop-the-world phase. Once the world is stopped, the
1426 // background collector checks the phase of the collection. If the
1427 // phase has not changed, it proceeds with the collection. If the
1428 // phase has changed, it skips that phase of the collection. See
1429 // the comments on the use of the Heap_lock in collect_in_background().
1430 //
1431 // Variable used in baton passing.
1432 // _foregroundGCIsActive - Set to true by the foreground collector when
1433 // it wants the baton. The foreground clears it when it has finished
1434 // the collection.
1435 // _foregroundGCShouldWait - Set to true by the background collector
1436 // when it is running. The foreground collector waits while
1437 // _foregroundGCShouldWait is true.
1438 // CGC_lock - monitor used to protect access to the above variables
1439 // and to notify the foreground and background collectors.
1440 // _collectorState - current state of the CMS collection.
1441 //
1442 // The foreground collector
1443 // acquires the CGC_lock
1444 // sets _foregroundGCIsActive
1445 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1446 // various locks acquired in preparation for the collection
1447 // are released so as not to block the background collector
1448 // that is in the midst of a collection
1449 // proceeds with the collection
1450 // clears _foregroundGCIsActive
1451 // returns
1452 //
1453 // The background collector in a loop iterating on the phases of the
1454 // collection
1455 // acquires the CGC_lock
1456 // sets _foregroundGCShouldWait
1457 // if _foregroundGCIsActive is set
1458 // clears _foregroundGCShouldWait, notifies _CGC_lock
1459 // waits on _CGC_lock for _foregroundGCIsActive to become false
1460 // and exits the loop.
1461 // otherwise
1462 // proceed with that phase of the collection
1463 // if the phase is a stop-the-world phase,
1464 // yield the baton once more just before enqueueing
1465 // the stop-world CMS operation (executed by the VM thread).
1466 // returns after all phases of the collection are done
1467 //
1468
1469 void CMSCollector::acquire_control_and_collect(bool full,
1470 bool clear_all_soft_refs) {
1471 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1472 assert(!Thread::current()->is_ConcurrentGC_thread(),
1473 "shouldn't try to acquire control from self!");
1474
1475 // Start the protocol for acquiring control of the
1476 // collection from the background collector (aka CMS thread).
1477 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1478 "VM thread should have CMS token");
1479 // Remember the possibly interrupted state of an ongoing
1480 // concurrent collection
1481 CollectorState first_state = _collectorState;
1482
1483 // Signal to a possibly ongoing concurrent collection that
1484 // we want to do a foreground collection.
1485 _foregroundGCIsActive = true;
1486
1487 // release locks and wait for a notify from the background collector
1488 // releasing the locks in only necessary for phases which
1489 // do yields to improve the granularity of the collection.
1490 assert_lock_strong(bitMapLock());
1491 // We need to lock the Free list lock for the space that we are
1492 // currently collecting.
1493 assert(haveFreelistLocks(), "Must be holding free list locks");
1494 bitMapLock()->unlock();
1495 releaseFreelistLocks();
1496 {
1497 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1498 if (_foregroundGCShouldWait) {
1499 // We are going to be waiting for action for the CMS thread;
1500 // it had better not be gone (for instance at shutdown)!
1501 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1502 "CMS thread must be running");
1503 // Wait here until the background collector gives us the go-ahead
1504 ConcurrentMarkSweepThread::clear_CMS_flag(
1505 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1506 // Get a possibly blocked CMS thread going:
1507 // Note that we set _foregroundGCIsActive true above,
1508 // without protection of the CGC_lock.
1509 CGC_lock->notify();
1510 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1511 "Possible deadlock");
1512 while (_foregroundGCShouldWait) {
1513 // wait for notification
1514 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1515 // Possibility of delay/starvation here, since CMS token does
1516 // not know to give priority to VM thread? Actually, i think
1517 // there wouldn't be any delay/starvation, but the proof of
1518 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1519 }
1520 ConcurrentMarkSweepThread::set_CMS_flag(
1521 ConcurrentMarkSweepThread::CMS_vm_has_token);
1522 }
1523 }
1524 // The CMS_token is already held. Get back the other locks.
1525 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1526 "VM thread should have CMS token");
1527 getFreelistLocks();
1528 bitMapLock()->lock_without_safepoint_check();
1529 if (TraceCMSState) {
1530 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1531 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1532 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1533 }
1534
1535 // Inform cms gen if this was due to partial collection failing.
1536 // The CMS gen may use this fact to determine its expansion policy.
1537 GenCollectedHeap* gch = GenCollectedHeap::heap();
1538 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1539 assert(!_cmsGen->incremental_collection_failed(),
1540 "Should have been noticed, reacted to and cleared");
1541 _cmsGen->set_incremental_collection_failed();
1542 }
1543
1544 if (first_state > Idling) {
1545 report_concurrent_mode_interruption();
1546 }
1547
1548 set_did_compact(true);
1549
1550 // If the collection is being acquired from the background
1551 // collector, there may be references on the discovered
1552 // references lists. Abandon those references, since some
1553 // of them may have become unreachable after concurrent
1554 // discovery; the STW compacting collector will redo discovery
1555 // more precisely, without being subject to floating garbage.
1556 // Leaving otherwise unreachable references in the discovered
1557 // lists would require special handling.
1558 ref_processor()->disable_discovery();
1559 ref_processor()->abandon_partial_discovery();
1560 ref_processor()->verify_no_references_recorded();
1561
1562 if (first_state > Idling) {
1563 save_heap_summary();
1564 }
1565
1566 do_compaction_work(clear_all_soft_refs);
1567
1568 // Has the GC time limit been exceeded?
1569 size_t max_eden_size = _young_gen->max_capacity() -
1570 _young_gen->to()->capacity() -
1571 _young_gen->from()->capacity();
1572 GCCause::Cause gc_cause = gch->gc_cause();
1573 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1574 _young_gen->eden()->used(),
1575 _cmsGen->max_capacity(),
1576 max_eden_size,
1577 full,
1578 gc_cause,
1579 gch->collector_policy());
1580
1581 // Reset the expansion cause, now that we just completed
1582 // a collection cycle.
1583 clear_expansion_cause();
1584 _foregroundGCIsActive = false;
1585 return;
1586 }
1587
1588 // Resize the tenured generation
1589 // after obtaining the free list locks for the
1590 // two generations.
1591 void CMSCollector::compute_new_size() {
1592 assert_locked_or_safepoint(Heap_lock);
1593 FreelistLocker z(this);
1594 MetaspaceGC::compute_new_size();
1595 _cmsGen->compute_new_size_free_list();
1596 }
1597
1598 // A work method used by the foreground collector to do
1599 // a mark-sweep-compact.
1600 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1601 GenCollectedHeap* gch = GenCollectedHeap::heap();
1602
1603 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1604 gc_timer->register_gc_start();
1605
1606 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1607 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1608
1609 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1610
1611 // Temporarily widen the span of the weak reference processing to
1612 // the entire heap.
1613 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1614 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1615 // Temporarily, clear the "is_alive_non_header" field of the
1616 // reference processor.
1617 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1618 // Temporarily make reference _processing_ single threaded (non-MT).
1619 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1620 // Temporarily make refs discovery atomic
1621 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1622 // Temporarily make reference _discovery_ single threaded (non-MT)
1623 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1624
1625 ref_processor()->set_enqueuing_is_done(false);
1626 ref_processor()->enable_discovery();
1627 ref_processor()->setup_policy(clear_all_soft_refs);
1628 // If an asynchronous collection finishes, the _modUnionTable is
1629 // all clear. If we are assuming the collection from an asynchronous
1630 // collection, clear the _modUnionTable.
1631 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1632 "_modUnionTable should be clear if the baton was not passed");
1633 _modUnionTable.clear_all();
1634 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1635 "mod union for klasses should be clear if the baton was passed");
1636 _ct->klass_rem_set()->clear_mod_union();
1637
1638 // We must adjust the allocation statistics being maintained
1639 // in the free list space. We do so by reading and clearing
1640 // the sweep timer and updating the block flux rate estimates below.
1641 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1642 if (_inter_sweep_timer.is_active()) {
1643 _inter_sweep_timer.stop();
1644 // Note that we do not use this sample to update the _inter_sweep_estimate.
1645 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1646 _inter_sweep_estimate.padded_average(),
1647 _intra_sweep_estimate.padded_average());
1648 }
1649
1650 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1651 #ifdef ASSERT
1652 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1653 size_t free_size = cms_space->free();
1654 assert(free_size ==
1655 pointer_delta(cms_space->end(), cms_space->compaction_top())
1656 * HeapWordSize,
1657 "All the free space should be compacted into one chunk at top");
1658 assert(cms_space->dictionary()->total_chunk_size(
1659 debug_only(cms_space->freelistLock())) == 0 ||
1660 cms_space->totalSizeInIndexedFreeLists() == 0,
1661 "All the free space should be in a single chunk");
1662 size_t num = cms_space->totalCount();
1663 assert((free_size == 0 && num == 0) ||
1664 (free_size > 0 && (num == 1 || num == 2)),
1665 "There should be at most 2 free chunks after compaction");
1666 #endif // ASSERT
1667 _collectorState = Resetting;
1668 assert(_restart_addr == NULL,
1669 "Should have been NULL'd before baton was passed");
1670 reset(false /* == !concurrent */);
1671 _cmsGen->reset_after_compaction();
1672 _concurrent_cycles_since_last_unload = 0;
1673
1674 // Clear any data recorded in the PLAB chunk arrays.
1675 if (_survivor_plab_array != NULL) {
1676 reset_survivor_plab_arrays();
1677 }
1678
1679 // Adjust the per-size allocation stats for the next epoch.
1680 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1681 // Restart the "inter sweep timer" for the next epoch.
1682 _inter_sweep_timer.reset();
1683 _inter_sweep_timer.start();
1684
1685 gc_timer->register_gc_end();
1686
1687 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1688
1689 // For a mark-sweep-compact, compute_new_size() will be called
1690 // in the heap's do_collection() method.
1691 }
1692
1693 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1694 ContiguousSpace* eden_space = _young_gen->eden();
1695 ContiguousSpace* from_space = _young_gen->from();
1696 ContiguousSpace* to_space = _young_gen->to();
1697 // Eden
1698 if (_eden_chunk_array != NULL) {
1699 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1700 p2i(eden_space->bottom()), p2i(eden_space->top()),
1701 p2i(eden_space->end()), eden_space->capacity());
1702 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1703 "_eden_chunk_capacity=" SIZE_FORMAT,
1704 _eden_chunk_index, _eden_chunk_capacity);
1705 for (size_t i = 0; i < _eden_chunk_index; i++) {
1706 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1707 i, p2i(_eden_chunk_array[i]));
1708 }
1709 }
1710 // Survivor
1711 if (_survivor_chunk_array != NULL) {
1712 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1713 p2i(from_space->bottom()), p2i(from_space->top()),
1714 p2i(from_space->end()), from_space->capacity());
1715 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1716 "_survivor_chunk_capacity=" SIZE_FORMAT,
1717 _survivor_chunk_index, _survivor_chunk_capacity);
1718 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1719 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1720 i, p2i(_survivor_chunk_array[i]));
1721 }
1722 }
1723 }
1724
1725 void CMSCollector::getFreelistLocks() const {
1726 // Get locks for all free lists in all generations that this
1727 // collector is responsible for
1728 _cmsGen->freelistLock()->lock_without_safepoint_check();
1729 }
1730
1731 void CMSCollector::releaseFreelistLocks() const {
1732 // Release locks for all free lists in all generations that this
1733 // collector is responsible for
1734 _cmsGen->freelistLock()->unlock();
1735 }
1736
1737 bool CMSCollector::haveFreelistLocks() const {
1738 // Check locks for all free lists in all generations that this
1739 // collector is responsible for
1740 assert_lock_strong(_cmsGen->freelistLock());
1741 PRODUCT_ONLY(ShouldNotReachHere());
1742 return true;
1743 }
1744
1745 // A utility class that is used by the CMS collector to
1746 // temporarily "release" the foreground collector from its
1747 // usual obligation to wait for the background collector to
1748 // complete an ongoing phase before proceeding.
1749 class ReleaseForegroundGC: public StackObj {
1750 private:
1751 CMSCollector* _c;
1752 public:
1753 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1754 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1755 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1756 // allow a potentially blocked foreground collector to proceed
1757 _c->_foregroundGCShouldWait = false;
1758 if (_c->_foregroundGCIsActive) {
1759 CGC_lock->notify();
1760 }
1761 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1762 "Possible deadlock");
1763 }
1764
1765 ~ReleaseForegroundGC() {
1766 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1767 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1768 _c->_foregroundGCShouldWait = true;
1769 }
1770 };
1771
1772 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1773 assert(Thread::current()->is_ConcurrentGC_thread(),
1774 "A CMS asynchronous collection is only allowed on a CMS thread.");
1775
1776 GenCollectedHeap* gch = GenCollectedHeap::heap();
1777 {
1778 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1779 MutexLockerEx hl(Heap_lock, safepoint_check);
1780 FreelistLocker fll(this);
1781 MutexLockerEx x(CGC_lock, safepoint_check);
1782 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1783 // The foreground collector is active or we're
1784 // not using asynchronous collections. Skip this
1785 // background collection.
1786 assert(!_foregroundGCShouldWait, "Should be clear");
1787 return;
1788 } else {
1789 assert(_collectorState == Idling, "Should be idling before start.");
1790 _collectorState = InitialMarking;
1791 register_gc_start(cause);
1792 // Reset the expansion cause, now that we are about to begin
1793 // a new cycle.
1794 clear_expansion_cause();
1795
1796 // Clear the MetaspaceGC flag since a concurrent collection
1797 // is starting but also clear it after the collection.
1798 MetaspaceGC::set_should_concurrent_collect(false);
1799 }
1800 // Decide if we want to enable class unloading as part of the
1801 // ensuing concurrent GC cycle.
1802 update_should_unload_classes();
1803 _full_gc_requested = false; // acks all outstanding full gc requests
1804 _full_gc_cause = GCCause::_no_gc;
1805 // Signal that we are about to start a collection
1806 gch->increment_total_full_collections(); // ... starting a collection cycle
1807 _collection_count_start = gch->total_full_collections();
1808 }
1809
1810 // Used for PrintGC
1811 size_t prev_used;
1812 if (PrintGC && Verbose) {
1813 prev_used = _cmsGen->used();
1814 }
1815
1816 // The change of the collection state is normally done at this level;
1817 // the exceptions are phases that are executed while the world is
1818 // stopped. For those phases the change of state is done while the
1819 // world is stopped. For baton passing purposes this allows the
1820 // background collector to finish the phase and change state atomically.
1821 // The foreground collector cannot wait on a phase that is done
1822 // while the world is stopped because the foreground collector already
1823 // has the world stopped and would deadlock.
1824 while (_collectorState != Idling) {
1825 if (TraceCMSState) {
1826 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1827 p2i(Thread::current()), _collectorState);
1828 }
1829 // The foreground collector
1830 // holds the Heap_lock throughout its collection.
1831 // holds the CMS token (but not the lock)
1832 // except while it is waiting for the background collector to yield.
1833 //
1834 // The foreground collector should be blocked (not for long)
1835 // if the background collector is about to start a phase
1836 // executed with world stopped. If the background
1837 // collector has already started such a phase, the
1838 // foreground collector is blocked waiting for the
1839 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1840 // are executed in the VM thread.
1841 //
1842 // The locking order is
1843 // PendingListLock (PLL) -- if applicable (FinalMarking)
1844 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1845 // CMS token (claimed in
1846 // stop_world_and_do() -->
1847 // safepoint_synchronize() -->
1848 // CMSThread::synchronize())
1849
1850 {
1851 // Check if the FG collector wants us to yield.
1852 CMSTokenSync x(true); // is cms thread
1853 if (waitForForegroundGC()) {
1854 // We yielded to a foreground GC, nothing more to be
1855 // done this round.
1856 assert(_foregroundGCShouldWait == false, "We set it to false in "
1857 "waitForForegroundGC()");
1858 if (TraceCMSState) {
1859 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1860 " exiting collection CMS state %d",
1861 p2i(Thread::current()), _collectorState);
1862 }
1863 return;
1864 } else {
1865 // The background collector can run but check to see if the
1866 // foreground collector has done a collection while the
1867 // background collector was waiting to get the CGC_lock
1868 // above. If yes, break so that _foregroundGCShouldWait
1869 // is cleared before returning.
1870 if (_collectorState == Idling) {
1871 break;
1872 }
1873 }
1874 }
1875
1876 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1877 "should be waiting");
1878
1879 switch (_collectorState) {
1880 case InitialMarking:
1881 {
1882 ReleaseForegroundGC x(this);
1883 stats().record_cms_begin();
1884 VM_CMS_Initial_Mark initial_mark_op(this);
1885 VMThread::execute(&initial_mark_op);
1886 }
1887 // The collector state may be any legal state at this point
1888 // since the background collector may have yielded to the
1889 // foreground collector.
1890 break;
1891 case Marking:
1892 // initial marking in checkpointRootsInitialWork has been completed
1893 if (markFromRoots()) { // we were successful
1894 assert(_collectorState == Precleaning, "Collector state should "
1895 "have changed");
1896 } else {
1897 assert(_foregroundGCIsActive, "Internal state inconsistency");
1898 }
1899 break;
1900 case Precleaning:
1901 // marking from roots in markFromRoots has been completed
1902 preclean();
1903 assert(_collectorState == AbortablePreclean ||
1904 _collectorState == FinalMarking,
1905 "Collector state should have changed");
1906 break;
1907 case AbortablePreclean:
1908 abortable_preclean();
1909 assert(_collectorState == FinalMarking, "Collector state should "
1910 "have changed");
1911 break;
1912 case FinalMarking:
1913 {
1914 ReleaseForegroundGC x(this);
1915
1916 VM_CMS_Final_Remark final_remark_op(this);
1917 VMThread::execute(&final_remark_op);
1918 }
1919 assert(_foregroundGCShouldWait, "block post-condition");
1920 break;
1921 case Sweeping:
1922 // final marking in checkpointRootsFinal has been completed
1923 sweep();
1924 assert(_collectorState == Resizing, "Collector state change "
1925 "to Resizing must be done under the free_list_lock");
1926
1927 case Resizing: {
1928 // Sweeping has been completed...
1929 // At this point the background collection has completed.
1930 // Don't move the call to compute_new_size() down
1931 // into code that might be executed if the background
1932 // collection was preempted.
1933 {
1934 ReleaseForegroundGC x(this); // unblock FG collection
1935 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1936 CMSTokenSync z(true); // not strictly needed.
1937 if (_collectorState == Resizing) {
1938 compute_new_size();
1939 save_heap_summary();
1940 _collectorState = Resetting;
1941 } else {
1942 assert(_collectorState == Idling, "The state should only change"
1943 " because the foreground collector has finished the collection");
1944 }
1945 }
1946 break;
1947 }
1948 case Resetting:
1949 // CMS heap resizing has been completed
1950 reset(true);
1951 assert(_collectorState == Idling, "Collector state should "
1952 "have changed");
1953
1954 MetaspaceGC::set_should_concurrent_collect(false);
1955
1956 stats().record_cms_end();
1957 // Don't move the concurrent_phases_end() and compute_new_size()
1958 // calls to here because a preempted background collection
1959 // has it's state set to "Resetting".
1960 break;
1961 case Idling:
1962 default:
1963 ShouldNotReachHere();
1964 break;
1965 }
1966 if (TraceCMSState) {
1967 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1968 p2i(Thread::current()), _collectorState);
1969 }
1970 assert(_foregroundGCShouldWait, "block post-condition");
1971 }
1972
1973 // Should this be in gc_epilogue?
1974 collector_policy()->counters()->update_counters();
1975
1976 {
1977 // Clear _foregroundGCShouldWait and, in the event that the
1978 // foreground collector is waiting, notify it, before
1979 // returning.
1980 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1981 _foregroundGCShouldWait = false;
1982 if (_foregroundGCIsActive) {
1983 CGC_lock->notify();
1984 }
1985 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1986 "Possible deadlock");
1987 }
1988 if (TraceCMSState) {
1989 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1990 " exiting collection CMS state %d",
1991 p2i(Thread::current()), _collectorState);
1992 }
1993 if (PrintGC && Verbose) {
1994 _cmsGen->print_heap_change(prev_used);
1995 }
1996 }
1997
1998 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1999 _cms_start_registered = true;
2000 _gc_timer_cm->register_gc_start();
2001 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2002 }
2003
2004 void CMSCollector::register_gc_end() {
2005 if (_cms_start_registered) {
2006 report_heap_summary(GCWhen::AfterGC);
2007
2008 _gc_timer_cm->register_gc_end();
2009 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2010 _cms_start_registered = false;
2011 }
2012 }
2013
2014 void CMSCollector::save_heap_summary() {
2015 GenCollectedHeap* gch = GenCollectedHeap::heap();
2016 _last_heap_summary = gch->create_heap_summary();
2017 _last_metaspace_summary = gch->create_metaspace_summary();
2018 }
2019
2020 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2021 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2022 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2023 }
2024
2025 bool CMSCollector::waitForForegroundGC() {
2026 bool res = false;
2027 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2028 "CMS thread should have CMS token");
2029 // Block the foreground collector until the
2030 // background collectors decides whether to
2031 // yield.
2032 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2033 _foregroundGCShouldWait = true;
2034 if (_foregroundGCIsActive) {
2035 // The background collector yields to the
2036 // foreground collector and returns a value
2037 // indicating that it has yielded. The foreground
2038 // collector can proceed.
2039 res = true;
2040 _foregroundGCShouldWait = false;
2041 ConcurrentMarkSweepThread::clear_CMS_flag(
2042 ConcurrentMarkSweepThread::CMS_cms_has_token);
2043 ConcurrentMarkSweepThread::set_CMS_flag(
2044 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2045 // Get a possibly blocked foreground thread going
2046 CGC_lock->notify();
2047 if (TraceCMSState) {
2048 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2049 p2i(Thread::current()), _collectorState);
2050 }
2051 while (_foregroundGCIsActive) {
2052 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2053 }
2054 ConcurrentMarkSweepThread::set_CMS_flag(
2055 ConcurrentMarkSweepThread::CMS_cms_has_token);
2056 ConcurrentMarkSweepThread::clear_CMS_flag(
2057 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2058 }
2059 if (TraceCMSState) {
2060 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2061 p2i(Thread::current()), _collectorState);
2062 }
2063 return res;
2064 }
2065
2066 // Because of the need to lock the free lists and other structures in
2067 // the collector, common to all the generations that the collector is
2068 // collecting, we need the gc_prologues of individual CMS generations
2069 // delegate to their collector. It may have been simpler had the
2070 // current infrastructure allowed one to call a prologue on a
2071 // collector. In the absence of that we have the generation's
2072 // prologue delegate to the collector, which delegates back
2073 // some "local" work to a worker method in the individual generations
2074 // that it's responsible for collecting, while itself doing any
2075 // work common to all generations it's responsible for. A similar
2076 // comment applies to the gc_epilogue()'s.
2077 // The role of the variable _between_prologue_and_epilogue is to
2078 // enforce the invocation protocol.
2079 void CMSCollector::gc_prologue(bool full) {
2080 // Call gc_prologue_work() for the CMSGen
2081 // we are responsible for.
2082
2083 // The following locking discipline assumes that we are only called
2084 // when the world is stopped.
2085 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2086
2087 // The CMSCollector prologue must call the gc_prologues for the
2088 // "generations" that it's responsible
2089 // for.
2090
2091 assert( Thread::current()->is_VM_thread()
2092 || ( CMSScavengeBeforeRemark
2093 && Thread::current()->is_ConcurrentGC_thread()),
2094 "Incorrect thread type for prologue execution");
2095
2096 if (_between_prologue_and_epilogue) {
2097 // We have already been invoked; this is a gc_prologue delegation
2098 // from yet another CMS generation that we are responsible for, just
2099 // ignore it since all relevant work has already been done.
2100 return;
2101 }
2102
2103 // set a bit saying prologue has been called; cleared in epilogue
2104 _between_prologue_and_epilogue = true;
2105 // Claim locks for common data structures, then call gc_prologue_work()
2106 // for each CMSGen.
2107
2108 getFreelistLocks(); // gets free list locks on constituent spaces
2109 bitMapLock()->lock_without_safepoint_check();
2110
2111 // Should call gc_prologue_work() for all cms gens we are responsible for
2112 bool duringMarking = _collectorState >= Marking
2113 && _collectorState < Sweeping;
2114
2115 // The young collections clear the modified oops state, which tells if
2116 // there are any modified oops in the class. The remark phase also needs
2117 // that information. Tell the young collection to save the union of all
2118 // modified klasses.
2119 if (duringMarking) {
2120 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2121 }
2122
2123 bool registerClosure = duringMarking;
2124
2125 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2126 &_modUnionClosurePar
2127 : &_modUnionClosure;
2128 _cmsGen->gc_prologue_work(full, registerClosure, muc);
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 gch->gen_process_roots(Generation::Old,
2431 true, // younger gens are roots
2432 true, // activate StrongRootsScope
2433 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2434 should_unload_classes(),
2435 ¬Older,
2436 NULL,
2437 NULL); // SSS: Provide correct closure
2438
2439 // Now mark from the roots
2440 MarkFromRootsClosure markFromRootsClosure(this, _span,
2441 verification_mark_bm(), verification_mark_stack(),
2442 false /* don't yield */, true /* verifying */);
2443 assert(_restart_addr == NULL, "Expected pre-condition");
2444 verification_mark_bm()->iterate(&markFromRootsClosure);
2445 while (_restart_addr != NULL) {
2446 // Deal with stack overflow: by restarting at the indicated
2447 // address.
2448 HeapWord* ra = _restart_addr;
2449 markFromRootsClosure.reset(ra);
2450 _restart_addr = NULL;
2451 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2452 }
2453 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2454 verify_work_stacks_empty();
2455
2456 // Marking completed -- now verify that each bit marked in
2457 // verification_mark_bm() is also marked in markBitMap(); flag all
2458 // errors by printing corresponding objects.
2459 VerifyMarkedClosure vcl(markBitMap());
2460 verification_mark_bm()->iterate(&vcl);
2461 if (vcl.failed()) {
2462 gclog_or_tty->print("Verification failed");
2463 gch->print_on(gclog_or_tty);
2464 fatal("CMS: failed marking verification after remark");
2465 }
2466 }
2467
2468 class VerifyKlassOopsKlassClosure : public KlassClosure {
2469 class VerifyKlassOopsClosure : public OopClosure {
2470 CMSBitMap* _bitmap;
2471 public:
2472 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2473 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2474 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2475 } _oop_closure;
2476 public:
2477 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2478 void do_klass(Klass* k) {
2479 k->oops_do(&_oop_closure);
2480 }
2481 };
2482
2483 void CMSCollector::verify_after_remark_work_2() {
2484 ResourceMark rm;
2485 HandleMark hm;
2486 GenCollectedHeap* gch = GenCollectedHeap::heap();
2487
2488 // Get a clear set of claim bits for the roots processing to work with.
2489 ClassLoaderDataGraph::clear_claimed_marks();
2490
2491 // Mark from roots one level into CMS
2492 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2493 markBitMap());
2494 CLDToOopClosure cld_closure(¬Older, true);
2495
2496 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2497
2498 gch->gen_process_roots(Generation::Old,
2499 true, // younger gens are roots
2500 true, // activate StrongRootsScope
2501 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2502 should_unload_classes(),
2503 ¬Older,
2504 NULL,
2505 &cld_closure);
2506
2507 // Now mark from the roots
2508 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2509 verification_mark_bm(), markBitMap(), verification_mark_stack());
2510 assert(_restart_addr == NULL, "Expected pre-condition");
2511 verification_mark_bm()->iterate(&markFromRootsClosure);
2512 while (_restart_addr != NULL) {
2513 // Deal with stack overflow: by restarting at the indicated
2514 // address.
2515 HeapWord* ra = _restart_addr;
2516 markFromRootsClosure.reset(ra);
2517 _restart_addr = NULL;
2518 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2519 }
2520 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2521 verify_work_stacks_empty();
2522
2523 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2524 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2525
2526 // Marking completed -- now verify that each bit marked in
2527 // verification_mark_bm() is also marked in markBitMap(); flag all
2528 // errors by printing corresponding objects.
2529 VerifyMarkedClosure vcl(markBitMap());
2530 verification_mark_bm()->iterate(&vcl);
2531 assert(!vcl.failed(), "Else verification above should not have succeeded");
2532 }
2533
2534 void ConcurrentMarkSweepGeneration::save_marks() {
2535 // delegate to CMS space
2536 cmsSpace()->save_marks();
2537 for (uint i = 0; i < ParallelGCThreads; i++) {
2538 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2539 }
2540 }
2541
2542 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2543 return cmsSpace()->no_allocs_since_save_marks();
2544 }
2545
2546 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2547 \
2548 void ConcurrentMarkSweepGeneration:: \
2549 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2550 cl->set_generation(this); \
2551 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2552 cl->reset_generation(); \
2553 save_marks(); \
2554 }
2555
2556 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2557
2558 void
2559 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2560 if (freelistLock()->owned_by_self()) {
2561 Generation::oop_iterate(cl);
2562 } else {
2563 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2564 Generation::oop_iterate(cl);
2565 }
2566 }
2567
2568 void
2569 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2570 if (freelistLock()->owned_by_self()) {
2571 Generation::object_iterate(cl);
2572 } else {
2573 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2574 Generation::object_iterate(cl);
2575 }
2576 }
2577
2578 void
2579 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2580 if (freelistLock()->owned_by_self()) {
2581 Generation::safe_object_iterate(cl);
2582 } else {
2583 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2584 Generation::safe_object_iterate(cl);
2585 }
2586 }
2587
2588 void
2589 ConcurrentMarkSweepGeneration::post_compact() {
2590 }
2591
2592 void
2593 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2594 // Fix the linear allocation blocks to look like free blocks.
2595
2596 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2597 // are not called when the heap is verified during universe initialization and
2598 // at vm shutdown.
2599 if (freelistLock()->owned_by_self()) {
2600 cmsSpace()->prepare_for_verify();
2601 } else {
2602 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2603 cmsSpace()->prepare_for_verify();
2604 }
2605 }
2606
2607 void
2608 ConcurrentMarkSweepGeneration::verify() {
2609 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2610 // are not called when the heap is verified during universe initialization and
2611 // at vm shutdown.
2612 if (freelistLock()->owned_by_self()) {
2613 cmsSpace()->verify();
2614 } else {
2615 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2616 cmsSpace()->verify();
2617 }
2618 }
2619
2620 void CMSCollector::verify() {
2621 _cmsGen->verify();
2622 }
2623
2624 #ifndef PRODUCT
2625 bool CMSCollector::overflow_list_is_empty() const {
2626 assert(_num_par_pushes >= 0, "Inconsistency");
2627 if (_overflow_list == NULL) {
2628 assert(_num_par_pushes == 0, "Inconsistency");
2629 }
2630 return _overflow_list == NULL;
2631 }
2632
2633 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2634 // merely consolidate assertion checks that appear to occur together frequently.
2635 void CMSCollector::verify_work_stacks_empty() const {
2636 assert(_markStack.isEmpty(), "Marking stack should be empty");
2637 assert(overflow_list_is_empty(), "Overflow list should be empty");
2638 }
2639
2640 void CMSCollector::verify_overflow_empty() const {
2641 assert(overflow_list_is_empty(), "Overflow list should be empty");
2642 assert(no_preserved_marks(), "No preserved marks");
2643 }
2644 #endif // PRODUCT
2645
2646 // Decide if we want to enable class unloading as part of the
2647 // ensuing concurrent GC cycle. We will collect and
2648 // unload classes if it's the case that:
2649 // (1) an explicit gc request has been made and the flag
2650 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2651 // (2) (a) class unloading is enabled at the command line, and
2652 // (b) old gen is getting really full
2653 // NOTE: Provided there is no change in the state of the heap between
2654 // calls to this method, it should have idempotent results. Moreover,
2655 // its results should be monotonically increasing (i.e. going from 0 to 1,
2656 // but not 1 to 0) between successive calls between which the heap was
2657 // not collected. For the implementation below, it must thus rely on
2658 // the property that concurrent_cycles_since_last_unload()
2659 // will not decrease unless a collection cycle happened and that
2660 // _cmsGen->is_too_full() are
2661 // themselves also monotonic in that sense. See check_monotonicity()
2662 // below.
2663 void CMSCollector::update_should_unload_classes() {
2664 _should_unload_classes = false;
2665 // Condition 1 above
2666 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2667 _should_unload_classes = true;
2668 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2669 // Disjuncts 2.b.(i,ii,iii) above
2670 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2671 CMSClassUnloadingMaxInterval)
2672 || _cmsGen->is_too_full();
2673 }
2674 }
2675
2676 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2677 bool res = should_concurrent_collect();
2678 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2679 return res;
2680 }
2681
2682 void CMSCollector::setup_cms_unloading_and_verification_state() {
2683 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2684 || VerifyBeforeExit;
2685 const int rso = GenCollectedHeap::SO_AllCodeCache;
2686
2687 // We set the proper root for this CMS cycle here.
2688 if (should_unload_classes()) { // Should unload classes this cycle
2689 remove_root_scanning_option(rso); // Shrink the root set appropriately
2690 set_verifying(should_verify); // Set verification state for this cycle
2691 return; // Nothing else needs to be done at this time
2692 }
2693
2694 // Not unloading classes this cycle
2695 assert(!should_unload_classes(), "Inconsistency!");
2696
2697 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2698 // Include symbols, strings and code cache elements to prevent their resurrection.
2699 add_root_scanning_option(rso);
2700 set_verifying(true);
2701 } else if (verifying() && !should_verify) {
2702 // We were verifying, but some verification flags got disabled.
2703 set_verifying(false);
2704 // Exclude symbols, strings and code cache elements from root scanning to
2705 // reduce IM and RM pauses.
2706 remove_root_scanning_option(rso);
2707 }
2708 }
2709
2710
2711 #ifndef PRODUCT
2712 HeapWord* CMSCollector::block_start(const void* p) const {
2713 const HeapWord* addr = (HeapWord*)p;
2714 if (_span.contains(p)) {
2715 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2716 return _cmsGen->cmsSpace()->block_start(p);
2717 }
2718 }
2719 return NULL;
2720 }
2721 #endif
2722
2723 HeapWord*
2724 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2725 bool tlab,
2726 bool parallel) {
2727 CMSSynchronousYieldRequest yr;
2728 assert(!tlab, "Can't deal with TLAB allocation");
2729 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2730 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2731 if (GCExpandToAllocateDelayMillis > 0) {
2732 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2733 }
2734 return have_lock_and_allocate(word_size, tlab);
2735 }
2736
2737 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2738 size_t bytes,
2739 size_t expand_bytes,
2740 CMSExpansionCause::Cause cause)
2741 {
2742
2743 bool success = expand(bytes, expand_bytes);
2744
2745 // remember why we expanded; this information is used
2746 // by shouldConcurrentCollect() when making decisions on whether to start
2747 // a new CMS cycle.
2748 if (success) {
2749 set_expansion_cause(cause);
2750 if (PrintGCDetails && Verbose) {
2751 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2752 CMSExpansionCause::to_string(cause));
2753 }
2754 }
2755 }
2756
2757 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2758 HeapWord* res = NULL;
2759 MutexLocker x(ParGCRareEvent_lock);
2760 while (true) {
2761 // Expansion by some other thread might make alloc OK now:
2762 res = ps->lab.alloc(word_sz);
2763 if (res != NULL) return res;
2764 // If there's not enough expansion space available, give up.
2765 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2766 return NULL;
2767 }
2768 // Otherwise, we try expansion.
2769 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2770 // Now go around the loop and try alloc again;
2771 // A competing par_promote might beat us to the expansion space,
2772 // so we may go around the loop again if promotion fails again.
2773 if (GCExpandToAllocateDelayMillis > 0) {
2774 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2775 }
2776 }
2777 }
2778
2779
2780 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2781 PromotionInfo* promo) {
2782 MutexLocker x(ParGCRareEvent_lock);
2783 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2784 while (true) {
2785 // Expansion by some other thread might make alloc OK now:
2786 if (promo->ensure_spooling_space()) {
2787 assert(promo->has_spooling_space(),
2788 "Post-condition of successful ensure_spooling_space()");
2789 return true;
2790 }
2791 // If there's not enough expansion space available, give up.
2792 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2793 return false;
2794 }
2795 // Otherwise, we try expansion.
2796 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2797 // Now go around the loop and try alloc again;
2798 // A competing allocation might beat us to the expansion space,
2799 // so we may go around the loop again if allocation fails again.
2800 if (GCExpandToAllocateDelayMillis > 0) {
2801 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2802 }
2803 }
2804 }
2805
2806 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2807 // Only shrink if a compaction was done so that all the free space
2808 // in the generation is in a contiguous block at the end.
2809 if (did_compact()) {
2810 CardGeneration::shrink(bytes);
2811 }
2812 }
2813
2814 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2815 assert_locked_or_safepoint(Heap_lock);
2816 }
2817
2818 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2819 assert_locked_or_safepoint(Heap_lock);
2820 assert_lock_strong(freelistLock());
2821 if (PrintGCDetails && Verbose) {
2822 warning("Shrinking of CMS not yet implemented");
2823 }
2824 return;
2825 }
2826
2827
2828 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2829 // phases.
2830 class CMSPhaseAccounting: public StackObj {
2831 public:
2832 CMSPhaseAccounting(CMSCollector *collector,
2833 const char *phase,
2834 const GCId gc_id,
2835 bool print_cr = true);
2836 ~CMSPhaseAccounting();
2837
2838 private:
2839 CMSCollector *_collector;
2840 const char *_phase;
2841 elapsedTimer _wallclock;
2842 bool _print_cr;
2843 const GCId _gc_id;
2844
2845 public:
2846 // Not MT-safe; so do not pass around these StackObj's
2847 // where they may be accessed by other threads.
2848 jlong wallclock_millis() {
2849 assert(_wallclock.is_active(), "Wall clock should not stop");
2850 _wallclock.stop(); // to record time
2851 jlong ret = _wallclock.milliseconds();
2852 _wallclock.start(); // restart
2853 return ret;
2854 }
2855 };
2856
2857 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2858 const char *phase,
2859 const GCId gc_id,
2860 bool print_cr) :
2861 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2862
2863 if (PrintCMSStatistics != 0) {
2864 _collector->resetYields();
2865 }
2866 if (PrintGCDetails) {
2867 gclog_or_tty->gclog_stamp(_gc_id);
2868 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2869 _collector->cmsGen()->short_name(), _phase);
2870 }
2871 _collector->resetTimer();
2872 _wallclock.start();
2873 _collector->startTimer();
2874 }
2875
2876 CMSPhaseAccounting::~CMSPhaseAccounting() {
2877 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2878 _collector->stopTimer();
2879 _wallclock.stop();
2880 if (PrintGCDetails) {
2881 gclog_or_tty->gclog_stamp(_gc_id);
2882 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2883 _collector->cmsGen()->short_name(),
2884 _phase, _collector->timerValue(), _wallclock.seconds());
2885 if (_print_cr) {
2886 gclog_or_tty->cr();
2887 }
2888 if (PrintCMSStatistics != 0) {
2889 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2890 _collector->yields());
2891 }
2892 }
2893 }
2894
2895 // CMS work
2896
2897 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2898 class CMSParMarkTask : public AbstractGangTask {
2899 protected:
2900 CMSCollector* _collector;
2901 int _n_workers;
2902 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
2903 AbstractGangTask(name),
2904 _collector(collector),
2905 _n_workers(n_workers) {}
2906 // Work method in support of parallel rescan ... of young gen spaces
2907 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2908 ContiguousSpace* space,
2909 HeapWord** chunk_array, size_t chunk_top);
2910 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2911 };
2912
2913 // Parallel initial mark task
2914 class CMSParInitialMarkTask: public CMSParMarkTask {
2915 public:
2916 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
2917 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
2918 collector, n_workers) {}
2919 void work(uint worker_id);
2920 };
2921
2922 // Checkpoint the roots into this generation from outside
2923 // this generation. [Note this initial checkpoint need only
2924 // be approximate -- we'll do a catch up phase subsequently.]
2925 void CMSCollector::checkpointRootsInitial() {
2926 assert(_collectorState == InitialMarking, "Wrong collector state");
2927 check_correct_thread_executing();
2928 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2929
2930 save_heap_summary();
2931 report_heap_summary(GCWhen::BeforeGC);
2932
2933 ReferenceProcessor* rp = ref_processor();
2934 assert(_restart_addr == NULL, "Control point invariant");
2935 {
2936 // acquire locks for subsequent manipulations
2937 MutexLockerEx x(bitMapLock(),
2938 Mutex::_no_safepoint_check_flag);
2939 checkpointRootsInitialWork();
2940 // enable ("weak") refs discovery
2941 rp->enable_discovery();
2942 _collectorState = Marking;
2943 }
2944 }
2945
2946 void CMSCollector::checkpointRootsInitialWork() {
2947 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2948 assert(_collectorState == InitialMarking, "just checking");
2949
2950 // If there has not been a GC[n-1] since last GC[n] cycle completed,
2951 // precede our marking with a collection of all
2952 // younger generations to keep floating garbage to a minimum.
2953 // XXX: we won't do this for now -- it's an optimization to be done later.
2954
2955 // already have locks
2956 assert_lock_strong(bitMapLock());
2957 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2958
2959 // Setup the verification and class unloading state for this
2960 // CMS collection cycle.
2961 setup_cms_unloading_and_verification_state();
2962
2963 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2964 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
2965
2966 // Reset all the PLAB chunk arrays if necessary.
2967 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2968 reset_survivor_plab_arrays();
2969 }
2970
2971 ResourceMark rm;
2972 HandleMark hm;
2973
2974 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2975 GenCollectedHeap* gch = GenCollectedHeap::heap();
2976
2977 verify_work_stacks_empty();
2978 verify_overflow_empty();
2979
2980 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2981 // Update the saved marks which may affect the root scans.
2982 gch->save_marks();
2983
2984 // weak reference processing has not started yet.
2985 ref_processor()->set_enqueuing_is_done(false);
2986
2987 // Need to remember all newly created CLDs,
2988 // so that we can guarantee that the remark finds them.
2989 ClassLoaderDataGraph::remember_new_clds(true);
2990
2991 // Whenever a CLD is found, it will be claimed before proceeding to mark
2992 // the klasses. The claimed marks need to be cleared before marking starts.
2993 ClassLoaderDataGraph::clear_claimed_marks();
2994
2995 if (CMSPrintEdenSurvivorChunks) {
2996 print_eden_and_survivor_chunk_arrays();
2997 }
2998
2999 {
3000 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3001 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
3002 // The parallel version.
3003 FlexibleWorkGang* workers = gch->workers();
3004 assert(workers != NULL, "Need parallel worker threads.");
3005 int n_workers = workers->active_workers();
3006 CMSParInitialMarkTask tsk(this, n_workers);
3007 gch->set_par_threads(n_workers);
3008 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3009 if (n_workers > 1) {
3010 StrongRootsScope srs;
3011 workers->run_task(&tsk);
3012 } else {
3013 StrongRootsScope srs;
3014 tsk.work(0);
3015 }
3016 gch->set_par_threads(0);
3017 } else {
3018 // The serial version.
3019 CLDToOopClosure cld_closure(¬Older, true);
3020 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3021 gch->gen_process_roots(Generation::Old,
3022 true, // younger gens are roots
3023 true, // activate StrongRootsScope
3024 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3025 should_unload_classes(),
3026 ¬Older,
3027 NULL,
3028 &cld_closure);
3029 }
3030 }
3031
3032 // Clear mod-union table; it will be dirtied in the prologue of
3033 // CMS generation per each younger generation collection.
3034
3035 assert(_modUnionTable.isAllClear(),
3036 "Was cleared in most recent final checkpoint phase"
3037 " or no bits are set in the gc_prologue before the start of the next "
3038 "subsequent marking phase.");
3039
3040 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3041
3042 // Save the end of the used_region of the constituent generations
3043 // to be used to limit the extent of sweep in each generation.
3044 save_sweep_limits();
3045 verify_overflow_empty();
3046 }
3047
3048 bool CMSCollector::markFromRoots() {
3049 // we might be tempted to assert that:
3050 // assert(!SafepointSynchronize::is_at_safepoint(),
3051 // "inconsistent argument?");
3052 // However that wouldn't be right, because it's possible that
3053 // a safepoint is indeed in progress as a younger generation
3054 // stop-the-world GC happens even as we mark in this generation.
3055 assert(_collectorState == Marking, "inconsistent state?");
3056 check_correct_thread_executing();
3057 verify_overflow_empty();
3058
3059 // Weak ref discovery note: We may be discovering weak
3060 // refs in this generation concurrent (but interleaved) with
3061 // weak ref discovery by a younger generation collector.
3062
3063 CMSTokenSyncWithLocks ts(true, bitMapLock());
3064 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3065 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3066 bool res = markFromRootsWork();
3067 if (res) {
3068 _collectorState = Precleaning;
3069 } else { // We failed and a foreground collection wants to take over
3070 assert(_foregroundGCIsActive, "internal state inconsistency");
3071 assert(_restart_addr == NULL, "foreground will restart from scratch");
3072 if (PrintGCDetails) {
3073 gclog_or_tty->print_cr("bailing out to foreground collection");
3074 }
3075 }
3076 verify_overflow_empty();
3077 return res;
3078 }
3079
3080 bool CMSCollector::markFromRootsWork() {
3081 // iterate over marked bits in bit map, doing a full scan and mark
3082 // from these roots using the following algorithm:
3083 // . if oop is to the right of the current scan pointer,
3084 // mark corresponding bit (we'll process it later)
3085 // . else (oop is to left of current scan pointer)
3086 // push oop on marking stack
3087 // . drain the marking stack
3088
3089 // Note that when we do a marking step we need to hold the
3090 // bit map lock -- recall that direct allocation (by mutators)
3091 // and promotion (by younger generation collectors) is also
3092 // marking the bit map. [the so-called allocate live policy.]
3093 // Because the implementation of bit map marking is not
3094 // robust wrt simultaneous marking of bits in the same word,
3095 // we need to make sure that there is no such interference
3096 // between concurrent such updates.
3097
3098 // already have locks
3099 assert_lock_strong(bitMapLock());
3100
3101 verify_work_stacks_empty();
3102 verify_overflow_empty();
3103 bool result = false;
3104 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3105 result = do_marking_mt();
3106 } else {
3107 result = do_marking_st();
3108 }
3109 return result;
3110 }
3111
3112 // Forward decl
3113 class CMSConcMarkingTask;
3114
3115 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3116 CMSCollector* _collector;
3117 CMSConcMarkingTask* _task;
3118 public:
3119 virtual void yield();
3120
3121 // "n_threads" is the number of threads to be terminated.
3122 // "queue_set" is a set of work queues of other threads.
3123 // "collector" is the CMS collector associated with this task terminator.
3124 // "yield" indicates whether we need the gang as a whole to yield.
3125 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3126 ParallelTaskTerminator(n_threads, queue_set),
3127 _collector(collector) { }
3128
3129 void set_task(CMSConcMarkingTask* task) {
3130 _task = task;
3131 }
3132 };
3133
3134 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3135 CMSConcMarkingTask* _task;
3136 public:
3137 bool should_exit_termination();
3138 void set_task(CMSConcMarkingTask* task) {
3139 _task = task;
3140 }
3141 };
3142
3143 // MT Concurrent Marking Task
3144 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3145 CMSCollector* _collector;
3146 int _n_workers; // requested/desired # workers
3147 bool _result;
3148 CompactibleFreeListSpace* _cms_space;
3149 char _pad_front[64]; // padding to ...
3150 HeapWord* _global_finger; // ... avoid sharing cache line
3151 char _pad_back[64];
3152 HeapWord* _restart_addr;
3153
3154 // Exposed here for yielding support
3155 Mutex* const _bit_map_lock;
3156
3157 // The per thread work queues, available here for stealing
3158 OopTaskQueueSet* _task_queues;
3159
3160 // Termination (and yielding) support
3161 CMSConcMarkingTerminator _term;
3162 CMSConcMarkingTerminatorTerminator _term_term;
3163
3164 public:
3165 CMSConcMarkingTask(CMSCollector* collector,
3166 CompactibleFreeListSpace* cms_space,
3167 YieldingFlexibleWorkGang* workers,
3168 OopTaskQueueSet* task_queues):
3169 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3170 _collector(collector),
3171 _cms_space(cms_space),
3172 _n_workers(0), _result(true),
3173 _task_queues(task_queues),
3174 _term(_n_workers, task_queues, _collector),
3175 _bit_map_lock(collector->bitMapLock())
3176 {
3177 _requested_size = _n_workers;
3178 _term.set_task(this);
3179 _term_term.set_task(this);
3180 _restart_addr = _global_finger = _cms_space->bottom();
3181 }
3182
3183
3184 OopTaskQueueSet* task_queues() { return _task_queues; }
3185
3186 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3187
3188 HeapWord** global_finger_addr() { return &_global_finger; }
3189
3190 CMSConcMarkingTerminator* terminator() { return &_term; }
3191
3192 virtual void set_for_termination(int active_workers) {
3193 terminator()->reset_for_reuse(active_workers);
3194 }
3195
3196 void work(uint worker_id);
3197 bool should_yield() {
3198 return ConcurrentMarkSweepThread::should_yield()
3199 && !_collector->foregroundGCIsActive();
3200 }
3201
3202 virtual void coordinator_yield(); // stuff done by coordinator
3203 bool result() { return _result; }
3204
3205 void reset(HeapWord* ra) {
3206 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3207 _restart_addr = _global_finger = ra;
3208 _term.reset_for_reuse();
3209 }
3210
3211 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3212 OopTaskQueue* work_q);
3213
3214 private:
3215 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3216 void do_work_steal(int i);
3217 void bump_global_finger(HeapWord* f);
3218 };
3219
3220 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3221 assert(_task != NULL, "Error");
3222 return _task->yielding();
3223 // Note that we do not need the disjunct || _task->should_yield() above
3224 // because we want terminating threads to yield only if the task
3225 // is already in the midst of yielding, which happens only after at least one
3226 // thread has yielded.
3227 }
3228
3229 void CMSConcMarkingTerminator::yield() {
3230 if (_task->should_yield()) {
3231 _task->yield();
3232 } else {
3233 ParallelTaskTerminator::yield();
3234 }
3235 }
3236
3237 ////////////////////////////////////////////////////////////////
3238 // Concurrent Marking Algorithm Sketch
3239 ////////////////////////////////////////////////////////////////
3240 // Until all tasks exhausted (both spaces):
3241 // -- claim next available chunk
3242 // -- bump global finger via CAS
3243 // -- find first object that starts in this chunk
3244 // and start scanning bitmap from that position
3245 // -- scan marked objects for oops
3246 // -- CAS-mark target, and if successful:
3247 // . if target oop is above global finger (volatile read)
3248 // nothing to do
3249 // . if target oop is in chunk and above local finger
3250 // then nothing to do
3251 // . else push on work-queue
3252 // -- Deal with possible overflow issues:
3253 // . local work-queue overflow causes stuff to be pushed on
3254 // global (common) overflow queue
3255 // . always first empty local work queue
3256 // . then get a batch of oops from global work queue if any
3257 // . then do work stealing
3258 // -- When all tasks claimed (both spaces)
3259 // and local work queue empty,
3260 // then in a loop do:
3261 // . check global overflow stack; steal a batch of oops and trace
3262 // . try to steal from other threads oif GOS is empty
3263 // . if neither is available, offer termination
3264 // -- Terminate and return result
3265 //
3266 void CMSConcMarkingTask::work(uint worker_id) {
3267 elapsedTimer _timer;
3268 ResourceMark rm;
3269 HandleMark hm;
3270
3271 DEBUG_ONLY(_collector->verify_overflow_empty();)
3272
3273 // Before we begin work, our work queue should be empty
3274 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3275 // Scan the bitmap covering _cms_space, tracing through grey objects.
3276 _timer.start();
3277 do_scan_and_mark(worker_id, _cms_space);
3278 _timer.stop();
3279 if (PrintCMSStatistics != 0) {
3280 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3281 worker_id, _timer.seconds());
3282 // XXX: need xxx/xxx type of notation, two timers
3283 }
3284
3285 // ... do work stealing
3286 _timer.reset();
3287 _timer.start();
3288 do_work_steal(worker_id);
3289 _timer.stop();
3290 if (PrintCMSStatistics != 0) {
3291 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3292 worker_id, _timer.seconds());
3293 // XXX: need xxx/xxx type of notation, two timers
3294 }
3295 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3296 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3297 // Note that under the current task protocol, the
3298 // following assertion is true even of the spaces
3299 // expanded since the completion of the concurrent
3300 // marking. XXX This will likely change under a strict
3301 // ABORT semantics.
3302 // After perm removal the comparison was changed to
3303 // greater than or equal to from strictly greater than.
3304 // Before perm removal the highest address sweep would
3305 // have been at the end of perm gen but now is at the
3306 // end of the tenured gen.
3307 assert(_global_finger >= _cms_space->end(),
3308 "All tasks have been completed");
3309 DEBUG_ONLY(_collector->verify_overflow_empty();)
3310 }
3311
3312 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3313 HeapWord* read = _global_finger;
3314 HeapWord* cur = read;
3315 while (f > read) {
3316 cur = read;
3317 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3318 if (cur == read) {
3319 // our cas succeeded
3320 assert(_global_finger >= f, "protocol consistency");
3321 break;
3322 }
3323 }
3324 }
3325
3326 // This is really inefficient, and should be redone by
3327 // using (not yet available) block-read and -write interfaces to the
3328 // stack and the work_queue. XXX FIX ME !!!
3329 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3330 OopTaskQueue* work_q) {
3331 // Fast lock-free check
3332 if (ovflw_stk->length() == 0) {
3333 return false;
3334 }
3335 assert(work_q->size() == 0, "Shouldn't steal");
3336 MutexLockerEx ml(ovflw_stk->par_lock(),
3337 Mutex::_no_safepoint_check_flag);
3338 // Grab up to 1/4 the size of the work queue
3339 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3340 (size_t)ParGCDesiredObjsFromOverflowList);
3341 num = MIN2(num, ovflw_stk->length());
3342 for (int i = (int) num; i > 0; i--) {
3343 oop cur = ovflw_stk->pop();
3344 assert(cur != NULL, "Counted wrong?");
3345 work_q->push(cur);
3346 }
3347 return num > 0;
3348 }
3349
3350 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3351 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3352 int n_tasks = pst->n_tasks();
3353 // We allow that there may be no tasks to do here because
3354 // we are restarting after a stack overflow.
3355 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3356 uint nth_task = 0;
3357
3358 HeapWord* aligned_start = sp->bottom();
3359 if (sp->used_region().contains(_restart_addr)) {
3360 // Align down to a card boundary for the start of 0th task
3361 // for this space.
3362 aligned_start =
3363 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3364 CardTableModRefBS::card_size);
3365 }
3366
3367 size_t chunk_size = sp->marking_task_size();
3368 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3369 // Having claimed the nth task in this space,
3370 // compute the chunk that it corresponds to:
3371 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3372 aligned_start + (nth_task+1)*chunk_size);
3373 // Try and bump the global finger via a CAS;
3374 // note that we need to do the global finger bump
3375 // _before_ taking the intersection below, because
3376 // the task corresponding to that region will be
3377 // deemed done even if the used_region() expands
3378 // because of allocation -- as it almost certainly will
3379 // during start-up while the threads yield in the
3380 // closure below.
3381 HeapWord* finger = span.end();
3382 bump_global_finger(finger); // atomically
3383 // There are null tasks here corresponding to chunks
3384 // beyond the "top" address of the space.
3385 span = span.intersection(sp->used_region());
3386 if (!span.is_empty()) { // Non-null task
3387 HeapWord* prev_obj;
3388 assert(!span.contains(_restart_addr) || nth_task == 0,
3389 "Inconsistency");
3390 if (nth_task == 0) {
3391 // For the 0th task, we'll not need to compute a block_start.
3392 if (span.contains(_restart_addr)) {
3393 // In the case of a restart because of stack overflow,
3394 // we might additionally skip a chunk prefix.
3395 prev_obj = _restart_addr;
3396 } else {
3397 prev_obj = span.start();
3398 }
3399 } else {
3400 // We want to skip the first object because
3401 // the protocol is to scan any object in its entirety
3402 // that _starts_ in this span; a fortiori, any
3403 // object starting in an earlier span is scanned
3404 // as part of an earlier claimed task.
3405 // Below we use the "careful" version of block_start
3406 // so we do not try to navigate uninitialized objects.
3407 prev_obj = sp->block_start_careful(span.start());
3408 // Below we use a variant of block_size that uses the
3409 // Printezis bits to avoid waiting for allocated
3410 // objects to become initialized/parsable.
3411 while (prev_obj < span.start()) {
3412 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3413 if (sz > 0) {
3414 prev_obj += sz;
3415 } else {
3416 // In this case we may end up doing a bit of redundant
3417 // scanning, but that appears unavoidable, short of
3418 // locking the free list locks; see bug 6324141.
3419 break;
3420 }
3421 }
3422 }
3423 if (prev_obj < span.end()) {
3424 MemRegion my_span = MemRegion(prev_obj, span.end());
3425 // Do the marking work within a non-empty span --
3426 // the last argument to the constructor indicates whether the
3427 // iteration should be incremental with periodic yields.
3428 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3429 &_collector->_markBitMap,
3430 work_queue(i),
3431 &_collector->_markStack);
3432 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3433 } // else nothing to do for this task
3434 } // else nothing to do for this task
3435 }
3436 // We'd be tempted to assert here that since there are no
3437 // more tasks left to claim in this space, the global_finger
3438 // must exceed space->top() and a fortiori space->end(). However,
3439 // that would not quite be correct because the bumping of
3440 // global_finger occurs strictly after the claiming of a task,
3441 // so by the time we reach here the global finger may not yet
3442 // have been bumped up by the thread that claimed the last
3443 // task.
3444 pst->all_tasks_completed();
3445 }
3446
3447 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3448 private:
3449 CMSCollector* _collector;
3450 CMSConcMarkingTask* _task;
3451 MemRegion _span;
3452 CMSBitMap* _bit_map;
3453 CMSMarkStack* _overflow_stack;
3454 OopTaskQueue* _work_queue;
3455 protected:
3456 DO_OOP_WORK_DEFN
3457 public:
3458 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3459 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3460 MetadataAwareOopClosure(collector->ref_processor()),
3461 _collector(collector),
3462 _task(task),
3463 _span(collector->_span),
3464 _work_queue(work_queue),
3465 _bit_map(bit_map),
3466 _overflow_stack(overflow_stack)
3467 { }
3468 virtual void do_oop(oop* p);
3469 virtual void do_oop(narrowOop* p);
3470
3471 void trim_queue(size_t max);
3472 void handle_stack_overflow(HeapWord* lost);
3473 void do_yield_check() {
3474 if (_task->should_yield()) {
3475 _task->yield();
3476 }
3477 }
3478 };
3479
3480 // Grey object scanning during work stealing phase --
3481 // the salient assumption here is that any references
3482 // that are in these stolen objects being scanned must
3483 // already have been initialized (else they would not have
3484 // been published), so we do not need to check for
3485 // uninitialized objects before pushing here.
3486 void Par_ConcMarkingClosure::do_oop(oop obj) {
3487 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3488 HeapWord* addr = (HeapWord*)obj;
3489 // Check if oop points into the CMS generation
3490 // and is not marked
3491 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3492 // a white object ...
3493 // If we manage to "claim" the object, by being the
3494 // first thread to mark it, then we push it on our
3495 // marking stack
3496 if (_bit_map->par_mark(addr)) { // ... now grey
3497 // push on work queue (grey set)
3498 bool simulate_overflow = false;
3499 NOT_PRODUCT(
3500 if (CMSMarkStackOverflowALot &&
3501 _collector->simulate_overflow()) {
3502 // simulate a stack overflow
3503 simulate_overflow = true;
3504 }
3505 )
3506 if (simulate_overflow ||
3507 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3508 // stack overflow
3509 if (PrintCMSStatistics != 0) {
3510 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3511 SIZE_FORMAT, _overflow_stack->capacity());
3512 }
3513 // We cannot assert that the overflow stack is full because
3514 // it may have been emptied since.
3515 assert(simulate_overflow ||
3516 _work_queue->size() == _work_queue->max_elems(),
3517 "Else push should have succeeded");
3518 handle_stack_overflow(addr);
3519 }
3520 } // Else, some other thread got there first
3521 do_yield_check();
3522 }
3523 }
3524
3525 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3526 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3527
3528 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3529 while (_work_queue->size() > max) {
3530 oop new_oop;
3531 if (_work_queue->pop_local(new_oop)) {
3532 assert(new_oop->is_oop(), "Should be an oop");
3533 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3534 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3535 new_oop->oop_iterate(this); // do_oop() above
3536 do_yield_check();
3537 }
3538 }
3539 }
3540
3541 // Upon stack overflow, we discard (part of) the stack,
3542 // remembering the least address amongst those discarded
3543 // in CMSCollector's _restart_address.
3544 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3545 // We need to do this under a mutex to prevent other
3546 // workers from interfering with the work done below.
3547 MutexLockerEx ml(_overflow_stack->par_lock(),
3548 Mutex::_no_safepoint_check_flag);
3549 // Remember the least grey address discarded
3550 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3551 _collector->lower_restart_addr(ra);
3552 _overflow_stack->reset(); // discard stack contents
3553 _overflow_stack->expand(); // expand the stack if possible
3554 }
3555
3556
3557 void CMSConcMarkingTask::do_work_steal(int i) {
3558 OopTaskQueue* work_q = work_queue(i);
3559 oop obj_to_scan;
3560 CMSBitMap* bm = &(_collector->_markBitMap);
3561 CMSMarkStack* ovflw = &(_collector->_markStack);
3562 int* seed = _collector->hash_seed(i);
3563 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3564 while (true) {
3565 cl.trim_queue(0);
3566 assert(work_q->size() == 0, "Should have been emptied above");
3567 if (get_work_from_overflow_stack(ovflw, work_q)) {
3568 // Can't assert below because the work obtained from the
3569 // overflow stack may already have been stolen from us.
3570 // assert(work_q->size() > 0, "Work from overflow stack");
3571 continue;
3572 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3573 assert(obj_to_scan->is_oop(), "Should be an oop");
3574 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3575 obj_to_scan->oop_iterate(&cl);
3576 } else if (terminator()->offer_termination(&_term_term)) {
3577 assert(work_q->size() == 0, "Impossible!");
3578 break;
3579 } else if (yielding() || should_yield()) {
3580 yield();
3581 }
3582 }
3583 }
3584
3585 // This is run by the CMS (coordinator) thread.
3586 void CMSConcMarkingTask::coordinator_yield() {
3587 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3588 "CMS thread should hold CMS token");
3589 // First give up the locks, then yield, then re-lock
3590 // We should probably use a constructor/destructor idiom to
3591 // do this unlock/lock or modify the MutexUnlocker class to
3592 // serve our purpose. XXX
3593 assert_lock_strong(_bit_map_lock);
3594 _bit_map_lock->unlock();
3595 ConcurrentMarkSweepThread::desynchronize(true);
3596 _collector->stopTimer();
3597 if (PrintCMSStatistics != 0) {
3598 _collector->incrementYields();
3599 }
3600
3601 // It is possible for whichever thread initiated the yield request
3602 // not to get a chance to wake up and take the bitmap lock between
3603 // this thread releasing it and reacquiring it. So, while the
3604 // should_yield() flag is on, let's sleep for a bit to give the
3605 // other thread a chance to wake up. The limit imposed on the number
3606 // of iterations is defensive, to avoid any unforseen circumstances
3607 // putting us into an infinite loop. Since it's always been this
3608 // (coordinator_yield()) method that was observed to cause the
3609 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3610 // which is by default non-zero. For the other seven methods that
3611 // also perform the yield operation, as are using a different
3612 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3613 // can enable the sleeping for those methods too, if necessary.
3614 // See 6442774.
3615 //
3616 // We really need to reconsider the synchronization between the GC
3617 // thread and the yield-requesting threads in the future and we
3618 // should really use wait/notify, which is the recommended
3619 // way of doing this type of interaction. Additionally, we should
3620 // consolidate the eight methods that do the yield operation and they
3621 // are almost identical into one for better maintainability and
3622 // readability. See 6445193.
3623 //
3624 // Tony 2006.06.29
3625 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3626 ConcurrentMarkSweepThread::should_yield() &&
3627 !CMSCollector::foregroundGCIsActive(); ++i) {
3628 os::sleep(Thread::current(), 1, false);
3629 }
3630
3631 ConcurrentMarkSweepThread::synchronize(true);
3632 _bit_map_lock->lock_without_safepoint_check();
3633 _collector->startTimer();
3634 }
3635
3636 bool CMSCollector::do_marking_mt() {
3637 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3638 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
3639 conc_workers()->total_workers(),
3640 conc_workers()->active_workers(),
3641 Threads::number_of_non_daemon_threads());
3642 conc_workers()->set_active_workers(num_workers);
3643
3644 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3645
3646 CMSConcMarkingTask tsk(this,
3647 cms_space,
3648 conc_workers(),
3649 task_queues());
3650
3651 // Since the actual number of workers we get may be different
3652 // from the number we requested above, do we need to do anything different
3653 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3654 // class?? XXX
3655 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3656
3657 // Refs discovery is already non-atomic.
3658 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3659 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3660 conc_workers()->start_task(&tsk);
3661 while (tsk.yielded()) {
3662 tsk.coordinator_yield();
3663 conc_workers()->continue_task(&tsk);
3664 }
3665 // If the task was aborted, _restart_addr will be non-NULL
3666 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3667 while (_restart_addr != NULL) {
3668 // XXX For now we do not make use of ABORTED state and have not
3669 // yet implemented the right abort semantics (even in the original
3670 // single-threaded CMS case). That needs some more investigation
3671 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3672 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3673 // If _restart_addr is non-NULL, a marking stack overflow
3674 // occurred; we need to do a fresh marking iteration from the
3675 // indicated restart address.
3676 if (_foregroundGCIsActive) {
3677 // We may be running into repeated stack overflows, having
3678 // reached the limit of the stack size, while making very
3679 // slow forward progress. It may be best to bail out and
3680 // let the foreground collector do its job.
3681 // Clear _restart_addr, so that foreground GC
3682 // works from scratch. This avoids the headache of
3683 // a "rescan" which would otherwise be needed because
3684 // of the dirty mod union table & card table.
3685 _restart_addr = NULL;
3686 return false;
3687 }
3688 // Adjust the task to restart from _restart_addr
3689 tsk.reset(_restart_addr);
3690 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3691 _restart_addr);
3692 _restart_addr = NULL;
3693 // Get the workers going again
3694 conc_workers()->start_task(&tsk);
3695 while (tsk.yielded()) {
3696 tsk.coordinator_yield();
3697 conc_workers()->continue_task(&tsk);
3698 }
3699 }
3700 assert(tsk.completed(), "Inconsistency");
3701 assert(tsk.result() == true, "Inconsistency");
3702 return true;
3703 }
3704
3705 bool CMSCollector::do_marking_st() {
3706 ResourceMark rm;
3707 HandleMark hm;
3708
3709 // Temporarily make refs discovery single threaded (non-MT)
3710 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3711 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3712 &_markStack, CMSYield);
3713 // the last argument to iterate indicates whether the iteration
3714 // should be incremental with periodic yields.
3715 _markBitMap.iterate(&markFromRootsClosure);
3716 // If _restart_addr is non-NULL, a marking stack overflow
3717 // occurred; we need to do a fresh iteration from the
3718 // indicated restart address.
3719 while (_restart_addr != NULL) {
3720 if (_foregroundGCIsActive) {
3721 // We may be running into repeated stack overflows, having
3722 // reached the limit of the stack size, while making very
3723 // slow forward progress. It may be best to bail out and
3724 // let the foreground collector do its job.
3725 // Clear _restart_addr, so that foreground GC
3726 // works from scratch. This avoids the headache of
3727 // a "rescan" which would otherwise be needed because
3728 // of the dirty mod union table & card table.
3729 _restart_addr = NULL;
3730 return false; // indicating failure to complete marking
3731 }
3732 // Deal with stack overflow:
3733 // we restart marking from _restart_addr
3734 HeapWord* ra = _restart_addr;
3735 markFromRootsClosure.reset(ra);
3736 _restart_addr = NULL;
3737 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3738 }
3739 return true;
3740 }
3741
3742 void CMSCollector::preclean() {
3743 check_correct_thread_executing();
3744 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3745 verify_work_stacks_empty();
3746 verify_overflow_empty();
3747 _abort_preclean = false;
3748 if (CMSPrecleaningEnabled) {
3749 if (!CMSEdenChunksRecordAlways) {
3750 _eden_chunk_index = 0;
3751 }
3752 size_t used = get_eden_used();
3753 size_t capacity = get_eden_capacity();
3754 // Don't start sampling unless we will get sufficiently
3755 // many samples.
3756 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3757 * CMSScheduleRemarkEdenPenetration)) {
3758 _start_sampling = true;
3759 } else {
3760 _start_sampling = false;
3761 }
3762 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3763 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3764 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3765 }
3766 CMSTokenSync x(true); // is cms thread
3767 if (CMSPrecleaningEnabled) {
3768 sample_eden();
3769 _collectorState = AbortablePreclean;
3770 } else {
3771 _collectorState = FinalMarking;
3772 }
3773 verify_work_stacks_empty();
3774 verify_overflow_empty();
3775 }
3776
3777 // Try and schedule the remark such that young gen
3778 // occupancy is CMSScheduleRemarkEdenPenetration %.
3779 void CMSCollector::abortable_preclean() {
3780 check_correct_thread_executing();
3781 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3782 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3783
3784 // If Eden's current occupancy is below this threshold,
3785 // immediately schedule the remark; else preclean
3786 // past the next scavenge in an effort to
3787 // schedule the pause as described above. By choosing
3788 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3789 // we will never do an actual abortable preclean cycle.
3790 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3791 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3792 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3793 // We need more smarts in the abortable preclean
3794 // loop below to deal with cases where allocation
3795 // in young gen is very very slow, and our precleaning
3796 // is running a losing race against a horde of
3797 // mutators intent on flooding us with CMS updates
3798 // (dirty cards).
3799 // One, admittedly dumb, strategy is to give up
3800 // after a certain number of abortable precleaning loops
3801 // or after a certain maximum time. We want to make
3802 // this smarter in the next iteration.
3803 // XXX FIX ME!!! YSR
3804 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3805 while (!(should_abort_preclean() ||
3806 ConcurrentMarkSweepThread::should_terminate())) {
3807 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3808 cumworkdone += workdone;
3809 loops++;
3810 // Voluntarily terminate abortable preclean phase if we have
3811 // been at it for too long.
3812 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3813 loops >= CMSMaxAbortablePrecleanLoops) {
3814 if (PrintGCDetails) {
3815 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3816 }
3817 break;
3818 }
3819 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3820 if (PrintGCDetails) {
3821 gclog_or_tty->print(" CMS: abort preclean due to time ");
3822 }
3823 break;
3824 }
3825 // If we are doing little work each iteration, we should
3826 // take a short break.
3827 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3828 // Sleep for some time, waiting for work to accumulate
3829 stopTimer();
3830 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3831 startTimer();
3832 waited++;
3833 }
3834 }
3835 if (PrintCMSStatistics > 0) {
3836 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3837 loops, waited, cumworkdone);
3838 }
3839 }
3840 CMSTokenSync x(true); // is cms thread
3841 if (_collectorState != Idling) {
3842 assert(_collectorState == AbortablePreclean,
3843 "Spontaneous state transition?");
3844 _collectorState = FinalMarking;
3845 } // Else, a foreground collection completed this CMS cycle.
3846 return;
3847 }
3848
3849 // Respond to an Eden sampling opportunity
3850 void CMSCollector::sample_eden() {
3851 // Make sure a young gc cannot sneak in between our
3852 // reading and recording of a sample.
3853 assert(Thread::current()->is_ConcurrentGC_thread(),
3854 "Only the cms thread may collect Eden samples");
3855 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3856 "Should collect samples while holding CMS token");
3857 if (!_start_sampling) {
3858 return;
3859 }
3860 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3861 // is populated by the young generation.
3862 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3863 if (_eden_chunk_index < _eden_chunk_capacity) {
3864 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3865 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3866 "Unexpected state of Eden");
3867 // We'd like to check that what we just sampled is an oop-start address;
3868 // however, we cannot do that here since the object may not yet have been
3869 // initialized. So we'll instead do the check when we _use_ this sample
3870 // later.
3871 if (_eden_chunk_index == 0 ||
3872 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3873 _eden_chunk_array[_eden_chunk_index-1])
3874 >= CMSSamplingGrain)) {
3875 _eden_chunk_index++; // commit sample
3876 }
3877 }
3878 }
3879 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3880 size_t used = get_eden_used();
3881 size_t capacity = get_eden_capacity();
3882 assert(used <= capacity, "Unexpected state of Eden");
3883 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3884 _abort_preclean = true;
3885 }
3886 }
3887 }
3888
3889
3890 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3891 assert(_collectorState == Precleaning ||
3892 _collectorState == AbortablePreclean, "incorrect state");
3893 ResourceMark rm;
3894 HandleMark hm;
3895
3896 // Precleaning is currently not MT but the reference processor
3897 // may be set for MT. Disable it temporarily here.
3898 ReferenceProcessor* rp = ref_processor();
3899 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3900
3901 // Do one pass of scrubbing the discovered reference lists
3902 // to remove any reference objects with strongly-reachable
3903 // referents.
3904 if (clean_refs) {
3905 CMSPrecleanRefsYieldClosure yield_cl(this);
3906 assert(rp->span().equals(_span), "Spans should be equal");
3907 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3908 &_markStack, true /* preclean */);
3909 CMSDrainMarkingStackClosure complete_trace(this,
3910 _span, &_markBitMap, &_markStack,
3911 &keep_alive, true /* preclean */);
3912
3913 // We don't want this step to interfere with a young
3914 // collection because we don't want to take CPU
3915 // or memory bandwidth away from the young GC threads
3916 // (which may be as many as there are CPUs).
3917 // Note that we don't need to protect ourselves from
3918 // interference with mutators because they can't
3919 // manipulate the discovered reference lists nor affect
3920 // the computed reachability of the referents, the
3921 // only properties manipulated by the precleaning
3922 // of these reference lists.
3923 stopTimer();
3924 CMSTokenSyncWithLocks x(true /* is cms thread */,
3925 bitMapLock());
3926 startTimer();
3927 sample_eden();
3928
3929 // The following will yield to allow foreground
3930 // collection to proceed promptly. XXX YSR:
3931 // The code in this method may need further
3932 // tweaking for better performance and some restructuring
3933 // for cleaner interfaces.
3934 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3935 rp->preclean_discovered_references(
3936 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3937 gc_timer, _gc_tracer_cm->gc_id());
3938 }
3939
3940 if (clean_survivor) { // preclean the active survivor space(s)
3941 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3942 &_markBitMap, &_modUnionTable,
3943 &_markStack, true /* precleaning phase */);
3944 stopTimer();
3945 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3946 bitMapLock());
3947 startTimer();
3948 unsigned int before_count =
3949 GenCollectedHeap::heap()->total_collections();
3950 SurvivorSpacePrecleanClosure
3951 sss_cl(this, _span, &_markBitMap, &_markStack,
3952 &pam_cl, before_count, CMSYield);
3953 _young_gen->from()->object_iterate_careful(&sss_cl);
3954 _young_gen->to()->object_iterate_careful(&sss_cl);
3955 }
3956 MarkRefsIntoAndScanClosure
3957 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3958 &_markStack, this, CMSYield,
3959 true /* precleaning phase */);
3960 // CAUTION: The following closure has persistent state that may need to
3961 // be reset upon a decrease in the sequence of addresses it
3962 // processes.
3963 ScanMarkedObjectsAgainCarefullyClosure
3964 smoac_cl(this, _span,
3965 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3966
3967 // Preclean dirty cards in ModUnionTable and CardTable using
3968 // appropriate convergence criterion;
3969 // repeat CMSPrecleanIter times unless we find that
3970 // we are losing.
3971 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3972 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3973 "Bad convergence multiplier");
3974 assert(CMSPrecleanThreshold >= 100,
3975 "Unreasonably low CMSPrecleanThreshold");
3976
3977 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3978 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3979 numIter < CMSPrecleanIter;
3980 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3981 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3982 if (Verbose && PrintGCDetails) {
3983 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3984 }
3985 // Either there are very few dirty cards, so re-mark
3986 // pause will be small anyway, or our pre-cleaning isn't
3987 // that much faster than the rate at which cards are being
3988 // dirtied, so we might as well stop and re-mark since
3989 // precleaning won't improve our re-mark time by much.
3990 if (curNumCards <= CMSPrecleanThreshold ||
3991 (numIter > 0 &&
3992 (curNumCards * CMSPrecleanDenominator >
3993 lastNumCards * CMSPrecleanNumerator))) {
3994 numIter++;
3995 cumNumCards += curNumCards;
3996 break;
3997 }
3998 }
3999
4000 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4001
4002 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4003 cumNumCards += curNumCards;
4004 if (PrintGCDetails && PrintCMSStatistics != 0) {
4005 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4006 curNumCards, cumNumCards, numIter);
4007 }
4008 return cumNumCards; // as a measure of useful work done
4009 }
4010
4011 // PRECLEANING NOTES:
4012 // Precleaning involves:
4013 // . reading the bits of the modUnionTable and clearing the set bits.
4014 // . For the cards corresponding to the set bits, we scan the
4015 // objects on those cards. This means we need the free_list_lock
4016 // so that we can safely iterate over the CMS space when scanning
4017 // for oops.
4018 // . When we scan the objects, we'll be both reading and setting
4019 // marks in the marking bit map, so we'll need the marking bit map.
4020 // . For protecting _collector_state transitions, we take the CGC_lock.
4021 // Note that any races in the reading of of card table entries by the
4022 // CMS thread on the one hand and the clearing of those entries by the
4023 // VM thread or the setting of those entries by the mutator threads on the
4024 // other are quite benign. However, for efficiency it makes sense to keep
4025 // the VM thread from racing with the CMS thread while the latter is
4026 // dirty card info to the modUnionTable. We therefore also use the
4027 // CGC_lock to protect the reading of the card table and the mod union
4028 // table by the CM thread.
4029 // . We run concurrently with mutator updates, so scanning
4030 // needs to be done carefully -- we should not try to scan
4031 // potentially uninitialized objects.
4032 //
4033 // Locking strategy: While holding the CGC_lock, we scan over and
4034 // reset a maximal dirty range of the mod union / card tables, then lock
4035 // the free_list_lock and bitmap lock to do a full marking, then
4036 // release these locks; and repeat the cycle. This allows for a
4037 // certain amount of fairness in the sharing of these locks between
4038 // the CMS collector on the one hand, and the VM thread and the
4039 // mutators on the other.
4040
4041 // NOTE: preclean_mod_union_table() and preclean_card_table()
4042 // further below are largely identical; if you need to modify
4043 // one of these methods, please check the other method too.
4044
4045 size_t CMSCollector::preclean_mod_union_table(
4046 ConcurrentMarkSweepGeneration* gen,
4047 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4048 verify_work_stacks_empty();
4049 verify_overflow_empty();
4050
4051 // strategy: starting with the first card, accumulate contiguous
4052 // ranges of dirty cards; clear these cards, then scan the region
4053 // covered by these cards.
4054
4055 // Since all of the MUT is committed ahead, we can just use
4056 // that, in case the generations expand while we are precleaning.
4057 // It might also be fine to just use the committed part of the
4058 // generation, but we might potentially miss cards when the
4059 // generation is rapidly expanding while we are in the midst
4060 // of precleaning.
4061 HeapWord* startAddr = gen->reserved().start();
4062 HeapWord* endAddr = gen->reserved().end();
4063
4064 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4065
4066 size_t numDirtyCards, cumNumDirtyCards;
4067 HeapWord *nextAddr, *lastAddr;
4068 for (cumNumDirtyCards = numDirtyCards = 0,
4069 nextAddr = lastAddr = startAddr;
4070 nextAddr < endAddr;
4071 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4072
4073 ResourceMark rm;
4074 HandleMark hm;
4075
4076 MemRegion dirtyRegion;
4077 {
4078 stopTimer();
4079 // Potential yield point
4080 CMSTokenSync ts(true);
4081 startTimer();
4082 sample_eden();
4083 // Get dirty region starting at nextOffset (inclusive),
4084 // simultaneously clearing it.
4085 dirtyRegion =
4086 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4087 assert(dirtyRegion.start() >= nextAddr,
4088 "returned region inconsistent?");
4089 }
4090 // Remember where the next search should begin.
4091 // The returned region (if non-empty) is a right open interval,
4092 // so lastOffset is obtained from the right end of that
4093 // interval.
4094 lastAddr = dirtyRegion.end();
4095 // Should do something more transparent and less hacky XXX
4096 numDirtyCards =
4097 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4098
4099 // We'll scan the cards in the dirty region (with periodic
4100 // yields for foreground GC as needed).
4101 if (!dirtyRegion.is_empty()) {
4102 assert(numDirtyCards > 0, "consistency check");
4103 HeapWord* stop_point = NULL;
4104 stopTimer();
4105 // Potential yield point
4106 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4107 bitMapLock());
4108 startTimer();
4109 {
4110 verify_work_stacks_empty();
4111 verify_overflow_empty();
4112 sample_eden();
4113 stop_point =
4114 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4115 }
4116 if (stop_point != NULL) {
4117 // The careful iteration stopped early either because it found an
4118 // uninitialized object, or because we were in the midst of an
4119 // "abortable preclean", which should now be aborted. Redirty
4120 // the bits corresponding to the partially-scanned or unscanned
4121 // cards. We'll either restart at the next block boundary or
4122 // abort the preclean.
4123 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4124 "Should only be AbortablePreclean.");
4125 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4126 if (should_abort_preclean()) {
4127 break; // out of preclean loop
4128 } else {
4129 // Compute the next address at which preclean should pick up;
4130 // might need bitMapLock in order to read P-bits.
4131 lastAddr = next_card_start_after_block(stop_point);
4132 }
4133 }
4134 } else {
4135 assert(lastAddr == endAddr, "consistency check");
4136 assert(numDirtyCards == 0, "consistency check");
4137 break;
4138 }
4139 }
4140 verify_work_stacks_empty();
4141 verify_overflow_empty();
4142 return cumNumDirtyCards;
4143 }
4144
4145 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4146 // below are largely identical; if you need to modify
4147 // one of these methods, please check the other method too.
4148
4149 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4150 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4151 // strategy: it's similar to precleamModUnionTable above, in that
4152 // we accumulate contiguous ranges of dirty cards, mark these cards
4153 // precleaned, then scan the region covered by these cards.
4154 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4155 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4156
4157 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4158
4159 size_t numDirtyCards, cumNumDirtyCards;
4160 HeapWord *lastAddr, *nextAddr;
4161
4162 for (cumNumDirtyCards = numDirtyCards = 0,
4163 nextAddr = lastAddr = startAddr;
4164 nextAddr < endAddr;
4165 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4166
4167 ResourceMark rm;
4168 HandleMark hm;
4169
4170 MemRegion dirtyRegion;
4171 {
4172 // See comments in "Precleaning notes" above on why we
4173 // do this locking. XXX Could the locking overheads be
4174 // too high when dirty cards are sparse? [I don't think so.]
4175 stopTimer();
4176 CMSTokenSync x(true); // is cms thread
4177 startTimer();
4178 sample_eden();
4179 // Get and clear dirty region from card table
4180 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4181 MemRegion(nextAddr, endAddr),
4182 true,
4183 CardTableModRefBS::precleaned_card_val());
4184
4185 assert(dirtyRegion.start() >= nextAddr,
4186 "returned region inconsistent?");
4187 }
4188 lastAddr = dirtyRegion.end();
4189 numDirtyCards =
4190 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4191
4192 if (!dirtyRegion.is_empty()) {
4193 stopTimer();
4194 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4195 startTimer();
4196 sample_eden();
4197 verify_work_stacks_empty();
4198 verify_overflow_empty();
4199 HeapWord* stop_point =
4200 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4201 if (stop_point != NULL) {
4202 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4203 "Should only be AbortablePreclean.");
4204 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4205 if (should_abort_preclean()) {
4206 break; // out of preclean loop
4207 } else {
4208 // Compute the next address at which preclean should pick up.
4209 lastAddr = next_card_start_after_block(stop_point);
4210 }
4211 }
4212 } else {
4213 break;
4214 }
4215 }
4216 verify_work_stacks_empty();
4217 verify_overflow_empty();
4218 return cumNumDirtyCards;
4219 }
4220
4221 class PrecleanKlassClosure : public KlassClosure {
4222 KlassToOopClosure _cm_klass_closure;
4223 public:
4224 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4225 void do_klass(Klass* k) {
4226 if (k->has_accumulated_modified_oops()) {
4227 k->clear_accumulated_modified_oops();
4228
4229 _cm_klass_closure.do_klass(k);
4230 }
4231 }
4232 };
4233
4234 // The freelist lock is needed to prevent asserts, is it really needed?
4235 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4236
4237 cl->set_freelistLock(freelistLock);
4238
4239 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4240
4241 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4242 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4243 PrecleanKlassClosure preclean_klass_closure(cl);
4244 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4245
4246 verify_work_stacks_empty();
4247 verify_overflow_empty();
4248 }
4249
4250 void CMSCollector::checkpointRootsFinal() {
4251 assert(_collectorState == FinalMarking, "incorrect state transition?");
4252 check_correct_thread_executing();
4253 // world is stopped at this checkpoint
4254 assert(SafepointSynchronize::is_at_safepoint(),
4255 "world should be stopped");
4256 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4257
4258 verify_work_stacks_empty();
4259 verify_overflow_empty();
4260
4261 if (PrintGCDetails) {
4262 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4263 _young_gen->used() / K,
4264 _young_gen->capacity() / K);
4265 }
4266 {
4267 if (CMSScavengeBeforeRemark) {
4268 GenCollectedHeap* gch = GenCollectedHeap::heap();
4269 // Temporarily set flag to false, GCH->do_collection will
4270 // expect it to be false and set to true
4271 FlagSetting fl(gch->_is_gc_active, false);
4272 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4273 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4274 gch->do_collection(true, // full (i.e. force, see below)
4275 false, // !clear_all_soft_refs
4276 0, // size
4277 false, // is_tlab
4278 Generation::Young // type
4279 );
4280 }
4281 FreelistLocker x(this);
4282 MutexLockerEx y(bitMapLock(),
4283 Mutex::_no_safepoint_check_flag);
4284 checkpointRootsFinalWork();
4285 }
4286 verify_work_stacks_empty();
4287 verify_overflow_empty();
4288 }
4289
4290 void CMSCollector::checkpointRootsFinalWork() {
4291 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4292
4293 assert(haveFreelistLocks(), "must have free list locks");
4294 assert_lock_strong(bitMapLock());
4295
4296 ResourceMark rm;
4297 HandleMark hm;
4298
4299 GenCollectedHeap* gch = GenCollectedHeap::heap();
4300
4301 if (should_unload_classes()) {
4302 CodeCache::gc_prologue();
4303 }
4304 assert(haveFreelistLocks(), "must have free list locks");
4305 assert_lock_strong(bitMapLock());
4306
4307 // We might assume that we need not fill TLAB's when
4308 // CMSScavengeBeforeRemark is set, because we may have just done
4309 // a scavenge which would have filled all TLAB's -- and besides
4310 // Eden would be empty. This however may not always be the case --
4311 // for instance although we asked for a scavenge, it may not have
4312 // happened because of a JNI critical section. We probably need
4313 // a policy for deciding whether we can in that case wait until
4314 // the critical section releases and then do the remark following
4315 // the scavenge, and skip it here. In the absence of that policy,
4316 // or of an indication of whether the scavenge did indeed occur,
4317 // we cannot rely on TLAB's having been filled and must do
4318 // so here just in case a scavenge did not happen.
4319 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4320 // Update the saved marks which may affect the root scans.
4321 gch->save_marks();
4322
4323 if (CMSPrintEdenSurvivorChunks) {
4324 print_eden_and_survivor_chunk_arrays();
4325 }
4326
4327 {
4328 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4329
4330 // Note on the role of the mod union table:
4331 // Since the marker in "markFromRoots" marks concurrently with
4332 // mutators, it is possible for some reachable objects not to have been
4333 // scanned. For instance, an only reference to an object A was
4334 // placed in object B after the marker scanned B. Unless B is rescanned,
4335 // A would be collected. Such updates to references in marked objects
4336 // are detected via the mod union table which is the set of all cards
4337 // dirtied since the first checkpoint in this GC cycle and prior to
4338 // the most recent young generation GC, minus those cleaned up by the
4339 // concurrent precleaning.
4340 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
4341 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4342 do_remark_parallel();
4343 } else {
4344 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4345 _gc_timer_cm, _gc_tracer_cm->gc_id());
4346 do_remark_non_parallel();
4347 }
4348 }
4349 verify_work_stacks_empty();
4350 verify_overflow_empty();
4351
4352 {
4353 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4354 refProcessingWork();
4355 }
4356 verify_work_stacks_empty();
4357 verify_overflow_empty();
4358
4359 if (should_unload_classes()) {
4360 CodeCache::gc_epilogue();
4361 }
4362 JvmtiExport::gc_epilogue();
4363
4364 // If we encountered any (marking stack / work queue) overflow
4365 // events during the current CMS cycle, take appropriate
4366 // remedial measures, where possible, so as to try and avoid
4367 // recurrence of that condition.
4368 assert(_markStack.isEmpty(), "No grey objects");
4369 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4370 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4371 if (ser_ovflw > 0) {
4372 if (PrintCMSStatistics != 0) {
4373 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4374 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4375 ", kac_preclean="SIZE_FORMAT")",
4376 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4377 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4378 }
4379 _markStack.expand();
4380 _ser_pmc_remark_ovflw = 0;
4381 _ser_pmc_preclean_ovflw = 0;
4382 _ser_kac_preclean_ovflw = 0;
4383 _ser_kac_ovflw = 0;
4384 }
4385 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4386 if (PrintCMSStatistics != 0) {
4387 gclog_or_tty->print_cr("Work queue overflow (benign) "
4388 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4389 _par_pmc_remark_ovflw, _par_kac_ovflw);
4390 }
4391 _par_pmc_remark_ovflw = 0;
4392 _par_kac_ovflw = 0;
4393 }
4394 if (PrintCMSStatistics != 0) {
4395 if (_markStack._hit_limit > 0) {
4396 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4397 _markStack._hit_limit);
4398 }
4399 if (_markStack._failed_double > 0) {
4400 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4401 " current capacity "SIZE_FORMAT,
4402 _markStack._failed_double,
4403 _markStack.capacity());
4404 }
4405 }
4406 _markStack._hit_limit = 0;
4407 _markStack._failed_double = 0;
4408
4409 if ((VerifyAfterGC || VerifyDuringGC) &&
4410 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4411 verify_after_remark();
4412 }
4413
4414 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4415
4416 // Change under the freelistLocks.
4417 _collectorState = Sweeping;
4418 // Call isAllClear() under bitMapLock
4419 assert(_modUnionTable.isAllClear(),
4420 "Should be clear by end of the final marking");
4421 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4422 "Should be clear by end of the final marking");
4423 }
4424
4425 void CMSParInitialMarkTask::work(uint worker_id) {
4426 elapsedTimer _timer;
4427 ResourceMark rm;
4428 HandleMark hm;
4429
4430 // ---------- scan from roots --------------
4431 _timer.start();
4432 GenCollectedHeap* gch = GenCollectedHeap::heap();
4433 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4434
4435 // ---------- young gen roots --------------
4436 {
4437 work_on_young_gen_roots(worker_id, &par_mri_cl);
4438 _timer.stop();
4439 if (PrintCMSStatistics != 0) {
4440 gclog_or_tty->print_cr(
4441 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4442 worker_id, _timer.seconds());
4443 }
4444 }
4445
4446 // ---------- remaining roots --------------
4447 _timer.reset();
4448 _timer.start();
4449
4450 CLDToOopClosure cld_closure(&par_mri_cl, true);
4451
4452 gch->gen_process_roots(Generation::Old,
4453 false, // yg was scanned above
4454 false, // this is parallel code
4455 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4456 _collector->should_unload_classes(),
4457 &par_mri_cl,
4458 NULL,
4459 &cld_closure);
4460 assert(_collector->should_unload_classes()
4461 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4462 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4463 _timer.stop();
4464 if (PrintCMSStatistics != 0) {
4465 gclog_or_tty->print_cr(
4466 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4467 worker_id, _timer.seconds());
4468 }
4469 }
4470
4471 // Parallel remark task
4472 class CMSParRemarkTask: public CMSParMarkTask {
4473 CompactibleFreeListSpace* _cms_space;
4474
4475 // The per-thread work queues, available here for stealing.
4476 OopTaskQueueSet* _task_queues;
4477 ParallelTaskTerminator _term;
4478
4479 public:
4480 // A value of 0 passed to n_workers will cause the number of
4481 // workers to be taken from the active workers in the work gang.
4482 CMSParRemarkTask(CMSCollector* collector,
4483 CompactibleFreeListSpace* cms_space,
4484 int n_workers, FlexibleWorkGang* workers,
4485 OopTaskQueueSet* task_queues):
4486 CMSParMarkTask("Rescan roots and grey objects in parallel",
4487 collector, n_workers),
4488 _cms_space(cms_space),
4489 _task_queues(task_queues),
4490 _term(n_workers, task_queues) { }
4491
4492 OopTaskQueueSet* task_queues() { return _task_queues; }
4493
4494 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4495
4496 ParallelTaskTerminator* terminator() { return &_term; }
4497 int n_workers() { return _n_workers; }
4498
4499 void work(uint worker_id);
4500
4501 private:
4502 // ... of dirty cards in old space
4503 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4504 Par_MarkRefsIntoAndScanClosure* cl);
4505
4506 // ... work stealing for the above
4507 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4508 };
4509
4510 class RemarkKlassClosure : public KlassClosure {
4511 KlassToOopClosure _cm_klass_closure;
4512 public:
4513 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4514 void do_klass(Klass* k) {
4515 // Check if we have modified any oops in the Klass during the concurrent marking.
4516 if (k->has_accumulated_modified_oops()) {
4517 k->clear_accumulated_modified_oops();
4518
4519 // We could have transfered the current modified marks to the accumulated marks,
4520 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4521 } else if (k->has_modified_oops()) {
4522 // Don't clear anything, this info is needed by the next young collection.
4523 } else {
4524 // No modified oops in the Klass.
4525 return;
4526 }
4527
4528 // The klass has modified fields, need to scan the klass.
4529 _cm_klass_closure.do_klass(k);
4530 }
4531 };
4532
4533 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4534 ParNewGeneration* young_gen = _collector->_young_gen;
4535 ContiguousSpace* eden_space = young_gen->eden();
4536 ContiguousSpace* from_space = young_gen->from();
4537 ContiguousSpace* to_space = young_gen->to();
4538
4539 HeapWord** eca = _collector->_eden_chunk_array;
4540 size_t ect = _collector->_eden_chunk_index;
4541 HeapWord** sca = _collector->_survivor_chunk_array;
4542 size_t sct = _collector->_survivor_chunk_index;
4543
4544 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4545 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4546
4547 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4548 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4549 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4550 }
4551
4552 // work_queue(i) is passed to the closure
4553 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4554 // also is passed to do_dirty_card_rescan_tasks() and to
4555 // do_work_steal() to select the i-th task_queue.
4556
4557 void CMSParRemarkTask::work(uint worker_id) {
4558 elapsedTimer _timer;
4559 ResourceMark rm;
4560 HandleMark hm;
4561
4562 // ---------- rescan from roots --------------
4563 _timer.start();
4564 GenCollectedHeap* gch = GenCollectedHeap::heap();
4565 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4566 _collector->_span, _collector->ref_processor(),
4567 &(_collector->_markBitMap),
4568 work_queue(worker_id));
4569
4570 // Rescan young gen roots first since these are likely
4571 // coarsely partitioned and may, on that account, constitute
4572 // the critical path; thus, it's best to start off that
4573 // work first.
4574 // ---------- young gen roots --------------
4575 {
4576 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4577 _timer.stop();
4578 if (PrintCMSStatistics != 0) {
4579 gclog_or_tty->print_cr(
4580 "Finished young gen rescan work in %dth thread: %3.3f sec",
4581 worker_id, _timer.seconds());
4582 }
4583 }
4584
4585 // ---------- remaining roots --------------
4586 _timer.reset();
4587 _timer.start();
4588 gch->gen_process_roots(Generation::Old,
4589 false, // yg was scanned above
4590 false, // this is parallel code
4591 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4592 _collector->should_unload_classes(),
4593 &par_mrias_cl,
4594 NULL,
4595 NULL); // The dirty klasses will be handled below
4596
4597 assert(_collector->should_unload_classes()
4598 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4599 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4600 _timer.stop();
4601 if (PrintCMSStatistics != 0) {
4602 gclog_or_tty->print_cr(
4603 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4604 worker_id, _timer.seconds());
4605 }
4606
4607 // ---------- unhandled CLD scanning ----------
4608 if (worker_id == 0) { // Single threaded at the moment.
4609 _timer.reset();
4610 _timer.start();
4611
4612 // Scan all new class loader data objects and new dependencies that were
4613 // introduced during concurrent marking.
4614 ResourceMark rm;
4615 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4616 for (int i = 0; i < array->length(); i++) {
4617 par_mrias_cl.do_class_loader_data(array->at(i));
4618 }
4619
4620 // We don't need to keep track of new CLDs anymore.
4621 ClassLoaderDataGraph::remember_new_clds(false);
4622
4623 _timer.stop();
4624 if (PrintCMSStatistics != 0) {
4625 gclog_or_tty->print_cr(
4626 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4627 worker_id, _timer.seconds());
4628 }
4629 }
4630
4631 // ---------- dirty klass scanning ----------
4632 if (worker_id == 0) { // Single threaded at the moment.
4633 _timer.reset();
4634 _timer.start();
4635
4636 // Scan all classes that was dirtied during the concurrent marking phase.
4637 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4638 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4639
4640 _timer.stop();
4641 if (PrintCMSStatistics != 0) {
4642 gclog_or_tty->print_cr(
4643 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4644 worker_id, _timer.seconds());
4645 }
4646 }
4647
4648 // We might have added oops to ClassLoaderData::_handles during the
4649 // concurrent marking phase. These oops point to newly allocated objects
4650 // that are guaranteed to be kept alive. Either by the direct allocation
4651 // code, or when the young collector processes the roots. Hence,
4652 // we don't have to revisit the _handles block during the remark phase.
4653
4654 // ---------- rescan dirty cards ------------
4655 _timer.reset();
4656 _timer.start();
4657
4658 // Do the rescan tasks for each of the two spaces
4659 // (cms_space) in turn.
4660 // "worker_id" is passed to select the task_queue for "worker_id"
4661 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4662 _timer.stop();
4663 if (PrintCMSStatistics != 0) {
4664 gclog_or_tty->print_cr(
4665 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4666 worker_id, _timer.seconds());
4667 }
4668
4669 // ---------- steal work from other threads ...
4670 // ---------- ... and drain overflow list.
4671 _timer.reset();
4672 _timer.start();
4673 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4674 _timer.stop();
4675 if (PrintCMSStatistics != 0) {
4676 gclog_or_tty->print_cr(
4677 "Finished work stealing in %dth thread: %3.3f sec",
4678 worker_id, _timer.seconds());
4679 }
4680 }
4681
4682 // Note that parameter "i" is not used.
4683 void
4684 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4685 OopsInGenClosure* cl, ContiguousSpace* space,
4686 HeapWord** chunk_array, size_t chunk_top) {
4687 // Until all tasks completed:
4688 // . claim an unclaimed task
4689 // . compute region boundaries corresponding to task claimed
4690 // using chunk_array
4691 // . par_oop_iterate(cl) over that region
4692
4693 ResourceMark rm;
4694 HandleMark hm;
4695
4696 SequentialSubTasksDone* pst = space->par_seq_tasks();
4697
4698 uint nth_task = 0;
4699 uint n_tasks = pst->n_tasks();
4700
4701 if (n_tasks > 0) {
4702 assert(pst->valid(), "Uninitialized use?");
4703 HeapWord *start, *end;
4704 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4705 // We claimed task # nth_task; compute its boundaries.
4706 if (chunk_top == 0) { // no samples were taken
4707 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4708 start = space->bottom();
4709 end = space->top();
4710 } else if (nth_task == 0) {
4711 start = space->bottom();
4712 end = chunk_array[nth_task];
4713 } else if (nth_task < (uint)chunk_top) {
4714 assert(nth_task >= 1, "Control point invariant");
4715 start = chunk_array[nth_task - 1];
4716 end = chunk_array[nth_task];
4717 } else {
4718 assert(nth_task == (uint)chunk_top, "Control point invariant");
4719 start = chunk_array[chunk_top - 1];
4720 end = space->top();
4721 }
4722 MemRegion mr(start, end);
4723 // Verify that mr is in space
4724 assert(mr.is_empty() || space->used_region().contains(mr),
4725 "Should be in space");
4726 // Verify that "start" is an object boundary
4727 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4728 "Should be an oop");
4729 space->par_oop_iterate(mr, cl);
4730 }
4731 pst->all_tasks_completed();
4732 }
4733 }
4734
4735 void
4736 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4737 CompactibleFreeListSpace* sp, int i,
4738 Par_MarkRefsIntoAndScanClosure* cl) {
4739 // Until all tasks completed:
4740 // . claim an unclaimed task
4741 // . compute region boundaries corresponding to task claimed
4742 // . transfer dirty bits ct->mut for that region
4743 // . apply rescanclosure to dirty mut bits for that region
4744
4745 ResourceMark rm;
4746 HandleMark hm;
4747
4748 OopTaskQueue* work_q = work_queue(i);
4749 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4750 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4751 // CAUTION: This closure has state that persists across calls to
4752 // the work method dirty_range_iterate_clear() in that it has
4753 // embedded in it a (subtype of) UpwardsObjectClosure. The
4754 // use of that state in the embedded UpwardsObjectClosure instance
4755 // assumes that the cards are always iterated (even if in parallel
4756 // by several threads) in monotonically increasing order per each
4757 // thread. This is true of the implementation below which picks
4758 // card ranges (chunks) in monotonically increasing order globally
4759 // and, a-fortiori, in monotonically increasing order per thread
4760 // (the latter order being a subsequence of the former).
4761 // If the work code below is ever reorganized into a more chaotic
4762 // work-partitioning form than the current "sequential tasks"
4763 // paradigm, the use of that persistent state will have to be
4764 // revisited and modified appropriately. See also related
4765 // bug 4756801 work on which should examine this code to make
4766 // sure that the changes there do not run counter to the
4767 // assumptions made here and necessary for correctness and
4768 // efficiency. Note also that this code might yield inefficient
4769 // behavior in the case of very large objects that span one or
4770 // more work chunks. Such objects would potentially be scanned
4771 // several times redundantly. Work on 4756801 should try and
4772 // address that performance anomaly if at all possible. XXX
4773 MemRegion full_span = _collector->_span;
4774 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4775 MarkFromDirtyCardsClosure
4776 greyRescanClosure(_collector, full_span, // entire span of interest
4777 sp, bm, work_q, cl);
4778
4779 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4780 assert(pst->valid(), "Uninitialized use?");
4781 uint nth_task = 0;
4782 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4783 MemRegion span = sp->used_region();
4784 HeapWord* start_addr = span.start();
4785 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4786 alignment);
4787 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4788 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4789 start_addr, "Check alignment");
4790 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4791 chunk_size, "Check alignment");
4792
4793 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4794 // Having claimed the nth_task, compute corresponding mem-region,
4795 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4796 // The alignment restriction ensures that we do not need any
4797 // synchronization with other gang-workers while setting or
4798 // clearing bits in thus chunk of the MUT.
4799 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4800 start_addr + (nth_task+1)*chunk_size);
4801 // The last chunk's end might be way beyond end of the
4802 // used region. In that case pull back appropriately.
4803 if (this_span.end() > end_addr) {
4804 this_span.set_end(end_addr);
4805 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4806 }
4807 // Iterate over the dirty cards covering this chunk, marking them
4808 // precleaned, and setting the corresponding bits in the mod union
4809 // table. Since we have been careful to partition at Card and MUT-word
4810 // boundaries no synchronization is needed between parallel threads.
4811 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4812 &modUnionClosure);
4813
4814 // Having transferred these marks into the modUnionTable,
4815 // rescan the marked objects on the dirty cards in the modUnionTable.
4816 // Even if this is at a synchronous collection, the initial marking
4817 // may have been done during an asynchronous collection so there
4818 // may be dirty bits in the mod-union table.
4819 _collector->_modUnionTable.dirty_range_iterate_clear(
4820 this_span, &greyRescanClosure);
4821 _collector->_modUnionTable.verifyNoOneBitsInRange(
4822 this_span.start(),
4823 this_span.end());
4824 }
4825 pst->all_tasks_completed(); // declare that i am done
4826 }
4827
4828 // . see if we can share work_queues with ParNew? XXX
4829 void
4830 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4831 int* seed) {
4832 OopTaskQueue* work_q = work_queue(i);
4833 NOT_PRODUCT(int num_steals = 0;)
4834 oop obj_to_scan;
4835 CMSBitMap* bm = &(_collector->_markBitMap);
4836
4837 while (true) {
4838 // Completely finish any left over work from (an) earlier round(s)
4839 cl->trim_queue(0);
4840 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4841 (size_t)ParGCDesiredObjsFromOverflowList);
4842 // Now check if there's any work in the overflow list
4843 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4844 // only affects the number of attempts made to get work from the
4845 // overflow list and does not affect the number of workers. Just
4846 // pass ParallelGCThreads so this behavior is unchanged.
4847 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4848 work_q,
4849 ParallelGCThreads)) {
4850 // found something in global overflow list;
4851 // not yet ready to go stealing work from others.
4852 // We'd like to assert(work_q->size() != 0, ...)
4853 // because we just took work from the overflow list,
4854 // but of course we can't since all of that could have
4855 // been already stolen from us.
4856 // "He giveth and He taketh away."
4857 continue;
4858 }
4859 // Verify that we have no work before we resort to stealing
4860 assert(work_q->size() == 0, "Have work, shouldn't steal");
4861 // Try to steal from other queues that have work
4862 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4863 NOT_PRODUCT(num_steals++;)
4864 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4865 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4866 // Do scanning work
4867 obj_to_scan->oop_iterate(cl);
4868 // Loop around, finish this work, and try to steal some more
4869 } else if (terminator()->offer_termination()) {
4870 break; // nirvana from the infinite cycle
4871 }
4872 }
4873 NOT_PRODUCT(
4874 if (PrintCMSStatistics != 0) {
4875 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4876 }
4877 )
4878 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4879 "Else our work is not yet done");
4880 }
4881
4882 // Record object boundaries in _eden_chunk_array by sampling the eden
4883 // top in the slow-path eden object allocation code path and record
4884 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4885 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4886 // sampling in sample_eden() that activates during the part of the
4887 // preclean phase.
4888 void CMSCollector::sample_eden_chunk() {
4889 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4890 if (_eden_chunk_lock->try_lock()) {
4891 // Record a sample. This is the critical section. The contents
4892 // of the _eden_chunk_array have to be non-decreasing in the
4893 // address order.
4894 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4895 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4896 "Unexpected state of Eden");
4897 if (_eden_chunk_index == 0 ||
4898 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4899 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4900 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4901 _eden_chunk_index++; // commit sample
4902 }
4903 _eden_chunk_lock->unlock();
4904 }
4905 }
4906 }
4907
4908 // Return a thread-local PLAB recording array, as appropriate.
4909 void* CMSCollector::get_data_recorder(int thr_num) {
4910 if (_survivor_plab_array != NULL &&
4911 (CMSPLABRecordAlways ||
4912 (_collectorState > Marking && _collectorState < FinalMarking))) {
4913 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4914 ChunkArray* ca = &_survivor_plab_array[thr_num];
4915 ca->reset(); // clear it so that fresh data is recorded
4916 return (void*) ca;
4917 } else {
4918 return NULL;
4919 }
4920 }
4921
4922 // Reset all the thread-local PLAB recording arrays
4923 void CMSCollector::reset_survivor_plab_arrays() {
4924 for (uint i = 0; i < ParallelGCThreads; i++) {
4925 _survivor_plab_array[i].reset();
4926 }
4927 }
4928
4929 // Merge the per-thread plab arrays into the global survivor chunk
4930 // array which will provide the partitioning of the survivor space
4931 // for CMS initial scan and rescan.
4932 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4933 int no_of_gc_threads) {
4934 assert(_survivor_plab_array != NULL, "Error");
4935 assert(_survivor_chunk_array != NULL, "Error");
4936 assert(_collectorState == FinalMarking ||
4937 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4938 for (int j = 0; j < no_of_gc_threads; j++) {
4939 _cursor[j] = 0;
4940 }
4941 HeapWord* top = surv->top();
4942 size_t i;
4943 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4944 HeapWord* min_val = top; // Higher than any PLAB address
4945 uint min_tid = 0; // position of min_val this round
4946 for (int j = 0; j < no_of_gc_threads; j++) {
4947 ChunkArray* cur_sca = &_survivor_plab_array[j];
4948 if (_cursor[j] == cur_sca->end()) {
4949 continue;
4950 }
4951 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4952 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4953 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4954 if (cur_val < min_val) {
4955 min_tid = j;
4956 min_val = cur_val;
4957 } else {
4958 assert(cur_val < top, "All recorded addresses should be less");
4959 }
4960 }
4961 // At this point min_val and min_tid are respectively
4962 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4963 // and the thread (j) that witnesses that address.
4964 // We record this address in the _survivor_chunk_array[i]
4965 // and increment _cursor[min_tid] prior to the next round i.
4966 if (min_val == top) {
4967 break;
4968 }
4969 _survivor_chunk_array[i] = min_val;
4970 _cursor[min_tid]++;
4971 }
4972 // We are all done; record the size of the _survivor_chunk_array
4973 _survivor_chunk_index = i; // exclusive: [0, i)
4974 if (PrintCMSStatistics > 0) {
4975 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4976 }
4977 // Verify that we used up all the recorded entries
4978 #ifdef ASSERT
4979 size_t total = 0;
4980 for (int j = 0; j < no_of_gc_threads; j++) {
4981 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4982 total += _cursor[j];
4983 }
4984 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4985 // Check that the merged array is in sorted order
4986 if (total > 0) {
4987 for (size_t i = 0; i < total - 1; i++) {
4988 if (PrintCMSStatistics > 0) {
4989 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4990 i, p2i(_survivor_chunk_array[i]));
4991 }
4992 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4993 "Not sorted");
4994 }
4995 }
4996 #endif // ASSERT
4997 }
4998
4999 // Set up the space's par_seq_tasks structure for work claiming
5000 // for parallel initial scan and rescan of young gen.
5001 // See ParRescanTask where this is currently used.
5002 void
5003 CMSCollector::
5004 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5005 assert(n_threads > 0, "Unexpected n_threads argument");
5006
5007 // Eden space
5008 if (!_young_gen->eden()->is_empty()) {
5009 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5010 assert(!pst->valid(), "Clobbering existing data?");
5011 // Each valid entry in [0, _eden_chunk_index) represents a task.
5012 size_t n_tasks = _eden_chunk_index + 1;
5013 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5014 // Sets the condition for completion of the subtask (how many threads
5015 // need to finish in order to be done).
5016 pst->set_n_threads(n_threads);
5017 pst->set_n_tasks((int)n_tasks);
5018 }
5019
5020 // Merge the survivor plab arrays into _survivor_chunk_array
5021 if (_survivor_plab_array != NULL) {
5022 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5023 } else {
5024 assert(_survivor_chunk_index == 0, "Error");
5025 }
5026
5027 // To space
5028 {
5029 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5030 assert(!pst->valid(), "Clobbering existing data?");
5031 // Sets the condition for completion of the subtask (how many threads
5032 // need to finish in order to be done).
5033 pst->set_n_threads(n_threads);
5034 pst->set_n_tasks(1);
5035 assert(pst->valid(), "Error");
5036 }
5037
5038 // From space
5039 {
5040 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5041 assert(!pst->valid(), "Clobbering existing data?");
5042 size_t n_tasks = _survivor_chunk_index + 1;
5043 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
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((int)n_tasks);
5048 assert(pst->valid(), "Error");
5049 }
5050 }
5051
5052 // Parallel version of remark
5053 void CMSCollector::do_remark_parallel() {
5054 GenCollectedHeap* gch = GenCollectedHeap::heap();
5055 FlexibleWorkGang* workers = gch->workers();
5056 assert(workers != NULL, "Need parallel worker threads.");
5057 // Choose to use the number of GC workers most recently set
5058 // into "active_workers". If active_workers is not set, set it
5059 // to ParallelGCThreads.
5060 int n_workers = workers->active_workers();
5061 if (n_workers == 0) {
5062 assert(n_workers > 0, "Should have been set during scavenge");
5063 n_workers = ParallelGCThreads;
5064 workers->set_active_workers(n_workers);
5065 }
5066 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5067
5068 CMSParRemarkTask tsk(this,
5069 cms_space,
5070 n_workers, workers, task_queues());
5071
5072 // Set up for parallel process_roots work.
5073 gch->set_par_threads(n_workers);
5074 // We won't be iterating over the cards in the card table updating
5075 // the younger_gen cards, so we shouldn't call the following else
5076 // the verification code as well as subsequent younger_refs_iterate
5077 // code would get confused. XXX
5078 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5079
5080 // The young gen rescan work will not be done as part of
5081 // process_roots (which currently doesn't know how to
5082 // parallelize such a scan), but rather will be broken up into
5083 // a set of parallel tasks (via the sampling that the [abortable]
5084 // preclean phase did of eden, plus the [two] tasks of
5085 // scanning the [two] survivor spaces. Further fine-grain
5086 // parallelization of the scanning of the survivor spaces
5087 // themselves, and of precleaning of the younger gen itself
5088 // is deferred to the future.
5089 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5090
5091 // The dirty card rescan work is broken up into a "sequence"
5092 // of parallel tasks (per constituent space) that are dynamically
5093 // claimed by the parallel threads.
5094 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5095
5096 // It turns out that even when we're using 1 thread, doing the work in a
5097 // separate thread causes wide variance in run times. We can't help this
5098 // in the multi-threaded case, but we special-case n=1 here to get
5099 // repeatable measurements of the 1-thread overhead of the parallel code.
5100 if (n_workers > 1) {
5101 // Make refs discovery MT-safe, if it isn't already: it may not
5102 // necessarily be so, since it's possible that we are doing
5103 // ST marking.
5104 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5105 StrongRootsScope srs;
5106 workers->run_task(&tsk);
5107 } else {
5108 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5109 StrongRootsScope srs;
5110 tsk.work(0);
5111 }
5112
5113 gch->set_par_threads(0); // 0 ==> non-parallel.
5114 // restore, single-threaded for now, any preserved marks
5115 // as a result of work_q overflow
5116 restore_preserved_marks_if_any();
5117 }
5118
5119 // Non-parallel version of remark
5120 void CMSCollector::do_remark_non_parallel() {
5121 ResourceMark rm;
5122 HandleMark hm;
5123 GenCollectedHeap* gch = GenCollectedHeap::heap();
5124 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5125
5126 MarkRefsIntoAndScanClosure
5127 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5128 &_markStack, this,
5129 false /* should_yield */, false /* not precleaning */);
5130 MarkFromDirtyCardsClosure
5131 markFromDirtyCardsClosure(this, _span,
5132 NULL, // space is set further below
5133 &_markBitMap, &_markStack, &mrias_cl);
5134 {
5135 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5136 // Iterate over the dirty cards, setting the corresponding bits in the
5137 // mod union table.
5138 {
5139 ModUnionClosure modUnionClosure(&_modUnionTable);
5140 _ct->ct_bs()->dirty_card_iterate(
5141 _cmsGen->used_region(),
5142 &modUnionClosure);
5143 }
5144 // Having transferred these marks into the modUnionTable, we just need
5145 // to rescan the marked objects on the dirty cards in the modUnionTable.
5146 // The initial marking may have been done during an asynchronous
5147 // collection so there may be dirty bits in the mod-union table.
5148 const int alignment =
5149 CardTableModRefBS::card_size * BitsPerWord;
5150 {
5151 // ... First handle dirty cards in CMS gen
5152 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5153 MemRegion ur = _cmsGen->used_region();
5154 HeapWord* lb = ur.start();
5155 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5156 MemRegion cms_span(lb, ub);
5157 _modUnionTable.dirty_range_iterate_clear(cms_span,
5158 &markFromDirtyCardsClosure);
5159 verify_work_stacks_empty();
5160 if (PrintCMSStatistics != 0) {
5161 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5162 markFromDirtyCardsClosure.num_dirty_cards());
5163 }
5164 }
5165 }
5166 if (VerifyDuringGC &&
5167 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5168 HandleMark hm; // Discard invalid handles created during verification
5169 Universe::verify();
5170 }
5171 {
5172 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5173
5174 verify_work_stacks_empty();
5175
5176 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5177 StrongRootsScope srs;
5178
5179 gch->gen_process_roots(Generation::Old,
5180 true, // younger gens as roots
5181 false, // use the local StrongRootsScope
5182 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5183 should_unload_classes(),
5184 &mrias_cl,
5185 NULL,
5186 NULL); // The dirty klasses will be handled below
5187
5188 assert(should_unload_classes()
5189 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5190 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5191 }
5192
5193 {
5194 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5195
5196 verify_work_stacks_empty();
5197
5198 // Scan all class loader data objects that might have been introduced
5199 // during concurrent marking.
5200 ResourceMark rm;
5201 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5202 for (int i = 0; i < array->length(); i++) {
5203 mrias_cl.do_class_loader_data(array->at(i));
5204 }
5205
5206 // We don't need to keep track of new CLDs anymore.
5207 ClassLoaderDataGraph::remember_new_clds(false);
5208
5209 verify_work_stacks_empty();
5210 }
5211
5212 {
5213 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5214
5215 verify_work_stacks_empty();
5216
5217 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5218 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5219
5220 verify_work_stacks_empty();
5221 }
5222
5223 // We might have added oops to ClassLoaderData::_handles during the
5224 // concurrent marking phase. These oops point to newly allocated objects
5225 // that are guaranteed to be kept alive. Either by the direct allocation
5226 // code, or when the young collector processes the roots. Hence,
5227 // we don't have to revisit the _handles block during the remark phase.
5228
5229 verify_work_stacks_empty();
5230 // Restore evacuated mark words, if any, used for overflow list links
5231 if (!CMSOverflowEarlyRestoration) {
5232 restore_preserved_marks_if_any();
5233 }
5234 verify_overflow_empty();
5235 }
5236
5237 ////////////////////////////////////////////////////////
5238 // Parallel Reference Processing Task Proxy Class
5239 ////////////////////////////////////////////////////////
5240 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5241 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5242 CMSCollector* _collector;
5243 CMSBitMap* _mark_bit_map;
5244 const MemRegion _span;
5245 ProcessTask& _task;
5246
5247 public:
5248 CMSRefProcTaskProxy(ProcessTask& task,
5249 CMSCollector* collector,
5250 const MemRegion& span,
5251 CMSBitMap* mark_bit_map,
5252 AbstractWorkGang* workers,
5253 OopTaskQueueSet* task_queues):
5254 // XXX Should superclass AGTWOQ also know about AWG since it knows
5255 // about the task_queues used by the AWG? Then it could initialize
5256 // the terminator() object. See 6984287. The set_for_termination()
5257 // below is a temporary band-aid for the regression in 6984287.
5258 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5259 task_queues),
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 set_for_termination(workers->active_workers());
5266 }
5267
5268 OopTaskQueueSet* task_queues() { return queues(); }
5269
5270 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5271
5272 void do_work_steal(int i,
5273 CMSParDrainMarkingStackClosure* drain,
5274 CMSParKeepAliveClosure* keep_alive,
5275 int* seed);
5276
5277 virtual void work(uint worker_id);
5278 };
5279
5280 void CMSRefProcTaskProxy::work(uint worker_id) {
5281 ResourceMark rm;
5282 HandleMark hm;
5283 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5284 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5285 _mark_bit_map,
5286 work_queue(worker_id));
5287 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5288 _mark_bit_map,
5289 work_queue(worker_id));
5290 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5291 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5292 if (_task.marks_oops_alive()) {
5293 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5294 _collector->hash_seed(worker_id));
5295 }
5296 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5297 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5298 }
5299
5300 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5301 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5302 EnqueueTask& _task;
5303
5304 public:
5305 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5306 : AbstractGangTask("Enqueue reference objects in parallel"),
5307 _task(task)
5308 { }
5309
5310 virtual void work(uint worker_id)
5311 {
5312 _task.work(worker_id);
5313 }
5314 };
5315
5316 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5317 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5318 _span(span),
5319 _bit_map(bit_map),
5320 _work_queue(work_queue),
5321 _mark_and_push(collector, span, bit_map, work_queue),
5322 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5323 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5324 { }
5325
5326 // . see if we can share work_queues with ParNew? XXX
5327 void CMSRefProcTaskProxy::do_work_steal(int i,
5328 CMSParDrainMarkingStackClosure* drain,
5329 CMSParKeepAliveClosure* keep_alive,
5330 int* seed) {
5331 OopTaskQueue* work_q = work_queue(i);
5332 NOT_PRODUCT(int num_steals = 0;)
5333 oop obj_to_scan;
5334
5335 while (true) {
5336 // Completely finish any left over work from (an) earlier round(s)
5337 drain->trim_queue(0);
5338 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5339 (size_t)ParGCDesiredObjsFromOverflowList);
5340 // Now check if there's any work in the overflow list
5341 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5342 // only affects the number of attempts made to get work from the
5343 // overflow list and does not affect the number of workers. Just
5344 // pass ParallelGCThreads so this behavior is unchanged.
5345 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5346 work_q,
5347 ParallelGCThreads)) {
5348 // Found something in global overflow list;
5349 // not yet ready to go stealing work from others.
5350 // We'd like to assert(work_q->size() != 0, ...)
5351 // because we just took work from the overflow list,
5352 // but of course we can't, since all of that might have
5353 // been already stolen from us.
5354 continue;
5355 }
5356 // Verify that we have no work before we resort to stealing
5357 assert(work_q->size() == 0, "Have work, shouldn't steal");
5358 // Try to steal from other queues that have work
5359 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5360 NOT_PRODUCT(num_steals++;)
5361 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5362 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5363 // Do scanning work
5364 obj_to_scan->oop_iterate(keep_alive);
5365 // Loop around, finish this work, and try to steal some more
5366 } else if (terminator()->offer_termination()) {
5367 break; // nirvana from the infinite cycle
5368 }
5369 }
5370 NOT_PRODUCT(
5371 if (PrintCMSStatistics != 0) {
5372 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5373 }
5374 )
5375 }
5376
5377 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5378 {
5379 GenCollectedHeap* gch = GenCollectedHeap::heap();
5380 FlexibleWorkGang* workers = gch->workers();
5381 assert(workers != NULL, "Need parallel worker threads.");
5382 CMSRefProcTaskProxy rp_task(task, &_collector,
5383 _collector.ref_processor()->span(),
5384 _collector.markBitMap(),
5385 workers, _collector.task_queues());
5386 workers->run_task(&rp_task);
5387 }
5388
5389 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5390 {
5391
5392 GenCollectedHeap* gch = GenCollectedHeap::heap();
5393 FlexibleWorkGang* workers = gch->workers();
5394 assert(workers != NULL, "Need parallel worker threads.");
5395 CMSRefEnqueueTaskProxy enq_task(task);
5396 workers->run_task(&enq_task);
5397 }
5398
5399 void CMSCollector::refProcessingWork() {
5400 ResourceMark rm;
5401 HandleMark hm;
5402
5403 ReferenceProcessor* rp = ref_processor();
5404 assert(rp->span().equals(_span), "Spans should be equal");
5405 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5406 // Process weak references.
5407 rp->setup_policy(false);
5408 verify_work_stacks_empty();
5409
5410 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5411 &_markStack, false /* !preclean */);
5412 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5413 _span, &_markBitMap, &_markStack,
5414 &cmsKeepAliveClosure, false /* !preclean */);
5415 {
5416 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5417
5418 ReferenceProcessorStats stats;
5419 if (rp->processing_is_mt()) {
5420 // Set the degree of MT here. If the discovery is done MT, there
5421 // may have been a different number of threads doing the discovery
5422 // and a different number of discovered lists may have Ref objects.
5423 // That is OK as long as the Reference lists are balanced (see
5424 // balance_all_queues() and balance_queues()).
5425 GenCollectedHeap* gch = GenCollectedHeap::heap();
5426 int active_workers = ParallelGCThreads;
5427 FlexibleWorkGang* workers = gch->workers();
5428 if (workers != NULL) {
5429 active_workers = workers->active_workers();
5430 // The expectation is that active_workers will have already
5431 // been set to a reasonable value. If it has not been set,
5432 // investigate.
5433 assert(active_workers > 0, "Should have been set during scavenge");
5434 }
5435 rp->set_active_mt_degree(active_workers);
5436 CMSRefProcTaskExecutor task_executor(*this);
5437 stats = rp->process_discovered_references(&_is_alive_closure,
5438 &cmsKeepAliveClosure,
5439 &cmsDrainMarkingStackClosure,
5440 &task_executor,
5441 _gc_timer_cm,
5442 _gc_tracer_cm->gc_id());
5443 } else {
5444 stats = rp->process_discovered_references(&_is_alive_closure,
5445 &cmsKeepAliveClosure,
5446 &cmsDrainMarkingStackClosure,
5447 NULL,
5448 _gc_timer_cm,
5449 _gc_tracer_cm->gc_id());
5450 }
5451 _gc_tracer_cm->report_gc_reference_stats(stats);
5452
5453 }
5454
5455 // This is the point where the entire marking should have completed.
5456 verify_work_stacks_empty();
5457
5458 if (should_unload_classes()) {
5459 {
5460 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5461
5462 // Unload classes and purge the SystemDictionary.
5463 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5464
5465 // Unload nmethods.
5466 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5467
5468 // Prune dead klasses from subklass/sibling/implementor lists.
5469 Klass::clean_weak_klass_links(&_is_alive_closure);
5470 }
5471
5472 {
5473 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5474 // Clean up unreferenced symbols in symbol table.
5475 SymbolTable::unlink();
5476 }
5477
5478 {
5479 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5480 // Delete entries for dead interned strings.
5481 StringTable::unlink(&_is_alive_closure);
5482 }
5483 }
5484
5485
5486 // Restore any preserved marks as a result of mark stack or
5487 // work queue overflow
5488 restore_preserved_marks_if_any(); // done single-threaded for now
5489
5490 rp->set_enqueuing_is_done(true);
5491 if (rp->processing_is_mt()) {
5492 rp->balance_all_queues();
5493 CMSRefProcTaskExecutor task_executor(*this);
5494 rp->enqueue_discovered_references(&task_executor);
5495 } else {
5496 rp->enqueue_discovered_references(NULL);
5497 }
5498 rp->verify_no_references_recorded();
5499 assert(!rp->discovery_enabled(), "should have been disabled");
5500 }
5501
5502 #ifndef PRODUCT
5503 void CMSCollector::check_correct_thread_executing() {
5504 Thread* t = Thread::current();
5505 // Only the VM thread or the CMS thread should be here.
5506 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5507 "Unexpected thread type");
5508 // If this is the vm thread, the foreground process
5509 // should not be waiting. Note that _foregroundGCIsActive is
5510 // true while the foreground collector is waiting.
5511 if (_foregroundGCShouldWait) {
5512 // We cannot be the VM thread
5513 assert(t->is_ConcurrentGC_thread(),
5514 "Should be CMS thread");
5515 } else {
5516 // We can be the CMS thread only if we are in a stop-world
5517 // phase of CMS collection.
5518 if (t->is_ConcurrentGC_thread()) {
5519 assert(_collectorState == InitialMarking ||
5520 _collectorState == FinalMarking,
5521 "Should be a stop-world phase");
5522 // The CMS thread should be holding the CMS_token.
5523 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5524 "Potential interference with concurrently "
5525 "executing VM thread");
5526 }
5527 }
5528 }
5529 #endif
5530
5531 void CMSCollector::sweep() {
5532 assert(_collectorState == Sweeping, "just checking");
5533 check_correct_thread_executing();
5534 verify_work_stacks_empty();
5535 verify_overflow_empty();
5536 increment_sweep_count();
5537 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5538
5539 _inter_sweep_timer.stop();
5540 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5541
5542 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5543 _intra_sweep_timer.reset();
5544 _intra_sweep_timer.start();
5545 {
5546 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5547 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5548 // First sweep the old gen
5549 {
5550 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5551 bitMapLock());
5552 sweepWork(_cmsGen);
5553 }
5554
5555 // Update Universe::_heap_*_at_gc figures.
5556 // We need all the free list locks to make the abstract state
5557 // transition from Sweeping to Resetting. See detailed note
5558 // further below.
5559 {
5560 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5561 // Update heap occupancy information which is used as
5562 // input to soft ref clearing policy at the next gc.
5563 Universe::update_heap_info_at_gc();
5564 _collectorState = Resizing;
5565 }
5566 }
5567 verify_work_stacks_empty();
5568 verify_overflow_empty();
5569
5570 if (should_unload_classes()) {
5571 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5572 // requires that the virtual spaces are stable and not deleted.
5573 ClassLoaderDataGraph::set_should_purge(true);
5574 }
5575
5576 _intra_sweep_timer.stop();
5577 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5578
5579 _inter_sweep_timer.reset();
5580 _inter_sweep_timer.start();
5581
5582 // We need to use a monotonically non-decreasing time in ms
5583 // or we will see time-warp warnings and os::javaTimeMillis()
5584 // does not guarantee monotonicity.
5585 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5586 update_time_of_last_gc(now);
5587
5588 // NOTE on abstract state transitions:
5589 // Mutators allocate-live and/or mark the mod-union table dirty
5590 // based on the state of the collection. The former is done in
5591 // the interval [Marking, Sweeping] and the latter in the interval
5592 // [Marking, Sweeping). Thus the transitions into the Marking state
5593 // and out of the Sweeping state must be synchronously visible
5594 // globally to the mutators.
5595 // The transition into the Marking state happens with the world
5596 // stopped so the mutators will globally see it. Sweeping is
5597 // done asynchronously by the background collector so the transition
5598 // from the Sweeping state to the Resizing state must be done
5599 // under the freelistLock (as is the check for whether to
5600 // allocate-live and whether to dirty the mod-union table).
5601 assert(_collectorState == Resizing, "Change of collector state to"
5602 " Resizing must be done under the freelistLocks (plural)");
5603
5604 // Now that sweeping has been completed, we clear
5605 // the incremental_collection_failed flag,
5606 // thus inviting a younger gen collection to promote into
5607 // this generation. If such a promotion may still fail,
5608 // the flag will be set again when a young collection is
5609 // attempted.
5610 GenCollectedHeap* gch = GenCollectedHeap::heap();
5611 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5612 gch->update_full_collections_completed(_collection_count_start);
5613 }
5614
5615 // FIX ME!!! Looks like this belongs in CFLSpace, with
5616 // CMSGen merely delegating to it.
5617 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5618 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5619 HeapWord* minAddr = _cmsSpace->bottom();
5620 HeapWord* largestAddr =
5621 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5622 if (largestAddr == NULL) {
5623 // The dictionary appears to be empty. In this case
5624 // try to coalesce at the end of the heap.
5625 largestAddr = _cmsSpace->end();
5626 }
5627 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5628 size_t nearLargestOffset =
5629 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5630 if (PrintFLSStatistics != 0) {
5631 gclog_or_tty->print_cr(
5632 "CMS: Large Block: " PTR_FORMAT ";"
5633 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5634 p2i(largestAddr),
5635 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5636 }
5637 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5638 }
5639
5640 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5641 return addr >= _cmsSpace->nearLargestChunk();
5642 }
5643
5644 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5645 return _cmsSpace->find_chunk_at_end();
5646 }
5647
5648 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5649 bool full) {
5650 // If the young generation has been collected. Gather any statistics
5651 // that are of interest at this point.
5652 bool current_is_young = (current_generation == GenCollectedHeap::heap()->young_gen());
5653 if (!full && current_is_young) {
5654 // Gather statistics on the young generation collection.
5655 collector()->stats().record_gc0_end(used());
5656 }
5657 }
5658
5659 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) {
5660 // We iterate over the space(s) underlying this generation,
5661 // checking the mark bit map to see if the bits corresponding
5662 // to specific blocks are marked or not. Blocks that are
5663 // marked are live and are not swept up. All remaining blocks
5664 // are swept up, with coalescing on-the-fly as we sweep up
5665 // contiguous free and/or garbage blocks:
5666 // We need to ensure that the sweeper synchronizes with allocators
5667 // and stop-the-world collectors. In particular, the following
5668 // locks are used:
5669 // . CMS token: if this is held, a stop the world collection cannot occur
5670 // . freelistLock: if this is held no allocation can occur from this
5671 // generation by another thread
5672 // . bitMapLock: if this is held, no other thread can access or update
5673 //
5674
5675 // Note that we need to hold the freelistLock if we use
5676 // block iterate below; else the iterator might go awry if
5677 // a mutator (or promotion) causes block contents to change
5678 // (for instance if the allocator divvies up a block).
5679 // If we hold the free list lock, for all practical purposes
5680 // young generation GC's can't occur (they'll usually need to
5681 // promote), so we might as well prevent all young generation
5682 // GC's while we do a sweeping step. For the same reason, we might
5683 // as well take the bit map lock for the entire duration
5684
5685 // check that we hold the requisite locks
5686 assert(have_cms_token(), "Should hold cms token");
5687 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5688 assert_lock_strong(gen->freelistLock());
5689 assert_lock_strong(bitMapLock());
5690
5691 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5692 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5693 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5694 _inter_sweep_estimate.padded_average(),
5695 _intra_sweep_estimate.padded_average());
5696 gen->setNearLargestChunk();
5697
5698 {
5699 SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield);
5700 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5701 // We need to free-up/coalesce garbage/blocks from a
5702 // co-terminal free run. This is done in the SweepClosure
5703 // destructor; so, do not remove this scope, else the
5704 // end-of-sweep-census below will be off by a little bit.
5705 }
5706 gen->cmsSpace()->sweep_completed();
5707 gen->cmsSpace()->endSweepFLCensus(sweep_count());
5708 if (should_unload_classes()) { // unloaded classes this cycle,
5709 _concurrent_cycles_since_last_unload = 0; // ... reset count
5710 } else { // did not unload classes,
5711 _concurrent_cycles_since_last_unload++; // ... increment count
5712 }
5713 }
5714
5715 // Reset CMS data structures (for now just the marking bit map)
5716 // preparatory for the next cycle.
5717 void CMSCollector::reset(bool concurrent) {
5718 if (concurrent) {
5719 CMSTokenSyncWithLocks ts(true, bitMapLock());
5720
5721 // If the state is not "Resetting", the foreground thread
5722 // has done a collection and the resetting.
5723 if (_collectorState != Resetting) {
5724 assert(_collectorState == Idling, "The state should only change"
5725 " because the foreground collector has finished the collection");
5726 return;
5727 }
5728
5729 // Clear the mark bitmap (no grey objects to start with)
5730 // for the next cycle.
5731 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5732 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5733
5734 HeapWord* curAddr = _markBitMap.startWord();
5735 while (curAddr < _markBitMap.endWord()) {
5736 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5737 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5738 _markBitMap.clear_large_range(chunk);
5739 if (ConcurrentMarkSweepThread::should_yield() &&
5740 !foregroundGCIsActive() &&
5741 CMSYield) {
5742 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5743 "CMS thread should hold CMS token");
5744 assert_lock_strong(bitMapLock());
5745 bitMapLock()->unlock();
5746 ConcurrentMarkSweepThread::desynchronize(true);
5747 stopTimer();
5748 if (PrintCMSStatistics != 0) {
5749 incrementYields();
5750 }
5751
5752 // See the comment in coordinator_yield()
5753 for (unsigned i = 0; i < CMSYieldSleepCount &&
5754 ConcurrentMarkSweepThread::should_yield() &&
5755 !CMSCollector::foregroundGCIsActive(); ++i) {
5756 os::sleep(Thread::current(), 1, false);
5757 }
5758
5759 ConcurrentMarkSweepThread::synchronize(true);
5760 bitMapLock()->lock_without_safepoint_check();
5761 startTimer();
5762 }
5763 curAddr = chunk.end();
5764 }
5765 // A successful mostly concurrent collection has been done.
5766 // Because only the full (i.e., concurrent mode failure) collections
5767 // are being measured for gc overhead limits, clean the "near" flag
5768 // and count.
5769 size_policy()->reset_gc_overhead_limit_count();
5770 _collectorState = Idling;
5771 } else {
5772 // already have the lock
5773 assert(_collectorState == Resetting, "just checking");
5774 assert_lock_strong(bitMapLock());
5775 _markBitMap.clear_all();
5776 _collectorState = Idling;
5777 }
5778
5779 register_gc_end();
5780 }
5781
5782 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5783 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5784 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5785 TraceCollectorStats tcs(counters());
5786
5787 switch (op) {
5788 case CMS_op_checkpointRootsInitial: {
5789 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5790 checkpointRootsInitial();
5791 if (PrintGC) {
5792 _cmsGen->printOccupancy("initial-mark");
5793 }
5794 break;
5795 }
5796 case CMS_op_checkpointRootsFinal: {
5797 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5798 checkpointRootsFinal();
5799 if (PrintGC) {
5800 _cmsGen->printOccupancy("remark");
5801 }
5802 break;
5803 }
5804 default:
5805 fatal("No such CMS_op");
5806 }
5807 }
5808
5809 #ifndef PRODUCT
5810 size_t const CMSCollector::skip_header_HeapWords() {
5811 return FreeChunk::header_size();
5812 }
5813
5814 // Try and collect here conditions that should hold when
5815 // CMS thread is exiting. The idea is that the foreground GC
5816 // thread should not be blocked if it wants to terminate
5817 // the CMS thread and yet continue to run the VM for a while
5818 // after that.
5819 void CMSCollector::verify_ok_to_terminate() const {
5820 assert(Thread::current()->is_ConcurrentGC_thread(),
5821 "should be called by CMS thread");
5822 assert(!_foregroundGCShouldWait, "should be false");
5823 // We could check here that all the various low-level locks
5824 // are not held by the CMS thread, but that is overkill; see
5825 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5826 // is checked.
5827 }
5828 #endif
5829
5830 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5831 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5832 "missing Printezis mark?");
5833 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5834 size_t size = pointer_delta(nextOneAddr + 1, addr);
5835 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5836 "alignment problem");
5837 assert(size >= 3, "Necessary for Printezis marks to work");
5838 return size;
5839 }
5840
5841 // A variant of the above (block_size_using_printezis_bits()) except
5842 // that we return 0 if the P-bits are not yet set.
5843 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5844 if (_markBitMap.isMarked(addr + 1)) {
5845 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5846 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5847 size_t size = pointer_delta(nextOneAddr + 1, addr);
5848 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5849 "alignment problem");
5850 assert(size >= 3, "Necessary for Printezis marks to work");
5851 return size;
5852 }
5853 return 0;
5854 }
5855
5856 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5857 size_t sz = 0;
5858 oop p = (oop)addr;
5859 if (p->klass_or_null() != NULL) {
5860 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5861 } else {
5862 sz = block_size_using_printezis_bits(addr);
5863 }
5864 assert(sz > 0, "size must be nonzero");
5865 HeapWord* next_block = addr + sz;
5866 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5867 CardTableModRefBS::card_size);
5868 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5869 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5870 "must be different cards");
5871 return next_card;
5872 }
5873
5874
5875 // CMS Bit Map Wrapper /////////////////////////////////////////
5876
5877 // Construct a CMS bit map infrastructure, but don't create the
5878 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5879 // further below.
5880 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5881 _bm(),
5882 _shifter(shifter),
5883 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5884 Monitor::_safepoint_check_sometimes) : NULL)
5885 {
5886 _bmStartWord = 0;
5887 _bmWordSize = 0;
5888 }
5889
5890 bool CMSBitMap::allocate(MemRegion mr) {
5891 _bmStartWord = mr.start();
5892 _bmWordSize = mr.word_size();
5893 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5894 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5895 if (!brs.is_reserved()) {
5896 warning("CMS bit map allocation failure");
5897 return false;
5898 }
5899 // For now we'll just commit all of the bit map up front.
5900 // Later on we'll try to be more parsimonious with swap.
5901 if (!_virtual_space.initialize(brs, brs.size())) {
5902 warning("CMS bit map backing store failure");
5903 return false;
5904 }
5905 assert(_virtual_space.committed_size() == brs.size(),
5906 "didn't reserve backing store for all of CMS bit map?");
5907 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5908 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5909 _bmWordSize, "inconsistency in bit map sizing");
5910 _bm.set_size(_bmWordSize >> _shifter);
5911
5912 // bm.clear(); // can we rely on getting zero'd memory? verify below
5913 assert(isAllClear(),
5914 "Expected zero'd memory from ReservedSpace constructor");
5915 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5916 "consistency check");
5917 return true;
5918 }
5919
5920 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5921 HeapWord *next_addr, *end_addr, *last_addr;
5922 assert_locked();
5923 assert(covers(mr), "out-of-range error");
5924 // XXX assert that start and end are appropriately aligned
5925 for (next_addr = mr.start(), end_addr = mr.end();
5926 next_addr < end_addr; next_addr = last_addr) {
5927 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5928 last_addr = dirty_region.end();
5929 if (!dirty_region.is_empty()) {
5930 cl->do_MemRegion(dirty_region);
5931 } else {
5932 assert(last_addr == end_addr, "program logic");
5933 return;
5934 }
5935 }
5936 }
5937
5938 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5939 _bm.print_on_error(st, prefix);
5940 }
5941
5942 #ifndef PRODUCT
5943 void CMSBitMap::assert_locked() const {
5944 CMSLockVerifier::assert_locked(lock());
5945 }
5946
5947 bool CMSBitMap::covers(MemRegion mr) const {
5948 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5949 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5950 "size inconsistency");
5951 return (mr.start() >= _bmStartWord) &&
5952 (mr.end() <= endWord());
5953 }
5954
5955 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5956 return (start >= _bmStartWord && (start + size) <= endWord());
5957 }
5958
5959 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5960 // verify that there are no 1 bits in the interval [left, right)
5961 FalseBitMapClosure falseBitMapClosure;
5962 iterate(&falseBitMapClosure, left, right);
5963 }
5964
5965 void CMSBitMap::region_invariant(MemRegion mr)
5966 {
5967 assert_locked();
5968 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5969 assert(!mr.is_empty(), "unexpected empty region");
5970 assert(covers(mr), "mr should be covered by bit map");
5971 // convert address range into offset range
5972 size_t start_ofs = heapWordToOffset(mr.start());
5973 // Make sure that end() is appropriately aligned
5974 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5975 (1 << (_shifter+LogHeapWordSize))),
5976 "Misaligned mr.end()");
5977 size_t end_ofs = heapWordToOffset(mr.end());
5978 assert(end_ofs > start_ofs, "Should mark at least one bit");
5979 }
5980
5981 #endif
5982
5983 bool CMSMarkStack::allocate(size_t size) {
5984 // allocate a stack of the requisite depth
5985 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5986 size * sizeof(oop)));
5987 if (!rs.is_reserved()) {
5988 warning("CMSMarkStack allocation failure");
5989 return false;
5990 }
5991 if (!_virtual_space.initialize(rs, rs.size())) {
5992 warning("CMSMarkStack backing store failure");
5993 return false;
5994 }
5995 assert(_virtual_space.committed_size() == rs.size(),
5996 "didn't reserve backing store for all of CMS stack?");
5997 _base = (oop*)(_virtual_space.low());
5998 _index = 0;
5999 _capacity = size;
6000 NOT_PRODUCT(_max_depth = 0);
6001 return true;
6002 }
6003
6004 // XXX FIX ME !!! In the MT case we come in here holding a
6005 // leaf lock. For printing we need to take a further lock
6006 // which has lower rank. We need to recalibrate the two
6007 // lock-ranks involved in order to be able to print the
6008 // messages below. (Or defer the printing to the caller.
6009 // For now we take the expedient path of just disabling the
6010 // messages for the problematic case.)
6011 void CMSMarkStack::expand() {
6012 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6013 if (_capacity == MarkStackSizeMax) {
6014 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6015 // We print a warning message only once per CMS cycle.
6016 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6017 }
6018 return;
6019 }
6020 // Double capacity if possible
6021 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6022 // Do not give up existing stack until we have managed to
6023 // get the double capacity that we desired.
6024 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6025 new_capacity * sizeof(oop)));
6026 if (rs.is_reserved()) {
6027 // Release the backing store associated with old stack
6028 _virtual_space.release();
6029 // Reinitialize virtual space for new stack
6030 if (!_virtual_space.initialize(rs, rs.size())) {
6031 fatal("Not enough swap for expanded marking stack");
6032 }
6033 _base = (oop*)(_virtual_space.low());
6034 _index = 0;
6035 _capacity = new_capacity;
6036 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6037 // Failed to double capacity, continue;
6038 // we print a detail message only once per CMS cycle.
6039 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6040 SIZE_FORMAT"K",
6041 _capacity / K, new_capacity / K);
6042 }
6043 }
6044
6045
6046 // Closures
6047 // XXX: there seems to be a lot of code duplication here;
6048 // should refactor and consolidate common code.
6049
6050 // This closure is used to mark refs into the CMS generation in
6051 // the CMS bit map. Called at the first checkpoint. This closure
6052 // assumes that we do not need to re-mark dirty cards; if the CMS
6053 // generation on which this is used is not an oldest
6054 // generation then this will lose younger_gen cards!
6055
6056 MarkRefsIntoClosure::MarkRefsIntoClosure(
6057 MemRegion span, CMSBitMap* bitMap):
6058 _span(span),
6059 _bitMap(bitMap)
6060 {
6061 assert(_ref_processor == NULL, "deliberately left NULL");
6062 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6063 }
6064
6065 void MarkRefsIntoClosure::do_oop(oop obj) {
6066 // if p points into _span, then mark corresponding bit in _markBitMap
6067 assert(obj->is_oop(), "expected an oop");
6068 HeapWord* addr = (HeapWord*)obj;
6069 if (_span.contains(addr)) {
6070 // this should be made more efficient
6071 _bitMap->mark(addr);
6072 }
6073 }
6074
6075 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6076 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6077
6078 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6079 MemRegion span, CMSBitMap* bitMap):
6080 _span(span),
6081 _bitMap(bitMap)
6082 {
6083 assert(_ref_processor == NULL, "deliberately left NULL");
6084 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6085 }
6086
6087 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6088 // if p points into _span, then mark corresponding bit in _markBitMap
6089 assert(obj->is_oop(), "expected an oop");
6090 HeapWord* addr = (HeapWord*)obj;
6091 if (_span.contains(addr)) {
6092 // this should be made more efficient
6093 _bitMap->par_mark(addr);
6094 }
6095 }
6096
6097 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6098 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6099
6100 // A variant of the above, used for CMS marking verification.
6101 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6102 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6103 _span(span),
6104 _verification_bm(verification_bm),
6105 _cms_bm(cms_bm)
6106 {
6107 assert(_ref_processor == NULL, "deliberately left NULL");
6108 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6109 }
6110
6111 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6112 // if p points into _span, then mark corresponding bit in _markBitMap
6113 assert(obj->is_oop(), "expected an oop");
6114 HeapWord* addr = (HeapWord*)obj;
6115 if (_span.contains(addr)) {
6116 _verification_bm->mark(addr);
6117 if (!_cms_bm->isMarked(addr)) {
6118 oop(addr)->print();
6119 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6120 fatal("... aborting");
6121 }
6122 }
6123 }
6124
6125 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6126 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6127
6128 //////////////////////////////////////////////////
6129 // MarkRefsIntoAndScanClosure
6130 //////////////////////////////////////////////////
6131
6132 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6133 ReferenceProcessor* rp,
6134 CMSBitMap* bit_map,
6135 CMSBitMap* mod_union_table,
6136 CMSMarkStack* mark_stack,
6137 CMSCollector* collector,
6138 bool should_yield,
6139 bool concurrent_precleaning):
6140 _collector(collector),
6141 _span(span),
6142 _bit_map(bit_map),
6143 _mark_stack(mark_stack),
6144 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6145 mark_stack, concurrent_precleaning),
6146 _yield(should_yield),
6147 _concurrent_precleaning(concurrent_precleaning),
6148 _freelistLock(NULL)
6149 {
6150 _ref_processor = rp;
6151 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6152 }
6153
6154 // This closure is used to mark refs into the CMS generation at the
6155 // second (final) checkpoint, and to scan and transitively follow
6156 // the unmarked oops. It is also used during the concurrent precleaning
6157 // phase while scanning objects on dirty cards in the CMS generation.
6158 // The marks are made in the marking bit map and the marking stack is
6159 // used for keeping the (newly) grey objects during the scan.
6160 // The parallel version (Par_...) appears further below.
6161 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6162 if (obj != NULL) {
6163 assert(obj->is_oop(), "expected an oop");
6164 HeapWord* addr = (HeapWord*)obj;
6165 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6166 assert(_collector->overflow_list_is_empty(),
6167 "overflow list should be empty");
6168 if (_span.contains(addr) &&
6169 !_bit_map->isMarked(addr)) {
6170 // mark bit map (object is now grey)
6171 _bit_map->mark(addr);
6172 // push on marking stack (stack should be empty), and drain the
6173 // stack by applying this closure to the oops in the oops popped
6174 // from the stack (i.e. blacken the grey objects)
6175 bool res = _mark_stack->push(obj);
6176 assert(res, "Should have space to push on empty stack");
6177 do {
6178 oop new_oop = _mark_stack->pop();
6179 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6180 assert(_bit_map->isMarked((HeapWord*)new_oop),
6181 "only grey objects on this stack");
6182 // iterate over the oops in this oop, marking and pushing
6183 // the ones in CMS heap (i.e. in _span).
6184 new_oop->oop_iterate(&_pushAndMarkClosure);
6185 // check if it's time to yield
6186 do_yield_check();
6187 } while (!_mark_stack->isEmpty() ||
6188 (!_concurrent_precleaning && take_from_overflow_list()));
6189 // if marking stack is empty, and we are not doing this
6190 // during precleaning, then check the overflow list
6191 }
6192 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6193 assert(_collector->overflow_list_is_empty(),
6194 "overflow list was drained above");
6195 // We could restore evacuated mark words, if any, used for
6196 // overflow list links here because the overflow list is
6197 // provably empty here. That would reduce the maximum
6198 // size requirements for preserved_{oop,mark}_stack.
6199 // But we'll just postpone it until we are all done
6200 // so we can just stream through.
6201 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6202 _collector->restore_preserved_marks_if_any();
6203 assert(_collector->no_preserved_marks(), "No preserved marks");
6204 }
6205 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6206 "All preserved marks should have been restored above");
6207 }
6208 }
6209
6210 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6211 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6212
6213 void MarkRefsIntoAndScanClosure::do_yield_work() {
6214 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6215 "CMS thread should hold CMS token");
6216 assert_lock_strong(_freelistLock);
6217 assert_lock_strong(_bit_map->lock());
6218 // relinquish the free_list_lock and bitMaplock()
6219 _bit_map->lock()->unlock();
6220 _freelistLock->unlock();
6221 ConcurrentMarkSweepThread::desynchronize(true);
6222 _collector->stopTimer();
6223 if (PrintCMSStatistics != 0) {
6224 _collector->incrementYields();
6225 }
6226
6227 // See the comment in coordinator_yield()
6228 for (unsigned i = 0;
6229 i < CMSYieldSleepCount &&
6230 ConcurrentMarkSweepThread::should_yield() &&
6231 !CMSCollector::foregroundGCIsActive();
6232 ++i) {
6233 os::sleep(Thread::current(), 1, false);
6234 }
6235
6236 ConcurrentMarkSweepThread::synchronize(true);
6237 _freelistLock->lock_without_safepoint_check();
6238 _bit_map->lock()->lock_without_safepoint_check();
6239 _collector->startTimer();
6240 }
6241
6242 ///////////////////////////////////////////////////////////
6243 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6244 // MarkRefsIntoAndScanClosure
6245 ///////////////////////////////////////////////////////////
6246 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6247 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6248 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6249 _span(span),
6250 _bit_map(bit_map),
6251 _work_queue(work_queue),
6252 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6253 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6254 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6255 {
6256 _ref_processor = rp;
6257 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6258 }
6259
6260 // This closure is used to mark refs into the CMS generation at the
6261 // second (final) checkpoint, and to scan and transitively follow
6262 // the unmarked oops. The marks are made in the marking bit map and
6263 // the work_queue is used for keeping the (newly) grey objects during
6264 // the scan phase whence they are also available for stealing by parallel
6265 // threads. Since the marking bit map is shared, updates are
6266 // synchronized (via CAS).
6267 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6268 if (obj != NULL) {
6269 // Ignore mark word because this could be an already marked oop
6270 // that may be chained at the end of the overflow list.
6271 assert(obj->is_oop(true), "expected an oop");
6272 HeapWord* addr = (HeapWord*)obj;
6273 if (_span.contains(addr) &&
6274 !_bit_map->isMarked(addr)) {
6275 // mark bit map (object will become grey):
6276 // It is possible for several threads to be
6277 // trying to "claim" this object concurrently;
6278 // the unique thread that succeeds in marking the
6279 // object first will do the subsequent push on
6280 // to the work queue (or overflow list).
6281 if (_bit_map->par_mark(addr)) {
6282 // push on work_queue (which may not be empty), and trim the
6283 // queue to an appropriate length by applying this closure to
6284 // the oops in the oops popped from the stack (i.e. blacken the
6285 // grey objects)
6286 bool res = _work_queue->push(obj);
6287 assert(res, "Low water mark should be less than capacity?");
6288 trim_queue(_low_water_mark);
6289 } // Else, another thread claimed the object
6290 }
6291 }
6292 }
6293
6294 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6295 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6296
6297 // This closure is used to rescan the marked objects on the dirty cards
6298 // in the mod union table and the card table proper.
6299 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6300 oop p, MemRegion mr) {
6301
6302 size_t size = 0;
6303 HeapWord* addr = (HeapWord*)p;
6304 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6305 assert(_span.contains(addr), "we are scanning the CMS generation");
6306 // check if it's time to yield
6307 if (do_yield_check()) {
6308 // We yielded for some foreground stop-world work,
6309 // and we have been asked to abort this ongoing preclean cycle.
6310 return 0;
6311 }
6312 if (_bitMap->isMarked(addr)) {
6313 // it's marked; is it potentially uninitialized?
6314 if (p->klass_or_null() != NULL) {
6315 // an initialized object; ignore mark word in verification below
6316 // since we are running concurrent with mutators
6317 assert(p->is_oop(true), "should be an oop");
6318 if (p->is_objArray()) {
6319 // objArrays are precisely marked; restrict scanning
6320 // to dirty cards only.
6321 size = CompactibleFreeListSpace::adjustObjectSize(
6322 p->oop_iterate(_scanningClosure, mr));
6323 } else {
6324 // A non-array may have been imprecisely marked; we need
6325 // to scan object in its entirety.
6326 size = CompactibleFreeListSpace::adjustObjectSize(
6327 p->oop_iterate(_scanningClosure));
6328 }
6329 #ifdef ASSERT
6330 size_t direct_size =
6331 CompactibleFreeListSpace::adjustObjectSize(p->size());
6332 assert(size == direct_size, "Inconsistency in size");
6333 assert(size >= 3, "Necessary for Printezis marks to work");
6334 if (!_bitMap->isMarked(addr+1)) {
6335 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6336 } else {
6337 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6338 assert(_bitMap->isMarked(addr+size-1),
6339 "inconsistent Printezis mark");
6340 }
6341 #endif // ASSERT
6342 } else {
6343 // An uninitialized object.
6344 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6345 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6346 size = pointer_delta(nextOneAddr + 1, addr);
6347 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6348 "alignment problem");
6349 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6350 // will dirty the card when the klass pointer is installed in the
6351 // object (signaling the completion of initialization).
6352 }
6353 } else {
6354 // Either a not yet marked object or an uninitialized object
6355 if (p->klass_or_null() == NULL) {
6356 // An uninitialized object, skip to the next card, since
6357 // we may not be able to read its P-bits yet.
6358 assert(size == 0, "Initial value");
6359 } else {
6360 // An object not (yet) reached by marking: we merely need to
6361 // compute its size so as to go look at the next block.
6362 assert(p->is_oop(true), "should be an oop");
6363 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6364 }
6365 }
6366 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6367 return size;
6368 }
6369
6370 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6371 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6372 "CMS thread should hold CMS token");
6373 assert_lock_strong(_freelistLock);
6374 assert_lock_strong(_bitMap->lock());
6375 // relinquish the free_list_lock and bitMaplock()
6376 _bitMap->lock()->unlock();
6377 _freelistLock->unlock();
6378 ConcurrentMarkSweepThread::desynchronize(true);
6379 _collector->stopTimer();
6380 if (PrintCMSStatistics != 0) {
6381 _collector->incrementYields();
6382 }
6383
6384 // See the comment in coordinator_yield()
6385 for (unsigned i = 0; i < CMSYieldSleepCount &&
6386 ConcurrentMarkSweepThread::should_yield() &&
6387 !CMSCollector::foregroundGCIsActive(); ++i) {
6388 os::sleep(Thread::current(), 1, false);
6389 }
6390
6391 ConcurrentMarkSweepThread::synchronize(true);
6392 _freelistLock->lock_without_safepoint_check();
6393 _bitMap->lock()->lock_without_safepoint_check();
6394 _collector->startTimer();
6395 }
6396
6397
6398 //////////////////////////////////////////////////////////////////
6399 // SurvivorSpacePrecleanClosure
6400 //////////////////////////////////////////////////////////////////
6401 // This (single-threaded) closure is used to preclean the oops in
6402 // the survivor spaces.
6403 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6404
6405 HeapWord* addr = (HeapWord*)p;
6406 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6407 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6408 assert(p->klass_or_null() != NULL, "object should be initialized");
6409 // an initialized object; ignore mark word in verification below
6410 // since we are running concurrent with mutators
6411 assert(p->is_oop(true), "should be an oop");
6412 // Note that we do not yield while we iterate over
6413 // the interior oops of p, pushing the relevant ones
6414 // on our marking stack.
6415 size_t size = p->oop_iterate(_scanning_closure);
6416 do_yield_check();
6417 // Observe that below, we do not abandon the preclean
6418 // phase as soon as we should; rather we empty the
6419 // marking stack before returning. This is to satisfy
6420 // some existing assertions. In general, it may be a
6421 // good idea to abort immediately and complete the marking
6422 // from the grey objects at a later time.
6423 while (!_mark_stack->isEmpty()) {
6424 oop new_oop = _mark_stack->pop();
6425 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6426 assert(_bit_map->isMarked((HeapWord*)new_oop),
6427 "only grey objects on this stack");
6428 // iterate over the oops in this oop, marking and pushing
6429 // the ones in CMS heap (i.e. in _span).
6430 new_oop->oop_iterate(_scanning_closure);
6431 // check if it's time to yield
6432 do_yield_check();
6433 }
6434 unsigned int after_count =
6435 GenCollectedHeap::heap()->total_collections();
6436 bool abort = (_before_count != after_count) ||
6437 _collector->should_abort_preclean();
6438 return abort ? 0 : size;
6439 }
6440
6441 void SurvivorSpacePrecleanClosure::do_yield_work() {
6442 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6443 "CMS thread should hold CMS token");
6444 assert_lock_strong(_bit_map->lock());
6445 // Relinquish the bit map lock
6446 _bit_map->lock()->unlock();
6447 ConcurrentMarkSweepThread::desynchronize(true);
6448 _collector->stopTimer();
6449 if (PrintCMSStatistics != 0) {
6450 _collector->incrementYields();
6451 }
6452
6453 // See the comment in coordinator_yield()
6454 for (unsigned i = 0; i < CMSYieldSleepCount &&
6455 ConcurrentMarkSweepThread::should_yield() &&
6456 !CMSCollector::foregroundGCIsActive(); ++i) {
6457 os::sleep(Thread::current(), 1, false);
6458 }
6459
6460 ConcurrentMarkSweepThread::synchronize(true);
6461 _bit_map->lock()->lock_without_safepoint_check();
6462 _collector->startTimer();
6463 }
6464
6465 // This closure is used to rescan the marked objects on the dirty cards
6466 // in the mod union table and the card table proper. In the parallel
6467 // case, although the bitMap is shared, we do a single read so the
6468 // isMarked() query is "safe".
6469 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6470 // Ignore mark word because we are running concurrent with mutators
6471 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6472 HeapWord* addr = (HeapWord*)p;
6473 assert(_span.contains(addr), "we are scanning the CMS generation");
6474 bool is_obj_array = false;
6475 #ifdef ASSERT
6476 if (!_parallel) {
6477 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6478 assert(_collector->overflow_list_is_empty(),
6479 "overflow list should be empty");
6480
6481 }
6482 #endif // ASSERT
6483 if (_bit_map->isMarked(addr)) {
6484 // Obj arrays are precisely marked, non-arrays are not;
6485 // so we scan objArrays precisely and non-arrays in their
6486 // entirety.
6487 if (p->is_objArray()) {
6488 is_obj_array = true;
6489 if (_parallel) {
6490 p->oop_iterate(_par_scan_closure, mr);
6491 } else {
6492 p->oop_iterate(_scan_closure, mr);
6493 }
6494 } else {
6495 if (_parallel) {
6496 p->oop_iterate(_par_scan_closure);
6497 } else {
6498 p->oop_iterate(_scan_closure);
6499 }
6500 }
6501 }
6502 #ifdef ASSERT
6503 if (!_parallel) {
6504 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6505 assert(_collector->overflow_list_is_empty(),
6506 "overflow list should be empty");
6507
6508 }
6509 #endif // ASSERT
6510 return is_obj_array;
6511 }
6512
6513 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6514 MemRegion span,
6515 CMSBitMap* bitMap, CMSMarkStack* markStack,
6516 bool should_yield, bool verifying):
6517 _collector(collector),
6518 _span(span),
6519 _bitMap(bitMap),
6520 _mut(&collector->_modUnionTable),
6521 _markStack(markStack),
6522 _yield(should_yield),
6523 _skipBits(0)
6524 {
6525 assert(_markStack->isEmpty(), "stack should be empty");
6526 _finger = _bitMap->startWord();
6527 _threshold = _finger;
6528 assert(_collector->_restart_addr == NULL, "Sanity check");
6529 assert(_span.contains(_finger), "Out of bounds _finger?");
6530 DEBUG_ONLY(_verifying = verifying;)
6531 }
6532
6533 void MarkFromRootsClosure::reset(HeapWord* addr) {
6534 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6535 assert(_span.contains(addr), "Out of bounds _finger?");
6536 _finger = addr;
6537 _threshold = (HeapWord*)round_to(
6538 (intptr_t)_finger, CardTableModRefBS::card_size);
6539 }
6540
6541 // Should revisit to see if this should be restructured for
6542 // greater efficiency.
6543 bool MarkFromRootsClosure::do_bit(size_t offset) {
6544 if (_skipBits > 0) {
6545 _skipBits--;
6546 return true;
6547 }
6548 // convert offset into a HeapWord*
6549 HeapWord* addr = _bitMap->startWord() + offset;
6550 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6551 "address out of range");
6552 assert(_bitMap->isMarked(addr), "tautology");
6553 if (_bitMap->isMarked(addr+1)) {
6554 // this is an allocated but not yet initialized object
6555 assert(_skipBits == 0, "tautology");
6556 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6557 oop p = oop(addr);
6558 if (p->klass_or_null() == NULL) {
6559 DEBUG_ONLY(if (!_verifying) {)
6560 // We re-dirty the cards on which this object lies and increase
6561 // the _threshold so that we'll come back to scan this object
6562 // during the preclean or remark phase. (CMSCleanOnEnter)
6563 if (CMSCleanOnEnter) {
6564 size_t sz = _collector->block_size_using_printezis_bits(addr);
6565 HeapWord* end_card_addr = (HeapWord*)round_to(
6566 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6567 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6568 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6569 // Bump _threshold to end_card_addr; note that
6570 // _threshold cannot possibly exceed end_card_addr, anyhow.
6571 // This prevents future clearing of the card as the scan proceeds
6572 // to the right.
6573 assert(_threshold <= end_card_addr,
6574 "Because we are just scanning into this object");
6575 if (_threshold < end_card_addr) {
6576 _threshold = end_card_addr;
6577 }
6578 if (p->klass_or_null() != NULL) {
6579 // Redirty the range of cards...
6580 _mut->mark_range(redirty_range);
6581 } // ...else the setting of klass will dirty the card anyway.
6582 }
6583 DEBUG_ONLY(})
6584 return true;
6585 }
6586 }
6587 scanOopsInOop(addr);
6588 return true;
6589 }
6590
6591 // We take a break if we've been at this for a while,
6592 // so as to avoid monopolizing the locks involved.
6593 void MarkFromRootsClosure::do_yield_work() {
6594 // First give up the locks, then yield, then re-lock
6595 // We should probably use a constructor/destructor idiom to
6596 // do this unlock/lock or modify the MutexUnlocker class to
6597 // serve our purpose. XXX
6598 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6599 "CMS thread should hold CMS token");
6600 assert_lock_strong(_bitMap->lock());
6601 _bitMap->lock()->unlock();
6602 ConcurrentMarkSweepThread::desynchronize(true);
6603 _collector->stopTimer();
6604 if (PrintCMSStatistics != 0) {
6605 _collector->incrementYields();
6606 }
6607
6608 // See the comment in coordinator_yield()
6609 for (unsigned i = 0; i < CMSYieldSleepCount &&
6610 ConcurrentMarkSweepThread::should_yield() &&
6611 !CMSCollector::foregroundGCIsActive(); ++i) {
6612 os::sleep(Thread::current(), 1, false);
6613 }
6614
6615 ConcurrentMarkSweepThread::synchronize(true);
6616 _bitMap->lock()->lock_without_safepoint_check();
6617 _collector->startTimer();
6618 }
6619
6620 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6621 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6622 assert(_markStack->isEmpty(),
6623 "should drain stack to limit stack usage");
6624 // convert ptr to an oop preparatory to scanning
6625 oop obj = oop(ptr);
6626 // Ignore mark word in verification below, since we
6627 // may be running concurrent with mutators.
6628 assert(obj->is_oop(true), "should be an oop");
6629 assert(_finger <= ptr, "_finger runneth ahead");
6630 // advance the finger to right end of this object
6631 _finger = ptr + obj->size();
6632 assert(_finger > ptr, "we just incremented it above");
6633 // On large heaps, it may take us some time to get through
6634 // the marking phase. During
6635 // this time it's possible that a lot of mutations have
6636 // accumulated in the card table and the mod union table --
6637 // these mutation records are redundant until we have
6638 // actually traced into the corresponding card.
6639 // Here, we check whether advancing the finger would make
6640 // us cross into a new card, and if so clear corresponding
6641 // cards in the MUT (preclean them in the card-table in the
6642 // future).
6643
6644 DEBUG_ONLY(if (!_verifying) {)
6645 // The clean-on-enter optimization is disabled by default,
6646 // until we fix 6178663.
6647 if (CMSCleanOnEnter && (_finger > _threshold)) {
6648 // [_threshold, _finger) represents the interval
6649 // of cards to be cleared in MUT (or precleaned in card table).
6650 // The set of cards to be cleared is all those that overlap
6651 // with the interval [_threshold, _finger); note that
6652 // _threshold is always kept card-aligned but _finger isn't
6653 // always card-aligned.
6654 HeapWord* old_threshold = _threshold;
6655 assert(old_threshold == (HeapWord*)round_to(
6656 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6657 "_threshold should always be card-aligned");
6658 _threshold = (HeapWord*)round_to(
6659 (intptr_t)_finger, CardTableModRefBS::card_size);
6660 MemRegion mr(old_threshold, _threshold);
6661 assert(!mr.is_empty(), "Control point invariant");
6662 assert(_span.contains(mr), "Should clear within span");
6663 _mut->clear_range(mr);
6664 }
6665 DEBUG_ONLY(})
6666 // Note: the finger doesn't advance while we drain
6667 // the stack below.
6668 PushOrMarkClosure pushOrMarkClosure(_collector,
6669 _span, _bitMap, _markStack,
6670 _finger, this);
6671 bool res = _markStack->push(obj);
6672 assert(res, "Empty non-zero size stack should have space for single push");
6673 while (!_markStack->isEmpty()) {
6674 oop new_oop = _markStack->pop();
6675 // Skip verifying header mark word below because we are
6676 // running concurrent with mutators.
6677 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6678 // now scan this oop's oops
6679 new_oop->oop_iterate(&pushOrMarkClosure);
6680 do_yield_check();
6681 }
6682 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6683 }
6684
6685 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6686 CMSCollector* collector, MemRegion span,
6687 CMSBitMap* bit_map,
6688 OopTaskQueue* work_queue,
6689 CMSMarkStack* overflow_stack):
6690 _collector(collector),
6691 _whole_span(collector->_span),
6692 _span(span),
6693 _bit_map(bit_map),
6694 _mut(&collector->_modUnionTable),
6695 _work_queue(work_queue),
6696 _overflow_stack(overflow_stack),
6697 _skip_bits(0),
6698 _task(task)
6699 {
6700 assert(_work_queue->size() == 0, "work_queue should be empty");
6701 _finger = span.start();
6702 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6703 assert(_span.contains(_finger), "Out of bounds _finger?");
6704 }
6705
6706 // Should revisit to see if this should be restructured for
6707 // greater efficiency.
6708 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6709 if (_skip_bits > 0) {
6710 _skip_bits--;
6711 return true;
6712 }
6713 // convert offset into a HeapWord*
6714 HeapWord* addr = _bit_map->startWord() + offset;
6715 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6716 "address out of range");
6717 assert(_bit_map->isMarked(addr), "tautology");
6718 if (_bit_map->isMarked(addr+1)) {
6719 // this is an allocated object that might not yet be initialized
6720 assert(_skip_bits == 0, "tautology");
6721 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6722 oop p = oop(addr);
6723 if (p->klass_or_null() == NULL) {
6724 // in the case of Clean-on-Enter optimization, redirty card
6725 // and avoid clearing card by increasing the threshold.
6726 return true;
6727 }
6728 }
6729 scan_oops_in_oop(addr);
6730 return true;
6731 }
6732
6733 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6734 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6735 // Should we assert that our work queue is empty or
6736 // below some drain limit?
6737 assert(_work_queue->size() == 0,
6738 "should drain stack to limit stack usage");
6739 // convert ptr to an oop preparatory to scanning
6740 oop obj = oop(ptr);
6741 // Ignore mark word in verification below, since we
6742 // may be running concurrent with mutators.
6743 assert(obj->is_oop(true), "should be an oop");
6744 assert(_finger <= ptr, "_finger runneth ahead");
6745 // advance the finger to right end of this object
6746 _finger = ptr + obj->size();
6747 assert(_finger > ptr, "we just incremented it above");
6748 // On large heaps, it may take us some time to get through
6749 // the marking phase. During
6750 // this time it's possible that a lot of mutations have
6751 // accumulated in the card table and the mod union table --
6752 // these mutation records are redundant until we have
6753 // actually traced into the corresponding card.
6754 // Here, we check whether advancing the finger would make
6755 // us cross into a new card, and if so clear corresponding
6756 // cards in the MUT (preclean them in the card-table in the
6757 // future).
6758
6759 // The clean-on-enter optimization is disabled by default,
6760 // until we fix 6178663.
6761 if (CMSCleanOnEnter && (_finger > _threshold)) {
6762 // [_threshold, _finger) represents the interval
6763 // of cards to be cleared in MUT (or precleaned in card table).
6764 // The set of cards to be cleared is all those that overlap
6765 // with the interval [_threshold, _finger); note that
6766 // _threshold is always kept card-aligned but _finger isn't
6767 // always card-aligned.
6768 HeapWord* old_threshold = _threshold;
6769 assert(old_threshold == (HeapWord*)round_to(
6770 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6771 "_threshold should always be card-aligned");
6772 _threshold = (HeapWord*)round_to(
6773 (intptr_t)_finger, CardTableModRefBS::card_size);
6774 MemRegion mr(old_threshold, _threshold);
6775 assert(!mr.is_empty(), "Control point invariant");
6776 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6777 _mut->clear_range(mr);
6778 }
6779
6780 // Note: the local finger doesn't advance while we drain
6781 // the stack below, but the global finger sure can and will.
6782 HeapWord** gfa = _task->global_finger_addr();
6783 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6784 _span, _bit_map,
6785 _work_queue,
6786 _overflow_stack,
6787 _finger,
6788 gfa, this);
6789 bool res = _work_queue->push(obj); // overflow could occur here
6790 assert(res, "Will hold once we use workqueues");
6791 while (true) {
6792 oop new_oop;
6793 if (!_work_queue->pop_local(new_oop)) {
6794 // We emptied our work_queue; check if there's stuff that can
6795 // be gotten from the overflow stack.
6796 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6797 _overflow_stack, _work_queue)) {
6798 do_yield_check();
6799 continue;
6800 } else { // done
6801 break;
6802 }
6803 }
6804 // Skip verifying header mark word below because we are
6805 // running concurrent with mutators.
6806 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6807 // now scan this oop's oops
6808 new_oop->oop_iterate(&pushOrMarkClosure);
6809 do_yield_check();
6810 }
6811 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6812 }
6813
6814 // Yield in response to a request from VM Thread or
6815 // from mutators.
6816 void Par_MarkFromRootsClosure::do_yield_work() {
6817 assert(_task != NULL, "sanity");
6818 _task->yield();
6819 }
6820
6821 // A variant of the above used for verifying CMS marking work.
6822 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6823 MemRegion span,
6824 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6825 CMSMarkStack* mark_stack):
6826 _collector(collector),
6827 _span(span),
6828 _verification_bm(verification_bm),
6829 _cms_bm(cms_bm),
6830 _mark_stack(mark_stack),
6831 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6832 mark_stack)
6833 {
6834 assert(_mark_stack->isEmpty(), "stack should be empty");
6835 _finger = _verification_bm->startWord();
6836 assert(_collector->_restart_addr == NULL, "Sanity check");
6837 assert(_span.contains(_finger), "Out of bounds _finger?");
6838 }
6839
6840 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6841 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6842 assert(_span.contains(addr), "Out of bounds _finger?");
6843 _finger = addr;
6844 }
6845
6846 // Should revisit to see if this should be restructured for
6847 // greater efficiency.
6848 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6849 // convert offset into a HeapWord*
6850 HeapWord* addr = _verification_bm->startWord() + offset;
6851 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6852 "address out of range");
6853 assert(_verification_bm->isMarked(addr), "tautology");
6854 assert(_cms_bm->isMarked(addr), "tautology");
6855
6856 assert(_mark_stack->isEmpty(),
6857 "should drain stack to limit stack usage");
6858 // convert addr to an oop preparatory to scanning
6859 oop obj = oop(addr);
6860 assert(obj->is_oop(), "should be an oop");
6861 assert(_finger <= addr, "_finger runneth ahead");
6862 // advance the finger to right end of this object
6863 _finger = addr + obj->size();
6864 assert(_finger > addr, "we just incremented it above");
6865 // Note: the finger doesn't advance while we drain
6866 // the stack below.
6867 bool res = _mark_stack->push(obj);
6868 assert(res, "Empty non-zero size stack should have space for single push");
6869 while (!_mark_stack->isEmpty()) {
6870 oop new_oop = _mark_stack->pop();
6871 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6872 // now scan this oop's oops
6873 new_oop->oop_iterate(&_pam_verify_closure);
6874 }
6875 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6876 return true;
6877 }
6878
6879 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6880 CMSCollector* collector, MemRegion span,
6881 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6882 CMSMarkStack* mark_stack):
6883 MetadataAwareOopClosure(collector->ref_processor()),
6884 _collector(collector),
6885 _span(span),
6886 _verification_bm(verification_bm),
6887 _cms_bm(cms_bm),
6888 _mark_stack(mark_stack)
6889 { }
6890
6891 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6892 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6893
6894 // Upon stack overflow, we discard (part of) the stack,
6895 // remembering the least address amongst those discarded
6896 // in CMSCollector's _restart_address.
6897 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6898 // Remember the least grey address discarded
6899 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6900 _collector->lower_restart_addr(ra);
6901 _mark_stack->reset(); // discard stack contents
6902 _mark_stack->expand(); // expand the stack if possible
6903 }
6904
6905 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6906 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6907 HeapWord* addr = (HeapWord*)obj;
6908 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6909 // Oop lies in _span and isn't yet grey or black
6910 _verification_bm->mark(addr); // now grey
6911 if (!_cms_bm->isMarked(addr)) {
6912 oop(addr)->print();
6913 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6914 p2i(addr));
6915 fatal("... aborting");
6916 }
6917
6918 if (!_mark_stack->push(obj)) { // stack overflow
6919 if (PrintCMSStatistics != 0) {
6920 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6921 SIZE_FORMAT, _mark_stack->capacity());
6922 }
6923 assert(_mark_stack->isFull(), "Else push should have succeeded");
6924 handle_stack_overflow(addr);
6925 }
6926 // anything including and to the right of _finger
6927 // will be scanned as we iterate over the remainder of the
6928 // bit map
6929 }
6930 }
6931
6932 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6933 MemRegion span,
6934 CMSBitMap* bitMap, CMSMarkStack* markStack,
6935 HeapWord* finger, MarkFromRootsClosure* parent) :
6936 MetadataAwareOopClosure(collector->ref_processor()),
6937 _collector(collector),
6938 _span(span),
6939 _bitMap(bitMap),
6940 _markStack(markStack),
6941 _finger(finger),
6942 _parent(parent)
6943 { }
6944
6945 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6946 MemRegion span,
6947 CMSBitMap* bit_map,
6948 OopTaskQueue* work_queue,
6949 CMSMarkStack* overflow_stack,
6950 HeapWord* finger,
6951 HeapWord** global_finger_addr,
6952 Par_MarkFromRootsClosure* parent) :
6953 MetadataAwareOopClosure(collector->ref_processor()),
6954 _collector(collector),
6955 _whole_span(collector->_span),
6956 _span(span),
6957 _bit_map(bit_map),
6958 _work_queue(work_queue),
6959 _overflow_stack(overflow_stack),
6960 _finger(finger),
6961 _global_finger_addr(global_finger_addr),
6962 _parent(parent)
6963 { }
6964
6965 // Assumes thread-safe access by callers, who are
6966 // responsible for mutual exclusion.
6967 void CMSCollector::lower_restart_addr(HeapWord* low) {
6968 assert(_span.contains(low), "Out of bounds addr");
6969 if (_restart_addr == NULL) {
6970 _restart_addr = low;
6971 } else {
6972 _restart_addr = MIN2(_restart_addr, low);
6973 }
6974 }
6975
6976 // Upon stack overflow, we discard (part of) the stack,
6977 // remembering the least address amongst those discarded
6978 // in CMSCollector's _restart_address.
6979 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6980 // Remember the least grey address discarded
6981 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6982 _collector->lower_restart_addr(ra);
6983 _markStack->reset(); // discard stack contents
6984 _markStack->expand(); // expand the stack if possible
6985 }
6986
6987 // Upon stack overflow, we discard (part of) the stack,
6988 // remembering the least address amongst those discarded
6989 // in CMSCollector's _restart_address.
6990 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6991 // We need to do this under a mutex to prevent other
6992 // workers from interfering with the work done below.
6993 MutexLockerEx ml(_overflow_stack->par_lock(),
6994 Mutex::_no_safepoint_check_flag);
6995 // Remember the least grey address discarded
6996 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6997 _collector->lower_restart_addr(ra);
6998 _overflow_stack->reset(); // discard stack contents
6999 _overflow_stack->expand(); // expand the stack if possible
7000 }
7001
7002 void PushOrMarkClosure::do_oop(oop obj) {
7003 // Ignore mark word because we are running concurrent with mutators.
7004 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7005 HeapWord* addr = (HeapWord*)obj;
7006 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7007 // Oop lies in _span and isn't yet grey or black
7008 _bitMap->mark(addr); // now grey
7009 if (addr < _finger) {
7010 // the bit map iteration has already either passed, or
7011 // sampled, this bit in the bit map; we'll need to
7012 // use the marking stack to scan this oop's oops.
7013 bool simulate_overflow = false;
7014 NOT_PRODUCT(
7015 if (CMSMarkStackOverflowALot &&
7016 _collector->simulate_overflow()) {
7017 // simulate a stack overflow
7018 simulate_overflow = true;
7019 }
7020 )
7021 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7022 if (PrintCMSStatistics != 0) {
7023 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7024 SIZE_FORMAT, _markStack->capacity());
7025 }
7026 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7027 handle_stack_overflow(addr);
7028 }
7029 }
7030 // anything including and to the right of _finger
7031 // will be scanned as we iterate over the remainder of the
7032 // bit map
7033 do_yield_check();
7034 }
7035 }
7036
7037 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7038 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7039
7040 void Par_PushOrMarkClosure::do_oop(oop obj) {
7041 // Ignore mark word because we are running concurrent with mutators.
7042 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7043 HeapWord* addr = (HeapWord*)obj;
7044 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7045 // Oop lies in _span and isn't yet grey or black
7046 // We read the global_finger (volatile read) strictly after marking oop
7047 bool res = _bit_map->par_mark(addr); // now grey
7048 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7049 // Should we push this marked oop on our stack?
7050 // -- if someone else marked it, nothing to do
7051 // -- if target oop is above global finger nothing to do
7052 // -- if target oop is in chunk and above local finger
7053 // then nothing to do
7054 // -- else push on work queue
7055 if ( !res // someone else marked it, they will deal with it
7056 || (addr >= *gfa) // will be scanned in a later task
7057 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7058 return;
7059 }
7060 // the bit map iteration has already either passed, or
7061 // sampled, this bit in the bit map; we'll need to
7062 // use the marking stack to scan this oop's oops.
7063 bool simulate_overflow = false;
7064 NOT_PRODUCT(
7065 if (CMSMarkStackOverflowALot &&
7066 _collector->simulate_overflow()) {
7067 // simulate a stack overflow
7068 simulate_overflow = true;
7069 }
7070 )
7071 if (simulate_overflow ||
7072 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7073 // stack overflow
7074 if (PrintCMSStatistics != 0) {
7075 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7076 SIZE_FORMAT, _overflow_stack->capacity());
7077 }
7078 // We cannot assert that the overflow stack is full because
7079 // it may have been emptied since.
7080 assert(simulate_overflow ||
7081 _work_queue->size() == _work_queue->max_elems(),
7082 "Else push should have succeeded");
7083 handle_stack_overflow(addr);
7084 }
7085 do_yield_check();
7086 }
7087 }
7088
7089 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7090 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7091
7092 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7093 MemRegion span,
7094 ReferenceProcessor* rp,
7095 CMSBitMap* bit_map,
7096 CMSBitMap* mod_union_table,
7097 CMSMarkStack* mark_stack,
7098 bool concurrent_precleaning):
7099 MetadataAwareOopClosure(rp),
7100 _collector(collector),
7101 _span(span),
7102 _bit_map(bit_map),
7103 _mod_union_table(mod_union_table),
7104 _mark_stack(mark_stack),
7105 _concurrent_precleaning(concurrent_precleaning)
7106 {
7107 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7108 }
7109
7110 // Grey object rescan during pre-cleaning and second checkpoint phases --
7111 // the non-parallel version (the parallel version appears further below.)
7112 void PushAndMarkClosure::do_oop(oop obj) {
7113 // Ignore mark word verification. If during concurrent precleaning,
7114 // the object monitor may be locked. If during the checkpoint
7115 // phases, the object may already have been reached by a different
7116 // path and may be at the end of the global overflow list (so
7117 // the mark word may be NULL).
7118 assert(obj->is_oop_or_null(true /* ignore mark word */),
7119 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7120 HeapWord* addr = (HeapWord*)obj;
7121 // Check if oop points into the CMS generation
7122 // and is not marked
7123 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7124 // a white object ...
7125 _bit_map->mark(addr); // ... now grey
7126 // push on the marking stack (grey set)
7127 bool simulate_overflow = false;
7128 NOT_PRODUCT(
7129 if (CMSMarkStackOverflowALot &&
7130 _collector->simulate_overflow()) {
7131 // simulate a stack overflow
7132 simulate_overflow = true;
7133 }
7134 )
7135 if (simulate_overflow || !_mark_stack->push(obj)) {
7136 if (_concurrent_precleaning) {
7137 // During precleaning we can just dirty the appropriate card(s)
7138 // in the mod union table, thus ensuring that the object remains
7139 // in the grey set and continue. In the case of object arrays
7140 // we need to dirty all of the cards that the object spans,
7141 // since the rescan of object arrays will be limited to the
7142 // dirty cards.
7143 // Note that no one can be interfering with us in this action
7144 // of dirtying the mod union table, so no locking or atomics
7145 // are required.
7146 if (obj->is_objArray()) {
7147 size_t sz = obj->size();
7148 HeapWord* end_card_addr = (HeapWord*)round_to(
7149 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7150 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7151 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7152 _mod_union_table->mark_range(redirty_range);
7153 } else {
7154 _mod_union_table->mark(addr);
7155 }
7156 _collector->_ser_pmc_preclean_ovflw++;
7157 } else {
7158 // During the remark phase, we need to remember this oop
7159 // in the overflow list.
7160 _collector->push_on_overflow_list(obj);
7161 _collector->_ser_pmc_remark_ovflw++;
7162 }
7163 }
7164 }
7165 }
7166
7167 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7168 MemRegion span,
7169 ReferenceProcessor* rp,
7170 CMSBitMap* bit_map,
7171 OopTaskQueue* work_queue):
7172 MetadataAwareOopClosure(rp),
7173 _collector(collector),
7174 _span(span),
7175 _bit_map(bit_map),
7176 _work_queue(work_queue)
7177 {
7178 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7179 }
7180
7181 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7182 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7183
7184 // Grey object rescan during second checkpoint phase --
7185 // the parallel version.
7186 void Par_PushAndMarkClosure::do_oop(oop obj) {
7187 // In the assert below, we ignore the mark word because
7188 // this oop may point to an already visited object that is
7189 // on the overflow stack (in which case the mark word has
7190 // been hijacked for chaining into the overflow stack --
7191 // if this is the last object in the overflow stack then
7192 // its mark word will be NULL). Because this object may
7193 // have been subsequently popped off the global overflow
7194 // stack, and the mark word possibly restored to the prototypical
7195 // value, by the time we get to examined this failing assert in
7196 // the debugger, is_oop_or_null(false) may subsequently start
7197 // to hold.
7198 assert(obj->is_oop_or_null(true),
7199 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7200 HeapWord* addr = (HeapWord*)obj;
7201 // Check if oop points into the CMS generation
7202 // and is not marked
7203 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7204 // a white object ...
7205 // If we manage to "claim" the object, by being the
7206 // first thread to mark it, then we push it on our
7207 // marking stack
7208 if (_bit_map->par_mark(addr)) { // ... now grey
7209 // push on work queue (grey set)
7210 bool simulate_overflow = false;
7211 NOT_PRODUCT(
7212 if (CMSMarkStackOverflowALot &&
7213 _collector->par_simulate_overflow()) {
7214 // simulate a stack overflow
7215 simulate_overflow = true;
7216 }
7217 )
7218 if (simulate_overflow || !_work_queue->push(obj)) {
7219 _collector->par_push_on_overflow_list(obj);
7220 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7221 }
7222 } // Else, some other thread got there first
7223 }
7224 }
7225
7226 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7227 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7228
7229 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7230 Mutex* bml = _collector->bitMapLock();
7231 assert_lock_strong(bml);
7232 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7233 "CMS thread should hold CMS token");
7234
7235 bml->unlock();
7236 ConcurrentMarkSweepThread::desynchronize(true);
7237
7238 _collector->stopTimer();
7239 if (PrintCMSStatistics != 0) {
7240 _collector->incrementYields();
7241 }
7242
7243 // See the comment in coordinator_yield()
7244 for (unsigned i = 0; i < CMSYieldSleepCount &&
7245 ConcurrentMarkSweepThread::should_yield() &&
7246 !CMSCollector::foregroundGCIsActive(); ++i) {
7247 os::sleep(Thread::current(), 1, false);
7248 }
7249
7250 ConcurrentMarkSweepThread::synchronize(true);
7251 bml->lock();
7252
7253 _collector->startTimer();
7254 }
7255
7256 bool CMSPrecleanRefsYieldClosure::should_return() {
7257 if (ConcurrentMarkSweepThread::should_yield()) {
7258 do_yield_work();
7259 }
7260 return _collector->foregroundGCIsActive();
7261 }
7262
7263 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7264 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7265 "mr should be aligned to start at a card boundary");
7266 // We'd like to assert:
7267 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7268 // "mr should be a range of cards");
7269 // However, that would be too strong in one case -- the last
7270 // partition ends at _unallocated_block which, in general, can be
7271 // an arbitrary boundary, not necessarily card aligned.
7272 if (PrintCMSStatistics != 0) {
7273 _num_dirty_cards +=
7274 mr.word_size()/CardTableModRefBS::card_size_in_words;
7275 }
7276 _space->object_iterate_mem(mr, &_scan_cl);
7277 }
7278
7279 SweepClosure::SweepClosure(CMSCollector* collector,
7280 ConcurrentMarkSweepGeneration* g,
7281 CMSBitMap* bitMap, bool should_yield) :
7282 _collector(collector),
7283 _g(g),
7284 _sp(g->cmsSpace()),
7285 _limit(_sp->sweep_limit()),
7286 _freelistLock(_sp->freelistLock()),
7287 _bitMap(bitMap),
7288 _yield(should_yield),
7289 _inFreeRange(false), // No free range at beginning of sweep
7290 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7291 _lastFreeRangeCoalesced(false),
7292 _freeFinger(g->used_region().start())
7293 {
7294 NOT_PRODUCT(
7295 _numObjectsFreed = 0;
7296 _numWordsFreed = 0;
7297 _numObjectsLive = 0;
7298 _numWordsLive = 0;
7299 _numObjectsAlreadyFree = 0;
7300 _numWordsAlreadyFree = 0;
7301 _last_fc = NULL;
7302
7303 _sp->initializeIndexedFreeListArrayReturnedBytes();
7304 _sp->dictionary()->initialize_dict_returned_bytes();
7305 )
7306 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7307 "sweep _limit out of bounds");
7308 if (CMSTraceSweeper) {
7309 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7310 p2i(_limit));
7311 }
7312 }
7313
7314 void SweepClosure::print_on(outputStream* st) const {
7315 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7316 p2i(_sp->bottom()), p2i(_sp->end()));
7317 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7318 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7319 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7320 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7321 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7322 }
7323
7324 #ifndef PRODUCT
7325 // Assertion checking only: no useful work in product mode --
7326 // however, if any of the flags below become product flags,
7327 // you may need to review this code to see if it needs to be
7328 // enabled in product mode.
7329 SweepClosure::~SweepClosure() {
7330 assert_lock_strong(_freelistLock);
7331 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7332 "sweep _limit out of bounds");
7333 if (inFreeRange()) {
7334 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7335 print();
7336 ShouldNotReachHere();
7337 }
7338 if (Verbose && PrintGC) {
7339 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7340 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7341 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7342 SIZE_FORMAT" bytes "
7343 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7344 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7345 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7346 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7347 * sizeof(HeapWord);
7348 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7349
7350 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7351 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7352 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7353 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7354 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
7355 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7356 indexListReturnedBytes);
7357 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7358 dict_returned_bytes);
7359 }
7360 }
7361 if (CMSTraceSweeper) {
7362 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7363 p2i(_limit));
7364 }
7365 }
7366 #endif // PRODUCT
7367
7368 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7369 bool freeRangeInFreeLists) {
7370 if (CMSTraceSweeper) {
7371 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7372 p2i(freeFinger), freeRangeInFreeLists);
7373 }
7374 assert(!inFreeRange(), "Trampling existing free range");
7375 set_inFreeRange(true);
7376 set_lastFreeRangeCoalesced(false);
7377
7378 set_freeFinger(freeFinger);
7379 set_freeRangeInFreeLists(freeRangeInFreeLists);
7380 if (CMSTestInFreeList) {
7381 if (freeRangeInFreeLists) {
7382 FreeChunk* fc = (FreeChunk*) freeFinger;
7383 assert(fc->is_free(), "A chunk on the free list should be free.");
7384 assert(fc->size() > 0, "Free range should have a size");
7385 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7386 }
7387 }
7388 }
7389
7390 // Note that the sweeper runs concurrently with mutators. Thus,
7391 // it is possible for direct allocation in this generation to happen
7392 // in the middle of the sweep. Note that the sweeper also coalesces
7393 // contiguous free blocks. Thus, unless the sweeper and the allocator
7394 // synchronize appropriately freshly allocated blocks may get swept up.
7395 // This is accomplished by the sweeper locking the free lists while
7396 // it is sweeping. Thus blocks that are determined to be free are
7397 // indeed free. There is however one additional complication:
7398 // blocks that have been allocated since the final checkpoint and
7399 // mark, will not have been marked and so would be treated as
7400 // unreachable and swept up. To prevent this, the allocator marks
7401 // the bit map when allocating during the sweep phase. This leads,
7402 // however, to a further complication -- objects may have been allocated
7403 // but not yet initialized -- in the sense that the header isn't yet
7404 // installed. The sweeper can not then determine the size of the block
7405 // in order to skip over it. To deal with this case, we use a technique
7406 // (due to Printezis) to encode such uninitialized block sizes in the
7407 // bit map. Since the bit map uses a bit per every HeapWord, but the
7408 // CMS generation has a minimum object size of 3 HeapWords, it follows
7409 // that "normal marks" won't be adjacent in the bit map (there will
7410 // always be at least two 0 bits between successive 1 bits). We make use
7411 // of these "unused" bits to represent uninitialized blocks -- the bit
7412 // corresponding to the start of the uninitialized object and the next
7413 // bit are both set. Finally, a 1 bit marks the end of the object that
7414 // started with the two consecutive 1 bits to indicate its potentially
7415 // uninitialized state.
7416
7417 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7418 FreeChunk* fc = (FreeChunk*)addr;
7419 size_t res;
7420
7421 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7422 // than "addr == _limit" because although _limit was a block boundary when
7423 // we started the sweep, it may no longer be one because heap expansion
7424 // may have caused us to coalesce the block ending at the address _limit
7425 // with a newly expanded chunk (this happens when _limit was set to the
7426 // previous _end of the space), so we may have stepped past _limit:
7427 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7428 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7429 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7430 "sweep _limit out of bounds");
7431 assert(addr < _sp->end(), "addr out of bounds");
7432 // Flush any free range we might be holding as a single
7433 // coalesced chunk to the appropriate free list.
7434 if (inFreeRange()) {
7435 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7436 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", p2i(freeFinger())));
7437 flush_cur_free_chunk(freeFinger(),
7438 pointer_delta(addr, freeFinger()));
7439 if (CMSTraceSweeper) {
7440 gclog_or_tty->print("Sweep: last chunk: ");
7441 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") "
7442 "[coalesced:%d]\n",
7443 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7444 lastFreeRangeCoalesced() ? 1 : 0);
7445 }
7446 }
7447
7448 // help the iterator loop finish
7449 return pointer_delta(_sp->end(), addr);
7450 }
7451
7452 assert(addr < _limit, "sweep invariant");
7453 // check if we should yield
7454 do_yield_check(addr);
7455 if (fc->is_free()) {
7456 // Chunk that is already free
7457 res = fc->size();
7458 do_already_free_chunk(fc);
7459 debug_only(_sp->verifyFreeLists());
7460 // If we flush the chunk at hand in lookahead_and_flush()
7461 // and it's coalesced with a preceding chunk, then the
7462 // process of "mangling" the payload of the coalesced block
7463 // will cause erasure of the size information from the
7464 // (erstwhile) header of all the coalesced blocks but the
7465 // first, so the first disjunct in the assert will not hold
7466 // in that specific case (in which case the second disjunct
7467 // will hold).
7468 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7469 "Otherwise the size info doesn't change at this step");
7470 NOT_PRODUCT(
7471 _numObjectsAlreadyFree++;
7472 _numWordsAlreadyFree += res;
7473 )
7474 NOT_PRODUCT(_last_fc = fc;)
7475 } else if (!_bitMap->isMarked(addr)) {
7476 // Chunk is fresh garbage
7477 res = do_garbage_chunk(fc);
7478 debug_only(_sp->verifyFreeLists());
7479 NOT_PRODUCT(
7480 _numObjectsFreed++;
7481 _numWordsFreed += res;
7482 )
7483 } else {
7484 // Chunk that is alive.
7485 res = do_live_chunk(fc);
7486 debug_only(_sp->verifyFreeLists());
7487 NOT_PRODUCT(
7488 _numObjectsLive++;
7489 _numWordsLive += res;
7490 )
7491 }
7492 return res;
7493 }
7494
7495 // For the smart allocation, record following
7496 // split deaths - a free chunk is removed from its free list because
7497 // it is being split into two or more chunks.
7498 // split birth - a free chunk is being added to its free list because
7499 // a larger free chunk has been split and resulted in this free chunk.
7500 // coal death - a free chunk is being removed from its free list because
7501 // it is being coalesced into a large free chunk.
7502 // coal birth - a free chunk is being added to its free list because
7503 // it was created when two or more free chunks where coalesced into
7504 // this free chunk.
7505 //
7506 // These statistics are used to determine the desired number of free
7507 // chunks of a given size. The desired number is chosen to be relative
7508 // to the end of a CMS sweep. The desired number at the end of a sweep
7509 // is the
7510 // count-at-end-of-previous-sweep (an amount that was enough)
7511 // - count-at-beginning-of-current-sweep (the excess)
7512 // + split-births (gains in this size during interval)
7513 // - split-deaths (demands on this size during interval)
7514 // where the interval is from the end of one sweep to the end of the
7515 // next.
7516 //
7517 // When sweeping the sweeper maintains an accumulated chunk which is
7518 // the chunk that is made up of chunks that have been coalesced. That
7519 // will be termed the left-hand chunk. A new chunk of garbage that
7520 // is being considered for coalescing will be referred to as the
7521 // right-hand chunk.
7522 //
7523 // When making a decision on whether to coalesce a right-hand chunk with
7524 // the current left-hand chunk, the current count vs. the desired count
7525 // of the left-hand chunk is considered. Also if the right-hand chunk
7526 // is near the large chunk at the end of the heap (see
7527 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7528 // left-hand chunk is coalesced.
7529 //
7530 // When making a decision about whether to split a chunk, the desired count
7531 // vs. the current count of the candidate to be split is also considered.
7532 // If the candidate is underpopulated (currently fewer chunks than desired)
7533 // a chunk of an overpopulated (currently more chunks than desired) size may
7534 // be chosen. The "hint" associated with a free list, if non-null, points
7535 // to a free list which may be overpopulated.
7536 //
7537
7538 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7539 const size_t size = fc->size();
7540 // Chunks that cannot be coalesced are not in the
7541 // free lists.
7542 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7543 assert(_sp->verify_chunk_in_free_list(fc),
7544 "free chunk should be in free lists");
7545 }
7546 // a chunk that is already free, should not have been
7547 // marked in the bit map
7548 HeapWord* const addr = (HeapWord*) fc;
7549 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7550 // Verify that the bit map has no bits marked between
7551 // addr and purported end of this block.
7552 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7553
7554 // Some chunks cannot be coalesced under any circumstances.
7555 // See the definition of cantCoalesce().
7556 if (!fc->cantCoalesce()) {
7557 // This chunk can potentially be coalesced.
7558 if (_sp->adaptive_freelists()) {
7559 // All the work is done in
7560 do_post_free_or_garbage_chunk(fc, size);
7561 } else { // Not adaptive free lists
7562 // this is a free chunk that can potentially be coalesced by the sweeper;
7563 if (!inFreeRange()) {
7564 // if the next chunk is a free block that can't be coalesced
7565 // it doesn't make sense to remove this chunk from the free lists
7566 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7567 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7568 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
7569 nextChunk->is_free() && // ... which is free...
7570 nextChunk->cantCoalesce()) { // ... but can't be coalesced
7571 // nothing to do
7572 } else {
7573 // Potentially the start of a new free range:
7574 // Don't eagerly remove it from the free lists.
7575 // No need to remove it if it will just be put
7576 // back again. (Also from a pragmatic point of view
7577 // if it is a free block in a region that is beyond
7578 // any allocated blocks, an assertion will fail)
7579 // Remember the start of a free run.
7580 initialize_free_range(addr, true);
7581 // end - can coalesce with next chunk
7582 }
7583 } else {
7584 // the midst of a free range, we are coalescing
7585 print_free_block_coalesced(fc);
7586 if (CMSTraceSweeper) {
7587 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7588 }
7589 // remove it from the free lists
7590 _sp->removeFreeChunkFromFreeLists(fc);
7591 set_lastFreeRangeCoalesced(true);
7592 // If the chunk is being coalesced and the current free range is
7593 // in the free lists, remove the current free range so that it
7594 // will be returned to the free lists in its entirety - all
7595 // the coalesced pieces included.
7596 if (freeRangeInFreeLists()) {
7597 FreeChunk* ffc = (FreeChunk*) freeFinger();
7598 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7599 "Size of free range is inconsistent with chunk size.");
7600 if (CMSTestInFreeList) {
7601 assert(_sp->verify_chunk_in_free_list(ffc),
7602 "free range is not in free lists");
7603 }
7604 _sp->removeFreeChunkFromFreeLists(ffc);
7605 set_freeRangeInFreeLists(false);
7606 }
7607 }
7608 }
7609 // Note that if the chunk is not coalescable (the else arm
7610 // below), we unconditionally flush, without needing to do
7611 // a "lookahead," as we do below.
7612 if (inFreeRange()) lookahead_and_flush(fc, size);
7613 } else {
7614 // Code path common to both original and adaptive free lists.
7615
7616 // cant coalesce with previous block; this should be treated
7617 // as the end of a free run if any
7618 if (inFreeRange()) {
7619 // we kicked some butt; time to pick up the garbage
7620 assert(freeFinger() < addr, "freeFinger points too high");
7621 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7622 }
7623 // else, nothing to do, just continue
7624 }
7625 }
7626
7627 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7628 // This is a chunk of garbage. It is not in any free list.
7629 // Add it to a free list or let it possibly be coalesced into
7630 // a larger chunk.
7631 HeapWord* const addr = (HeapWord*) fc;
7632 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7633
7634 if (_sp->adaptive_freelists()) {
7635 // Verify that the bit map has no bits marked between
7636 // addr and purported end of just dead object.
7637 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7638
7639 do_post_free_or_garbage_chunk(fc, size);
7640 } else {
7641 if (!inFreeRange()) {
7642 // start of a new free range
7643 assert(size > 0, "A free range should have a size");
7644 initialize_free_range(addr, false);
7645 } else {
7646 // this will be swept up when we hit the end of the
7647 // free range
7648 if (CMSTraceSweeper) {
7649 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7650 }
7651 // If the chunk is being coalesced and the current free range is
7652 // in the free lists, remove the current free range so that it
7653 // will be returned to the free lists in its entirety - all
7654 // the coalesced pieces included.
7655 if (freeRangeInFreeLists()) {
7656 FreeChunk* ffc = (FreeChunk*)freeFinger();
7657 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7658 "Size of free range is inconsistent with chunk size.");
7659 if (CMSTestInFreeList) {
7660 assert(_sp->verify_chunk_in_free_list(ffc),
7661 "free range is not in free lists");
7662 }
7663 _sp->removeFreeChunkFromFreeLists(ffc);
7664 set_freeRangeInFreeLists(false);
7665 }
7666 set_lastFreeRangeCoalesced(true);
7667 }
7668 // this will be swept up when we hit the end of the free range
7669
7670 // Verify that the bit map has no bits marked between
7671 // addr and purported end of just dead object.
7672 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7673 }
7674 assert(_limit >= addr + size,
7675 "A freshly garbage chunk can't possibly straddle over _limit");
7676 if (inFreeRange()) lookahead_and_flush(fc, size);
7677 return size;
7678 }
7679
7680 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7681 HeapWord* addr = (HeapWord*) fc;
7682 // The sweeper has just found a live object. Return any accumulated
7683 // left hand chunk to the free lists.
7684 if (inFreeRange()) {
7685 assert(freeFinger() < addr, "freeFinger points too high");
7686 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7687 }
7688
7689 // This object is live: we'd normally expect this to be
7690 // an oop, and like to assert the following:
7691 // assert(oop(addr)->is_oop(), "live block should be an oop");
7692 // However, as we commented above, this may be an object whose
7693 // header hasn't yet been initialized.
7694 size_t size;
7695 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7696 if (_bitMap->isMarked(addr + 1)) {
7697 // Determine the size from the bit map, rather than trying to
7698 // compute it from the object header.
7699 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7700 size = pointer_delta(nextOneAddr + 1, addr);
7701 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7702 "alignment problem");
7703
7704 #ifdef ASSERT
7705 if (oop(addr)->klass_or_null() != NULL) {
7706 // Ignore mark word because we are running concurrent with mutators
7707 assert(oop(addr)->is_oop(true), "live block should be an oop");
7708 assert(size ==
7709 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7710 "P-mark and computed size do not agree");
7711 }
7712 #endif
7713
7714 } else {
7715 // This should be an initialized object that's alive.
7716 assert(oop(addr)->klass_or_null() != NULL,
7717 "Should be an initialized object");
7718 // Ignore mark word because we are running concurrent with mutators
7719 assert(oop(addr)->is_oop(true), "live block should be an oop");
7720 // Verify that the bit map has no bits marked between
7721 // addr and purported end of this block.
7722 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7723 assert(size >= 3, "Necessary for Printezis marks to work");
7724 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7725 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7726 }
7727 return size;
7728 }
7729
7730 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7731 size_t chunkSize) {
7732 // do_post_free_or_garbage_chunk() should only be called in the case
7733 // of the adaptive free list allocator.
7734 const bool fcInFreeLists = fc->is_free();
7735 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7736 assert((HeapWord*)fc <= _limit, "sweep invariant");
7737 if (CMSTestInFreeList && fcInFreeLists) {
7738 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7739 }
7740
7741 if (CMSTraceSweeper) {
7742 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7743 }
7744
7745 HeapWord* const fc_addr = (HeapWord*) fc;
7746
7747 bool coalesce;
7748 const size_t left = pointer_delta(fc_addr, freeFinger());
7749 const size_t right = chunkSize;
7750 switch (FLSCoalescePolicy) {
7751 // numeric value forms a coalition aggressiveness metric
7752 case 0: { // never coalesce
7753 coalesce = false;
7754 break;
7755 }
7756 case 1: { // coalesce if left & right chunks on overpopulated lists
7757 coalesce = _sp->coalOverPopulated(left) &&
7758 _sp->coalOverPopulated(right);
7759 break;
7760 }
7761 case 2: { // coalesce if left chunk on overpopulated list (default)
7762 coalesce = _sp->coalOverPopulated(left);
7763 break;
7764 }
7765 case 3: { // coalesce if left OR right chunk on overpopulated list
7766 coalesce = _sp->coalOverPopulated(left) ||
7767 _sp->coalOverPopulated(right);
7768 break;
7769 }
7770 case 4: { // always coalesce
7771 coalesce = true;
7772 break;
7773 }
7774 default:
7775 ShouldNotReachHere();
7776 }
7777
7778 // Should the current free range be coalesced?
7779 // If the chunk is in a free range and either we decided to coalesce above
7780 // or the chunk is near the large block at the end of the heap
7781 // (isNearLargestChunk() returns true), then coalesce this chunk.
7782 const bool doCoalesce = inFreeRange()
7783 && (coalesce || _g->isNearLargestChunk(fc_addr));
7784 if (doCoalesce) {
7785 // Coalesce the current free range on the left with the new
7786 // chunk on the right. If either is on a free list,
7787 // it must be removed from the list and stashed in the closure.
7788 if (freeRangeInFreeLists()) {
7789 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7790 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7791 "Size of free range is inconsistent with chunk size.");
7792 if (CMSTestInFreeList) {
7793 assert(_sp->verify_chunk_in_free_list(ffc),
7794 "Chunk is not in free lists");
7795 }
7796 _sp->coalDeath(ffc->size());
7797 _sp->removeFreeChunkFromFreeLists(ffc);
7798 set_freeRangeInFreeLists(false);
7799 }
7800 if (fcInFreeLists) {
7801 _sp->coalDeath(chunkSize);
7802 assert(fc->size() == chunkSize,
7803 "The chunk has the wrong size or is not in the free lists");
7804 _sp->removeFreeChunkFromFreeLists(fc);
7805 }
7806 set_lastFreeRangeCoalesced(true);
7807 print_free_block_coalesced(fc);
7808 } else { // not in a free range and/or should not coalesce
7809 // Return the current free range and start a new one.
7810 if (inFreeRange()) {
7811 // In a free range but cannot coalesce with the right hand chunk.
7812 // Put the current free range into the free lists.
7813 flush_cur_free_chunk(freeFinger(),
7814 pointer_delta(fc_addr, freeFinger()));
7815 }
7816 // Set up for new free range. Pass along whether the right hand
7817 // chunk is in the free lists.
7818 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7819 }
7820 }
7821
7822 // Lookahead flush:
7823 // If we are tracking a free range, and this is the last chunk that
7824 // we'll look at because its end crosses past _limit, we'll preemptively
7825 // flush it along with any free range we may be holding on to. Note that
7826 // this can be the case only for an already free or freshly garbage
7827 // chunk. If this block is an object, it can never straddle
7828 // over _limit. The "straddling" occurs when _limit is set at
7829 // the previous end of the space when this cycle started, and
7830 // a subsequent heap expansion caused the previously co-terminal
7831 // free block to be coalesced with the newly expanded portion,
7832 // thus rendering _limit a non-block-boundary making it dangerous
7833 // for the sweeper to step over and examine.
7834 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7835 assert(inFreeRange(), "Should only be called if currently in a free range.");
7836 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7837 assert(_sp->used_region().contains(eob - 1),
7838 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7839 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7840 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7841 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size));
7842 if (eob >= _limit) {
7843 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7844 if (CMSTraceSweeper) {
7845 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7846 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7847 "[" PTR_FORMAT "," PTR_FORMAT ")",
7848 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7849 }
7850 // Return the storage we are tracking back into the free lists.
7851 if (CMSTraceSweeper) {
7852 gclog_or_tty->print_cr("Flushing ... ");
7853 }
7854 assert(freeFinger() < eob, "Error");
7855 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7856 }
7857 }
7858
7859 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7860 assert(inFreeRange(), "Should only be called if currently in a free range.");
7861 assert(size > 0,
7862 "A zero sized chunk cannot be added to the free lists.");
7863 if (!freeRangeInFreeLists()) {
7864 if (CMSTestInFreeList) {
7865 FreeChunk* fc = (FreeChunk*) chunk;
7866 fc->set_size(size);
7867 assert(!_sp->verify_chunk_in_free_list(fc),
7868 "chunk should not be in free lists yet");
7869 }
7870 if (CMSTraceSweeper) {
7871 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7872 p2i(chunk), size);
7873 }
7874 // A new free range is going to be starting. The current
7875 // free range has not been added to the free lists yet or
7876 // was removed so add it back.
7877 // If the current free range was coalesced, then the death
7878 // of the free range was recorded. Record a birth now.
7879 if (lastFreeRangeCoalesced()) {
7880 _sp->coalBirth(size);
7881 }
7882 _sp->addChunkAndRepairOffsetTable(chunk, size,
7883 lastFreeRangeCoalesced());
7884 } else if (CMSTraceSweeper) {
7885 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7886 }
7887 set_inFreeRange(false);
7888 set_freeRangeInFreeLists(false);
7889 }
7890
7891 // We take a break if we've been at this for a while,
7892 // so as to avoid monopolizing the locks involved.
7893 void SweepClosure::do_yield_work(HeapWord* addr) {
7894 // Return current free chunk being used for coalescing (if any)
7895 // to the appropriate freelist. After yielding, the next
7896 // free block encountered will start a coalescing range of
7897 // free blocks. If the next free block is adjacent to the
7898 // chunk just flushed, they will need to wait for the next
7899 // sweep to be coalesced.
7900 if (inFreeRange()) {
7901 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7902 }
7903
7904 // First give up the locks, then yield, then re-lock.
7905 // We should probably use a constructor/destructor idiom to
7906 // do this unlock/lock or modify the MutexUnlocker class to
7907 // serve our purpose. XXX
7908 assert_lock_strong(_bitMap->lock());
7909 assert_lock_strong(_freelistLock);
7910 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7911 "CMS thread should hold CMS token");
7912 _bitMap->lock()->unlock();
7913 _freelistLock->unlock();
7914 ConcurrentMarkSweepThread::desynchronize(true);
7915 _collector->stopTimer();
7916 if (PrintCMSStatistics != 0) {
7917 _collector->incrementYields();
7918 }
7919
7920 // See the comment in coordinator_yield()
7921 for (unsigned i = 0; i < CMSYieldSleepCount &&
7922 ConcurrentMarkSweepThread::should_yield() &&
7923 !CMSCollector::foregroundGCIsActive(); ++i) {
7924 os::sleep(Thread::current(), 1, false);
7925 }
7926
7927 ConcurrentMarkSweepThread::synchronize(true);
7928 _freelistLock->lock();
7929 _bitMap->lock()->lock_without_safepoint_check();
7930 _collector->startTimer();
7931 }
7932
7933 #ifndef PRODUCT
7934 // This is actually very useful in a product build if it can
7935 // be called from the debugger. Compile it into the product
7936 // as needed.
7937 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7938 return debug_cms_space->verify_chunk_in_free_list(fc);
7939 }
7940 #endif
7941
7942 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7943 if (CMSTraceSweeper) {
7944 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7945 p2i(fc), fc->size());
7946 }
7947 }
7948
7949 // CMSIsAliveClosure
7950 bool CMSIsAliveClosure::do_object_b(oop obj) {
7951 HeapWord* addr = (HeapWord*)obj;
7952 return addr != NULL &&
7953 (!_span.contains(addr) || _bit_map->isMarked(addr));
7954 }
7955
7956
7957 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7958 MemRegion span,
7959 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7960 bool cpc):
7961 _collector(collector),
7962 _span(span),
7963 _bit_map(bit_map),
7964 _mark_stack(mark_stack),
7965 _concurrent_precleaning(cpc) {
7966 assert(!_span.is_empty(), "Empty span could spell trouble");
7967 }
7968
7969
7970 // CMSKeepAliveClosure: the serial version
7971 void CMSKeepAliveClosure::do_oop(oop obj) {
7972 HeapWord* addr = (HeapWord*)obj;
7973 if (_span.contains(addr) &&
7974 !_bit_map->isMarked(addr)) {
7975 _bit_map->mark(addr);
7976 bool simulate_overflow = false;
7977 NOT_PRODUCT(
7978 if (CMSMarkStackOverflowALot &&
7979 _collector->simulate_overflow()) {
7980 // simulate a stack overflow
7981 simulate_overflow = true;
7982 }
7983 )
7984 if (simulate_overflow || !_mark_stack->push(obj)) {
7985 if (_concurrent_precleaning) {
7986 // We dirty the overflown object and let the remark
7987 // phase deal with it.
7988 assert(_collector->overflow_list_is_empty(), "Error");
7989 // In the case of object arrays, we need to dirty all of
7990 // the cards that the object spans. No locking or atomics
7991 // are needed since no one else can be mutating the mod union
7992 // table.
7993 if (obj->is_objArray()) {
7994 size_t sz = obj->size();
7995 HeapWord* end_card_addr =
7996 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7997 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7998 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7999 _collector->_modUnionTable.mark_range(redirty_range);
8000 } else {
8001 _collector->_modUnionTable.mark(addr);
8002 }
8003 _collector->_ser_kac_preclean_ovflw++;
8004 } else {
8005 _collector->push_on_overflow_list(obj);
8006 _collector->_ser_kac_ovflw++;
8007 }
8008 }
8009 }
8010 }
8011
8012 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8013 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8014
8015 // CMSParKeepAliveClosure: a parallel version of the above.
8016 // The work queues are private to each closure (thread),
8017 // but (may be) available for stealing by other threads.
8018 void CMSParKeepAliveClosure::do_oop(oop obj) {
8019 HeapWord* addr = (HeapWord*)obj;
8020 if (_span.contains(addr) &&
8021 !_bit_map->isMarked(addr)) {
8022 // In general, during recursive tracing, several threads
8023 // may be concurrently getting here; the first one to
8024 // "tag" it, claims it.
8025 if (_bit_map->par_mark(addr)) {
8026 bool res = _work_queue->push(obj);
8027 assert(res, "Low water mark should be much less than capacity");
8028 // Do a recursive trim in the hope that this will keep
8029 // stack usage lower, but leave some oops for potential stealers
8030 trim_queue(_low_water_mark);
8031 } // Else, another thread got there first
8032 }
8033 }
8034
8035 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8036 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8037
8038 void CMSParKeepAliveClosure::trim_queue(uint max) {
8039 while (_work_queue->size() > max) {
8040 oop new_oop;
8041 if (_work_queue->pop_local(new_oop)) {
8042 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8043 assert(_bit_map->isMarked((HeapWord*)new_oop),
8044 "no white objects on this stack!");
8045 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8046 // iterate over the oops in this oop, marking and pushing
8047 // the ones in CMS heap (i.e. in _span).
8048 new_oop->oop_iterate(&_mark_and_push);
8049 }
8050 }
8051 }
8052
8053 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8054 CMSCollector* collector,
8055 MemRegion span, CMSBitMap* bit_map,
8056 OopTaskQueue* work_queue):
8057 _collector(collector),
8058 _span(span),
8059 _bit_map(bit_map),
8060 _work_queue(work_queue) { }
8061
8062 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8063 HeapWord* addr = (HeapWord*)obj;
8064 if (_span.contains(addr) &&
8065 !_bit_map->isMarked(addr)) {
8066 if (_bit_map->par_mark(addr)) {
8067 bool simulate_overflow = false;
8068 NOT_PRODUCT(
8069 if (CMSMarkStackOverflowALot &&
8070 _collector->par_simulate_overflow()) {
8071 // simulate a stack overflow
8072 simulate_overflow = true;
8073 }
8074 )
8075 if (simulate_overflow || !_work_queue->push(obj)) {
8076 _collector->par_push_on_overflow_list(obj);
8077 _collector->_par_kac_ovflw++;
8078 }
8079 } // Else another thread got there already
8080 }
8081 }
8082
8083 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8084 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8085
8086 //////////////////////////////////////////////////////////////////
8087 // CMSExpansionCause /////////////////////////////
8088 //////////////////////////////////////////////////////////////////
8089 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8090 switch (cause) {
8091 case _no_expansion:
8092 return "No expansion";
8093 case _satisfy_free_ratio:
8094 return "Free ratio";
8095 case _satisfy_promotion:
8096 return "Satisfy promotion";
8097 case _satisfy_allocation:
8098 return "allocation";
8099 case _allocate_par_lab:
8100 return "Par LAB";
8101 case _allocate_par_spooling_space:
8102 return "Par Spooling Space";
8103 case _adaptive_size_policy:
8104 return "Ergonomics";
8105 default:
8106 return "unknown";
8107 }
8108 }
8109
8110 void CMSDrainMarkingStackClosure::do_void() {
8111 // the max number to take from overflow list at a time
8112 const size_t num = _mark_stack->capacity()/4;
8113 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8114 "Overflow list should be NULL during concurrent phases");
8115 while (!_mark_stack->isEmpty() ||
8116 // if stack is empty, check the overflow list
8117 _collector->take_from_overflow_list(num, _mark_stack)) {
8118 oop obj = _mark_stack->pop();
8119 HeapWord* addr = (HeapWord*)obj;
8120 assert(_span.contains(addr), "Should be within span");
8121 assert(_bit_map->isMarked(addr), "Should be marked");
8122 assert(obj->is_oop(), "Should be an oop");
8123 obj->oop_iterate(_keep_alive);
8124 }
8125 }
8126
8127 void CMSParDrainMarkingStackClosure::do_void() {
8128 // drain queue
8129 trim_queue(0);
8130 }
8131
8132 // Trim our work_queue so its length is below max at return
8133 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8134 while (_work_queue->size() > max) {
8135 oop new_oop;
8136 if (_work_queue->pop_local(new_oop)) {
8137 assert(new_oop->is_oop(), "Expected an oop");
8138 assert(_bit_map->isMarked((HeapWord*)new_oop),
8139 "no white objects on this stack!");
8140 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8141 // iterate over the oops in this oop, marking and pushing
8142 // the ones in CMS heap (i.e. in _span).
8143 new_oop->oop_iterate(&_mark_and_push);
8144 }
8145 }
8146 }
8147
8148 ////////////////////////////////////////////////////////////////////
8149 // Support for Marking Stack Overflow list handling and related code
8150 ////////////////////////////////////////////////////////////////////
8151 // Much of the following code is similar in shape and spirit to the
8152 // code used in ParNewGC. We should try and share that code
8153 // as much as possible in the future.
8154
8155 #ifndef PRODUCT
8156 // Debugging support for CMSStackOverflowALot
8157
8158 // It's OK to call this multi-threaded; the worst thing
8159 // that can happen is that we'll get a bunch of closely
8160 // spaced simulated overflows, but that's OK, in fact
8161 // probably good as it would exercise the overflow code
8162 // under contention.
8163 bool CMSCollector::simulate_overflow() {
8164 if (_overflow_counter-- <= 0) { // just being defensive
8165 _overflow_counter = CMSMarkStackOverflowInterval;
8166 return true;
8167 } else {
8168 return false;
8169 }
8170 }
8171
8172 bool CMSCollector::par_simulate_overflow() {
8173 return simulate_overflow();
8174 }
8175 #endif
8176
8177 // Single-threaded
8178 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8179 assert(stack->isEmpty(), "Expected precondition");
8180 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8181 size_t i = num;
8182 oop cur = _overflow_list;
8183 const markOop proto = markOopDesc::prototype();
8184 NOT_PRODUCT(ssize_t n = 0;)
8185 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8186 next = oop(cur->mark());
8187 cur->set_mark(proto); // until proven otherwise
8188 assert(cur->is_oop(), "Should be an oop");
8189 bool res = stack->push(cur);
8190 assert(res, "Bit off more than can chew?");
8191 NOT_PRODUCT(n++;)
8192 }
8193 _overflow_list = cur;
8194 #ifndef PRODUCT
8195 assert(_num_par_pushes >= n, "Too many pops?");
8196 _num_par_pushes -=n;
8197 #endif
8198 return !stack->isEmpty();
8199 }
8200
8201 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8202 // (MT-safe) Get a prefix of at most "num" from the list.
8203 // The overflow list is chained through the mark word of
8204 // each object in the list. We fetch the entire list,
8205 // break off a prefix of the right size and return the
8206 // remainder. If other threads try to take objects from
8207 // the overflow list at that time, they will wait for
8208 // some time to see if data becomes available. If (and
8209 // only if) another thread places one or more object(s)
8210 // on the global list before we have returned the suffix
8211 // to the global list, we will walk down our local list
8212 // to find its end and append the global list to
8213 // our suffix before returning it. This suffix walk can
8214 // prove to be expensive (quadratic in the amount of traffic)
8215 // when there are many objects in the overflow list and
8216 // there is much producer-consumer contention on the list.
8217 // *NOTE*: The overflow list manipulation code here and
8218 // in ParNewGeneration:: are very similar in shape,
8219 // except that in the ParNew case we use the old (from/eden)
8220 // copy of the object to thread the list via its klass word.
8221 // Because of the common code, if you make any changes in
8222 // the code below, please check the ParNew version to see if
8223 // similar changes might be needed.
8224 // CR 6797058 has been filed to consolidate the common code.
8225 bool CMSCollector::par_take_from_overflow_list(size_t num,
8226 OopTaskQueue* work_q,
8227 int no_of_gc_threads) {
8228 assert(work_q->size() == 0, "First empty local work queue");
8229 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8230 if (_overflow_list == NULL) {
8231 return false;
8232 }
8233 // Grab the entire list; we'll put back a suffix
8234 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8235 Thread* tid = Thread::current();
8236 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8237 // set to ParallelGCThreads.
8238 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8239 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8240 // If the list is busy, we spin for a short while,
8241 // sleeping between attempts to get the list.
8242 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8243 os::sleep(tid, sleep_time_millis, false);
8244 if (_overflow_list == NULL) {
8245 // Nothing left to take
8246 return false;
8247 } else if (_overflow_list != BUSY) {
8248 // Try and grab the prefix
8249 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8250 }
8251 }
8252 // If the list was found to be empty, or we spun long
8253 // enough, we give up and return empty-handed. If we leave
8254 // the list in the BUSY state below, it must be the case that
8255 // some other thread holds the overflow list and will set it
8256 // to a non-BUSY state in the future.
8257 if (prefix == NULL || prefix == BUSY) {
8258 // Nothing to take or waited long enough
8259 if (prefix == NULL) {
8260 // Write back the NULL in case we overwrote it with BUSY above
8261 // and it is still the same value.
8262 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8263 }
8264 return false;
8265 }
8266 assert(prefix != NULL && prefix != BUSY, "Error");
8267 size_t i = num;
8268 oop cur = prefix;
8269 // Walk down the first "num" objects, unless we reach the end.
8270 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8271 if (cur->mark() == NULL) {
8272 // We have "num" or fewer elements in the list, so there
8273 // is nothing to return to the global list.
8274 // Write back the NULL in lieu of the BUSY we wrote
8275 // above, if it is still the same value.
8276 if (_overflow_list == BUSY) {
8277 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8278 }
8279 } else {
8280 // Chop off the suffix and return it to the global list.
8281 assert(cur->mark() != BUSY, "Error");
8282 oop suffix_head = cur->mark(); // suffix will be put back on global list
8283 cur->set_mark(NULL); // break off suffix
8284 // It's possible that the list is still in the empty(busy) state
8285 // we left it in a short while ago; in that case we may be
8286 // able to place back the suffix without incurring the cost
8287 // of a walk down the list.
8288 oop observed_overflow_list = _overflow_list;
8289 oop cur_overflow_list = observed_overflow_list;
8290 bool attached = false;
8291 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8292 observed_overflow_list =
8293 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8294 if (cur_overflow_list == observed_overflow_list) {
8295 attached = true;
8296 break;
8297 } else cur_overflow_list = observed_overflow_list;
8298 }
8299 if (!attached) {
8300 // Too bad, someone else sneaked in (at least) an element; we'll need
8301 // to do a splice. Find tail of suffix so we can prepend suffix to global
8302 // list.
8303 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8304 oop suffix_tail = cur;
8305 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8306 "Tautology");
8307 observed_overflow_list = _overflow_list;
8308 do {
8309 cur_overflow_list = observed_overflow_list;
8310 if (cur_overflow_list != BUSY) {
8311 // Do the splice ...
8312 suffix_tail->set_mark(markOop(cur_overflow_list));
8313 } else { // cur_overflow_list == BUSY
8314 suffix_tail->set_mark(NULL);
8315 }
8316 // ... and try to place spliced list back on overflow_list ...
8317 observed_overflow_list =
8318 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8319 } while (cur_overflow_list != observed_overflow_list);
8320 // ... until we have succeeded in doing so.
8321 }
8322 }
8323
8324 // Push the prefix elements on work_q
8325 assert(prefix != NULL, "control point invariant");
8326 const markOop proto = markOopDesc::prototype();
8327 oop next;
8328 NOT_PRODUCT(ssize_t n = 0;)
8329 for (cur = prefix; cur != NULL; cur = next) {
8330 next = oop(cur->mark());
8331 cur->set_mark(proto); // until proven otherwise
8332 assert(cur->is_oop(), "Should be an oop");
8333 bool res = work_q->push(cur);
8334 assert(res, "Bit off more than we can chew?");
8335 NOT_PRODUCT(n++;)
8336 }
8337 #ifndef PRODUCT
8338 assert(_num_par_pushes >= n, "Too many pops?");
8339 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8340 #endif
8341 return true;
8342 }
8343
8344 // Single-threaded
8345 void CMSCollector::push_on_overflow_list(oop p) {
8346 NOT_PRODUCT(_num_par_pushes++;)
8347 assert(p->is_oop(), "Not an oop");
8348 preserve_mark_if_necessary(p);
8349 p->set_mark((markOop)_overflow_list);
8350 _overflow_list = p;
8351 }
8352
8353 // Multi-threaded; use CAS to prepend to overflow list
8354 void CMSCollector::par_push_on_overflow_list(oop p) {
8355 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8356 assert(p->is_oop(), "Not an oop");
8357 par_preserve_mark_if_necessary(p);
8358 oop observed_overflow_list = _overflow_list;
8359 oop cur_overflow_list;
8360 do {
8361 cur_overflow_list = observed_overflow_list;
8362 if (cur_overflow_list != BUSY) {
8363 p->set_mark(markOop(cur_overflow_list));
8364 } else {
8365 p->set_mark(NULL);
8366 }
8367 observed_overflow_list =
8368 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8369 } while (cur_overflow_list != observed_overflow_list);
8370 }
8371 #undef BUSY
8372
8373 // Single threaded
8374 // General Note on GrowableArray: pushes may silently fail
8375 // because we are (temporarily) out of C-heap for expanding
8376 // the stack. The problem is quite ubiquitous and affects
8377 // a lot of code in the JVM. The prudent thing for GrowableArray
8378 // to do (for now) is to exit with an error. However, that may
8379 // be too draconian in some cases because the caller may be
8380 // able to recover without much harm. For such cases, we
8381 // should probably introduce a "soft_push" method which returns
8382 // an indication of success or failure with the assumption that
8383 // the caller may be able to recover from a failure; code in
8384 // the VM can then be changed, incrementally, to deal with such
8385 // failures where possible, thus, incrementally hardening the VM
8386 // in such low resource situations.
8387 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8388 _preserved_oop_stack.push(p);
8389 _preserved_mark_stack.push(m);
8390 assert(m == p->mark(), "Mark word changed");
8391 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8392 "bijection");
8393 }
8394
8395 // Single threaded
8396 void CMSCollector::preserve_mark_if_necessary(oop p) {
8397 markOop m = p->mark();
8398 if (m->must_be_preserved(p)) {
8399 preserve_mark_work(p, m);
8400 }
8401 }
8402
8403 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8404 markOop m = p->mark();
8405 if (m->must_be_preserved(p)) {
8406 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8407 // Even though we read the mark word without holding
8408 // the lock, we are assured that it will not change
8409 // because we "own" this oop, so no other thread can
8410 // be trying to push it on the overflow list; see
8411 // the assertion in preserve_mark_work() that checks
8412 // that m == p->mark().
8413 preserve_mark_work(p, m);
8414 }
8415 }
8416
8417 // We should be able to do this multi-threaded,
8418 // a chunk of stack being a task (this is
8419 // correct because each oop only ever appears
8420 // once in the overflow list. However, it's
8421 // not very easy to completely overlap this with
8422 // other operations, so will generally not be done
8423 // until all work's been completed. Because we
8424 // expect the preserved oop stack (set) to be small,
8425 // it's probably fine to do this single-threaded.
8426 // We can explore cleverer concurrent/overlapped/parallel
8427 // processing of preserved marks if we feel the
8428 // need for this in the future. Stack overflow should
8429 // be so rare in practice and, when it happens, its
8430 // effect on performance so great that this will
8431 // likely just be in the noise anyway.
8432 void CMSCollector::restore_preserved_marks_if_any() {
8433 assert(SafepointSynchronize::is_at_safepoint(),
8434 "world should be stopped");
8435 assert(Thread::current()->is_ConcurrentGC_thread() ||
8436 Thread::current()->is_VM_thread(),
8437 "should be single-threaded");
8438 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8439 "bijection");
8440
8441 while (!_preserved_oop_stack.is_empty()) {
8442 oop p = _preserved_oop_stack.pop();
8443 assert(p->is_oop(), "Should be an oop");
8444 assert(_span.contains(p), "oop should be in _span");
8445 assert(p->mark() == markOopDesc::prototype(),
8446 "Set when taken from overflow list");
8447 markOop m = _preserved_mark_stack.pop();
8448 p->set_mark(m);
8449 }
8450 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8451 "stacks were cleared above");
8452 }
8453
8454 #ifndef PRODUCT
8455 bool CMSCollector::no_preserved_marks() const {
8456 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8457 }
8458 #endif
8459
8460 // Transfer some number of overflown objects to usual marking
8461 // stack. Return true if some objects were transferred.
8462 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8463 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8464 (size_t)ParGCDesiredObjsFromOverflowList);
8465
8466 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8467 assert(_collector->overflow_list_is_empty() || res,
8468 "If list is not empty, we should have taken something");
8469 assert(!res || !_mark_stack->isEmpty(),
8470 "If we took something, it should now be on our stack");
8471 return res;
8472 }
8473
8474 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8475 size_t res = _sp->block_size_no_stall(addr, _collector);
8476 if (_sp->block_is_obj(addr)) {
8477 if (_live_bit_map->isMarked(addr)) {
8478 // It can't have been dead in a previous cycle
8479 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8480 } else {
8481 _dead_bit_map->mark(addr); // mark the dead object
8482 }
8483 }
8484 // Could be 0, if the block size could not be computed without stalling.
8485 return res;
8486 }
8487
8488 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8489
8490 switch (phase) {
8491 case CMSCollector::InitialMarking:
8492 initialize(true /* fullGC */ ,
8493 cause /* cause of the GC */,
8494 true /* recordGCBeginTime */,
8495 true /* recordPreGCUsage */,
8496 false /* recordPeakUsage */,
8497 false /* recordPostGCusage */,
8498 true /* recordAccumulatedGCTime */,
8499 false /* recordGCEndTime */,
8500 false /* countCollection */ );
8501 break;
8502
8503 case CMSCollector::FinalMarking:
8504 initialize(true /* fullGC */ ,
8505 cause /* cause of the GC */,
8506 false /* recordGCBeginTime */,
8507 false /* recordPreGCUsage */,
8508 false /* recordPeakUsage */,
8509 false /* recordPostGCusage */,
8510 true /* recordAccumulatedGCTime */,
8511 false /* recordGCEndTime */,
8512 false /* countCollection */ );
8513 break;
8514
8515 case CMSCollector::Sweeping:
8516 initialize(true /* fullGC */ ,
8517 cause /* cause of the GC */,
8518 false /* recordGCBeginTime */,
8519 false /* recordPreGCUsage */,
8520 true /* recordPeakUsage */,
8521 true /* recordPostGCusage */,
8522 false /* recordAccumulatedGCTime */,
8523 true /* recordGCEndTime */,
8524 true /* countCollection */ );
8525 break;
8526
8527 default:
8528 ShouldNotReachHere();
8529 }
8530 }