1 /*
   2  * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/classLoaderData.hpp"
  27 #include "classfile/stringTable.hpp"
  28 #include "classfile/systemDictionary.hpp"
  29 #include "code/codeCache.hpp"
  30 #include "gc/cms/cmsCollectorPolicy.hpp"
  31 #include "gc/cms/cmsOopClosures.inline.hpp"
  32 #include "gc/cms/compactibleFreeListSpace.hpp"
  33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
  34 #include "gc/cms/concurrentMarkSweepThread.hpp"
  35 #include "gc/cms/parNewGeneration.hpp"
  36 #include "gc/cms/vmCMSOperations.hpp"
  37 #include "gc/serial/genMarkSweep.hpp"
  38 #include "gc/serial/tenuredGeneration.hpp"
  39 #include "gc/shared/adaptiveSizePolicy.hpp"
  40 #include "gc/shared/cardGeneration.inline.hpp"
  41 #include "gc/shared/cardTableRS.hpp"
  42 #include "gc/shared/collectedHeap.inline.hpp"
  43 #include "gc/shared/collectorCounters.hpp"
  44 #include "gc/shared/collectorPolicy.hpp"
  45 #include "gc/shared/gcLocker.inline.hpp"
  46 #include "gc/shared/gcPolicyCounters.hpp"
  47 #include "gc/shared/gcTimer.hpp"
  48 #include "gc/shared/gcTrace.hpp"
  49 #include "gc/shared/gcTraceTime.hpp"
  50 #include "gc/shared/genCollectedHeap.hpp"
  51 #include "gc/shared/genOopClosures.inline.hpp"
  52 #include "gc/shared/isGCActiveMark.hpp"
  53 #include "gc/shared/referencePolicy.hpp"
  54 #include "gc/shared/strongRootsScope.hpp"
  55 #include "gc/shared/taskqueue.inline.hpp"
  56 #include "memory/allocation.hpp"
  57 #include "memory/iterator.inline.hpp"
  58 #include "memory/padded.hpp"
  59 #include "memory/resourceArea.hpp"
  60 #include "oops/oop.inline.hpp"
  61 #include "prims/jvmtiExport.hpp"
  62 #include "runtime/atomic.inline.hpp"
  63 #include "runtime/globals_extension.hpp"
  64 #include "runtime/handles.inline.hpp"
  65 #include "runtime/java.hpp"
  66 #include "runtime/orderAccess.inline.hpp"
  67 #include "runtime/vmThread.hpp"
  68 #include "services/memoryService.hpp"
  69 #include "services/runtimeService.hpp"
  70 #include "utilities/stack.inline.hpp"
  71 
  72 // statics
  73 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
  74 bool CMSCollector::_full_gc_requested = false;
  75 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
  76 
  77 //////////////////////////////////////////////////////////////////
  78 // In support of CMS/VM thread synchronization
  79 //////////////////////////////////////////////////////////////////
  80 // We split use of the CGC_lock into 2 "levels".
  81 // The low-level locking is of the usual CGC_lock monitor. We introduce
  82 // a higher level "token" (hereafter "CMS token") built on top of the
  83 // low level monitor (hereafter "CGC lock").
  84 // The token-passing protocol gives priority to the VM thread. The
  85 // CMS-lock doesn't provide any fairness guarantees, but clients
  86 // should ensure that it is only held for very short, bounded
  87 // durations.
  88 //
  89 // When either of the CMS thread or the VM thread is involved in
  90 // collection operations during which it does not want the other
  91 // thread to interfere, it obtains the CMS token.
  92 //
  93 // If either thread tries to get the token while the other has
  94 // it, that thread waits. However, if the VM thread and CMS thread
  95 // both want the token, then the VM thread gets priority while the
  96 // CMS thread waits. This ensures, for instance, that the "concurrent"
  97 // phases of the CMS thread's work do not block out the VM thread
  98 // for long periods of time as the CMS thread continues to hog
  99 // the token. (See bug 4616232).
 100 //
 101 // The baton-passing functions are, however, controlled by the
 102 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
 103 // and here the low-level CMS lock, not the high level token,
 104 // ensures mutual exclusion.
 105 //
 106 // Two important conditions that we have to satisfy:
 107 // 1. if a thread does a low-level wait on the CMS lock, then it
 108 //    relinquishes the CMS token if it were holding that token
 109 //    when it acquired the low-level CMS lock.
 110 // 2. any low-level notifications on the low-level lock
 111 //    should only be sent when a thread has relinquished the token.
 112 //
 113 // In the absence of either property, we'd have potential deadlock.
 114 //
 115 // We protect each of the CMS (concurrent and sequential) phases
 116 // with the CMS _token_, not the CMS _lock_.
 117 //
 118 // The only code protected by CMS lock is the token acquisition code
 119 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
 120 // baton-passing code.
 121 //
 122 // Unfortunately, i couldn't come up with a good abstraction to factor and
 123 // hide the naked CGC_lock manipulation in the baton-passing code
 124 // further below. That's something we should try to do. Also, the proof
 125 // of correctness of this 2-level locking scheme is far from obvious,
 126 // and potentially quite slippery. We have an uneasy suspicion, for instance,
 127 // that there may be a theoretical possibility of delay/starvation in the
 128 // low-level lock/wait/notify scheme used for the baton-passing because of
 129 // potential interference with the priority scheme embodied in the
 130 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
 131 // invocation further below and marked with "XXX 20011219YSR".
 132 // Indeed, as we note elsewhere, this may become yet more slippery
 133 // in the presence of multiple CMS and/or multiple VM threads. XXX
 134 
 135 class CMSTokenSync: public StackObj {
 136  private:
 137   bool _is_cms_thread;
 138  public:
 139   CMSTokenSync(bool is_cms_thread):
 140     _is_cms_thread(is_cms_thread) {
 141     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
 142            "Incorrect argument to constructor");
 143     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
 144   }
 145 
 146   ~CMSTokenSync() {
 147     assert(_is_cms_thread ?
 148              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
 149              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
 150           "Incorrect state");
 151     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
 152   }
 153 };
 154 
 155 // Convenience class that does a CMSTokenSync, and then acquires
 156 // upto three locks.
 157 class CMSTokenSyncWithLocks: public CMSTokenSync {
 158  private:
 159   // Note: locks are acquired in textual declaration order
 160   // and released in the opposite order
 161   MutexLockerEx _locker1, _locker2, _locker3;
 162  public:
 163   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
 164                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
 165     CMSTokenSync(is_cms_thread),
 166     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
 167     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
 168     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
 169   { }
 170 };
 171 
 172 
 173 //////////////////////////////////////////////////////////////////
 174 //  Concurrent Mark-Sweep Generation /////////////////////////////
 175 //////////////////////////////////////////////////////////////////
 176 
 177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
 178 
 179 // This struct contains per-thread things necessary to support parallel
 180 // young-gen collection.
 181 class CMSParGCThreadState: public CHeapObj<mtGC> {
 182  public:
 183   CFLS_LAB lab;
 184   PromotionInfo promo;
 185 
 186   // Constructor.
 187   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
 188     promo.setSpace(cfls);
 189   }
 190 };
 191 
 192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
 193      ReservedSpace rs, size_t initial_byte_size,
 194      CardTableRS* ct, bool use_adaptive_freelists,
 195      FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
 196   CardGeneration(rs, initial_byte_size, ct),
 197   _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
 198   _did_compact(false)
 199 {
 200   HeapWord* bottom = (HeapWord*) _virtual_space.low();
 201   HeapWord* end    = (HeapWord*) _virtual_space.high();
 202 
 203   _direct_allocated_words = 0;
 204   NOT_PRODUCT(
 205     _numObjectsPromoted = 0;
 206     _numWordsPromoted = 0;
 207     _numObjectsAllocated = 0;
 208     _numWordsAllocated = 0;
 209   )
 210 
 211   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
 212                                            use_adaptive_freelists,
 213                                            dictionaryChoice);
 214   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
 215   _cmsSpace->_gen = this;
 216 
 217   _gc_stats = new CMSGCStats();
 218 
 219   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
 220   // offsets match. The ability to tell free chunks from objects
 221   // depends on this property.
 222   debug_only(
 223     FreeChunk* junk = NULL;
 224     assert(UseCompressedClassPointers ||
 225            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
 226            "Offset of FreeChunk::_prev within FreeChunk must match"
 227            "  that of OopDesc::_klass within OopDesc");
 228   )
 229 
 230   _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
 231   for (uint i = 0; i < ParallelGCThreads; i++) {
 232     _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
 233   }
 234 
 235   _incremental_collection_failed = false;
 236   // The "dilatation_factor" is the expansion that can occur on
 237   // account of the fact that the minimum object size in the CMS
 238   // generation may be larger than that in, say, a contiguous young
 239   //  generation.
 240   // Ideally, in the calculation below, we'd compute the dilatation
 241   // factor as: MinChunkSize/(promoting_gen's min object size)
 242   // Since we do not have such a general query interface for the
 243   // promoting generation, we'll instead just use the minimum
 244   // object size (which today is a header's worth of space);
 245   // note that all arithmetic is in units of HeapWords.
 246   assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
 247   assert(_dilatation_factor >= 1.0, "from previous assert");
 248 }
 249 
 250 
 251 // The field "_initiating_occupancy" represents the occupancy percentage
 252 // at which we trigger a new collection cycle.  Unless explicitly specified
 253 // via CMSInitiatingOccupancyFraction (argument "io" below), it
 254 // is calculated by:
 255 //
 256 //   Let "f" be MinHeapFreeRatio in
 257 //
 258 //    _initiating_occupancy = 100-f +
 259 //                           f * (CMSTriggerRatio/100)
 260 //   where CMSTriggerRatio is the argument "tr" below.
 261 //
 262 // That is, if we assume the heap is at its desired maximum occupancy at the
 263 // end of a collection, we let CMSTriggerRatio of the (purported) free
 264 // space be allocated before initiating a new collection cycle.
 265 //
 266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
 267   assert(io <= 100 && tr <= 100, "Check the arguments");
 268   if (io >= 0) {
 269     _initiating_occupancy = (double)io / 100.0;
 270   } else {
 271     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
 272                              (double)(tr * MinHeapFreeRatio) / 100.0)
 273                             / 100.0;
 274   }
 275 }
 276 
 277 void ConcurrentMarkSweepGeneration::ref_processor_init() {
 278   assert(collector() != NULL, "no collector");
 279   collector()->ref_processor_init();
 280 }
 281 
 282 void CMSCollector::ref_processor_init() {
 283   if (_ref_processor == NULL) {
 284     // Allocate and initialize a reference processor
 285     _ref_processor =
 286       new ReferenceProcessor(_span,                               // span
 287                              (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
 288                              ParallelGCThreads,                   // mt processing degree
 289                              _cmsGen->refs_discovery_is_mt(),     // mt discovery
 290                              MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
 291                              _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
 292                              &_is_alive_closure);                 // closure for liveness info
 293     // Initialize the _ref_processor field of CMSGen
 294     _cmsGen->set_ref_processor(_ref_processor);
 295 
 296   }
 297 }
 298 
 299 AdaptiveSizePolicy* CMSCollector::size_policy() {
 300   GenCollectedHeap* gch = GenCollectedHeap::heap();
 301   return gch->gen_policy()->size_policy();
 302 }
 303 
 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
 305 
 306   const char* gen_name = "old";
 307   GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy();
 308 
 309   // Generation Counters - generation 1, 1 subspace
 310   _gen_counters = new GenerationCounters(gen_name, 1, 1,
 311       gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
 312 
 313   _space_counters = new GSpaceCounters(gen_name, 0,
 314                                        _virtual_space.reserved_size(),
 315                                        this, _gen_counters);
 316 }
 317 
 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
 319   _cms_gen(cms_gen)
 320 {
 321   assert(alpha <= 100, "bad value");
 322   _saved_alpha = alpha;
 323 
 324   // Initialize the alphas to the bootstrap value of 100.
 325   _gc0_alpha = _cms_alpha = 100;
 326 
 327   _cms_begin_time.update();
 328   _cms_end_time.update();
 329 
 330   _gc0_duration = 0.0;
 331   _gc0_period = 0.0;
 332   _gc0_promoted = 0;
 333 
 334   _cms_duration = 0.0;
 335   _cms_period = 0.0;
 336   _cms_allocated = 0;
 337 
 338   _cms_used_at_gc0_begin = 0;
 339   _cms_used_at_gc0_end = 0;
 340   _allow_duty_cycle_reduction = false;
 341   _valid_bits = 0;
 342 }
 343 
 344 double CMSStats::cms_free_adjustment_factor(size_t free) const {
 345   // TBD: CR 6909490
 346   return 1.0;
 347 }
 348 
 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
 350 }
 351 
 352 // If promotion failure handling is on use
 353 // the padded average size of the promotion for each
 354 // young generation collection.
 355 double CMSStats::time_until_cms_gen_full() const {
 356   size_t cms_free = _cms_gen->cmsSpace()->free();
 357   GenCollectedHeap* gch = GenCollectedHeap::heap();
 358   size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
 359                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
 360   if (cms_free > expected_promotion) {
 361     // Start a cms collection if there isn't enough space to promote
 362     // for the next minor collection.  Use the padded average as
 363     // a safety factor.
 364     cms_free -= expected_promotion;
 365 
 366     // Adjust by the safety factor.
 367     double cms_free_dbl = (double)cms_free;
 368     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
 369     // Apply a further correction factor which tries to adjust
 370     // for recent occurance of concurrent mode failures.
 371     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
 372     cms_free_dbl = cms_free_dbl * cms_adjustment;
 373 
 374     if (PrintGCDetails && Verbose) {
 375       gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
 376         SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
 377         cms_free, expected_promotion);
 378       gclog_or_tty->print_cr("  cms_free_dbl %f cms_consumption_rate %f",
 379         cms_free_dbl, cms_consumption_rate() + 1.0);
 380     }
 381     // Add 1 in case the consumption rate goes to zero.
 382     return cms_free_dbl / (cms_consumption_rate() + 1.0);
 383   }
 384   return 0.0;
 385 }
 386 
 387 // Compare the duration of the cms collection to the
 388 // time remaining before the cms generation is empty.
 389 // Note that the time from the start of the cms collection
 390 // to the start of the cms sweep (less than the total
 391 // duration of the cms collection) can be used.  This
 392 // has been tried and some applications experienced
 393 // promotion failures early in execution.  This was
 394 // possibly because the averages were not accurate
 395 // enough at the beginning.
 396 double CMSStats::time_until_cms_start() const {
 397   // We add "gc0_period" to the "work" calculation
 398   // below because this query is done (mostly) at the
 399   // end of a scavenge, so we need to conservatively
 400   // account for that much possible delay
 401   // in the query so as to avoid concurrent mode failures
 402   // due to starting the collection just a wee bit too
 403   // late.
 404   double work = cms_duration() + gc0_period();
 405   double deadline = time_until_cms_gen_full();
 406   // If a concurrent mode failure occurred recently, we want to be
 407   // more conservative and halve our expected time_until_cms_gen_full()
 408   if (work > deadline) {
 409     if (Verbose && PrintGCDetails) {
 410       gclog_or_tty->print(
 411         " CMSCollector: collect because of anticipated promotion "
 412         "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
 413         gc0_period(), time_until_cms_gen_full());
 414     }
 415     return 0.0;
 416   }
 417   return work - deadline;
 418 }
 419 
 420 #ifndef PRODUCT
 421 void CMSStats::print_on(outputStream *st) const {
 422   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
 423   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
 424                gc0_duration(), gc0_period(), gc0_promoted());
 425   st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
 426             cms_duration(), cms_period(), cms_allocated());
 427   st->print(",cms_since_beg=%g,cms_since_end=%g",
 428             cms_time_since_begin(), cms_time_since_end());
 429   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
 430             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
 431 
 432   if (valid()) {
 433     st->print(",promo_rate=%g,cms_alloc_rate=%g",
 434               promotion_rate(), cms_allocation_rate());
 435     st->print(",cms_consumption_rate=%g,time_until_full=%g",
 436               cms_consumption_rate(), time_until_cms_gen_full());
 437   }
 438   st->print(" ");
 439 }
 440 #endif // #ifndef PRODUCT
 441 
 442 CMSCollector::CollectorState CMSCollector::_collectorState =
 443                              CMSCollector::Idling;
 444 bool CMSCollector::_foregroundGCIsActive = false;
 445 bool CMSCollector::_foregroundGCShouldWait = false;
 446 
 447 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
 448                            CardTableRS*                   ct,
 449                            ConcurrentMarkSweepPolicy*     cp):
 450   _cmsGen(cmsGen),
 451   _ct(ct),
 452   _ref_processor(NULL),    // will be set later
 453   _conc_workers(NULL),     // may be set later
 454   _abort_preclean(false),
 455   _start_sampling(false),
 456   _between_prologue_and_epilogue(false),
 457   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
 458   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
 459                  -1 /* lock-free */, "No_lock" /* dummy */),
 460   _modUnionClosurePar(&_modUnionTable),
 461   // Adjust my span to cover old (cms) gen
 462   _span(cmsGen->reserved()),
 463   // Construct the is_alive_closure with _span & markBitMap
 464   _is_alive_closure(_span, &_markBitMap),
 465   _restart_addr(NULL),
 466   _overflow_list(NULL),
 467   _stats(cmsGen),
 468   _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
 469                              //verify that this lock should be acquired with safepoint check.
 470                              Monitor::_safepoint_check_sometimes)),
 471   _eden_chunk_array(NULL),     // may be set in ctor body
 472   _eden_chunk_capacity(0),     // -- ditto --
 473   _eden_chunk_index(0),        // -- ditto --
 474   _survivor_plab_array(NULL),  // -- ditto --
 475   _survivor_chunk_array(NULL), // -- ditto --
 476   _survivor_chunk_capacity(0), // -- ditto --
 477   _survivor_chunk_index(0),    // -- ditto --
 478   _ser_pmc_preclean_ovflw(0),
 479   _ser_kac_preclean_ovflw(0),
 480   _ser_pmc_remark_ovflw(0),
 481   _par_pmc_remark_ovflw(0),
 482   _ser_kac_ovflw(0),
 483   _par_kac_ovflw(0),
 484 #ifndef PRODUCT
 485   _num_par_pushes(0),
 486 #endif
 487   _collection_count_start(0),
 488   _verifying(false),
 489   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
 490   _completed_initialization(false),
 491   _collector_policy(cp),
 492   _should_unload_classes(CMSClassUnloadingEnabled),
 493   _concurrent_cycles_since_last_unload(0),
 494   _roots_scanning_options(GenCollectedHeap::SO_None),
 495   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 496   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 497   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
 498   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 499   _cms_start_registered(false)
 500 {
 501   if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
 502     ExplicitGCInvokesConcurrent = true;
 503   }
 504   // Now expand the span and allocate the collection support structures
 505   // (MUT, marking bit map etc.) to cover both generations subject to
 506   // collection.
 507 
 508   // For use by dirty card to oop closures.
 509   _cmsGen->cmsSpace()->set_collector(this);
 510 
 511   // Allocate MUT and marking bit map
 512   {
 513     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
 514     if (!_markBitMap.allocate(_span)) {
 515       warning("Failed to allocate CMS Bit Map");
 516       return;
 517     }
 518     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
 519   }
 520   {
 521     _modUnionTable.allocate(_span);
 522     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
 523   }
 524 
 525   if (!_markStack.allocate(MarkStackSize)) {
 526     warning("Failed to allocate CMS Marking Stack");
 527     return;
 528   }
 529 
 530   // Support for multi-threaded concurrent phases
 531   if (CMSConcurrentMTEnabled) {
 532     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
 533       // just for now
 534       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
 535     }
 536     if (ConcGCThreads > 1) {
 537       _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
 538                                  ConcGCThreads, true);
 539       if (_conc_workers == NULL) {
 540         warning("GC/CMS: _conc_workers allocation failure: "
 541               "forcing -CMSConcurrentMTEnabled");
 542         CMSConcurrentMTEnabled = false;
 543       } else {
 544         _conc_workers->initialize_workers();
 545       }
 546     } else {
 547       CMSConcurrentMTEnabled = false;
 548     }
 549   }
 550   if (!CMSConcurrentMTEnabled) {
 551     ConcGCThreads = 0;
 552   } else {
 553     // Turn off CMSCleanOnEnter optimization temporarily for
 554     // the MT case where it's not fixed yet; see 6178663.
 555     CMSCleanOnEnter = false;
 556   }
 557   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
 558          "Inconsistency");
 559 
 560   // Parallel task queues; these are shared for the
 561   // concurrent and stop-world phases of CMS, but
 562   // are not shared with parallel scavenge (ParNew).
 563   {
 564     uint i;
 565     uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
 566 
 567     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
 568          || ParallelRefProcEnabled)
 569         && num_queues > 0) {
 570       _task_queues = new OopTaskQueueSet(num_queues);
 571       if (_task_queues == NULL) {
 572         warning("task_queues allocation failure.");
 573         return;
 574       }
 575       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
 576       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
 577       for (i = 0; i < num_queues; i++) {
 578         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
 579         if (q == NULL) {
 580           warning("work_queue allocation failure.");
 581           return;
 582         }
 583         _task_queues->register_queue(i, q);
 584       }
 585       for (i = 0; i < num_queues; i++) {
 586         _task_queues->queue(i)->initialize();
 587         _hash_seed[i] = 17;  // copied from ParNew
 588       }
 589     }
 590   }
 591 
 592   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
 593 
 594   // Clip CMSBootstrapOccupancy between 0 and 100.
 595   _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
 596 
 597   // Now tell CMS generations the identity of their collector
 598   ConcurrentMarkSweepGeneration::set_collector(this);
 599 
 600   // Create & start a CMS thread for this CMS collector
 601   _cmsThread = ConcurrentMarkSweepThread::start(this);
 602   assert(cmsThread() != NULL, "CMS Thread should have been created");
 603   assert(cmsThread()->collector() == this,
 604          "CMS Thread should refer to this gen");
 605   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 606 
 607   // Support for parallelizing young gen rescan
 608   GenCollectedHeap* gch = GenCollectedHeap::heap();
 609   assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
 610   _young_gen = (ParNewGeneration*)gch->young_gen();
 611   if (gch->supports_inline_contig_alloc()) {
 612     _top_addr = gch->top_addr();
 613     _end_addr = gch->end_addr();
 614     assert(_young_gen != NULL, "no _young_gen");
 615     _eden_chunk_index = 0;
 616     _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
 617     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
 618   }
 619 
 620   // Support for parallelizing survivor space rescan
 621   if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
 622     const size_t max_plab_samples =
 623       ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size();
 624 
 625     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
 626     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
 627     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
 628     _survivor_chunk_capacity = 2*max_plab_samples;
 629     for (uint i = 0; i < ParallelGCThreads; i++) {
 630       HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 631       ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
 632       assert(cur->end() == 0, "Should be 0");
 633       assert(cur->array() == vec, "Should be vec");
 634       assert(cur->capacity() == max_plab_samples, "Error");
 635     }
 636   }
 637 
 638   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
 639   _gc_counters = new CollectorCounters("CMS", 1);
 640   _completed_initialization = true;
 641   _inter_sweep_timer.start();  // start of time
 642 }
 643 
 644 size_t CMSCollector::plab_sample_minimum_size() {
 645   // The default value of MinTLABSize is 2k, but there is
 646   // no way to get the default value if the flag has been overridden.
 647   return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K);
 648 }
 649 
 650 const char* ConcurrentMarkSweepGeneration::name() const {
 651   return "concurrent mark-sweep generation";
 652 }
 653 void ConcurrentMarkSweepGeneration::update_counters() {
 654   if (UsePerfData) {
 655     _space_counters->update_all();
 656     _gen_counters->update_all();
 657   }
 658 }
 659 
 660 // this is an optimized version of update_counters(). it takes the
 661 // used value as a parameter rather than computing it.
 662 //
 663 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
 664   if (UsePerfData) {
 665     _space_counters->update_used(used);
 666     _space_counters->update_capacity();
 667     _gen_counters->update_all();
 668   }
 669 }
 670 
 671 void ConcurrentMarkSweepGeneration::print() const {
 672   Generation::print();
 673   cmsSpace()->print();
 674 }
 675 
 676 #ifndef PRODUCT
 677 void ConcurrentMarkSweepGeneration::print_statistics() {
 678   cmsSpace()->printFLCensus(0);
 679 }
 680 #endif
 681 
 682 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
 683   GenCollectedHeap* gch = GenCollectedHeap::heap();
 684   if (PrintGCDetails) {
 685     // I didn't want to change the logging when removing the level concept,
 686     // but I guess this logging could say "old" or something instead of "1".
 687     assert(gch->is_old_gen(this),
 688            "The CMS generation should be the old generation");
 689     uint level = 1;
 690     if (Verbose) {
 691       gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]",
 692         level, short_name(), s, used(), capacity());
 693     } else {
 694       gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]",
 695         level, short_name(), s, used() / K, capacity() / K);
 696     }
 697   }
 698   if (Verbose) {
 699     gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")",
 700               gch->used(), gch->capacity());
 701   } else {
 702     gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)",
 703               gch->used() / K, gch->capacity() / K);
 704   }
 705 }
 706 
 707 size_t
 708 ConcurrentMarkSweepGeneration::contiguous_available() const {
 709   // dld proposes an improvement in precision here. If the committed
 710   // part of the space ends in a free block we should add that to
 711   // uncommitted size in the calculation below. Will make this
 712   // change later, staying with the approximation below for the
 713   // time being. -- ysr.
 714   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
 715 }
 716 
 717 size_t
 718 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
 719   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
 720 }
 721 
 722 size_t ConcurrentMarkSweepGeneration::max_available() const {
 723   return free() + _virtual_space.uncommitted_size();
 724 }
 725 
 726 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
 727   size_t available = max_available();
 728   size_t av_promo  = (size_t)gc_stats()->avg_promoted()->padded_average();
 729   bool   res = (available >= av_promo) || (available >= max_promotion_in_bytes);
 730   if (Verbose && PrintGCDetails) {
 731     gclog_or_tty->print_cr(
 732       "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "),"
 733       "max_promo(" SIZE_FORMAT ")",
 734       res? "":" not", available, res? ">=":"<",
 735       av_promo, max_promotion_in_bytes);
 736   }
 737   return res;
 738 }
 739 
 740 // At a promotion failure dump information on block layout in heap
 741 // (cms old generation).
 742 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
 743   if (CMSDumpAtPromotionFailure) {
 744     cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
 745   }
 746 }
 747 
 748 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
 749   // Clear the promotion information.  These pointers can be adjusted
 750   // along with all the other pointers into the heap but
 751   // compaction is expected to be a rare event with
 752   // a heap using cms so don't do it without seeing the need.
 753   for (uint i = 0; i < ParallelGCThreads; i++) {
 754     _par_gc_thread_states[i]->promo.reset();
 755   }
 756 }
 757 
 758 void ConcurrentMarkSweepGeneration::compute_new_size() {
 759   assert_locked_or_safepoint(Heap_lock);
 760 
 761   // If incremental collection failed, we just want to expand
 762   // to the limit.
 763   if (incremental_collection_failed()) {
 764     clear_incremental_collection_failed();
 765     grow_to_reserved();
 766     return;
 767   }
 768 
 769   // The heap has been compacted but not reset yet.
 770   // Any metric such as free() or used() will be incorrect.
 771 
 772   CardGeneration::compute_new_size();
 773 
 774   // Reset again after a possible resizing
 775   if (did_compact()) {
 776     cmsSpace()->reset_after_compaction();
 777   }
 778 }
 779 
 780 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
 781   assert_locked_or_safepoint(Heap_lock);
 782 
 783   // If incremental collection failed, we just want to expand
 784   // to the limit.
 785   if (incremental_collection_failed()) {
 786     clear_incremental_collection_failed();
 787     grow_to_reserved();
 788     return;
 789   }
 790 
 791   double free_percentage = ((double) free()) / capacity();
 792   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
 793   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
 794 
 795   // compute expansion delta needed for reaching desired free percentage
 796   if (free_percentage < desired_free_percentage) {
 797     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 798     assert(desired_capacity >= capacity(), "invalid expansion size");
 799     size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
 800     if (PrintGCDetails && Verbose) {
 801       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 802       gclog_or_tty->print_cr("\nFrom compute_new_size: ");
 803       gclog_or_tty->print_cr("  Free fraction %f", free_percentage);
 804       gclog_or_tty->print_cr("  Desired free fraction %f",
 805               desired_free_percentage);
 806       gclog_or_tty->print_cr("  Maximum free fraction %f",
 807               maximum_free_percentage);
 808       gclog_or_tty->print_cr("  Capacity " SIZE_FORMAT, capacity()/1000);
 809       gclog_or_tty->print_cr("  Desired capacity " SIZE_FORMAT,
 810               desired_capacity/1000);
 811       GenCollectedHeap* gch = GenCollectedHeap::heap();
 812       assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
 813       size_t young_size = gch->young_gen()->capacity();
 814       gclog_or_tty->print_cr("  Young gen size " SIZE_FORMAT, young_size / 1000);
 815       gclog_or_tty->print_cr("  unsafe_max_alloc_nogc " SIZE_FORMAT,
 816               unsafe_max_alloc_nogc()/1000);
 817       gclog_or_tty->print_cr("  contiguous available " SIZE_FORMAT,
 818               contiguous_available()/1000);
 819       gclog_or_tty->print_cr("  Expand by " SIZE_FORMAT " (bytes)",
 820               expand_bytes);
 821     }
 822     // safe if expansion fails
 823     expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
 824     if (PrintGCDetails && Verbose) {
 825       gclog_or_tty->print_cr("  Expanded free fraction %f",
 826         ((double) free()) / capacity());
 827     }
 828   } else {
 829     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 830     assert(desired_capacity <= capacity(), "invalid expansion size");
 831     size_t shrink_bytes = capacity() - desired_capacity;
 832     // Don't shrink unless the delta is greater than the minimum shrink we want
 833     if (shrink_bytes >= MinHeapDeltaBytes) {
 834       shrink_free_list_by(shrink_bytes);
 835     }
 836   }
 837 }
 838 
 839 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
 840   return cmsSpace()->freelistLock();
 841 }
 842 
 843 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
 844                                                   bool   tlab) {
 845   CMSSynchronousYieldRequest yr;
 846   MutexLockerEx x(freelistLock(),
 847                   Mutex::_no_safepoint_check_flag);
 848   return have_lock_and_allocate(size, tlab);
 849 }
 850 
 851 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
 852                                                   bool   tlab /* ignored */) {
 853   assert_lock_strong(freelistLock());
 854   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
 855   HeapWord* res = cmsSpace()->allocate(adjustedSize);
 856   // Allocate the object live (grey) if the background collector has
 857   // started marking. This is necessary because the marker may
 858   // have passed this address and consequently this object will
 859   // not otherwise be greyed and would be incorrectly swept up.
 860   // Note that if this object contains references, the writing
 861   // of those references will dirty the card containing this object
 862   // allowing the object to be blackened (and its references scanned)
 863   // either during a preclean phase or at the final checkpoint.
 864   if (res != NULL) {
 865     // We may block here with an uninitialized object with
 866     // its mark-bit or P-bits not yet set. Such objects need
 867     // to be safely navigable by block_start().
 868     assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
 869     assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
 870     collector()->direct_allocated(res, adjustedSize);
 871     _direct_allocated_words += adjustedSize;
 872     // allocation counters
 873     NOT_PRODUCT(
 874       _numObjectsAllocated++;
 875       _numWordsAllocated += (int)adjustedSize;
 876     )
 877   }
 878   return res;
 879 }
 880 
 881 // In the case of direct allocation by mutators in a generation that
 882 // is being concurrently collected, the object must be allocated
 883 // live (grey) if the background collector has started marking.
 884 // This is necessary because the marker may
 885 // have passed this address and consequently this object will
 886 // not otherwise be greyed and would be incorrectly swept up.
 887 // Note that if this object contains references, the writing
 888 // of those references will dirty the card containing this object
 889 // allowing the object to be blackened (and its references scanned)
 890 // either during a preclean phase or at the final checkpoint.
 891 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
 892   assert(_markBitMap.covers(start, size), "Out of bounds");
 893   if (_collectorState >= Marking) {
 894     MutexLockerEx y(_markBitMap.lock(),
 895                     Mutex::_no_safepoint_check_flag);
 896     // [see comments preceding SweepClosure::do_blk() below for details]
 897     //
 898     // Can the P-bits be deleted now?  JJJ
 899     //
 900     // 1. need to mark the object as live so it isn't collected
 901     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
 902     // 3. need to mark the end of the object so marking, precleaning or sweeping
 903     //    can skip over uninitialized or unparsable objects. An allocated
 904     //    object is considered uninitialized for our purposes as long as
 905     //    its klass word is NULL.  All old gen objects are parsable
 906     //    as soon as they are initialized.)
 907     _markBitMap.mark(start);          // object is live
 908     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
 909     _markBitMap.mark(start + size - 1);
 910                                       // mark end of object
 911   }
 912   // check that oop looks uninitialized
 913   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
 914 }
 915 
 916 void CMSCollector::promoted(bool par, HeapWord* start,
 917                             bool is_obj_array, size_t obj_size) {
 918   assert(_markBitMap.covers(start), "Out of bounds");
 919   // See comment in direct_allocated() about when objects should
 920   // be allocated live.
 921   if (_collectorState >= Marking) {
 922     // we already hold the marking bit map lock, taken in
 923     // the prologue
 924     if (par) {
 925       _markBitMap.par_mark(start);
 926     } else {
 927       _markBitMap.mark(start);
 928     }
 929     // We don't need to mark the object as uninitialized (as
 930     // in direct_allocated above) because this is being done with the
 931     // world stopped and the object will be initialized by the
 932     // time the marking, precleaning or sweeping get to look at it.
 933     // But see the code for copying objects into the CMS generation,
 934     // where we need to ensure that concurrent readers of the
 935     // block offset table are able to safely navigate a block that
 936     // is in flux from being free to being allocated (and in
 937     // transition while being copied into) and subsequently
 938     // becoming a bona-fide object when the copy/promotion is complete.
 939     assert(SafepointSynchronize::is_at_safepoint(),
 940            "expect promotion only at safepoints");
 941 
 942     if (_collectorState < Sweeping) {
 943       // Mark the appropriate cards in the modUnionTable, so that
 944       // this object gets scanned before the sweep. If this is
 945       // not done, CMS generation references in the object might
 946       // not get marked.
 947       // For the case of arrays, which are otherwise precisely
 948       // marked, we need to dirty the entire array, not just its head.
 949       if (is_obj_array) {
 950         // The [par_]mark_range() method expects mr.end() below to
 951         // be aligned to the granularity of a bit's representation
 952         // in the heap. In the case of the MUT below, that's a
 953         // card size.
 954         MemRegion mr(start,
 955                      (HeapWord*)round_to((intptr_t)(start + obj_size),
 956                         CardTableModRefBS::card_size /* bytes */));
 957         if (par) {
 958           _modUnionTable.par_mark_range(mr);
 959         } else {
 960           _modUnionTable.mark_range(mr);
 961         }
 962       } else {  // not an obj array; we can just mark the head
 963         if (par) {
 964           _modUnionTable.par_mark(start);
 965         } else {
 966           _modUnionTable.mark(start);
 967         }
 968       }
 969     }
 970   }
 971 }
 972 
 973 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
 974   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
 975   // allocate, copy and if necessary update promoinfo --
 976   // delegate to underlying space.
 977   assert_lock_strong(freelistLock());
 978 
 979 #ifndef PRODUCT
 980   if (GenCollectedHeap::heap()->promotion_should_fail()) {
 981     return NULL;
 982   }
 983 #endif  // #ifndef PRODUCT
 984 
 985   oop res = _cmsSpace->promote(obj, obj_size);
 986   if (res == NULL) {
 987     // expand and retry
 988     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
 989     expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
 990     // Since this is the old generation, we don't try to promote
 991     // into a more senior generation.
 992     res = _cmsSpace->promote(obj, obj_size);
 993   }
 994   if (res != NULL) {
 995     // See comment in allocate() about when objects should
 996     // be allocated live.
 997     assert(obj->is_oop(), "Will dereference klass pointer below");
 998     collector()->promoted(false,           // Not parallel
 999                           (HeapWord*)res, obj->is_objArray(), obj_size);
1000     // promotion counters
1001     NOT_PRODUCT(
1002       _numObjectsPromoted++;
1003       _numWordsPromoted +=
1004         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1005     )
1006   }
1007   return res;
1008 }
1009 
1010 
1011 // IMPORTANT: Notes on object size recognition in CMS.
1012 // ---------------------------------------------------
1013 // A block of storage in the CMS generation is always in
1014 // one of three states. A free block (FREE), an allocated
1015 // object (OBJECT) whose size() method reports the correct size,
1016 // and an intermediate state (TRANSIENT) in which its size cannot
1017 // be accurately determined.
1018 // STATE IDENTIFICATION:   (32 bit and 64 bit w/o COOPS)
1019 // -----------------------------------------------------
1020 // FREE:      klass_word & 1 == 1; mark_word holds block size
1021 //
1022 // OBJECT:    klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1023 //            obj->size() computes correct size
1024 //
1025 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1026 //
1027 // STATE IDENTIFICATION: (64 bit+COOPS)
1028 // ------------------------------------
1029 // FREE:      mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1030 //
1031 // OBJECT:    klass_word installed; klass_word != 0;
1032 //            obj->size() computes correct size
1033 //
1034 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1035 //
1036 //
1037 // STATE TRANSITION DIAGRAM
1038 //
1039 //        mut / parnew                     mut  /  parnew
1040 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1041 //  ^                                                                   |
1042 //  |------------------------ DEAD <------------------------------------|
1043 //         sweep                            mut
1044 //
1045 // While a block is in TRANSIENT state its size cannot be determined
1046 // so readers will either need to come back later or stall until
1047 // the size can be determined. Note that for the case of direct
1048 // allocation, P-bits, when available, may be used to determine the
1049 // size of an object that may not yet have been initialized.
1050 
1051 // Things to support parallel young-gen collection.
1052 oop
1053 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1054                                            oop old, markOop m,
1055                                            size_t word_sz) {
1056 #ifndef PRODUCT
1057   if (GenCollectedHeap::heap()->promotion_should_fail()) {
1058     return NULL;
1059   }
1060 #endif  // #ifndef PRODUCT
1061 
1062   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1063   PromotionInfo* promoInfo = &ps->promo;
1064   // if we are tracking promotions, then first ensure space for
1065   // promotion (including spooling space for saving header if necessary).
1066   // then allocate and copy, then track promoted info if needed.
1067   // When tracking (see PromotionInfo::track()), the mark word may
1068   // be displaced and in this case restoration of the mark word
1069   // occurs in the (oop_since_save_marks_)iterate phase.
1070   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1071     // Out of space for allocating spooling buffers;
1072     // try expanding and allocating spooling buffers.
1073     if (!expand_and_ensure_spooling_space(promoInfo)) {
1074       return NULL;
1075     }
1076   }
1077   assert(promoInfo->has_spooling_space(), "Control point invariant");
1078   const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1079   HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1080   if (obj_ptr == NULL) {
1081      obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1082      if (obj_ptr == NULL) {
1083        return NULL;
1084      }
1085   }
1086   oop obj = oop(obj_ptr);
1087   OrderAccess::storestore();
1088   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1089   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1090   // IMPORTANT: See note on object initialization for CMS above.
1091   // Otherwise, copy the object.  Here we must be careful to insert the
1092   // klass pointer last, since this marks the block as an allocated object.
1093   // Except with compressed oops it's the mark word.
1094   HeapWord* old_ptr = (HeapWord*)old;
1095   // Restore the mark word copied above.
1096   obj->set_mark(m);
1097   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1098   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1099   OrderAccess::storestore();
1100 
1101   if (UseCompressedClassPointers) {
1102     // Copy gap missed by (aligned) header size calculation below
1103     obj->set_klass_gap(old->klass_gap());
1104   }
1105   if (word_sz > (size_t)oopDesc::header_size()) {
1106     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1107                                  obj_ptr + oopDesc::header_size(),
1108                                  word_sz - oopDesc::header_size());
1109   }
1110 
1111   // Now we can track the promoted object, if necessary.  We take care
1112   // to delay the transition from uninitialized to full object
1113   // (i.e., insertion of klass pointer) until after, so that it
1114   // atomically becomes a promoted object.
1115   if (promoInfo->tracking()) {
1116     promoInfo->track((PromotedObject*)obj, old->klass());
1117   }
1118   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1119   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1120   assert(old->is_oop(), "Will use and dereference old klass ptr below");
1121 
1122   // Finally, install the klass pointer (this should be volatile).
1123   OrderAccess::storestore();
1124   obj->set_klass(old->klass());
1125   // We should now be able to calculate the right size for this object
1126   assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1127 
1128   collector()->promoted(true,          // parallel
1129                         obj_ptr, old->is_objArray(), word_sz);
1130 
1131   NOT_PRODUCT(
1132     Atomic::inc_ptr(&_numObjectsPromoted);
1133     Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1134   )
1135 
1136   return obj;
1137 }
1138 
1139 void
1140 ConcurrentMarkSweepGeneration::
1141 par_promote_alloc_done(int thread_num) {
1142   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1143   ps->lab.retire(thread_num);
1144 }
1145 
1146 void
1147 ConcurrentMarkSweepGeneration::
1148 par_oop_since_save_marks_iterate_done(int thread_num) {
1149   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1150   ParScanWithoutBarrierClosure* dummy_cl = NULL;
1151   ps->promo.promoted_oops_iterate_nv(dummy_cl);
1152 }
1153 
1154 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
1155                                                    size_t size,
1156                                                    bool   tlab)
1157 {
1158   // We allow a STW collection only if a full
1159   // collection was requested.
1160   return full || should_allocate(size, tlab); // FIX ME !!!
1161   // This and promotion failure handling are connected at the
1162   // hip and should be fixed by untying them.
1163 }
1164 
1165 bool CMSCollector::shouldConcurrentCollect() {
1166   if (_full_gc_requested) {
1167     if (Verbose && PrintGCDetails) {
1168       gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1169                              " gc request (or gc_locker)");
1170     }
1171     return true;
1172   }
1173 
1174   FreelistLocker x(this);
1175   // ------------------------------------------------------------------
1176   // Print out lots of information which affects the initiation of
1177   // a collection.
1178   if (PrintCMSInitiationStatistics && stats().valid()) {
1179     gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1180     gclog_or_tty->stamp();
1181     gclog_or_tty->cr();
1182     stats().print_on(gclog_or_tty);
1183     gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1184       stats().time_until_cms_gen_full());
1185     gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free());
1186     gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT,
1187                            _cmsGen->contiguous_available());
1188     gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1189     gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1190     gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1191     gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1192     gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1193     gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1194     gclog_or_tty->print_cr("metadata initialized %d",
1195       MetaspaceGC::should_concurrent_collect());
1196   }
1197   // ------------------------------------------------------------------
1198 
1199   // If the estimated time to complete a cms collection (cms_duration())
1200   // is less than the estimated time remaining until the cms generation
1201   // is full, start a collection.
1202   if (!UseCMSInitiatingOccupancyOnly) {
1203     if (stats().valid()) {
1204       if (stats().time_until_cms_start() == 0.0) {
1205         return true;
1206       }
1207     } else {
1208       // We want to conservatively collect somewhat early in order
1209       // to try and "bootstrap" our CMS/promotion statistics;
1210       // this branch will not fire after the first successful CMS
1211       // collection because the stats should then be valid.
1212       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1213         if (Verbose && PrintGCDetails) {
1214           gclog_or_tty->print_cr(
1215             " CMSCollector: collect for bootstrapping statistics:"
1216             " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1217             _bootstrap_occupancy);
1218         }
1219         return true;
1220       }
1221     }
1222   }
1223 
1224   // Otherwise, we start a collection cycle if
1225   // old gen want a collection cycle started. Each may use
1226   // an appropriate criterion for making this decision.
1227   // XXX We need to make sure that the gen expansion
1228   // criterion dovetails well with this. XXX NEED TO FIX THIS
1229   if (_cmsGen->should_concurrent_collect()) {
1230     if (Verbose && PrintGCDetails) {
1231       gclog_or_tty->print_cr("CMS old gen initiated");
1232     }
1233     return true;
1234   }
1235 
1236   // We start a collection if we believe an incremental collection may fail;
1237   // this is not likely to be productive in practice because it's probably too
1238   // late anyway.
1239   GenCollectedHeap* gch = GenCollectedHeap::heap();
1240   assert(gch->collector_policy()->is_generation_policy(),
1241          "You may want to check the correctness of the following");
1242   if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1243     if (Verbose && PrintGCDetails) {
1244       gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1245     }
1246     return true;
1247   }
1248 
1249   if (MetaspaceGC::should_concurrent_collect()) {
1250     if (Verbose && PrintGCDetails) {
1251       gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1252     }
1253     return true;
1254   }
1255 
1256   // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1257   if (CMSTriggerInterval >= 0) {
1258     if (CMSTriggerInterval == 0) {
1259       // Trigger always
1260       return true;
1261     }
1262 
1263     // Check the CMS time since begin (we do not check the stats validity
1264     // as we want to be able to trigger the first CMS cycle as well)
1265     if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1266       if (Verbose && PrintGCDetails) {
1267         if (stats().valid()) {
1268           gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1269                                  stats().cms_time_since_begin());
1270         } else {
1271           gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1272         }
1273       }
1274       return true;
1275     }
1276   }
1277 
1278   return false;
1279 }
1280 
1281 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1282 
1283 // Clear _expansion_cause fields of constituent generations
1284 void CMSCollector::clear_expansion_cause() {
1285   _cmsGen->clear_expansion_cause();
1286 }
1287 
1288 // We should be conservative in starting a collection cycle.  To
1289 // start too eagerly runs the risk of collecting too often in the
1290 // extreme.  To collect too rarely falls back on full collections,
1291 // which works, even if not optimum in terms of concurrent work.
1292 // As a work around for too eagerly collecting, use the flag
1293 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
1294 // giving the user an easily understandable way of controlling the
1295 // collections.
1296 // We want to start a new collection cycle if any of the following
1297 // conditions hold:
1298 // . our current occupancy exceeds the configured initiating occupancy
1299 //   for this generation, or
1300 // . we recently needed to expand this space and have not, since that
1301 //   expansion, done a collection of this generation, or
1302 // . the underlying space believes that it may be a good idea to initiate
1303 //   a concurrent collection (this may be based on criteria such as the
1304 //   following: the space uses linear allocation and linear allocation is
1305 //   going to fail, or there is believed to be excessive fragmentation in
1306 //   the generation, etc... or ...
1307 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1308 //   the case of the old generation; see CR 6543076):
1309 //   we may be approaching a point at which allocation requests may fail because
1310 //   we will be out of sufficient free space given allocation rate estimates.]
1311 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1312 
1313   assert_lock_strong(freelistLock());
1314   if (occupancy() > initiating_occupancy()) {
1315     if (PrintGCDetails && Verbose) {
1316       gclog_or_tty->print(" %s: collect because of occupancy %f / %f  ",
1317         short_name(), occupancy(), initiating_occupancy());
1318     }
1319     return true;
1320   }
1321   if (UseCMSInitiatingOccupancyOnly) {
1322     return false;
1323   }
1324   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1325     if (PrintGCDetails && Verbose) {
1326       gclog_or_tty->print(" %s: collect because expanded for allocation ",
1327         short_name());
1328     }
1329     return true;
1330   }
1331   if (_cmsSpace->should_concurrent_collect()) {
1332     if (PrintGCDetails && Verbose) {
1333       gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1334         short_name());
1335     }
1336     return true;
1337   }
1338   return false;
1339 }
1340 
1341 void ConcurrentMarkSweepGeneration::collect(bool   full,
1342                                             bool   clear_all_soft_refs,
1343                                             size_t size,
1344                                             bool   tlab)
1345 {
1346   collector()->collect(full, clear_all_soft_refs, size, tlab);
1347 }
1348 
1349 void CMSCollector::collect(bool   full,
1350                            bool   clear_all_soft_refs,
1351                            size_t size,
1352                            bool   tlab)
1353 {
1354   // The following "if" branch is present for defensive reasons.
1355   // In the current uses of this interface, it can be replaced with:
1356   // assert(!GC_locker.is_active(), "Can't be called otherwise");
1357   // But I am not placing that assert here to allow future
1358   // generality in invoking this interface.
1359   if (GC_locker::is_active()) {
1360     // A consistency test for GC_locker
1361     assert(GC_locker::needs_gc(), "Should have been set already");
1362     // Skip this foreground collection, instead
1363     // expanding the heap if necessary.
1364     // Need the free list locks for the call to free() in compute_new_size()
1365     compute_new_size();
1366     return;
1367   }
1368   acquire_control_and_collect(full, clear_all_soft_refs);
1369 }
1370 
1371 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1372   GenCollectedHeap* gch = GenCollectedHeap::heap();
1373   unsigned int gc_count = gch->total_full_collections();
1374   if (gc_count == full_gc_count) {
1375     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1376     _full_gc_requested = true;
1377     _full_gc_cause = cause;
1378     CGC_lock->notify();   // nudge CMS thread
1379   } else {
1380     assert(gc_count > full_gc_count, "Error: causal loop");
1381   }
1382 }
1383 
1384 bool CMSCollector::is_external_interruption() {
1385   GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1386   return GCCause::is_user_requested_gc(cause) ||
1387          GCCause::is_serviceability_requested_gc(cause);
1388 }
1389 
1390 void CMSCollector::report_concurrent_mode_interruption() {
1391   if (is_external_interruption()) {
1392     if (PrintGCDetails) {
1393       gclog_or_tty->print(" (concurrent mode interrupted)");
1394     }
1395   } else {
1396     if (PrintGCDetails) {
1397       gclog_or_tty->print(" (concurrent mode failure)");
1398     }
1399     _gc_tracer_cm->report_concurrent_mode_failure();
1400   }
1401 }
1402 
1403 
1404 // The foreground and background collectors need to coordinate in order
1405 // to make sure that they do not mutually interfere with CMS collections.
1406 // When a background collection is active,
1407 // the foreground collector may need to take over (preempt) and
1408 // synchronously complete an ongoing collection. Depending on the
1409 // frequency of the background collections and the heap usage
1410 // of the application, this preemption can be seldom or frequent.
1411 // There are only certain
1412 // points in the background collection that the "collection-baton"
1413 // can be passed to the foreground collector.
1414 //
1415 // The foreground collector will wait for the baton before
1416 // starting any part of the collection.  The foreground collector
1417 // will only wait at one location.
1418 //
1419 // The background collector will yield the baton before starting a new
1420 // phase of the collection (e.g., before initial marking, marking from roots,
1421 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
1422 // of the loop which switches the phases. The background collector does some
1423 // of the phases (initial mark, final re-mark) with the world stopped.
1424 // Because of locking involved in stopping the world,
1425 // the foreground collector should not block waiting for the background
1426 // collector when it is doing a stop-the-world phase.  The background
1427 // collector will yield the baton at an additional point just before
1428 // it enters a stop-the-world phase.  Once the world is stopped, the
1429 // background collector checks the phase of the collection.  If the
1430 // phase has not changed, it proceeds with the collection.  If the
1431 // phase has changed, it skips that phase of the collection.  See
1432 // the comments on the use of the Heap_lock in collect_in_background().
1433 //
1434 // Variable used in baton passing.
1435 //   _foregroundGCIsActive - Set to true by the foreground collector when
1436 //      it wants the baton.  The foreground clears it when it has finished
1437 //      the collection.
1438 //   _foregroundGCShouldWait - Set to true by the background collector
1439 //        when it is running.  The foreground collector waits while
1440 //      _foregroundGCShouldWait is true.
1441 //  CGC_lock - monitor used to protect access to the above variables
1442 //      and to notify the foreground and background collectors.
1443 //  _collectorState - current state of the CMS collection.
1444 //
1445 // The foreground collector
1446 //   acquires the CGC_lock
1447 //   sets _foregroundGCIsActive
1448 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
1449 //     various locks acquired in preparation for the collection
1450 //     are released so as not to block the background collector
1451 //     that is in the midst of a collection
1452 //   proceeds with the collection
1453 //   clears _foregroundGCIsActive
1454 //   returns
1455 //
1456 // The background collector in a loop iterating on the phases of the
1457 //      collection
1458 //   acquires the CGC_lock
1459 //   sets _foregroundGCShouldWait
1460 //   if _foregroundGCIsActive is set
1461 //     clears _foregroundGCShouldWait, notifies _CGC_lock
1462 //     waits on _CGC_lock for _foregroundGCIsActive to become false
1463 //     and exits the loop.
1464 //   otherwise
1465 //     proceed with that phase of the collection
1466 //     if the phase is a stop-the-world phase,
1467 //       yield the baton once more just before enqueueing
1468 //       the stop-world CMS operation (executed by the VM thread).
1469 //   returns after all phases of the collection are done
1470 //
1471 
1472 void CMSCollector::acquire_control_and_collect(bool full,
1473         bool clear_all_soft_refs) {
1474   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1475   assert(!Thread::current()->is_ConcurrentGC_thread(),
1476          "shouldn't try to acquire control from self!");
1477 
1478   // Start the protocol for acquiring control of the
1479   // collection from the background collector (aka CMS thread).
1480   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1481          "VM thread should have CMS token");
1482   // Remember the possibly interrupted state of an ongoing
1483   // concurrent collection
1484   CollectorState first_state = _collectorState;
1485 
1486   // Signal to a possibly ongoing concurrent collection that
1487   // we want to do a foreground collection.
1488   _foregroundGCIsActive = true;
1489 
1490   // release locks and wait for a notify from the background collector
1491   // releasing the locks in only necessary for phases which
1492   // do yields to improve the granularity of the collection.
1493   assert_lock_strong(bitMapLock());
1494   // We need to lock the Free list lock for the space that we are
1495   // currently collecting.
1496   assert(haveFreelistLocks(), "Must be holding free list locks");
1497   bitMapLock()->unlock();
1498   releaseFreelistLocks();
1499   {
1500     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1501     if (_foregroundGCShouldWait) {
1502       // We are going to be waiting for action for the CMS thread;
1503       // it had better not be gone (for instance at shutdown)!
1504       assert(ConcurrentMarkSweepThread::cmst() != NULL,
1505              "CMS thread must be running");
1506       // Wait here until the background collector gives us the go-ahead
1507       ConcurrentMarkSweepThread::clear_CMS_flag(
1508         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
1509       // Get a possibly blocked CMS thread going:
1510       //   Note that we set _foregroundGCIsActive true above,
1511       //   without protection of the CGC_lock.
1512       CGC_lock->notify();
1513       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1514              "Possible deadlock");
1515       while (_foregroundGCShouldWait) {
1516         // wait for notification
1517         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1518         // Possibility of delay/starvation here, since CMS token does
1519         // not know to give priority to VM thread? Actually, i think
1520         // there wouldn't be any delay/starvation, but the proof of
1521         // that "fact" (?) appears non-trivial. XXX 20011219YSR
1522       }
1523       ConcurrentMarkSweepThread::set_CMS_flag(
1524         ConcurrentMarkSweepThread::CMS_vm_has_token);
1525     }
1526   }
1527   // The CMS_token is already held.  Get back the other locks.
1528   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1529          "VM thread should have CMS token");
1530   getFreelistLocks();
1531   bitMapLock()->lock_without_safepoint_check();
1532   if (TraceCMSState) {
1533     gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1534       INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1535     gclog_or_tty->print_cr("    gets control with state %d", _collectorState);
1536   }
1537 
1538   // Inform cms gen if this was due to partial collection failing.
1539   // The CMS gen may use this fact to determine its expansion policy.
1540   GenCollectedHeap* gch = GenCollectedHeap::heap();
1541   if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1542     assert(!_cmsGen->incremental_collection_failed(),
1543            "Should have been noticed, reacted to and cleared");
1544     _cmsGen->set_incremental_collection_failed();
1545   }
1546 
1547   if (first_state > Idling) {
1548     report_concurrent_mode_interruption();
1549   }
1550 
1551   set_did_compact(true);
1552 
1553   // If the collection is being acquired from the background
1554   // collector, there may be references on the discovered
1555   // references lists.  Abandon those references, since some
1556   // of them may have become unreachable after concurrent
1557   // discovery; the STW compacting collector will redo discovery
1558   // more precisely, without being subject to floating garbage.
1559   // Leaving otherwise unreachable references in the discovered
1560   // lists would require special handling.
1561   ref_processor()->disable_discovery();
1562   ref_processor()->abandon_partial_discovery();
1563   ref_processor()->verify_no_references_recorded();
1564 
1565   if (first_state > Idling) {
1566     save_heap_summary();
1567   }
1568 
1569   do_compaction_work(clear_all_soft_refs);
1570 
1571   // Has the GC time limit been exceeded?
1572   size_t max_eden_size = _young_gen->max_capacity() -
1573                          _young_gen->to()->capacity() -
1574                          _young_gen->from()->capacity();
1575   GCCause::Cause gc_cause = gch->gc_cause();
1576   size_policy()->check_gc_overhead_limit(_young_gen->used(),
1577                                          _young_gen->eden()->used(),
1578                                          _cmsGen->max_capacity(),
1579                                          max_eden_size,
1580                                          full,
1581                                          gc_cause,
1582                                          gch->collector_policy());
1583 
1584   // Reset the expansion cause, now that we just completed
1585   // a collection cycle.
1586   clear_expansion_cause();
1587   _foregroundGCIsActive = false;
1588   return;
1589 }
1590 
1591 // Resize the tenured generation
1592 // after obtaining the free list locks for the
1593 // two generations.
1594 void CMSCollector::compute_new_size() {
1595   assert_locked_or_safepoint(Heap_lock);
1596   FreelistLocker z(this);
1597   MetaspaceGC::compute_new_size();
1598   _cmsGen->compute_new_size_free_list();
1599 }
1600 
1601 // A work method used by the foreground collector to do
1602 // a mark-sweep-compact.
1603 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1604   GenCollectedHeap* gch = GenCollectedHeap::heap();
1605 
1606   STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1607   gc_timer->register_gc_start();
1608 
1609   SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1610   gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1611 
1612   GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1613 
1614   // Temporarily widen the span of the weak reference processing to
1615   // the entire heap.
1616   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1617   ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1618   // Temporarily, clear the "is_alive_non_header" field of the
1619   // reference processor.
1620   ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1621   // Temporarily make reference _processing_ single threaded (non-MT).
1622   ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1623   // Temporarily make refs discovery atomic
1624   ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1625   // Temporarily make reference _discovery_ single threaded (non-MT)
1626   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1627 
1628   ref_processor()->set_enqueuing_is_done(false);
1629   ref_processor()->enable_discovery();
1630   ref_processor()->setup_policy(clear_all_soft_refs);
1631   // If an asynchronous collection finishes, the _modUnionTable is
1632   // all clear.  If we are assuming the collection from an asynchronous
1633   // collection, clear the _modUnionTable.
1634   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1635     "_modUnionTable should be clear if the baton was not passed");
1636   _modUnionTable.clear_all();
1637   assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1638     "mod union for klasses should be clear if the baton was passed");
1639   _ct->klass_rem_set()->clear_mod_union();
1640 
1641   // We must adjust the allocation statistics being maintained
1642   // in the free list space. We do so by reading and clearing
1643   // the sweep timer and updating the block flux rate estimates below.
1644   assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1645   if (_inter_sweep_timer.is_active()) {
1646     _inter_sweep_timer.stop();
1647     // Note that we do not use this sample to update the _inter_sweep_estimate.
1648     _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1649                                             _inter_sweep_estimate.padded_average(),
1650                                             _intra_sweep_estimate.padded_average());
1651   }
1652 
1653   GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1654   #ifdef ASSERT
1655     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1656     size_t free_size = cms_space->free();
1657     assert(free_size ==
1658            pointer_delta(cms_space->end(), cms_space->compaction_top())
1659            * HeapWordSize,
1660       "All the free space should be compacted into one chunk at top");
1661     assert(cms_space->dictionary()->total_chunk_size(
1662                                       debug_only(cms_space->freelistLock())) == 0 ||
1663            cms_space->totalSizeInIndexedFreeLists() == 0,
1664       "All the free space should be in a single chunk");
1665     size_t num = cms_space->totalCount();
1666     assert((free_size == 0 && num == 0) ||
1667            (free_size > 0  && (num == 1 || num == 2)),
1668          "There should be at most 2 free chunks after compaction");
1669   #endif // ASSERT
1670   _collectorState = Resetting;
1671   assert(_restart_addr == NULL,
1672          "Should have been NULL'd before baton was passed");
1673   reset(false /* == !concurrent */);
1674   _cmsGen->reset_after_compaction();
1675   _concurrent_cycles_since_last_unload = 0;
1676 
1677   // Clear any data recorded in the PLAB chunk arrays.
1678   if (_survivor_plab_array != NULL) {
1679     reset_survivor_plab_arrays();
1680   }
1681 
1682   // Adjust the per-size allocation stats for the next epoch.
1683   _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1684   // Restart the "inter sweep timer" for the next epoch.
1685   _inter_sweep_timer.reset();
1686   _inter_sweep_timer.start();
1687 
1688   gc_timer->register_gc_end();
1689 
1690   gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1691 
1692   // For a mark-sweep-compact, compute_new_size() will be called
1693   // in the heap's do_collection() method.
1694 }
1695 
1696 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1697   ContiguousSpace* eden_space = _young_gen->eden();
1698   ContiguousSpace* from_space = _young_gen->from();
1699   ContiguousSpace* to_space   = _young_gen->to();
1700   // Eden
1701   if (_eden_chunk_array != NULL) {
1702     gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1703                            p2i(eden_space->bottom()), p2i(eden_space->top()),
1704                            p2i(eden_space->end()), eden_space->capacity());
1705     gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1706                            "_eden_chunk_capacity=" SIZE_FORMAT,
1707                            _eden_chunk_index, _eden_chunk_capacity);
1708     for (size_t i = 0; i < _eden_chunk_index; i++) {
1709       gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1710                              i, p2i(_eden_chunk_array[i]));
1711     }
1712   }
1713   // Survivor
1714   if (_survivor_chunk_array != NULL) {
1715     gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1716                            p2i(from_space->bottom()), p2i(from_space->top()),
1717                            p2i(from_space->end()), from_space->capacity());
1718     gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1719                            "_survivor_chunk_capacity=" SIZE_FORMAT,
1720                            _survivor_chunk_index, _survivor_chunk_capacity);
1721     for (size_t i = 0; i < _survivor_chunk_index; i++) {
1722       gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1723                              i, p2i(_survivor_chunk_array[i]));
1724     }
1725   }
1726 }
1727 
1728 void CMSCollector::getFreelistLocks() const {
1729   // Get locks for all free lists in all generations that this
1730   // collector is responsible for
1731   _cmsGen->freelistLock()->lock_without_safepoint_check();
1732 }
1733 
1734 void CMSCollector::releaseFreelistLocks() const {
1735   // Release locks for all free lists in all generations that this
1736   // collector is responsible for
1737   _cmsGen->freelistLock()->unlock();
1738 }
1739 
1740 bool CMSCollector::haveFreelistLocks() const {
1741   // Check locks for all free lists in all generations that this
1742   // collector is responsible for
1743   assert_lock_strong(_cmsGen->freelistLock());
1744   PRODUCT_ONLY(ShouldNotReachHere());
1745   return true;
1746 }
1747 
1748 // A utility class that is used by the CMS collector to
1749 // temporarily "release" the foreground collector from its
1750 // usual obligation to wait for the background collector to
1751 // complete an ongoing phase before proceeding.
1752 class ReleaseForegroundGC: public StackObj {
1753  private:
1754   CMSCollector* _c;
1755  public:
1756   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1757     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1758     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1759     // allow a potentially blocked foreground collector to proceed
1760     _c->_foregroundGCShouldWait = false;
1761     if (_c->_foregroundGCIsActive) {
1762       CGC_lock->notify();
1763     }
1764     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1765            "Possible deadlock");
1766   }
1767 
1768   ~ReleaseForegroundGC() {
1769     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1770     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1771     _c->_foregroundGCShouldWait = true;
1772   }
1773 };
1774 
1775 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1776   assert(Thread::current()->is_ConcurrentGC_thread(),
1777     "A CMS asynchronous collection is only allowed on a CMS thread.");
1778 
1779   GenCollectedHeap* gch = GenCollectedHeap::heap();
1780   {
1781     bool safepoint_check = Mutex::_no_safepoint_check_flag;
1782     MutexLockerEx hl(Heap_lock, safepoint_check);
1783     FreelistLocker fll(this);
1784     MutexLockerEx x(CGC_lock, safepoint_check);
1785     if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1786       // The foreground collector is active or we're
1787       // not using asynchronous collections.  Skip this
1788       // background collection.
1789       assert(!_foregroundGCShouldWait, "Should be clear");
1790       return;
1791     } else {
1792       assert(_collectorState == Idling, "Should be idling before start.");
1793       _collectorState = InitialMarking;
1794       register_gc_start(cause);
1795       // Reset the expansion cause, now that we are about to begin
1796       // a new cycle.
1797       clear_expansion_cause();
1798 
1799       // Clear the MetaspaceGC flag since a concurrent collection
1800       // is starting but also clear it after the collection.
1801       MetaspaceGC::set_should_concurrent_collect(false);
1802     }
1803     // Decide if we want to enable class unloading as part of the
1804     // ensuing concurrent GC cycle.
1805     update_should_unload_classes();
1806     _full_gc_requested = false;           // acks all outstanding full gc requests
1807     _full_gc_cause = GCCause::_no_gc;
1808     // Signal that we are about to start a collection
1809     gch->increment_total_full_collections();  // ... starting a collection cycle
1810     _collection_count_start = gch->total_full_collections();
1811   }
1812 
1813   // Used for PrintGC
1814   size_t prev_used;
1815   if (PrintGC && Verbose) {
1816     prev_used = _cmsGen->used();
1817   }
1818 
1819   // The change of the collection state is normally done at this level;
1820   // the exceptions are phases that are executed while the world is
1821   // stopped.  For those phases the change of state is done while the
1822   // world is stopped.  For baton passing purposes this allows the
1823   // background collector to finish the phase and change state atomically.
1824   // The foreground collector cannot wait on a phase that is done
1825   // while the world is stopped because the foreground collector already
1826   // has the world stopped and would deadlock.
1827   while (_collectorState != Idling) {
1828     if (TraceCMSState) {
1829       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1830         p2i(Thread::current()), _collectorState);
1831     }
1832     // The foreground collector
1833     //   holds the Heap_lock throughout its collection.
1834     //   holds the CMS token (but not the lock)
1835     //     except while it is waiting for the background collector to yield.
1836     //
1837     // The foreground collector should be blocked (not for long)
1838     //   if the background collector is about to start a phase
1839     //   executed with world stopped.  If the background
1840     //   collector has already started such a phase, the
1841     //   foreground collector is blocked waiting for the
1842     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
1843     //   are executed in the VM thread.
1844     //
1845     // The locking order is
1846     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
1847     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
1848     //   CMS token  (claimed in
1849     //                stop_world_and_do() -->
1850     //                  safepoint_synchronize() -->
1851     //                    CMSThread::synchronize())
1852 
1853     {
1854       // Check if the FG collector wants us to yield.
1855       CMSTokenSync x(true); // is cms thread
1856       if (waitForForegroundGC()) {
1857         // We yielded to a foreground GC, nothing more to be
1858         // done this round.
1859         assert(_foregroundGCShouldWait == false, "We set it to false in "
1860                "waitForForegroundGC()");
1861         if (TraceCMSState) {
1862           gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1863             " exiting collection CMS state %d",
1864             p2i(Thread::current()), _collectorState);
1865         }
1866         return;
1867       } else {
1868         // The background collector can run but check to see if the
1869         // foreground collector has done a collection while the
1870         // background collector was waiting to get the CGC_lock
1871         // above.  If yes, break so that _foregroundGCShouldWait
1872         // is cleared before returning.
1873         if (_collectorState == Idling) {
1874           break;
1875         }
1876       }
1877     }
1878 
1879     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1880       "should be waiting");
1881 
1882     switch (_collectorState) {
1883       case InitialMarking:
1884         {
1885           ReleaseForegroundGC x(this);
1886           stats().record_cms_begin();
1887           VM_CMS_Initial_Mark initial_mark_op(this);
1888           VMThread::execute(&initial_mark_op);
1889         }
1890         // The collector state may be any legal state at this point
1891         // since the background collector may have yielded to the
1892         // foreground collector.
1893         break;
1894       case Marking:
1895         // initial marking in checkpointRootsInitialWork has been completed
1896         if (markFromRoots()) { // we were successful
1897           assert(_collectorState == Precleaning, "Collector state should "
1898             "have changed");
1899         } else {
1900           assert(_foregroundGCIsActive, "Internal state inconsistency");
1901         }
1902         break;
1903       case Precleaning:
1904         // marking from roots in markFromRoots has been completed
1905         preclean();
1906         assert(_collectorState == AbortablePreclean ||
1907                _collectorState == FinalMarking,
1908                "Collector state should have changed");
1909         break;
1910       case AbortablePreclean:
1911         abortable_preclean();
1912         assert(_collectorState == FinalMarking, "Collector state should "
1913           "have changed");
1914         break;
1915       case FinalMarking:
1916         {
1917           ReleaseForegroundGC x(this);
1918 
1919           VM_CMS_Final_Remark final_remark_op(this);
1920           VMThread::execute(&final_remark_op);
1921         }
1922         assert(_foregroundGCShouldWait, "block post-condition");
1923         break;
1924       case Sweeping:
1925         // final marking in checkpointRootsFinal has been completed
1926         sweep();
1927         assert(_collectorState == Resizing, "Collector state change "
1928           "to Resizing must be done under the free_list_lock");
1929 
1930       case Resizing: {
1931         // Sweeping has been completed...
1932         // At this point the background collection has completed.
1933         // Don't move the call to compute_new_size() down
1934         // into code that might be executed if the background
1935         // collection was preempted.
1936         {
1937           ReleaseForegroundGC x(this);   // unblock FG collection
1938           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
1939           CMSTokenSync        z(true);   // not strictly needed.
1940           if (_collectorState == Resizing) {
1941             compute_new_size();
1942             save_heap_summary();
1943             _collectorState = Resetting;
1944           } else {
1945             assert(_collectorState == Idling, "The state should only change"
1946                    " because the foreground collector has finished the collection");
1947           }
1948         }
1949         break;
1950       }
1951       case Resetting:
1952         // CMS heap resizing has been completed
1953         reset(true);
1954         assert(_collectorState == Idling, "Collector state should "
1955           "have changed");
1956 
1957         MetaspaceGC::set_should_concurrent_collect(false);
1958 
1959         stats().record_cms_end();
1960         // Don't move the concurrent_phases_end() and compute_new_size()
1961         // calls to here because a preempted background collection
1962         // has it's state set to "Resetting".
1963         break;
1964       case Idling:
1965       default:
1966         ShouldNotReachHere();
1967         break;
1968     }
1969     if (TraceCMSState) {
1970       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
1971         p2i(Thread::current()), _collectorState);
1972     }
1973     assert(_foregroundGCShouldWait, "block post-condition");
1974   }
1975 
1976   // Should this be in gc_epilogue?
1977   collector_policy()->counters()->update_counters();
1978 
1979   {
1980     // Clear _foregroundGCShouldWait and, in the event that the
1981     // foreground collector is waiting, notify it, before
1982     // returning.
1983     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1984     _foregroundGCShouldWait = false;
1985     if (_foregroundGCIsActive) {
1986       CGC_lock->notify();
1987     }
1988     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1989            "Possible deadlock");
1990   }
1991   if (TraceCMSState) {
1992     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1993       " exiting collection CMS state %d",
1994       p2i(Thread::current()), _collectorState);
1995   }
1996   if (PrintGC && Verbose) {
1997     _cmsGen->print_heap_change(prev_used);
1998   }
1999 }
2000 
2001 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2002   _cms_start_registered = true;
2003   _gc_timer_cm->register_gc_start();
2004   _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2005 }
2006 
2007 void CMSCollector::register_gc_end() {
2008   if (_cms_start_registered) {
2009     report_heap_summary(GCWhen::AfterGC);
2010 
2011     _gc_timer_cm->register_gc_end();
2012     _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2013     _cms_start_registered = false;
2014   }
2015 }
2016 
2017 void CMSCollector::save_heap_summary() {
2018   GenCollectedHeap* gch = GenCollectedHeap::heap();
2019   _last_heap_summary = gch->create_heap_summary();
2020   _last_metaspace_summary = gch->create_metaspace_summary();
2021 }
2022 
2023 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2024   _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2025   _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2026 }
2027 
2028 bool CMSCollector::waitForForegroundGC() {
2029   bool res = false;
2030   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2031          "CMS thread should have CMS token");
2032   // Block the foreground collector until the
2033   // background collectors decides whether to
2034   // yield.
2035   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2036   _foregroundGCShouldWait = true;
2037   if (_foregroundGCIsActive) {
2038     // The background collector yields to the
2039     // foreground collector and returns a value
2040     // indicating that it has yielded.  The foreground
2041     // collector can proceed.
2042     res = true;
2043     _foregroundGCShouldWait = false;
2044     ConcurrentMarkSweepThread::clear_CMS_flag(
2045       ConcurrentMarkSweepThread::CMS_cms_has_token);
2046     ConcurrentMarkSweepThread::set_CMS_flag(
2047       ConcurrentMarkSweepThread::CMS_cms_wants_token);
2048     // Get a possibly blocked foreground thread going
2049     CGC_lock->notify();
2050     if (TraceCMSState) {
2051       gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2052         p2i(Thread::current()), _collectorState);
2053     }
2054     while (_foregroundGCIsActive) {
2055       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2056     }
2057     ConcurrentMarkSweepThread::set_CMS_flag(
2058       ConcurrentMarkSweepThread::CMS_cms_has_token);
2059     ConcurrentMarkSweepThread::clear_CMS_flag(
2060       ConcurrentMarkSweepThread::CMS_cms_wants_token);
2061   }
2062   if (TraceCMSState) {
2063     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2064       p2i(Thread::current()), _collectorState);
2065   }
2066   return res;
2067 }
2068 
2069 // Because of the need to lock the free lists and other structures in
2070 // the collector, common to all the generations that the collector is
2071 // collecting, we need the gc_prologues of individual CMS generations
2072 // delegate to their collector. It may have been simpler had the
2073 // current infrastructure allowed one to call a prologue on a
2074 // collector. In the absence of that we have the generation's
2075 // prologue delegate to the collector, which delegates back
2076 // some "local" work to a worker method in the individual generations
2077 // that it's responsible for collecting, while itself doing any
2078 // work common to all generations it's responsible for. A similar
2079 // comment applies to the  gc_epilogue()'s.
2080 // The role of the variable _between_prologue_and_epilogue is to
2081 // enforce the invocation protocol.
2082 void CMSCollector::gc_prologue(bool full) {
2083   // Call gc_prologue_work() for the CMSGen
2084   // we are responsible for.
2085 
2086   // The following locking discipline assumes that we are only called
2087   // when the world is stopped.
2088   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2089 
2090   // The CMSCollector prologue must call the gc_prologues for the
2091   // "generations" that it's responsible
2092   // for.
2093 
2094   assert(   Thread::current()->is_VM_thread()
2095          || (   CMSScavengeBeforeRemark
2096              && Thread::current()->is_ConcurrentGC_thread()),
2097          "Incorrect thread type for prologue execution");
2098 
2099   if (_between_prologue_and_epilogue) {
2100     // We have already been invoked; this is a gc_prologue delegation
2101     // from yet another CMS generation that we are responsible for, just
2102     // ignore it since all relevant work has already been done.
2103     return;
2104   }
2105 
2106   // set a bit saying prologue has been called; cleared in epilogue
2107   _between_prologue_and_epilogue = true;
2108   // Claim locks for common data structures, then call gc_prologue_work()
2109   // for each CMSGen.
2110 
2111   getFreelistLocks();   // gets free list locks on constituent spaces
2112   bitMapLock()->lock_without_safepoint_check();
2113 
2114   // Should call gc_prologue_work() for all cms gens we are responsible for
2115   bool duringMarking =    _collectorState >= Marking
2116                          && _collectorState < Sweeping;
2117 
2118   // The young collections clear the modified oops state, which tells if
2119   // there are any modified oops in the class. The remark phase also needs
2120   // that information. Tell the young collection to save the union of all
2121   // modified klasses.
2122   if (duringMarking) {
2123     _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2124   }
2125 
2126   bool registerClosure = duringMarking;
2127 
2128   _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2129 
2130   if (!full) {
2131     stats().record_gc0_begin();
2132   }
2133 }
2134 
2135 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2136 
2137   _capacity_at_prologue = capacity();
2138   _used_at_prologue = used();
2139 
2140   // Delegate to CMScollector which knows how to coordinate between
2141   // this and any other CMS generations that it is responsible for
2142   // collecting.
2143   collector()->gc_prologue(full);
2144 }
2145 
2146 // This is a "private" interface for use by this generation's CMSCollector.
2147 // Not to be called directly by any other entity (for instance,
2148 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2149 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2150   bool registerClosure, ModUnionClosure* modUnionClosure) {
2151   assert(!incremental_collection_failed(), "Shouldn't be set yet");
2152   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2153     "Should be NULL");
2154   if (registerClosure) {
2155     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2156   }
2157   cmsSpace()->gc_prologue();
2158   // Clear stat counters
2159   NOT_PRODUCT(
2160     assert(_numObjectsPromoted == 0, "check");
2161     assert(_numWordsPromoted   == 0, "check");
2162     if (Verbose && PrintGC) {
2163       gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, "
2164                           SIZE_FORMAT " bytes concurrently",
2165       _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2166     }
2167     _numObjectsAllocated = 0;
2168     _numWordsAllocated   = 0;
2169   )
2170 }
2171 
2172 void CMSCollector::gc_epilogue(bool full) {
2173   // The following locking discipline assumes that we are only called
2174   // when the world is stopped.
2175   assert(SafepointSynchronize::is_at_safepoint(),
2176          "world is stopped assumption");
2177 
2178   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2179   // if linear allocation blocks need to be appropriately marked to allow the
2180   // the blocks to be parsable. We also check here whether we need to nudge the
2181   // CMS collector thread to start a new cycle (if it's not already active).
2182   assert(   Thread::current()->is_VM_thread()
2183          || (   CMSScavengeBeforeRemark
2184              && Thread::current()->is_ConcurrentGC_thread()),
2185          "Incorrect thread type for epilogue execution");
2186 
2187   if (!_between_prologue_and_epilogue) {
2188     // We have already been invoked; this is a gc_epilogue delegation
2189     // from yet another CMS generation that we are responsible for, just
2190     // ignore it since all relevant work has already been done.
2191     return;
2192   }
2193   assert(haveFreelistLocks(), "must have freelist locks");
2194   assert_lock_strong(bitMapLock());
2195 
2196   _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2197 
2198   _cmsGen->gc_epilogue_work(full);
2199 
2200   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2201     // in case sampling was not already enabled, enable it
2202     _start_sampling = true;
2203   }
2204   // reset _eden_chunk_array so sampling starts afresh
2205   _eden_chunk_index = 0;
2206 
2207   size_t cms_used   = _cmsGen->cmsSpace()->used();
2208 
2209   // update performance counters - this uses a special version of
2210   // update_counters() that allows the utilization to be passed as a
2211   // parameter, avoiding multiple calls to used().
2212   //
2213   _cmsGen->update_counters(cms_used);
2214 
2215   bitMapLock()->unlock();
2216   releaseFreelistLocks();
2217 
2218   if (!CleanChunkPoolAsync) {
2219     Chunk::clean_chunk_pool();
2220   }
2221 
2222   set_did_compact(false);
2223   _between_prologue_and_epilogue = false;  // ready for next cycle
2224 }
2225 
2226 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2227   collector()->gc_epilogue(full);
2228 
2229   // Also reset promotion tracking in par gc thread states.
2230   for (uint i = 0; i < ParallelGCThreads; i++) {
2231     _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2232   }
2233 }
2234 
2235 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2236   assert(!incremental_collection_failed(), "Should have been cleared");
2237   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2238   cmsSpace()->gc_epilogue();
2239     // Print stat counters
2240   NOT_PRODUCT(
2241     assert(_numObjectsAllocated == 0, "check");
2242     assert(_numWordsAllocated == 0, "check");
2243     if (Verbose && PrintGC) {
2244       gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, "
2245                           SIZE_FORMAT " bytes",
2246                  _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2247     }
2248     _numObjectsPromoted = 0;
2249     _numWordsPromoted   = 0;
2250   )
2251 
2252   if (PrintGC && Verbose) {
2253     // Call down the chain in contiguous_available needs the freelistLock
2254     // so print this out before releasing the freeListLock.
2255     gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ",
2256                         contiguous_available());
2257   }
2258 }
2259 
2260 #ifndef PRODUCT
2261 bool CMSCollector::have_cms_token() {
2262   Thread* thr = Thread::current();
2263   if (thr->is_VM_thread()) {
2264     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2265   } else if (thr->is_ConcurrentGC_thread()) {
2266     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2267   } else if (thr->is_GC_task_thread()) {
2268     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2269            ParGCRareEvent_lock->owned_by_self();
2270   }
2271   return false;
2272 }
2273 #endif
2274 
2275 // Check reachability of the given heap address in CMS generation,
2276 // treating all other generations as roots.
2277 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2278   // We could "guarantee" below, rather than assert, but I'll
2279   // leave these as "asserts" so that an adventurous debugger
2280   // could try this in the product build provided some subset of
2281   // the conditions were met, provided they were interested in the
2282   // results and knew that the computation below wouldn't interfere
2283   // with other concurrent computations mutating the structures
2284   // being read or written.
2285   assert(SafepointSynchronize::is_at_safepoint(),
2286          "Else mutations in object graph will make answer suspect");
2287   assert(have_cms_token(), "Should hold cms token");
2288   assert(haveFreelistLocks(), "must hold free list locks");
2289   assert_lock_strong(bitMapLock());
2290 
2291   // Clear the marking bit map array before starting, but, just
2292   // for kicks, first report if the given address is already marked
2293   gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2294                 _markBitMap.isMarked(addr) ? "" : " not");
2295 
2296   if (verify_after_remark()) {
2297     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2298     bool result = verification_mark_bm()->isMarked(addr);
2299     gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2300                            result ? "IS" : "is NOT");
2301     return result;
2302   } else {
2303     gclog_or_tty->print_cr("Could not compute result");
2304     return false;
2305   }
2306 }
2307 
2308 
2309 void
2310 CMSCollector::print_on_error(outputStream* st) {
2311   CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2312   if (collector != NULL) {
2313     CMSBitMap* bitmap = &collector->_markBitMap;
2314     st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2315     bitmap->print_on_error(st, " Bits: ");
2316 
2317     st->cr();
2318 
2319     CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2320     st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2321     mut_bitmap->print_on_error(st, " Bits: ");
2322   }
2323 }
2324 
2325 ////////////////////////////////////////////////////////
2326 // CMS Verification Support
2327 ////////////////////////////////////////////////////////
2328 // Following the remark phase, the following invariant
2329 // should hold -- each object in the CMS heap which is
2330 // marked in markBitMap() should be marked in the verification_mark_bm().
2331 
2332 class VerifyMarkedClosure: public BitMapClosure {
2333   CMSBitMap* _marks;
2334   bool       _failed;
2335 
2336  public:
2337   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2338 
2339   bool do_bit(size_t offset) {
2340     HeapWord* addr = _marks->offsetToHeapWord(offset);
2341     if (!_marks->isMarked(addr)) {
2342       oop(addr)->print_on(gclog_or_tty);
2343       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2344       _failed = true;
2345     }
2346     return true;
2347   }
2348 
2349   bool failed() { return _failed; }
2350 };
2351 
2352 bool CMSCollector::verify_after_remark(bool silent) {
2353   if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2354   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2355   static bool init = false;
2356 
2357   assert(SafepointSynchronize::is_at_safepoint(),
2358          "Else mutations in object graph will make answer suspect");
2359   assert(have_cms_token(),
2360          "Else there may be mutual interference in use of "
2361          " verification data structures");
2362   assert(_collectorState > Marking && _collectorState <= Sweeping,
2363          "Else marking info checked here may be obsolete");
2364   assert(haveFreelistLocks(), "must hold free list locks");
2365   assert_lock_strong(bitMapLock());
2366 
2367 
2368   // Allocate marking bit map if not already allocated
2369   if (!init) { // first time
2370     if (!verification_mark_bm()->allocate(_span)) {
2371       return false;
2372     }
2373     init = true;
2374   }
2375 
2376   assert(verification_mark_stack()->isEmpty(), "Should be empty");
2377 
2378   // Turn off refs discovery -- so we will be tracing through refs.
2379   // This is as intended, because by this time
2380   // GC must already have cleared any refs that need to be cleared,
2381   // and traced those that need to be marked; moreover,
2382   // the marking done here is not going to interfere in any
2383   // way with the marking information used by GC.
2384   NoRefDiscovery no_discovery(ref_processor());
2385 
2386   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2387 
2388   // Clear any marks from a previous round
2389   verification_mark_bm()->clear_all();
2390   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2391   verify_work_stacks_empty();
2392 
2393   GenCollectedHeap* gch = GenCollectedHeap::heap();
2394   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2395   // Update the saved marks which may affect the root scans.
2396   gch->save_marks();
2397 
2398   if (CMSRemarkVerifyVariant == 1) {
2399     // In this first variant of verification, we complete
2400     // all marking, then check if the new marks-vector is
2401     // a subset of the CMS marks-vector.
2402     verify_after_remark_work_1();
2403   } else if (CMSRemarkVerifyVariant == 2) {
2404     // In this second variant of verification, we flag an error
2405     // (i.e. an object reachable in the new marks-vector not reachable
2406     // in the CMS marks-vector) immediately, also indicating the
2407     // identify of an object (A) that references the unmarked object (B) --
2408     // presumably, a mutation to A failed to be picked up by preclean/remark?
2409     verify_after_remark_work_2();
2410   } else {
2411     warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2412             CMSRemarkVerifyVariant);
2413   }
2414   if (!silent) gclog_or_tty->print(" done] ");
2415   return true;
2416 }
2417 
2418 void CMSCollector::verify_after_remark_work_1() {
2419   ResourceMark rm;
2420   HandleMark  hm;
2421   GenCollectedHeap* gch = GenCollectedHeap::heap();
2422 
2423   // Get a clear set of claim bits for the roots processing to work with.
2424   ClassLoaderDataGraph::clear_claimed_marks();
2425 
2426   // Mark from roots one level into CMS
2427   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2428   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2429 
2430   {
2431     StrongRootsScope srs(1);
2432 
2433     gch->gen_process_roots(&srs,
2434                            GenCollectedHeap::OldGen,
2435                            true,   // younger gens are roots
2436                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2437                            should_unload_classes(),
2438                            &notOlder,
2439                            NULL,
2440                            NULL);
2441   }
2442 
2443   // Now mark from the roots
2444   MarkFromRootsClosure markFromRootsClosure(this, _span,
2445     verification_mark_bm(), verification_mark_stack(),
2446     false /* don't yield */, true /* verifying */);
2447   assert(_restart_addr == NULL, "Expected pre-condition");
2448   verification_mark_bm()->iterate(&markFromRootsClosure);
2449   while (_restart_addr != NULL) {
2450     // Deal with stack overflow: by restarting at the indicated
2451     // address.
2452     HeapWord* ra = _restart_addr;
2453     markFromRootsClosure.reset(ra);
2454     _restart_addr = NULL;
2455     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2456   }
2457   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2458   verify_work_stacks_empty();
2459 
2460   // Marking completed -- now verify that each bit marked in
2461   // verification_mark_bm() is also marked in markBitMap(); flag all
2462   // errors by printing corresponding objects.
2463   VerifyMarkedClosure vcl(markBitMap());
2464   verification_mark_bm()->iterate(&vcl);
2465   if (vcl.failed()) {
2466     gclog_or_tty->print("Verification failed");
2467     gch->print_on(gclog_or_tty);
2468     fatal("CMS: failed marking verification after remark");
2469   }
2470 }
2471 
2472 class VerifyKlassOopsKlassClosure : public KlassClosure {
2473   class VerifyKlassOopsClosure : public OopClosure {
2474     CMSBitMap* _bitmap;
2475    public:
2476     VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2477     void do_oop(oop* p)       { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2478     void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2479   } _oop_closure;
2480  public:
2481   VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2482   void do_klass(Klass* k) {
2483     k->oops_do(&_oop_closure);
2484   }
2485 };
2486 
2487 void CMSCollector::verify_after_remark_work_2() {
2488   ResourceMark rm;
2489   HandleMark  hm;
2490   GenCollectedHeap* gch = GenCollectedHeap::heap();
2491 
2492   // Get a clear set of claim bits for the roots processing to work with.
2493   ClassLoaderDataGraph::clear_claimed_marks();
2494 
2495   // Mark from roots one level into CMS
2496   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2497                                      markBitMap());
2498   CLDToOopClosure cld_closure(&notOlder, true);
2499 
2500   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2501 
2502   {
2503     StrongRootsScope srs(1);
2504 
2505     gch->gen_process_roots(&srs,
2506                            GenCollectedHeap::OldGen,
2507                            true,   // younger gens are roots
2508                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2509                            should_unload_classes(),
2510                            &notOlder,
2511                            NULL,
2512                            &cld_closure);
2513   }
2514 
2515   // Now mark from the roots
2516   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2517     verification_mark_bm(), markBitMap(), verification_mark_stack());
2518   assert(_restart_addr == NULL, "Expected pre-condition");
2519   verification_mark_bm()->iterate(&markFromRootsClosure);
2520   while (_restart_addr != NULL) {
2521     // Deal with stack overflow: by restarting at the indicated
2522     // address.
2523     HeapWord* ra = _restart_addr;
2524     markFromRootsClosure.reset(ra);
2525     _restart_addr = NULL;
2526     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2527   }
2528   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2529   verify_work_stacks_empty();
2530 
2531   VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2532   ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2533 
2534   // Marking completed -- now verify that each bit marked in
2535   // verification_mark_bm() is also marked in markBitMap(); flag all
2536   // errors by printing corresponding objects.
2537   VerifyMarkedClosure vcl(markBitMap());
2538   verification_mark_bm()->iterate(&vcl);
2539   assert(!vcl.failed(), "Else verification above should not have succeeded");
2540 }
2541 
2542 void ConcurrentMarkSweepGeneration::save_marks() {
2543   // delegate to CMS space
2544   cmsSpace()->save_marks();
2545   for (uint i = 0; i < ParallelGCThreads; i++) {
2546     _par_gc_thread_states[i]->promo.startTrackingPromotions();
2547   }
2548 }
2549 
2550 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2551   return cmsSpace()->no_allocs_since_save_marks();
2552 }
2553 
2554 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
2555                                                                 \
2556 void ConcurrentMarkSweepGeneration::                            \
2557 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
2558   cl->set_generation(this);                                     \
2559   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
2560   cl->reset_generation();                                       \
2561   save_marks();                                                 \
2562 }
2563 
2564 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2565 
2566 void
2567 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2568   if (freelistLock()->owned_by_self()) {
2569     Generation::oop_iterate(cl);
2570   } else {
2571     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2572     Generation::oop_iterate(cl);
2573   }
2574 }
2575 
2576 void
2577 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2578   if (freelistLock()->owned_by_self()) {
2579     Generation::object_iterate(cl);
2580   } else {
2581     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2582     Generation::object_iterate(cl);
2583   }
2584 }
2585 
2586 void
2587 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2588   if (freelistLock()->owned_by_self()) {
2589     Generation::safe_object_iterate(cl);
2590   } else {
2591     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2592     Generation::safe_object_iterate(cl);
2593   }
2594 }
2595 
2596 void
2597 ConcurrentMarkSweepGeneration::post_compact() {
2598 }
2599 
2600 void
2601 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2602   // Fix the linear allocation blocks to look like free blocks.
2603 
2604   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2605   // are not called when the heap is verified during universe initialization and
2606   // at vm shutdown.
2607   if (freelistLock()->owned_by_self()) {
2608     cmsSpace()->prepare_for_verify();
2609   } else {
2610     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2611     cmsSpace()->prepare_for_verify();
2612   }
2613 }
2614 
2615 void
2616 ConcurrentMarkSweepGeneration::verify() {
2617   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2618   // are not called when the heap is verified during universe initialization and
2619   // at vm shutdown.
2620   if (freelistLock()->owned_by_self()) {
2621     cmsSpace()->verify();
2622   } else {
2623     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2624     cmsSpace()->verify();
2625   }
2626 }
2627 
2628 void CMSCollector::verify() {
2629   _cmsGen->verify();
2630 }
2631 
2632 #ifndef PRODUCT
2633 bool CMSCollector::overflow_list_is_empty() const {
2634   assert(_num_par_pushes >= 0, "Inconsistency");
2635   if (_overflow_list == NULL) {
2636     assert(_num_par_pushes == 0, "Inconsistency");
2637   }
2638   return _overflow_list == NULL;
2639 }
2640 
2641 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2642 // merely consolidate assertion checks that appear to occur together frequently.
2643 void CMSCollector::verify_work_stacks_empty() const {
2644   assert(_markStack.isEmpty(), "Marking stack should be empty");
2645   assert(overflow_list_is_empty(), "Overflow list should be empty");
2646 }
2647 
2648 void CMSCollector::verify_overflow_empty() const {
2649   assert(overflow_list_is_empty(), "Overflow list should be empty");
2650   assert(no_preserved_marks(), "No preserved marks");
2651 }
2652 #endif // PRODUCT
2653 
2654 // Decide if we want to enable class unloading as part of the
2655 // ensuing concurrent GC cycle. We will collect and
2656 // unload classes if it's the case that:
2657 // (1) an explicit gc request has been made and the flag
2658 //     ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2659 // (2) (a) class unloading is enabled at the command line, and
2660 //     (b) old gen is getting really full
2661 // NOTE: Provided there is no change in the state of the heap between
2662 // calls to this method, it should have idempotent results. Moreover,
2663 // its results should be monotonically increasing (i.e. going from 0 to 1,
2664 // but not 1 to 0) between successive calls between which the heap was
2665 // not collected. For the implementation below, it must thus rely on
2666 // the property that concurrent_cycles_since_last_unload()
2667 // will not decrease unless a collection cycle happened and that
2668 // _cmsGen->is_too_full() are
2669 // themselves also monotonic in that sense. See check_monotonicity()
2670 // below.
2671 void CMSCollector::update_should_unload_classes() {
2672   _should_unload_classes = false;
2673   // Condition 1 above
2674   if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2675     _should_unload_classes = true;
2676   } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2677     // Disjuncts 2.b.(i,ii,iii) above
2678     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2679                               CMSClassUnloadingMaxInterval)
2680                            || _cmsGen->is_too_full();
2681   }
2682 }
2683 
2684 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2685   bool res = should_concurrent_collect();
2686   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2687   return res;
2688 }
2689 
2690 void CMSCollector::setup_cms_unloading_and_verification_state() {
2691   const  bool should_verify =   VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2692                              || VerifyBeforeExit;
2693   const  int  rso           =   GenCollectedHeap::SO_AllCodeCache;
2694 
2695   // We set the proper root for this CMS cycle here.
2696   if (should_unload_classes()) {   // Should unload classes this cycle
2697     remove_root_scanning_option(rso);  // Shrink the root set appropriately
2698     set_verifying(should_verify);    // Set verification state for this cycle
2699     return;                            // Nothing else needs to be done at this time
2700   }
2701 
2702   // Not unloading classes this cycle
2703   assert(!should_unload_classes(), "Inconsistency!");
2704 
2705   // If we are not unloading classes then add SO_AllCodeCache to root
2706   // scanning options.
2707   add_root_scanning_option(rso);
2708 
2709   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2710     set_verifying(true);
2711   } else if (verifying() && !should_verify) {
2712     // We were verifying, but some verification flags got disabled.
2713     set_verifying(false);
2714     // Exclude symbols, strings and code cache elements from root scanning to
2715     // reduce IM and RM pauses.
2716     remove_root_scanning_option(rso);
2717   }
2718 }
2719 
2720 
2721 #ifndef PRODUCT
2722 HeapWord* CMSCollector::block_start(const void* p) const {
2723   const HeapWord* addr = (HeapWord*)p;
2724   if (_span.contains(p)) {
2725     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2726       return _cmsGen->cmsSpace()->block_start(p);
2727     }
2728   }
2729   return NULL;
2730 }
2731 #endif
2732 
2733 HeapWord*
2734 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2735                                                    bool   tlab,
2736                                                    bool   parallel) {
2737   CMSSynchronousYieldRequest yr;
2738   assert(!tlab, "Can't deal with TLAB allocation");
2739   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2740   expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2741   if (GCExpandToAllocateDelayMillis > 0) {
2742     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2743   }
2744   return have_lock_and_allocate(word_size, tlab);
2745 }
2746 
2747 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2748     size_t bytes,
2749     size_t expand_bytes,
2750     CMSExpansionCause::Cause cause)
2751 {
2752 
2753   bool success = expand(bytes, expand_bytes);
2754 
2755   // remember why we expanded; this information is used
2756   // by shouldConcurrentCollect() when making decisions on whether to start
2757   // a new CMS cycle.
2758   if (success) {
2759     set_expansion_cause(cause);
2760     if (PrintGCDetails && Verbose) {
2761       gclog_or_tty->print_cr("Expanded CMS gen for %s",
2762         CMSExpansionCause::to_string(cause));
2763     }
2764   }
2765 }
2766 
2767 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2768   HeapWord* res = NULL;
2769   MutexLocker x(ParGCRareEvent_lock);
2770   while (true) {
2771     // Expansion by some other thread might make alloc OK now:
2772     res = ps->lab.alloc(word_sz);
2773     if (res != NULL) return res;
2774     // If there's not enough expansion space available, give up.
2775     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2776       return NULL;
2777     }
2778     // Otherwise, we try expansion.
2779     expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2780     // Now go around the loop and try alloc again;
2781     // A competing par_promote might beat us to the expansion space,
2782     // so we may go around the loop again if promotion fails again.
2783     if (GCExpandToAllocateDelayMillis > 0) {
2784       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2785     }
2786   }
2787 }
2788 
2789 
2790 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2791   PromotionInfo* promo) {
2792   MutexLocker x(ParGCRareEvent_lock);
2793   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2794   while (true) {
2795     // Expansion by some other thread might make alloc OK now:
2796     if (promo->ensure_spooling_space()) {
2797       assert(promo->has_spooling_space(),
2798              "Post-condition of successful ensure_spooling_space()");
2799       return true;
2800     }
2801     // If there's not enough expansion space available, give up.
2802     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2803       return false;
2804     }
2805     // Otherwise, we try expansion.
2806     expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2807     // Now go around the loop and try alloc again;
2808     // A competing allocation might beat us to the expansion space,
2809     // so we may go around the loop again if allocation fails again.
2810     if (GCExpandToAllocateDelayMillis > 0) {
2811       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2812     }
2813   }
2814 }
2815 
2816 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2817   // Only shrink if a compaction was done so that all the free space
2818   // in the generation is in a contiguous block at the end.
2819   if (did_compact()) {
2820     CardGeneration::shrink(bytes);
2821   }
2822 }
2823 
2824 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2825   assert_locked_or_safepoint(Heap_lock);
2826 }
2827 
2828 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2829   assert_locked_or_safepoint(Heap_lock);
2830   assert_lock_strong(freelistLock());
2831   if (PrintGCDetails && Verbose) {
2832     warning("Shrinking of CMS not yet implemented");
2833   }
2834   return;
2835 }
2836 
2837 
2838 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2839 // phases.
2840 class CMSPhaseAccounting: public StackObj {
2841  public:
2842   CMSPhaseAccounting(CMSCollector *collector,
2843                      const char *phase,
2844                      const GCId gc_id,
2845                      bool print_cr = true);
2846   ~CMSPhaseAccounting();
2847 
2848  private:
2849   CMSCollector *_collector;
2850   const char *_phase;
2851   elapsedTimer _wallclock;
2852   bool _print_cr;
2853   const GCId _gc_id;
2854 
2855  public:
2856   // Not MT-safe; so do not pass around these StackObj's
2857   // where they may be accessed by other threads.
2858   jlong wallclock_millis() {
2859     assert(_wallclock.is_active(), "Wall clock should not stop");
2860     _wallclock.stop();  // to record time
2861     jlong ret = _wallclock.milliseconds();
2862     _wallclock.start(); // restart
2863     return ret;
2864   }
2865 };
2866 
2867 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2868                                        const char *phase,
2869                                        const GCId gc_id,
2870                                        bool print_cr) :
2871   _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2872 
2873   if (PrintCMSStatistics != 0) {
2874     _collector->resetYields();
2875   }
2876   if (PrintGCDetails) {
2877     gclog_or_tty->gclog_stamp(_gc_id);
2878     gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2879       _collector->cmsGen()->short_name(), _phase);
2880   }
2881   _collector->resetTimer();
2882   _wallclock.start();
2883   _collector->startTimer();
2884 }
2885 
2886 CMSPhaseAccounting::~CMSPhaseAccounting() {
2887   assert(_wallclock.is_active(), "Wall clock should not have stopped");
2888   _collector->stopTimer();
2889   _wallclock.stop();
2890   if (PrintGCDetails) {
2891     gclog_or_tty->gclog_stamp(_gc_id);
2892     gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2893                  _collector->cmsGen()->short_name(),
2894                  _phase, _collector->timerValue(), _wallclock.seconds());
2895     if (_print_cr) {
2896       gclog_or_tty->cr();
2897     }
2898     if (PrintCMSStatistics != 0) {
2899       gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2900                     _collector->yields());
2901     }
2902   }
2903 }
2904 
2905 // CMS work
2906 
2907 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2908 class CMSParMarkTask : public AbstractGangTask {
2909  protected:
2910   CMSCollector*     _collector;
2911   uint              _n_workers;
2912   CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2913       AbstractGangTask(name),
2914       _collector(collector),
2915       _n_workers(n_workers) {}
2916   // Work method in support of parallel rescan ... of young gen spaces
2917   void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2918                              ContiguousSpace* space,
2919                              HeapWord** chunk_array, size_t chunk_top);
2920   void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2921 };
2922 
2923 // Parallel initial mark task
2924 class CMSParInitialMarkTask: public CMSParMarkTask {
2925   StrongRootsScope* _strong_roots_scope;
2926  public:
2927   CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2928       CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2929       _strong_roots_scope(strong_roots_scope) {}
2930   void work(uint worker_id);
2931 };
2932 
2933 // Checkpoint the roots into this generation from outside
2934 // this generation. [Note this initial checkpoint need only
2935 // be approximate -- we'll do a catch up phase subsequently.]
2936 void CMSCollector::checkpointRootsInitial() {
2937   assert(_collectorState == InitialMarking, "Wrong collector state");
2938   check_correct_thread_executing();
2939   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2940 
2941   save_heap_summary();
2942   report_heap_summary(GCWhen::BeforeGC);
2943 
2944   ReferenceProcessor* rp = ref_processor();
2945   assert(_restart_addr == NULL, "Control point invariant");
2946   {
2947     // acquire locks for subsequent manipulations
2948     MutexLockerEx x(bitMapLock(),
2949                     Mutex::_no_safepoint_check_flag);
2950     checkpointRootsInitialWork();
2951     // enable ("weak") refs discovery
2952     rp->enable_discovery();
2953     _collectorState = Marking;
2954   }
2955 }
2956 
2957 void CMSCollector::checkpointRootsInitialWork() {
2958   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2959   assert(_collectorState == InitialMarking, "just checking");
2960 
2961   // If there has not been a GC[n-1] since last GC[n] cycle completed,
2962   // precede our marking with a collection of all
2963   // younger generations to keep floating garbage to a minimum.
2964   // XXX: we won't do this for now -- it's an optimization to be done later.
2965 
2966   // already have locks
2967   assert_lock_strong(bitMapLock());
2968   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2969 
2970   // Setup the verification and class unloading state for this
2971   // CMS collection cycle.
2972   setup_cms_unloading_and_verification_state();
2973 
2974   NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2975     PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
2976 
2977   // Reset all the PLAB chunk arrays if necessary.
2978   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2979     reset_survivor_plab_arrays();
2980   }
2981 
2982   ResourceMark rm;
2983   HandleMark  hm;
2984 
2985   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2986   GenCollectedHeap* gch = GenCollectedHeap::heap();
2987 
2988   verify_work_stacks_empty();
2989   verify_overflow_empty();
2990 
2991   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2992   // Update the saved marks which may affect the root scans.
2993   gch->save_marks();
2994 
2995   // weak reference processing has not started yet.
2996   ref_processor()->set_enqueuing_is_done(false);
2997 
2998   // Need to remember all newly created CLDs,
2999   // so that we can guarantee that the remark finds them.
3000   ClassLoaderDataGraph::remember_new_clds(true);
3001 
3002   // Whenever a CLD is found, it will be claimed before proceeding to mark
3003   // the klasses. The claimed marks need to be cleared before marking starts.
3004   ClassLoaderDataGraph::clear_claimed_marks();
3005 
3006   if (CMSPrintEdenSurvivorChunks) {
3007     print_eden_and_survivor_chunk_arrays();
3008   }
3009 
3010   {
3011     COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3012     if (CMSParallelInitialMarkEnabled) {
3013       // The parallel version.
3014       FlexibleWorkGang* workers = gch->workers();
3015       assert(workers != NULL, "Need parallel worker threads.");
3016       uint n_workers = workers->active_workers();
3017 
3018       StrongRootsScope srs(n_workers);
3019 
3020       CMSParInitialMarkTask tsk(this, &srs, n_workers);
3021       initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3022       if (n_workers > 1) {
3023         workers->run_task(&tsk);
3024       } else {
3025         tsk.work(0);
3026       }
3027     } else {
3028       // The serial version.
3029       CLDToOopClosure cld_closure(&notOlder, true);
3030       gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3031 
3032       StrongRootsScope srs(1);
3033 
3034       gch->gen_process_roots(&srs,
3035                              GenCollectedHeap::OldGen,
3036                              true,   // younger gens are roots
3037                              GenCollectedHeap::ScanningOption(roots_scanning_options()),
3038                              should_unload_classes(),
3039                              &notOlder,
3040                              NULL,
3041                              &cld_closure);
3042     }
3043   }
3044 
3045   // Clear mod-union table; it will be dirtied in the prologue of
3046   // CMS generation per each younger generation collection.
3047 
3048   assert(_modUnionTable.isAllClear(),
3049        "Was cleared in most recent final checkpoint phase"
3050        " or no bits are set in the gc_prologue before the start of the next "
3051        "subsequent marking phase.");
3052 
3053   assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3054 
3055   // Save the end of the used_region of the constituent generations
3056   // to be used to limit the extent of sweep in each generation.
3057   save_sweep_limits();
3058   verify_overflow_empty();
3059 }
3060 
3061 bool CMSCollector::markFromRoots() {
3062   // we might be tempted to assert that:
3063   // assert(!SafepointSynchronize::is_at_safepoint(),
3064   //        "inconsistent argument?");
3065   // However that wouldn't be right, because it's possible that
3066   // a safepoint is indeed in progress as a younger generation
3067   // stop-the-world GC happens even as we mark in this generation.
3068   assert(_collectorState == Marking, "inconsistent state?");
3069   check_correct_thread_executing();
3070   verify_overflow_empty();
3071 
3072   // Weak ref discovery note: We may be discovering weak
3073   // refs in this generation concurrent (but interleaved) with
3074   // weak ref discovery by a younger generation collector.
3075 
3076   CMSTokenSyncWithLocks ts(true, bitMapLock());
3077   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3078   CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3079   bool res = markFromRootsWork();
3080   if (res) {
3081     _collectorState = Precleaning;
3082   } else { // We failed and a foreground collection wants to take over
3083     assert(_foregroundGCIsActive, "internal state inconsistency");
3084     assert(_restart_addr == NULL,  "foreground will restart from scratch");
3085     if (PrintGCDetails) {
3086       gclog_or_tty->print_cr("bailing out to foreground collection");
3087     }
3088   }
3089   verify_overflow_empty();
3090   return res;
3091 }
3092 
3093 bool CMSCollector::markFromRootsWork() {
3094   // iterate over marked bits in bit map, doing a full scan and mark
3095   // from these roots using the following algorithm:
3096   // . if oop is to the right of the current scan pointer,
3097   //   mark corresponding bit (we'll process it later)
3098   // . else (oop is to left of current scan pointer)
3099   //   push oop on marking stack
3100   // . drain the marking stack
3101 
3102   // Note that when we do a marking step we need to hold the
3103   // bit map lock -- recall that direct allocation (by mutators)
3104   // and promotion (by younger generation collectors) is also
3105   // marking the bit map. [the so-called allocate live policy.]
3106   // Because the implementation of bit map marking is not
3107   // robust wrt simultaneous marking of bits in the same word,
3108   // we need to make sure that there is no such interference
3109   // between concurrent such updates.
3110 
3111   // already have locks
3112   assert_lock_strong(bitMapLock());
3113 
3114   verify_work_stacks_empty();
3115   verify_overflow_empty();
3116   bool result = false;
3117   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3118     result = do_marking_mt();
3119   } else {
3120     result = do_marking_st();
3121   }
3122   return result;
3123 }
3124 
3125 // Forward decl
3126 class CMSConcMarkingTask;
3127 
3128 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3129   CMSCollector*       _collector;
3130   CMSConcMarkingTask* _task;
3131  public:
3132   virtual void yield();
3133 
3134   // "n_threads" is the number of threads to be terminated.
3135   // "queue_set" is a set of work queues of other threads.
3136   // "collector" is the CMS collector associated with this task terminator.
3137   // "yield" indicates whether we need the gang as a whole to yield.
3138   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3139     ParallelTaskTerminator(n_threads, queue_set),
3140     _collector(collector) { }
3141 
3142   void set_task(CMSConcMarkingTask* task) {
3143     _task = task;
3144   }
3145 };
3146 
3147 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3148   CMSConcMarkingTask* _task;
3149  public:
3150   bool should_exit_termination();
3151   void set_task(CMSConcMarkingTask* task) {
3152     _task = task;
3153   }
3154 };
3155 
3156 // MT Concurrent Marking Task
3157 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3158   CMSCollector* _collector;
3159   uint          _n_workers;       // requested/desired # workers
3160   bool          _result;
3161   CompactibleFreeListSpace*  _cms_space;
3162   char          _pad_front[64];   // padding to ...
3163   HeapWord*     _global_finger;   // ... avoid sharing cache line
3164   char          _pad_back[64];
3165   HeapWord*     _restart_addr;
3166 
3167   //  Exposed here for yielding support
3168   Mutex* const _bit_map_lock;
3169 
3170   // The per thread work queues, available here for stealing
3171   OopTaskQueueSet*  _task_queues;
3172 
3173   // Termination (and yielding) support
3174   CMSConcMarkingTerminator _term;
3175   CMSConcMarkingTerminatorTerminator _term_term;
3176 
3177  public:
3178   CMSConcMarkingTask(CMSCollector* collector,
3179                  CompactibleFreeListSpace* cms_space,
3180                  YieldingFlexibleWorkGang* workers,
3181                  OopTaskQueueSet* task_queues):
3182     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3183     _collector(collector),
3184     _cms_space(cms_space),
3185     _n_workers(0), _result(true),
3186     _task_queues(task_queues),
3187     _term(_n_workers, task_queues, _collector),
3188     _bit_map_lock(collector->bitMapLock())
3189   {
3190     _requested_size = _n_workers;
3191     _term.set_task(this);
3192     _term_term.set_task(this);
3193     _restart_addr = _global_finger = _cms_space->bottom();
3194   }
3195 
3196 
3197   OopTaskQueueSet* task_queues()  { return _task_queues; }
3198 
3199   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3200 
3201   HeapWord** global_finger_addr() { return &_global_finger; }
3202 
3203   CMSConcMarkingTerminator* terminator() { return &_term; }
3204 
3205   virtual void set_for_termination(uint active_workers) {
3206     terminator()->reset_for_reuse(active_workers);
3207   }
3208 
3209   void work(uint worker_id);
3210   bool should_yield() {
3211     return    ConcurrentMarkSweepThread::should_yield()
3212            && !_collector->foregroundGCIsActive();
3213   }
3214 
3215   virtual void coordinator_yield();  // stuff done by coordinator
3216   bool result() { return _result; }
3217 
3218   void reset(HeapWord* ra) {
3219     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
3220     _restart_addr = _global_finger = ra;
3221     _term.reset_for_reuse();
3222   }
3223 
3224   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3225                                            OopTaskQueue* work_q);
3226 
3227  private:
3228   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3229   void do_work_steal(int i);
3230   void bump_global_finger(HeapWord* f);
3231 };
3232 
3233 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3234   assert(_task != NULL, "Error");
3235   return _task->yielding();
3236   // Note that we do not need the disjunct || _task->should_yield() above
3237   // because we want terminating threads to yield only if the task
3238   // is already in the midst of yielding, which happens only after at least one
3239   // thread has yielded.
3240 }
3241 
3242 void CMSConcMarkingTerminator::yield() {
3243   if (_task->should_yield()) {
3244     _task->yield();
3245   } else {
3246     ParallelTaskTerminator::yield();
3247   }
3248 }
3249 
3250 ////////////////////////////////////////////////////////////////
3251 // Concurrent Marking Algorithm Sketch
3252 ////////////////////////////////////////////////////////////////
3253 // Until all tasks exhausted (both spaces):
3254 // -- claim next available chunk
3255 // -- bump global finger via CAS
3256 // -- find first object that starts in this chunk
3257 //    and start scanning bitmap from that position
3258 // -- scan marked objects for oops
3259 // -- CAS-mark target, and if successful:
3260 //    . if target oop is above global finger (volatile read)
3261 //      nothing to do
3262 //    . if target oop is in chunk and above local finger
3263 //        then nothing to do
3264 //    . else push on work-queue
3265 // -- Deal with possible overflow issues:
3266 //    . local work-queue overflow causes stuff to be pushed on
3267 //      global (common) overflow queue
3268 //    . always first empty local work queue
3269 //    . then get a batch of oops from global work queue if any
3270 //    . then do work stealing
3271 // -- When all tasks claimed (both spaces)
3272 //    and local work queue empty,
3273 //    then in a loop do:
3274 //    . check global overflow stack; steal a batch of oops and trace
3275 //    . try to steal from other threads oif GOS is empty
3276 //    . if neither is available, offer termination
3277 // -- Terminate and return result
3278 //
3279 void CMSConcMarkingTask::work(uint worker_id) {
3280   elapsedTimer _timer;
3281   ResourceMark rm;
3282   HandleMark hm;
3283 
3284   DEBUG_ONLY(_collector->verify_overflow_empty();)
3285 
3286   // Before we begin work, our work queue should be empty
3287   assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3288   // Scan the bitmap covering _cms_space, tracing through grey objects.
3289   _timer.start();
3290   do_scan_and_mark(worker_id, _cms_space);
3291   _timer.stop();
3292   if (PrintCMSStatistics != 0) {
3293     gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3294       worker_id, _timer.seconds());
3295       // XXX: need xxx/xxx type of notation, two timers
3296   }
3297 
3298   // ... do work stealing
3299   _timer.reset();
3300   _timer.start();
3301   do_work_steal(worker_id);
3302   _timer.stop();
3303   if (PrintCMSStatistics != 0) {
3304     gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3305       worker_id, _timer.seconds());
3306       // XXX: need xxx/xxx type of notation, two timers
3307   }
3308   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3309   assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3310   // Note that under the current task protocol, the
3311   // following assertion is true even of the spaces
3312   // expanded since the completion of the concurrent
3313   // marking. XXX This will likely change under a strict
3314   // ABORT semantics.
3315   // After perm removal the comparison was changed to
3316   // greater than or equal to from strictly greater than.
3317   // Before perm removal the highest address sweep would
3318   // have been at the end of perm gen but now is at the
3319   // end of the tenured gen.
3320   assert(_global_finger >=  _cms_space->end(),
3321          "All tasks have been completed");
3322   DEBUG_ONLY(_collector->verify_overflow_empty();)
3323 }
3324 
3325 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3326   HeapWord* read = _global_finger;
3327   HeapWord* cur  = read;
3328   while (f > read) {
3329     cur = read;
3330     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3331     if (cur == read) {
3332       // our cas succeeded
3333       assert(_global_finger >= f, "protocol consistency");
3334       break;
3335     }
3336   }
3337 }
3338 
3339 // This is really inefficient, and should be redone by
3340 // using (not yet available) block-read and -write interfaces to the
3341 // stack and the work_queue. XXX FIX ME !!!
3342 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3343                                                       OopTaskQueue* work_q) {
3344   // Fast lock-free check
3345   if (ovflw_stk->length() == 0) {
3346     return false;
3347   }
3348   assert(work_q->size() == 0, "Shouldn't steal");
3349   MutexLockerEx ml(ovflw_stk->par_lock(),
3350                    Mutex::_no_safepoint_check_flag);
3351   // Grab up to 1/4 the size of the work queue
3352   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3353                     (size_t)ParGCDesiredObjsFromOverflowList);
3354   num = MIN2(num, ovflw_stk->length());
3355   for (int i = (int) num; i > 0; i--) {
3356     oop cur = ovflw_stk->pop();
3357     assert(cur != NULL, "Counted wrong?");
3358     work_q->push(cur);
3359   }
3360   return num > 0;
3361 }
3362 
3363 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3364   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3365   int n_tasks = pst->n_tasks();
3366   // We allow that there may be no tasks to do here because
3367   // we are restarting after a stack overflow.
3368   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3369   uint nth_task = 0;
3370 
3371   HeapWord* aligned_start = sp->bottom();
3372   if (sp->used_region().contains(_restart_addr)) {
3373     // Align down to a card boundary for the start of 0th task
3374     // for this space.
3375     aligned_start =
3376       (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3377                                  CardTableModRefBS::card_size);
3378   }
3379 
3380   size_t chunk_size = sp->marking_task_size();
3381   while (!pst->is_task_claimed(/* reference */ nth_task)) {
3382     // Having claimed the nth task in this space,
3383     // compute the chunk that it corresponds to:
3384     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3385                                aligned_start + (nth_task+1)*chunk_size);
3386     // Try and bump the global finger via a CAS;
3387     // note that we need to do the global finger bump
3388     // _before_ taking the intersection below, because
3389     // the task corresponding to that region will be
3390     // deemed done even if the used_region() expands
3391     // because of allocation -- as it almost certainly will
3392     // during start-up while the threads yield in the
3393     // closure below.
3394     HeapWord* finger = span.end();
3395     bump_global_finger(finger);   // atomically
3396     // There are null tasks here corresponding to chunks
3397     // beyond the "top" address of the space.
3398     span = span.intersection(sp->used_region());
3399     if (!span.is_empty()) {  // Non-null task
3400       HeapWord* prev_obj;
3401       assert(!span.contains(_restart_addr) || nth_task == 0,
3402              "Inconsistency");
3403       if (nth_task == 0) {
3404         // For the 0th task, we'll not need to compute a block_start.
3405         if (span.contains(_restart_addr)) {
3406           // In the case of a restart because of stack overflow,
3407           // we might additionally skip a chunk prefix.
3408           prev_obj = _restart_addr;
3409         } else {
3410           prev_obj = span.start();
3411         }
3412       } else {
3413         // We want to skip the first object because
3414         // the protocol is to scan any object in its entirety
3415         // that _starts_ in this span; a fortiori, any
3416         // object starting in an earlier span is scanned
3417         // as part of an earlier claimed task.
3418         // Below we use the "careful" version of block_start
3419         // so we do not try to navigate uninitialized objects.
3420         prev_obj = sp->block_start_careful(span.start());
3421         // Below we use a variant of block_size that uses the
3422         // Printezis bits to avoid waiting for allocated
3423         // objects to become initialized/parsable.
3424         while (prev_obj < span.start()) {
3425           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3426           if (sz > 0) {
3427             prev_obj += sz;
3428           } else {
3429             // In this case we may end up doing a bit of redundant
3430             // scanning, but that appears unavoidable, short of
3431             // locking the free list locks; see bug 6324141.
3432             break;
3433           }
3434         }
3435       }
3436       if (prev_obj < span.end()) {
3437         MemRegion my_span = MemRegion(prev_obj, span.end());
3438         // Do the marking work within a non-empty span --
3439         // the last argument to the constructor indicates whether the
3440         // iteration should be incremental with periodic yields.
3441         Par_MarkFromRootsClosure cl(this, _collector, my_span,
3442                                     &_collector->_markBitMap,
3443                                     work_queue(i),
3444                                     &_collector->_markStack);
3445         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3446       } // else nothing to do for this task
3447     }   // else nothing to do for this task
3448   }
3449   // We'd be tempted to assert here that since there are no
3450   // more tasks left to claim in this space, the global_finger
3451   // must exceed space->top() and a fortiori space->end(). However,
3452   // that would not quite be correct because the bumping of
3453   // global_finger occurs strictly after the claiming of a task,
3454   // so by the time we reach here the global finger may not yet
3455   // have been bumped up by the thread that claimed the last
3456   // task.
3457   pst->all_tasks_completed();
3458 }
3459 
3460 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3461  private:
3462   CMSCollector* _collector;
3463   CMSConcMarkingTask* _task;
3464   MemRegion     _span;
3465   CMSBitMap*    _bit_map;
3466   CMSMarkStack* _overflow_stack;
3467   OopTaskQueue* _work_queue;
3468  protected:
3469   DO_OOP_WORK_DEFN
3470  public:
3471   Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3472                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3473     MetadataAwareOopClosure(collector->ref_processor()),
3474     _collector(collector),
3475     _task(task),
3476     _span(collector->_span),
3477     _work_queue(work_queue),
3478     _bit_map(bit_map),
3479     _overflow_stack(overflow_stack)
3480   { }
3481   virtual void do_oop(oop* p);
3482   virtual void do_oop(narrowOop* p);
3483 
3484   void trim_queue(size_t max);
3485   void handle_stack_overflow(HeapWord* lost);
3486   void do_yield_check() {
3487     if (_task->should_yield()) {
3488       _task->yield();
3489     }
3490   }
3491 };
3492 
3493 // Grey object scanning during work stealing phase --
3494 // the salient assumption here is that any references
3495 // that are in these stolen objects being scanned must
3496 // already have been initialized (else they would not have
3497 // been published), so we do not need to check for
3498 // uninitialized objects before pushing here.
3499 void Par_ConcMarkingClosure::do_oop(oop obj) {
3500   assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3501   HeapWord* addr = (HeapWord*)obj;
3502   // Check if oop points into the CMS generation
3503   // and is not marked
3504   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3505     // a white object ...
3506     // If we manage to "claim" the object, by being the
3507     // first thread to mark it, then we push it on our
3508     // marking stack
3509     if (_bit_map->par_mark(addr)) {     // ... now grey
3510       // push on work queue (grey set)
3511       bool simulate_overflow = false;
3512       NOT_PRODUCT(
3513         if (CMSMarkStackOverflowALot &&
3514             _collector->simulate_overflow()) {
3515           // simulate a stack overflow
3516           simulate_overflow = true;
3517         }
3518       )
3519       if (simulate_overflow ||
3520           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3521         // stack overflow
3522         if (PrintCMSStatistics != 0) {
3523           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3524                                  SIZE_FORMAT, _overflow_stack->capacity());
3525         }
3526         // We cannot assert that the overflow stack is full because
3527         // it may have been emptied since.
3528         assert(simulate_overflow ||
3529                _work_queue->size() == _work_queue->max_elems(),
3530               "Else push should have succeeded");
3531         handle_stack_overflow(addr);
3532       }
3533     } // Else, some other thread got there first
3534     do_yield_check();
3535   }
3536 }
3537 
3538 void Par_ConcMarkingClosure::do_oop(oop* p)       { Par_ConcMarkingClosure::do_oop_work(p); }
3539 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3540 
3541 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3542   while (_work_queue->size() > max) {
3543     oop new_oop;
3544     if (_work_queue->pop_local(new_oop)) {
3545       assert(new_oop->is_oop(), "Should be an oop");
3546       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3547       assert(_span.contains((HeapWord*)new_oop), "Not in span");
3548       new_oop->oop_iterate(this);  // do_oop() above
3549       do_yield_check();
3550     }
3551   }
3552 }
3553 
3554 // Upon stack overflow, we discard (part of) the stack,
3555 // remembering the least address amongst those discarded
3556 // in CMSCollector's _restart_address.
3557 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3558   // We need to do this under a mutex to prevent other
3559   // workers from interfering with the work done below.
3560   MutexLockerEx ml(_overflow_stack->par_lock(),
3561                    Mutex::_no_safepoint_check_flag);
3562   // Remember the least grey address discarded
3563   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3564   _collector->lower_restart_addr(ra);
3565   _overflow_stack->reset();  // discard stack contents
3566   _overflow_stack->expand(); // expand the stack if possible
3567 }
3568 
3569 
3570 void CMSConcMarkingTask::do_work_steal(int i) {
3571   OopTaskQueue* work_q = work_queue(i);
3572   oop obj_to_scan;
3573   CMSBitMap* bm = &(_collector->_markBitMap);
3574   CMSMarkStack* ovflw = &(_collector->_markStack);
3575   int* seed = _collector->hash_seed(i);
3576   Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3577   while (true) {
3578     cl.trim_queue(0);
3579     assert(work_q->size() == 0, "Should have been emptied above");
3580     if (get_work_from_overflow_stack(ovflw, work_q)) {
3581       // Can't assert below because the work obtained from the
3582       // overflow stack may already have been stolen from us.
3583       // assert(work_q->size() > 0, "Work from overflow stack");
3584       continue;
3585     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3586       assert(obj_to_scan->is_oop(), "Should be an oop");
3587       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3588       obj_to_scan->oop_iterate(&cl);
3589     } else if (terminator()->offer_termination(&_term_term)) {
3590       assert(work_q->size() == 0, "Impossible!");
3591       break;
3592     } else if (yielding() || should_yield()) {
3593       yield();
3594     }
3595   }
3596 }
3597 
3598 // This is run by the CMS (coordinator) thread.
3599 void CMSConcMarkingTask::coordinator_yield() {
3600   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3601          "CMS thread should hold CMS token");
3602   // First give up the locks, then yield, then re-lock
3603   // We should probably use a constructor/destructor idiom to
3604   // do this unlock/lock or modify the MutexUnlocker class to
3605   // serve our purpose. XXX
3606   assert_lock_strong(_bit_map_lock);
3607   _bit_map_lock->unlock();
3608   ConcurrentMarkSweepThread::desynchronize(true);
3609   _collector->stopTimer();
3610   if (PrintCMSStatistics != 0) {
3611     _collector->incrementYields();
3612   }
3613 
3614   // It is possible for whichever thread initiated the yield request
3615   // not to get a chance to wake up and take the bitmap lock between
3616   // this thread releasing it and reacquiring it. So, while the
3617   // should_yield() flag is on, let's sleep for a bit to give the
3618   // other thread a chance to wake up. The limit imposed on the number
3619   // of iterations is defensive, to avoid any unforseen circumstances
3620   // putting us into an infinite loop. Since it's always been this
3621   // (coordinator_yield()) method that was observed to cause the
3622   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3623   // which is by default non-zero. For the other seven methods that
3624   // also perform the yield operation, as are using a different
3625   // parameter (CMSYieldSleepCount) which is by default zero. This way we
3626   // can enable the sleeping for those methods too, if necessary.
3627   // See 6442774.
3628   //
3629   // We really need to reconsider the synchronization between the GC
3630   // thread and the yield-requesting threads in the future and we
3631   // should really use wait/notify, which is the recommended
3632   // way of doing this type of interaction. Additionally, we should
3633   // consolidate the eight methods that do the yield operation and they
3634   // are almost identical into one for better maintainability and
3635   // readability. See 6445193.
3636   //
3637   // Tony 2006.06.29
3638   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3639                    ConcurrentMarkSweepThread::should_yield() &&
3640                    !CMSCollector::foregroundGCIsActive(); ++i) {
3641     os::sleep(Thread::current(), 1, false);
3642   }
3643 
3644   ConcurrentMarkSweepThread::synchronize(true);
3645   _bit_map_lock->lock_without_safepoint_check();
3646   _collector->startTimer();
3647 }
3648 
3649 bool CMSCollector::do_marking_mt() {
3650   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3651   uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3652                                                                   conc_workers()->active_workers(),
3653                                                                   Threads::number_of_non_daemon_threads());
3654   conc_workers()->set_active_workers(num_workers);
3655 
3656   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
3657 
3658   CMSConcMarkingTask tsk(this,
3659                          cms_space,
3660                          conc_workers(),
3661                          task_queues());
3662 
3663   // Since the actual number of workers we get may be different
3664   // from the number we requested above, do we need to do anything different
3665   // below? In particular, may be we need to subclass the SequantialSubTasksDone
3666   // class?? XXX
3667   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3668 
3669   // Refs discovery is already non-atomic.
3670   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3671   assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3672   conc_workers()->start_task(&tsk);
3673   while (tsk.yielded()) {
3674     tsk.coordinator_yield();
3675     conc_workers()->continue_task(&tsk);
3676   }
3677   // If the task was aborted, _restart_addr will be non-NULL
3678   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3679   while (_restart_addr != NULL) {
3680     // XXX For now we do not make use of ABORTED state and have not
3681     // yet implemented the right abort semantics (even in the original
3682     // single-threaded CMS case). That needs some more investigation
3683     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3684     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3685     // If _restart_addr is non-NULL, a marking stack overflow
3686     // occurred; we need to do a fresh marking iteration from the
3687     // indicated restart address.
3688     if (_foregroundGCIsActive) {
3689       // We may be running into repeated stack overflows, having
3690       // reached the limit of the stack size, while making very
3691       // slow forward progress. It may be best to bail out and
3692       // let the foreground collector do its job.
3693       // Clear _restart_addr, so that foreground GC
3694       // works from scratch. This avoids the headache of
3695       // a "rescan" which would otherwise be needed because
3696       // of the dirty mod union table & card table.
3697       _restart_addr = NULL;
3698       return false;
3699     }
3700     // Adjust the task to restart from _restart_addr
3701     tsk.reset(_restart_addr);
3702     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3703                   _restart_addr);
3704     _restart_addr = NULL;
3705     // Get the workers going again
3706     conc_workers()->start_task(&tsk);
3707     while (tsk.yielded()) {
3708       tsk.coordinator_yield();
3709       conc_workers()->continue_task(&tsk);
3710     }
3711   }
3712   assert(tsk.completed(), "Inconsistency");
3713   assert(tsk.result() == true, "Inconsistency");
3714   return true;
3715 }
3716 
3717 bool CMSCollector::do_marking_st() {
3718   ResourceMark rm;
3719   HandleMark   hm;
3720 
3721   // Temporarily make refs discovery single threaded (non-MT)
3722   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3723   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3724     &_markStack, CMSYield);
3725   // the last argument to iterate indicates whether the iteration
3726   // should be incremental with periodic yields.
3727   _markBitMap.iterate(&markFromRootsClosure);
3728   // If _restart_addr is non-NULL, a marking stack overflow
3729   // occurred; we need to do a fresh iteration from the
3730   // indicated restart address.
3731   while (_restart_addr != NULL) {
3732     if (_foregroundGCIsActive) {
3733       // We may be running into repeated stack overflows, having
3734       // reached the limit of the stack size, while making very
3735       // slow forward progress. It may be best to bail out and
3736       // let the foreground collector do its job.
3737       // Clear _restart_addr, so that foreground GC
3738       // works from scratch. This avoids the headache of
3739       // a "rescan" which would otherwise be needed because
3740       // of the dirty mod union table & card table.
3741       _restart_addr = NULL;
3742       return false;  // indicating failure to complete marking
3743     }
3744     // Deal with stack overflow:
3745     // we restart marking from _restart_addr
3746     HeapWord* ra = _restart_addr;
3747     markFromRootsClosure.reset(ra);
3748     _restart_addr = NULL;
3749     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3750   }
3751   return true;
3752 }
3753 
3754 void CMSCollector::preclean() {
3755   check_correct_thread_executing();
3756   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3757   verify_work_stacks_empty();
3758   verify_overflow_empty();
3759   _abort_preclean = false;
3760   if (CMSPrecleaningEnabled) {
3761     if (!CMSEdenChunksRecordAlways) {
3762       _eden_chunk_index = 0;
3763     }
3764     size_t used = get_eden_used();
3765     size_t capacity = get_eden_capacity();
3766     // Don't start sampling unless we will get sufficiently
3767     // many samples.
3768     if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3769                 * CMSScheduleRemarkEdenPenetration)) {
3770       _start_sampling = true;
3771     } else {
3772       _start_sampling = false;
3773     }
3774     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3775     CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3776     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3777   }
3778   CMSTokenSync x(true); // is cms thread
3779   if (CMSPrecleaningEnabled) {
3780     sample_eden();
3781     _collectorState = AbortablePreclean;
3782   } else {
3783     _collectorState = FinalMarking;
3784   }
3785   verify_work_stacks_empty();
3786   verify_overflow_empty();
3787 }
3788 
3789 // Try and schedule the remark such that young gen
3790 // occupancy is CMSScheduleRemarkEdenPenetration %.
3791 void CMSCollector::abortable_preclean() {
3792   check_correct_thread_executing();
3793   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
3794   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3795 
3796   // If Eden's current occupancy is below this threshold,
3797   // immediately schedule the remark; else preclean
3798   // past the next scavenge in an effort to
3799   // schedule the pause as described above. By choosing
3800   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3801   // we will never do an actual abortable preclean cycle.
3802   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3803     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3804     CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3805     // We need more smarts in the abortable preclean
3806     // loop below to deal with cases where allocation
3807     // in young gen is very very slow, and our precleaning
3808     // is running a losing race against a horde of
3809     // mutators intent on flooding us with CMS updates
3810     // (dirty cards).
3811     // One, admittedly dumb, strategy is to give up
3812     // after a certain number of abortable precleaning loops
3813     // or after a certain maximum time. We want to make
3814     // this smarter in the next iteration.
3815     // XXX FIX ME!!! YSR
3816     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3817     while (!(should_abort_preclean() ||
3818              ConcurrentMarkSweepThread::should_terminate())) {
3819       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3820       cumworkdone += workdone;
3821       loops++;
3822       // Voluntarily terminate abortable preclean phase if we have
3823       // been at it for too long.
3824       if ((CMSMaxAbortablePrecleanLoops != 0) &&
3825           loops >= CMSMaxAbortablePrecleanLoops) {
3826         if (PrintGCDetails) {
3827           gclog_or_tty->print(" CMS: abort preclean due to loops ");
3828         }
3829         break;
3830       }
3831       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3832         if (PrintGCDetails) {
3833           gclog_or_tty->print(" CMS: abort preclean due to time ");
3834         }
3835         break;
3836       }
3837       // If we are doing little work each iteration, we should
3838       // take a short break.
3839       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3840         // Sleep for some time, waiting for work to accumulate
3841         stopTimer();
3842         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3843         startTimer();
3844         waited++;
3845       }
3846     }
3847     if (PrintCMSStatistics > 0) {
3848       gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3849                           loops, waited, cumworkdone);
3850     }
3851   }
3852   CMSTokenSync x(true); // is cms thread
3853   if (_collectorState != Idling) {
3854     assert(_collectorState == AbortablePreclean,
3855            "Spontaneous state transition?");
3856     _collectorState = FinalMarking;
3857   } // Else, a foreground collection completed this CMS cycle.
3858   return;
3859 }
3860 
3861 // Respond to an Eden sampling opportunity
3862 void CMSCollector::sample_eden() {
3863   // Make sure a young gc cannot sneak in between our
3864   // reading and recording of a sample.
3865   assert(Thread::current()->is_ConcurrentGC_thread(),
3866          "Only the cms thread may collect Eden samples");
3867   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3868          "Should collect samples while holding CMS token");
3869   if (!_start_sampling) {
3870     return;
3871   }
3872   // When CMSEdenChunksRecordAlways is true, the eden chunk array
3873   // is populated by the young generation.
3874   if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3875     if (_eden_chunk_index < _eden_chunk_capacity) {
3876       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
3877       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3878              "Unexpected state of Eden");
3879       // We'd like to check that what we just sampled is an oop-start address;
3880       // however, we cannot do that here since the object may not yet have been
3881       // initialized. So we'll instead do the check when we _use_ this sample
3882       // later.
3883       if (_eden_chunk_index == 0 ||
3884           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3885                          _eden_chunk_array[_eden_chunk_index-1])
3886            >= CMSSamplingGrain)) {
3887         _eden_chunk_index++;  // commit sample
3888       }
3889     }
3890   }
3891   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3892     size_t used = get_eden_used();
3893     size_t capacity = get_eden_capacity();
3894     assert(used <= capacity, "Unexpected state of Eden");
3895     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3896       _abort_preclean = true;
3897     }
3898   }
3899 }
3900 
3901 
3902 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3903   assert(_collectorState == Precleaning ||
3904          _collectorState == AbortablePreclean, "incorrect state");
3905   ResourceMark rm;
3906   HandleMark   hm;
3907 
3908   // Precleaning is currently not MT but the reference processor
3909   // may be set for MT.  Disable it temporarily here.
3910   ReferenceProcessor* rp = ref_processor();
3911   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3912 
3913   // Do one pass of scrubbing the discovered reference lists
3914   // to remove any reference objects with strongly-reachable
3915   // referents.
3916   if (clean_refs) {
3917     CMSPrecleanRefsYieldClosure yield_cl(this);
3918     assert(rp->span().equals(_span), "Spans should be equal");
3919     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3920                                    &_markStack, true /* preclean */);
3921     CMSDrainMarkingStackClosure complete_trace(this,
3922                                    _span, &_markBitMap, &_markStack,
3923                                    &keep_alive, true /* preclean */);
3924 
3925     // We don't want this step to interfere with a young
3926     // collection because we don't want to take CPU
3927     // or memory bandwidth away from the young GC threads
3928     // (which may be as many as there are CPUs).
3929     // Note that we don't need to protect ourselves from
3930     // interference with mutators because they can't
3931     // manipulate the discovered reference lists nor affect
3932     // the computed reachability of the referents, the
3933     // only properties manipulated by the precleaning
3934     // of these reference lists.
3935     stopTimer();
3936     CMSTokenSyncWithLocks x(true /* is cms thread */,
3937                             bitMapLock());
3938     startTimer();
3939     sample_eden();
3940 
3941     // The following will yield to allow foreground
3942     // collection to proceed promptly. XXX YSR:
3943     // The code in this method may need further
3944     // tweaking for better performance and some restructuring
3945     // for cleaner interfaces.
3946     GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3947     rp->preclean_discovered_references(
3948           rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3949           gc_timer, _gc_tracer_cm->gc_id());
3950   }
3951 
3952   if (clean_survivor) {  // preclean the active survivor space(s)
3953     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3954                              &_markBitMap, &_modUnionTable,
3955                              &_markStack, true /* precleaning phase */);
3956     stopTimer();
3957     CMSTokenSyncWithLocks ts(true /* is cms thread */,
3958                              bitMapLock());
3959     startTimer();
3960     unsigned int before_count =
3961       GenCollectedHeap::heap()->total_collections();
3962     SurvivorSpacePrecleanClosure
3963       sss_cl(this, _span, &_markBitMap, &_markStack,
3964              &pam_cl, before_count, CMSYield);
3965     _young_gen->from()->object_iterate_careful(&sss_cl);
3966     _young_gen->to()->object_iterate_careful(&sss_cl);
3967   }
3968   MarkRefsIntoAndScanClosure
3969     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3970              &_markStack, this, CMSYield,
3971              true /* precleaning phase */);
3972   // CAUTION: The following closure has persistent state that may need to
3973   // be reset upon a decrease in the sequence of addresses it
3974   // processes.
3975   ScanMarkedObjectsAgainCarefullyClosure
3976     smoac_cl(this, _span,
3977       &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3978 
3979   // Preclean dirty cards in ModUnionTable and CardTable using
3980   // appropriate convergence criterion;
3981   // repeat CMSPrecleanIter times unless we find that
3982   // we are losing.
3983   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3984   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3985          "Bad convergence multiplier");
3986   assert(CMSPrecleanThreshold >= 100,
3987          "Unreasonably low CMSPrecleanThreshold");
3988 
3989   size_t numIter, cumNumCards, lastNumCards, curNumCards;
3990   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3991        numIter < CMSPrecleanIter;
3992        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3993     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
3994     if (Verbose && PrintGCDetails) {
3995       gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3996     }
3997     // Either there are very few dirty cards, so re-mark
3998     // pause will be small anyway, or our pre-cleaning isn't
3999     // that much faster than the rate at which cards are being
4000     // dirtied, so we might as well stop and re-mark since
4001     // precleaning won't improve our re-mark time by much.
4002     if (curNumCards <= CMSPrecleanThreshold ||
4003         (numIter > 0 &&
4004          (curNumCards * CMSPrecleanDenominator >
4005          lastNumCards * CMSPrecleanNumerator))) {
4006       numIter++;
4007       cumNumCards += curNumCards;
4008       break;
4009     }
4010   }
4011 
4012   preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4013 
4014   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4015   cumNumCards += curNumCards;
4016   if (PrintGCDetails && PrintCMSStatistics != 0) {
4017     gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4018                   curNumCards, cumNumCards, numIter);
4019   }
4020   return cumNumCards;   // as a measure of useful work done
4021 }
4022 
4023 // PRECLEANING NOTES:
4024 // Precleaning involves:
4025 // . reading the bits of the modUnionTable and clearing the set bits.
4026 // . For the cards corresponding to the set bits, we scan the
4027 //   objects on those cards. This means we need the free_list_lock
4028 //   so that we can safely iterate over the CMS space when scanning
4029 //   for oops.
4030 // . When we scan the objects, we'll be both reading and setting
4031 //   marks in the marking bit map, so we'll need the marking bit map.
4032 // . For protecting _collector_state transitions, we take the CGC_lock.
4033 //   Note that any races in the reading of of card table entries by the
4034 //   CMS thread on the one hand and the clearing of those entries by the
4035 //   VM thread or the setting of those entries by the mutator threads on the
4036 //   other are quite benign. However, for efficiency it makes sense to keep
4037 //   the VM thread from racing with the CMS thread while the latter is
4038 //   dirty card info to the modUnionTable. We therefore also use the
4039 //   CGC_lock to protect the reading of the card table and the mod union
4040 //   table by the CM thread.
4041 // . We run concurrently with mutator updates, so scanning
4042 //   needs to be done carefully  -- we should not try to scan
4043 //   potentially uninitialized objects.
4044 //
4045 // Locking strategy: While holding the CGC_lock, we scan over and
4046 // reset a maximal dirty range of the mod union / card tables, then lock
4047 // the free_list_lock and bitmap lock to do a full marking, then
4048 // release these locks; and repeat the cycle. This allows for a
4049 // certain amount of fairness in the sharing of these locks between
4050 // the CMS collector on the one hand, and the VM thread and the
4051 // mutators on the other.
4052 
4053 // NOTE: preclean_mod_union_table() and preclean_card_table()
4054 // further below are largely identical; if you need to modify
4055 // one of these methods, please check the other method too.
4056 
4057 size_t CMSCollector::preclean_mod_union_table(
4058   ConcurrentMarkSweepGeneration* gen,
4059   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4060   verify_work_stacks_empty();
4061   verify_overflow_empty();
4062 
4063   // strategy: starting with the first card, accumulate contiguous
4064   // ranges of dirty cards; clear these cards, then scan the region
4065   // covered by these cards.
4066 
4067   // Since all of the MUT is committed ahead, we can just use
4068   // that, in case the generations expand while we are precleaning.
4069   // It might also be fine to just use the committed part of the
4070   // generation, but we might potentially miss cards when the
4071   // generation is rapidly expanding while we are in the midst
4072   // of precleaning.
4073   HeapWord* startAddr = gen->reserved().start();
4074   HeapWord* endAddr   = gen->reserved().end();
4075 
4076   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
4077 
4078   size_t numDirtyCards, cumNumDirtyCards;
4079   HeapWord *nextAddr, *lastAddr;
4080   for (cumNumDirtyCards = numDirtyCards = 0,
4081        nextAddr = lastAddr = startAddr;
4082        nextAddr < endAddr;
4083        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4084 
4085     ResourceMark rm;
4086     HandleMark   hm;
4087 
4088     MemRegion dirtyRegion;
4089     {
4090       stopTimer();
4091       // Potential yield point
4092       CMSTokenSync ts(true);
4093       startTimer();
4094       sample_eden();
4095       // Get dirty region starting at nextOffset (inclusive),
4096       // simultaneously clearing it.
4097       dirtyRegion =
4098         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4099       assert(dirtyRegion.start() >= nextAddr,
4100              "returned region inconsistent?");
4101     }
4102     // Remember where the next search should begin.
4103     // The returned region (if non-empty) is a right open interval,
4104     // so lastOffset is obtained from the right end of that
4105     // interval.
4106     lastAddr = dirtyRegion.end();
4107     // Should do something more transparent and less hacky XXX
4108     numDirtyCards =
4109       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4110 
4111     // We'll scan the cards in the dirty region (with periodic
4112     // yields for foreground GC as needed).
4113     if (!dirtyRegion.is_empty()) {
4114       assert(numDirtyCards > 0, "consistency check");
4115       HeapWord* stop_point = NULL;
4116       stopTimer();
4117       // Potential yield point
4118       CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4119                                bitMapLock());
4120       startTimer();
4121       {
4122         verify_work_stacks_empty();
4123         verify_overflow_empty();
4124         sample_eden();
4125         stop_point =
4126           gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4127       }
4128       if (stop_point != NULL) {
4129         // The careful iteration stopped early either because it found an
4130         // uninitialized object, or because we were in the midst of an
4131         // "abortable preclean", which should now be aborted. Redirty
4132         // the bits corresponding to the partially-scanned or unscanned
4133         // cards. We'll either restart at the next block boundary or
4134         // abort the preclean.
4135         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4136                "Should only be AbortablePreclean.");
4137         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4138         if (should_abort_preclean()) {
4139           break; // out of preclean loop
4140         } else {
4141           // Compute the next address at which preclean should pick up;
4142           // might need bitMapLock in order to read P-bits.
4143           lastAddr = next_card_start_after_block(stop_point);
4144         }
4145       }
4146     } else {
4147       assert(lastAddr == endAddr, "consistency check");
4148       assert(numDirtyCards == 0, "consistency check");
4149       break;
4150     }
4151   }
4152   verify_work_stacks_empty();
4153   verify_overflow_empty();
4154   return cumNumDirtyCards;
4155 }
4156 
4157 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4158 // below are largely identical; if you need to modify
4159 // one of these methods, please check the other method too.
4160 
4161 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4162   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4163   // strategy: it's similar to precleamModUnionTable above, in that
4164   // we accumulate contiguous ranges of dirty cards, mark these cards
4165   // precleaned, then scan the region covered by these cards.
4166   HeapWord* endAddr   = (HeapWord*)(gen->_virtual_space.high());
4167   HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4168 
4169   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
4170 
4171   size_t numDirtyCards, cumNumDirtyCards;
4172   HeapWord *lastAddr, *nextAddr;
4173 
4174   for (cumNumDirtyCards = numDirtyCards = 0,
4175        nextAddr = lastAddr = startAddr;
4176        nextAddr < endAddr;
4177        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4178 
4179     ResourceMark rm;
4180     HandleMark   hm;
4181 
4182     MemRegion dirtyRegion;
4183     {
4184       // See comments in "Precleaning notes" above on why we
4185       // do this locking. XXX Could the locking overheads be
4186       // too high when dirty cards are sparse? [I don't think so.]
4187       stopTimer();
4188       CMSTokenSync x(true); // is cms thread
4189       startTimer();
4190       sample_eden();
4191       // Get and clear dirty region from card table
4192       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4193                                     MemRegion(nextAddr, endAddr),
4194                                     true,
4195                                     CardTableModRefBS::precleaned_card_val());
4196 
4197       assert(dirtyRegion.start() >= nextAddr,
4198              "returned region inconsistent?");
4199     }
4200     lastAddr = dirtyRegion.end();
4201     numDirtyCards =
4202       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4203 
4204     if (!dirtyRegion.is_empty()) {
4205       stopTimer();
4206       CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4207       startTimer();
4208       sample_eden();
4209       verify_work_stacks_empty();
4210       verify_overflow_empty();
4211       HeapWord* stop_point =
4212         gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4213       if (stop_point != NULL) {
4214         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4215                "Should only be AbortablePreclean.");
4216         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4217         if (should_abort_preclean()) {
4218           break; // out of preclean loop
4219         } else {
4220           // Compute the next address at which preclean should pick up.
4221           lastAddr = next_card_start_after_block(stop_point);
4222         }
4223       }
4224     } else {
4225       break;
4226     }
4227   }
4228   verify_work_stacks_empty();
4229   verify_overflow_empty();
4230   return cumNumDirtyCards;
4231 }
4232 
4233 class PrecleanKlassClosure : public KlassClosure {
4234   KlassToOopClosure _cm_klass_closure;
4235  public:
4236   PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4237   void do_klass(Klass* k) {
4238     if (k->has_accumulated_modified_oops()) {
4239       k->clear_accumulated_modified_oops();
4240 
4241       _cm_klass_closure.do_klass(k);
4242     }
4243   }
4244 };
4245 
4246 // The freelist lock is needed to prevent asserts, is it really needed?
4247 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4248 
4249   cl->set_freelistLock(freelistLock);
4250 
4251   CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4252 
4253   // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4254   // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4255   PrecleanKlassClosure preclean_klass_closure(cl);
4256   ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4257 
4258   verify_work_stacks_empty();
4259   verify_overflow_empty();
4260 }
4261 
4262 void CMSCollector::checkpointRootsFinal() {
4263   assert(_collectorState == FinalMarking, "incorrect state transition?");
4264   check_correct_thread_executing();
4265   // world is stopped at this checkpoint
4266   assert(SafepointSynchronize::is_at_safepoint(),
4267          "world should be stopped");
4268   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4269 
4270   verify_work_stacks_empty();
4271   verify_overflow_empty();
4272 
4273   if (PrintGCDetails) {
4274     gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]",
4275                         _young_gen->used() / K,
4276                         _young_gen->capacity() / K);
4277   }
4278   {
4279     if (CMSScavengeBeforeRemark) {
4280       GenCollectedHeap* gch = GenCollectedHeap::heap();
4281       // Temporarily set flag to false, GCH->do_collection will
4282       // expect it to be false and set to true
4283       FlagSetting fl(gch->_is_gc_active, false);
4284       NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4285         PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4286       gch->do_collection(true,                      // full (i.e. force, see below)
4287                          false,                     // !clear_all_soft_refs
4288                          0,                         // size
4289                          false,                     // is_tlab
4290                          GenCollectedHeap::YoungGen // type
4291         );
4292     }
4293     FreelistLocker x(this);
4294     MutexLockerEx y(bitMapLock(),
4295                     Mutex::_no_safepoint_check_flag);
4296     checkpointRootsFinalWork();
4297   }
4298   verify_work_stacks_empty();
4299   verify_overflow_empty();
4300 }
4301 
4302 void CMSCollector::checkpointRootsFinalWork() {
4303   NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4304 
4305   assert(haveFreelistLocks(), "must have free list locks");
4306   assert_lock_strong(bitMapLock());
4307 
4308   ResourceMark rm;
4309   HandleMark   hm;
4310 
4311   GenCollectedHeap* gch = GenCollectedHeap::heap();
4312 
4313   if (should_unload_classes()) {
4314     CodeCache::gc_prologue();
4315   }
4316   assert(haveFreelistLocks(), "must have free list locks");
4317   assert_lock_strong(bitMapLock());
4318 
4319   // We might assume that we need not fill TLAB's when
4320   // CMSScavengeBeforeRemark is set, because we may have just done
4321   // a scavenge which would have filled all TLAB's -- and besides
4322   // Eden would be empty. This however may not always be the case --
4323   // for instance although we asked for a scavenge, it may not have
4324   // happened because of a JNI critical section. We probably need
4325   // a policy for deciding whether we can in that case wait until
4326   // the critical section releases and then do the remark following
4327   // the scavenge, and skip it here. In the absence of that policy,
4328   // or of an indication of whether the scavenge did indeed occur,
4329   // we cannot rely on TLAB's having been filled and must do
4330   // so here just in case a scavenge did not happen.
4331   gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
4332   // Update the saved marks which may affect the root scans.
4333   gch->save_marks();
4334 
4335   if (CMSPrintEdenSurvivorChunks) {
4336     print_eden_and_survivor_chunk_arrays();
4337   }
4338 
4339   {
4340     COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4341 
4342     // Note on the role of the mod union table:
4343     // Since the marker in "markFromRoots" marks concurrently with
4344     // mutators, it is possible for some reachable objects not to have been
4345     // scanned. For instance, an only reference to an object A was
4346     // placed in object B after the marker scanned B. Unless B is rescanned,
4347     // A would be collected. Such updates to references in marked objects
4348     // are detected via the mod union table which is the set of all cards
4349     // dirtied since the first checkpoint in this GC cycle and prior to
4350     // the most recent young generation GC, minus those cleaned up by the
4351     // concurrent precleaning.
4352     if (CMSParallelRemarkEnabled) {
4353       GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4354       do_remark_parallel();
4355     } else {
4356       GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4357                   _gc_timer_cm, _gc_tracer_cm->gc_id());
4358       do_remark_non_parallel();
4359     }
4360   }
4361   verify_work_stacks_empty();
4362   verify_overflow_empty();
4363 
4364   {
4365     NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4366     refProcessingWork();
4367   }
4368   verify_work_stacks_empty();
4369   verify_overflow_empty();
4370 
4371   if (should_unload_classes()) {
4372     CodeCache::gc_epilogue();
4373   }
4374   JvmtiExport::gc_epilogue();
4375 
4376   // If we encountered any (marking stack / work queue) overflow
4377   // events during the current CMS cycle, take appropriate
4378   // remedial measures, where possible, so as to try and avoid
4379   // recurrence of that condition.
4380   assert(_markStack.isEmpty(), "No grey objects");
4381   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4382                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
4383   if (ser_ovflw > 0) {
4384     if (PrintCMSStatistics != 0) {
4385       gclog_or_tty->print_cr("Marking stack overflow (benign) "
4386         "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT
4387         ", kac_preclean=" SIZE_FORMAT ")",
4388         _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4389         _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4390     }
4391     _markStack.expand();
4392     _ser_pmc_remark_ovflw = 0;
4393     _ser_pmc_preclean_ovflw = 0;
4394     _ser_kac_preclean_ovflw = 0;
4395     _ser_kac_ovflw = 0;
4396   }
4397   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4398     if (PrintCMSStatistics != 0) {
4399       gclog_or_tty->print_cr("Work queue overflow (benign) "
4400         "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4401         _par_pmc_remark_ovflw, _par_kac_ovflw);
4402     }
4403     _par_pmc_remark_ovflw = 0;
4404     _par_kac_ovflw = 0;
4405   }
4406   if (PrintCMSStatistics != 0) {
4407      if (_markStack._hit_limit > 0) {
4408        gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4409                               _markStack._hit_limit);
4410      }
4411      if (_markStack._failed_double > 0) {
4412        gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT "),"
4413                               " current capacity " SIZE_FORMAT,
4414                               _markStack._failed_double,
4415                               _markStack.capacity());
4416      }
4417   }
4418   _markStack._hit_limit = 0;
4419   _markStack._failed_double = 0;
4420 
4421   if ((VerifyAfterGC || VerifyDuringGC) &&
4422       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4423     verify_after_remark();
4424   }
4425 
4426   _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4427 
4428   // Change under the freelistLocks.
4429   _collectorState = Sweeping;
4430   // Call isAllClear() under bitMapLock
4431   assert(_modUnionTable.isAllClear(),
4432       "Should be clear by end of the final marking");
4433   assert(_ct->klass_rem_set()->mod_union_is_clear(),
4434       "Should be clear by end of the final marking");
4435 }
4436 
4437 void CMSParInitialMarkTask::work(uint worker_id) {
4438   elapsedTimer _timer;
4439   ResourceMark rm;
4440   HandleMark   hm;
4441 
4442   // ---------- scan from roots --------------
4443   _timer.start();
4444   GenCollectedHeap* gch = GenCollectedHeap::heap();
4445   Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4446 
4447   // ---------- young gen roots --------------
4448   {
4449     work_on_young_gen_roots(worker_id, &par_mri_cl);
4450     _timer.stop();
4451     if (PrintCMSStatistics != 0) {
4452       gclog_or_tty->print_cr(
4453         "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4454         worker_id, _timer.seconds());
4455     }
4456   }
4457 
4458   // ---------- remaining roots --------------
4459   _timer.reset();
4460   _timer.start();
4461 
4462   CLDToOopClosure cld_closure(&par_mri_cl, true);
4463 
4464   gch->gen_process_roots(_strong_roots_scope,
4465                          GenCollectedHeap::OldGen,
4466                          false,     // yg was scanned above
4467                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4468                          _collector->should_unload_classes(),
4469                          &par_mri_cl,
4470                          NULL,
4471                          &cld_closure);
4472   assert(_collector->should_unload_classes()
4473          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4474          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4475   _timer.stop();
4476   if (PrintCMSStatistics != 0) {
4477     gclog_or_tty->print_cr(
4478       "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4479       worker_id, _timer.seconds());
4480   }
4481 }
4482 
4483 // Parallel remark task
4484 class CMSParRemarkTask: public CMSParMarkTask {
4485   CompactibleFreeListSpace* _cms_space;
4486 
4487   // The per-thread work queues, available here for stealing.
4488   OopTaskQueueSet*       _task_queues;
4489   ParallelTaskTerminator _term;
4490   StrongRootsScope*      _strong_roots_scope;
4491 
4492  public:
4493   // A value of 0 passed to n_workers will cause the number of
4494   // workers to be taken from the active workers in the work gang.
4495   CMSParRemarkTask(CMSCollector* collector,
4496                    CompactibleFreeListSpace* cms_space,
4497                    uint n_workers, FlexibleWorkGang* workers,
4498                    OopTaskQueueSet* task_queues,
4499                    StrongRootsScope* strong_roots_scope):
4500     CMSParMarkTask("Rescan roots and grey objects in parallel",
4501                    collector, n_workers),
4502     _cms_space(cms_space),
4503     _task_queues(task_queues),
4504     _term(n_workers, task_queues),
4505     _strong_roots_scope(strong_roots_scope) { }
4506 
4507   OopTaskQueueSet* task_queues() { return _task_queues; }
4508 
4509   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4510 
4511   ParallelTaskTerminator* terminator() { return &_term; }
4512   uint n_workers() { return _n_workers; }
4513 
4514   void work(uint worker_id);
4515 
4516  private:
4517   // ... of  dirty cards in old space
4518   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4519                                   Par_MarkRefsIntoAndScanClosure* cl);
4520 
4521   // ... work stealing for the above
4522   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4523 };
4524 
4525 class RemarkKlassClosure : public KlassClosure {
4526   KlassToOopClosure _cm_klass_closure;
4527  public:
4528   RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4529   void do_klass(Klass* k) {
4530     // Check if we have modified any oops in the Klass during the concurrent marking.
4531     if (k->has_accumulated_modified_oops()) {
4532       k->clear_accumulated_modified_oops();
4533 
4534       // We could have transfered the current modified marks to the accumulated marks,
4535       // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4536     } else if (k->has_modified_oops()) {
4537       // Don't clear anything, this info is needed by the next young collection.
4538     } else {
4539       // No modified oops in the Klass.
4540       return;
4541     }
4542 
4543     // The klass has modified fields, need to scan the klass.
4544     _cm_klass_closure.do_klass(k);
4545   }
4546 };
4547 
4548 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4549   ParNewGeneration* young_gen = _collector->_young_gen;
4550   ContiguousSpace* eden_space = young_gen->eden();
4551   ContiguousSpace* from_space = young_gen->from();
4552   ContiguousSpace* to_space   = young_gen->to();
4553 
4554   HeapWord** eca = _collector->_eden_chunk_array;
4555   size_t     ect = _collector->_eden_chunk_index;
4556   HeapWord** sca = _collector->_survivor_chunk_array;
4557   size_t     sct = _collector->_survivor_chunk_index;
4558 
4559   assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4560   assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4561 
4562   do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4563   do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4564   do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4565 }
4566 
4567 // work_queue(i) is passed to the closure
4568 // Par_MarkRefsIntoAndScanClosure.  The "i" parameter
4569 // also is passed to do_dirty_card_rescan_tasks() and to
4570 // do_work_steal() to select the i-th task_queue.
4571 
4572 void CMSParRemarkTask::work(uint worker_id) {
4573   elapsedTimer _timer;
4574   ResourceMark rm;
4575   HandleMark   hm;
4576 
4577   // ---------- rescan from roots --------------
4578   _timer.start();
4579   GenCollectedHeap* gch = GenCollectedHeap::heap();
4580   Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4581     _collector->_span, _collector->ref_processor(),
4582     &(_collector->_markBitMap),
4583     work_queue(worker_id));
4584 
4585   // Rescan young gen roots first since these are likely
4586   // coarsely partitioned and may, on that account, constitute
4587   // the critical path; thus, it's best to start off that
4588   // work first.
4589   // ---------- young gen roots --------------
4590   {
4591     work_on_young_gen_roots(worker_id, &par_mrias_cl);
4592     _timer.stop();
4593     if (PrintCMSStatistics != 0) {
4594       gclog_or_tty->print_cr(
4595         "Finished young gen rescan work in %dth thread: %3.3f sec",
4596         worker_id, _timer.seconds());
4597     }
4598   }
4599 
4600   // ---------- remaining roots --------------
4601   _timer.reset();
4602   _timer.start();
4603   gch->gen_process_roots(_strong_roots_scope,
4604                          GenCollectedHeap::OldGen,
4605                          false,     // yg was scanned above
4606                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4607                          _collector->should_unload_classes(),
4608                          &par_mrias_cl,
4609                          NULL,
4610                          NULL);     // The dirty klasses will be handled below
4611 
4612   assert(_collector->should_unload_classes()
4613          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4614          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4615   _timer.stop();
4616   if (PrintCMSStatistics != 0) {
4617     gclog_or_tty->print_cr(
4618       "Finished remaining root rescan work in %dth thread: %3.3f sec",
4619       worker_id, _timer.seconds());
4620   }
4621 
4622   // ---------- unhandled CLD scanning ----------
4623   if (worker_id == 0) { // Single threaded at the moment.
4624     _timer.reset();
4625     _timer.start();
4626 
4627     // Scan all new class loader data objects and new dependencies that were
4628     // introduced during concurrent marking.
4629     ResourceMark rm;
4630     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4631     for (int i = 0; i < array->length(); i++) {
4632       par_mrias_cl.do_class_loader_data(array->at(i));
4633     }
4634 
4635     // We don't need to keep track of new CLDs anymore.
4636     ClassLoaderDataGraph::remember_new_clds(false);
4637 
4638     _timer.stop();
4639     if (PrintCMSStatistics != 0) {
4640       gclog_or_tty->print_cr(
4641           "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4642           worker_id, _timer.seconds());
4643     }
4644   }
4645 
4646   // ---------- dirty klass scanning ----------
4647   if (worker_id == 0) { // Single threaded at the moment.
4648     _timer.reset();
4649     _timer.start();
4650 
4651     // Scan all classes that was dirtied during the concurrent marking phase.
4652     RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4653     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4654 
4655     _timer.stop();
4656     if (PrintCMSStatistics != 0) {
4657       gclog_or_tty->print_cr(
4658           "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4659           worker_id, _timer.seconds());
4660     }
4661   }
4662 
4663   // We might have added oops to ClassLoaderData::_handles during the
4664   // concurrent marking phase. These oops point to newly allocated objects
4665   // that are guaranteed to be kept alive. Either by the direct allocation
4666   // code, or when the young collector processes the roots. Hence,
4667   // we don't have to revisit the _handles block during the remark phase.
4668 
4669   // ---------- rescan dirty cards ------------
4670   _timer.reset();
4671   _timer.start();
4672 
4673   // Do the rescan tasks for each of the two spaces
4674   // (cms_space) in turn.
4675   // "worker_id" is passed to select the task_queue for "worker_id"
4676   do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4677   _timer.stop();
4678   if (PrintCMSStatistics != 0) {
4679     gclog_or_tty->print_cr(
4680       "Finished dirty card rescan work in %dth thread: %3.3f sec",
4681       worker_id, _timer.seconds());
4682   }
4683 
4684   // ---------- steal work from other threads ...
4685   // ---------- ... and drain overflow list.
4686   _timer.reset();
4687   _timer.start();
4688   do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4689   _timer.stop();
4690   if (PrintCMSStatistics != 0) {
4691     gclog_or_tty->print_cr(
4692       "Finished work stealing in %dth thread: %3.3f sec",
4693       worker_id, _timer.seconds());
4694   }
4695 }
4696 
4697 // Note that parameter "i" is not used.
4698 void
4699 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4700   OopsInGenClosure* cl, ContiguousSpace* space,
4701   HeapWord** chunk_array, size_t chunk_top) {
4702   // Until all tasks completed:
4703   // . claim an unclaimed task
4704   // . compute region boundaries corresponding to task claimed
4705   //   using chunk_array
4706   // . par_oop_iterate(cl) over that region
4707 
4708   ResourceMark rm;
4709   HandleMark   hm;
4710 
4711   SequentialSubTasksDone* pst = space->par_seq_tasks();
4712 
4713   uint nth_task = 0;
4714   uint n_tasks  = pst->n_tasks();
4715 
4716   if (n_tasks > 0) {
4717     assert(pst->valid(), "Uninitialized use?");
4718     HeapWord *start, *end;
4719     while (!pst->is_task_claimed(/* reference */ nth_task)) {
4720       // We claimed task # nth_task; compute its boundaries.
4721       if (chunk_top == 0) {  // no samples were taken
4722         assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4723         start = space->bottom();
4724         end   = space->top();
4725       } else if (nth_task == 0) {
4726         start = space->bottom();
4727         end   = chunk_array[nth_task];
4728       } else if (nth_task < (uint)chunk_top) {
4729         assert(nth_task >= 1, "Control point invariant");
4730         start = chunk_array[nth_task - 1];
4731         end   = chunk_array[nth_task];
4732       } else {
4733         assert(nth_task == (uint)chunk_top, "Control point invariant");
4734         start = chunk_array[chunk_top - 1];
4735         end   = space->top();
4736       }
4737       MemRegion mr(start, end);
4738       // Verify that mr is in space
4739       assert(mr.is_empty() || space->used_region().contains(mr),
4740              "Should be in space");
4741       // Verify that "start" is an object boundary
4742       assert(mr.is_empty() || oop(mr.start())->is_oop(),
4743              "Should be an oop");
4744       // Inspect that there is good distribution of rescan tasks
4745       #ifdef ASSERT
4746         if (PrintCMSStatistics != 0) {
4747           tty->print_cr("rescan task=" UINT32_FORMAT "/" UINT32_FORMAT ", " PTR_FORMAT "-" PTR_FORMAT ", size=" SIZE_FORMAT".",
4748                         nth_task, n_tasks, p2i(mr.start()), p2i(mr.end()), mr.word_size());
4749         }
4750       #endif
4751       space->par_oop_iterate(mr, cl);
4752     }
4753     pst->all_tasks_completed();
4754   }
4755 }
4756 
4757 void
4758 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4759   CompactibleFreeListSpace* sp, int i,
4760   Par_MarkRefsIntoAndScanClosure* cl) {
4761   // Until all tasks completed:
4762   // . claim an unclaimed task
4763   // . compute region boundaries corresponding to task claimed
4764   // . transfer dirty bits ct->mut for that region
4765   // . apply rescanclosure to dirty mut bits for that region
4766 
4767   ResourceMark rm;
4768   HandleMark   hm;
4769 
4770   OopTaskQueue* work_q = work_queue(i);
4771   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4772   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4773   // CAUTION: This closure has state that persists across calls to
4774   // the work method dirty_range_iterate_clear() in that it has
4775   // embedded in it a (subtype of) UpwardsObjectClosure. The
4776   // use of that state in the embedded UpwardsObjectClosure instance
4777   // assumes that the cards are always iterated (even if in parallel
4778   // by several threads) in monotonically increasing order per each
4779   // thread. This is true of the implementation below which picks
4780   // card ranges (chunks) in monotonically increasing order globally
4781   // and, a-fortiori, in monotonically increasing order per thread
4782   // (the latter order being a subsequence of the former).
4783   // If the work code below is ever reorganized into a more chaotic
4784   // work-partitioning form than the current "sequential tasks"
4785   // paradigm, the use of that persistent state will have to be
4786   // revisited and modified appropriately. See also related
4787   // bug 4756801 work on which should examine this code to make
4788   // sure that the changes there do not run counter to the
4789   // assumptions made here and necessary for correctness and
4790   // efficiency. Note also that this code might yield inefficient
4791   // behavior in the case of very large objects that span one or
4792   // more work chunks. Such objects would potentially be scanned
4793   // several times redundantly. Work on 4756801 should try and
4794   // address that performance anomaly if at all possible. XXX
4795   MemRegion  full_span  = _collector->_span;
4796   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
4797   MarkFromDirtyCardsClosure
4798     greyRescanClosure(_collector, full_span, // entire span of interest
4799                       sp, bm, work_q, cl);
4800 
4801   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4802   assert(pst->valid(), "Uninitialized use?");
4803   uint nth_task = 0;
4804   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4805   MemRegion span = sp->used_region();
4806   HeapWord* start_addr = span.start();
4807   HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4808                                            alignment);
4809   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4810   assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4811          start_addr, "Check alignment");
4812   assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4813          chunk_size, "Check alignment");
4814 
4815   while (!pst->is_task_claimed(/* reference */ nth_task)) {
4816     // Having claimed the nth_task, compute corresponding mem-region,
4817     // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4818     // The alignment restriction ensures that we do not need any
4819     // synchronization with other gang-workers while setting or
4820     // clearing bits in thus chunk of the MUT.
4821     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4822                                     start_addr + (nth_task+1)*chunk_size);
4823     // The last chunk's end might be way beyond end of the
4824     // used region. In that case pull back appropriately.
4825     if (this_span.end() > end_addr) {
4826       this_span.set_end(end_addr);
4827       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4828     }
4829     // Iterate over the dirty cards covering this chunk, marking them
4830     // precleaned, and setting the corresponding bits in the mod union
4831     // table. Since we have been careful to partition at Card and MUT-word
4832     // boundaries no synchronization is needed between parallel threads.
4833     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4834                                                  &modUnionClosure);
4835 
4836     // Having transferred these marks into the modUnionTable,
4837     // rescan the marked objects on the dirty cards in the modUnionTable.
4838     // Even if this is at a synchronous collection, the initial marking
4839     // may have been done during an asynchronous collection so there
4840     // may be dirty bits in the mod-union table.
4841     _collector->_modUnionTable.dirty_range_iterate_clear(
4842                   this_span, &greyRescanClosure);
4843     _collector->_modUnionTable.verifyNoOneBitsInRange(
4844                                  this_span.start(),
4845                                  this_span.end());
4846   }
4847   pst->all_tasks_completed();  // declare that i am done
4848 }
4849 
4850 // . see if we can share work_queues with ParNew? XXX
4851 void
4852 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4853                                 int* seed) {
4854   OopTaskQueue* work_q = work_queue(i);
4855   NOT_PRODUCT(int num_steals = 0;)
4856   oop obj_to_scan;
4857   CMSBitMap* bm = &(_collector->_markBitMap);
4858 
4859   while (true) {
4860     // Completely finish any left over work from (an) earlier round(s)
4861     cl->trim_queue(0);
4862     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4863                                          (size_t)ParGCDesiredObjsFromOverflowList);
4864     // Now check if there's any work in the overflow list
4865     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4866     // only affects the number of attempts made to get work from the
4867     // overflow list and does not affect the number of workers.  Just
4868     // pass ParallelGCThreads so this behavior is unchanged.
4869     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4870                                                 work_q,
4871                                                 ParallelGCThreads)) {
4872       // found something in global overflow list;
4873       // not yet ready to go stealing work from others.
4874       // We'd like to assert(work_q->size() != 0, ...)
4875       // because we just took work from the overflow list,
4876       // but of course we can't since all of that could have
4877       // been already stolen from us.
4878       // "He giveth and He taketh away."
4879       continue;
4880     }
4881     // Verify that we have no work before we resort to stealing
4882     assert(work_q->size() == 0, "Have work, shouldn't steal");
4883     // Try to steal from other queues that have work
4884     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4885       NOT_PRODUCT(num_steals++;)
4886       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4887       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4888       // Do scanning work
4889       obj_to_scan->oop_iterate(cl);
4890       // Loop around, finish this work, and try to steal some more
4891     } else if (terminator()->offer_termination()) {
4892         break;  // nirvana from the infinite cycle
4893     }
4894   }
4895   NOT_PRODUCT(
4896     if (PrintCMSStatistics != 0) {
4897       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4898     }
4899   )
4900   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4901          "Else our work is not yet done");
4902 }
4903 
4904 // Record object boundaries in _eden_chunk_array by sampling the eden
4905 // top in the slow-path eden object allocation code path and record
4906 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4907 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4908 // sampling in sample_eden() that activates during the part of the
4909 // preclean phase.
4910 void CMSCollector::sample_eden_chunk() {
4911   if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4912     if (_eden_chunk_lock->try_lock()) {
4913       // Record a sample. This is the critical section. The contents
4914       // of the _eden_chunk_array have to be non-decreasing in the
4915       // address order.
4916       _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4917       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4918              "Unexpected state of Eden");
4919       if (_eden_chunk_index == 0 ||
4920           ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4921            (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4922                           _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4923         _eden_chunk_index++;  // commit sample
4924       }
4925       _eden_chunk_lock->unlock();
4926     }
4927   }
4928 }
4929 
4930 // Return a thread-local PLAB recording array, as appropriate.
4931 void* CMSCollector::get_data_recorder(int thr_num) {
4932   if (_survivor_plab_array != NULL &&
4933       (CMSPLABRecordAlways ||
4934        (_collectorState > Marking && _collectorState < FinalMarking))) {
4935     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4936     ChunkArray* ca = &_survivor_plab_array[thr_num];
4937     ca->reset();   // clear it so that fresh data is recorded
4938     return (void*) ca;
4939   } else {
4940     return NULL;
4941   }
4942 }
4943 
4944 // Reset all the thread-local PLAB recording arrays
4945 void CMSCollector::reset_survivor_plab_arrays() {
4946   for (uint i = 0; i < ParallelGCThreads; i++) {
4947     _survivor_plab_array[i].reset();
4948   }
4949 }
4950 
4951 // Merge the per-thread plab arrays into the global survivor chunk
4952 // array which will provide the partitioning of the survivor space
4953 // for CMS initial scan and rescan.
4954 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4955                                               int no_of_gc_threads) {
4956   assert(_survivor_plab_array  != NULL, "Error");
4957   assert(_survivor_chunk_array != NULL, "Error");
4958   assert(_collectorState == FinalMarking ||
4959          (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4960   for (int j = 0; j < no_of_gc_threads; j++) {
4961     _cursor[j] = 0;
4962   }
4963   HeapWord* top = surv->top();
4964   size_t i;
4965   // Stride through the _survivor_plab_arrays based on the number of gc threads so
4966   // the whole structure is scanned regardless of the min plab size. This will
4967   // produce a more even distribution of scan tasks for the parallel phase.
4968   size_t stride = ParallelGCThreads > 4 ? ParallelGCThreads/2 : ParallelGCThreads;
4969   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
4970     HeapWord* min_val = top;          // Higher than any PLAB address
4971     uint      min_tid = 0;            // position of min_val this round
4972     for (int j = 0; j < no_of_gc_threads; j++) {
4973       ChunkArray* cur_sca = &_survivor_plab_array[j];
4974       if (_cursor[j] >= cur_sca->end()) {
4975         continue;
4976       }
4977       assert(_cursor[j] < (cur_sca->end()), "ctl pt invariant");
4978       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4979       assert(surv->used_region().contains(cur_val), "Out of bounds value");
4980       if (cur_val < min_val) {
4981         min_tid = j;
4982         min_val = cur_val;
4983       } else {
4984         assert(cur_val < top, "All recorded addresses should be less");
4985       }
4986     }
4987     // At this point min_val and min_tid are respectively
4988     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4989     // and the thread (j) that witnesses that address.
4990     // We record this address in the _survivor_chunk_array[i]
4991     // and increment _cursor[min_tid] prior to the next round i.
4992     if (min_val == top) {
4993       break;
4994     }
4995     _survivor_chunk_array[i] = min_val;
4996     _cursor[min_tid] += stride;
4997   }
4998   // We are all done; record the size of the _survivor_chunk_array
4999   _survivor_chunk_index = i; // exclusive: [0, i)
5000   if (PrintCMSStatistics > 0) {
5001     gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5002   }
5003   // Verify that we used up all the recorded entries
5004   #ifdef ASSERT
5005     size_t total = 0;
5006     for (int j = 0; j < no_of_gc_threads; j++) {
5007       assert(_cursor[j] >= _survivor_plab_array[j].end() && _cursor[j] < (_survivor_plab_array[j].end() + stride) ,
5008              err_msg("Did not use all entries of thread  " INT32_FORMAT ": _cursor[]=" SIZE_FORMAT
5009                       " != survivor_plab_array[]=" SIZE_FORMAT,
5010                       j, _cursor[j], _survivor_plab_array[j].end()));
5011       total += _cursor[j];
5012     }
5013     assert(total == (_survivor_chunk_index * stride),
5014            err_msg("Survivor stride error: total=" SIZE_FORMAT ": _survivor_chunk_index=" SIZE_FORMAT,
5015                    total, _survivor_chunk_index));
5016     // Check that the merged array is in sorted order
5017     if (_survivor_chunk_index > 0) {
5018       for (size_t i = 0; i < _survivor_chunk_index - 1; i++) {
5019         if (PrintCMSStatistics > 0) {
5020           gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5021                               i, p2i(_survivor_chunk_array[i]));
5022         }
5023         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5024                "Not sorted");
5025       }
5026     }
5027   #endif // ASSERT
5028 }
5029 
5030 // Set up the space's par_seq_tasks structure for work claiming
5031 // for parallel initial scan and rescan of young gen.
5032 // See ParRescanTask where this is currently used.
5033 void
5034 CMSCollector::
5035 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5036   assert(n_threads > 0, "Unexpected n_threads argument");
5037 
5038   // Eden space
5039   if (!_young_gen->eden()->is_empty()) {
5040     SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5041     assert(!pst->valid(), "Clobbering existing data?");
5042     // Each valid entry in [0, _eden_chunk_index) represents a task.
5043     size_t n_tasks = _eden_chunk_index + 1;
5044     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5045     // Sets the condition for completion of the subtask (how many threads
5046     // need to finish in order to be done).
5047     pst->set_n_threads(n_threads);
5048     pst->set_n_tasks((int)n_tasks);
5049   }
5050 
5051   // Merge the survivor plab arrays into _survivor_chunk_array
5052   if (_survivor_plab_array != NULL) {
5053     merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5054   } else {
5055     assert(_survivor_chunk_index == 0, "Error");
5056   }
5057 
5058   // To space
5059   {
5060     SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5061     assert(!pst->valid(), "Clobbering existing data?");
5062     // Sets the condition for completion of the subtask (how many threads
5063     // need to finish in order to be done).
5064     pst->set_n_threads(n_threads);
5065     pst->set_n_tasks(1);
5066     assert(pst->valid(), "Error");
5067   }
5068 
5069   // From space
5070   {
5071     SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5072     assert(!pst->valid(), "Clobbering existing data?");
5073     size_t n_tasks = _survivor_chunk_index + 1;
5074     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5075     // Sets the condition for completion of the subtask (how many threads
5076     // need to finish in order to be done).
5077     pst->set_n_threads(n_threads);
5078     pst->set_n_tasks((int)n_tasks);
5079     assert(pst->valid(), "Error");
5080   }
5081 }
5082 
5083 // Parallel version of remark
5084 void CMSCollector::do_remark_parallel() {
5085   GenCollectedHeap* gch = GenCollectedHeap::heap();
5086   FlexibleWorkGang* workers = gch->workers();
5087   assert(workers != NULL, "Need parallel worker threads.");
5088   // Choose to use the number of GC workers most recently set
5089   // into "active_workers".
5090   uint n_workers = workers->active_workers();
5091 
5092   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
5093 
5094   StrongRootsScope srs(n_workers);
5095 
5096   CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
5097 
5098   // We won't be iterating over the cards in the card table updating
5099   // the younger_gen cards, so we shouldn't call the following else
5100   // the verification code as well as subsequent younger_refs_iterate
5101   // code would get confused. XXX
5102   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5103 
5104   // The young gen rescan work will not be done as part of
5105   // process_roots (which currently doesn't know how to
5106   // parallelize such a scan), but rather will be broken up into
5107   // a set of parallel tasks (via the sampling that the [abortable]
5108   // preclean phase did of eden, plus the [two] tasks of
5109   // scanning the [two] survivor spaces. Further fine-grain
5110   // parallelization of the scanning of the survivor spaces
5111   // themselves, and of precleaning of the younger gen itself
5112   // is deferred to the future.
5113   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5114 
5115   // The dirty card rescan work is broken up into a "sequence"
5116   // of parallel tasks (per constituent space) that are dynamically
5117   // claimed by the parallel threads.
5118   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5119 
5120   // It turns out that even when we're using 1 thread, doing the work in a
5121   // separate thread causes wide variance in run times.  We can't help this
5122   // in the multi-threaded case, but we special-case n=1 here to get
5123   // repeatable measurements of the 1-thread overhead of the parallel code.
5124   if (n_workers > 1) {
5125     // Make refs discovery MT-safe, if it isn't already: it may not
5126     // necessarily be so, since it's possible that we are doing
5127     // ST marking.
5128     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5129     workers->run_task(&tsk);
5130   } else {
5131     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5132     tsk.work(0);
5133   }
5134 
5135   // restore, single-threaded for now, any preserved marks
5136   // as a result of work_q overflow
5137   restore_preserved_marks_if_any();
5138 }
5139 
5140 // Non-parallel version of remark
5141 void CMSCollector::do_remark_non_parallel() {
5142   ResourceMark rm;
5143   HandleMark   hm;
5144   GenCollectedHeap* gch = GenCollectedHeap::heap();
5145   ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5146 
5147   MarkRefsIntoAndScanClosure
5148     mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5149              &_markStack, this,
5150              false /* should_yield */, false /* not precleaning */);
5151   MarkFromDirtyCardsClosure
5152     markFromDirtyCardsClosure(this, _span,
5153                               NULL,  // space is set further below
5154                               &_markBitMap, &_markStack, &mrias_cl);
5155   {
5156     GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5157     // Iterate over the dirty cards, setting the corresponding bits in the
5158     // mod union table.
5159     {
5160       ModUnionClosure modUnionClosure(&_modUnionTable);
5161       _ct->ct_bs()->dirty_card_iterate(
5162                       _cmsGen->used_region(),
5163                       &modUnionClosure);
5164     }
5165     // Having transferred these marks into the modUnionTable, we just need
5166     // to rescan the marked objects on the dirty cards in the modUnionTable.
5167     // The initial marking may have been done during an asynchronous
5168     // collection so there may be dirty bits in the mod-union table.
5169     const int alignment =
5170       CardTableModRefBS::card_size * BitsPerWord;
5171     {
5172       // ... First handle dirty cards in CMS gen
5173       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5174       MemRegion ur = _cmsGen->used_region();
5175       HeapWord* lb = ur.start();
5176       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5177       MemRegion cms_span(lb, ub);
5178       _modUnionTable.dirty_range_iterate_clear(cms_span,
5179                                                &markFromDirtyCardsClosure);
5180       verify_work_stacks_empty();
5181       if (PrintCMSStatistics != 0) {
5182         gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ",
5183           markFromDirtyCardsClosure.num_dirty_cards());
5184       }
5185     }
5186   }
5187   if (VerifyDuringGC &&
5188       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5189     HandleMark hm;  // Discard invalid handles created during verification
5190     Universe::verify();
5191   }
5192   {
5193     GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5194 
5195     verify_work_stacks_empty();
5196 
5197     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5198     StrongRootsScope srs(1);
5199 
5200     gch->gen_process_roots(&srs,
5201                            GenCollectedHeap::OldGen,
5202                            true,  // younger gens as roots
5203                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
5204                            should_unload_classes(),
5205                            &mrias_cl,
5206                            NULL,
5207                            NULL); // The dirty klasses will be handled below
5208 
5209     assert(should_unload_classes()
5210            || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5211            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5212   }
5213 
5214   {
5215     GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5216 
5217     verify_work_stacks_empty();
5218 
5219     // Scan all class loader data objects that might have been introduced
5220     // during concurrent marking.
5221     ResourceMark rm;
5222     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5223     for (int i = 0; i < array->length(); i++) {
5224       mrias_cl.do_class_loader_data(array->at(i));
5225     }
5226 
5227     // We don't need to keep track of new CLDs anymore.
5228     ClassLoaderDataGraph::remember_new_clds(false);
5229 
5230     verify_work_stacks_empty();
5231   }
5232 
5233   {
5234     GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5235 
5236     verify_work_stacks_empty();
5237 
5238     RemarkKlassClosure remark_klass_closure(&mrias_cl);
5239     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5240 
5241     verify_work_stacks_empty();
5242   }
5243 
5244   // We might have added oops to ClassLoaderData::_handles during the
5245   // concurrent marking phase. These oops point to newly allocated objects
5246   // that are guaranteed to be kept alive. Either by the direct allocation
5247   // code, or when the young collector processes the roots. Hence,
5248   // we don't have to revisit the _handles block during the remark phase.
5249 
5250   verify_work_stacks_empty();
5251   // Restore evacuated mark words, if any, used for overflow list links
5252   if (!CMSOverflowEarlyRestoration) {
5253     restore_preserved_marks_if_any();
5254   }
5255   verify_overflow_empty();
5256 }
5257 
5258 ////////////////////////////////////////////////////////
5259 // Parallel Reference Processing Task Proxy Class
5260 ////////////////////////////////////////////////////////
5261 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5262   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5263   CMSCollector*          _collector;
5264   CMSBitMap*             _mark_bit_map;
5265   const MemRegion        _span;
5266   ProcessTask&           _task;
5267 
5268 public:
5269   CMSRefProcTaskProxy(ProcessTask&     task,
5270                       CMSCollector*    collector,
5271                       const MemRegion& span,
5272                       CMSBitMap*       mark_bit_map,
5273                       AbstractWorkGang* workers,
5274                       OopTaskQueueSet* task_queues):
5275     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5276       task_queues,
5277       workers->active_workers()),
5278     _task(task),
5279     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5280   {
5281     assert(_collector->_span.equals(_span) && !_span.is_empty(),
5282            "Inconsistency in _span");
5283   }
5284 
5285   OopTaskQueueSet* task_queues() { return queues(); }
5286 
5287   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5288 
5289   void do_work_steal(int i,
5290                      CMSParDrainMarkingStackClosure* drain,
5291                      CMSParKeepAliveClosure* keep_alive,
5292                      int* seed);
5293 
5294   virtual void work(uint worker_id);
5295 };
5296 
5297 void CMSRefProcTaskProxy::work(uint worker_id) {
5298   ResourceMark rm;
5299   HandleMark hm;
5300   assert(_collector->_span.equals(_span), "Inconsistency in _span");
5301   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5302                                         _mark_bit_map,
5303                                         work_queue(worker_id));
5304   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5305                                                  _mark_bit_map,
5306                                                  work_queue(worker_id));
5307   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5308   _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5309   if (_task.marks_oops_alive()) {
5310     do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5311                   _collector->hash_seed(worker_id));
5312   }
5313   assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5314   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5315 }
5316 
5317 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5318   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5319   EnqueueTask& _task;
5320 
5321 public:
5322   CMSRefEnqueueTaskProxy(EnqueueTask& task)
5323     : AbstractGangTask("Enqueue reference objects in parallel"),
5324       _task(task)
5325   { }
5326 
5327   virtual void work(uint worker_id)
5328   {
5329     _task.work(worker_id);
5330   }
5331 };
5332 
5333 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5334   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5335    _span(span),
5336    _bit_map(bit_map),
5337    _work_queue(work_queue),
5338    _mark_and_push(collector, span, bit_map, work_queue),
5339    _low_water_mark(MIN2((work_queue->max_elems()/4),
5340                         ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5341 { }
5342 
5343 // . see if we can share work_queues with ParNew? XXX
5344 void CMSRefProcTaskProxy::do_work_steal(int i,
5345   CMSParDrainMarkingStackClosure* drain,
5346   CMSParKeepAliveClosure* keep_alive,
5347   int* seed) {
5348   OopTaskQueue* work_q = work_queue(i);
5349   NOT_PRODUCT(int num_steals = 0;)
5350   oop obj_to_scan;
5351 
5352   while (true) {
5353     // Completely finish any left over work from (an) earlier round(s)
5354     drain->trim_queue(0);
5355     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5356                                          (size_t)ParGCDesiredObjsFromOverflowList);
5357     // Now check if there's any work in the overflow list
5358     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5359     // only affects the number of attempts made to get work from the
5360     // overflow list and does not affect the number of workers.  Just
5361     // pass ParallelGCThreads so this behavior is unchanged.
5362     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5363                                                 work_q,
5364                                                 ParallelGCThreads)) {
5365       // Found something in global overflow list;
5366       // not yet ready to go stealing work from others.
5367       // We'd like to assert(work_q->size() != 0, ...)
5368       // because we just took work from the overflow list,
5369       // but of course we can't, since all of that might have
5370       // been already stolen from us.
5371       continue;
5372     }
5373     // Verify that we have no work before we resort to stealing
5374     assert(work_q->size() == 0, "Have work, shouldn't steal");
5375     // Try to steal from other queues that have work
5376     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5377       NOT_PRODUCT(num_steals++;)
5378       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5379       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5380       // Do scanning work
5381       obj_to_scan->oop_iterate(keep_alive);
5382       // Loop around, finish this work, and try to steal some more
5383     } else if (terminator()->offer_termination()) {
5384       break;  // nirvana from the infinite cycle
5385     }
5386   }
5387   NOT_PRODUCT(
5388     if (PrintCMSStatistics != 0) {
5389       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5390     }
5391   )
5392 }
5393 
5394 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5395 {
5396   GenCollectedHeap* gch = GenCollectedHeap::heap();
5397   FlexibleWorkGang* workers = gch->workers();
5398   assert(workers != NULL, "Need parallel worker threads.");
5399   CMSRefProcTaskProxy rp_task(task, &_collector,
5400                               _collector.ref_processor()->span(),
5401                               _collector.markBitMap(),
5402                               workers, _collector.task_queues());
5403   workers->run_task(&rp_task);
5404 }
5405 
5406 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5407 {
5408 
5409   GenCollectedHeap* gch = GenCollectedHeap::heap();
5410   FlexibleWorkGang* workers = gch->workers();
5411   assert(workers != NULL, "Need parallel worker threads.");
5412   CMSRefEnqueueTaskProxy enq_task(task);
5413   workers->run_task(&enq_task);
5414 }
5415 
5416 void CMSCollector::refProcessingWork() {
5417   ResourceMark rm;
5418   HandleMark   hm;
5419 
5420   ReferenceProcessor* rp = ref_processor();
5421   assert(rp->span().equals(_span), "Spans should be equal");
5422   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5423   // Process weak references.
5424   rp->setup_policy(false);
5425   verify_work_stacks_empty();
5426 
5427   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5428                                           &_markStack, false /* !preclean */);
5429   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5430                                 _span, &_markBitMap, &_markStack,
5431                                 &cmsKeepAliveClosure, false /* !preclean */);
5432   {
5433     GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5434 
5435     ReferenceProcessorStats stats;
5436     if (rp->processing_is_mt()) {
5437       // Set the degree of MT here.  If the discovery is done MT, there
5438       // may have been a different number of threads doing the discovery
5439       // and a different number of discovered lists may have Ref objects.
5440       // That is OK as long as the Reference lists are balanced (see
5441       // balance_all_queues() and balance_queues()).
5442       GenCollectedHeap* gch = GenCollectedHeap::heap();
5443       uint active_workers = ParallelGCThreads;
5444       FlexibleWorkGang* workers = gch->workers();
5445       if (workers != NULL) {
5446         active_workers = workers->active_workers();
5447         // The expectation is that active_workers will have already
5448         // been set to a reasonable value.  If it has not been set,
5449         // investigate.
5450         assert(active_workers > 0, "Should have been set during scavenge");
5451       }
5452       rp->set_active_mt_degree(active_workers);
5453       CMSRefProcTaskExecutor task_executor(*this);
5454       stats = rp->process_discovered_references(&_is_alive_closure,
5455                                         &cmsKeepAliveClosure,
5456                                         &cmsDrainMarkingStackClosure,
5457                                         &task_executor,
5458                                         _gc_timer_cm,
5459                                         _gc_tracer_cm->gc_id());
5460     } else {
5461       stats = rp->process_discovered_references(&_is_alive_closure,
5462                                         &cmsKeepAliveClosure,
5463                                         &cmsDrainMarkingStackClosure,
5464                                         NULL,
5465                                         _gc_timer_cm,
5466                                         _gc_tracer_cm->gc_id());
5467     }
5468     _gc_tracer_cm->report_gc_reference_stats(stats);
5469 
5470   }
5471 
5472   // This is the point where the entire marking should have completed.
5473   verify_work_stacks_empty();
5474 
5475   if (should_unload_classes()) {
5476     {
5477       GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5478 
5479       // Unload classes and purge the SystemDictionary.
5480       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5481 
5482       // Unload nmethods.
5483       CodeCache::do_unloading(&_is_alive_closure, purged_class);
5484 
5485       // Prune dead klasses from subklass/sibling/implementor lists.
5486       Klass::clean_weak_klass_links(&_is_alive_closure);
5487     }
5488 
5489     {
5490       GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5491       // Clean up unreferenced symbols in symbol table.
5492       SymbolTable::unlink();
5493     }
5494 
5495     {
5496       GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5497       // Delete entries for dead interned strings.
5498       StringTable::unlink(&_is_alive_closure);
5499     }
5500   }
5501 
5502 
5503   // Restore any preserved marks as a result of mark stack or
5504   // work queue overflow
5505   restore_preserved_marks_if_any();  // done single-threaded for now
5506 
5507   rp->set_enqueuing_is_done(true);
5508   if (rp->processing_is_mt()) {
5509     rp->balance_all_queues();
5510     CMSRefProcTaskExecutor task_executor(*this);
5511     rp->enqueue_discovered_references(&task_executor);
5512   } else {
5513     rp->enqueue_discovered_references(NULL);
5514   }
5515   rp->verify_no_references_recorded();
5516   assert(!rp->discovery_enabled(), "should have been disabled");
5517 }
5518 
5519 #ifndef PRODUCT
5520 void CMSCollector::check_correct_thread_executing() {
5521   Thread* t = Thread::current();
5522   // Only the VM thread or the CMS thread should be here.
5523   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5524          "Unexpected thread type");
5525   // If this is the vm thread, the foreground process
5526   // should not be waiting.  Note that _foregroundGCIsActive is
5527   // true while the foreground collector is waiting.
5528   if (_foregroundGCShouldWait) {
5529     // We cannot be the VM thread
5530     assert(t->is_ConcurrentGC_thread(),
5531            "Should be CMS thread");
5532   } else {
5533     // We can be the CMS thread only if we are in a stop-world
5534     // phase of CMS collection.
5535     if (t->is_ConcurrentGC_thread()) {
5536       assert(_collectorState == InitialMarking ||
5537              _collectorState == FinalMarking,
5538              "Should be a stop-world phase");
5539       // The CMS thread should be holding the CMS_token.
5540       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5541              "Potential interference with concurrently "
5542              "executing VM thread");
5543     }
5544   }
5545 }
5546 #endif
5547 
5548 void CMSCollector::sweep() {
5549   assert(_collectorState == Sweeping, "just checking");
5550   check_correct_thread_executing();
5551   verify_work_stacks_empty();
5552   verify_overflow_empty();
5553   increment_sweep_count();
5554   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5555 
5556   _inter_sweep_timer.stop();
5557   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5558 
5559   assert(!_intra_sweep_timer.is_active(), "Should not be active");
5560   _intra_sweep_timer.reset();
5561   _intra_sweep_timer.start();
5562   {
5563     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5564     CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5565     // First sweep the old gen
5566     {
5567       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5568                                bitMapLock());
5569       sweepWork(_cmsGen);
5570     }
5571 
5572     // Update Universe::_heap_*_at_gc figures.
5573     // We need all the free list locks to make the abstract state
5574     // transition from Sweeping to Resetting. See detailed note
5575     // further below.
5576     {
5577       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5578       // Update heap occupancy information which is used as
5579       // input to soft ref clearing policy at the next gc.
5580       Universe::update_heap_info_at_gc();
5581       _collectorState = Resizing;
5582     }
5583   }
5584   verify_work_stacks_empty();
5585   verify_overflow_empty();
5586 
5587   if (should_unload_classes()) {
5588     // Delay purge to the beginning of the next safepoint.  Metaspace::contains
5589     // requires that the virtual spaces are stable and not deleted.
5590     ClassLoaderDataGraph::set_should_purge(true);
5591   }
5592 
5593   _intra_sweep_timer.stop();
5594   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5595 
5596   _inter_sweep_timer.reset();
5597   _inter_sweep_timer.start();
5598 
5599   // We need to use a monotonically non-decreasing time in ms
5600   // or we will see time-warp warnings and os::javaTimeMillis()
5601   // does not guarantee monotonicity.
5602   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5603   update_time_of_last_gc(now);
5604 
5605   // NOTE on abstract state transitions:
5606   // Mutators allocate-live and/or mark the mod-union table dirty
5607   // based on the state of the collection.  The former is done in
5608   // the interval [Marking, Sweeping] and the latter in the interval
5609   // [Marking, Sweeping).  Thus the transitions into the Marking state
5610   // and out of the Sweeping state must be synchronously visible
5611   // globally to the mutators.
5612   // The transition into the Marking state happens with the world
5613   // stopped so the mutators will globally see it.  Sweeping is
5614   // done asynchronously by the background collector so the transition
5615   // from the Sweeping state to the Resizing state must be done
5616   // under the freelistLock (as is the check for whether to
5617   // allocate-live and whether to dirty the mod-union table).
5618   assert(_collectorState == Resizing, "Change of collector state to"
5619     " Resizing must be done under the freelistLocks (plural)");
5620 
5621   // Now that sweeping has been completed, we clear
5622   // the incremental_collection_failed flag,
5623   // thus inviting a younger gen collection to promote into
5624   // this generation. If such a promotion may still fail,
5625   // the flag will be set again when a young collection is
5626   // attempted.
5627   GenCollectedHeap* gch = GenCollectedHeap::heap();
5628   gch->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
5629   gch->update_full_collections_completed(_collection_count_start);
5630 }
5631 
5632 // FIX ME!!! Looks like this belongs in CFLSpace, with
5633 // CMSGen merely delegating to it.
5634 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5635   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5636   HeapWord*  minAddr        = _cmsSpace->bottom();
5637   HeapWord*  largestAddr    =
5638     (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5639   if (largestAddr == NULL) {
5640     // The dictionary appears to be empty.  In this case
5641     // try to coalesce at the end of the heap.
5642     largestAddr = _cmsSpace->end();
5643   }
5644   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
5645   size_t nearLargestOffset =
5646     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5647   if (PrintFLSStatistics != 0) {
5648     gclog_or_tty->print_cr(
5649       "CMS: Large Block: " PTR_FORMAT ";"
5650       " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5651       p2i(largestAddr),
5652       p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5653   }
5654   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5655 }
5656 
5657 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5658   return addr >= _cmsSpace->nearLargestChunk();
5659 }
5660 
5661 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5662   return _cmsSpace->find_chunk_at_end();
5663 }
5664 
5665 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5666                                                     bool full) {
5667   // If the young generation has been collected, gather any statistics
5668   // that are of interest at this point.
5669   bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5670   if (!full && current_is_young) {
5671     // Gather statistics on the young generation collection.
5672     collector()->stats().record_gc0_end(used());
5673   }
5674 }
5675 
5676 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen) {
5677   // We iterate over the space(s) underlying this generation,
5678   // checking the mark bit map to see if the bits corresponding
5679   // to specific blocks are marked or not. Blocks that are
5680   // marked are live and are not swept up. All remaining blocks
5681   // are swept up, with coalescing on-the-fly as we sweep up
5682   // contiguous free and/or garbage blocks:
5683   // We need to ensure that the sweeper synchronizes with allocators
5684   // and stop-the-world collectors. In particular, the following
5685   // locks are used:
5686   // . CMS token: if this is held, a stop the world collection cannot occur
5687   // . freelistLock: if this is held no allocation can occur from this
5688   //                 generation by another thread
5689   // . bitMapLock: if this is held, no other thread can access or update
5690   //
5691 
5692   // Note that we need to hold the freelistLock if we use
5693   // block iterate below; else the iterator might go awry if
5694   // a mutator (or promotion) causes block contents to change
5695   // (for instance if the allocator divvies up a block).
5696   // If we hold the free list lock, for all practical purposes
5697   // young generation GC's can't occur (they'll usually need to
5698   // promote), so we might as well prevent all young generation
5699   // GC's while we do a sweeping step. For the same reason, we might
5700   // as well take the bit map lock for the entire duration
5701 
5702   // check that we hold the requisite locks
5703   assert(have_cms_token(), "Should hold cms token");
5704   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5705   assert_lock_strong(gen->freelistLock());
5706   assert_lock_strong(bitMapLock());
5707 
5708   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5709   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
5710   gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5711                                       _inter_sweep_estimate.padded_average(),
5712                                       _intra_sweep_estimate.padded_average());
5713   gen->setNearLargestChunk();
5714 
5715   {
5716     SweepClosure sweepClosure(this, gen, &_markBitMap, CMSYield);
5717     gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5718     // We need to free-up/coalesce garbage/blocks from a
5719     // co-terminal free run. This is done in the SweepClosure
5720     // destructor; so, do not remove this scope, else the
5721     // end-of-sweep-census below will be off by a little bit.
5722   }
5723   gen->cmsSpace()->sweep_completed();
5724   gen->cmsSpace()->endSweepFLCensus(sweep_count());
5725   if (should_unload_classes()) {                // unloaded classes this cycle,
5726     _concurrent_cycles_since_last_unload = 0;   // ... reset count
5727   } else {                                      // did not unload classes,
5728     _concurrent_cycles_since_last_unload++;     // ... increment count
5729   }
5730 }
5731 
5732 // Reset CMS data structures (for now just the marking bit map)
5733 // preparatory for the next cycle.
5734 void CMSCollector::reset(bool concurrent) {
5735   if (concurrent) {
5736     CMSTokenSyncWithLocks ts(true, bitMapLock());
5737 
5738     // If the state is not "Resetting", the foreground  thread
5739     // has done a collection and the resetting.
5740     if (_collectorState != Resetting) {
5741       assert(_collectorState == Idling, "The state should only change"
5742         " because the foreground collector has finished the collection");
5743       return;
5744     }
5745 
5746     // Clear the mark bitmap (no grey objects to start with)
5747     // for the next cycle.
5748     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5749     CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5750 
5751     HeapWord* curAddr = _markBitMap.startWord();
5752     while (curAddr < _markBitMap.endWord()) {
5753       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
5754       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5755       _markBitMap.clear_large_range(chunk);
5756       if (ConcurrentMarkSweepThread::should_yield() &&
5757           !foregroundGCIsActive() &&
5758           CMSYield) {
5759         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5760                "CMS thread should hold CMS token");
5761         assert_lock_strong(bitMapLock());
5762         bitMapLock()->unlock();
5763         ConcurrentMarkSweepThread::desynchronize(true);
5764         stopTimer();
5765         if (PrintCMSStatistics != 0) {
5766           incrementYields();
5767         }
5768 
5769         // See the comment in coordinator_yield()
5770         for (unsigned i = 0; i < CMSYieldSleepCount &&
5771                          ConcurrentMarkSweepThread::should_yield() &&
5772                          !CMSCollector::foregroundGCIsActive(); ++i) {
5773           os::sleep(Thread::current(), 1, false);
5774         }
5775 
5776         ConcurrentMarkSweepThread::synchronize(true);
5777         bitMapLock()->lock_without_safepoint_check();
5778         startTimer();
5779       }
5780       curAddr = chunk.end();
5781     }
5782     // A successful mostly concurrent collection has been done.
5783     // Because only the full (i.e., concurrent mode failure) collections
5784     // are being measured for gc overhead limits, clean the "near" flag
5785     // and count.
5786     size_policy()->reset_gc_overhead_limit_count();
5787     _collectorState = Idling;
5788   } else {
5789     // already have the lock
5790     assert(_collectorState == Resetting, "just checking");
5791     assert_lock_strong(bitMapLock());
5792     _markBitMap.clear_all();
5793     _collectorState = Idling;
5794   }
5795 
5796   register_gc_end();
5797 }
5798 
5799 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5800   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5801   GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5802   TraceCollectorStats tcs(counters());
5803 
5804   switch (op) {
5805     case CMS_op_checkpointRootsInitial: {
5806       SvcGCMarker sgcm(SvcGCMarker::OTHER);
5807       checkpointRootsInitial();
5808       if (PrintGC) {
5809         _cmsGen->printOccupancy("initial-mark");
5810       }
5811       break;
5812     }
5813     case CMS_op_checkpointRootsFinal: {
5814       SvcGCMarker sgcm(SvcGCMarker::OTHER);
5815       checkpointRootsFinal();
5816       if (PrintGC) {
5817         _cmsGen->printOccupancy("remark");
5818       }
5819       break;
5820     }
5821     default:
5822       fatal("No such CMS_op");
5823   }
5824 }
5825 
5826 #ifndef PRODUCT
5827 size_t const CMSCollector::skip_header_HeapWords() {
5828   return FreeChunk::header_size();
5829 }
5830 
5831 // Try and collect here conditions that should hold when
5832 // CMS thread is exiting. The idea is that the foreground GC
5833 // thread should not be blocked if it wants to terminate
5834 // the CMS thread and yet continue to run the VM for a while
5835 // after that.
5836 void CMSCollector::verify_ok_to_terminate() const {
5837   assert(Thread::current()->is_ConcurrentGC_thread(),
5838          "should be called by CMS thread");
5839   assert(!_foregroundGCShouldWait, "should be false");
5840   // We could check here that all the various low-level locks
5841   // are not held by the CMS thread, but that is overkill; see
5842   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5843   // is checked.
5844 }
5845 #endif
5846 
5847 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5848    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5849           "missing Printezis mark?");
5850   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5851   size_t size = pointer_delta(nextOneAddr + 1, addr);
5852   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5853          "alignment problem");
5854   assert(size >= 3, "Necessary for Printezis marks to work");
5855   return size;
5856 }
5857 
5858 // A variant of the above (block_size_using_printezis_bits()) except
5859 // that we return 0 if the P-bits are not yet set.
5860 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5861   if (_markBitMap.isMarked(addr + 1)) {
5862     assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5863     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5864     size_t size = pointer_delta(nextOneAddr + 1, addr);
5865     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5866            "alignment problem");
5867     assert(size >= 3, "Necessary for Printezis marks to work");
5868     return size;
5869   }
5870   return 0;
5871 }
5872 
5873 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5874   size_t sz = 0;
5875   oop p = (oop)addr;
5876   if (p->klass_or_null() != NULL) {
5877     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5878   } else {
5879     sz = block_size_using_printezis_bits(addr);
5880   }
5881   assert(sz > 0, "size must be nonzero");
5882   HeapWord* next_block = addr + sz;
5883   HeapWord* next_card  = (HeapWord*)round_to((uintptr_t)next_block,
5884                                              CardTableModRefBS::card_size);
5885   assert(round_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
5886          round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5887          "must be different cards");
5888   return next_card;
5889 }
5890 
5891 
5892 // CMS Bit Map Wrapper /////////////////////////////////////////
5893 
5894 // Construct a CMS bit map infrastructure, but don't create the
5895 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5896 // further below.
5897 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5898   _bm(),
5899   _shifter(shifter),
5900   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5901                                     Monitor::_safepoint_check_sometimes) : NULL)
5902 {
5903   _bmStartWord = 0;
5904   _bmWordSize  = 0;
5905 }
5906 
5907 bool CMSBitMap::allocate(MemRegion mr) {
5908   _bmStartWord = mr.start();
5909   _bmWordSize  = mr.word_size();
5910   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5911                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5912   if (!brs.is_reserved()) {
5913     warning("CMS bit map allocation failure");
5914     return false;
5915   }
5916   // For now we'll just commit all of the bit map up front.
5917   // Later on we'll try to be more parsimonious with swap.
5918   if (!_virtual_space.initialize(brs, brs.size())) {
5919     warning("CMS bit map backing store failure");
5920     return false;
5921   }
5922   assert(_virtual_space.committed_size() == brs.size(),
5923          "didn't reserve backing store for all of CMS bit map?");
5924   _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5925   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5926          _bmWordSize, "inconsistency in bit map sizing");
5927   _bm.set_size(_bmWordSize >> _shifter);
5928 
5929   // bm.clear(); // can we rely on getting zero'd memory? verify below
5930   assert(isAllClear(),
5931          "Expected zero'd memory from ReservedSpace constructor");
5932   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5933          "consistency check");
5934   return true;
5935 }
5936 
5937 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5938   HeapWord *next_addr, *end_addr, *last_addr;
5939   assert_locked();
5940   assert(covers(mr), "out-of-range error");
5941   // XXX assert that start and end are appropriately aligned
5942   for (next_addr = mr.start(), end_addr = mr.end();
5943        next_addr < end_addr; next_addr = last_addr) {
5944     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5945     last_addr = dirty_region.end();
5946     if (!dirty_region.is_empty()) {
5947       cl->do_MemRegion(dirty_region);
5948     } else {
5949       assert(last_addr == end_addr, "program logic");
5950       return;
5951     }
5952   }
5953 }
5954 
5955 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5956   _bm.print_on_error(st, prefix);
5957 }
5958 
5959 #ifndef PRODUCT
5960 void CMSBitMap::assert_locked() const {
5961   CMSLockVerifier::assert_locked(lock());
5962 }
5963 
5964 bool CMSBitMap::covers(MemRegion mr) const {
5965   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5966   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5967          "size inconsistency");
5968   return (mr.start() >= _bmStartWord) &&
5969          (mr.end()   <= endWord());
5970 }
5971 
5972 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5973     return (start >= _bmStartWord && (start + size) <= endWord());
5974 }
5975 
5976 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5977   // verify that there are no 1 bits in the interval [left, right)
5978   FalseBitMapClosure falseBitMapClosure;
5979   iterate(&falseBitMapClosure, left, right);
5980 }
5981 
5982 void CMSBitMap::region_invariant(MemRegion mr)
5983 {
5984   assert_locked();
5985   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5986   assert(!mr.is_empty(), "unexpected empty region");
5987   assert(covers(mr), "mr should be covered by bit map");
5988   // convert address range into offset range
5989   size_t start_ofs = heapWordToOffset(mr.start());
5990   // Make sure that end() is appropriately aligned
5991   assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5992                         (1 << (_shifter+LogHeapWordSize))),
5993          "Misaligned mr.end()");
5994   size_t end_ofs   = heapWordToOffset(mr.end());
5995   assert(end_ofs > start_ofs, "Should mark at least one bit");
5996 }
5997 
5998 #endif
5999 
6000 bool CMSMarkStack::allocate(size_t size) {
6001   // allocate a stack of the requisite depth
6002   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6003                    size * sizeof(oop)));
6004   if (!rs.is_reserved()) {
6005     warning("CMSMarkStack allocation failure");
6006     return false;
6007   }
6008   if (!_virtual_space.initialize(rs, rs.size())) {
6009     warning("CMSMarkStack backing store failure");
6010     return false;
6011   }
6012   assert(_virtual_space.committed_size() == rs.size(),
6013          "didn't reserve backing store for all of CMS stack?");
6014   _base = (oop*)(_virtual_space.low());
6015   _index = 0;
6016   _capacity = size;
6017   NOT_PRODUCT(_max_depth = 0);
6018   return true;
6019 }
6020 
6021 // XXX FIX ME !!! In the MT case we come in here holding a
6022 // leaf lock. For printing we need to take a further lock
6023 // which has lower rank. We need to recalibrate the two
6024 // lock-ranks involved in order to be able to print the
6025 // messages below. (Or defer the printing to the caller.
6026 // For now we take the expedient path of just disabling the
6027 // messages for the problematic case.)
6028 void CMSMarkStack::expand() {
6029   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6030   if (_capacity == MarkStackSizeMax) {
6031     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6032       // We print a warning message only once per CMS cycle.
6033       gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6034     }
6035     return;
6036   }
6037   // Double capacity if possible
6038   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6039   // Do not give up existing stack until we have managed to
6040   // get the double capacity that we desired.
6041   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6042                    new_capacity * sizeof(oop)));
6043   if (rs.is_reserved()) {
6044     // Release the backing store associated with old stack
6045     _virtual_space.release();
6046     // Reinitialize virtual space for new stack
6047     if (!_virtual_space.initialize(rs, rs.size())) {
6048       fatal("Not enough swap for expanded marking stack");
6049     }
6050     _base = (oop*)(_virtual_space.low());
6051     _index = 0;
6052     _capacity = new_capacity;
6053   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6054     // Failed to double capacity, continue;
6055     // we print a detail message only once per CMS cycle.
6056     gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to "
6057             SIZE_FORMAT "K",
6058             _capacity / K, new_capacity / K);
6059   }
6060 }
6061 
6062 
6063 // Closures
6064 // XXX: there seems to be a lot of code  duplication here;
6065 // should refactor and consolidate common code.
6066 
6067 // This closure is used to mark refs into the CMS generation in
6068 // the CMS bit map. Called at the first checkpoint. This closure
6069 // assumes that we do not need to re-mark dirty cards; if the CMS
6070 // generation on which this is used is not an oldest
6071 // generation then this will lose younger_gen cards!
6072 
6073 MarkRefsIntoClosure::MarkRefsIntoClosure(
6074   MemRegion span, CMSBitMap* bitMap):
6075     _span(span),
6076     _bitMap(bitMap)
6077 {
6078     assert(_ref_processor == NULL, "deliberately left NULL");
6079     assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6080 }
6081 
6082 void MarkRefsIntoClosure::do_oop(oop obj) {
6083   // if p points into _span, then mark corresponding bit in _markBitMap
6084   assert(obj->is_oop(), "expected an oop");
6085   HeapWord* addr = (HeapWord*)obj;
6086   if (_span.contains(addr)) {
6087     // this should be made more efficient
6088     _bitMap->mark(addr);
6089   }
6090 }
6091 
6092 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
6093 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6094 
6095 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6096   MemRegion span, CMSBitMap* bitMap):
6097     _span(span),
6098     _bitMap(bitMap)
6099 {
6100     assert(_ref_processor == NULL, "deliberately left NULL");
6101     assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6102 }
6103 
6104 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6105   // if p points into _span, then mark corresponding bit in _markBitMap
6106   assert(obj->is_oop(), "expected an oop");
6107   HeapWord* addr = (HeapWord*)obj;
6108   if (_span.contains(addr)) {
6109     // this should be made more efficient
6110     _bitMap->par_mark(addr);
6111   }
6112 }
6113 
6114 void Par_MarkRefsIntoClosure::do_oop(oop* p)       { Par_MarkRefsIntoClosure::do_oop_work(p); }
6115 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6116 
6117 // A variant of the above, used for CMS marking verification.
6118 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6119   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6120     _span(span),
6121     _verification_bm(verification_bm),
6122     _cms_bm(cms_bm)
6123 {
6124     assert(_ref_processor == NULL, "deliberately left NULL");
6125     assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6126 }
6127 
6128 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6129   // if p points into _span, then mark corresponding bit in _markBitMap
6130   assert(obj->is_oop(), "expected an oop");
6131   HeapWord* addr = (HeapWord*)obj;
6132   if (_span.contains(addr)) {
6133     _verification_bm->mark(addr);
6134     if (!_cms_bm->isMarked(addr)) {
6135       oop(addr)->print();
6136       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6137       fatal("... aborting");
6138     }
6139   }
6140 }
6141 
6142 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6143 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6144 
6145 //////////////////////////////////////////////////
6146 // MarkRefsIntoAndScanClosure
6147 //////////////////////////////////////////////////
6148 
6149 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6150                                                        ReferenceProcessor* rp,
6151                                                        CMSBitMap* bit_map,
6152                                                        CMSBitMap* mod_union_table,
6153                                                        CMSMarkStack*  mark_stack,
6154                                                        CMSCollector* collector,
6155                                                        bool should_yield,
6156                                                        bool concurrent_precleaning):
6157   _collector(collector),
6158   _span(span),
6159   _bit_map(bit_map),
6160   _mark_stack(mark_stack),
6161   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6162                       mark_stack, concurrent_precleaning),
6163   _yield(should_yield),
6164   _concurrent_precleaning(concurrent_precleaning),
6165   _freelistLock(NULL)
6166 {
6167   _ref_processor = rp;
6168   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6169 }
6170 
6171 // This closure is used to mark refs into the CMS generation at the
6172 // second (final) checkpoint, and to scan and transitively follow
6173 // the unmarked oops. It is also used during the concurrent precleaning
6174 // phase while scanning objects on dirty cards in the CMS generation.
6175 // The marks are made in the marking bit map and the marking stack is
6176 // used for keeping the (newly) grey objects during the scan.
6177 // The parallel version (Par_...) appears further below.
6178 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6179   if (obj != NULL) {
6180     assert(obj->is_oop(), "expected an oop");
6181     HeapWord* addr = (HeapWord*)obj;
6182     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6183     assert(_collector->overflow_list_is_empty(),
6184            "overflow list should be empty");
6185     if (_span.contains(addr) &&
6186         !_bit_map->isMarked(addr)) {
6187       // mark bit map (object is now grey)
6188       _bit_map->mark(addr);
6189       // push on marking stack (stack should be empty), and drain the
6190       // stack by applying this closure to the oops in the oops popped
6191       // from the stack (i.e. blacken the grey objects)
6192       bool res = _mark_stack->push(obj);
6193       assert(res, "Should have space to push on empty stack");
6194       do {
6195         oop new_oop = _mark_stack->pop();
6196         assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6197         assert(_bit_map->isMarked((HeapWord*)new_oop),
6198                "only grey objects on this stack");
6199         // iterate over the oops in this oop, marking and pushing
6200         // the ones in CMS heap (i.e. in _span).
6201         new_oop->oop_iterate(&_pushAndMarkClosure);
6202         // check if it's time to yield
6203         do_yield_check();
6204       } while (!_mark_stack->isEmpty() ||
6205                (!_concurrent_precleaning && take_from_overflow_list()));
6206         // if marking stack is empty, and we are not doing this
6207         // during precleaning, then check the overflow list
6208     }
6209     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6210     assert(_collector->overflow_list_is_empty(),
6211            "overflow list was drained above");
6212     // We could restore evacuated mark words, if any, used for
6213     // overflow list links here because the overflow list is
6214     // provably empty here. That would reduce the maximum
6215     // size requirements for preserved_{oop,mark}_stack.
6216     // But we'll just postpone it until we are all done
6217     // so we can just stream through.
6218     if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6219       _collector->restore_preserved_marks_if_any();
6220       assert(_collector->no_preserved_marks(), "No preserved marks");
6221     }
6222     assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6223            "All preserved marks should have been restored above");
6224   }
6225 }
6226 
6227 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6228 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6229 
6230 void MarkRefsIntoAndScanClosure::do_yield_work() {
6231   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6232          "CMS thread should hold CMS token");
6233   assert_lock_strong(_freelistLock);
6234   assert_lock_strong(_bit_map->lock());
6235   // relinquish the free_list_lock and bitMaplock()
6236   _bit_map->lock()->unlock();
6237   _freelistLock->unlock();
6238   ConcurrentMarkSweepThread::desynchronize(true);
6239   _collector->stopTimer();
6240   if (PrintCMSStatistics != 0) {
6241     _collector->incrementYields();
6242   }
6243 
6244   // See the comment in coordinator_yield()
6245   for (unsigned i = 0;
6246        i < CMSYieldSleepCount &&
6247        ConcurrentMarkSweepThread::should_yield() &&
6248        !CMSCollector::foregroundGCIsActive();
6249        ++i) {
6250     os::sleep(Thread::current(), 1, false);
6251   }
6252 
6253   ConcurrentMarkSweepThread::synchronize(true);
6254   _freelistLock->lock_without_safepoint_check();
6255   _bit_map->lock()->lock_without_safepoint_check();
6256   _collector->startTimer();
6257 }
6258 
6259 ///////////////////////////////////////////////////////////
6260 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6261 //                                 MarkRefsIntoAndScanClosure
6262 ///////////////////////////////////////////////////////////
6263 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6264   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6265   CMSBitMap* bit_map, OopTaskQueue* work_queue):
6266   _span(span),
6267   _bit_map(bit_map),
6268   _work_queue(work_queue),
6269   _low_water_mark(MIN2((work_queue->max_elems()/4),
6270                        ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6271   _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6272 {
6273   _ref_processor = rp;
6274   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6275 }
6276 
6277 // This closure is used to mark refs into the CMS generation at the
6278 // second (final) checkpoint, and to scan and transitively follow
6279 // the unmarked oops. The marks are made in the marking bit map and
6280 // the work_queue is used for keeping the (newly) grey objects during
6281 // the scan phase whence they are also available for stealing by parallel
6282 // threads. Since the marking bit map is shared, updates are
6283 // synchronized (via CAS).
6284 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6285   if (obj != NULL) {
6286     // Ignore mark word because this could be an already marked oop
6287     // that may be chained at the end of the overflow list.
6288     assert(obj->is_oop(true), "expected an oop");
6289     HeapWord* addr = (HeapWord*)obj;
6290     if (_span.contains(addr) &&
6291         !_bit_map->isMarked(addr)) {
6292       // mark bit map (object will become grey):
6293       // It is possible for several threads to be
6294       // trying to "claim" this object concurrently;
6295       // the unique thread that succeeds in marking the
6296       // object first will do the subsequent push on
6297       // to the work queue (or overflow list).
6298       if (_bit_map->par_mark(addr)) {
6299         // push on work_queue (which may not be empty), and trim the
6300         // queue to an appropriate length by applying this closure to
6301         // the oops in the oops popped from the stack (i.e. blacken the
6302         // grey objects)
6303         bool res = _work_queue->push(obj);
6304         assert(res, "Low water mark should be less than capacity?");
6305         trim_queue(_low_water_mark);
6306       } // Else, another thread claimed the object
6307     }
6308   }
6309 }
6310 
6311 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p)       { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6312 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6313 
6314 // This closure is used to rescan the marked objects on the dirty cards
6315 // in the mod union table and the card table proper.
6316 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6317   oop p, MemRegion mr) {
6318 
6319   size_t size = 0;
6320   HeapWord* addr = (HeapWord*)p;
6321   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6322   assert(_span.contains(addr), "we are scanning the CMS generation");
6323   // check if it's time to yield
6324   if (do_yield_check()) {
6325     // We yielded for some foreground stop-world work,
6326     // and we have been asked to abort this ongoing preclean cycle.
6327     return 0;
6328   }
6329   if (_bitMap->isMarked(addr)) {
6330     // it's marked; is it potentially uninitialized?
6331     if (p->klass_or_null() != NULL) {
6332         // an initialized object; ignore mark word in verification below
6333         // since we are running concurrent with mutators
6334         assert(p->is_oop(true), "should be an oop");
6335         if (p->is_objArray()) {
6336           // objArrays are precisely marked; restrict scanning
6337           // to dirty cards only.
6338           size = CompactibleFreeListSpace::adjustObjectSize(
6339                    p->oop_iterate(_scanningClosure, mr));
6340         } else {
6341           // A non-array may have been imprecisely marked; we need
6342           // to scan object in its entirety.
6343           size = CompactibleFreeListSpace::adjustObjectSize(
6344                    p->oop_iterate(_scanningClosure));
6345         }
6346         #ifdef ASSERT
6347           size_t direct_size =
6348             CompactibleFreeListSpace::adjustObjectSize(p->size());
6349           assert(size == direct_size, "Inconsistency in size");
6350           assert(size >= 3, "Necessary for Printezis marks to work");
6351           if (!_bitMap->isMarked(addr+1)) {
6352             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6353           } else {
6354             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6355             assert(_bitMap->isMarked(addr+size-1),
6356                    "inconsistent Printezis mark");
6357           }
6358         #endif // ASSERT
6359     } else {
6360       // An uninitialized object.
6361       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6362       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6363       size = pointer_delta(nextOneAddr + 1, addr);
6364       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6365              "alignment problem");
6366       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6367       // will dirty the card when the klass pointer is installed in the
6368       // object (signaling the completion of initialization).
6369     }
6370   } else {
6371     // Either a not yet marked object or an uninitialized object
6372     if (p->klass_or_null() == NULL) {
6373       // An uninitialized object, skip to the next card, since
6374       // we may not be able to read its P-bits yet.
6375       assert(size == 0, "Initial value");
6376     } else {
6377       // An object not (yet) reached by marking: we merely need to
6378       // compute its size so as to go look at the next block.
6379       assert(p->is_oop(true), "should be an oop");
6380       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6381     }
6382   }
6383   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6384   return size;
6385 }
6386 
6387 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6388   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6389          "CMS thread should hold CMS token");
6390   assert_lock_strong(_freelistLock);
6391   assert_lock_strong(_bitMap->lock());
6392   // relinquish the free_list_lock and bitMaplock()
6393   _bitMap->lock()->unlock();
6394   _freelistLock->unlock();
6395   ConcurrentMarkSweepThread::desynchronize(true);
6396   _collector->stopTimer();
6397   if (PrintCMSStatistics != 0) {
6398     _collector->incrementYields();
6399   }
6400 
6401   // See the comment in coordinator_yield()
6402   for (unsigned i = 0; i < CMSYieldSleepCount &&
6403                    ConcurrentMarkSweepThread::should_yield() &&
6404                    !CMSCollector::foregroundGCIsActive(); ++i) {
6405     os::sleep(Thread::current(), 1, false);
6406   }
6407 
6408   ConcurrentMarkSweepThread::synchronize(true);
6409   _freelistLock->lock_without_safepoint_check();
6410   _bitMap->lock()->lock_without_safepoint_check();
6411   _collector->startTimer();
6412 }
6413 
6414 
6415 //////////////////////////////////////////////////////////////////
6416 // SurvivorSpacePrecleanClosure
6417 //////////////////////////////////////////////////////////////////
6418 // This (single-threaded) closure is used to preclean the oops in
6419 // the survivor spaces.
6420 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6421 
6422   HeapWord* addr = (HeapWord*)p;
6423   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6424   assert(!_span.contains(addr), "we are scanning the survivor spaces");
6425   assert(p->klass_or_null() != NULL, "object should be initialized");
6426   // an initialized object; ignore mark word in verification below
6427   // since we are running concurrent with mutators
6428   assert(p->is_oop(true), "should be an oop");
6429   // Note that we do not yield while we iterate over
6430   // the interior oops of p, pushing the relevant ones
6431   // on our marking stack.
6432   size_t size = p->oop_iterate(_scanning_closure);
6433   do_yield_check();
6434   // Observe that below, we do not abandon the preclean
6435   // phase as soon as we should; rather we empty the
6436   // marking stack before returning. This is to satisfy
6437   // some existing assertions. In general, it may be a
6438   // good idea to abort immediately and complete the marking
6439   // from the grey objects at a later time.
6440   while (!_mark_stack->isEmpty()) {
6441     oop new_oop = _mark_stack->pop();
6442     assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6443     assert(_bit_map->isMarked((HeapWord*)new_oop),
6444            "only grey objects on this stack");
6445     // iterate over the oops in this oop, marking and pushing
6446     // the ones in CMS heap (i.e. in _span).
6447     new_oop->oop_iterate(_scanning_closure);
6448     // check if it's time to yield
6449     do_yield_check();
6450   }
6451   unsigned int after_count =
6452     GenCollectedHeap::heap()->total_collections();
6453   bool abort = (_before_count != after_count) ||
6454                _collector->should_abort_preclean();
6455   return abort ? 0 : size;
6456 }
6457 
6458 void SurvivorSpacePrecleanClosure::do_yield_work() {
6459   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6460          "CMS thread should hold CMS token");
6461   assert_lock_strong(_bit_map->lock());
6462   // Relinquish the bit map lock
6463   _bit_map->lock()->unlock();
6464   ConcurrentMarkSweepThread::desynchronize(true);
6465   _collector->stopTimer();
6466   if (PrintCMSStatistics != 0) {
6467     _collector->incrementYields();
6468   }
6469 
6470   // See the comment in coordinator_yield()
6471   for (unsigned i = 0; i < CMSYieldSleepCount &&
6472                        ConcurrentMarkSweepThread::should_yield() &&
6473                        !CMSCollector::foregroundGCIsActive(); ++i) {
6474     os::sleep(Thread::current(), 1, false);
6475   }
6476 
6477   ConcurrentMarkSweepThread::synchronize(true);
6478   _bit_map->lock()->lock_without_safepoint_check();
6479   _collector->startTimer();
6480 }
6481 
6482 // This closure is used to rescan the marked objects on the dirty cards
6483 // in the mod union table and the card table proper. In the parallel
6484 // case, although the bitMap is shared, we do a single read so the
6485 // isMarked() query is "safe".
6486 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6487   // Ignore mark word because we are running concurrent with mutators
6488   assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6489   HeapWord* addr = (HeapWord*)p;
6490   assert(_span.contains(addr), "we are scanning the CMS generation");
6491   bool is_obj_array = false;
6492   #ifdef ASSERT
6493     if (!_parallel) {
6494       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6495       assert(_collector->overflow_list_is_empty(),
6496              "overflow list should be empty");
6497 
6498     }
6499   #endif // ASSERT
6500   if (_bit_map->isMarked(addr)) {
6501     // Obj arrays are precisely marked, non-arrays are not;
6502     // so we scan objArrays precisely and non-arrays in their
6503     // entirety.
6504     if (p->is_objArray()) {
6505       is_obj_array = true;
6506       if (_parallel) {
6507         p->oop_iterate(_par_scan_closure, mr);
6508       } else {
6509         p->oop_iterate(_scan_closure, mr);
6510       }
6511     } else {
6512       if (_parallel) {
6513         p->oop_iterate(_par_scan_closure);
6514       } else {
6515         p->oop_iterate(_scan_closure);
6516       }
6517     }
6518   }
6519   #ifdef ASSERT
6520     if (!_parallel) {
6521       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6522       assert(_collector->overflow_list_is_empty(),
6523              "overflow list should be empty");
6524 
6525     }
6526   #endif // ASSERT
6527   return is_obj_array;
6528 }
6529 
6530 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6531                         MemRegion span,
6532                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
6533                         bool should_yield, bool verifying):
6534   _collector(collector),
6535   _span(span),
6536   _bitMap(bitMap),
6537   _mut(&collector->_modUnionTable),
6538   _markStack(markStack),
6539   _yield(should_yield),
6540   _skipBits(0)
6541 {
6542   assert(_markStack->isEmpty(), "stack should be empty");
6543   _finger = _bitMap->startWord();
6544   _threshold = _finger;
6545   assert(_collector->_restart_addr == NULL, "Sanity check");
6546   assert(_span.contains(_finger), "Out of bounds _finger?");
6547   DEBUG_ONLY(_verifying = verifying;)
6548 }
6549 
6550 void MarkFromRootsClosure::reset(HeapWord* addr) {
6551   assert(_markStack->isEmpty(), "would cause duplicates on stack");
6552   assert(_span.contains(addr), "Out of bounds _finger?");
6553   _finger = addr;
6554   _threshold = (HeapWord*)round_to(
6555                  (intptr_t)_finger, CardTableModRefBS::card_size);
6556 }
6557 
6558 // Should revisit to see if this should be restructured for
6559 // greater efficiency.
6560 bool MarkFromRootsClosure::do_bit(size_t offset) {
6561   if (_skipBits > 0) {
6562     _skipBits--;
6563     return true;
6564   }
6565   // convert offset into a HeapWord*
6566   HeapWord* addr = _bitMap->startWord() + offset;
6567   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6568          "address out of range");
6569   assert(_bitMap->isMarked(addr), "tautology");
6570   if (_bitMap->isMarked(addr+1)) {
6571     // this is an allocated but not yet initialized object
6572     assert(_skipBits == 0, "tautology");
6573     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
6574     oop p = oop(addr);
6575     if (p->klass_or_null() == NULL) {
6576       DEBUG_ONLY(if (!_verifying) {)
6577         // We re-dirty the cards on which this object lies and increase
6578         // the _threshold so that we'll come back to scan this object
6579         // during the preclean or remark phase. (CMSCleanOnEnter)
6580         if (CMSCleanOnEnter) {
6581           size_t sz = _collector->block_size_using_printezis_bits(addr);
6582           HeapWord* end_card_addr   = (HeapWord*)round_to(
6583                                          (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6584           MemRegion redirty_range = MemRegion(addr, end_card_addr);
6585           assert(!redirty_range.is_empty(), "Arithmetical tautology");
6586           // Bump _threshold to end_card_addr; note that
6587           // _threshold cannot possibly exceed end_card_addr, anyhow.
6588           // This prevents future clearing of the card as the scan proceeds
6589           // to the right.
6590           assert(_threshold <= end_card_addr,
6591                  "Because we are just scanning into this object");
6592           if (_threshold < end_card_addr) {
6593             _threshold = end_card_addr;
6594           }
6595           if (p->klass_or_null() != NULL) {
6596             // Redirty the range of cards...
6597             _mut->mark_range(redirty_range);
6598           } // ...else the setting of klass will dirty the card anyway.
6599         }
6600       DEBUG_ONLY(})
6601       return true;
6602     }
6603   }
6604   scanOopsInOop(addr);
6605   return true;
6606 }
6607 
6608 // We take a break if we've been at this for a while,
6609 // so as to avoid monopolizing the locks involved.
6610 void MarkFromRootsClosure::do_yield_work() {
6611   // First give up the locks, then yield, then re-lock
6612   // We should probably use a constructor/destructor idiom to
6613   // do this unlock/lock or modify the MutexUnlocker class to
6614   // serve our purpose. XXX
6615   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6616          "CMS thread should hold CMS token");
6617   assert_lock_strong(_bitMap->lock());
6618   _bitMap->lock()->unlock();
6619   ConcurrentMarkSweepThread::desynchronize(true);
6620   _collector->stopTimer();
6621   if (PrintCMSStatistics != 0) {
6622     _collector->incrementYields();
6623   }
6624 
6625   // See the comment in coordinator_yield()
6626   for (unsigned i = 0; i < CMSYieldSleepCount &&
6627                        ConcurrentMarkSweepThread::should_yield() &&
6628                        !CMSCollector::foregroundGCIsActive(); ++i) {
6629     os::sleep(Thread::current(), 1, false);
6630   }
6631 
6632   ConcurrentMarkSweepThread::synchronize(true);
6633   _bitMap->lock()->lock_without_safepoint_check();
6634   _collector->startTimer();
6635 }
6636 
6637 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6638   assert(_bitMap->isMarked(ptr), "expected bit to be set");
6639   assert(_markStack->isEmpty(),
6640          "should drain stack to limit stack usage");
6641   // convert ptr to an oop preparatory to scanning
6642   oop obj = oop(ptr);
6643   // Ignore mark word in verification below, since we
6644   // may be running concurrent with mutators.
6645   assert(obj->is_oop(true), "should be an oop");
6646   assert(_finger <= ptr, "_finger runneth ahead");
6647   // advance the finger to right end of this object
6648   _finger = ptr + obj->size();
6649   assert(_finger > ptr, "we just incremented it above");
6650   // On large heaps, it may take us some time to get through
6651   // the marking phase. During
6652   // this time it's possible that a lot of mutations have
6653   // accumulated in the card table and the mod union table --
6654   // these mutation records are redundant until we have
6655   // actually traced into the corresponding card.
6656   // Here, we check whether advancing the finger would make
6657   // us cross into a new card, and if so clear corresponding
6658   // cards in the MUT (preclean them in the card-table in the
6659   // future).
6660 
6661   DEBUG_ONLY(if (!_verifying) {)
6662     // The clean-on-enter optimization is disabled by default,
6663     // until we fix 6178663.
6664     if (CMSCleanOnEnter && (_finger > _threshold)) {
6665       // [_threshold, _finger) represents the interval
6666       // of cards to be cleared  in MUT (or precleaned in card table).
6667       // The set of cards to be cleared is all those that overlap
6668       // with the interval [_threshold, _finger); note that
6669       // _threshold is always kept card-aligned but _finger isn't
6670       // always card-aligned.
6671       HeapWord* old_threshold = _threshold;
6672       assert(old_threshold == (HeapWord*)round_to(
6673               (intptr_t)old_threshold, CardTableModRefBS::card_size),
6674              "_threshold should always be card-aligned");
6675       _threshold = (HeapWord*)round_to(
6676                      (intptr_t)_finger, CardTableModRefBS::card_size);
6677       MemRegion mr(old_threshold, _threshold);
6678       assert(!mr.is_empty(), "Control point invariant");
6679       assert(_span.contains(mr), "Should clear within span");
6680       _mut->clear_range(mr);
6681     }
6682   DEBUG_ONLY(})
6683   // Note: the finger doesn't advance while we drain
6684   // the stack below.
6685   PushOrMarkClosure pushOrMarkClosure(_collector,
6686                                       _span, _bitMap, _markStack,
6687                                       _finger, this);
6688   bool res = _markStack->push(obj);
6689   assert(res, "Empty non-zero size stack should have space for single push");
6690   while (!_markStack->isEmpty()) {
6691     oop new_oop = _markStack->pop();
6692     // Skip verifying header mark word below because we are
6693     // running concurrent with mutators.
6694     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6695     // now scan this oop's oops
6696     new_oop->oop_iterate(&pushOrMarkClosure);
6697     do_yield_check();
6698   }
6699   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6700 }
6701 
6702 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6703                        CMSCollector* collector, MemRegion span,
6704                        CMSBitMap* bit_map,
6705                        OopTaskQueue* work_queue,
6706                        CMSMarkStack*  overflow_stack):
6707   _collector(collector),
6708   _whole_span(collector->_span),
6709   _span(span),
6710   _bit_map(bit_map),
6711   _mut(&collector->_modUnionTable),
6712   _work_queue(work_queue),
6713   _overflow_stack(overflow_stack),
6714   _skip_bits(0),
6715   _task(task)
6716 {
6717   assert(_work_queue->size() == 0, "work_queue should be empty");
6718   _finger = span.start();
6719   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
6720   assert(_span.contains(_finger), "Out of bounds _finger?");
6721 }
6722 
6723 // Should revisit to see if this should be restructured for
6724 // greater efficiency.
6725 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6726   if (_skip_bits > 0) {
6727     _skip_bits--;
6728     return true;
6729   }
6730   // convert offset into a HeapWord*
6731   HeapWord* addr = _bit_map->startWord() + offset;
6732   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6733          "address out of range");
6734   assert(_bit_map->isMarked(addr), "tautology");
6735   if (_bit_map->isMarked(addr+1)) {
6736     // this is an allocated object that might not yet be initialized
6737     assert(_skip_bits == 0, "tautology");
6738     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
6739     oop p = oop(addr);
6740     if (p->klass_or_null() == NULL) {
6741       // in the case of Clean-on-Enter optimization, redirty card
6742       // and avoid clearing card by increasing  the threshold.
6743       return true;
6744     }
6745   }
6746   scan_oops_in_oop(addr);
6747   return true;
6748 }
6749 
6750 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6751   assert(_bit_map->isMarked(ptr), "expected bit to be set");
6752   // Should we assert that our work queue is empty or
6753   // below some drain limit?
6754   assert(_work_queue->size() == 0,
6755          "should drain stack to limit stack usage");
6756   // convert ptr to an oop preparatory to scanning
6757   oop obj = oop(ptr);
6758   // Ignore mark word in verification below, since we
6759   // may be running concurrent with mutators.
6760   assert(obj->is_oop(true), "should be an oop");
6761   assert(_finger <= ptr, "_finger runneth ahead");
6762   // advance the finger to right end of this object
6763   _finger = ptr + obj->size();
6764   assert(_finger > ptr, "we just incremented it above");
6765   // On large heaps, it may take us some time to get through
6766   // the marking phase. During
6767   // this time it's possible that a lot of mutations have
6768   // accumulated in the card table and the mod union table --
6769   // these mutation records are redundant until we have
6770   // actually traced into the corresponding card.
6771   // Here, we check whether advancing the finger would make
6772   // us cross into a new card, and if so clear corresponding
6773   // cards in the MUT (preclean them in the card-table in the
6774   // future).
6775 
6776   // The clean-on-enter optimization is disabled by default,
6777   // until we fix 6178663.
6778   if (CMSCleanOnEnter && (_finger > _threshold)) {
6779     // [_threshold, _finger) represents the interval
6780     // of cards to be cleared  in MUT (or precleaned in card table).
6781     // The set of cards to be cleared is all those that overlap
6782     // with the interval [_threshold, _finger); note that
6783     // _threshold is always kept card-aligned but _finger isn't
6784     // always card-aligned.
6785     HeapWord* old_threshold = _threshold;
6786     assert(old_threshold == (HeapWord*)round_to(
6787             (intptr_t)old_threshold, CardTableModRefBS::card_size),
6788            "_threshold should always be card-aligned");
6789     _threshold = (HeapWord*)round_to(
6790                    (intptr_t)_finger, CardTableModRefBS::card_size);
6791     MemRegion mr(old_threshold, _threshold);
6792     assert(!mr.is_empty(), "Control point invariant");
6793     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6794     _mut->clear_range(mr);
6795   }
6796 
6797   // Note: the local finger doesn't advance while we drain
6798   // the stack below, but the global finger sure can and will.
6799   HeapWord** gfa = _task->global_finger_addr();
6800   Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6801                                       _span, _bit_map,
6802                                       _work_queue,
6803                                       _overflow_stack,
6804                                       _finger,
6805                                       gfa, this);
6806   bool res = _work_queue->push(obj);   // overflow could occur here
6807   assert(res, "Will hold once we use workqueues");
6808   while (true) {
6809     oop new_oop;
6810     if (!_work_queue->pop_local(new_oop)) {
6811       // We emptied our work_queue; check if there's stuff that can
6812       // be gotten from the overflow stack.
6813       if (CMSConcMarkingTask::get_work_from_overflow_stack(
6814             _overflow_stack, _work_queue)) {
6815         do_yield_check();
6816         continue;
6817       } else {  // done
6818         break;
6819       }
6820     }
6821     // Skip verifying header mark word below because we are
6822     // running concurrent with mutators.
6823     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6824     // now scan this oop's oops
6825     new_oop->oop_iterate(&pushOrMarkClosure);
6826     do_yield_check();
6827   }
6828   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6829 }
6830 
6831 // Yield in response to a request from VM Thread or
6832 // from mutators.
6833 void Par_MarkFromRootsClosure::do_yield_work() {
6834   assert(_task != NULL, "sanity");
6835   _task->yield();
6836 }
6837 
6838 // A variant of the above used for verifying CMS marking work.
6839 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6840                         MemRegion span,
6841                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6842                         CMSMarkStack*  mark_stack):
6843   _collector(collector),
6844   _span(span),
6845   _verification_bm(verification_bm),
6846   _cms_bm(cms_bm),
6847   _mark_stack(mark_stack),
6848   _pam_verify_closure(collector, span, verification_bm, cms_bm,
6849                       mark_stack)
6850 {
6851   assert(_mark_stack->isEmpty(), "stack should be empty");
6852   _finger = _verification_bm->startWord();
6853   assert(_collector->_restart_addr == NULL, "Sanity check");
6854   assert(_span.contains(_finger), "Out of bounds _finger?");
6855 }
6856 
6857 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6858   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6859   assert(_span.contains(addr), "Out of bounds _finger?");
6860   _finger = addr;
6861 }
6862 
6863 // Should revisit to see if this should be restructured for
6864 // greater efficiency.
6865 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6866   // convert offset into a HeapWord*
6867   HeapWord* addr = _verification_bm->startWord() + offset;
6868   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6869          "address out of range");
6870   assert(_verification_bm->isMarked(addr), "tautology");
6871   assert(_cms_bm->isMarked(addr), "tautology");
6872 
6873   assert(_mark_stack->isEmpty(),
6874          "should drain stack to limit stack usage");
6875   // convert addr to an oop preparatory to scanning
6876   oop obj = oop(addr);
6877   assert(obj->is_oop(), "should be an oop");
6878   assert(_finger <= addr, "_finger runneth ahead");
6879   // advance the finger to right end of this object
6880   _finger = addr + obj->size();
6881   assert(_finger > addr, "we just incremented it above");
6882   // Note: the finger doesn't advance while we drain
6883   // the stack below.
6884   bool res = _mark_stack->push(obj);
6885   assert(res, "Empty non-zero size stack should have space for single push");
6886   while (!_mark_stack->isEmpty()) {
6887     oop new_oop = _mark_stack->pop();
6888     assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6889     // now scan this oop's oops
6890     new_oop->oop_iterate(&_pam_verify_closure);
6891   }
6892   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6893   return true;
6894 }
6895 
6896 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6897   CMSCollector* collector, MemRegion span,
6898   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6899   CMSMarkStack*  mark_stack):
6900   MetadataAwareOopClosure(collector->ref_processor()),
6901   _collector(collector),
6902   _span(span),
6903   _verification_bm(verification_bm),
6904   _cms_bm(cms_bm),
6905   _mark_stack(mark_stack)
6906 { }
6907 
6908 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
6909 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6910 
6911 // Upon stack overflow, we discard (part of) the stack,
6912 // remembering the least address amongst those discarded
6913 // in CMSCollector's _restart_address.
6914 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6915   // Remember the least grey address discarded
6916   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6917   _collector->lower_restart_addr(ra);
6918   _mark_stack->reset();  // discard stack contents
6919   _mark_stack->expand(); // expand the stack if possible
6920 }
6921 
6922 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6923   assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6924   HeapWord* addr = (HeapWord*)obj;
6925   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6926     // Oop lies in _span and isn't yet grey or black
6927     _verification_bm->mark(addr);            // now grey
6928     if (!_cms_bm->isMarked(addr)) {
6929       oop(addr)->print();
6930       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6931                              p2i(addr));
6932       fatal("... aborting");
6933     }
6934 
6935     if (!_mark_stack->push(obj)) { // stack overflow
6936       if (PrintCMSStatistics != 0) {
6937         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6938                                SIZE_FORMAT, _mark_stack->capacity());
6939       }
6940       assert(_mark_stack->isFull(), "Else push should have succeeded");
6941       handle_stack_overflow(addr);
6942     }
6943     // anything including and to the right of _finger
6944     // will be scanned as we iterate over the remainder of the
6945     // bit map
6946   }
6947 }
6948 
6949 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6950                      MemRegion span,
6951                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
6952                      HeapWord* finger, MarkFromRootsClosure* parent) :
6953   MetadataAwareOopClosure(collector->ref_processor()),
6954   _collector(collector),
6955   _span(span),
6956   _bitMap(bitMap),
6957   _markStack(markStack),
6958   _finger(finger),
6959   _parent(parent)
6960 { }
6961 
6962 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6963                      MemRegion span,
6964                      CMSBitMap* bit_map,
6965                      OopTaskQueue* work_queue,
6966                      CMSMarkStack*  overflow_stack,
6967                      HeapWord* finger,
6968                      HeapWord** global_finger_addr,
6969                      Par_MarkFromRootsClosure* parent) :
6970   MetadataAwareOopClosure(collector->ref_processor()),
6971   _collector(collector),
6972   _whole_span(collector->_span),
6973   _span(span),
6974   _bit_map(bit_map),
6975   _work_queue(work_queue),
6976   _overflow_stack(overflow_stack),
6977   _finger(finger),
6978   _global_finger_addr(global_finger_addr),
6979   _parent(parent)
6980 { }
6981 
6982 // Assumes thread-safe access by callers, who are
6983 // responsible for mutual exclusion.
6984 void CMSCollector::lower_restart_addr(HeapWord* low) {
6985   assert(_span.contains(low), "Out of bounds addr");
6986   if (_restart_addr == NULL) {
6987     _restart_addr = low;
6988   } else {
6989     _restart_addr = MIN2(_restart_addr, low);
6990   }
6991 }
6992 
6993 // Upon stack overflow, we discard (part of) the stack,
6994 // remembering the least address amongst those discarded
6995 // in CMSCollector's _restart_address.
6996 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6997   // Remember the least grey address discarded
6998   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6999   _collector->lower_restart_addr(ra);
7000   _markStack->reset();  // discard stack contents
7001   _markStack->expand(); // expand the stack if possible
7002 }
7003 
7004 // Upon stack overflow, we discard (part of) the stack,
7005 // remembering the least address amongst those discarded
7006 // in CMSCollector's _restart_address.
7007 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7008   // We need to do this under a mutex to prevent other
7009   // workers from interfering with the work done below.
7010   MutexLockerEx ml(_overflow_stack->par_lock(),
7011                    Mutex::_no_safepoint_check_flag);
7012   // Remember the least grey address discarded
7013   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7014   _collector->lower_restart_addr(ra);
7015   _overflow_stack->reset();  // discard stack contents
7016   _overflow_stack->expand(); // expand the stack if possible
7017 }
7018 
7019 void PushOrMarkClosure::do_oop(oop obj) {
7020   // Ignore mark word because we are running concurrent with mutators.
7021   assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7022   HeapWord* addr = (HeapWord*)obj;
7023   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7024     // Oop lies in _span and isn't yet grey or black
7025     _bitMap->mark(addr);            // now grey
7026     if (addr < _finger) {
7027       // the bit map iteration has already either passed, or
7028       // sampled, this bit in the bit map; we'll need to
7029       // use the marking stack to scan this oop's oops.
7030       bool simulate_overflow = false;
7031       NOT_PRODUCT(
7032         if (CMSMarkStackOverflowALot &&
7033             _collector->simulate_overflow()) {
7034           // simulate a stack overflow
7035           simulate_overflow = true;
7036         }
7037       )
7038       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7039         if (PrintCMSStatistics != 0) {
7040           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7041                                  SIZE_FORMAT, _markStack->capacity());
7042         }
7043         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7044         handle_stack_overflow(addr);
7045       }
7046     }
7047     // anything including and to the right of _finger
7048     // will be scanned as we iterate over the remainder of the
7049     // bit map
7050     do_yield_check();
7051   }
7052 }
7053 
7054 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
7055 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7056 
7057 void Par_PushOrMarkClosure::do_oop(oop obj) {
7058   // Ignore mark word because we are running concurrent with mutators.
7059   assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7060   HeapWord* addr = (HeapWord*)obj;
7061   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7062     // Oop lies in _span and isn't yet grey or black
7063     // We read the global_finger (volatile read) strictly after marking oop
7064     bool res = _bit_map->par_mark(addr);    // now grey
7065     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7066     // Should we push this marked oop on our stack?
7067     // -- if someone else marked it, nothing to do
7068     // -- if target oop is above global finger nothing to do
7069     // -- if target oop is in chunk and above local finger
7070     //      then nothing to do
7071     // -- else push on work queue
7072     if (   !res       // someone else marked it, they will deal with it
7073         || (addr >= *gfa)  // will be scanned in a later task
7074         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7075       return;
7076     }
7077     // the bit map iteration has already either passed, or
7078     // sampled, this bit in the bit map; we'll need to
7079     // use the marking stack to scan this oop's oops.
7080     bool simulate_overflow = false;
7081     NOT_PRODUCT(
7082       if (CMSMarkStackOverflowALot &&
7083           _collector->simulate_overflow()) {
7084         // simulate a stack overflow
7085         simulate_overflow = true;
7086       }
7087     )
7088     if (simulate_overflow ||
7089         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7090       // stack overflow
7091       if (PrintCMSStatistics != 0) {
7092         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7093                                SIZE_FORMAT, _overflow_stack->capacity());
7094       }
7095       // We cannot assert that the overflow stack is full because
7096       // it may have been emptied since.
7097       assert(simulate_overflow ||
7098              _work_queue->size() == _work_queue->max_elems(),
7099             "Else push should have succeeded");
7100       handle_stack_overflow(addr);
7101     }
7102     do_yield_check();
7103   }
7104 }
7105 
7106 void Par_PushOrMarkClosure::do_oop(oop* p)       { Par_PushOrMarkClosure::do_oop_work(p); }
7107 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7108 
7109 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7110                                        MemRegion span,
7111                                        ReferenceProcessor* rp,
7112                                        CMSBitMap* bit_map,
7113                                        CMSBitMap* mod_union_table,
7114                                        CMSMarkStack*  mark_stack,
7115                                        bool           concurrent_precleaning):
7116   MetadataAwareOopClosure(rp),
7117   _collector(collector),
7118   _span(span),
7119   _bit_map(bit_map),
7120   _mod_union_table(mod_union_table),
7121   _mark_stack(mark_stack),
7122   _concurrent_precleaning(concurrent_precleaning)
7123 {
7124   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7125 }
7126 
7127 // Grey object rescan during pre-cleaning and second checkpoint phases --
7128 // the non-parallel version (the parallel version appears further below.)
7129 void PushAndMarkClosure::do_oop(oop obj) {
7130   // Ignore mark word verification. If during concurrent precleaning,
7131   // the object monitor may be locked. If during the checkpoint
7132   // phases, the object may already have been reached by a  different
7133   // path and may be at the end of the global overflow list (so
7134   // the mark word may be NULL).
7135   assert(obj->is_oop_or_null(true /* ignore mark word */),
7136          err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7137   HeapWord* addr = (HeapWord*)obj;
7138   // Check if oop points into the CMS generation
7139   // and is not marked
7140   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7141     // a white object ...
7142     _bit_map->mark(addr);         // ... now grey
7143     // push on the marking stack (grey set)
7144     bool simulate_overflow = false;
7145     NOT_PRODUCT(
7146       if (CMSMarkStackOverflowALot &&
7147           _collector->simulate_overflow()) {
7148         // simulate a stack overflow
7149         simulate_overflow = true;
7150       }
7151     )
7152     if (simulate_overflow || !_mark_stack->push(obj)) {
7153       if (_concurrent_precleaning) {
7154          // During precleaning we can just dirty the appropriate card(s)
7155          // in the mod union table, thus ensuring that the object remains
7156          // in the grey set  and continue. In the case of object arrays
7157          // we need to dirty all of the cards that the object spans,
7158          // since the rescan of object arrays will be limited to the
7159          // dirty cards.
7160          // Note that no one can be interfering with us in this action
7161          // of dirtying the mod union table, so no locking or atomics
7162          // are required.
7163          if (obj->is_objArray()) {
7164            size_t sz = obj->size();
7165            HeapWord* end_card_addr = (HeapWord*)round_to(
7166                                         (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7167            MemRegion redirty_range = MemRegion(addr, end_card_addr);
7168            assert(!redirty_range.is_empty(), "Arithmetical tautology");
7169            _mod_union_table->mark_range(redirty_range);
7170          } else {
7171            _mod_union_table->mark(addr);
7172          }
7173          _collector->_ser_pmc_preclean_ovflw++;
7174       } else {
7175          // During the remark phase, we need to remember this oop
7176          // in the overflow list.
7177          _collector->push_on_overflow_list(obj);
7178          _collector->_ser_pmc_remark_ovflw++;
7179       }
7180     }
7181   }
7182 }
7183 
7184 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7185                                                MemRegion span,
7186                                                ReferenceProcessor* rp,
7187                                                CMSBitMap* bit_map,
7188                                                OopTaskQueue* work_queue):
7189   MetadataAwareOopClosure(rp),
7190   _collector(collector),
7191   _span(span),
7192   _bit_map(bit_map),
7193   _work_queue(work_queue)
7194 {
7195   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7196 }
7197 
7198 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
7199 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7200 
7201 // Grey object rescan during second checkpoint phase --
7202 // the parallel version.
7203 void Par_PushAndMarkClosure::do_oop(oop obj) {
7204   // In the assert below, we ignore the mark word because
7205   // this oop may point to an already visited object that is
7206   // on the overflow stack (in which case the mark word has
7207   // been hijacked for chaining into the overflow stack --
7208   // if this is the last object in the overflow stack then
7209   // its mark word will be NULL). Because this object may
7210   // have been subsequently popped off the global overflow
7211   // stack, and the mark word possibly restored to the prototypical
7212   // value, by the time we get to examined this failing assert in
7213   // the debugger, is_oop_or_null(false) may subsequently start
7214   // to hold.
7215   assert(obj->is_oop_or_null(true),
7216          err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7217   HeapWord* addr = (HeapWord*)obj;
7218   // Check if oop points into the CMS generation
7219   // and is not marked
7220   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7221     // a white object ...
7222     // If we manage to "claim" the object, by being the
7223     // first thread to mark it, then we push it on our
7224     // marking stack
7225     if (_bit_map->par_mark(addr)) {     // ... now grey
7226       // push on work queue (grey set)
7227       bool simulate_overflow = false;
7228       NOT_PRODUCT(
7229         if (CMSMarkStackOverflowALot &&
7230             _collector->par_simulate_overflow()) {
7231           // simulate a stack overflow
7232           simulate_overflow = true;
7233         }
7234       )
7235       if (simulate_overflow || !_work_queue->push(obj)) {
7236         _collector->par_push_on_overflow_list(obj);
7237         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
7238       }
7239     } // Else, some other thread got there first
7240   }
7241 }
7242 
7243 void Par_PushAndMarkClosure::do_oop(oop* p)       { Par_PushAndMarkClosure::do_oop_work(p); }
7244 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7245 
7246 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7247   Mutex* bml = _collector->bitMapLock();
7248   assert_lock_strong(bml);
7249   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7250          "CMS thread should hold CMS token");
7251 
7252   bml->unlock();
7253   ConcurrentMarkSweepThread::desynchronize(true);
7254 
7255   _collector->stopTimer();
7256   if (PrintCMSStatistics != 0) {
7257     _collector->incrementYields();
7258   }
7259 
7260   // See the comment in coordinator_yield()
7261   for (unsigned i = 0; i < CMSYieldSleepCount &&
7262                        ConcurrentMarkSweepThread::should_yield() &&
7263                        !CMSCollector::foregroundGCIsActive(); ++i) {
7264     os::sleep(Thread::current(), 1, false);
7265   }
7266 
7267   ConcurrentMarkSweepThread::synchronize(true);
7268   bml->lock();
7269 
7270   _collector->startTimer();
7271 }
7272 
7273 bool CMSPrecleanRefsYieldClosure::should_return() {
7274   if (ConcurrentMarkSweepThread::should_yield()) {
7275     do_yield_work();
7276   }
7277   return _collector->foregroundGCIsActive();
7278 }
7279 
7280 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7281   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7282          "mr should be aligned to start at a card boundary");
7283   // We'd like to assert:
7284   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7285   //        "mr should be a range of cards");
7286   // However, that would be too strong in one case -- the last
7287   // partition ends at _unallocated_block which, in general, can be
7288   // an arbitrary boundary, not necessarily card aligned.
7289   if (PrintCMSStatistics != 0) {
7290     _num_dirty_cards +=
7291          mr.word_size()/CardTableModRefBS::card_size_in_words;
7292   }
7293   _space->object_iterate_mem(mr, &_scan_cl);
7294 }
7295 
7296 SweepClosure::SweepClosure(CMSCollector* collector,
7297                            ConcurrentMarkSweepGeneration* g,
7298                            CMSBitMap* bitMap, bool should_yield) :
7299   _collector(collector),
7300   _g(g),
7301   _sp(g->cmsSpace()),
7302   _limit(_sp->sweep_limit()),
7303   _freelistLock(_sp->freelistLock()),
7304   _bitMap(bitMap),
7305   _yield(should_yield),
7306   _inFreeRange(false),           // No free range at beginning of sweep
7307   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
7308   _lastFreeRangeCoalesced(false),
7309   _freeFinger(g->used_region().start())
7310 {
7311   NOT_PRODUCT(
7312     _numObjectsFreed = 0;
7313     _numWordsFreed   = 0;
7314     _numObjectsLive = 0;
7315     _numWordsLive = 0;
7316     _numObjectsAlreadyFree = 0;
7317     _numWordsAlreadyFree = 0;
7318     _last_fc = NULL;
7319 
7320     _sp->initializeIndexedFreeListArrayReturnedBytes();
7321     _sp->dictionary()->initialize_dict_returned_bytes();
7322   )
7323   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7324          "sweep _limit out of bounds");
7325   if (CMSTraceSweeper) {
7326     gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7327                         p2i(_limit));
7328   }
7329 }
7330 
7331 void SweepClosure::print_on(outputStream* st) const {
7332   tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7333                 p2i(_sp->bottom()), p2i(_sp->end()));
7334   tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7335   tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7336   NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7337   tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7338                 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7339 }
7340 
7341 #ifndef PRODUCT
7342 // Assertion checking only:  no useful work in product mode --
7343 // however, if any of the flags below become product flags,
7344 // you may need to review this code to see if it needs to be
7345 // enabled in product mode.
7346 SweepClosure::~SweepClosure() {
7347   assert_lock_strong(_freelistLock);
7348   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7349          "sweep _limit out of bounds");
7350   if (inFreeRange()) {
7351     warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7352     print();
7353     ShouldNotReachHere();
7354   }
7355   if (Verbose && PrintGC) {
7356     gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7357                         _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7358     gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects,  "
7359                            SIZE_FORMAT " bytes  "
7360       "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7361       _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7362       _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7363     size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7364                         * sizeof(HeapWord);
7365     gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7366 
7367     if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7368       size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7369       size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7370       size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7371       gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes);
7372       gclog_or_tty->print("   Indexed List Returned " SIZE_FORMAT " bytes",
7373         indexListReturnedBytes);
7374       gclog_or_tty->print_cr("        Dictionary Returned " SIZE_FORMAT " bytes",
7375         dict_returned_bytes);
7376     }
7377   }
7378   if (CMSTraceSweeper) {
7379     gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7380                            p2i(_limit));
7381   }
7382 }
7383 #endif  // PRODUCT
7384 
7385 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7386     bool freeRangeInFreeLists) {
7387   if (CMSTraceSweeper) {
7388     gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7389                p2i(freeFinger), freeRangeInFreeLists);
7390   }
7391   assert(!inFreeRange(), "Trampling existing free range");
7392   set_inFreeRange(true);
7393   set_lastFreeRangeCoalesced(false);
7394 
7395   set_freeFinger(freeFinger);
7396   set_freeRangeInFreeLists(freeRangeInFreeLists);
7397   if (CMSTestInFreeList) {
7398     if (freeRangeInFreeLists) {
7399       FreeChunk* fc = (FreeChunk*) freeFinger;
7400       assert(fc->is_free(), "A chunk on the free list should be free.");
7401       assert(fc->size() > 0, "Free range should have a size");
7402       assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7403     }
7404   }
7405 }
7406 
7407 // Note that the sweeper runs concurrently with mutators. Thus,
7408 // it is possible for direct allocation in this generation to happen
7409 // in the middle of the sweep. Note that the sweeper also coalesces
7410 // contiguous free blocks. Thus, unless the sweeper and the allocator
7411 // synchronize appropriately freshly allocated blocks may get swept up.
7412 // This is accomplished by the sweeper locking the free lists while
7413 // it is sweeping. Thus blocks that are determined to be free are
7414 // indeed free. There is however one additional complication:
7415 // blocks that have been allocated since the final checkpoint and
7416 // mark, will not have been marked and so would be treated as
7417 // unreachable and swept up. To prevent this, the allocator marks
7418 // the bit map when allocating during the sweep phase. This leads,
7419 // however, to a further complication -- objects may have been allocated
7420 // but not yet initialized -- in the sense that the header isn't yet
7421 // installed. The sweeper can not then determine the size of the block
7422 // in order to skip over it. To deal with this case, we use a technique
7423 // (due to Printezis) to encode such uninitialized block sizes in the
7424 // bit map. Since the bit map uses a bit per every HeapWord, but the
7425 // CMS generation has a minimum object size of 3 HeapWords, it follows
7426 // that "normal marks" won't be adjacent in the bit map (there will
7427 // always be at least two 0 bits between successive 1 bits). We make use
7428 // of these "unused" bits to represent uninitialized blocks -- the bit
7429 // corresponding to the start of the uninitialized object and the next
7430 // bit are both set. Finally, a 1 bit marks the end of the object that
7431 // started with the two consecutive 1 bits to indicate its potentially
7432 // uninitialized state.
7433 
7434 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7435   FreeChunk* fc = (FreeChunk*)addr;
7436   size_t res;
7437 
7438   // Check if we are done sweeping. Below we check "addr >= _limit" rather
7439   // than "addr == _limit" because although _limit was a block boundary when
7440   // we started the sweep, it may no longer be one because heap expansion
7441   // may have caused us to coalesce the block ending at the address _limit
7442   // with a newly expanded chunk (this happens when _limit was set to the
7443   // previous _end of the space), so we may have stepped past _limit:
7444   // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7445   if (addr >= _limit) { // we have swept up to or past the limit: finish up
7446     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7447            "sweep _limit out of bounds");
7448     assert(addr < _sp->end(), "addr out of bounds");
7449     // Flush any free range we might be holding as a single
7450     // coalesced chunk to the appropriate free list.
7451     if (inFreeRange()) {
7452       assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7453              err_msg("freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())));
7454       flush_cur_free_chunk(freeFinger(),
7455                            pointer_delta(addr, freeFinger()));
7456       if (CMSTraceSweeper) {
7457         gclog_or_tty->print("Sweep: last chunk: ");
7458         gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") "
7459                    "[coalesced:%d]\n",
7460                    p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7461                    lastFreeRangeCoalesced() ? 1 : 0);
7462       }
7463     }
7464 
7465     // help the iterator loop finish
7466     return pointer_delta(_sp->end(), addr);
7467   }
7468 
7469   assert(addr < _limit, "sweep invariant");
7470   // check if we should yield
7471   do_yield_check(addr);
7472   if (fc->is_free()) {
7473     // Chunk that is already free
7474     res = fc->size();
7475     do_already_free_chunk(fc);
7476     debug_only(_sp->verifyFreeLists());
7477     // If we flush the chunk at hand in lookahead_and_flush()
7478     // and it's coalesced with a preceding chunk, then the
7479     // process of "mangling" the payload of the coalesced block
7480     // will cause erasure of the size information from the
7481     // (erstwhile) header of all the coalesced blocks but the
7482     // first, so the first disjunct in the assert will not hold
7483     // in that specific case (in which case the second disjunct
7484     // will hold).
7485     assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7486            "Otherwise the size info doesn't change at this step");
7487     NOT_PRODUCT(
7488       _numObjectsAlreadyFree++;
7489       _numWordsAlreadyFree += res;
7490     )
7491     NOT_PRODUCT(_last_fc = fc;)
7492   } else if (!_bitMap->isMarked(addr)) {
7493     // Chunk is fresh garbage
7494     res = do_garbage_chunk(fc);
7495     debug_only(_sp->verifyFreeLists());
7496     NOT_PRODUCT(
7497       _numObjectsFreed++;
7498       _numWordsFreed += res;
7499     )
7500   } else {
7501     // Chunk that is alive.
7502     res = do_live_chunk(fc);
7503     debug_only(_sp->verifyFreeLists());
7504     NOT_PRODUCT(
7505         _numObjectsLive++;
7506         _numWordsLive += res;
7507     )
7508   }
7509   return res;
7510 }
7511 
7512 // For the smart allocation, record following
7513 //  split deaths - a free chunk is removed from its free list because
7514 //      it is being split into two or more chunks.
7515 //  split birth - a free chunk is being added to its free list because
7516 //      a larger free chunk has been split and resulted in this free chunk.
7517 //  coal death - a free chunk is being removed from its free list because
7518 //      it is being coalesced into a large free chunk.
7519 //  coal birth - a free chunk is being added to its free list because
7520 //      it was created when two or more free chunks where coalesced into
7521 //      this free chunk.
7522 //
7523 // These statistics are used to determine the desired number of free
7524 // chunks of a given size.  The desired number is chosen to be relative
7525 // to the end of a CMS sweep.  The desired number at the end of a sweep
7526 // is the
7527 //      count-at-end-of-previous-sweep (an amount that was enough)
7528 //              - count-at-beginning-of-current-sweep  (the excess)
7529 //              + split-births  (gains in this size during interval)
7530 //              - split-deaths  (demands on this size during interval)
7531 // where the interval is from the end of one sweep to the end of the
7532 // next.
7533 //
7534 // When sweeping the sweeper maintains an accumulated chunk which is
7535 // the chunk that is made up of chunks that have been coalesced.  That
7536 // will be termed the left-hand chunk.  A new chunk of garbage that
7537 // is being considered for coalescing will be referred to as the
7538 // right-hand chunk.
7539 //
7540 // When making a decision on whether to coalesce a right-hand chunk with
7541 // the current left-hand chunk, the current count vs. the desired count
7542 // of the left-hand chunk is considered.  Also if the right-hand chunk
7543 // is near the large chunk at the end of the heap (see
7544 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7545 // left-hand chunk is coalesced.
7546 //
7547 // When making a decision about whether to split a chunk, the desired count
7548 // vs. the current count of the candidate to be split is also considered.
7549 // If the candidate is underpopulated (currently fewer chunks than desired)
7550 // a chunk of an overpopulated (currently more chunks than desired) size may
7551 // be chosen.  The "hint" associated with a free list, if non-null, points
7552 // to a free list which may be overpopulated.
7553 //
7554 
7555 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7556   const size_t size = fc->size();
7557   // Chunks that cannot be coalesced are not in the
7558   // free lists.
7559   if (CMSTestInFreeList && !fc->cantCoalesce()) {
7560     assert(_sp->verify_chunk_in_free_list(fc),
7561       "free chunk should be in free lists");
7562   }
7563   // a chunk that is already free, should not have been
7564   // marked in the bit map
7565   HeapWord* const addr = (HeapWord*) fc;
7566   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7567   // Verify that the bit map has no bits marked between
7568   // addr and purported end of this block.
7569   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7570 
7571   // Some chunks cannot be coalesced under any circumstances.
7572   // See the definition of cantCoalesce().
7573   if (!fc->cantCoalesce()) {
7574     // This chunk can potentially be coalesced.
7575     if (_sp->adaptive_freelists()) {
7576       // All the work is done in
7577       do_post_free_or_garbage_chunk(fc, size);
7578     } else {  // Not adaptive free lists
7579       // this is a free chunk that can potentially be coalesced by the sweeper;
7580       if (!inFreeRange()) {
7581         // if the next chunk is a free block that can't be coalesced
7582         // it doesn't make sense to remove this chunk from the free lists
7583         FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7584         assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7585         if ((HeapWord*)nextChunk < _sp->end() &&     // There is another free chunk to the right ...
7586             nextChunk->is_free()               &&     // ... which is free...
7587             nextChunk->cantCoalesce()) {             // ... but can't be coalesced
7588           // nothing to do
7589         } else {
7590           // Potentially the start of a new free range:
7591           // Don't eagerly remove it from the free lists.
7592           // No need to remove it if it will just be put
7593           // back again.  (Also from a pragmatic point of view
7594           // if it is a free block in a region that is beyond
7595           // any allocated blocks, an assertion will fail)
7596           // Remember the start of a free run.
7597           initialize_free_range(addr, true);
7598           // end - can coalesce with next chunk
7599         }
7600       } else {
7601         // the midst of a free range, we are coalescing
7602         print_free_block_coalesced(fc);
7603         if (CMSTraceSweeper) {
7604           gclog_or_tty->print("  -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7605         }
7606         // remove it from the free lists
7607         _sp->removeFreeChunkFromFreeLists(fc);
7608         set_lastFreeRangeCoalesced(true);
7609         // If the chunk is being coalesced and the current free range is
7610         // in the free lists, remove the current free range so that it
7611         // will be returned to the free lists in its entirety - all
7612         // the coalesced pieces included.
7613         if (freeRangeInFreeLists()) {
7614           FreeChunk* ffc = (FreeChunk*) freeFinger();
7615           assert(ffc->size() == pointer_delta(addr, freeFinger()),
7616             "Size of free range is inconsistent with chunk size.");
7617           if (CMSTestInFreeList) {
7618             assert(_sp->verify_chunk_in_free_list(ffc),
7619               "free range is not in free lists");
7620           }
7621           _sp->removeFreeChunkFromFreeLists(ffc);
7622           set_freeRangeInFreeLists(false);
7623         }
7624       }
7625     }
7626     // Note that if the chunk is not coalescable (the else arm
7627     // below), we unconditionally flush, without needing to do
7628     // a "lookahead," as we do below.
7629     if (inFreeRange()) lookahead_and_flush(fc, size);
7630   } else {
7631     // Code path common to both original and adaptive free lists.
7632 
7633     // cant coalesce with previous block; this should be treated
7634     // as the end of a free run if any
7635     if (inFreeRange()) {
7636       // we kicked some butt; time to pick up the garbage
7637       assert(freeFinger() < addr, "freeFinger points too high");
7638       flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7639     }
7640     // else, nothing to do, just continue
7641   }
7642 }
7643 
7644 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7645   // This is a chunk of garbage.  It is not in any free list.
7646   // Add it to a free list or let it possibly be coalesced into
7647   // a larger chunk.
7648   HeapWord* const addr = (HeapWord*) fc;
7649   const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7650 
7651   if (_sp->adaptive_freelists()) {
7652     // Verify that the bit map has no bits marked between
7653     // addr and purported end of just dead object.
7654     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7655 
7656     do_post_free_or_garbage_chunk(fc, size);
7657   } else {
7658     if (!inFreeRange()) {
7659       // start of a new free range
7660       assert(size > 0, "A free range should have a size");
7661       initialize_free_range(addr, false);
7662     } else {
7663       // this will be swept up when we hit the end of the
7664       // free range
7665       if (CMSTraceSweeper) {
7666         gclog_or_tty->print("  -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7667       }
7668       // If the chunk is being coalesced and the current free range is
7669       // in the free lists, remove the current free range so that it
7670       // will be returned to the free lists in its entirety - all
7671       // the coalesced pieces included.
7672       if (freeRangeInFreeLists()) {
7673         FreeChunk* ffc = (FreeChunk*)freeFinger();
7674         assert(ffc->size() == pointer_delta(addr, freeFinger()),
7675           "Size of free range is inconsistent with chunk size.");
7676         if (CMSTestInFreeList) {
7677           assert(_sp->verify_chunk_in_free_list(ffc),
7678             "free range is not in free lists");
7679         }
7680         _sp->removeFreeChunkFromFreeLists(ffc);
7681         set_freeRangeInFreeLists(false);
7682       }
7683       set_lastFreeRangeCoalesced(true);
7684     }
7685     // this will be swept up when we hit the end of the free range
7686 
7687     // Verify that the bit map has no bits marked between
7688     // addr and purported end of just dead object.
7689     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7690   }
7691   assert(_limit >= addr + size,
7692          "A freshly garbage chunk can't possibly straddle over _limit");
7693   if (inFreeRange()) lookahead_and_flush(fc, size);
7694   return size;
7695 }
7696 
7697 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7698   HeapWord* addr = (HeapWord*) fc;
7699   // The sweeper has just found a live object. Return any accumulated
7700   // left hand chunk to the free lists.
7701   if (inFreeRange()) {
7702     assert(freeFinger() < addr, "freeFinger points too high");
7703     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7704   }
7705 
7706   // This object is live: we'd normally expect this to be
7707   // an oop, and like to assert the following:
7708   // assert(oop(addr)->is_oop(), "live block should be an oop");
7709   // However, as we commented above, this may be an object whose
7710   // header hasn't yet been initialized.
7711   size_t size;
7712   assert(_bitMap->isMarked(addr), "Tautology for this control point");
7713   if (_bitMap->isMarked(addr + 1)) {
7714     // Determine the size from the bit map, rather than trying to
7715     // compute it from the object header.
7716     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7717     size = pointer_delta(nextOneAddr + 1, addr);
7718     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7719            "alignment problem");
7720 
7721 #ifdef ASSERT
7722       if (oop(addr)->klass_or_null() != NULL) {
7723         // Ignore mark word because we are running concurrent with mutators
7724         assert(oop(addr)->is_oop(true), "live block should be an oop");
7725         assert(size ==
7726                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7727                "P-mark and computed size do not agree");
7728       }
7729 #endif
7730 
7731   } else {
7732     // This should be an initialized object that's alive.
7733     assert(oop(addr)->klass_or_null() != NULL,
7734            "Should be an initialized object");
7735     // Ignore mark word because we are running concurrent with mutators
7736     assert(oop(addr)->is_oop(true), "live block should be an oop");
7737     // Verify that the bit map has no bits marked between
7738     // addr and purported end of this block.
7739     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7740     assert(size >= 3, "Necessary for Printezis marks to work");
7741     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7742     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7743   }
7744   return size;
7745 }
7746 
7747 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7748                                                  size_t chunkSize) {
7749   // do_post_free_or_garbage_chunk() should only be called in the case
7750   // of the adaptive free list allocator.
7751   const bool fcInFreeLists = fc->is_free();
7752   assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7753   assert((HeapWord*)fc <= _limit, "sweep invariant");
7754   if (CMSTestInFreeList && fcInFreeLists) {
7755     assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7756   }
7757 
7758   if (CMSTraceSweeper) {
7759     gclog_or_tty->print_cr("  -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7760   }
7761 
7762   HeapWord* const fc_addr = (HeapWord*) fc;
7763 
7764   bool coalesce;
7765   const size_t left  = pointer_delta(fc_addr, freeFinger());
7766   const size_t right = chunkSize;
7767   switch (FLSCoalescePolicy) {
7768     // numeric value forms a coalition aggressiveness metric
7769     case 0:  { // never coalesce
7770       coalesce = false;
7771       break;
7772     }
7773     case 1: { // coalesce if left & right chunks on overpopulated lists
7774       coalesce = _sp->coalOverPopulated(left) &&
7775                  _sp->coalOverPopulated(right);
7776       break;
7777     }
7778     case 2: { // coalesce if left chunk on overpopulated list (default)
7779       coalesce = _sp->coalOverPopulated(left);
7780       break;
7781     }
7782     case 3: { // coalesce if left OR right chunk on overpopulated list
7783       coalesce = _sp->coalOverPopulated(left) ||
7784                  _sp->coalOverPopulated(right);
7785       break;
7786     }
7787     case 4: { // always coalesce
7788       coalesce = true;
7789       break;
7790     }
7791     default:
7792      ShouldNotReachHere();
7793   }
7794 
7795   // Should the current free range be coalesced?
7796   // If the chunk is in a free range and either we decided to coalesce above
7797   // or the chunk is near the large block at the end of the heap
7798   // (isNearLargestChunk() returns true), then coalesce this chunk.
7799   const bool doCoalesce = inFreeRange()
7800                           && (coalesce || _g->isNearLargestChunk(fc_addr));
7801   if (doCoalesce) {
7802     // Coalesce the current free range on the left with the new
7803     // chunk on the right.  If either is on a free list,
7804     // it must be removed from the list and stashed in the closure.
7805     if (freeRangeInFreeLists()) {
7806       FreeChunk* const ffc = (FreeChunk*)freeFinger();
7807       assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7808         "Size of free range is inconsistent with chunk size.");
7809       if (CMSTestInFreeList) {
7810         assert(_sp->verify_chunk_in_free_list(ffc),
7811           "Chunk is not in free lists");
7812       }
7813       _sp->coalDeath(ffc->size());
7814       _sp->removeFreeChunkFromFreeLists(ffc);
7815       set_freeRangeInFreeLists(false);
7816     }
7817     if (fcInFreeLists) {
7818       _sp->coalDeath(chunkSize);
7819       assert(fc->size() == chunkSize,
7820         "The chunk has the wrong size or is not in the free lists");
7821       _sp->removeFreeChunkFromFreeLists(fc);
7822     }
7823     set_lastFreeRangeCoalesced(true);
7824     print_free_block_coalesced(fc);
7825   } else {  // not in a free range and/or should not coalesce
7826     // Return the current free range and start a new one.
7827     if (inFreeRange()) {
7828       // In a free range but cannot coalesce with the right hand chunk.
7829       // Put the current free range into the free lists.
7830       flush_cur_free_chunk(freeFinger(),
7831                            pointer_delta(fc_addr, freeFinger()));
7832     }
7833     // Set up for new free range.  Pass along whether the right hand
7834     // chunk is in the free lists.
7835     initialize_free_range((HeapWord*)fc, fcInFreeLists);
7836   }
7837 }
7838 
7839 // Lookahead flush:
7840 // If we are tracking a free range, and this is the last chunk that
7841 // we'll look at because its end crosses past _limit, we'll preemptively
7842 // flush it along with any free range we may be holding on to. Note that
7843 // this can be the case only for an already free or freshly garbage
7844 // chunk. If this block is an object, it can never straddle
7845 // over _limit. The "straddling" occurs when _limit is set at
7846 // the previous end of the space when this cycle started, and
7847 // a subsequent heap expansion caused the previously co-terminal
7848 // free block to be coalesced with the newly expanded portion,
7849 // thus rendering _limit a non-block-boundary making it dangerous
7850 // for the sweeper to step over and examine.
7851 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7852   assert(inFreeRange(), "Should only be called if currently in a free range.");
7853   HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7854   assert(_sp->used_region().contains(eob - 1),
7855          err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7856                  " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7857                  " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7858                  p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size));
7859   if (eob >= _limit) {
7860     assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7861     if (CMSTraceSweeper) {
7862       gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7863                              "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7864                              "[" PTR_FORMAT "," PTR_FORMAT ")",
7865                              p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7866     }
7867     // Return the storage we are tracking back into the free lists.
7868     if (CMSTraceSweeper) {
7869       gclog_or_tty->print_cr("Flushing ... ");
7870     }
7871     assert(freeFinger() < eob, "Error");
7872     flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7873   }
7874 }
7875 
7876 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7877   assert(inFreeRange(), "Should only be called if currently in a free range.");
7878   assert(size > 0,
7879     "A zero sized chunk cannot be added to the free lists.");
7880   if (!freeRangeInFreeLists()) {
7881     if (CMSTestInFreeList) {
7882       FreeChunk* fc = (FreeChunk*) chunk;
7883       fc->set_size(size);
7884       assert(!_sp->verify_chunk_in_free_list(fc),
7885         "chunk should not be in free lists yet");
7886     }
7887     if (CMSTraceSweeper) {
7888       gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7889                     p2i(chunk), size);
7890     }
7891     // A new free range is going to be starting.  The current
7892     // free range has not been added to the free lists yet or
7893     // was removed so add it back.
7894     // If the current free range was coalesced, then the death
7895     // of the free range was recorded.  Record a birth now.
7896     if (lastFreeRangeCoalesced()) {
7897       _sp->coalBirth(size);
7898     }
7899     _sp->addChunkAndRepairOffsetTable(chunk, size,
7900             lastFreeRangeCoalesced());
7901   } else if (CMSTraceSweeper) {
7902     gclog_or_tty->print_cr("Already in free list: nothing to flush");
7903   }
7904   set_inFreeRange(false);
7905   set_freeRangeInFreeLists(false);
7906 }
7907 
7908 // We take a break if we've been at this for a while,
7909 // so as to avoid monopolizing the locks involved.
7910 void SweepClosure::do_yield_work(HeapWord* addr) {
7911   // Return current free chunk being used for coalescing (if any)
7912   // to the appropriate freelist.  After yielding, the next
7913   // free block encountered will start a coalescing range of
7914   // free blocks.  If the next free block is adjacent to the
7915   // chunk just flushed, they will need to wait for the next
7916   // sweep to be coalesced.
7917   if (inFreeRange()) {
7918     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7919   }
7920 
7921   // First give up the locks, then yield, then re-lock.
7922   // We should probably use a constructor/destructor idiom to
7923   // do this unlock/lock or modify the MutexUnlocker class to
7924   // serve our purpose. XXX
7925   assert_lock_strong(_bitMap->lock());
7926   assert_lock_strong(_freelistLock);
7927   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7928          "CMS thread should hold CMS token");
7929   _bitMap->lock()->unlock();
7930   _freelistLock->unlock();
7931   ConcurrentMarkSweepThread::desynchronize(true);
7932   _collector->stopTimer();
7933   if (PrintCMSStatistics != 0) {
7934     _collector->incrementYields();
7935   }
7936 
7937   // See the comment in coordinator_yield()
7938   for (unsigned i = 0; i < CMSYieldSleepCount &&
7939                        ConcurrentMarkSweepThread::should_yield() &&
7940                        !CMSCollector::foregroundGCIsActive(); ++i) {
7941     os::sleep(Thread::current(), 1, false);
7942   }
7943 
7944   ConcurrentMarkSweepThread::synchronize(true);
7945   _freelistLock->lock();
7946   _bitMap->lock()->lock_without_safepoint_check();
7947   _collector->startTimer();
7948 }
7949 
7950 #ifndef PRODUCT
7951 // This is actually very useful in a product build if it can
7952 // be called from the debugger.  Compile it into the product
7953 // as needed.
7954 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7955   return debug_cms_space->verify_chunk_in_free_list(fc);
7956 }
7957 #endif
7958 
7959 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7960   if (CMSTraceSweeper) {
7961     gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7962                            p2i(fc), fc->size());
7963   }
7964 }
7965 
7966 // CMSIsAliveClosure
7967 bool CMSIsAliveClosure::do_object_b(oop obj) {
7968   HeapWord* addr = (HeapWord*)obj;
7969   return addr != NULL &&
7970          (!_span.contains(addr) || _bit_map->isMarked(addr));
7971 }
7972 
7973 
7974 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7975                       MemRegion span,
7976                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7977                       bool cpc):
7978   _collector(collector),
7979   _span(span),
7980   _bit_map(bit_map),
7981   _mark_stack(mark_stack),
7982   _concurrent_precleaning(cpc) {
7983   assert(!_span.is_empty(), "Empty span could spell trouble");
7984 }
7985 
7986 
7987 // CMSKeepAliveClosure: the serial version
7988 void CMSKeepAliveClosure::do_oop(oop obj) {
7989   HeapWord* addr = (HeapWord*)obj;
7990   if (_span.contains(addr) &&
7991       !_bit_map->isMarked(addr)) {
7992     _bit_map->mark(addr);
7993     bool simulate_overflow = false;
7994     NOT_PRODUCT(
7995       if (CMSMarkStackOverflowALot &&
7996           _collector->simulate_overflow()) {
7997         // simulate a stack overflow
7998         simulate_overflow = true;
7999       }
8000     )
8001     if (simulate_overflow || !_mark_stack->push(obj)) {
8002       if (_concurrent_precleaning) {
8003         // We dirty the overflown object and let the remark
8004         // phase deal with it.
8005         assert(_collector->overflow_list_is_empty(), "Error");
8006         // In the case of object arrays, we need to dirty all of
8007         // the cards that the object spans. No locking or atomics
8008         // are needed since no one else can be mutating the mod union
8009         // table.
8010         if (obj->is_objArray()) {
8011           size_t sz = obj->size();
8012           HeapWord* end_card_addr =
8013             (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8014           MemRegion redirty_range = MemRegion(addr, end_card_addr);
8015           assert(!redirty_range.is_empty(), "Arithmetical tautology");
8016           _collector->_modUnionTable.mark_range(redirty_range);
8017         } else {
8018           _collector->_modUnionTable.mark(addr);
8019         }
8020         _collector->_ser_kac_preclean_ovflw++;
8021       } else {
8022         _collector->push_on_overflow_list(obj);
8023         _collector->_ser_kac_ovflw++;
8024       }
8025     }
8026   }
8027 }
8028 
8029 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
8030 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8031 
8032 // CMSParKeepAliveClosure: a parallel version of the above.
8033 // The work queues are private to each closure (thread),
8034 // but (may be) available for stealing by other threads.
8035 void CMSParKeepAliveClosure::do_oop(oop obj) {
8036   HeapWord* addr = (HeapWord*)obj;
8037   if (_span.contains(addr) &&
8038       !_bit_map->isMarked(addr)) {
8039     // In general, during recursive tracing, several threads
8040     // may be concurrently getting here; the first one to
8041     // "tag" it, claims it.
8042     if (_bit_map->par_mark(addr)) {
8043       bool res = _work_queue->push(obj);
8044       assert(res, "Low water mark should be much less than capacity");
8045       // Do a recursive trim in the hope that this will keep
8046       // stack usage lower, but leave some oops for potential stealers
8047       trim_queue(_low_water_mark);
8048     } // Else, another thread got there first
8049   }
8050 }
8051 
8052 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
8053 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8054 
8055 void CMSParKeepAliveClosure::trim_queue(uint max) {
8056   while (_work_queue->size() > max) {
8057     oop new_oop;
8058     if (_work_queue->pop_local(new_oop)) {
8059       assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8060       assert(_bit_map->isMarked((HeapWord*)new_oop),
8061              "no white objects on this stack!");
8062       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8063       // iterate over the oops in this oop, marking and pushing
8064       // the ones in CMS heap (i.e. in _span).
8065       new_oop->oop_iterate(&_mark_and_push);
8066     }
8067   }
8068 }
8069 
8070 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8071                                 CMSCollector* collector,
8072                                 MemRegion span, CMSBitMap* bit_map,
8073                                 OopTaskQueue* work_queue):
8074   _collector(collector),
8075   _span(span),
8076   _bit_map(bit_map),
8077   _work_queue(work_queue) { }
8078 
8079 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8080   HeapWord* addr = (HeapWord*)obj;
8081   if (_span.contains(addr) &&
8082       !_bit_map->isMarked(addr)) {
8083     if (_bit_map->par_mark(addr)) {
8084       bool simulate_overflow = false;
8085       NOT_PRODUCT(
8086         if (CMSMarkStackOverflowALot &&
8087             _collector->par_simulate_overflow()) {
8088           // simulate a stack overflow
8089           simulate_overflow = true;
8090         }
8091       )
8092       if (simulate_overflow || !_work_queue->push(obj)) {
8093         _collector->par_push_on_overflow_list(obj);
8094         _collector->_par_kac_ovflw++;
8095       }
8096     } // Else another thread got there already
8097   }
8098 }
8099 
8100 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8101 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8102 
8103 //////////////////////////////////////////////////////////////////
8104 //  CMSExpansionCause                /////////////////////////////
8105 //////////////////////////////////////////////////////////////////
8106 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8107   switch (cause) {
8108     case _no_expansion:
8109       return "No expansion";
8110     case _satisfy_free_ratio:
8111       return "Free ratio";
8112     case _satisfy_promotion:
8113       return "Satisfy promotion";
8114     case _satisfy_allocation:
8115       return "allocation";
8116     case _allocate_par_lab:
8117       return "Par LAB";
8118     case _allocate_par_spooling_space:
8119       return "Par Spooling Space";
8120     case _adaptive_size_policy:
8121       return "Ergonomics";
8122     default:
8123       return "unknown";
8124   }
8125 }
8126 
8127 void CMSDrainMarkingStackClosure::do_void() {
8128   // the max number to take from overflow list at a time
8129   const size_t num = _mark_stack->capacity()/4;
8130   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8131          "Overflow list should be NULL during concurrent phases");
8132   while (!_mark_stack->isEmpty() ||
8133          // if stack is empty, check the overflow list
8134          _collector->take_from_overflow_list(num, _mark_stack)) {
8135     oop obj = _mark_stack->pop();
8136     HeapWord* addr = (HeapWord*)obj;
8137     assert(_span.contains(addr), "Should be within span");
8138     assert(_bit_map->isMarked(addr), "Should be marked");
8139     assert(obj->is_oop(), "Should be an oop");
8140     obj->oop_iterate(_keep_alive);
8141   }
8142 }
8143 
8144 void CMSParDrainMarkingStackClosure::do_void() {
8145   // drain queue
8146   trim_queue(0);
8147 }
8148 
8149 // Trim our work_queue so its length is below max at return
8150 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8151   while (_work_queue->size() > max) {
8152     oop new_oop;
8153     if (_work_queue->pop_local(new_oop)) {
8154       assert(new_oop->is_oop(), "Expected an oop");
8155       assert(_bit_map->isMarked((HeapWord*)new_oop),
8156              "no white objects on this stack!");
8157       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8158       // iterate over the oops in this oop, marking and pushing
8159       // the ones in CMS heap (i.e. in _span).
8160       new_oop->oop_iterate(&_mark_and_push);
8161     }
8162   }
8163 }
8164 
8165 ////////////////////////////////////////////////////////////////////
8166 // Support for Marking Stack Overflow list handling and related code
8167 ////////////////////////////////////////////////////////////////////
8168 // Much of the following code is similar in shape and spirit to the
8169 // code used in ParNewGC. We should try and share that code
8170 // as much as possible in the future.
8171 
8172 #ifndef PRODUCT
8173 // Debugging support for CMSStackOverflowALot
8174 
8175 // It's OK to call this multi-threaded;  the worst thing
8176 // that can happen is that we'll get a bunch of closely
8177 // spaced simulated overflows, but that's OK, in fact
8178 // probably good as it would exercise the overflow code
8179 // under contention.
8180 bool CMSCollector::simulate_overflow() {
8181   if (_overflow_counter-- <= 0) { // just being defensive
8182     _overflow_counter = CMSMarkStackOverflowInterval;
8183     return true;
8184   } else {
8185     return false;
8186   }
8187 }
8188 
8189 bool CMSCollector::par_simulate_overflow() {
8190   return simulate_overflow();
8191 }
8192 #endif
8193 
8194 // Single-threaded
8195 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8196   assert(stack->isEmpty(), "Expected precondition");
8197   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8198   size_t i = num;
8199   oop  cur = _overflow_list;
8200   const markOop proto = markOopDesc::prototype();
8201   NOT_PRODUCT(ssize_t n = 0;)
8202   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8203     next = oop(cur->mark());
8204     cur->set_mark(proto);   // until proven otherwise
8205     assert(cur->is_oop(), "Should be an oop");
8206     bool res = stack->push(cur);
8207     assert(res, "Bit off more than can chew?");
8208     NOT_PRODUCT(n++;)
8209   }
8210   _overflow_list = cur;
8211 #ifndef PRODUCT
8212   assert(_num_par_pushes >= n, "Too many pops?");
8213   _num_par_pushes -=n;
8214 #endif
8215   return !stack->isEmpty();
8216 }
8217 
8218 #define BUSY  (cast_to_oop<intptr_t>(0x1aff1aff))
8219 // (MT-safe) Get a prefix of at most "num" from the list.
8220 // The overflow list is chained through the mark word of
8221 // each object in the list. We fetch the entire list,
8222 // break off a prefix of the right size and return the
8223 // remainder. If other threads try to take objects from
8224 // the overflow list at that time, they will wait for
8225 // some time to see if data becomes available. If (and
8226 // only if) another thread places one or more object(s)
8227 // on the global list before we have returned the suffix
8228 // to the global list, we will walk down our local list
8229 // to find its end and append the global list to
8230 // our suffix before returning it. This suffix walk can
8231 // prove to be expensive (quadratic in the amount of traffic)
8232 // when there are many objects in the overflow list and
8233 // there is much producer-consumer contention on the list.
8234 // *NOTE*: The overflow list manipulation code here and
8235 // in ParNewGeneration:: are very similar in shape,
8236 // except that in the ParNew case we use the old (from/eden)
8237 // copy of the object to thread the list via its klass word.
8238 // Because of the common code, if you make any changes in
8239 // the code below, please check the ParNew version to see if
8240 // similar changes might be needed.
8241 // CR 6797058 has been filed to consolidate the common code.
8242 bool CMSCollector::par_take_from_overflow_list(size_t num,
8243                                                OopTaskQueue* work_q,
8244                                                int no_of_gc_threads) {
8245   assert(work_q->size() == 0, "First empty local work queue");
8246   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8247   if (_overflow_list == NULL) {
8248     return false;
8249   }
8250   // Grab the entire list; we'll put back a suffix
8251   oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8252   Thread* tid = Thread::current();
8253   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8254   // set to ParallelGCThreads.
8255   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8256   size_t sleep_time_millis = MAX2((size_t)1, num/100);
8257   // If the list is busy, we spin for a short while,
8258   // sleeping between attempts to get the list.
8259   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8260     os::sleep(tid, sleep_time_millis, false);
8261     if (_overflow_list == NULL) {
8262       // Nothing left to take
8263       return false;
8264     } else if (_overflow_list != BUSY) {
8265       // Try and grab the prefix
8266       prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8267     }
8268   }
8269   // If the list was found to be empty, or we spun long
8270   // enough, we give up and return empty-handed. If we leave
8271   // the list in the BUSY state below, it must be the case that
8272   // some other thread holds the overflow list and will set it
8273   // to a non-BUSY state in the future.
8274   if (prefix == NULL || prefix == BUSY) {
8275      // Nothing to take or waited long enough
8276      if (prefix == NULL) {
8277        // Write back the NULL in case we overwrote it with BUSY above
8278        // and it is still the same value.
8279        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8280      }
8281      return false;
8282   }
8283   assert(prefix != NULL && prefix != BUSY, "Error");
8284   size_t i = num;
8285   oop cur = prefix;
8286   // Walk down the first "num" objects, unless we reach the end.
8287   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8288   if (cur->mark() == NULL) {
8289     // We have "num" or fewer elements in the list, so there
8290     // is nothing to return to the global list.
8291     // Write back the NULL in lieu of the BUSY we wrote
8292     // above, if it is still the same value.
8293     if (_overflow_list == BUSY) {
8294       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8295     }
8296   } else {
8297     // Chop off the suffix and return it to the global list.
8298     assert(cur->mark() != BUSY, "Error");
8299     oop suffix_head = cur->mark(); // suffix will be put back on global list
8300     cur->set_mark(NULL);           // break off suffix
8301     // It's possible that the list is still in the empty(busy) state
8302     // we left it in a short while ago; in that case we may be
8303     // able to place back the suffix without incurring the cost
8304     // of a walk down the list.
8305     oop observed_overflow_list = _overflow_list;
8306     oop cur_overflow_list = observed_overflow_list;
8307     bool attached = false;
8308     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8309       observed_overflow_list =
8310         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8311       if (cur_overflow_list == observed_overflow_list) {
8312         attached = true;
8313         break;
8314       } else cur_overflow_list = observed_overflow_list;
8315     }
8316     if (!attached) {
8317       // Too bad, someone else sneaked in (at least) an element; we'll need
8318       // to do a splice. Find tail of suffix so we can prepend suffix to global
8319       // list.
8320       for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8321       oop suffix_tail = cur;
8322       assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8323              "Tautology");
8324       observed_overflow_list = _overflow_list;
8325       do {
8326         cur_overflow_list = observed_overflow_list;
8327         if (cur_overflow_list != BUSY) {
8328           // Do the splice ...
8329           suffix_tail->set_mark(markOop(cur_overflow_list));
8330         } else { // cur_overflow_list == BUSY
8331           suffix_tail->set_mark(NULL);
8332         }
8333         // ... and try to place spliced list back on overflow_list ...
8334         observed_overflow_list =
8335           (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8336       } while (cur_overflow_list != observed_overflow_list);
8337       // ... until we have succeeded in doing so.
8338     }
8339   }
8340 
8341   // Push the prefix elements on work_q
8342   assert(prefix != NULL, "control point invariant");
8343   const markOop proto = markOopDesc::prototype();
8344   oop next;
8345   NOT_PRODUCT(ssize_t n = 0;)
8346   for (cur = prefix; cur != NULL; cur = next) {
8347     next = oop(cur->mark());
8348     cur->set_mark(proto);   // until proven otherwise
8349     assert(cur->is_oop(), "Should be an oop");
8350     bool res = work_q->push(cur);
8351     assert(res, "Bit off more than we can chew?");
8352     NOT_PRODUCT(n++;)
8353   }
8354 #ifndef PRODUCT
8355   assert(_num_par_pushes >= n, "Too many pops?");
8356   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8357 #endif
8358   return true;
8359 }
8360 
8361 // Single-threaded
8362 void CMSCollector::push_on_overflow_list(oop p) {
8363   NOT_PRODUCT(_num_par_pushes++;)
8364   assert(p->is_oop(), "Not an oop");
8365   preserve_mark_if_necessary(p);
8366   p->set_mark((markOop)_overflow_list);
8367   _overflow_list = p;
8368 }
8369 
8370 // Multi-threaded; use CAS to prepend to overflow list
8371 void CMSCollector::par_push_on_overflow_list(oop p) {
8372   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8373   assert(p->is_oop(), "Not an oop");
8374   par_preserve_mark_if_necessary(p);
8375   oop observed_overflow_list = _overflow_list;
8376   oop cur_overflow_list;
8377   do {
8378     cur_overflow_list = observed_overflow_list;
8379     if (cur_overflow_list != BUSY) {
8380       p->set_mark(markOop(cur_overflow_list));
8381     } else {
8382       p->set_mark(NULL);
8383     }
8384     observed_overflow_list =
8385       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8386   } while (cur_overflow_list != observed_overflow_list);
8387 }
8388 #undef BUSY
8389 
8390 // Single threaded
8391 // General Note on GrowableArray: pushes may silently fail
8392 // because we are (temporarily) out of C-heap for expanding
8393 // the stack. The problem is quite ubiquitous and affects
8394 // a lot of code in the JVM. The prudent thing for GrowableArray
8395 // to do (for now) is to exit with an error. However, that may
8396 // be too draconian in some cases because the caller may be
8397 // able to recover without much harm. For such cases, we
8398 // should probably introduce a "soft_push" method which returns
8399 // an indication of success or failure with the assumption that
8400 // the caller may be able to recover from a failure; code in
8401 // the VM can then be changed, incrementally, to deal with such
8402 // failures where possible, thus, incrementally hardening the VM
8403 // in such low resource situations.
8404 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8405   _preserved_oop_stack.push(p);
8406   _preserved_mark_stack.push(m);
8407   assert(m == p->mark(), "Mark word changed");
8408   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8409          "bijection");
8410 }
8411 
8412 // Single threaded
8413 void CMSCollector::preserve_mark_if_necessary(oop p) {
8414   markOop m = p->mark();
8415   if (m->must_be_preserved(p)) {
8416     preserve_mark_work(p, m);
8417   }
8418 }
8419 
8420 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8421   markOop m = p->mark();
8422   if (m->must_be_preserved(p)) {
8423     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8424     // Even though we read the mark word without holding
8425     // the lock, we are assured that it will not change
8426     // because we "own" this oop, so no other thread can
8427     // be trying to push it on the overflow list; see
8428     // the assertion in preserve_mark_work() that checks
8429     // that m == p->mark().
8430     preserve_mark_work(p, m);
8431   }
8432 }
8433 
8434 // We should be able to do this multi-threaded,
8435 // a chunk of stack being a task (this is
8436 // correct because each oop only ever appears
8437 // once in the overflow list. However, it's
8438 // not very easy to completely overlap this with
8439 // other operations, so will generally not be done
8440 // until all work's been completed. Because we
8441 // expect the preserved oop stack (set) to be small,
8442 // it's probably fine to do this single-threaded.
8443 // We can explore cleverer concurrent/overlapped/parallel
8444 // processing of preserved marks if we feel the
8445 // need for this in the future. Stack overflow should
8446 // be so rare in practice and, when it happens, its
8447 // effect on performance so great that this will
8448 // likely just be in the noise anyway.
8449 void CMSCollector::restore_preserved_marks_if_any() {
8450   assert(SafepointSynchronize::is_at_safepoint(),
8451          "world should be stopped");
8452   assert(Thread::current()->is_ConcurrentGC_thread() ||
8453          Thread::current()->is_VM_thread(),
8454          "should be single-threaded");
8455   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8456          "bijection");
8457 
8458   while (!_preserved_oop_stack.is_empty()) {
8459     oop p = _preserved_oop_stack.pop();
8460     assert(p->is_oop(), "Should be an oop");
8461     assert(_span.contains(p), "oop should be in _span");
8462     assert(p->mark() == markOopDesc::prototype(),
8463            "Set when taken from overflow list");
8464     markOop m = _preserved_mark_stack.pop();
8465     p->set_mark(m);
8466   }
8467   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8468          "stacks were cleared above");
8469 }
8470 
8471 #ifndef PRODUCT
8472 bool CMSCollector::no_preserved_marks() const {
8473   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8474 }
8475 #endif
8476 
8477 // Transfer some number of overflown objects to usual marking
8478 // stack. Return true if some objects were transferred.
8479 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8480   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8481                     (size_t)ParGCDesiredObjsFromOverflowList);
8482 
8483   bool res = _collector->take_from_overflow_list(num, _mark_stack);
8484   assert(_collector->overflow_list_is_empty() || res,
8485          "If list is not empty, we should have taken something");
8486   assert(!res || !_mark_stack->isEmpty(),
8487          "If we took something, it should now be on our stack");
8488   return res;
8489 }
8490 
8491 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8492   size_t res = _sp->block_size_no_stall(addr, _collector);
8493   if (_sp->block_is_obj(addr)) {
8494     if (_live_bit_map->isMarked(addr)) {
8495       // It can't have been dead in a previous cycle
8496       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8497     } else {
8498       _dead_bit_map->mark(addr);      // mark the dead object
8499     }
8500   }
8501   // Could be 0, if the block size could not be computed without stalling.
8502   return res;
8503 }
8504 
8505 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8506 
8507   switch (phase) {
8508     case CMSCollector::InitialMarking:
8509       initialize(true  /* fullGC */ ,
8510                  cause /* cause of the GC */,
8511                  true  /* recordGCBeginTime */,
8512                  true  /* recordPreGCUsage */,
8513                  false /* recordPeakUsage */,
8514                  false /* recordPostGCusage */,
8515                  true  /* recordAccumulatedGCTime */,
8516                  false /* recordGCEndTime */,
8517                  false /* countCollection */  );
8518       break;
8519 
8520     case CMSCollector::FinalMarking:
8521       initialize(true  /* fullGC */ ,
8522                  cause /* cause of the GC */,
8523                  false /* recordGCBeginTime */,
8524                  false /* recordPreGCUsage */,
8525                  false /* recordPeakUsage */,
8526                  false /* recordPostGCusage */,
8527                  true  /* recordAccumulatedGCTime */,
8528                  false /* recordGCEndTime */,
8529                  false /* countCollection */  );
8530       break;
8531 
8532     case CMSCollector::Sweeping:
8533       initialize(true  /* fullGC */ ,
8534                  cause /* cause of the GC */,
8535                  false /* recordGCBeginTime */,
8536                  false /* recordPreGCUsage */,
8537                  true  /* recordPeakUsage */,
8538                  true  /* recordPostGCusage */,
8539                  false /* recordAccumulatedGCTime */,
8540                  true  /* recordGCEndTime */,
8541                  true  /* countCollection */  );
8542       break;
8543 
8544     default:
8545       ShouldNotReachHere();
8546   }
8547 }