1 /*
   2  * Copyright (c) 2001, 2017, 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/symbolTable.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/codeCache.hpp"
  31 #include "gc/cms/cmsCollectorPolicy.hpp"
  32 #include "gc/cms/cmsOopClosures.inline.hpp"
  33 #include "gc/cms/compactibleFreeListSpace.hpp"
  34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
  35 #include "gc/cms/concurrentMarkSweepThread.hpp"
  36 #include "gc/cms/parNewGeneration.hpp"
  37 #include "gc/cms/vmCMSOperations.hpp"
  38 #include "gc/serial/genMarkSweep.hpp"
  39 #include "gc/serial/tenuredGeneration.hpp"
  40 #include "gc/shared/adaptiveSizePolicy.hpp"
  41 #include "gc/shared/cardGeneration.inline.hpp"
  42 #include "gc/shared/cardTableRS.hpp"
  43 #include "gc/shared/collectedHeap.inline.hpp"
  44 #include "gc/shared/collectorCounters.hpp"
  45 #include "gc/shared/collectorPolicy.hpp"
  46 #include "gc/shared/gcLocker.inline.hpp"
  47 #include "gc/shared/gcPolicyCounters.hpp"
  48 #include "gc/shared/gcTimer.hpp"
  49 #include "gc/shared/gcTrace.hpp"
  50 #include "gc/shared/gcTraceTime.inline.hpp"
  51 #include "gc/shared/genCollectedHeap.hpp"
  52 #include "gc/shared/genOopClosures.inline.hpp"
  53 #include "gc/shared/isGCActiveMark.hpp"
  54 #include "gc/shared/referencePolicy.hpp"
  55 #include "gc/shared/strongRootsScope.hpp"
  56 #include "gc/shared/taskqueue.inline.hpp"
  57 #include "logging/log.hpp"
  58 #include "logging/logStream.hpp"
  59 #include "memory/allocation.hpp"
  60 #include "memory/iterator.inline.hpp"
  61 #include "memory/padded.hpp"
  62 #include "memory/resourceArea.hpp"
  63 #include "oops/oop.inline.hpp"
  64 #include "prims/jvmtiExport.hpp"
  65 #include "runtime/atomic.hpp"
  66 #include "runtime/globals_extension.hpp"
  67 #include "runtime/handles.inline.hpp"
  68 #include "runtime/java.hpp"
  69 #include "runtime/orderAccess.inline.hpp"
  70 #include "runtime/timer.hpp"
  71 #include "runtime/vmThread.hpp"
  72 #include "services/memoryService.hpp"
  73 #include "services/runtimeService.hpp"
  74 #include "utilities/align.hpp"
  75 #include "utilities/stack.inline.hpp"
  76 
  77 // statics
  78 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
  79 bool CMSCollector::_full_gc_requested = false;
  80 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
  81 
  82 //////////////////////////////////////////////////////////////////
  83 // In support of CMS/VM thread synchronization
  84 //////////////////////////////////////////////////////////////////
  85 // We split use of the CGC_lock into 2 "levels".
  86 // The low-level locking is of the usual CGC_lock monitor. We introduce
  87 // a higher level "token" (hereafter "CMS token") built on top of the
  88 // low level monitor (hereafter "CGC lock").
  89 // The token-passing protocol gives priority to the VM thread. The
  90 // CMS-lock doesn't provide any fairness guarantees, but clients
  91 // should ensure that it is only held for very short, bounded
  92 // durations.
  93 //
  94 // When either of the CMS thread or the VM thread is involved in
  95 // collection operations during which it does not want the other
  96 // thread to interfere, it obtains the CMS token.
  97 //
  98 // If either thread tries to get the token while the other has
  99 // it, that thread waits. However, if the VM thread and CMS thread
 100 // both want the token, then the VM thread gets priority while the
 101 // CMS thread waits. This ensures, for instance, that the "concurrent"
 102 // phases of the CMS thread's work do not block out the VM thread
 103 // for long periods of time as the CMS thread continues to hog
 104 // the token. (See bug 4616232).
 105 //
 106 // The baton-passing functions are, however, controlled by the
 107 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
 108 // and here the low-level CMS lock, not the high level token,
 109 // ensures mutual exclusion.
 110 //
 111 // Two important conditions that we have to satisfy:
 112 // 1. if a thread does a low-level wait on the CMS lock, then it
 113 //    relinquishes the CMS token if it were holding that token
 114 //    when it acquired the low-level CMS lock.
 115 // 2. any low-level notifications on the low-level lock
 116 //    should only be sent when a thread has relinquished the token.
 117 //
 118 // In the absence of either property, we'd have potential deadlock.
 119 //
 120 // We protect each of the CMS (concurrent and sequential) phases
 121 // with the CMS _token_, not the CMS _lock_.
 122 //
 123 // The only code protected by CMS lock is the token acquisition code
 124 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
 125 // baton-passing code.
 126 //
 127 // Unfortunately, i couldn't come up with a good abstraction to factor and
 128 // hide the naked CGC_lock manipulation in the baton-passing code
 129 // further below. That's something we should try to do. Also, the proof
 130 // of correctness of this 2-level locking scheme is far from obvious,
 131 // and potentially quite slippery. We have an uneasy suspicion, for instance,
 132 // that there may be a theoretical possibility of delay/starvation in the
 133 // low-level lock/wait/notify scheme used for the baton-passing because of
 134 // potential interference with the priority scheme embodied in the
 135 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
 136 // invocation further below and marked with "XXX 20011219YSR".
 137 // Indeed, as we note elsewhere, this may become yet more slippery
 138 // in the presence of multiple CMS and/or multiple VM threads. XXX
 139 
 140 class CMSTokenSync: public StackObj {
 141  private:
 142   bool _is_cms_thread;
 143  public:
 144   CMSTokenSync(bool is_cms_thread):
 145     _is_cms_thread(is_cms_thread) {
 146     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
 147            "Incorrect argument to constructor");
 148     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
 149   }
 150 
 151   ~CMSTokenSync() {
 152     assert(_is_cms_thread ?
 153              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
 154              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
 155           "Incorrect state");
 156     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
 157   }
 158 };
 159 
 160 // Convenience class that does a CMSTokenSync, and then acquires
 161 // upto three locks.
 162 class CMSTokenSyncWithLocks: public CMSTokenSync {
 163  private:
 164   // Note: locks are acquired in textual declaration order
 165   // and released in the opposite order
 166   MutexLockerEx _locker1, _locker2, _locker3;
 167  public:
 168   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
 169                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
 170     CMSTokenSync(is_cms_thread),
 171     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
 172     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
 173     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
 174   { }
 175 };
 176 
 177 
 178 //////////////////////////////////////////////////////////////////
 179 //  Concurrent Mark-Sweep Generation /////////////////////////////
 180 //////////////////////////////////////////////////////////////////
 181 
 182 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
 183 
 184 // This struct contains per-thread things necessary to support parallel
 185 // young-gen collection.
 186 class CMSParGCThreadState: public CHeapObj<mtGC> {
 187  public:
 188   CompactibleFreeListSpaceLAB lab;
 189   PromotionInfo promo;
 190 
 191   // Constructor.
 192   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
 193     promo.setSpace(cfls);
 194   }
 195 };
 196 
 197 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
 198      ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
 199   CardGeneration(rs, initial_byte_size, ct),
 200   _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
 201   _did_compact(false)
 202 {
 203   HeapWord* bottom = (HeapWord*) _virtual_space.low();
 204   HeapWord* end    = (HeapWord*) _virtual_space.high();
 205 
 206   _direct_allocated_words = 0;
 207   NOT_PRODUCT(
 208     _numObjectsPromoted = 0;
 209     _numWordsPromoted = 0;
 210     _numObjectsAllocated = 0;
 211     _numWordsAllocated = 0;
 212   )
 213 
 214   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
 215   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
 216   _cmsSpace->_old_gen = this;
 217 
 218   _gc_stats = new CMSGCStats();
 219 
 220   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
 221   // offsets match. The ability to tell free chunks from objects
 222   // depends on this property.
 223   debug_only(
 224     FreeChunk* junk = NULL;
 225     assert(UseCompressedClassPointers ||
 226            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
 227            "Offset of FreeChunk::_prev within FreeChunk must match"
 228            "  that of OopDesc::_klass within OopDesc");
 229   )
 230 
 231   _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
 232   for (uint i = 0; i < ParallelGCThreads; i++) {
 233     _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
 234   }
 235 
 236   _incremental_collection_failed = false;
 237   // The "dilatation_factor" is the expansion that can occur on
 238   // account of the fact that the minimum object size in the CMS
 239   // generation may be larger than that in, say, a contiguous young
 240   //  generation.
 241   // Ideally, in the calculation below, we'd compute the dilatation
 242   // factor as: MinChunkSize/(promoting_gen's min object size)
 243   // Since we do not have such a general query interface for the
 244   // promoting generation, we'll instead just use the minimum
 245   // object size (which today is a header's worth of space);
 246   // note that all arithmetic is in units of HeapWords.
 247   assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
 248   assert(_dilatation_factor >= 1.0, "from previous assert");
 249 }
 250 
 251 
 252 // The field "_initiating_occupancy" represents the occupancy percentage
 253 // at which we trigger a new collection cycle.  Unless explicitly specified
 254 // via CMSInitiatingOccupancyFraction (argument "io" below), it
 255 // is calculated by:
 256 //
 257 //   Let "f" be MinHeapFreeRatio in
 258 //
 259 //    _initiating_occupancy = 100-f +
 260 //                           f * (CMSTriggerRatio/100)
 261 //   where CMSTriggerRatio is the argument "tr" below.
 262 //
 263 // That is, if we assume the heap is at its desired maximum occupancy at the
 264 // end of a collection, we let CMSTriggerRatio of the (purported) free
 265 // space be allocated before initiating a new collection cycle.
 266 //
 267 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
 268   assert(io <= 100 && tr <= 100, "Check the arguments");
 269   if (io >= 0) {
 270     _initiating_occupancy = (double)io / 100.0;
 271   } else {
 272     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
 273                              (double)(tr * MinHeapFreeRatio) / 100.0)
 274                             / 100.0;
 275   }
 276 }
 277 
 278 void ConcurrentMarkSweepGeneration::ref_processor_init() {
 279   assert(collector() != NULL, "no collector");
 280   collector()->ref_processor_init();
 281 }
 282 
 283 void CMSCollector::ref_processor_init() {
 284   if (_ref_processor == NULL) {
 285     // Allocate and initialize a reference processor
 286     _ref_processor =
 287       new ReferenceProcessor(_span,                               // span
 288                              (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
 289                              ParallelGCThreads,                   // mt processing degree
 290                              _cmsGen->refs_discovery_is_mt(),     // mt discovery
 291                              MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
 292                              _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
 293                              &_is_alive_closure);                 // closure for liveness info
 294     // Initialize the _ref_processor field of CMSGen
 295     _cmsGen->set_ref_processor(_ref_processor);
 296 
 297   }
 298 }
 299 
 300 AdaptiveSizePolicy* CMSCollector::size_policy() {
 301   GenCollectedHeap* gch = GenCollectedHeap::heap();
 302   return gch->gen_policy()->size_policy();
 303 }
 304 
 305 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
 306 
 307   const char* gen_name = "old";
 308   GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy();
 309   // Generation Counters - generation 1, 1 subspace
 310   _gen_counters = new GenerationCounters(gen_name, 1, 1,
 311       gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
 312 
 313   _space_counters = new GSpaceCounters(gen_name, 0,
 314                                        _virtual_space.reserved_size(),
 315                                        this, _gen_counters);
 316 }
 317 
 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
 319   _cms_gen(cms_gen)
 320 {
 321   assert(alpha <= 100, "bad value");
 322   _saved_alpha = alpha;
 323 
 324   // Initialize the alphas to the bootstrap value of 100.
 325   _gc0_alpha = _cms_alpha = 100;
 326 
 327   _cms_begin_time.update();
 328   _cms_end_time.update();
 329 
 330   _gc0_duration = 0.0;
 331   _gc0_period = 0.0;
 332   _gc0_promoted = 0;
 333 
 334   _cms_duration = 0.0;
 335   _cms_period = 0.0;
 336   _cms_allocated = 0;
 337 
 338   _cms_used_at_gc0_begin = 0;
 339   _cms_used_at_gc0_end = 0;
 340   _allow_duty_cycle_reduction = false;
 341   _valid_bits = 0;
 342 }
 343 
 344 double CMSStats::cms_free_adjustment_factor(size_t free) const {
 345   // TBD: CR 6909490
 346   return 1.0;
 347 }
 348 
 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
 350 }
 351 
 352 // If promotion failure handling is on use
 353 // the padded average size of the promotion for each
 354 // young generation collection.
 355 double CMSStats::time_until_cms_gen_full() const {
 356   size_t cms_free = _cms_gen->cmsSpace()->free();
 357   GenCollectedHeap* gch = GenCollectedHeap::heap();
 358   size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
 359                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
 360   if (cms_free > expected_promotion) {
 361     // Start a cms collection if there isn't enough space to promote
 362     // for the next young collection.  Use the padded average as
 363     // a safety factor.
 364     cms_free -= expected_promotion;
 365 
 366     // Adjust by the safety factor.
 367     double cms_free_dbl = (double)cms_free;
 368     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
 369     // Apply a further correction factor which tries to adjust
 370     // for recent occurance of concurrent mode failures.
 371     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
 372     cms_free_dbl = cms_free_dbl * cms_adjustment;
 373 
 374     log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
 375                   cms_free, expected_promotion);
 376     log_trace(gc)("  cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
 377     // Add 1 in case the consumption rate goes to zero.
 378     return cms_free_dbl / (cms_consumption_rate() + 1.0);
 379   }
 380   return 0.0;
 381 }
 382 
 383 // Compare the duration of the cms collection to the
 384 // time remaining before the cms generation is empty.
 385 // Note that the time from the start of the cms collection
 386 // to the start of the cms sweep (less than the total
 387 // duration of the cms collection) can be used.  This
 388 // has been tried and some applications experienced
 389 // promotion failures early in execution.  This was
 390 // possibly because the averages were not accurate
 391 // enough at the beginning.
 392 double CMSStats::time_until_cms_start() const {
 393   // We add "gc0_period" to the "work" calculation
 394   // below because this query is done (mostly) at the
 395   // end of a scavenge, so we need to conservatively
 396   // account for that much possible delay
 397   // in the query so as to avoid concurrent mode failures
 398   // due to starting the collection just a wee bit too
 399   // late.
 400   double work = cms_duration() + gc0_period();
 401   double deadline = time_until_cms_gen_full();
 402   // If a concurrent mode failure occurred recently, we want to be
 403   // more conservative and halve our expected time_until_cms_gen_full()
 404   if (work > deadline) {
 405     log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
 406                           cms_duration(), gc0_period(), time_until_cms_gen_full());
 407     return 0.0;
 408   }
 409   return work - deadline;
 410 }
 411 
 412 #ifndef PRODUCT
 413 void CMSStats::print_on(outputStream *st) const {
 414   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
 415   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
 416                gc0_duration(), gc0_period(), gc0_promoted());
 417   st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
 418             cms_duration(), cms_period(), cms_allocated());
 419   st->print(",cms_since_beg=%g,cms_since_end=%g",
 420             cms_time_since_begin(), cms_time_since_end());
 421   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
 422             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
 423 
 424   if (valid()) {
 425     st->print(",promo_rate=%g,cms_alloc_rate=%g",
 426               promotion_rate(), cms_allocation_rate());
 427     st->print(",cms_consumption_rate=%g,time_until_full=%g",
 428               cms_consumption_rate(), time_until_cms_gen_full());
 429   }
 430   st->cr();
 431 }
 432 #endif // #ifndef PRODUCT
 433 
 434 CMSCollector::CollectorState CMSCollector::_collectorState =
 435                              CMSCollector::Idling;
 436 bool CMSCollector::_foregroundGCIsActive = false;
 437 bool CMSCollector::_foregroundGCShouldWait = false;
 438 
 439 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
 440                            CardTableRS*                   ct,
 441                            ConcurrentMarkSweepPolicy*     cp):
 442   _cmsGen(cmsGen),
 443   _ct(ct),
 444   _ref_processor(NULL),    // will be set later
 445   _conc_workers(NULL),     // may be set later
 446   _abort_preclean(false),
 447   _start_sampling(false),
 448   _between_prologue_and_epilogue(false),
 449   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
 450   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
 451                  -1 /* lock-free */, "No_lock" /* dummy */),
 452   _modUnionClosurePar(&_modUnionTable),
 453   // Adjust my span to cover old (cms) gen
 454   _span(cmsGen->reserved()),
 455   // Construct the is_alive_closure with _span & markBitMap
 456   _is_alive_closure(_span, &_markBitMap),
 457   _restart_addr(NULL),
 458   _overflow_list(NULL),
 459   _stats(cmsGen),
 460   _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
 461                              //verify that this lock should be acquired with safepoint check.
 462                              Monitor::_safepoint_check_sometimes)),
 463   _eden_chunk_array(NULL),     // may be set in ctor body
 464   _eden_chunk_capacity(0),     // -- ditto --
 465   _eden_chunk_index(0),        // -- ditto --
 466   _survivor_plab_array(NULL),  // -- ditto --
 467   _survivor_chunk_array(NULL), // -- ditto --
 468   _survivor_chunk_capacity(0), // -- ditto --
 469   _survivor_chunk_index(0),    // -- ditto --
 470   _ser_pmc_preclean_ovflw(0),
 471   _ser_kac_preclean_ovflw(0),
 472   _ser_pmc_remark_ovflw(0),
 473   _par_pmc_remark_ovflw(0),
 474   _ser_kac_ovflw(0),
 475   _par_kac_ovflw(0),
 476 #ifndef PRODUCT
 477   _num_par_pushes(0),
 478 #endif
 479   _collection_count_start(0),
 480   _verifying(false),
 481   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
 482   _completed_initialization(false),
 483   _collector_policy(cp),
 484   _should_unload_classes(CMSClassUnloadingEnabled),
 485   _concurrent_cycles_since_last_unload(0),
 486   _roots_scanning_options(GenCollectedHeap::SO_None),
 487   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 488   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 489   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
 490   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 491   _cms_start_registered(false)
 492 {
 493   // Now expand the span and allocate the collection support structures
 494   // (MUT, marking bit map etc.) to cover both generations subject to
 495   // collection.
 496 
 497   // For use by dirty card to oop closures.
 498   _cmsGen->cmsSpace()->set_collector(this);
 499 
 500   // Allocate MUT and marking bit map
 501   {
 502     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
 503     if (!_markBitMap.allocate(_span)) {
 504       log_warning(gc)("Failed to allocate CMS Bit Map");
 505       return;
 506     }
 507     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
 508   }
 509   {
 510     _modUnionTable.allocate(_span);
 511     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
 512   }
 513 
 514   if (!_markStack.allocate(MarkStackSize)) {
 515     log_warning(gc)("Failed to allocate CMS Marking Stack");
 516     return;
 517   }
 518 
 519   // Support for multi-threaded concurrent phases
 520   if (CMSConcurrentMTEnabled) {
 521     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
 522       // just for now
 523       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
 524     }
 525     if (ConcGCThreads > 1) {
 526       _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
 527                                  ConcGCThreads, true);
 528       if (_conc_workers == NULL) {
 529         log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
 530         CMSConcurrentMTEnabled = false;
 531       } else {
 532         _conc_workers->initialize_workers();
 533       }
 534     } else {
 535       CMSConcurrentMTEnabled = false;
 536     }
 537   }
 538   if (!CMSConcurrentMTEnabled) {
 539     ConcGCThreads = 0;
 540   } else {
 541     // Turn off CMSCleanOnEnter optimization temporarily for
 542     // the MT case where it's not fixed yet; see 6178663.
 543     CMSCleanOnEnter = false;
 544   }
 545   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
 546          "Inconsistency");
 547   log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
 548   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 549 
 550   // Parallel task queues; these are shared for the
 551   // concurrent and stop-world phases of CMS, but
 552   // are not shared with parallel scavenge (ParNew).
 553   {
 554     uint i;
 555     uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
 556 
 557     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
 558          || ParallelRefProcEnabled)
 559         && num_queues > 0) {
 560       _task_queues = new OopTaskQueueSet(num_queues);
 561       if (_task_queues == NULL) {
 562         log_warning(gc)("task_queues allocation failure.");
 563         return;
 564       }
 565       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
 566       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
 567       for (i = 0; i < num_queues; i++) {
 568         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
 569         if (q == NULL) {
 570           log_warning(gc)("work_queue allocation failure.");
 571           return;
 572         }
 573         _task_queues->register_queue(i, q);
 574       }
 575       for (i = 0; i < num_queues; i++) {
 576         _task_queues->queue(i)->initialize();
 577         _hash_seed[i] = 17;  // copied from ParNew
 578       }
 579     }
 580   }
 581 
 582   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
 583 
 584   // Clip CMSBootstrapOccupancy between 0 and 100.
 585   _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
 586 
 587   // Now tell CMS generations the identity of their collector
 588   ConcurrentMarkSweepGeneration::set_collector(this);
 589 
 590   // Create & start a CMS thread for this CMS collector
 591   _cmsThread = ConcurrentMarkSweepThread::start(this);
 592   assert(cmsThread() != NULL, "CMS Thread should have been created");
 593   assert(cmsThread()->collector() == this,
 594          "CMS Thread should refer to this gen");
 595   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 596 
 597   // Support for parallelizing young gen rescan
 598   GenCollectedHeap* gch = GenCollectedHeap::heap();
 599   assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
 600   _young_gen = (ParNewGeneration*)gch->young_gen();
 601   if (gch->supports_inline_contig_alloc()) {
 602     _top_addr = gch->top_addr();
 603     _end_addr = gch->end_addr();
 604     assert(_young_gen != NULL, "no _young_gen");
 605     _eden_chunk_index = 0;
 606     _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
 607     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
 608   }
 609 
 610   // Support for parallelizing survivor space rescan
 611   if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
 612     const size_t max_plab_samples =
 613       _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
 614 
 615     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
 616     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 617     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
 618     _survivor_chunk_capacity = max_plab_samples;
 619     for (uint i = 0; i < ParallelGCThreads; i++) {
 620       HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 621       ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
 622       assert(cur->end() == 0, "Should be 0");
 623       assert(cur->array() == vec, "Should be vec");
 624       assert(cur->capacity() == max_plab_samples, "Error");
 625     }
 626   }
 627 
 628   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
 629   _gc_counters = new CollectorCounters("CMS", 1);
 630   _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2);
 631   _completed_initialization = true;
 632   _inter_sweep_timer.start();  // start of time
 633 }
 634 
 635 const char* ConcurrentMarkSweepGeneration::name() const {
 636   return "concurrent mark-sweep generation";
 637 }
 638 void ConcurrentMarkSweepGeneration::update_counters() {
 639   if (UsePerfData) {
 640     _space_counters->update_all();
 641     _gen_counters->update_all();
 642   }
 643 }
 644 
 645 // this is an optimized version of update_counters(). it takes the
 646 // used value as a parameter rather than computing it.
 647 //
 648 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
 649   if (UsePerfData) {
 650     _space_counters->update_used(used);
 651     _space_counters->update_capacity();
 652     _gen_counters->update_all();
 653   }
 654 }
 655 
 656 void ConcurrentMarkSweepGeneration::print() const {
 657   Generation::print();
 658   cmsSpace()->print();
 659 }
 660 
 661 #ifndef PRODUCT
 662 void ConcurrentMarkSweepGeneration::print_statistics() {
 663   cmsSpace()->printFLCensus(0);
 664 }
 665 #endif
 666 
 667 size_t
 668 ConcurrentMarkSweepGeneration::contiguous_available() const {
 669   // dld proposes an improvement in precision here. If the committed
 670   // part of the space ends in a free block we should add that to
 671   // uncommitted size in the calculation below. Will make this
 672   // change later, staying with the approximation below for the
 673   // time being. -- ysr.
 674   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
 675 }
 676 
 677 size_t
 678 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
 679   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
 680 }
 681 
 682 size_t ConcurrentMarkSweepGeneration::max_available() const {
 683   return free() + _virtual_space.uncommitted_size();
 684 }
 685 
 686 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
 687   size_t available = max_available();
 688   size_t av_promo  = (size_t)gc_stats()->avg_promoted()->padded_average();
 689   bool   res = (available >= av_promo) || (available >= max_promotion_in_bytes);
 690   log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
 691                            res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
 692   return res;
 693 }
 694 
 695 // At a promotion failure dump information on block layout in heap
 696 // (cms old generation).
 697 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
 698   Log(gc, promotion) log;
 699   if (log.is_trace()) {
 700     LogStream ls(log.trace());
 701     cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
 702   }
 703 }
 704 
 705 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
 706   // Clear the promotion information.  These pointers can be adjusted
 707   // along with all the other pointers into the heap but
 708   // compaction is expected to be a rare event with
 709   // a heap using cms so don't do it without seeing the need.
 710   for (uint i = 0; i < ParallelGCThreads; i++) {
 711     _par_gc_thread_states[i]->promo.reset();
 712   }
 713 }
 714 
 715 void ConcurrentMarkSweepGeneration::compute_new_size() {
 716   assert_locked_or_safepoint(Heap_lock);
 717 
 718   // If incremental collection failed, we just want to expand
 719   // to the limit.
 720   if (incremental_collection_failed()) {
 721     clear_incremental_collection_failed();
 722     grow_to_reserved();
 723     return;
 724   }
 725 
 726   // The heap has been compacted but not reset yet.
 727   // Any metric such as free() or used() will be incorrect.
 728 
 729   CardGeneration::compute_new_size();
 730 
 731   // Reset again after a possible resizing
 732   if (did_compact()) {
 733     cmsSpace()->reset_after_compaction();
 734   }
 735 }
 736 
 737 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
 738   assert_locked_or_safepoint(Heap_lock);
 739 
 740   // If incremental collection failed, we just want to expand
 741   // to the limit.
 742   if (incremental_collection_failed()) {
 743     clear_incremental_collection_failed();
 744     grow_to_reserved();
 745     return;
 746   }
 747 
 748   double free_percentage = ((double) free()) / capacity();
 749   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
 750   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
 751 
 752   // compute expansion delta needed for reaching desired free percentage
 753   if (free_percentage < desired_free_percentage) {
 754     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 755     assert(desired_capacity >= capacity(), "invalid expansion size");
 756     size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
 757     Log(gc) log;
 758     if (log.is_trace()) {
 759       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 760       log.trace("From compute_new_size: ");
 761       log.trace("  Free fraction %f", free_percentage);
 762       log.trace("  Desired free fraction %f", desired_free_percentage);
 763       log.trace("  Maximum free fraction %f", maximum_free_percentage);
 764       log.trace("  Capacity " SIZE_FORMAT, capacity() / 1000);
 765       log.trace("  Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
 766       GenCollectedHeap* gch = GenCollectedHeap::heap();
 767       assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
 768       size_t young_size = gch->young_gen()->capacity();
 769       log.trace("  Young gen size " SIZE_FORMAT, young_size / 1000);
 770       log.trace("  unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
 771       log.trace("  contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
 772       log.trace("  Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
 773     }
 774     // safe if expansion fails
 775     expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
 776     log.trace("  Expanded free fraction %f", ((double) free()) / capacity());
 777   } else {
 778     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 779     assert(desired_capacity <= capacity(), "invalid expansion size");
 780     size_t shrink_bytes = capacity() - desired_capacity;
 781     // Don't shrink unless the delta is greater than the minimum shrink we want
 782     if (shrink_bytes >= MinHeapDeltaBytes) {
 783       shrink_free_list_by(shrink_bytes);
 784     }
 785   }
 786 }
 787 
 788 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
 789   return cmsSpace()->freelistLock();
 790 }
 791 
 792 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
 793   CMSSynchronousYieldRequest yr;
 794   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
 795   return have_lock_and_allocate(size, tlab);
 796 }
 797 
 798 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
 799                                                                 bool   tlab /* ignored */) {
 800   assert_lock_strong(freelistLock());
 801   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
 802   HeapWord* res = cmsSpace()->allocate(adjustedSize);
 803   // Allocate the object live (grey) if the background collector has
 804   // started marking. This is necessary because the marker may
 805   // have passed this address and consequently this object will
 806   // not otherwise be greyed and would be incorrectly swept up.
 807   // Note that if this object contains references, the writing
 808   // of those references will dirty the card containing this object
 809   // allowing the object to be blackened (and its references scanned)
 810   // either during a preclean phase or at the final checkpoint.
 811   if (res != NULL) {
 812     // We may block here with an uninitialized object with
 813     // its mark-bit or P-bits not yet set. Such objects need
 814     // to be safely navigable by block_start().
 815     assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
 816     assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
 817     collector()->direct_allocated(res, adjustedSize);
 818     _direct_allocated_words += adjustedSize;
 819     // allocation counters
 820     NOT_PRODUCT(
 821       _numObjectsAllocated++;
 822       _numWordsAllocated += (int)adjustedSize;
 823     )
 824   }
 825   return res;
 826 }
 827 
 828 // In the case of direct allocation by mutators in a generation that
 829 // is being concurrently collected, the object must be allocated
 830 // live (grey) if the background collector has started marking.
 831 // This is necessary because the marker may
 832 // have passed this address and consequently this object will
 833 // not otherwise be greyed and would be incorrectly swept up.
 834 // Note that if this object contains references, the writing
 835 // of those references will dirty the card containing this object
 836 // allowing the object to be blackened (and its references scanned)
 837 // either during a preclean phase or at the final checkpoint.
 838 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
 839   assert(_markBitMap.covers(start, size), "Out of bounds");
 840   if (_collectorState >= Marking) {
 841     MutexLockerEx y(_markBitMap.lock(),
 842                     Mutex::_no_safepoint_check_flag);
 843     // [see comments preceding SweepClosure::do_blk() below for details]
 844     //
 845     // Can the P-bits be deleted now?  JJJ
 846     //
 847     // 1. need to mark the object as live so it isn't collected
 848     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
 849     // 3. need to mark the end of the object so marking, precleaning or sweeping
 850     //    can skip over uninitialized or unparsable objects. An allocated
 851     //    object is considered uninitialized for our purposes as long as
 852     //    its klass word is NULL.  All old gen objects are parsable
 853     //    as soon as they are initialized.)
 854     _markBitMap.mark(start);          // object is live
 855     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
 856     _markBitMap.mark(start + size - 1);
 857                                       // mark end of object
 858   }
 859   // check that oop looks uninitialized
 860   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
 861 }
 862 
 863 void CMSCollector::promoted(bool par, HeapWord* start,
 864                             bool is_obj_array, size_t obj_size) {
 865   assert(_markBitMap.covers(start), "Out of bounds");
 866   // See comment in direct_allocated() about when objects should
 867   // be allocated live.
 868   if (_collectorState >= Marking) {
 869     // we already hold the marking bit map lock, taken in
 870     // the prologue
 871     if (par) {
 872       _markBitMap.par_mark(start);
 873     } else {
 874       _markBitMap.mark(start);
 875     }
 876     // We don't need to mark the object as uninitialized (as
 877     // in direct_allocated above) because this is being done with the
 878     // world stopped and the object will be initialized by the
 879     // time the marking, precleaning or sweeping get to look at it.
 880     // But see the code for copying objects into the CMS generation,
 881     // where we need to ensure that concurrent readers of the
 882     // block offset table are able to safely navigate a block that
 883     // is in flux from being free to being allocated (and in
 884     // transition while being copied into) and subsequently
 885     // becoming a bona-fide object when the copy/promotion is complete.
 886     assert(SafepointSynchronize::is_at_safepoint(),
 887            "expect promotion only at safepoints");
 888 
 889     if (_collectorState < Sweeping) {
 890       // Mark the appropriate cards in the modUnionTable, so that
 891       // this object gets scanned before the sweep. If this is
 892       // not done, CMS generation references in the object might
 893       // not get marked.
 894       // For the case of arrays, which are otherwise precisely
 895       // marked, we need to dirty the entire array, not just its head.
 896       if (is_obj_array) {
 897         // The [par_]mark_range() method expects mr.end() below to
 898         // be aligned to the granularity of a bit's representation
 899         // in the heap. In the case of the MUT below, that's a
 900         // card size.
 901         MemRegion mr(start,
 902                      align_up(start + obj_size,
 903                         CardTableModRefBS::card_size /* bytes */));
 904         if (par) {
 905           _modUnionTable.par_mark_range(mr);
 906         } else {
 907           _modUnionTable.mark_range(mr);
 908         }
 909       } else {  // not an obj array; we can just mark the head
 910         if (par) {
 911           _modUnionTable.par_mark(start);
 912         } else {
 913           _modUnionTable.mark(start);
 914         }
 915       }
 916     }
 917   }
 918 }
 919 
 920 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
 921   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
 922   // allocate, copy and if necessary update promoinfo --
 923   // delegate to underlying space.
 924   assert_lock_strong(freelistLock());
 925 
 926 #ifndef PRODUCT
 927   if (GenCollectedHeap::heap()->promotion_should_fail()) {
 928     return NULL;
 929   }
 930 #endif  // #ifndef PRODUCT
 931 
 932   oop res = _cmsSpace->promote(obj, obj_size);
 933   if (res == NULL) {
 934     // expand and retry
 935     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
 936     expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
 937     // Since this is the old generation, we don't try to promote
 938     // into a more senior generation.
 939     res = _cmsSpace->promote(obj, obj_size);
 940   }
 941   if (res != NULL) {
 942     // See comment in allocate() about when objects should
 943     // be allocated live.
 944     assert(oopDesc::is_oop(obj), "Will dereference klass pointer below");
 945     collector()->promoted(false,           // Not parallel
 946                           (HeapWord*)res, obj->is_objArray(), obj_size);
 947     // promotion counters
 948     NOT_PRODUCT(
 949       _numObjectsPromoted++;
 950       _numWordsPromoted +=
 951         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
 952     )
 953   }
 954   return res;
 955 }
 956 
 957 
 958 // IMPORTANT: Notes on object size recognition in CMS.
 959 // ---------------------------------------------------
 960 // A block of storage in the CMS generation is always in
 961 // one of three states. A free block (FREE), an allocated
 962 // object (OBJECT) whose size() method reports the correct size,
 963 // and an intermediate state (TRANSIENT) in which its size cannot
 964 // be accurately determined.
 965 // STATE IDENTIFICATION:   (32 bit and 64 bit w/o COOPS)
 966 // -----------------------------------------------------
 967 // FREE:      klass_word & 1 == 1; mark_word holds block size
 968 //
 969 // OBJECT:    klass_word installed; klass_word != 0 && klass_word & 1 == 0;
 970 //            obj->size() computes correct size
 971 //
 972 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
 973 //
 974 // STATE IDENTIFICATION: (64 bit+COOPS)
 975 // ------------------------------------
 976 // FREE:      mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
 977 //
 978 // OBJECT:    klass_word installed; klass_word != 0;
 979 //            obj->size() computes correct size
 980 //
 981 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
 982 //
 983 //
 984 // STATE TRANSITION DIAGRAM
 985 //
 986 //        mut / parnew                     mut  /  parnew
 987 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
 988 //  ^                                                                   |
 989 //  |------------------------ DEAD <------------------------------------|
 990 //         sweep                            mut
 991 //
 992 // While a block is in TRANSIENT state its size cannot be determined
 993 // so readers will either need to come back later or stall until
 994 // the size can be determined. Note that for the case of direct
 995 // allocation, P-bits, when available, may be used to determine the
 996 // size of an object that may not yet have been initialized.
 997 
 998 // Things to support parallel young-gen collection.
 999 oop
1000 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1001                                            oop old, markOop m,
1002                                            size_t word_sz) {
1003 #ifndef PRODUCT
1004   if (GenCollectedHeap::heap()->promotion_should_fail()) {
1005     return NULL;
1006   }
1007 #endif  // #ifndef PRODUCT
1008 
1009   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1010   PromotionInfo* promoInfo = &ps->promo;
1011   // if we are tracking promotions, then first ensure space for
1012   // promotion (including spooling space for saving header if necessary).
1013   // then allocate and copy, then track promoted info if needed.
1014   // When tracking (see PromotionInfo::track()), the mark word may
1015   // be displaced and in this case restoration of the mark word
1016   // occurs in the (oop_since_save_marks_)iterate phase.
1017   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1018     // Out of space for allocating spooling buffers;
1019     // try expanding and allocating spooling buffers.
1020     if (!expand_and_ensure_spooling_space(promoInfo)) {
1021       return NULL;
1022     }
1023   }
1024   assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1025   const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1026   HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1027   if (obj_ptr == NULL) {
1028      obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1029      if (obj_ptr == NULL) {
1030        return NULL;
1031      }
1032   }
1033   oop obj = oop(obj_ptr);
1034   OrderAccess::storestore();
1035   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1036   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1037   // IMPORTANT: See note on object initialization for CMS above.
1038   // Otherwise, copy the object.  Here we must be careful to insert the
1039   // klass pointer last, since this marks the block as an allocated object.
1040   // Except with compressed oops it's the mark word.
1041   HeapWord* old_ptr = (HeapWord*)old;
1042   // Restore the mark word copied above.
1043   obj->set_mark(m);
1044   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1045   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1046   OrderAccess::storestore();
1047 
1048   if (UseCompressedClassPointers) {
1049     // Copy gap missed by (aligned) header size calculation below
1050     obj->set_klass_gap(old->klass_gap());
1051   }
1052   if (word_sz > (size_t)oopDesc::header_size()) {
1053     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1054                                  obj_ptr + oopDesc::header_size(),
1055                                  word_sz - oopDesc::header_size());
1056   }
1057 
1058   // Now we can track the promoted object, if necessary.  We take care
1059   // to delay the transition from uninitialized to full object
1060   // (i.e., insertion of klass pointer) until after, so that it
1061   // atomically becomes a promoted object.
1062   if (promoInfo->tracking()) {
1063     promoInfo->track((PromotedObject*)obj, old->klass());
1064   }
1065   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1066   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1067   assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below");
1068 
1069   // Finally, install the klass pointer (this should be volatile).
1070   OrderAccess::storestore();
1071   obj->set_klass(old->klass());
1072   // We should now be able to calculate the right size for this object
1073   assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1074 
1075   collector()->promoted(true,          // parallel
1076                         obj_ptr, old->is_objArray(), word_sz);
1077 
1078   NOT_PRODUCT(
1079     Atomic::inc_ptr(&_numObjectsPromoted);
1080     Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1081   )
1082 
1083   return obj;
1084 }
1085 
1086 void
1087 ConcurrentMarkSweepGeneration::
1088 par_promote_alloc_done(int thread_num) {
1089   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1090   ps->lab.retire(thread_num);
1091 }
1092 
1093 void
1094 ConcurrentMarkSweepGeneration::
1095 par_oop_since_save_marks_iterate_done(int thread_num) {
1096   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1097   ParScanWithoutBarrierClosure* dummy_cl = NULL;
1098   ps->promo.promoted_oops_iterate_nv(dummy_cl);
1099 
1100   // Because card-scanning has been completed, subsequent phases
1101   // (e.g., reference processing) will not need to recognize which
1102   // objects have been promoted during this GC. So, we can now disable
1103   // promotion tracking.
1104   ps->promo.stopTrackingPromotions();
1105 }
1106 
1107 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
1108                                                    size_t size,
1109                                                    bool   tlab)
1110 {
1111   // We allow a STW collection only if a full
1112   // collection was requested.
1113   return full || should_allocate(size, tlab); // FIX ME !!!
1114   // This and promotion failure handling are connected at the
1115   // hip and should be fixed by untying them.
1116 }
1117 
1118 bool CMSCollector::shouldConcurrentCollect() {
1119   LogTarget(Trace, gc) log;
1120 
1121   if (_full_gc_requested) {
1122     log.print("CMSCollector: collect because of explicit  gc request (or GCLocker)");
1123     return true;
1124   }
1125 
1126   FreelistLocker x(this);
1127   // ------------------------------------------------------------------
1128   // Print out lots of information which affects the initiation of
1129   // a collection.
1130   if (log.is_enabled() && stats().valid()) {
1131     log.print("CMSCollector shouldConcurrentCollect: ");
1132 
1133     LogStream out(log);
1134     stats().print_on(&out);
1135 
1136     log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1137     log.print("free=" SIZE_FORMAT, _cmsGen->free());
1138     log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1139     log.print("promotion_rate=%g", stats().promotion_rate());
1140     log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1141     log.print("occupancy=%3.7f", _cmsGen->occupancy());
1142     log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1143     log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1144     log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1145     log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1146   }
1147   // ------------------------------------------------------------------
1148 
1149   // If the estimated time to complete a cms collection (cms_duration())
1150   // is less than the estimated time remaining until the cms generation
1151   // is full, start a collection.
1152   if (!UseCMSInitiatingOccupancyOnly) {
1153     if (stats().valid()) {
1154       if (stats().time_until_cms_start() == 0.0) {
1155         return true;
1156       }
1157     } else {
1158       // We want to conservatively collect somewhat early in order
1159       // to try and "bootstrap" our CMS/promotion statistics;
1160       // this branch will not fire after the first successful CMS
1161       // collection because the stats should then be valid.
1162       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1163         log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1164                   _cmsGen->occupancy(), _bootstrap_occupancy);
1165         return true;
1166       }
1167     }
1168   }
1169 
1170   // Otherwise, we start a collection cycle if
1171   // old gen want a collection cycle started. Each may use
1172   // an appropriate criterion for making this decision.
1173   // XXX We need to make sure that the gen expansion
1174   // criterion dovetails well with this. XXX NEED TO FIX THIS
1175   if (_cmsGen->should_concurrent_collect()) {
1176     log.print("CMS old gen initiated");
1177     return true;
1178   }
1179 
1180   // We start a collection if we believe an incremental collection may fail;
1181   // this is not likely to be productive in practice because it's probably too
1182   // late anyway.
1183   GenCollectedHeap* gch = GenCollectedHeap::heap();
1184   assert(gch->collector_policy()->is_generation_policy(),
1185          "You may want to check the correctness of the following");
1186   if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1187     log.print("CMSCollector: collect because incremental collection will fail ");
1188     return true;
1189   }
1190 
1191   if (MetaspaceGC::should_concurrent_collect()) {
1192     log.print("CMSCollector: collect for metadata allocation ");
1193     return true;
1194   }
1195 
1196   // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1197   if (CMSTriggerInterval >= 0) {
1198     if (CMSTriggerInterval == 0) {
1199       // Trigger always
1200       return true;
1201     }
1202 
1203     // Check the CMS time since begin (we do not check the stats validity
1204     // as we want to be able to trigger the first CMS cycle as well)
1205     if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1206       if (stats().valid()) {
1207         log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1208                   stats().cms_time_since_begin());
1209       } else {
1210         log.print("CMSCollector: collect because of trigger interval (first collection)");
1211       }
1212       return true;
1213     }
1214   }
1215 
1216   return false;
1217 }
1218 
1219 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1220 
1221 // Clear _expansion_cause fields of constituent generations
1222 void CMSCollector::clear_expansion_cause() {
1223   _cmsGen->clear_expansion_cause();
1224 }
1225 
1226 // We should be conservative in starting a collection cycle.  To
1227 // start too eagerly runs the risk of collecting too often in the
1228 // extreme.  To collect too rarely falls back on full collections,
1229 // which works, even if not optimum in terms of concurrent work.
1230 // As a work around for too eagerly collecting, use the flag
1231 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
1232 // giving the user an easily understandable way of controlling the
1233 // collections.
1234 // We want to start a new collection cycle if any of the following
1235 // conditions hold:
1236 // . our current occupancy exceeds the configured initiating occupancy
1237 //   for this generation, or
1238 // . we recently needed to expand this space and have not, since that
1239 //   expansion, done a collection of this generation, or
1240 // . the underlying space believes that it may be a good idea to initiate
1241 //   a concurrent collection (this may be based on criteria such as the
1242 //   following: the space uses linear allocation and linear allocation is
1243 //   going to fail, or there is believed to be excessive fragmentation in
1244 //   the generation, etc... or ...
1245 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1246 //   the case of the old generation; see CR 6543076):
1247 //   we may be approaching a point at which allocation requests may fail because
1248 //   we will be out of sufficient free space given allocation rate estimates.]
1249 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1250 
1251   assert_lock_strong(freelistLock());
1252   if (occupancy() > initiating_occupancy()) {
1253     log_trace(gc)(" %s: collect because of occupancy %f / %f  ",
1254                   short_name(), occupancy(), initiating_occupancy());
1255     return true;
1256   }
1257   if (UseCMSInitiatingOccupancyOnly) {
1258     return false;
1259   }
1260   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1261     log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1262     return true;
1263   }
1264   return false;
1265 }
1266 
1267 void ConcurrentMarkSweepGeneration::collect(bool   full,
1268                                             bool   clear_all_soft_refs,
1269                                             size_t size,
1270                                             bool   tlab)
1271 {
1272   collector()->collect(full, clear_all_soft_refs, size, tlab);
1273 }
1274 
1275 void CMSCollector::collect(bool   full,
1276                            bool   clear_all_soft_refs,
1277                            size_t size,
1278                            bool   tlab)
1279 {
1280   // The following "if" branch is present for defensive reasons.
1281   // In the current uses of this interface, it can be replaced with:
1282   // assert(!GCLocker.is_active(), "Can't be called otherwise");
1283   // But I am not placing that assert here to allow future
1284   // generality in invoking this interface.
1285   if (GCLocker::is_active()) {
1286     // A consistency test for GCLocker
1287     assert(GCLocker::needs_gc(), "Should have been set already");
1288     // Skip this foreground collection, instead
1289     // expanding the heap if necessary.
1290     // Need the free list locks for the call to free() in compute_new_size()
1291     compute_new_size();
1292     return;
1293   }
1294   acquire_control_and_collect(full, clear_all_soft_refs);
1295 }
1296 
1297 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1298   GenCollectedHeap* gch = GenCollectedHeap::heap();
1299   unsigned int gc_count = gch->total_full_collections();
1300   if (gc_count == full_gc_count) {
1301     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1302     _full_gc_requested = true;
1303     _full_gc_cause = cause;
1304     CGC_lock->notify();   // nudge CMS thread
1305   } else {
1306     assert(gc_count > full_gc_count, "Error: causal loop");
1307   }
1308 }
1309 
1310 bool CMSCollector::is_external_interruption() {
1311   GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1312   return GCCause::is_user_requested_gc(cause) ||
1313          GCCause::is_serviceability_requested_gc(cause);
1314 }
1315 
1316 void CMSCollector::report_concurrent_mode_interruption() {
1317   if (is_external_interruption()) {
1318     log_debug(gc)("Concurrent mode interrupted");
1319   } else {
1320     log_debug(gc)("Concurrent mode failure");
1321     _gc_tracer_cm->report_concurrent_mode_failure();
1322   }
1323 }
1324 
1325 
1326 // The foreground and background collectors need to coordinate in order
1327 // to make sure that they do not mutually interfere with CMS collections.
1328 // When a background collection is active,
1329 // the foreground collector may need to take over (preempt) and
1330 // synchronously complete an ongoing collection. Depending on the
1331 // frequency of the background collections and the heap usage
1332 // of the application, this preemption can be seldom or frequent.
1333 // There are only certain
1334 // points in the background collection that the "collection-baton"
1335 // can be passed to the foreground collector.
1336 //
1337 // The foreground collector will wait for the baton before
1338 // starting any part of the collection.  The foreground collector
1339 // will only wait at one location.
1340 //
1341 // The background collector will yield the baton before starting a new
1342 // phase of the collection (e.g., before initial marking, marking from roots,
1343 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
1344 // of the loop which switches the phases. The background collector does some
1345 // of the phases (initial mark, final re-mark) with the world stopped.
1346 // Because of locking involved in stopping the world,
1347 // the foreground collector should not block waiting for the background
1348 // collector when it is doing a stop-the-world phase.  The background
1349 // collector will yield the baton at an additional point just before
1350 // it enters a stop-the-world phase.  Once the world is stopped, the
1351 // background collector checks the phase of the collection.  If the
1352 // phase has not changed, it proceeds with the collection.  If the
1353 // phase has changed, it skips that phase of the collection.  See
1354 // the comments on the use of the Heap_lock in collect_in_background().
1355 //
1356 // Variable used in baton passing.
1357 //   _foregroundGCIsActive - Set to true by the foreground collector when
1358 //      it wants the baton.  The foreground clears it when it has finished
1359 //      the collection.
1360 //   _foregroundGCShouldWait - Set to true by the background collector
1361 //        when it is running.  The foreground collector waits while
1362 //      _foregroundGCShouldWait is true.
1363 //  CGC_lock - monitor used to protect access to the above variables
1364 //      and to notify the foreground and background collectors.
1365 //  _collectorState - current state of the CMS collection.
1366 //
1367 // The foreground collector
1368 //   acquires the CGC_lock
1369 //   sets _foregroundGCIsActive
1370 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
1371 //     various locks acquired in preparation for the collection
1372 //     are released so as not to block the background collector
1373 //     that is in the midst of a collection
1374 //   proceeds with the collection
1375 //   clears _foregroundGCIsActive
1376 //   returns
1377 //
1378 // The background collector in a loop iterating on the phases of the
1379 //      collection
1380 //   acquires the CGC_lock
1381 //   sets _foregroundGCShouldWait
1382 //   if _foregroundGCIsActive is set
1383 //     clears _foregroundGCShouldWait, notifies _CGC_lock
1384 //     waits on _CGC_lock for _foregroundGCIsActive to become false
1385 //     and exits the loop.
1386 //   otherwise
1387 //     proceed with that phase of the collection
1388 //     if the phase is a stop-the-world phase,
1389 //       yield the baton once more just before enqueueing
1390 //       the stop-world CMS operation (executed by the VM thread).
1391 //   returns after all phases of the collection are done
1392 //
1393 
1394 void CMSCollector::acquire_control_and_collect(bool full,
1395         bool clear_all_soft_refs) {
1396   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1397   assert(!Thread::current()->is_ConcurrentGC_thread(),
1398          "shouldn't try to acquire control from self!");
1399 
1400   // Start the protocol for acquiring control of the
1401   // collection from the background collector (aka CMS thread).
1402   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1403          "VM thread should have CMS token");
1404   // Remember the possibly interrupted state of an ongoing
1405   // concurrent collection
1406   CollectorState first_state = _collectorState;
1407 
1408   // Signal to a possibly ongoing concurrent collection that
1409   // we want to do a foreground collection.
1410   _foregroundGCIsActive = true;
1411 
1412   // release locks and wait for a notify from the background collector
1413   // releasing the locks in only necessary for phases which
1414   // do yields to improve the granularity of the collection.
1415   assert_lock_strong(bitMapLock());
1416   // We need to lock the Free list lock for the space that we are
1417   // currently collecting.
1418   assert(haveFreelistLocks(), "Must be holding free list locks");
1419   bitMapLock()->unlock();
1420   releaseFreelistLocks();
1421   {
1422     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1423     if (_foregroundGCShouldWait) {
1424       // We are going to be waiting for action for the CMS thread;
1425       // it had better not be gone (for instance at shutdown)!
1426       assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1427              "CMS thread must be running");
1428       // Wait here until the background collector gives us the go-ahead
1429       ConcurrentMarkSweepThread::clear_CMS_flag(
1430         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
1431       // Get a possibly blocked CMS thread going:
1432       //   Note that we set _foregroundGCIsActive true above,
1433       //   without protection of the CGC_lock.
1434       CGC_lock->notify();
1435       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1436              "Possible deadlock");
1437       while (_foregroundGCShouldWait) {
1438         // wait for notification
1439         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1440         // Possibility of delay/starvation here, since CMS token does
1441         // not know to give priority to VM thread? Actually, i think
1442         // there wouldn't be any delay/starvation, but the proof of
1443         // that "fact" (?) appears non-trivial. XXX 20011219YSR
1444       }
1445       ConcurrentMarkSweepThread::set_CMS_flag(
1446         ConcurrentMarkSweepThread::CMS_vm_has_token);
1447     }
1448   }
1449   // The CMS_token is already held.  Get back the other locks.
1450   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1451          "VM thread should have CMS token");
1452   getFreelistLocks();
1453   bitMapLock()->lock_without_safepoint_check();
1454   log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1455                        p2i(Thread::current()), first_state);
1456   log_debug(gc, state)("    gets control with state %d", _collectorState);
1457 
1458   // Inform cms gen if this was due to partial collection failing.
1459   // The CMS gen may use this fact to determine its expansion policy.
1460   GenCollectedHeap* gch = GenCollectedHeap::heap();
1461   if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1462     assert(!_cmsGen->incremental_collection_failed(),
1463            "Should have been noticed, reacted to and cleared");
1464     _cmsGen->set_incremental_collection_failed();
1465   }
1466 
1467   if (first_state > Idling) {
1468     report_concurrent_mode_interruption();
1469   }
1470 
1471   set_did_compact(true);
1472 
1473   // If the collection is being acquired from the background
1474   // collector, there may be references on the discovered
1475   // references lists.  Abandon those references, since some
1476   // of them may have become unreachable after concurrent
1477   // discovery; the STW compacting collector will redo discovery
1478   // more precisely, without being subject to floating garbage.
1479   // Leaving otherwise unreachable references in the discovered
1480   // lists would require special handling.
1481   ref_processor()->disable_discovery();
1482   ref_processor()->abandon_partial_discovery();
1483   ref_processor()->verify_no_references_recorded();
1484 
1485   if (first_state > Idling) {
1486     save_heap_summary();
1487   }
1488 
1489   do_compaction_work(clear_all_soft_refs);
1490 
1491   // Has the GC time limit been exceeded?
1492   size_t max_eden_size = _young_gen->max_eden_size();
1493   GCCause::Cause gc_cause = gch->gc_cause();
1494   size_policy()->check_gc_overhead_limit(_young_gen->used(),
1495                                          _young_gen->eden()->used(),
1496                                          _cmsGen->max_capacity(),
1497                                          max_eden_size,
1498                                          full,
1499                                          gc_cause,
1500                                          gch->collector_policy());
1501 
1502   // Reset the expansion cause, now that we just completed
1503   // a collection cycle.
1504   clear_expansion_cause();
1505   _foregroundGCIsActive = false;
1506   return;
1507 }
1508 
1509 // Resize the tenured generation
1510 // after obtaining the free list locks for the
1511 // two generations.
1512 void CMSCollector::compute_new_size() {
1513   assert_locked_or_safepoint(Heap_lock);
1514   FreelistLocker z(this);
1515   MetaspaceGC::compute_new_size();
1516   _cmsGen->compute_new_size_free_list();
1517 }
1518 
1519 // A work method used by the foreground collector to do
1520 // a mark-sweep-compact.
1521 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1522   GenCollectedHeap* gch = GenCollectedHeap::heap();
1523 
1524   STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1525   gc_timer->register_gc_start();
1526 
1527   SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1528   gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1529 
1530   gch->pre_full_gc_dump(gc_timer);
1531 
1532   GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1533 
1534   // Temporarily widen the span of the weak reference processing to
1535   // the entire heap.
1536   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1537   ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1538   // Temporarily, clear the "is_alive_non_header" field of the
1539   // reference processor.
1540   ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1541   // Temporarily make reference _processing_ single threaded (non-MT).
1542   ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1543   // Temporarily make refs discovery atomic
1544   ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1545   // Temporarily make reference _discovery_ single threaded (non-MT)
1546   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1547 
1548   ref_processor()->set_enqueuing_is_done(false);
1549   ref_processor()->enable_discovery();
1550   ref_processor()->setup_policy(clear_all_soft_refs);
1551   // If an asynchronous collection finishes, the _modUnionTable is
1552   // all clear.  If we are assuming the collection from an asynchronous
1553   // collection, clear the _modUnionTable.
1554   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1555     "_modUnionTable should be clear if the baton was not passed");
1556   _modUnionTable.clear_all();
1557   assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1558     "mod union for klasses should be clear if the baton was passed");
1559   _ct->klass_rem_set()->clear_mod_union();
1560 
1561   // We must adjust the allocation statistics being maintained
1562   // in the free list space. We do so by reading and clearing
1563   // the sweep timer and updating the block flux rate estimates below.
1564   assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1565   if (_inter_sweep_timer.is_active()) {
1566     _inter_sweep_timer.stop();
1567     // Note that we do not use this sample to update the _inter_sweep_estimate.
1568     _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1569                                             _inter_sweep_estimate.padded_average(),
1570                                             _intra_sweep_estimate.padded_average());
1571   }
1572 
1573   GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1574   #ifdef ASSERT
1575     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1576     size_t free_size = cms_space->free();
1577     assert(free_size ==
1578            pointer_delta(cms_space->end(), cms_space->compaction_top())
1579            * HeapWordSize,
1580       "All the free space should be compacted into one chunk at top");
1581     assert(cms_space->dictionary()->total_chunk_size(
1582                                       debug_only(cms_space->freelistLock())) == 0 ||
1583            cms_space->totalSizeInIndexedFreeLists() == 0,
1584       "All the free space should be in a single chunk");
1585     size_t num = cms_space->totalCount();
1586     assert((free_size == 0 && num == 0) ||
1587            (free_size > 0  && (num == 1 || num == 2)),
1588          "There should be at most 2 free chunks after compaction");
1589   #endif // ASSERT
1590   _collectorState = Resetting;
1591   assert(_restart_addr == NULL,
1592          "Should have been NULL'd before baton was passed");
1593   reset_stw();
1594   _cmsGen->reset_after_compaction();
1595   _concurrent_cycles_since_last_unload = 0;
1596 
1597   // Clear any data recorded in the PLAB chunk arrays.
1598   if (_survivor_plab_array != NULL) {
1599     reset_survivor_plab_arrays();
1600   }
1601 
1602   // Adjust the per-size allocation stats for the next epoch.
1603   _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1604   // Restart the "inter sweep timer" for the next epoch.
1605   _inter_sweep_timer.reset();
1606   _inter_sweep_timer.start();
1607 
1608   // No longer a need to do a concurrent collection for Metaspace.
1609   MetaspaceGC::set_should_concurrent_collect(false);
1610 
1611   gch->post_full_gc_dump(gc_timer);
1612 
1613   gc_timer->register_gc_end();
1614 
1615   gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1616 
1617   // For a mark-sweep-compact, compute_new_size() will be called
1618   // in the heap's do_collection() method.
1619 }
1620 
1621 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1622   Log(gc, heap) log;
1623   if (!log.is_trace()) {
1624     return;
1625   }
1626 
1627   ContiguousSpace* eden_space = _young_gen->eden();
1628   ContiguousSpace* from_space = _young_gen->from();
1629   ContiguousSpace* to_space   = _young_gen->to();
1630   // Eden
1631   if (_eden_chunk_array != NULL) {
1632     log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1633               p2i(eden_space->bottom()), p2i(eden_space->top()),
1634               p2i(eden_space->end()), eden_space->capacity());
1635     log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1636               _eden_chunk_index, _eden_chunk_capacity);
1637     for (size_t i = 0; i < _eden_chunk_index; i++) {
1638       log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1639     }
1640   }
1641   // Survivor
1642   if (_survivor_chunk_array != NULL) {
1643     log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1644               p2i(from_space->bottom()), p2i(from_space->top()),
1645               p2i(from_space->end()), from_space->capacity());
1646     log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1647               _survivor_chunk_index, _survivor_chunk_capacity);
1648     for (size_t i = 0; i < _survivor_chunk_index; i++) {
1649       log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1650     }
1651   }
1652 }
1653 
1654 void CMSCollector::getFreelistLocks() const {
1655   // Get locks for all free lists in all generations that this
1656   // collector is responsible for
1657   _cmsGen->freelistLock()->lock_without_safepoint_check();
1658 }
1659 
1660 void CMSCollector::releaseFreelistLocks() const {
1661   // Release locks for all free lists in all generations that this
1662   // collector is responsible for
1663   _cmsGen->freelistLock()->unlock();
1664 }
1665 
1666 bool CMSCollector::haveFreelistLocks() const {
1667   // Check locks for all free lists in all generations that this
1668   // collector is responsible for
1669   assert_lock_strong(_cmsGen->freelistLock());
1670   PRODUCT_ONLY(ShouldNotReachHere());
1671   return true;
1672 }
1673 
1674 // A utility class that is used by the CMS collector to
1675 // temporarily "release" the foreground collector from its
1676 // usual obligation to wait for the background collector to
1677 // complete an ongoing phase before proceeding.
1678 class ReleaseForegroundGC: public StackObj {
1679  private:
1680   CMSCollector* _c;
1681  public:
1682   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1683     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1684     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1685     // allow a potentially blocked foreground collector to proceed
1686     _c->_foregroundGCShouldWait = false;
1687     if (_c->_foregroundGCIsActive) {
1688       CGC_lock->notify();
1689     }
1690     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1691            "Possible deadlock");
1692   }
1693 
1694   ~ReleaseForegroundGC() {
1695     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1696     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1697     _c->_foregroundGCShouldWait = true;
1698   }
1699 };
1700 
1701 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1702   assert(Thread::current()->is_ConcurrentGC_thread(),
1703     "A CMS asynchronous collection is only allowed on a CMS thread.");
1704 
1705   GenCollectedHeap* gch = GenCollectedHeap::heap();
1706   {
1707     bool safepoint_check = Mutex::_no_safepoint_check_flag;
1708     MutexLockerEx hl(Heap_lock, safepoint_check);
1709     FreelistLocker fll(this);
1710     MutexLockerEx x(CGC_lock, safepoint_check);
1711     if (_foregroundGCIsActive) {
1712       // The foreground collector is. Skip this
1713       // background collection.
1714       assert(!_foregroundGCShouldWait, "Should be clear");
1715       return;
1716     } else {
1717       assert(_collectorState == Idling, "Should be idling before start.");
1718       _collectorState = InitialMarking;
1719       register_gc_start(cause);
1720       // Reset the expansion cause, now that we are about to begin
1721       // a new cycle.
1722       clear_expansion_cause();
1723 
1724       // Clear the MetaspaceGC flag since a concurrent collection
1725       // is starting but also clear it after the collection.
1726       MetaspaceGC::set_should_concurrent_collect(false);
1727     }
1728     // Decide if we want to enable class unloading as part of the
1729     // ensuing concurrent GC cycle.
1730     update_should_unload_classes();
1731     _full_gc_requested = false;           // acks all outstanding full gc requests
1732     _full_gc_cause = GCCause::_no_gc;
1733     // Signal that we are about to start a collection
1734     gch->increment_total_full_collections();  // ... starting a collection cycle
1735     _collection_count_start = gch->total_full_collections();
1736   }
1737 
1738   size_t prev_used = _cmsGen->used();
1739 
1740   // The change of the collection state is normally done at this level;
1741   // the exceptions are phases that are executed while the world is
1742   // stopped.  For those phases the change of state is done while the
1743   // world is stopped.  For baton passing purposes this allows the
1744   // background collector to finish the phase and change state atomically.
1745   // The foreground collector cannot wait on a phase that is done
1746   // while the world is stopped because the foreground collector already
1747   // has the world stopped and would deadlock.
1748   while (_collectorState != Idling) {
1749     log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1750                          p2i(Thread::current()), _collectorState);
1751     // The foreground collector
1752     //   holds the Heap_lock throughout its collection.
1753     //   holds the CMS token (but not the lock)
1754     //     except while it is waiting for the background collector to yield.
1755     //
1756     // The foreground collector should be blocked (not for long)
1757     //   if the background collector is about to start a phase
1758     //   executed with world stopped.  If the background
1759     //   collector has already started such a phase, the
1760     //   foreground collector is blocked waiting for the
1761     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
1762     //   are executed in the VM thread.
1763     //
1764     // The locking order is
1765     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
1766     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
1767     //   CMS token  (claimed in
1768     //                stop_world_and_do() -->
1769     //                  safepoint_synchronize() -->
1770     //                    CMSThread::synchronize())
1771 
1772     {
1773       // Check if the FG collector wants us to yield.
1774       CMSTokenSync x(true); // is cms thread
1775       if (waitForForegroundGC()) {
1776         // We yielded to a foreground GC, nothing more to be
1777         // done this round.
1778         assert(_foregroundGCShouldWait == false, "We set it to false in "
1779                "waitForForegroundGC()");
1780         log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1781                              p2i(Thread::current()), _collectorState);
1782         return;
1783       } else {
1784         // The background collector can run but check to see if the
1785         // foreground collector has done a collection while the
1786         // background collector was waiting to get the CGC_lock
1787         // above.  If yes, break so that _foregroundGCShouldWait
1788         // is cleared before returning.
1789         if (_collectorState == Idling) {
1790           break;
1791         }
1792       }
1793     }
1794 
1795     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1796       "should be waiting");
1797 
1798     switch (_collectorState) {
1799       case InitialMarking:
1800         {
1801           ReleaseForegroundGC x(this);
1802           stats().record_cms_begin();
1803           VM_CMS_Initial_Mark initial_mark_op(this);
1804           VMThread::execute(&initial_mark_op);
1805         }
1806         // The collector state may be any legal state at this point
1807         // since the background collector may have yielded to the
1808         // foreground collector.
1809         break;
1810       case Marking:
1811         // initial marking in checkpointRootsInitialWork has been completed
1812         if (markFromRoots()) { // we were successful
1813           assert(_collectorState == Precleaning, "Collector state should "
1814             "have changed");
1815         } else {
1816           assert(_foregroundGCIsActive, "Internal state inconsistency");
1817         }
1818         break;
1819       case Precleaning:
1820         // marking from roots in markFromRoots has been completed
1821         preclean();
1822         assert(_collectorState == AbortablePreclean ||
1823                _collectorState == FinalMarking,
1824                "Collector state should have changed");
1825         break;
1826       case AbortablePreclean:
1827         abortable_preclean();
1828         assert(_collectorState == FinalMarking, "Collector state should "
1829           "have changed");
1830         break;
1831       case FinalMarking:
1832         {
1833           ReleaseForegroundGC x(this);
1834 
1835           VM_CMS_Final_Remark final_remark_op(this);
1836           VMThread::execute(&final_remark_op);
1837         }
1838         assert(_foregroundGCShouldWait, "block post-condition");
1839         break;
1840       case Sweeping:
1841         // final marking in checkpointRootsFinal has been completed
1842         sweep();
1843         assert(_collectorState == Resizing, "Collector state change "
1844           "to Resizing must be done under the free_list_lock");
1845 
1846       case Resizing: {
1847         // Sweeping has been completed...
1848         // At this point the background collection has completed.
1849         // Don't move the call to compute_new_size() down
1850         // into code that might be executed if the background
1851         // collection was preempted.
1852         {
1853           ReleaseForegroundGC x(this);   // unblock FG collection
1854           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
1855           CMSTokenSync        z(true);   // not strictly needed.
1856           if (_collectorState == Resizing) {
1857             compute_new_size();
1858             save_heap_summary();
1859             _collectorState = Resetting;
1860           } else {
1861             assert(_collectorState == Idling, "The state should only change"
1862                    " because the foreground collector has finished the collection");
1863           }
1864         }
1865         break;
1866       }
1867       case Resetting:
1868         // CMS heap resizing has been completed
1869         reset_concurrent();
1870         assert(_collectorState == Idling, "Collector state should "
1871           "have changed");
1872 
1873         MetaspaceGC::set_should_concurrent_collect(false);
1874 
1875         stats().record_cms_end();
1876         // Don't move the concurrent_phases_end() and compute_new_size()
1877         // calls to here because a preempted background collection
1878         // has it's state set to "Resetting".
1879         break;
1880       case Idling:
1881       default:
1882         ShouldNotReachHere();
1883         break;
1884     }
1885     log_debug(gc, state)("  Thread " INTPTR_FORMAT " done - next CMS state %d",
1886                          p2i(Thread::current()), _collectorState);
1887     assert(_foregroundGCShouldWait, "block post-condition");
1888   }
1889 
1890   // Should this be in gc_epilogue?
1891   collector_policy()->counters()->update_counters();
1892 
1893   {
1894     // Clear _foregroundGCShouldWait and, in the event that the
1895     // foreground collector is waiting, notify it, before
1896     // returning.
1897     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1898     _foregroundGCShouldWait = false;
1899     if (_foregroundGCIsActive) {
1900       CGC_lock->notify();
1901     }
1902     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1903            "Possible deadlock");
1904   }
1905   log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1906                        p2i(Thread::current()), _collectorState);
1907   log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K("  SIZE_FORMAT "K)",
1908                      prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1909 }
1910 
1911 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1912   _cms_start_registered = true;
1913   _gc_timer_cm->register_gc_start();
1914   _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1915 }
1916 
1917 void CMSCollector::register_gc_end() {
1918   if (_cms_start_registered) {
1919     report_heap_summary(GCWhen::AfterGC);
1920 
1921     _gc_timer_cm->register_gc_end();
1922     _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1923     _cms_start_registered = false;
1924   }
1925 }
1926 
1927 void CMSCollector::save_heap_summary() {
1928   GenCollectedHeap* gch = GenCollectedHeap::heap();
1929   _last_heap_summary = gch->create_heap_summary();
1930   _last_metaspace_summary = gch->create_metaspace_summary();
1931 }
1932 
1933 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1934   _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1935   _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1936 }
1937 
1938 bool CMSCollector::waitForForegroundGC() {
1939   bool res = false;
1940   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1941          "CMS thread should have CMS token");
1942   // Block the foreground collector until the
1943   // background collectors decides whether to
1944   // yield.
1945   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1946   _foregroundGCShouldWait = true;
1947   if (_foregroundGCIsActive) {
1948     // The background collector yields to the
1949     // foreground collector and returns a value
1950     // indicating that it has yielded.  The foreground
1951     // collector can proceed.
1952     res = true;
1953     _foregroundGCShouldWait = false;
1954     ConcurrentMarkSweepThread::clear_CMS_flag(
1955       ConcurrentMarkSweepThread::CMS_cms_has_token);
1956     ConcurrentMarkSweepThread::set_CMS_flag(
1957       ConcurrentMarkSweepThread::CMS_cms_wants_token);
1958     // Get a possibly blocked foreground thread going
1959     CGC_lock->notify();
1960     log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1961                          p2i(Thread::current()), _collectorState);
1962     while (_foregroundGCIsActive) {
1963       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1964     }
1965     ConcurrentMarkSweepThread::set_CMS_flag(
1966       ConcurrentMarkSweepThread::CMS_cms_has_token);
1967     ConcurrentMarkSweepThread::clear_CMS_flag(
1968       ConcurrentMarkSweepThread::CMS_cms_wants_token);
1969   }
1970   log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1971                        p2i(Thread::current()), _collectorState);
1972   return res;
1973 }
1974 
1975 // Because of the need to lock the free lists and other structures in
1976 // the collector, common to all the generations that the collector is
1977 // collecting, we need the gc_prologues of individual CMS generations
1978 // delegate to their collector. It may have been simpler had the
1979 // current infrastructure allowed one to call a prologue on a
1980 // collector. In the absence of that we have the generation's
1981 // prologue delegate to the collector, which delegates back
1982 // some "local" work to a worker method in the individual generations
1983 // that it's responsible for collecting, while itself doing any
1984 // work common to all generations it's responsible for. A similar
1985 // comment applies to the  gc_epilogue()'s.
1986 // The role of the variable _between_prologue_and_epilogue is to
1987 // enforce the invocation protocol.
1988 void CMSCollector::gc_prologue(bool full) {
1989   // Call gc_prologue_work() for the CMSGen
1990   // we are responsible for.
1991 
1992   // The following locking discipline assumes that we are only called
1993   // when the world is stopped.
1994   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
1995 
1996   // The CMSCollector prologue must call the gc_prologues for the
1997   // "generations" that it's responsible
1998   // for.
1999 
2000   assert(   Thread::current()->is_VM_thread()
2001          || (   CMSScavengeBeforeRemark
2002              && Thread::current()->is_ConcurrentGC_thread()),
2003          "Incorrect thread type for prologue execution");
2004 
2005   if (_between_prologue_and_epilogue) {
2006     // We have already been invoked; this is a gc_prologue delegation
2007     // from yet another CMS generation that we are responsible for, just
2008     // ignore it since all relevant work has already been done.
2009     return;
2010   }
2011 
2012   // set a bit saying prologue has been called; cleared in epilogue
2013   _between_prologue_and_epilogue = true;
2014   // Claim locks for common data structures, then call gc_prologue_work()
2015   // for each CMSGen.
2016 
2017   getFreelistLocks();   // gets free list locks on constituent spaces
2018   bitMapLock()->lock_without_safepoint_check();
2019 
2020   // Should call gc_prologue_work() for all cms gens we are responsible for
2021   bool duringMarking =    _collectorState >= Marking
2022                          && _collectorState < Sweeping;
2023 
2024   // The young collections clear the modified oops state, which tells if
2025   // there are any modified oops in the class. The remark phase also needs
2026   // that information. Tell the young collection to save the union of all
2027   // modified klasses.
2028   if (duringMarking) {
2029     _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2030   }
2031 
2032   bool registerClosure = duringMarking;
2033 
2034   _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2035 
2036   if (!full) {
2037     stats().record_gc0_begin();
2038   }
2039 }
2040 
2041 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2042 
2043   _capacity_at_prologue = capacity();
2044   _used_at_prologue = used();
2045 
2046   // We enable promotion tracking so that card-scanning can recognize
2047   // which objects have been promoted during this GC and skip them.
2048   for (uint i = 0; i < ParallelGCThreads; i++) {
2049     _par_gc_thread_states[i]->promo.startTrackingPromotions();
2050   }
2051 
2052   // Delegate to CMScollector which knows how to coordinate between
2053   // this and any other CMS generations that it is responsible for
2054   // collecting.
2055   collector()->gc_prologue(full);
2056 }
2057 
2058 // This is a "private" interface for use by this generation's CMSCollector.
2059 // Not to be called directly by any other entity (for instance,
2060 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2061 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2062   bool registerClosure, ModUnionClosure* modUnionClosure) {
2063   assert(!incremental_collection_failed(), "Shouldn't be set yet");
2064   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2065     "Should be NULL");
2066   if (registerClosure) {
2067     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2068   }
2069   cmsSpace()->gc_prologue();
2070   // Clear stat counters
2071   NOT_PRODUCT(
2072     assert(_numObjectsPromoted == 0, "check");
2073     assert(_numWordsPromoted   == 0, "check");
2074     log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2075                                  _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2076     _numObjectsAllocated = 0;
2077     _numWordsAllocated   = 0;
2078   )
2079 }
2080 
2081 void CMSCollector::gc_epilogue(bool full) {
2082   // The following locking discipline assumes that we are only called
2083   // when the world is stopped.
2084   assert(SafepointSynchronize::is_at_safepoint(),
2085          "world is stopped assumption");
2086 
2087   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2088   // if linear allocation blocks need to be appropriately marked to allow the
2089   // the blocks to be parsable. We also check here whether we need to nudge the
2090   // CMS collector thread to start a new cycle (if it's not already active).
2091   assert(   Thread::current()->is_VM_thread()
2092          || (   CMSScavengeBeforeRemark
2093              && Thread::current()->is_ConcurrentGC_thread()),
2094          "Incorrect thread type for epilogue execution");
2095 
2096   if (!_between_prologue_and_epilogue) {
2097     // We have already been invoked; this is a gc_epilogue delegation
2098     // from yet another CMS generation that we are responsible for, just
2099     // ignore it since all relevant work has already been done.
2100     return;
2101   }
2102   assert(haveFreelistLocks(), "must have freelist locks");
2103   assert_lock_strong(bitMapLock());
2104 
2105   _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2106 
2107   _cmsGen->gc_epilogue_work(full);
2108 
2109   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2110     // in case sampling was not already enabled, enable it
2111     _start_sampling = true;
2112   }
2113   // reset _eden_chunk_array so sampling starts afresh
2114   _eden_chunk_index = 0;
2115 
2116   size_t cms_used   = _cmsGen->cmsSpace()->used();
2117 
2118   // update performance counters - this uses a special version of
2119   // update_counters() that allows the utilization to be passed as a
2120   // parameter, avoiding multiple calls to used().
2121   //
2122   _cmsGen->update_counters(cms_used);
2123 
2124   bitMapLock()->unlock();
2125   releaseFreelistLocks();
2126 
2127   if (!CleanChunkPoolAsync) {
2128     Chunk::clean_chunk_pool();
2129   }
2130 
2131   set_did_compact(false);
2132   _between_prologue_and_epilogue = false;  // ready for next cycle
2133 }
2134 
2135 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2136   collector()->gc_epilogue(full);
2137 
2138   // When using ParNew, promotion tracking should have already been
2139   // disabled. However, the prologue (which enables promotion
2140   // tracking) and epilogue are called irrespective of the type of
2141   // GC. So they will also be called before and after Full GCs, during
2142   // which promotion tracking will not be explicitly disabled. So,
2143   // it's safer to also disable it here too (to be symmetric with
2144   // enabling it in the prologue).
2145   for (uint i = 0; i < ParallelGCThreads; i++) {
2146     _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2147   }
2148 }
2149 
2150 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2151   assert(!incremental_collection_failed(), "Should have been cleared");
2152   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2153   cmsSpace()->gc_epilogue();
2154     // Print stat counters
2155   NOT_PRODUCT(
2156     assert(_numObjectsAllocated == 0, "check");
2157     assert(_numWordsAllocated == 0, "check");
2158     log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2159                                      _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2160     _numObjectsPromoted = 0;
2161     _numWordsPromoted   = 0;
2162   )
2163 
2164   // Call down the chain in contiguous_available needs the freelistLock
2165   // so print this out before releasing the freeListLock.
2166   log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2167 }
2168 
2169 #ifndef PRODUCT
2170 bool CMSCollector::have_cms_token() {
2171   Thread* thr = Thread::current();
2172   if (thr->is_VM_thread()) {
2173     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2174   } else if (thr->is_ConcurrentGC_thread()) {
2175     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2176   } else if (thr->is_GC_task_thread()) {
2177     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2178            ParGCRareEvent_lock->owned_by_self();
2179   }
2180   return false;
2181 }
2182 
2183 // Check reachability of the given heap address in CMS generation,
2184 // treating all other generations as roots.
2185 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2186   // We could "guarantee" below, rather than assert, but I'll
2187   // leave these as "asserts" so that an adventurous debugger
2188   // could try this in the product build provided some subset of
2189   // the conditions were met, provided they were interested in the
2190   // results and knew that the computation below wouldn't interfere
2191   // with other concurrent computations mutating the structures
2192   // being read or written.
2193   assert(SafepointSynchronize::is_at_safepoint(),
2194          "Else mutations in object graph will make answer suspect");
2195   assert(have_cms_token(), "Should hold cms token");
2196   assert(haveFreelistLocks(), "must hold free list locks");
2197   assert_lock_strong(bitMapLock());
2198 
2199   // Clear the marking bit map array before starting, but, just
2200   // for kicks, first report if the given address is already marked
2201   tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2202                 _markBitMap.isMarked(addr) ? "" : " not");
2203 
2204   if (verify_after_remark()) {
2205     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2206     bool result = verification_mark_bm()->isMarked(addr);
2207     tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2208                   result ? "IS" : "is NOT");
2209     return result;
2210   } else {
2211     tty->print_cr("Could not compute result");
2212     return false;
2213   }
2214 }
2215 #endif
2216 
2217 void
2218 CMSCollector::print_on_error(outputStream* st) {
2219   CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2220   if (collector != NULL) {
2221     CMSBitMap* bitmap = &collector->_markBitMap;
2222     st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2223     bitmap->print_on_error(st, " Bits: ");
2224 
2225     st->cr();
2226 
2227     CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2228     st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2229     mut_bitmap->print_on_error(st, " Bits: ");
2230   }
2231 }
2232 
2233 ////////////////////////////////////////////////////////
2234 // CMS Verification Support
2235 ////////////////////////////////////////////////////////
2236 // Following the remark phase, the following invariant
2237 // should hold -- each object in the CMS heap which is
2238 // marked in markBitMap() should be marked in the verification_mark_bm().
2239 
2240 class VerifyMarkedClosure: public BitMapClosure {
2241   CMSBitMap* _marks;
2242   bool       _failed;
2243 
2244  public:
2245   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2246 
2247   bool do_bit(size_t offset) {
2248     HeapWord* addr = _marks->offsetToHeapWord(offset);
2249     if (!_marks->isMarked(addr)) {
2250       Log(gc, verify) log;
2251       ResourceMark rm;
2252       LogStream ls(log.error());
2253       oop(addr)->print_on(&ls);
2254       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2255       _failed = true;
2256     }
2257     return true;
2258   }
2259 
2260   bool failed() { return _failed; }
2261 };
2262 
2263 bool CMSCollector::verify_after_remark() {
2264   GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2265   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2266   static bool init = false;
2267 
2268   assert(SafepointSynchronize::is_at_safepoint(),
2269          "Else mutations in object graph will make answer suspect");
2270   assert(have_cms_token(),
2271          "Else there may be mutual interference in use of "
2272          " verification data structures");
2273   assert(_collectorState > Marking && _collectorState <= Sweeping,
2274          "Else marking info checked here may be obsolete");
2275   assert(haveFreelistLocks(), "must hold free list locks");
2276   assert_lock_strong(bitMapLock());
2277 
2278 
2279   // Allocate marking bit map if not already allocated
2280   if (!init) { // first time
2281     if (!verification_mark_bm()->allocate(_span)) {
2282       return false;
2283     }
2284     init = true;
2285   }
2286 
2287   assert(verification_mark_stack()->isEmpty(), "Should be empty");
2288 
2289   // Turn off refs discovery -- so we will be tracing through refs.
2290   // This is as intended, because by this time
2291   // GC must already have cleared any refs that need to be cleared,
2292   // and traced those that need to be marked; moreover,
2293   // the marking done here is not going to interfere in any
2294   // way with the marking information used by GC.
2295   NoRefDiscovery no_discovery(ref_processor());
2296 
2297 #if defined(COMPILER2) || INCLUDE_JVMCI
2298   DerivedPointerTableDeactivate dpt_deact;
2299 #endif
2300 
2301   // Clear any marks from a previous round
2302   verification_mark_bm()->clear_all();
2303   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2304   verify_work_stacks_empty();
2305 
2306   GenCollectedHeap* gch = GenCollectedHeap::heap();
2307   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2308   // Update the saved marks which may affect the root scans.
2309   gch->save_marks();
2310 
2311   if (CMSRemarkVerifyVariant == 1) {
2312     // In this first variant of verification, we complete
2313     // all marking, then check if the new marks-vector is
2314     // a subset of the CMS marks-vector.
2315     verify_after_remark_work_1();
2316   } else {
2317     guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2318     // In this second variant of verification, we flag an error
2319     // (i.e. an object reachable in the new marks-vector not reachable
2320     // in the CMS marks-vector) immediately, also indicating the
2321     // identify of an object (A) that references the unmarked object (B) --
2322     // presumably, a mutation to A failed to be picked up by preclean/remark?
2323     verify_after_remark_work_2();
2324   }
2325 
2326   return true;
2327 }
2328 
2329 void CMSCollector::verify_after_remark_work_1() {
2330   ResourceMark rm;
2331   HandleMark  hm;
2332   GenCollectedHeap* gch = GenCollectedHeap::heap();
2333 
2334   // Get a clear set of claim bits for the roots processing to work with.
2335   ClassLoaderDataGraph::clear_claimed_marks();
2336 
2337   // Mark from roots one level into CMS
2338   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2339   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2340 
2341   {
2342     StrongRootsScope srs(1);
2343 
2344     gch->cms_process_roots(&srs,
2345                            true,   // young gen as roots
2346                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2347                            should_unload_classes(),
2348                            &notOlder,
2349                            NULL);
2350   }
2351 
2352   // Now mark from the roots
2353   MarkFromRootsClosure markFromRootsClosure(this, _span,
2354     verification_mark_bm(), verification_mark_stack(),
2355     false /* don't yield */, true /* verifying */);
2356   assert(_restart_addr == NULL, "Expected pre-condition");
2357   verification_mark_bm()->iterate(&markFromRootsClosure);
2358   while (_restart_addr != NULL) {
2359     // Deal with stack overflow: by restarting at the indicated
2360     // address.
2361     HeapWord* ra = _restart_addr;
2362     markFromRootsClosure.reset(ra);
2363     _restart_addr = NULL;
2364     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2365   }
2366   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2367   verify_work_stacks_empty();
2368 
2369   // Marking completed -- now verify that each bit marked in
2370   // verification_mark_bm() is also marked in markBitMap(); flag all
2371   // errors by printing corresponding objects.
2372   VerifyMarkedClosure vcl(markBitMap());
2373   verification_mark_bm()->iterate(&vcl);
2374   if (vcl.failed()) {
2375     Log(gc, verify) log;
2376     log.error("Failed marking verification after remark");
2377     ResourceMark rm;
2378     LogStream ls(log.error());
2379     gch->print_on(&ls);
2380     fatal("CMS: failed marking verification after remark");
2381   }
2382 }
2383 
2384 class VerifyKlassOopsKlassClosure : public KlassClosure {
2385   class VerifyKlassOopsClosure : public OopClosure {
2386     CMSBitMap* _bitmap;
2387    public:
2388     VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2389     void do_oop(oop* p)       { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2390     void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2391   } _oop_closure;
2392  public:
2393   VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2394   void do_klass(Klass* k) {
2395     k->oops_do(&_oop_closure);
2396   }
2397 };
2398 
2399 void CMSCollector::verify_after_remark_work_2() {
2400   ResourceMark rm;
2401   HandleMark  hm;
2402   GenCollectedHeap* gch = GenCollectedHeap::heap();
2403 
2404   // Get a clear set of claim bits for the roots processing to work with.
2405   ClassLoaderDataGraph::clear_claimed_marks();
2406 
2407   // Mark from roots one level into CMS
2408   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2409                                      markBitMap());
2410   CLDToOopClosure cld_closure(&notOlder, true);
2411 
2412   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2413 
2414   {
2415     StrongRootsScope srs(1);
2416 
2417     gch->cms_process_roots(&srs,
2418                            true,   // young gen as roots
2419                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2420                            should_unload_classes(),
2421                            &notOlder,
2422                            &cld_closure);
2423   }
2424 
2425   // Now mark from the roots
2426   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2427     verification_mark_bm(), markBitMap(), verification_mark_stack());
2428   assert(_restart_addr == NULL, "Expected pre-condition");
2429   verification_mark_bm()->iterate(&markFromRootsClosure);
2430   while (_restart_addr != NULL) {
2431     // Deal with stack overflow: by restarting at the indicated
2432     // address.
2433     HeapWord* ra = _restart_addr;
2434     markFromRootsClosure.reset(ra);
2435     _restart_addr = NULL;
2436     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2437   }
2438   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2439   verify_work_stacks_empty();
2440 
2441   VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2442   ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2443 
2444   // Marking completed -- now verify that each bit marked in
2445   // verification_mark_bm() is also marked in markBitMap(); flag all
2446   // errors by printing corresponding objects.
2447   VerifyMarkedClosure vcl(markBitMap());
2448   verification_mark_bm()->iterate(&vcl);
2449   assert(!vcl.failed(), "Else verification above should not have succeeded");
2450 }
2451 
2452 void ConcurrentMarkSweepGeneration::save_marks() {
2453   // delegate to CMS space
2454   cmsSpace()->save_marks();
2455 }
2456 
2457 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2458   return cmsSpace()->no_allocs_since_save_marks();
2459 }
2460 
2461 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
2462                                                                 \
2463 void ConcurrentMarkSweepGeneration::                            \
2464 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
2465   cl->set_generation(this);                                     \
2466   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
2467   cl->reset_generation();                                       \
2468   save_marks();                                                 \
2469 }
2470 
2471 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2472 
2473 void
2474 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2475   if (freelistLock()->owned_by_self()) {
2476     Generation::oop_iterate(cl);
2477   } else {
2478     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2479     Generation::oop_iterate(cl);
2480   }
2481 }
2482 
2483 void
2484 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2485   if (freelistLock()->owned_by_self()) {
2486     Generation::object_iterate(cl);
2487   } else {
2488     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2489     Generation::object_iterate(cl);
2490   }
2491 }
2492 
2493 void
2494 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2495   if (freelistLock()->owned_by_self()) {
2496     Generation::safe_object_iterate(cl);
2497   } else {
2498     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2499     Generation::safe_object_iterate(cl);
2500   }
2501 }
2502 
2503 void
2504 ConcurrentMarkSweepGeneration::post_compact() {
2505 }
2506 
2507 void
2508 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2509   // Fix the linear allocation blocks to look like free blocks.
2510 
2511   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2512   // are not called when the heap is verified during universe initialization and
2513   // at vm shutdown.
2514   if (freelistLock()->owned_by_self()) {
2515     cmsSpace()->prepare_for_verify();
2516   } else {
2517     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2518     cmsSpace()->prepare_for_verify();
2519   }
2520 }
2521 
2522 void
2523 ConcurrentMarkSweepGeneration::verify() {
2524   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2525   // are not called when the heap is verified during universe initialization and
2526   // at vm shutdown.
2527   if (freelistLock()->owned_by_self()) {
2528     cmsSpace()->verify();
2529   } else {
2530     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2531     cmsSpace()->verify();
2532   }
2533 }
2534 
2535 void CMSCollector::verify() {
2536   _cmsGen->verify();
2537 }
2538 
2539 #ifndef PRODUCT
2540 bool CMSCollector::overflow_list_is_empty() const {
2541   assert(_num_par_pushes >= 0, "Inconsistency");
2542   if (_overflow_list == NULL) {
2543     assert(_num_par_pushes == 0, "Inconsistency");
2544   }
2545   return _overflow_list == NULL;
2546 }
2547 
2548 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2549 // merely consolidate assertion checks that appear to occur together frequently.
2550 void CMSCollector::verify_work_stacks_empty() const {
2551   assert(_markStack.isEmpty(), "Marking stack should be empty");
2552   assert(overflow_list_is_empty(), "Overflow list should be empty");
2553 }
2554 
2555 void CMSCollector::verify_overflow_empty() const {
2556   assert(overflow_list_is_empty(), "Overflow list should be empty");
2557   assert(no_preserved_marks(), "No preserved marks");
2558 }
2559 #endif // PRODUCT
2560 
2561 // Decide if we want to enable class unloading as part of the
2562 // ensuing concurrent GC cycle. We will collect and
2563 // unload classes if it's the case that:
2564 //  (a) class unloading is enabled at the command line, and
2565 //  (b) old gen is getting really full
2566 // NOTE: Provided there is no change in the state of the heap between
2567 // calls to this method, it should have idempotent results. Moreover,
2568 // its results should be monotonically increasing (i.e. going from 0 to 1,
2569 // but not 1 to 0) between successive calls between which the heap was
2570 // not collected. For the implementation below, it must thus rely on
2571 // the property that concurrent_cycles_since_last_unload()
2572 // will not decrease unless a collection cycle happened and that
2573 // _cmsGen->is_too_full() are
2574 // themselves also monotonic in that sense. See check_monotonicity()
2575 // below.
2576 void CMSCollector::update_should_unload_classes() {
2577   _should_unload_classes = false;
2578   if (CMSClassUnloadingEnabled) {
2579     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2580                               CMSClassUnloadingMaxInterval)
2581                            || _cmsGen->is_too_full();
2582   }
2583 }
2584 
2585 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2586   bool res = should_concurrent_collect();
2587   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2588   return res;
2589 }
2590 
2591 void CMSCollector::setup_cms_unloading_and_verification_state() {
2592   const  bool should_verify =   VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2593                              || VerifyBeforeExit;
2594   const  int  rso           =   GenCollectedHeap::SO_AllCodeCache;
2595 
2596   // We set the proper root for this CMS cycle here.
2597   if (should_unload_classes()) {   // Should unload classes this cycle
2598     remove_root_scanning_option(rso);  // Shrink the root set appropriately
2599     set_verifying(should_verify);    // Set verification state for this cycle
2600     return;                            // Nothing else needs to be done at this time
2601   }
2602 
2603   // Not unloading classes this cycle
2604   assert(!should_unload_classes(), "Inconsistency!");
2605 
2606   // If we are not unloading classes then add SO_AllCodeCache to root
2607   // scanning options.
2608   add_root_scanning_option(rso);
2609 
2610   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2611     set_verifying(true);
2612   } else if (verifying() && !should_verify) {
2613     // We were verifying, but some verification flags got disabled.
2614     set_verifying(false);
2615     // Exclude symbols, strings and code cache elements from root scanning to
2616     // reduce IM and RM pauses.
2617     remove_root_scanning_option(rso);
2618   }
2619 }
2620 
2621 
2622 #ifndef PRODUCT
2623 HeapWord* CMSCollector::block_start(const void* p) const {
2624   const HeapWord* addr = (HeapWord*)p;
2625   if (_span.contains(p)) {
2626     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2627       return _cmsGen->cmsSpace()->block_start(p);
2628     }
2629   }
2630   return NULL;
2631 }
2632 #endif
2633 
2634 HeapWord*
2635 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2636                                                    bool   tlab,
2637                                                    bool   parallel) {
2638   CMSSynchronousYieldRequest yr;
2639   assert(!tlab, "Can't deal with TLAB allocation");
2640   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2641   expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2642   if (GCExpandToAllocateDelayMillis > 0) {
2643     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2644   }
2645   return have_lock_and_allocate(word_size, tlab);
2646 }
2647 
2648 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2649     size_t bytes,
2650     size_t expand_bytes,
2651     CMSExpansionCause::Cause cause)
2652 {
2653 
2654   bool success = expand(bytes, expand_bytes);
2655 
2656   // remember why we expanded; this information is used
2657   // by shouldConcurrentCollect() when making decisions on whether to start
2658   // a new CMS cycle.
2659   if (success) {
2660     set_expansion_cause(cause);
2661     log_trace(gc)("Expanded CMS gen for %s",  CMSExpansionCause::to_string(cause));
2662   }
2663 }
2664 
2665 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2666   HeapWord* res = NULL;
2667   MutexLocker x(ParGCRareEvent_lock);
2668   while (true) {
2669     // Expansion by some other thread might make alloc OK now:
2670     res = ps->lab.alloc(word_sz);
2671     if (res != NULL) return res;
2672     // If there's not enough expansion space available, give up.
2673     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2674       return NULL;
2675     }
2676     // Otherwise, we try expansion.
2677     expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2678     // Now go around the loop and try alloc again;
2679     // A competing par_promote might beat us to the expansion space,
2680     // so we may go around the loop again if promotion fails again.
2681     if (GCExpandToAllocateDelayMillis > 0) {
2682       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2683     }
2684   }
2685 }
2686 
2687 
2688 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2689   PromotionInfo* promo) {
2690   MutexLocker x(ParGCRareEvent_lock);
2691   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2692   while (true) {
2693     // Expansion by some other thread might make alloc OK now:
2694     if (promo->ensure_spooling_space()) {
2695       assert(promo->has_spooling_space(),
2696              "Post-condition of successful ensure_spooling_space()");
2697       return true;
2698     }
2699     // If there's not enough expansion space available, give up.
2700     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2701       return false;
2702     }
2703     // Otherwise, we try expansion.
2704     expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2705     // Now go around the loop and try alloc again;
2706     // A competing allocation might beat us to the expansion space,
2707     // so we may go around the loop again if allocation fails again.
2708     if (GCExpandToAllocateDelayMillis > 0) {
2709       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2710     }
2711   }
2712 }
2713 
2714 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2715   // Only shrink if a compaction was done so that all the free space
2716   // in the generation is in a contiguous block at the end.
2717   if (did_compact()) {
2718     CardGeneration::shrink(bytes);
2719   }
2720 }
2721 
2722 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2723   assert_locked_or_safepoint(Heap_lock);
2724 }
2725 
2726 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2727   assert_locked_or_safepoint(Heap_lock);
2728   assert_lock_strong(freelistLock());
2729   log_trace(gc)("Shrinking of CMS not yet implemented");
2730   return;
2731 }
2732 
2733 
2734 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2735 // phases.
2736 class CMSPhaseAccounting: public StackObj {
2737  public:
2738   CMSPhaseAccounting(CMSCollector *collector,
2739                      const char *title);
2740   ~CMSPhaseAccounting();
2741 
2742  private:
2743   CMSCollector *_collector;
2744   const char *_title;
2745   GCTraceConcTime(Info, gc) _trace_time;
2746 
2747  public:
2748   // Not MT-safe; so do not pass around these StackObj's
2749   // where they may be accessed by other threads.
2750   double wallclock_millis() {
2751     return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2752   }
2753 };
2754 
2755 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2756                                        const char *title) :
2757   _collector(collector), _title(title), _trace_time(title) {
2758 
2759   _collector->resetYields();
2760   _collector->resetTimer();
2761   _collector->startTimer();
2762   _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2763 }
2764 
2765 CMSPhaseAccounting::~CMSPhaseAccounting() {
2766   _collector->gc_timer_cm()->register_gc_concurrent_end();
2767   _collector->stopTimer();
2768   log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2769   log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2770 }
2771 
2772 // CMS work
2773 
2774 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2775 class CMSParMarkTask : public AbstractGangTask {
2776  protected:
2777   CMSCollector*     _collector;
2778   uint              _n_workers;
2779   CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2780       AbstractGangTask(name),
2781       _collector(collector),
2782       _n_workers(n_workers) {}
2783   // Work method in support of parallel rescan ... of young gen spaces
2784   void do_young_space_rescan(OopsInGenClosure* cl,
2785                              ContiguousSpace* space,
2786                              HeapWord** chunk_array, size_t chunk_top);
2787   void work_on_young_gen_roots(OopsInGenClosure* cl);
2788 };
2789 
2790 // Parallel initial mark task
2791 class CMSParInitialMarkTask: public CMSParMarkTask {
2792   StrongRootsScope* _strong_roots_scope;
2793  public:
2794   CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2795       CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2796       _strong_roots_scope(strong_roots_scope) {}
2797   void work(uint worker_id);
2798 };
2799 
2800 // Checkpoint the roots into this generation from outside
2801 // this generation. [Note this initial checkpoint need only
2802 // be approximate -- we'll do a catch up phase subsequently.]
2803 void CMSCollector::checkpointRootsInitial() {
2804   assert(_collectorState == InitialMarking, "Wrong collector state");
2805   check_correct_thread_executing();
2806   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2807 
2808   save_heap_summary();
2809   report_heap_summary(GCWhen::BeforeGC);
2810 
2811   ReferenceProcessor* rp = ref_processor();
2812   assert(_restart_addr == NULL, "Control point invariant");
2813   {
2814     // acquire locks for subsequent manipulations
2815     MutexLockerEx x(bitMapLock(),
2816                     Mutex::_no_safepoint_check_flag);
2817     checkpointRootsInitialWork();
2818     // enable ("weak") refs discovery
2819     rp->enable_discovery();
2820     _collectorState = Marking;
2821   }
2822 }
2823 
2824 void CMSCollector::checkpointRootsInitialWork() {
2825   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2826   assert(_collectorState == InitialMarking, "just checking");
2827 
2828   // Already have locks.
2829   assert_lock_strong(bitMapLock());
2830   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2831 
2832   // Setup the verification and class unloading state for this
2833   // CMS collection cycle.
2834   setup_cms_unloading_and_verification_state();
2835 
2836   GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2837 
2838   // Reset all the PLAB chunk arrays if necessary.
2839   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2840     reset_survivor_plab_arrays();
2841   }
2842 
2843   ResourceMark rm;
2844   HandleMark  hm;
2845 
2846   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2847   GenCollectedHeap* gch = GenCollectedHeap::heap();
2848 
2849   verify_work_stacks_empty();
2850   verify_overflow_empty();
2851 
2852   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2853   // Update the saved marks which may affect the root scans.
2854   gch->save_marks();
2855 
2856   // weak reference processing has not started yet.
2857   ref_processor()->set_enqueuing_is_done(false);
2858 
2859   // Need to remember all newly created CLDs,
2860   // so that we can guarantee that the remark finds them.
2861   ClassLoaderDataGraph::remember_new_clds(true);
2862 
2863   // Whenever a CLD is found, it will be claimed before proceeding to mark
2864   // the klasses. The claimed marks need to be cleared before marking starts.
2865   ClassLoaderDataGraph::clear_claimed_marks();
2866 
2867   print_eden_and_survivor_chunk_arrays();
2868 
2869   {
2870 #if defined(COMPILER2) || INCLUDE_JVMCI
2871     DerivedPointerTableDeactivate dpt_deact;
2872 #endif
2873     if (CMSParallelInitialMarkEnabled) {
2874       // The parallel version.
2875       WorkGang* workers = gch->workers();
2876       assert(workers != NULL, "Need parallel worker threads.");
2877       uint n_workers = workers->active_workers();
2878 
2879       StrongRootsScope srs(n_workers);
2880 
2881       CMSParInitialMarkTask tsk(this, &srs, n_workers);
2882       initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2883       // If the total workers is greater than 1, then multiple workers
2884       // may be used at some time and the initialization has been set
2885       // such that the single threaded path cannot be used.
2886       if (workers->total_workers() > 1) {
2887         workers->run_task(&tsk);
2888       } else {
2889         tsk.work(0);
2890       }
2891     } else {
2892       // The serial version.
2893       CLDToOopClosure cld_closure(&notOlder, true);
2894       gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2895 
2896       StrongRootsScope srs(1);
2897 
2898       gch->cms_process_roots(&srs,
2899                              true,   // young gen as roots
2900                              GenCollectedHeap::ScanningOption(roots_scanning_options()),
2901                              should_unload_classes(),
2902                              &notOlder,
2903                              &cld_closure);
2904     }
2905   }
2906 
2907   // Clear mod-union table; it will be dirtied in the prologue of
2908   // CMS generation per each young generation collection.
2909 
2910   assert(_modUnionTable.isAllClear(),
2911        "Was cleared in most recent final checkpoint phase"
2912        " or no bits are set in the gc_prologue before the start of the next "
2913        "subsequent marking phase.");
2914 
2915   assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
2916 
2917   // Save the end of the used_region of the constituent generations
2918   // to be used to limit the extent of sweep in each generation.
2919   save_sweep_limits();
2920   verify_overflow_empty();
2921 }
2922 
2923 bool CMSCollector::markFromRoots() {
2924   // we might be tempted to assert that:
2925   // assert(!SafepointSynchronize::is_at_safepoint(),
2926   //        "inconsistent argument?");
2927   // However that wouldn't be right, because it's possible that
2928   // a safepoint is indeed in progress as a young generation
2929   // stop-the-world GC happens even as we mark in this generation.
2930   assert(_collectorState == Marking, "inconsistent state?");
2931   check_correct_thread_executing();
2932   verify_overflow_empty();
2933 
2934   // Weak ref discovery note: We may be discovering weak
2935   // refs in this generation concurrent (but interleaved) with
2936   // weak ref discovery by the young generation collector.
2937 
2938   CMSTokenSyncWithLocks ts(true, bitMapLock());
2939   GCTraceCPUTime tcpu;
2940   CMSPhaseAccounting pa(this, "Concurrent Mark");
2941   bool res = markFromRootsWork();
2942   if (res) {
2943     _collectorState = Precleaning;
2944   } else { // We failed and a foreground collection wants to take over
2945     assert(_foregroundGCIsActive, "internal state inconsistency");
2946     assert(_restart_addr == NULL,  "foreground will restart from scratch");
2947     log_debug(gc)("bailing out to foreground collection");
2948   }
2949   verify_overflow_empty();
2950   return res;
2951 }
2952 
2953 bool CMSCollector::markFromRootsWork() {
2954   // iterate over marked bits in bit map, doing a full scan and mark
2955   // from these roots using the following algorithm:
2956   // . if oop is to the right of the current scan pointer,
2957   //   mark corresponding bit (we'll process it later)
2958   // . else (oop is to left of current scan pointer)
2959   //   push oop on marking stack
2960   // . drain the marking stack
2961 
2962   // Note that when we do a marking step we need to hold the
2963   // bit map lock -- recall that direct allocation (by mutators)
2964   // and promotion (by the young generation collector) is also
2965   // marking the bit map. [the so-called allocate live policy.]
2966   // Because the implementation of bit map marking is not
2967   // robust wrt simultaneous marking of bits in the same word,
2968   // we need to make sure that there is no such interference
2969   // between concurrent such updates.
2970 
2971   // already have locks
2972   assert_lock_strong(bitMapLock());
2973 
2974   verify_work_stacks_empty();
2975   verify_overflow_empty();
2976   bool result = false;
2977   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2978     result = do_marking_mt();
2979   } else {
2980     result = do_marking_st();
2981   }
2982   return result;
2983 }
2984 
2985 // Forward decl
2986 class CMSConcMarkingTask;
2987 
2988 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2989   CMSCollector*       _collector;
2990   CMSConcMarkingTask* _task;
2991  public:
2992   virtual void yield();
2993 
2994   // "n_threads" is the number of threads to be terminated.
2995   // "queue_set" is a set of work queues of other threads.
2996   // "collector" is the CMS collector associated with this task terminator.
2997   // "yield" indicates whether we need the gang as a whole to yield.
2998   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
2999     ParallelTaskTerminator(n_threads, queue_set),
3000     _collector(collector) { }
3001 
3002   void set_task(CMSConcMarkingTask* task) {
3003     _task = task;
3004   }
3005 };
3006 
3007 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3008   CMSConcMarkingTask* _task;
3009  public:
3010   bool should_exit_termination();
3011   void set_task(CMSConcMarkingTask* task) {
3012     _task = task;
3013   }
3014 };
3015 
3016 // MT Concurrent Marking Task
3017 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3018   CMSCollector*             _collector;
3019   uint                      _n_workers;      // requested/desired # workers
3020   bool                      _result;
3021   CompactibleFreeListSpace* _cms_space;
3022   char                      _pad_front[64];   // padding to ...
3023   HeapWord* volatile        _global_finger;   // ... avoid sharing cache line
3024   char                      _pad_back[64];
3025   HeapWord*                 _restart_addr;
3026 
3027   //  Exposed here for yielding support
3028   Mutex* const _bit_map_lock;
3029 
3030   // The per thread work queues, available here for stealing
3031   OopTaskQueueSet*  _task_queues;
3032 
3033   // Termination (and yielding) support
3034   CMSConcMarkingTerminator _term;
3035   CMSConcMarkingTerminatorTerminator _term_term;
3036 
3037  public:
3038   CMSConcMarkingTask(CMSCollector* collector,
3039                  CompactibleFreeListSpace* cms_space,
3040                  YieldingFlexibleWorkGang* workers,
3041                  OopTaskQueueSet* task_queues):
3042     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3043     _collector(collector),
3044     _cms_space(cms_space),
3045     _n_workers(0), _result(true),
3046     _task_queues(task_queues),
3047     _term(_n_workers, task_queues, _collector),
3048     _bit_map_lock(collector->bitMapLock())
3049   {
3050     _requested_size = _n_workers;
3051     _term.set_task(this);
3052     _term_term.set_task(this);
3053     _restart_addr = _global_finger = _cms_space->bottom();
3054   }
3055 
3056 
3057   OopTaskQueueSet* task_queues()  { return _task_queues; }
3058 
3059   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3060 
3061   HeapWord* volatile* global_finger_addr() { return &_global_finger; }
3062 
3063   CMSConcMarkingTerminator* terminator() { return &_term; }
3064 
3065   virtual void set_for_termination(uint active_workers) {
3066     terminator()->reset_for_reuse(active_workers);
3067   }
3068 
3069   void work(uint worker_id);
3070   bool should_yield() {
3071     return    ConcurrentMarkSweepThread::should_yield()
3072            && !_collector->foregroundGCIsActive();
3073   }
3074 
3075   virtual void coordinator_yield();  // stuff done by coordinator
3076   bool result() { return _result; }
3077 
3078   void reset(HeapWord* ra) {
3079     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
3080     _restart_addr = _global_finger = ra;
3081     _term.reset_for_reuse();
3082   }
3083 
3084   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3085                                            OopTaskQueue* work_q);
3086 
3087  private:
3088   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3089   void do_work_steal(int i);
3090   void bump_global_finger(HeapWord* f);
3091 };
3092 
3093 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3094   assert(_task != NULL, "Error");
3095   return _task->yielding();
3096   // Note that we do not need the disjunct || _task->should_yield() above
3097   // because we want terminating threads to yield only if the task
3098   // is already in the midst of yielding, which happens only after at least one
3099   // thread has yielded.
3100 }
3101 
3102 void CMSConcMarkingTerminator::yield() {
3103   if (_task->should_yield()) {
3104     _task->yield();
3105   } else {
3106     ParallelTaskTerminator::yield();
3107   }
3108 }
3109 
3110 ////////////////////////////////////////////////////////////////
3111 // Concurrent Marking Algorithm Sketch
3112 ////////////////////////////////////////////////////////////////
3113 // Until all tasks exhausted (both spaces):
3114 // -- claim next available chunk
3115 // -- bump global finger via CAS
3116 // -- find first object that starts in this chunk
3117 //    and start scanning bitmap from that position
3118 // -- scan marked objects for oops
3119 // -- CAS-mark target, and if successful:
3120 //    . if target oop is above global finger (volatile read)
3121 //      nothing to do
3122 //    . if target oop is in chunk and above local finger
3123 //        then nothing to do
3124 //    . else push on work-queue
3125 // -- Deal with possible overflow issues:
3126 //    . local work-queue overflow causes stuff to be pushed on
3127 //      global (common) overflow queue
3128 //    . always first empty local work queue
3129 //    . then get a batch of oops from global work queue if any
3130 //    . then do work stealing
3131 // -- When all tasks claimed (both spaces)
3132 //    and local work queue empty,
3133 //    then in a loop do:
3134 //    . check global overflow stack; steal a batch of oops and trace
3135 //    . try to steal from other threads oif GOS is empty
3136 //    . if neither is available, offer termination
3137 // -- Terminate and return result
3138 //
3139 void CMSConcMarkingTask::work(uint worker_id) {
3140   elapsedTimer _timer;
3141   ResourceMark rm;
3142   HandleMark hm;
3143 
3144   DEBUG_ONLY(_collector->verify_overflow_empty();)
3145 
3146   // Before we begin work, our work queue should be empty
3147   assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3148   // Scan the bitmap covering _cms_space, tracing through grey objects.
3149   _timer.start();
3150   do_scan_and_mark(worker_id, _cms_space);
3151   _timer.stop();
3152   log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3153 
3154   // ... do work stealing
3155   _timer.reset();
3156   _timer.start();
3157   do_work_steal(worker_id);
3158   _timer.stop();
3159   log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3160   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3161   assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3162   // Note that under the current task protocol, the
3163   // following assertion is true even of the spaces
3164   // expanded since the completion of the concurrent
3165   // marking. XXX This will likely change under a strict
3166   // ABORT semantics.
3167   // After perm removal the comparison was changed to
3168   // greater than or equal to from strictly greater than.
3169   // Before perm removal the highest address sweep would
3170   // have been at the end of perm gen but now is at the
3171   // end of the tenured gen.
3172   assert(_global_finger >=  _cms_space->end(),
3173          "All tasks have been completed");
3174   DEBUG_ONLY(_collector->verify_overflow_empty();)
3175 }
3176 
3177 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3178   HeapWord* read = _global_finger;
3179   HeapWord* cur  = read;
3180   while (f > read) {
3181     cur = read;
3182     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3183     if (cur == read) {
3184       // our cas succeeded
3185       assert(_global_finger >= f, "protocol consistency");
3186       break;
3187     }
3188   }
3189 }
3190 
3191 // This is really inefficient, and should be redone by
3192 // using (not yet available) block-read and -write interfaces to the
3193 // stack and the work_queue. XXX FIX ME !!!
3194 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3195                                                       OopTaskQueue* work_q) {
3196   // Fast lock-free check
3197   if (ovflw_stk->length() == 0) {
3198     return false;
3199   }
3200   assert(work_q->size() == 0, "Shouldn't steal");
3201   MutexLockerEx ml(ovflw_stk->par_lock(),
3202                    Mutex::_no_safepoint_check_flag);
3203   // Grab up to 1/4 the size of the work queue
3204   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3205                     (size_t)ParGCDesiredObjsFromOverflowList);
3206   num = MIN2(num, ovflw_stk->length());
3207   for (int i = (int) num; i > 0; i--) {
3208     oop cur = ovflw_stk->pop();
3209     assert(cur != NULL, "Counted wrong?");
3210     work_q->push(cur);
3211   }
3212   return num > 0;
3213 }
3214 
3215 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3216   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3217   int n_tasks = pst->n_tasks();
3218   // We allow that there may be no tasks to do here because
3219   // we are restarting after a stack overflow.
3220   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3221   uint nth_task = 0;
3222 
3223   HeapWord* aligned_start = sp->bottom();
3224   if (sp->used_region().contains(_restart_addr)) {
3225     // Align down to a card boundary for the start of 0th task
3226     // for this space.
3227     aligned_start = align_down(_restart_addr, CardTableModRefBS::card_size);
3228   }
3229 
3230   size_t chunk_size = sp->marking_task_size();
3231   while (!pst->is_task_claimed(/* reference */ nth_task)) {
3232     // Having claimed the nth task in this space,
3233     // compute the chunk that it corresponds to:
3234     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3235                                aligned_start + (nth_task+1)*chunk_size);
3236     // Try and bump the global finger via a CAS;
3237     // note that we need to do the global finger bump
3238     // _before_ taking the intersection below, because
3239     // the task corresponding to that region will be
3240     // deemed done even if the used_region() expands
3241     // because of allocation -- as it almost certainly will
3242     // during start-up while the threads yield in the
3243     // closure below.
3244     HeapWord* finger = span.end();
3245     bump_global_finger(finger);   // atomically
3246     // There are null tasks here corresponding to chunks
3247     // beyond the "top" address of the space.
3248     span = span.intersection(sp->used_region());
3249     if (!span.is_empty()) {  // Non-null task
3250       HeapWord* prev_obj;
3251       assert(!span.contains(_restart_addr) || nth_task == 0,
3252              "Inconsistency");
3253       if (nth_task == 0) {
3254         // For the 0th task, we'll not need to compute a block_start.
3255         if (span.contains(_restart_addr)) {
3256           // In the case of a restart because of stack overflow,
3257           // we might additionally skip a chunk prefix.
3258           prev_obj = _restart_addr;
3259         } else {
3260           prev_obj = span.start();
3261         }
3262       } else {
3263         // We want to skip the first object because
3264         // the protocol is to scan any object in its entirety
3265         // that _starts_ in this span; a fortiori, any
3266         // object starting in an earlier span is scanned
3267         // as part of an earlier claimed task.
3268         // Below we use the "careful" version of block_start
3269         // so we do not try to navigate uninitialized objects.
3270         prev_obj = sp->block_start_careful(span.start());
3271         // Below we use a variant of block_size that uses the
3272         // Printezis bits to avoid waiting for allocated
3273         // objects to become initialized/parsable.
3274         while (prev_obj < span.start()) {
3275           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3276           if (sz > 0) {
3277             prev_obj += sz;
3278           } else {
3279             // In this case we may end up doing a bit of redundant
3280             // scanning, but that appears unavoidable, short of
3281             // locking the free list locks; see bug 6324141.
3282             break;
3283           }
3284         }
3285       }
3286       if (prev_obj < span.end()) {
3287         MemRegion my_span = MemRegion(prev_obj, span.end());
3288         // Do the marking work within a non-empty span --
3289         // the last argument to the constructor indicates whether the
3290         // iteration should be incremental with periodic yields.
3291         ParMarkFromRootsClosure cl(this, _collector, my_span,
3292                                    &_collector->_markBitMap,
3293                                    work_queue(i),
3294                                    &_collector->_markStack);
3295         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3296       } // else nothing to do for this task
3297     }   // else nothing to do for this task
3298   }
3299   // We'd be tempted to assert here that since there are no
3300   // more tasks left to claim in this space, the global_finger
3301   // must exceed space->top() and a fortiori space->end(). However,
3302   // that would not quite be correct because the bumping of
3303   // global_finger occurs strictly after the claiming of a task,
3304   // so by the time we reach here the global finger may not yet
3305   // have been bumped up by the thread that claimed the last
3306   // task.
3307   pst->all_tasks_completed();
3308 }
3309 
3310 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3311  private:
3312   CMSCollector* _collector;
3313   CMSConcMarkingTask* _task;
3314   MemRegion     _span;
3315   CMSBitMap*    _bit_map;
3316   CMSMarkStack* _overflow_stack;
3317   OopTaskQueue* _work_queue;
3318  protected:
3319   DO_OOP_WORK_DEFN
3320  public:
3321   ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3322                         CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3323     MetadataAwareOopClosure(collector->ref_processor()),
3324     _collector(collector),
3325     _task(task),
3326     _span(collector->_span),
3327     _work_queue(work_queue),
3328     _bit_map(bit_map),
3329     _overflow_stack(overflow_stack)
3330   { }
3331   virtual void do_oop(oop* p);
3332   virtual void do_oop(narrowOop* p);
3333 
3334   void trim_queue(size_t max);
3335   void handle_stack_overflow(HeapWord* lost);
3336   void do_yield_check() {
3337     if (_task->should_yield()) {
3338       _task->yield();
3339     }
3340   }
3341 };
3342 
3343 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3344 
3345 // Grey object scanning during work stealing phase --
3346 // the salient assumption here is that any references
3347 // that are in these stolen objects being scanned must
3348 // already have been initialized (else they would not have
3349 // been published), so we do not need to check for
3350 // uninitialized objects before pushing here.
3351 void ParConcMarkingClosure::do_oop(oop obj) {
3352   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3353   HeapWord* addr = (HeapWord*)obj;
3354   // Check if oop points into the CMS generation
3355   // and is not marked
3356   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3357     // a white object ...
3358     // If we manage to "claim" the object, by being the
3359     // first thread to mark it, then we push it on our
3360     // marking stack
3361     if (_bit_map->par_mark(addr)) {     // ... now grey
3362       // push on work queue (grey set)
3363       bool simulate_overflow = false;
3364       NOT_PRODUCT(
3365         if (CMSMarkStackOverflowALot &&
3366             _collector->simulate_overflow()) {
3367           // simulate a stack overflow
3368           simulate_overflow = true;
3369         }
3370       )
3371       if (simulate_overflow ||
3372           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3373         // stack overflow
3374         log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3375         // We cannot assert that the overflow stack is full because
3376         // it may have been emptied since.
3377         assert(simulate_overflow ||
3378                _work_queue->size() == _work_queue->max_elems(),
3379               "Else push should have succeeded");
3380         handle_stack_overflow(addr);
3381       }
3382     } // Else, some other thread got there first
3383     do_yield_check();
3384   }
3385 }
3386 
3387 void ParConcMarkingClosure::do_oop(oop* p)       { ParConcMarkingClosure::do_oop_work(p); }
3388 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3389 
3390 void ParConcMarkingClosure::trim_queue(size_t max) {
3391   while (_work_queue->size() > max) {
3392     oop new_oop;
3393     if (_work_queue->pop_local(new_oop)) {
3394       assert(oopDesc::is_oop(new_oop), "Should be an oop");
3395       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3396       assert(_span.contains((HeapWord*)new_oop), "Not in span");
3397       new_oop->oop_iterate(this);  // do_oop() above
3398       do_yield_check();
3399     }
3400   }
3401 }
3402 
3403 // Upon stack overflow, we discard (part of) the stack,
3404 // remembering the least address amongst those discarded
3405 // in CMSCollector's _restart_address.
3406 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3407   // We need to do this under a mutex to prevent other
3408   // workers from interfering with the work done below.
3409   MutexLockerEx ml(_overflow_stack->par_lock(),
3410                    Mutex::_no_safepoint_check_flag);
3411   // Remember the least grey address discarded
3412   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3413   _collector->lower_restart_addr(ra);
3414   _overflow_stack->reset();  // discard stack contents
3415   _overflow_stack->expand(); // expand the stack if possible
3416 }
3417 
3418 
3419 void CMSConcMarkingTask::do_work_steal(int i) {
3420   OopTaskQueue* work_q = work_queue(i);
3421   oop obj_to_scan;
3422   CMSBitMap* bm = &(_collector->_markBitMap);
3423   CMSMarkStack* ovflw = &(_collector->_markStack);
3424   int* seed = _collector->hash_seed(i);
3425   ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3426   while (true) {
3427     cl.trim_queue(0);
3428     assert(work_q->size() == 0, "Should have been emptied above");
3429     if (get_work_from_overflow_stack(ovflw, work_q)) {
3430       // Can't assert below because the work obtained from the
3431       // overflow stack may already have been stolen from us.
3432       // assert(work_q->size() > 0, "Work from overflow stack");
3433       continue;
3434     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3435       assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3436       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3437       obj_to_scan->oop_iterate(&cl);
3438     } else if (terminator()->offer_termination(&_term_term)) {
3439       assert(work_q->size() == 0, "Impossible!");
3440       break;
3441     } else if (yielding() || should_yield()) {
3442       yield();
3443     }
3444   }
3445 }
3446 
3447 // This is run by the CMS (coordinator) thread.
3448 void CMSConcMarkingTask::coordinator_yield() {
3449   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3450          "CMS thread should hold CMS token");
3451   // First give up the locks, then yield, then re-lock
3452   // We should probably use a constructor/destructor idiom to
3453   // do this unlock/lock or modify the MutexUnlocker class to
3454   // serve our purpose. XXX
3455   assert_lock_strong(_bit_map_lock);
3456   _bit_map_lock->unlock();
3457   ConcurrentMarkSweepThread::desynchronize(true);
3458   _collector->stopTimer();
3459   _collector->incrementYields();
3460 
3461   // It is possible for whichever thread initiated the yield request
3462   // not to get a chance to wake up and take the bitmap lock between
3463   // this thread releasing it and reacquiring it. So, while the
3464   // should_yield() flag is on, let's sleep for a bit to give the
3465   // other thread a chance to wake up. The limit imposed on the number
3466   // of iterations is defensive, to avoid any unforseen circumstances
3467   // putting us into an infinite loop. Since it's always been this
3468   // (coordinator_yield()) method that was observed to cause the
3469   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3470   // which is by default non-zero. For the other seven methods that
3471   // also perform the yield operation, as are using a different
3472   // parameter (CMSYieldSleepCount) which is by default zero. This way we
3473   // can enable the sleeping for those methods too, if necessary.
3474   // See 6442774.
3475   //
3476   // We really need to reconsider the synchronization between the GC
3477   // thread and the yield-requesting threads in the future and we
3478   // should really use wait/notify, which is the recommended
3479   // way of doing this type of interaction. Additionally, we should
3480   // consolidate the eight methods that do the yield operation and they
3481   // are almost identical into one for better maintainability and
3482   // readability. See 6445193.
3483   //
3484   // Tony 2006.06.29
3485   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3486                    ConcurrentMarkSweepThread::should_yield() &&
3487                    !CMSCollector::foregroundGCIsActive(); ++i) {
3488     os::sleep(Thread::current(), 1, false);
3489   }
3490 
3491   ConcurrentMarkSweepThread::synchronize(true);
3492   _bit_map_lock->lock_without_safepoint_check();
3493   _collector->startTimer();
3494 }
3495 
3496 bool CMSCollector::do_marking_mt() {
3497   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3498   uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3499                                                                   conc_workers()->active_workers(),
3500                                                                   Threads::number_of_non_daemon_threads());
3501   num_workers = conc_workers()->update_active_workers(num_workers);
3502   log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3503 
3504   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
3505 
3506   CMSConcMarkingTask tsk(this,
3507                          cms_space,
3508                          conc_workers(),
3509                          task_queues());
3510 
3511   // Since the actual number of workers we get may be different
3512   // from the number we requested above, do we need to do anything different
3513   // below? In particular, may be we need to subclass the SequantialSubTasksDone
3514   // class?? XXX
3515   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3516 
3517   // Refs discovery is already non-atomic.
3518   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3519   assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3520   conc_workers()->start_task(&tsk);
3521   while (tsk.yielded()) {
3522     tsk.coordinator_yield();
3523     conc_workers()->continue_task(&tsk);
3524   }
3525   // If the task was aborted, _restart_addr will be non-NULL
3526   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3527   while (_restart_addr != NULL) {
3528     // XXX For now we do not make use of ABORTED state and have not
3529     // yet implemented the right abort semantics (even in the original
3530     // single-threaded CMS case). That needs some more investigation
3531     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3532     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3533     // If _restart_addr is non-NULL, a marking stack overflow
3534     // occurred; we need to do a fresh marking iteration from the
3535     // indicated restart address.
3536     if (_foregroundGCIsActive) {
3537       // We may be running into repeated stack overflows, having
3538       // reached the limit of the stack size, while making very
3539       // slow forward progress. It may be best to bail out and
3540       // let the foreground collector do its job.
3541       // Clear _restart_addr, so that foreground GC
3542       // works from scratch. This avoids the headache of
3543       // a "rescan" which would otherwise be needed because
3544       // of the dirty mod union table & card table.
3545       _restart_addr = NULL;
3546       return false;
3547     }
3548     // Adjust the task to restart from _restart_addr
3549     tsk.reset(_restart_addr);
3550     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3551                   _restart_addr);
3552     _restart_addr = NULL;
3553     // Get the workers going again
3554     conc_workers()->start_task(&tsk);
3555     while (tsk.yielded()) {
3556       tsk.coordinator_yield();
3557       conc_workers()->continue_task(&tsk);
3558     }
3559   }
3560   assert(tsk.completed(), "Inconsistency");
3561   assert(tsk.result() == true, "Inconsistency");
3562   return true;
3563 }
3564 
3565 bool CMSCollector::do_marking_st() {
3566   ResourceMark rm;
3567   HandleMark   hm;
3568 
3569   // Temporarily make refs discovery single threaded (non-MT)
3570   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3571   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3572     &_markStack, CMSYield);
3573   // the last argument to iterate indicates whether the iteration
3574   // should be incremental with periodic yields.
3575   _markBitMap.iterate(&markFromRootsClosure);
3576   // If _restart_addr is non-NULL, a marking stack overflow
3577   // occurred; we need to do a fresh iteration from the
3578   // indicated restart address.
3579   while (_restart_addr != NULL) {
3580     if (_foregroundGCIsActive) {
3581       // We may be running into repeated stack overflows, having
3582       // reached the limit of the stack size, while making very
3583       // slow forward progress. It may be best to bail out and
3584       // let the foreground collector do its job.
3585       // Clear _restart_addr, so that foreground GC
3586       // works from scratch. This avoids the headache of
3587       // a "rescan" which would otherwise be needed because
3588       // of the dirty mod union table & card table.
3589       _restart_addr = NULL;
3590       return false;  // indicating failure to complete marking
3591     }
3592     // Deal with stack overflow:
3593     // we restart marking from _restart_addr
3594     HeapWord* ra = _restart_addr;
3595     markFromRootsClosure.reset(ra);
3596     _restart_addr = NULL;
3597     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3598   }
3599   return true;
3600 }
3601 
3602 void CMSCollector::preclean() {
3603   check_correct_thread_executing();
3604   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3605   verify_work_stacks_empty();
3606   verify_overflow_empty();
3607   _abort_preclean = false;
3608   if (CMSPrecleaningEnabled) {
3609     if (!CMSEdenChunksRecordAlways) {
3610       _eden_chunk_index = 0;
3611     }
3612     size_t used = get_eden_used();
3613     size_t capacity = get_eden_capacity();
3614     // Don't start sampling unless we will get sufficiently
3615     // many samples.
3616     if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3617                 * CMSScheduleRemarkEdenPenetration)) {
3618       _start_sampling = true;
3619     } else {
3620       _start_sampling = false;
3621     }
3622     GCTraceCPUTime tcpu;
3623     CMSPhaseAccounting pa(this, "Concurrent Preclean");
3624     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3625   }
3626   CMSTokenSync x(true); // is cms thread
3627   if (CMSPrecleaningEnabled) {
3628     sample_eden();
3629     _collectorState = AbortablePreclean;
3630   } else {
3631     _collectorState = FinalMarking;
3632   }
3633   verify_work_stacks_empty();
3634   verify_overflow_empty();
3635 }
3636 
3637 // Try and schedule the remark such that young gen
3638 // occupancy is CMSScheduleRemarkEdenPenetration %.
3639 void CMSCollector::abortable_preclean() {
3640   check_correct_thread_executing();
3641   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
3642   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3643 
3644   // If Eden's current occupancy is below this threshold,
3645   // immediately schedule the remark; else preclean
3646   // past the next scavenge in an effort to
3647   // schedule the pause as described above. By choosing
3648   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3649   // we will never do an actual abortable preclean cycle.
3650   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3651     GCTraceCPUTime tcpu;
3652     CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3653     // We need more smarts in the abortable preclean
3654     // loop below to deal with cases where allocation
3655     // in young gen is very very slow, and our precleaning
3656     // is running a losing race against a horde of
3657     // mutators intent on flooding us with CMS updates
3658     // (dirty cards).
3659     // One, admittedly dumb, strategy is to give up
3660     // after a certain number of abortable precleaning loops
3661     // or after a certain maximum time. We want to make
3662     // this smarter in the next iteration.
3663     // XXX FIX ME!!! YSR
3664     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3665     while (!(should_abort_preclean() ||
3666              ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3667       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3668       cumworkdone += workdone;
3669       loops++;
3670       // Voluntarily terminate abortable preclean phase if we have
3671       // been at it for too long.
3672       if ((CMSMaxAbortablePrecleanLoops != 0) &&
3673           loops >= CMSMaxAbortablePrecleanLoops) {
3674         log_debug(gc)(" CMS: abort preclean due to loops ");
3675         break;
3676       }
3677       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3678         log_debug(gc)(" CMS: abort preclean due to time ");
3679         break;
3680       }
3681       // If we are doing little work each iteration, we should
3682       // take a short break.
3683       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3684         // Sleep for some time, waiting for work to accumulate
3685         stopTimer();
3686         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3687         startTimer();
3688         waited++;
3689       }
3690     }
3691     log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3692                                loops, waited, cumworkdone);
3693   }
3694   CMSTokenSync x(true); // is cms thread
3695   if (_collectorState != Idling) {
3696     assert(_collectorState == AbortablePreclean,
3697            "Spontaneous state transition?");
3698     _collectorState = FinalMarking;
3699   } // Else, a foreground collection completed this CMS cycle.
3700   return;
3701 }
3702 
3703 // Respond to an Eden sampling opportunity
3704 void CMSCollector::sample_eden() {
3705   // Make sure a young gc cannot sneak in between our
3706   // reading and recording of a sample.
3707   assert(Thread::current()->is_ConcurrentGC_thread(),
3708          "Only the cms thread may collect Eden samples");
3709   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3710          "Should collect samples while holding CMS token");
3711   if (!_start_sampling) {
3712     return;
3713   }
3714   // When CMSEdenChunksRecordAlways is true, the eden chunk array
3715   // is populated by the young generation.
3716   if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3717     if (_eden_chunk_index < _eden_chunk_capacity) {
3718       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
3719       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3720              "Unexpected state of Eden");
3721       // We'd like to check that what we just sampled is an oop-start address;
3722       // however, we cannot do that here since the object may not yet have been
3723       // initialized. So we'll instead do the check when we _use_ this sample
3724       // later.
3725       if (_eden_chunk_index == 0 ||
3726           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3727                          _eden_chunk_array[_eden_chunk_index-1])
3728            >= CMSSamplingGrain)) {
3729         _eden_chunk_index++;  // commit sample
3730       }
3731     }
3732   }
3733   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3734     size_t used = get_eden_used();
3735     size_t capacity = get_eden_capacity();
3736     assert(used <= capacity, "Unexpected state of Eden");
3737     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3738       _abort_preclean = true;
3739     }
3740   }
3741 }
3742 
3743 
3744 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3745   assert(_collectorState == Precleaning ||
3746          _collectorState == AbortablePreclean, "incorrect state");
3747   ResourceMark rm;
3748   HandleMark   hm;
3749 
3750   // Precleaning is currently not MT but the reference processor
3751   // may be set for MT.  Disable it temporarily here.
3752   ReferenceProcessor* rp = ref_processor();
3753   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3754 
3755   // Do one pass of scrubbing the discovered reference lists
3756   // to remove any reference objects with strongly-reachable
3757   // referents.
3758   if (clean_refs) {
3759     CMSPrecleanRefsYieldClosure yield_cl(this);
3760     assert(rp->span().equals(_span), "Spans should be equal");
3761     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3762                                    &_markStack, true /* preclean */);
3763     CMSDrainMarkingStackClosure complete_trace(this,
3764                                    _span, &_markBitMap, &_markStack,
3765                                    &keep_alive, true /* preclean */);
3766 
3767     // We don't want this step to interfere with a young
3768     // collection because we don't want to take CPU
3769     // or memory bandwidth away from the young GC threads
3770     // (which may be as many as there are CPUs).
3771     // Note that we don't need to protect ourselves from
3772     // interference with mutators because they can't
3773     // manipulate the discovered reference lists nor affect
3774     // the computed reachability of the referents, the
3775     // only properties manipulated by the precleaning
3776     // of these reference lists.
3777     stopTimer();
3778     CMSTokenSyncWithLocks x(true /* is cms thread */,
3779                             bitMapLock());
3780     startTimer();
3781     sample_eden();
3782 
3783     // The following will yield to allow foreground
3784     // collection to proceed promptly. XXX YSR:
3785     // The code in this method may need further
3786     // tweaking for better performance and some restructuring
3787     // for cleaner interfaces.
3788     GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3789     rp->preclean_discovered_references(
3790           rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3791           gc_timer);
3792   }
3793 
3794   if (clean_survivor) {  // preclean the active survivor space(s)
3795     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3796                              &_markBitMap, &_modUnionTable,
3797                              &_markStack, true /* precleaning phase */);
3798     stopTimer();
3799     CMSTokenSyncWithLocks ts(true /* is cms thread */,
3800                              bitMapLock());
3801     startTimer();
3802     unsigned int before_count =
3803       GenCollectedHeap::heap()->total_collections();
3804     SurvivorSpacePrecleanClosure
3805       sss_cl(this, _span, &_markBitMap, &_markStack,
3806              &pam_cl, before_count, CMSYield);
3807     _young_gen->from()->object_iterate_careful(&sss_cl);
3808     _young_gen->to()->object_iterate_careful(&sss_cl);
3809   }
3810   MarkRefsIntoAndScanClosure
3811     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3812              &_markStack, this, CMSYield,
3813              true /* precleaning phase */);
3814   // CAUTION: The following closure has persistent state that may need to
3815   // be reset upon a decrease in the sequence of addresses it
3816   // processes.
3817   ScanMarkedObjectsAgainCarefullyClosure
3818     smoac_cl(this, _span,
3819       &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3820 
3821   // Preclean dirty cards in ModUnionTable and CardTable using
3822   // appropriate convergence criterion;
3823   // repeat CMSPrecleanIter times unless we find that
3824   // we are losing.
3825   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3826   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3827          "Bad convergence multiplier");
3828   assert(CMSPrecleanThreshold >= 100,
3829          "Unreasonably low CMSPrecleanThreshold");
3830 
3831   size_t numIter, cumNumCards, lastNumCards, curNumCards;
3832   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3833        numIter < CMSPrecleanIter;
3834        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3835     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
3836     log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3837     // Either there are very few dirty cards, so re-mark
3838     // pause will be small anyway, or our pre-cleaning isn't
3839     // that much faster than the rate at which cards are being
3840     // dirtied, so we might as well stop and re-mark since
3841     // precleaning won't improve our re-mark time by much.
3842     if (curNumCards <= CMSPrecleanThreshold ||
3843         (numIter > 0 &&
3844          (curNumCards * CMSPrecleanDenominator >
3845          lastNumCards * CMSPrecleanNumerator))) {
3846       numIter++;
3847       cumNumCards += curNumCards;
3848       break;
3849     }
3850   }
3851 
3852   preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
3853 
3854   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3855   cumNumCards += curNumCards;
3856   log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3857                              curNumCards, cumNumCards, numIter);
3858   return cumNumCards;   // as a measure of useful work done
3859 }
3860 
3861 // PRECLEANING NOTES:
3862 // Precleaning involves:
3863 // . reading the bits of the modUnionTable and clearing the set bits.
3864 // . For the cards corresponding to the set bits, we scan the
3865 //   objects on those cards. This means we need the free_list_lock
3866 //   so that we can safely iterate over the CMS space when scanning
3867 //   for oops.
3868 // . When we scan the objects, we'll be both reading and setting
3869 //   marks in the marking bit map, so we'll need the marking bit map.
3870 // . For protecting _collector_state transitions, we take the CGC_lock.
3871 //   Note that any races in the reading of of card table entries by the
3872 //   CMS thread on the one hand and the clearing of those entries by the
3873 //   VM thread or the setting of those entries by the mutator threads on the
3874 //   other are quite benign. However, for efficiency it makes sense to keep
3875 //   the VM thread from racing with the CMS thread while the latter is
3876 //   dirty card info to the modUnionTable. We therefore also use the
3877 //   CGC_lock to protect the reading of the card table and the mod union
3878 //   table by the CM thread.
3879 // . We run concurrently with mutator updates, so scanning
3880 //   needs to be done carefully  -- we should not try to scan
3881 //   potentially uninitialized objects.
3882 //
3883 // Locking strategy: While holding the CGC_lock, we scan over and
3884 // reset a maximal dirty range of the mod union / card tables, then lock
3885 // the free_list_lock and bitmap lock to do a full marking, then
3886 // release these locks; and repeat the cycle. This allows for a
3887 // certain amount of fairness in the sharing of these locks between
3888 // the CMS collector on the one hand, and the VM thread and the
3889 // mutators on the other.
3890 
3891 // NOTE: preclean_mod_union_table() and preclean_card_table()
3892 // further below are largely identical; if you need to modify
3893 // one of these methods, please check the other method too.
3894 
3895 size_t CMSCollector::preclean_mod_union_table(
3896   ConcurrentMarkSweepGeneration* old_gen,
3897   ScanMarkedObjectsAgainCarefullyClosure* cl) {
3898   verify_work_stacks_empty();
3899   verify_overflow_empty();
3900 
3901   // strategy: starting with the first card, accumulate contiguous
3902   // ranges of dirty cards; clear these cards, then scan the region
3903   // covered by these cards.
3904 
3905   // Since all of the MUT is committed ahead, we can just use
3906   // that, in case the generations expand while we are precleaning.
3907   // It might also be fine to just use the committed part of the
3908   // generation, but we might potentially miss cards when the
3909   // generation is rapidly expanding while we are in the midst
3910   // of precleaning.
3911   HeapWord* startAddr = old_gen->reserved().start();
3912   HeapWord* endAddr   = old_gen->reserved().end();
3913 
3914   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
3915 
3916   size_t numDirtyCards, cumNumDirtyCards;
3917   HeapWord *nextAddr, *lastAddr;
3918   for (cumNumDirtyCards = numDirtyCards = 0,
3919        nextAddr = lastAddr = startAddr;
3920        nextAddr < endAddr;
3921        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3922 
3923     ResourceMark rm;
3924     HandleMark   hm;
3925 
3926     MemRegion dirtyRegion;
3927     {
3928       stopTimer();
3929       // Potential yield point
3930       CMSTokenSync ts(true);
3931       startTimer();
3932       sample_eden();
3933       // Get dirty region starting at nextOffset (inclusive),
3934       // simultaneously clearing it.
3935       dirtyRegion =
3936         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3937       assert(dirtyRegion.start() >= nextAddr,
3938              "returned region inconsistent?");
3939     }
3940     // Remember where the next search should begin.
3941     // The returned region (if non-empty) is a right open interval,
3942     // so lastOffset is obtained from the right end of that
3943     // interval.
3944     lastAddr = dirtyRegion.end();
3945     // Should do something more transparent and less hacky XXX
3946     numDirtyCards =
3947       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3948 
3949     // We'll scan the cards in the dirty region (with periodic
3950     // yields for foreground GC as needed).
3951     if (!dirtyRegion.is_empty()) {
3952       assert(numDirtyCards > 0, "consistency check");
3953       HeapWord* stop_point = NULL;
3954       stopTimer();
3955       // Potential yield point
3956       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3957                                bitMapLock());
3958       startTimer();
3959       {
3960         verify_work_stacks_empty();
3961         verify_overflow_empty();
3962         sample_eden();
3963         stop_point =
3964           old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3965       }
3966       if (stop_point != NULL) {
3967         // The careful iteration stopped early either because it found an
3968         // uninitialized object, or because we were in the midst of an
3969         // "abortable preclean", which should now be aborted. Redirty
3970         // the bits corresponding to the partially-scanned or unscanned
3971         // cards. We'll either restart at the next block boundary or
3972         // abort the preclean.
3973         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3974                "Should only be AbortablePreclean.");
3975         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3976         if (should_abort_preclean()) {
3977           break; // out of preclean loop
3978         } else {
3979           // Compute the next address at which preclean should pick up;
3980           // might need bitMapLock in order to read P-bits.
3981           lastAddr = next_card_start_after_block(stop_point);
3982         }
3983       }
3984     } else {
3985       assert(lastAddr == endAddr, "consistency check");
3986       assert(numDirtyCards == 0, "consistency check");
3987       break;
3988     }
3989   }
3990   verify_work_stacks_empty();
3991   verify_overflow_empty();
3992   return cumNumDirtyCards;
3993 }
3994 
3995 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3996 // below are largely identical; if you need to modify
3997 // one of these methods, please check the other method too.
3998 
3999 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4000   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4001   // strategy: it's similar to precleamModUnionTable above, in that
4002   // we accumulate contiguous ranges of dirty cards, mark these cards
4003   // precleaned, then scan the region covered by these cards.
4004   HeapWord* endAddr   = (HeapWord*)(old_gen->_virtual_space.high());
4005   HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4006 
4007   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
4008 
4009   size_t numDirtyCards, cumNumDirtyCards;
4010   HeapWord *lastAddr, *nextAddr;
4011 
4012   for (cumNumDirtyCards = numDirtyCards = 0,
4013        nextAddr = lastAddr = startAddr;
4014        nextAddr < endAddr;
4015        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4016 
4017     ResourceMark rm;
4018     HandleMark   hm;
4019 
4020     MemRegion dirtyRegion;
4021     {
4022       // See comments in "Precleaning notes" above on why we
4023       // do this locking. XXX Could the locking overheads be
4024       // too high when dirty cards are sparse? [I don't think so.]
4025       stopTimer();
4026       CMSTokenSync x(true); // is cms thread
4027       startTimer();
4028       sample_eden();
4029       // Get and clear dirty region from card table
4030       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4031                                     MemRegion(nextAddr, endAddr),
4032                                     true,
4033                                     CardTableModRefBS::precleaned_card_val());
4034 
4035       assert(dirtyRegion.start() >= nextAddr,
4036              "returned region inconsistent?");
4037     }
4038     lastAddr = dirtyRegion.end();
4039     numDirtyCards =
4040       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4041 
4042     if (!dirtyRegion.is_empty()) {
4043       stopTimer();
4044       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4045       startTimer();
4046       sample_eden();
4047       verify_work_stacks_empty();
4048       verify_overflow_empty();
4049       HeapWord* stop_point =
4050         old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4051       if (stop_point != NULL) {
4052         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4053                "Should only be AbortablePreclean.");
4054         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4055         if (should_abort_preclean()) {
4056           break; // out of preclean loop
4057         } else {
4058           // Compute the next address at which preclean should pick up.
4059           lastAddr = next_card_start_after_block(stop_point);
4060         }
4061       }
4062     } else {
4063       break;
4064     }
4065   }
4066   verify_work_stacks_empty();
4067   verify_overflow_empty();
4068   return cumNumDirtyCards;
4069 }
4070 
4071 class PrecleanKlassClosure : public KlassClosure {
4072   KlassToOopClosure _cm_klass_closure;
4073  public:
4074   PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4075   void do_klass(Klass* k) {
4076     if (k->has_accumulated_modified_oops()) {
4077       k->clear_accumulated_modified_oops();
4078 
4079       _cm_klass_closure.do_klass(k);
4080     }
4081   }
4082 };
4083 
4084 // The freelist lock is needed to prevent asserts, is it really needed?
4085 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4086 
4087   cl->set_freelistLock(freelistLock);
4088 
4089   CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4090 
4091   // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4092   // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4093   PrecleanKlassClosure preclean_klass_closure(cl);
4094   ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4095 
4096   verify_work_stacks_empty();
4097   verify_overflow_empty();
4098 }
4099 
4100 void CMSCollector::checkpointRootsFinal() {
4101   assert(_collectorState == FinalMarking, "incorrect state transition?");
4102   check_correct_thread_executing();
4103   // world is stopped at this checkpoint
4104   assert(SafepointSynchronize::is_at_safepoint(),
4105          "world should be stopped");
4106   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4107 
4108   verify_work_stacks_empty();
4109   verify_overflow_empty();
4110 
4111   log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4112                 _young_gen->used() / K, _young_gen->capacity() / K);
4113   {
4114     if (CMSScavengeBeforeRemark) {
4115       GenCollectedHeap* gch = GenCollectedHeap::heap();
4116       // Temporarily set flag to false, GCH->do_collection will
4117       // expect it to be false and set to true
4118       FlagSetting fl(gch->_is_gc_active, false);
4119 
4120       gch->do_collection(true,                      // full (i.e. force, see below)
4121                          false,                     // !clear_all_soft_refs
4122                          0,                         // size
4123                          false,                     // is_tlab
4124                          GenCollectedHeap::YoungGen // type
4125         );
4126     }
4127     FreelistLocker x(this);
4128     MutexLockerEx y(bitMapLock(),
4129                     Mutex::_no_safepoint_check_flag);
4130     checkpointRootsFinalWork();
4131   }
4132   verify_work_stacks_empty();
4133   verify_overflow_empty();
4134 }
4135 
4136 void CMSCollector::checkpointRootsFinalWork() {
4137   GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4138 
4139   assert(haveFreelistLocks(), "must have free list locks");
4140   assert_lock_strong(bitMapLock());
4141 
4142   ResourceMark rm;
4143   HandleMark   hm;
4144 
4145   GenCollectedHeap* gch = GenCollectedHeap::heap();
4146 
4147   if (should_unload_classes()) {
4148     CodeCache::gc_prologue();
4149   }
4150   assert(haveFreelistLocks(), "must have free list locks");
4151   assert_lock_strong(bitMapLock());
4152 
4153   // We might assume that we need not fill TLAB's when
4154   // CMSScavengeBeforeRemark is set, because we may have just done
4155   // a scavenge which would have filled all TLAB's -- and besides
4156   // Eden would be empty. This however may not always be the case --
4157   // for instance although we asked for a scavenge, it may not have
4158   // happened because of a JNI critical section. We probably need
4159   // a policy for deciding whether we can in that case wait until
4160   // the critical section releases and then do the remark following
4161   // the scavenge, and skip it here. In the absence of that policy,
4162   // or of an indication of whether the scavenge did indeed occur,
4163   // we cannot rely on TLAB's having been filled and must do
4164   // so here just in case a scavenge did not happen.
4165   gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
4166   // Update the saved marks which may affect the root scans.
4167   gch->save_marks();
4168 
4169   print_eden_and_survivor_chunk_arrays();
4170 
4171   {
4172 #if defined(COMPILER2) || INCLUDE_JVMCI
4173     DerivedPointerTableDeactivate dpt_deact;
4174 #endif
4175 
4176     // Note on the role of the mod union table:
4177     // Since the marker in "markFromRoots" marks concurrently with
4178     // mutators, it is possible for some reachable objects not to have been
4179     // scanned. For instance, an only reference to an object A was
4180     // placed in object B after the marker scanned B. Unless B is rescanned,
4181     // A would be collected. Such updates to references in marked objects
4182     // are detected via the mod union table which is the set of all cards
4183     // dirtied since the first checkpoint in this GC cycle and prior to
4184     // the most recent young generation GC, minus those cleaned up by the
4185     // concurrent precleaning.
4186     if (CMSParallelRemarkEnabled) {
4187       GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4188       do_remark_parallel();
4189     } else {
4190       GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4191       do_remark_non_parallel();
4192     }
4193   }
4194   verify_work_stacks_empty();
4195   verify_overflow_empty();
4196 
4197   {
4198     GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4199     refProcessingWork();
4200   }
4201   verify_work_stacks_empty();
4202   verify_overflow_empty();
4203 
4204   if (should_unload_classes()) {
4205     CodeCache::gc_epilogue();
4206   }
4207   JvmtiExport::gc_epilogue();
4208 
4209   // If we encountered any (marking stack / work queue) overflow
4210   // events during the current CMS cycle, take appropriate
4211   // remedial measures, where possible, so as to try and avoid
4212   // recurrence of that condition.
4213   assert(_markStack.isEmpty(), "No grey objects");
4214   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4215                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
4216   if (ser_ovflw > 0) {
4217     log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4218                          _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4219     _markStack.expand();
4220     _ser_pmc_remark_ovflw = 0;
4221     _ser_pmc_preclean_ovflw = 0;
4222     _ser_kac_preclean_ovflw = 0;
4223     _ser_kac_ovflw = 0;
4224   }
4225   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4226      log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4227                           _par_pmc_remark_ovflw, _par_kac_ovflw);
4228      _par_pmc_remark_ovflw = 0;
4229     _par_kac_ovflw = 0;
4230   }
4231    if (_markStack._hit_limit > 0) {
4232      log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4233                           _markStack._hit_limit);
4234    }
4235    if (_markStack._failed_double > 0) {
4236      log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4237                           _markStack._failed_double, _markStack.capacity());
4238    }
4239   _markStack._hit_limit = 0;
4240   _markStack._failed_double = 0;
4241 
4242   if ((VerifyAfterGC || VerifyDuringGC) &&
4243       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4244     verify_after_remark();
4245   }
4246 
4247   _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4248 
4249   // Change under the freelistLocks.
4250   _collectorState = Sweeping;
4251   // Call isAllClear() under bitMapLock
4252   assert(_modUnionTable.isAllClear(),
4253       "Should be clear by end of the final marking");
4254   assert(_ct->klass_rem_set()->mod_union_is_clear(),
4255       "Should be clear by end of the final marking");
4256 }
4257 
4258 void CMSParInitialMarkTask::work(uint worker_id) {
4259   elapsedTimer _timer;
4260   ResourceMark rm;
4261   HandleMark   hm;
4262 
4263   // ---------- scan from roots --------------
4264   _timer.start();
4265   GenCollectedHeap* gch = GenCollectedHeap::heap();
4266   ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4267 
4268   // ---------- young gen roots --------------
4269   {
4270     work_on_young_gen_roots(&par_mri_cl);
4271     _timer.stop();
4272     log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4273   }
4274 
4275   // ---------- remaining roots --------------
4276   _timer.reset();
4277   _timer.start();
4278 
4279   CLDToOopClosure cld_closure(&par_mri_cl, true);
4280 
4281   gch->cms_process_roots(_strong_roots_scope,
4282                          false,     // yg was scanned above
4283                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4284                          _collector->should_unload_classes(),
4285                          &par_mri_cl,
4286                          &cld_closure);
4287   assert(_collector->should_unload_classes()
4288          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4289          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4290   _timer.stop();
4291   log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4292 }
4293 
4294 // Parallel remark task
4295 class CMSParRemarkTask: public CMSParMarkTask {
4296   CompactibleFreeListSpace* _cms_space;
4297 
4298   // The per-thread work queues, available here for stealing.
4299   OopTaskQueueSet*       _task_queues;
4300   ParallelTaskTerminator _term;
4301   StrongRootsScope*      _strong_roots_scope;
4302 
4303  public:
4304   // A value of 0 passed to n_workers will cause the number of
4305   // workers to be taken from the active workers in the work gang.
4306   CMSParRemarkTask(CMSCollector* collector,
4307                    CompactibleFreeListSpace* cms_space,
4308                    uint n_workers, WorkGang* workers,
4309                    OopTaskQueueSet* task_queues,
4310                    StrongRootsScope* strong_roots_scope):
4311     CMSParMarkTask("Rescan roots and grey objects in parallel",
4312                    collector, n_workers),
4313     _cms_space(cms_space),
4314     _task_queues(task_queues),
4315     _term(n_workers, task_queues),
4316     _strong_roots_scope(strong_roots_scope) { }
4317 
4318   OopTaskQueueSet* task_queues() { return _task_queues; }
4319 
4320   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4321 
4322   ParallelTaskTerminator* terminator() { return &_term; }
4323   uint n_workers() { return _n_workers; }
4324 
4325   void work(uint worker_id);
4326 
4327  private:
4328   // ... of  dirty cards in old space
4329   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4330                                   ParMarkRefsIntoAndScanClosure* cl);
4331 
4332   // ... work stealing for the above
4333   void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4334 };
4335 
4336 class RemarkKlassClosure : public KlassClosure {
4337   KlassToOopClosure _cm_klass_closure;
4338  public:
4339   RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4340   void do_klass(Klass* k) {
4341     // Check if we have modified any oops in the Klass during the concurrent marking.
4342     if (k->has_accumulated_modified_oops()) {
4343       k->clear_accumulated_modified_oops();
4344 
4345       // We could have transfered the current modified marks to the accumulated marks,
4346       // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4347     } else if (k->has_modified_oops()) {
4348       // Don't clear anything, this info is needed by the next young collection.
4349     } else {
4350       // No modified oops in the Klass.
4351       return;
4352     }
4353 
4354     // The klass has modified fields, need to scan the klass.
4355     _cm_klass_closure.do_klass(k);
4356   }
4357 };
4358 
4359 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4360   ParNewGeneration* young_gen = _collector->_young_gen;
4361   ContiguousSpace* eden_space = young_gen->eden();
4362   ContiguousSpace* from_space = young_gen->from();
4363   ContiguousSpace* to_space   = young_gen->to();
4364 
4365   HeapWord** eca = _collector->_eden_chunk_array;
4366   size_t     ect = _collector->_eden_chunk_index;
4367   HeapWord** sca = _collector->_survivor_chunk_array;
4368   size_t     sct = _collector->_survivor_chunk_index;
4369 
4370   assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4371   assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4372 
4373   do_young_space_rescan(cl, to_space, NULL, 0);
4374   do_young_space_rescan(cl, from_space, sca, sct);
4375   do_young_space_rescan(cl, eden_space, eca, ect);
4376 }
4377 
4378 // work_queue(i) is passed to the closure
4379 // ParMarkRefsIntoAndScanClosure.  The "i" parameter
4380 // also is passed to do_dirty_card_rescan_tasks() and to
4381 // do_work_steal() to select the i-th task_queue.
4382 
4383 void CMSParRemarkTask::work(uint worker_id) {
4384   elapsedTimer _timer;
4385   ResourceMark rm;
4386   HandleMark   hm;
4387 
4388   // ---------- rescan from roots --------------
4389   _timer.start();
4390   GenCollectedHeap* gch = GenCollectedHeap::heap();
4391   ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4392     _collector->_span, _collector->ref_processor(),
4393     &(_collector->_markBitMap),
4394     work_queue(worker_id));
4395 
4396   // Rescan young gen roots first since these are likely
4397   // coarsely partitioned and may, on that account, constitute
4398   // the critical path; thus, it's best to start off that
4399   // work first.
4400   // ---------- young gen roots --------------
4401   {
4402     work_on_young_gen_roots(&par_mrias_cl);
4403     _timer.stop();
4404     log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4405   }
4406 
4407   // ---------- remaining roots --------------
4408   _timer.reset();
4409   _timer.start();
4410   gch->cms_process_roots(_strong_roots_scope,
4411                          false,     // yg was scanned above
4412                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4413                          _collector->should_unload_classes(),
4414                          &par_mrias_cl,
4415                          NULL);     // The dirty klasses will be handled below
4416 
4417   assert(_collector->should_unload_classes()
4418          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4419          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4420   _timer.stop();
4421   log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec",  worker_id, _timer.seconds());
4422 
4423   // ---------- unhandled CLD scanning ----------
4424   if (worker_id == 0) { // Single threaded at the moment.
4425     _timer.reset();
4426     _timer.start();
4427 
4428     // Scan all new class loader data objects and new dependencies that were
4429     // introduced during concurrent marking.
4430     ResourceMark rm;
4431     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4432     for (int i = 0; i < array->length(); i++) {
4433       par_mrias_cl.do_cld_nv(array->at(i));
4434     }
4435 
4436     // We don't need to keep track of new CLDs anymore.
4437     ClassLoaderDataGraph::remember_new_clds(false);
4438 
4439     _timer.stop();
4440     log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4441   }
4442 
4443   // ---------- dirty klass scanning ----------
4444   if (worker_id == 0) { // Single threaded at the moment.
4445     _timer.reset();
4446     _timer.start();
4447 
4448     // Scan all classes that was dirtied during the concurrent marking phase.
4449     RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4450     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4451 
4452     _timer.stop();
4453     log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4454   }
4455 
4456   // We might have added oops to ClassLoaderData::_handles during the
4457   // concurrent marking phase. These oops point to newly allocated objects
4458   // that are guaranteed to be kept alive. Either by the direct allocation
4459   // code, or when the young collector processes the roots. Hence,
4460   // we don't have to revisit the _handles block during the remark phase.
4461 
4462   // ---------- rescan dirty cards ------------
4463   _timer.reset();
4464   _timer.start();
4465 
4466   // Do the rescan tasks for each of the two spaces
4467   // (cms_space) in turn.
4468   // "worker_id" is passed to select the task_queue for "worker_id"
4469   do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4470   _timer.stop();
4471   log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4472 
4473   // ---------- steal work from other threads ...
4474   // ---------- ... and drain overflow list.
4475   _timer.reset();
4476   _timer.start();
4477   do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4478   _timer.stop();
4479   log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4480 }
4481 
4482 void
4483 CMSParMarkTask::do_young_space_rescan(
4484   OopsInGenClosure* cl, ContiguousSpace* space,
4485   HeapWord** chunk_array, size_t chunk_top) {
4486   // Until all tasks completed:
4487   // . claim an unclaimed task
4488   // . compute region boundaries corresponding to task claimed
4489   //   using chunk_array
4490   // . par_oop_iterate(cl) over that region
4491 
4492   ResourceMark rm;
4493   HandleMark   hm;
4494 
4495   SequentialSubTasksDone* pst = space->par_seq_tasks();
4496 
4497   uint nth_task = 0;
4498   uint n_tasks  = pst->n_tasks();
4499 
4500   if (n_tasks > 0) {
4501     assert(pst->valid(), "Uninitialized use?");
4502     HeapWord *start, *end;
4503     while (!pst->is_task_claimed(/* reference */ nth_task)) {
4504       // We claimed task # nth_task; compute its boundaries.
4505       if (chunk_top == 0) {  // no samples were taken
4506         assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4507         start = space->bottom();
4508         end   = space->top();
4509       } else if (nth_task == 0) {
4510         start = space->bottom();
4511         end   = chunk_array[nth_task];
4512       } else if (nth_task < (uint)chunk_top) {
4513         assert(nth_task >= 1, "Control point invariant");
4514         start = chunk_array[nth_task - 1];
4515         end   = chunk_array[nth_task];
4516       } else {
4517         assert(nth_task == (uint)chunk_top, "Control point invariant");
4518         start = chunk_array[chunk_top - 1];
4519         end   = space->top();
4520       }
4521       MemRegion mr(start, end);
4522       // Verify that mr is in space
4523       assert(mr.is_empty() || space->used_region().contains(mr),
4524              "Should be in space");
4525       // Verify that "start" is an object boundary
4526       assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4527              "Should be an oop");
4528       space->par_oop_iterate(mr, cl);
4529     }
4530     pst->all_tasks_completed();
4531   }
4532 }
4533 
4534 void
4535 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4536   CompactibleFreeListSpace* sp, int i,
4537   ParMarkRefsIntoAndScanClosure* cl) {
4538   // Until all tasks completed:
4539   // . claim an unclaimed task
4540   // . compute region boundaries corresponding to task claimed
4541   // . transfer dirty bits ct->mut for that region
4542   // . apply rescanclosure to dirty mut bits for that region
4543 
4544   ResourceMark rm;
4545   HandleMark   hm;
4546 
4547   OopTaskQueue* work_q = work_queue(i);
4548   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4549   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4550   // CAUTION: This closure has state that persists across calls to
4551   // the work method dirty_range_iterate_clear() in that it has
4552   // embedded in it a (subtype of) UpwardsObjectClosure. The
4553   // use of that state in the embedded UpwardsObjectClosure instance
4554   // assumes that the cards are always iterated (even if in parallel
4555   // by several threads) in monotonically increasing order per each
4556   // thread. This is true of the implementation below which picks
4557   // card ranges (chunks) in monotonically increasing order globally
4558   // and, a-fortiori, in monotonically increasing order per thread
4559   // (the latter order being a subsequence of the former).
4560   // If the work code below is ever reorganized into a more chaotic
4561   // work-partitioning form than the current "sequential tasks"
4562   // paradigm, the use of that persistent state will have to be
4563   // revisited and modified appropriately. See also related
4564   // bug 4756801 work on which should examine this code to make
4565   // sure that the changes there do not run counter to the
4566   // assumptions made here and necessary for correctness and
4567   // efficiency. Note also that this code might yield inefficient
4568   // behavior in the case of very large objects that span one or
4569   // more work chunks. Such objects would potentially be scanned
4570   // several times redundantly. Work on 4756801 should try and
4571   // address that performance anomaly if at all possible. XXX
4572   MemRegion  full_span  = _collector->_span;
4573   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
4574   MarkFromDirtyCardsClosure
4575     greyRescanClosure(_collector, full_span, // entire span of interest
4576                       sp, bm, work_q, cl);
4577 
4578   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4579   assert(pst->valid(), "Uninitialized use?");
4580   uint nth_task = 0;
4581   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4582   MemRegion span = sp->used_region();
4583   HeapWord* start_addr = span.start();
4584   HeapWord* end_addr = align_up(span.end(), alignment);
4585   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4586   assert(is_aligned(start_addr, alignment), "Check alignment");
4587   assert(is_aligned(chunk_size, alignment), "Check alignment");
4588 
4589   while (!pst->is_task_claimed(/* reference */ nth_task)) {
4590     // Having claimed the nth_task, compute corresponding mem-region,
4591     // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4592     // The alignment restriction ensures that we do not need any
4593     // synchronization with other gang-workers while setting or
4594     // clearing bits in thus chunk of the MUT.
4595     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4596                                     start_addr + (nth_task+1)*chunk_size);
4597     // The last chunk's end might be way beyond end of the
4598     // used region. In that case pull back appropriately.
4599     if (this_span.end() > end_addr) {
4600       this_span.set_end(end_addr);
4601       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4602     }
4603     // Iterate over the dirty cards covering this chunk, marking them
4604     // precleaned, and setting the corresponding bits in the mod union
4605     // table. Since we have been careful to partition at Card and MUT-word
4606     // boundaries no synchronization is needed between parallel threads.
4607     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4608                                                  &modUnionClosure);
4609 
4610     // Having transferred these marks into the modUnionTable,
4611     // rescan the marked objects on the dirty cards in the modUnionTable.
4612     // Even if this is at a synchronous collection, the initial marking
4613     // may have been done during an asynchronous collection so there
4614     // may be dirty bits in the mod-union table.
4615     _collector->_modUnionTable.dirty_range_iterate_clear(
4616                   this_span, &greyRescanClosure);
4617     _collector->_modUnionTable.verifyNoOneBitsInRange(
4618                                  this_span.start(),
4619                                  this_span.end());
4620   }
4621   pst->all_tasks_completed();  // declare that i am done
4622 }
4623 
4624 // . see if we can share work_queues with ParNew? XXX
4625 void
4626 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4627                                 int* seed) {
4628   OopTaskQueue* work_q = work_queue(i);
4629   NOT_PRODUCT(int num_steals = 0;)
4630   oop obj_to_scan;
4631   CMSBitMap* bm = &(_collector->_markBitMap);
4632 
4633   while (true) {
4634     // Completely finish any left over work from (an) earlier round(s)
4635     cl->trim_queue(0);
4636     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4637                                          (size_t)ParGCDesiredObjsFromOverflowList);
4638     // Now check if there's any work in the overflow list
4639     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4640     // only affects the number of attempts made to get work from the
4641     // overflow list and does not affect the number of workers.  Just
4642     // pass ParallelGCThreads so this behavior is unchanged.
4643     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4644                                                 work_q,
4645                                                 ParallelGCThreads)) {
4646       // found something in global overflow list;
4647       // not yet ready to go stealing work from others.
4648       // We'd like to assert(work_q->size() != 0, ...)
4649       // because we just took work from the overflow list,
4650       // but of course we can't since all of that could have
4651       // been already stolen from us.
4652       // "He giveth and He taketh away."
4653       continue;
4654     }
4655     // Verify that we have no work before we resort to stealing
4656     assert(work_q->size() == 0, "Have work, shouldn't steal");
4657     // Try to steal from other queues that have work
4658     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4659       NOT_PRODUCT(num_steals++;)
4660       assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4661       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4662       // Do scanning work
4663       obj_to_scan->oop_iterate(cl);
4664       // Loop around, finish this work, and try to steal some more
4665     } else if (terminator()->offer_termination()) {
4666         break;  // nirvana from the infinite cycle
4667     }
4668   }
4669   log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4670   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4671          "Else our work is not yet done");
4672 }
4673 
4674 // Record object boundaries in _eden_chunk_array by sampling the eden
4675 // top in the slow-path eden object allocation code path and record
4676 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4677 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4678 // sampling in sample_eden() that activates during the part of the
4679 // preclean phase.
4680 void CMSCollector::sample_eden_chunk() {
4681   if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4682     if (_eden_chunk_lock->try_lock()) {
4683       // Record a sample. This is the critical section. The contents
4684       // of the _eden_chunk_array have to be non-decreasing in the
4685       // address order.
4686       _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4687       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4688              "Unexpected state of Eden");
4689       if (_eden_chunk_index == 0 ||
4690           ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4691            (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4692                           _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4693         _eden_chunk_index++;  // commit sample
4694       }
4695       _eden_chunk_lock->unlock();
4696     }
4697   }
4698 }
4699 
4700 // Return a thread-local PLAB recording array, as appropriate.
4701 void* CMSCollector::get_data_recorder(int thr_num) {
4702   if (_survivor_plab_array != NULL &&
4703       (CMSPLABRecordAlways ||
4704        (_collectorState > Marking && _collectorState < FinalMarking))) {
4705     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4706     ChunkArray* ca = &_survivor_plab_array[thr_num];
4707     ca->reset();   // clear it so that fresh data is recorded
4708     return (void*) ca;
4709   } else {
4710     return NULL;
4711   }
4712 }
4713 
4714 // Reset all the thread-local PLAB recording arrays
4715 void CMSCollector::reset_survivor_plab_arrays() {
4716   for (uint i = 0; i < ParallelGCThreads; i++) {
4717     _survivor_plab_array[i].reset();
4718   }
4719 }
4720 
4721 // Merge the per-thread plab arrays into the global survivor chunk
4722 // array which will provide the partitioning of the survivor space
4723 // for CMS initial scan and rescan.
4724 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4725                                               int no_of_gc_threads) {
4726   assert(_survivor_plab_array  != NULL, "Error");
4727   assert(_survivor_chunk_array != NULL, "Error");
4728   assert(_collectorState == FinalMarking ||
4729          (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4730   for (int j = 0; j < no_of_gc_threads; j++) {
4731     _cursor[j] = 0;
4732   }
4733   HeapWord* top = surv->top();
4734   size_t i;
4735   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
4736     HeapWord* min_val = top;          // Higher than any PLAB address
4737     uint      min_tid = 0;            // position of min_val this round
4738     for (int j = 0; j < no_of_gc_threads; j++) {
4739       ChunkArray* cur_sca = &_survivor_plab_array[j];
4740       if (_cursor[j] == cur_sca->end()) {
4741         continue;
4742       }
4743       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4744       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4745       assert(surv->used_region().contains(cur_val), "Out of bounds value");
4746       if (cur_val < min_val) {
4747         min_tid = j;
4748         min_val = cur_val;
4749       } else {
4750         assert(cur_val < top, "All recorded addresses should be less");
4751       }
4752     }
4753     // At this point min_val and min_tid are respectively
4754     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4755     // and the thread (j) that witnesses that address.
4756     // We record this address in the _survivor_chunk_array[i]
4757     // and increment _cursor[min_tid] prior to the next round i.
4758     if (min_val == top) {
4759       break;
4760     }
4761     _survivor_chunk_array[i] = min_val;
4762     _cursor[min_tid]++;
4763   }
4764   // We are all done; record the size of the _survivor_chunk_array
4765   _survivor_chunk_index = i; // exclusive: [0, i)
4766   log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4767   // Verify that we used up all the recorded entries
4768   #ifdef ASSERT
4769     size_t total = 0;
4770     for (int j = 0; j < no_of_gc_threads; j++) {
4771       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4772       total += _cursor[j];
4773     }
4774     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4775     // Check that the merged array is in sorted order
4776     if (total > 0) {
4777       for (size_t i = 0; i < total - 1; i++) {
4778         log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4779                                      i, p2i(_survivor_chunk_array[i]));
4780         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4781                "Not sorted");
4782       }
4783     }
4784   #endif // ASSERT
4785 }
4786 
4787 // Set up the space's par_seq_tasks structure for work claiming
4788 // for parallel initial scan and rescan of young gen.
4789 // See ParRescanTask where this is currently used.
4790 void
4791 CMSCollector::
4792 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4793   assert(n_threads > 0, "Unexpected n_threads argument");
4794 
4795   // Eden space
4796   if (!_young_gen->eden()->is_empty()) {
4797     SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4798     assert(!pst->valid(), "Clobbering existing data?");
4799     // Each valid entry in [0, _eden_chunk_index) represents a task.
4800     size_t n_tasks = _eden_chunk_index + 1;
4801     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4802     // Sets the condition for completion of the subtask (how many threads
4803     // need to finish in order to be done).
4804     pst->set_n_threads(n_threads);
4805     pst->set_n_tasks((int)n_tasks);
4806   }
4807 
4808   // Merge the survivor plab arrays into _survivor_chunk_array
4809   if (_survivor_plab_array != NULL) {
4810     merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4811   } else {
4812     assert(_survivor_chunk_index == 0, "Error");
4813   }
4814 
4815   // To space
4816   {
4817     SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4818     assert(!pst->valid(), "Clobbering existing data?");
4819     // Sets the condition for completion of the subtask (how many threads
4820     // need to finish in order to be done).
4821     pst->set_n_threads(n_threads);
4822     pst->set_n_tasks(1);
4823     assert(pst->valid(), "Error");
4824   }
4825 
4826   // From space
4827   {
4828     SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4829     assert(!pst->valid(), "Clobbering existing data?");
4830     size_t n_tasks = _survivor_chunk_index + 1;
4831     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4832     // Sets the condition for completion of the subtask (how many threads
4833     // need to finish in order to be done).
4834     pst->set_n_threads(n_threads);
4835     pst->set_n_tasks((int)n_tasks);
4836     assert(pst->valid(), "Error");
4837   }
4838 }
4839 
4840 // Parallel version of remark
4841 void CMSCollector::do_remark_parallel() {
4842   GenCollectedHeap* gch = GenCollectedHeap::heap();
4843   WorkGang* workers = gch->workers();
4844   assert(workers != NULL, "Need parallel worker threads.");
4845   // Choose to use the number of GC workers most recently set
4846   // into "active_workers".
4847   uint n_workers = workers->active_workers();
4848 
4849   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
4850 
4851   StrongRootsScope srs(n_workers);
4852 
4853   CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4854 
4855   // We won't be iterating over the cards in the card table updating
4856   // the younger_gen cards, so we shouldn't call the following else
4857   // the verification code as well as subsequent younger_refs_iterate
4858   // code would get confused. XXX
4859   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4860 
4861   // The young gen rescan work will not be done as part of
4862   // process_roots (which currently doesn't know how to
4863   // parallelize such a scan), but rather will be broken up into
4864   // a set of parallel tasks (via the sampling that the [abortable]
4865   // preclean phase did of eden, plus the [two] tasks of
4866   // scanning the [two] survivor spaces. Further fine-grain
4867   // parallelization of the scanning of the survivor spaces
4868   // themselves, and of precleaning of the young gen itself
4869   // is deferred to the future.
4870   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4871 
4872   // The dirty card rescan work is broken up into a "sequence"
4873   // of parallel tasks (per constituent space) that are dynamically
4874   // claimed by the parallel threads.
4875   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4876 
4877   // It turns out that even when we're using 1 thread, doing the work in a
4878   // separate thread causes wide variance in run times.  We can't help this
4879   // in the multi-threaded case, but we special-case n=1 here to get
4880   // repeatable measurements of the 1-thread overhead of the parallel code.
4881   if (n_workers > 1) {
4882     // Make refs discovery MT-safe, if it isn't already: it may not
4883     // necessarily be so, since it's possible that we are doing
4884     // ST marking.
4885     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4886     workers->run_task(&tsk);
4887   } else {
4888     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4889     tsk.work(0);
4890   }
4891 
4892   // restore, single-threaded for now, any preserved marks
4893   // as a result of work_q overflow
4894   restore_preserved_marks_if_any();
4895 }
4896 
4897 // Non-parallel version of remark
4898 void CMSCollector::do_remark_non_parallel() {
4899   ResourceMark rm;
4900   HandleMark   hm;
4901   GenCollectedHeap* gch = GenCollectedHeap::heap();
4902   ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4903 
4904   MarkRefsIntoAndScanClosure
4905     mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4906              &_markStack, this,
4907              false /* should_yield */, false /* not precleaning */);
4908   MarkFromDirtyCardsClosure
4909     markFromDirtyCardsClosure(this, _span,
4910                               NULL,  // space is set further below
4911                               &_markBitMap, &_markStack, &mrias_cl);
4912   {
4913     GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4914     // Iterate over the dirty cards, setting the corresponding bits in the
4915     // mod union table.
4916     {
4917       ModUnionClosure modUnionClosure(&_modUnionTable);
4918       _ct->ct_bs()->dirty_card_iterate(
4919                       _cmsGen->used_region(),
4920                       &modUnionClosure);
4921     }
4922     // Having transferred these marks into the modUnionTable, we just need
4923     // to rescan the marked objects on the dirty cards in the modUnionTable.
4924     // The initial marking may have been done during an asynchronous
4925     // collection so there may be dirty bits in the mod-union table.
4926     const int alignment =
4927       CardTableModRefBS::card_size * BitsPerWord;
4928     {
4929       // ... First handle dirty cards in CMS gen
4930       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4931       MemRegion ur = _cmsGen->used_region();
4932       HeapWord* lb = ur.start();
4933       HeapWord* ub = align_up(ur.end(), alignment);
4934       MemRegion cms_span(lb, ub);
4935       _modUnionTable.dirty_range_iterate_clear(cms_span,
4936                                                &markFromDirtyCardsClosure);
4937       verify_work_stacks_empty();
4938       log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4939     }
4940   }
4941   if (VerifyDuringGC &&
4942       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4943     HandleMark hm;  // Discard invalid handles created during verification
4944     Universe::verify();
4945   }
4946   {
4947     GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4948 
4949     verify_work_stacks_empty();
4950 
4951     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4952     StrongRootsScope srs(1);
4953 
4954     gch->cms_process_roots(&srs,
4955                            true,  // young gen as roots
4956                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
4957                            should_unload_classes(),
4958                            &mrias_cl,
4959                            NULL); // The dirty klasses will be handled below
4960 
4961     assert(should_unload_classes()
4962            || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4963            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4964   }
4965 
4966   {
4967     GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4968 
4969     verify_work_stacks_empty();
4970 
4971     // Scan all class loader data objects that might have been introduced
4972     // during concurrent marking.
4973     ResourceMark rm;
4974     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4975     for (int i = 0; i < array->length(); i++) {
4976       mrias_cl.do_cld_nv(array->at(i));
4977     }
4978 
4979     // We don't need to keep track of new CLDs anymore.
4980     ClassLoaderDataGraph::remember_new_clds(false);
4981 
4982     verify_work_stacks_empty();
4983   }
4984 
4985   {
4986     GCTraceTime(Trace, gc, phases) t("Dirty Klass Scan", _gc_timer_cm);
4987 
4988     verify_work_stacks_empty();
4989 
4990     RemarkKlassClosure remark_klass_closure(&mrias_cl);
4991     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4992 
4993     verify_work_stacks_empty();
4994   }
4995 
4996   // We might have added oops to ClassLoaderData::_handles during the
4997   // concurrent marking phase. These oops point to newly allocated objects
4998   // that are guaranteed to be kept alive. Either by the direct allocation
4999   // code, or when the young collector processes the roots. Hence,
5000   // we don't have to revisit the _handles block during the remark phase.
5001 
5002   verify_work_stacks_empty();
5003   // Restore evacuated mark words, if any, used for overflow list links
5004   restore_preserved_marks_if_any();
5005 
5006   verify_overflow_empty();
5007 }
5008 
5009 ////////////////////////////////////////////////////////
5010 // Parallel Reference Processing Task Proxy Class
5011 ////////////////////////////////////////////////////////
5012 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5013   OopTaskQueueSet*       _queues;
5014   ParallelTaskTerminator _terminator;
5015  public:
5016   AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5017     AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5018   ParallelTaskTerminator* terminator() { return &_terminator; }
5019   OopTaskQueueSet* queues() { return _queues; }
5020 };
5021 
5022 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5023   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5024   CMSCollector*          _collector;
5025   CMSBitMap*             _mark_bit_map;
5026   const MemRegion        _span;
5027   ProcessTask&           _task;
5028 
5029 public:
5030   CMSRefProcTaskProxy(ProcessTask&     task,
5031                       CMSCollector*    collector,
5032                       const MemRegion& span,
5033                       CMSBitMap*       mark_bit_map,
5034                       AbstractWorkGang* workers,
5035                       OopTaskQueueSet* task_queues):
5036     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5037       task_queues,
5038       workers->active_workers()),
5039     _task(task),
5040     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5041   {
5042     assert(_collector->_span.equals(_span) && !_span.is_empty(),
5043            "Inconsistency in _span");
5044   }
5045 
5046   OopTaskQueueSet* task_queues() { return queues(); }
5047 
5048   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5049 
5050   void do_work_steal(int i,
5051                      CMSParDrainMarkingStackClosure* drain,
5052                      CMSParKeepAliveClosure* keep_alive,
5053                      int* seed);
5054 
5055   virtual void work(uint worker_id);
5056 };
5057 
5058 void CMSRefProcTaskProxy::work(uint worker_id) {
5059   ResourceMark rm;
5060   HandleMark hm;
5061   assert(_collector->_span.equals(_span), "Inconsistency in _span");
5062   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5063                                         _mark_bit_map,
5064                                         work_queue(worker_id));
5065   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5066                                                  _mark_bit_map,
5067                                                  work_queue(worker_id));
5068   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5069   _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5070   if (_task.marks_oops_alive()) {
5071     do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5072                   _collector->hash_seed(worker_id));
5073   }
5074   assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5075   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5076 }
5077 
5078 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5079   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5080   EnqueueTask& _task;
5081 
5082 public:
5083   CMSRefEnqueueTaskProxy(EnqueueTask& task)
5084     : AbstractGangTask("Enqueue reference objects in parallel"),
5085       _task(task)
5086   { }
5087 
5088   virtual void work(uint worker_id)
5089   {
5090     _task.work(worker_id);
5091   }
5092 };
5093 
5094 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5095   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5096    _span(span),
5097    _bit_map(bit_map),
5098    _work_queue(work_queue),
5099    _mark_and_push(collector, span, bit_map, work_queue),
5100    _low_water_mark(MIN2((work_queue->max_elems()/4),
5101                         ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5102 { }
5103 
5104 // . see if we can share work_queues with ParNew? XXX
5105 void CMSRefProcTaskProxy::do_work_steal(int i,
5106   CMSParDrainMarkingStackClosure* drain,
5107   CMSParKeepAliveClosure* keep_alive,
5108   int* seed) {
5109   OopTaskQueue* work_q = work_queue(i);
5110   NOT_PRODUCT(int num_steals = 0;)
5111   oop obj_to_scan;
5112 
5113   while (true) {
5114     // Completely finish any left over work from (an) earlier round(s)
5115     drain->trim_queue(0);
5116     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5117                                          (size_t)ParGCDesiredObjsFromOverflowList);
5118     // Now check if there's any work in the overflow list
5119     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5120     // only affects the number of attempts made to get work from the
5121     // overflow list and does not affect the number of workers.  Just
5122     // pass ParallelGCThreads so this behavior is unchanged.
5123     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5124                                                 work_q,
5125                                                 ParallelGCThreads)) {
5126       // Found something in global overflow list;
5127       // not yet ready to go stealing work from others.
5128       // We'd like to assert(work_q->size() != 0, ...)
5129       // because we just took work from the overflow list,
5130       // but of course we can't, since all of that might have
5131       // been already stolen from us.
5132       continue;
5133     }
5134     // Verify that we have no work before we resort to stealing
5135     assert(work_q->size() == 0, "Have work, shouldn't steal");
5136     // Try to steal from other queues that have work
5137     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5138       NOT_PRODUCT(num_steals++;)
5139       assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5140       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5141       // Do scanning work
5142       obj_to_scan->oop_iterate(keep_alive);
5143       // Loop around, finish this work, and try to steal some more
5144     } else if (terminator()->offer_termination()) {
5145       break;  // nirvana from the infinite cycle
5146     }
5147   }
5148   log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5149 }
5150 
5151 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5152 {
5153   GenCollectedHeap* gch = GenCollectedHeap::heap();
5154   WorkGang* workers = gch->workers();
5155   assert(workers != NULL, "Need parallel worker threads.");
5156   CMSRefProcTaskProxy rp_task(task, &_collector,
5157                               _collector.ref_processor()->span(),
5158                               _collector.markBitMap(),
5159                               workers, _collector.task_queues());
5160   workers->run_task(&rp_task);
5161 }
5162 
5163 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5164 {
5165 
5166   GenCollectedHeap* gch = GenCollectedHeap::heap();
5167   WorkGang* workers = gch->workers();
5168   assert(workers != NULL, "Need parallel worker threads.");
5169   CMSRefEnqueueTaskProxy enq_task(task);
5170   workers->run_task(&enq_task);
5171 }
5172 
5173 void CMSCollector::refProcessingWork() {
5174   ResourceMark rm;
5175   HandleMark   hm;
5176 
5177   ReferenceProcessor* rp = ref_processor();
5178   assert(rp->span().equals(_span), "Spans should be equal");
5179   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5180   // Process weak references.
5181   rp->setup_policy(false);
5182   verify_work_stacks_empty();
5183 
5184   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5185                                           &_markStack, false /* !preclean */);
5186   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5187                                 _span, &_markBitMap, &_markStack,
5188                                 &cmsKeepAliveClosure, false /* !preclean */);
5189   ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
5190   {
5191     GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5192 
5193     ReferenceProcessorStats stats;
5194     if (rp->processing_is_mt()) {
5195       // Set the degree of MT here.  If the discovery is done MT, there
5196       // may have been a different number of threads doing the discovery
5197       // and a different number of discovered lists may have Ref objects.
5198       // That is OK as long as the Reference lists are balanced (see
5199       // balance_all_queues() and balance_queues()).
5200       GenCollectedHeap* gch = GenCollectedHeap::heap();
5201       uint active_workers = ParallelGCThreads;
5202       WorkGang* workers = gch->workers();
5203       if (workers != NULL) {
5204         active_workers = workers->active_workers();
5205         // The expectation is that active_workers will have already
5206         // been set to a reasonable value.  If it has not been set,
5207         // investigate.
5208         assert(active_workers > 0, "Should have been set during scavenge");
5209       }
5210       rp->set_active_mt_degree(active_workers);
5211       CMSRefProcTaskExecutor task_executor(*this);
5212       stats = rp->process_discovered_references(&_is_alive_closure,
5213                                         &cmsKeepAliveClosure,
5214                                         &cmsDrainMarkingStackClosure,
5215                                         &task_executor,
5216                                         &pt);
5217     } else {
5218       stats = rp->process_discovered_references(&_is_alive_closure,
5219                                         &cmsKeepAliveClosure,
5220                                         &cmsDrainMarkingStackClosure,
5221                                         NULL,
5222                                         &pt);
5223     }
5224     _gc_tracer_cm->report_gc_reference_stats(stats);
5225     pt.print_all_references();
5226   }
5227 
5228   // This is the point where the entire marking should have completed.
5229   verify_work_stacks_empty();
5230 
5231   if (should_unload_classes()) {
5232     {
5233       GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5234 
5235       // Unload classes and purge the SystemDictionary.
5236       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm);
5237 
5238       // Unload nmethods.
5239       CodeCache::do_unloading(&_is_alive_closure, purged_class);
5240 
5241       // Prune dead klasses from subklass/sibling/implementor lists.
5242       Klass::clean_weak_klass_links(&_is_alive_closure);
5243     }
5244 
5245     {
5246       GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5247       // Clean up unreferenced symbols in symbol table.
5248       SymbolTable::unlink();
5249     }
5250 
5251     {
5252       GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5253       // Delete entries for dead interned strings.
5254       StringTable::unlink(&_is_alive_closure);
5255     }
5256   }
5257 
5258   // Restore any preserved marks as a result of mark stack or
5259   // work queue overflow
5260   restore_preserved_marks_if_any();  // done single-threaded for now
5261 
5262   rp->set_enqueuing_is_done(true);
5263   if (rp->processing_is_mt()) {
5264     rp->balance_all_queues();
5265     CMSRefProcTaskExecutor task_executor(*this);
5266     rp->enqueue_discovered_references(&task_executor, &pt);
5267   } else {
5268     rp->enqueue_discovered_references(NULL, &pt);
5269   }
5270   rp->verify_no_references_recorded();
5271   pt.print_enqueue_phase();
5272   assert(!rp->discovery_enabled(), "should have been disabled");
5273 }
5274 
5275 #ifndef PRODUCT
5276 void CMSCollector::check_correct_thread_executing() {
5277   Thread* t = Thread::current();
5278   // Only the VM thread or the CMS thread should be here.
5279   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5280          "Unexpected thread type");
5281   // If this is the vm thread, the foreground process
5282   // should not be waiting.  Note that _foregroundGCIsActive is
5283   // true while the foreground collector is waiting.
5284   if (_foregroundGCShouldWait) {
5285     // We cannot be the VM thread
5286     assert(t->is_ConcurrentGC_thread(),
5287            "Should be CMS thread");
5288   } else {
5289     // We can be the CMS thread only if we are in a stop-world
5290     // phase of CMS collection.
5291     if (t->is_ConcurrentGC_thread()) {
5292       assert(_collectorState == InitialMarking ||
5293              _collectorState == FinalMarking,
5294              "Should be a stop-world phase");
5295       // The CMS thread should be holding the CMS_token.
5296       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5297              "Potential interference with concurrently "
5298              "executing VM thread");
5299     }
5300   }
5301 }
5302 #endif
5303 
5304 void CMSCollector::sweep() {
5305   assert(_collectorState == Sweeping, "just checking");
5306   check_correct_thread_executing();
5307   verify_work_stacks_empty();
5308   verify_overflow_empty();
5309   increment_sweep_count();
5310   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5311 
5312   _inter_sweep_timer.stop();
5313   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5314 
5315   assert(!_intra_sweep_timer.is_active(), "Should not be active");
5316   _intra_sweep_timer.reset();
5317   _intra_sweep_timer.start();
5318   {
5319     GCTraceCPUTime tcpu;
5320     CMSPhaseAccounting pa(this, "Concurrent Sweep");
5321     // First sweep the old gen
5322     {
5323       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5324                                bitMapLock());
5325       sweepWork(_cmsGen);
5326     }
5327 
5328     // Update Universe::_heap_*_at_gc figures.
5329     // We need all the free list locks to make the abstract state
5330     // transition from Sweeping to Resetting. See detailed note
5331     // further below.
5332     {
5333       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5334       // Update heap occupancy information which is used as
5335       // input to soft ref clearing policy at the next gc.
5336       Universe::update_heap_info_at_gc();
5337       _collectorState = Resizing;
5338     }
5339   }
5340   verify_work_stacks_empty();
5341   verify_overflow_empty();
5342 
5343   if (should_unload_classes()) {
5344     // Delay purge to the beginning of the next safepoint.  Metaspace::contains
5345     // requires that the virtual spaces are stable and not deleted.
5346     ClassLoaderDataGraph::set_should_purge(true);
5347   }
5348 
5349   _intra_sweep_timer.stop();
5350   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5351 
5352   _inter_sweep_timer.reset();
5353   _inter_sweep_timer.start();
5354 
5355   // We need to use a monotonically non-decreasing time in ms
5356   // or we will see time-warp warnings and os::javaTimeMillis()
5357   // does not guarantee monotonicity.
5358   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5359   update_time_of_last_gc(now);
5360 
5361   // NOTE on abstract state transitions:
5362   // Mutators allocate-live and/or mark the mod-union table dirty
5363   // based on the state of the collection.  The former is done in
5364   // the interval [Marking, Sweeping] and the latter in the interval
5365   // [Marking, Sweeping).  Thus the transitions into the Marking state
5366   // and out of the Sweeping state must be synchronously visible
5367   // globally to the mutators.
5368   // The transition into the Marking state happens with the world
5369   // stopped so the mutators will globally see it.  Sweeping is
5370   // done asynchronously by the background collector so the transition
5371   // from the Sweeping state to the Resizing state must be done
5372   // under the freelistLock (as is the check for whether to
5373   // allocate-live and whether to dirty the mod-union table).
5374   assert(_collectorState == Resizing, "Change of collector state to"
5375     " Resizing must be done under the freelistLocks (plural)");
5376 
5377   // Now that sweeping has been completed, we clear
5378   // the incremental_collection_failed flag,
5379   // thus inviting a younger gen collection to promote into
5380   // this generation. If such a promotion may still fail,
5381   // the flag will be set again when a young collection is
5382   // attempted.
5383   GenCollectedHeap* gch = GenCollectedHeap::heap();
5384   gch->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
5385   gch->update_full_collections_completed(_collection_count_start);
5386 }
5387 
5388 // FIX ME!!! Looks like this belongs in CFLSpace, with
5389 // CMSGen merely delegating to it.
5390 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5391   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5392   HeapWord*  minAddr        = _cmsSpace->bottom();
5393   HeapWord*  largestAddr    =
5394     (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5395   if (largestAddr == NULL) {
5396     // The dictionary appears to be empty.  In this case
5397     // try to coalesce at the end of the heap.
5398     largestAddr = _cmsSpace->end();
5399   }
5400   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
5401   size_t nearLargestOffset =
5402     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5403   log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5404                           p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5405   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5406 }
5407 
5408 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5409   return addr >= _cmsSpace->nearLargestChunk();
5410 }
5411 
5412 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5413   return _cmsSpace->find_chunk_at_end();
5414 }
5415 
5416 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5417                                                     bool full) {
5418   // If the young generation has been collected, gather any statistics
5419   // that are of interest at this point.
5420   bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5421   if (!full && current_is_young) {
5422     // Gather statistics on the young generation collection.
5423     collector()->stats().record_gc0_end(used());
5424   }
5425 }
5426 
5427 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5428   // We iterate over the space(s) underlying this generation,
5429   // checking the mark bit map to see if the bits corresponding
5430   // to specific blocks are marked or not. Blocks that are
5431   // marked are live and are not swept up. All remaining blocks
5432   // are swept up, with coalescing on-the-fly as we sweep up
5433   // contiguous free and/or garbage blocks:
5434   // We need to ensure that the sweeper synchronizes with allocators
5435   // and stop-the-world collectors. In particular, the following
5436   // locks are used:
5437   // . CMS token: if this is held, a stop the world collection cannot occur
5438   // . freelistLock: if this is held no allocation can occur from this
5439   //                 generation by another thread
5440   // . bitMapLock: if this is held, no other thread can access or update
5441   //
5442 
5443   // Note that we need to hold the freelistLock if we use
5444   // block iterate below; else the iterator might go awry if
5445   // a mutator (or promotion) causes block contents to change
5446   // (for instance if the allocator divvies up a block).
5447   // If we hold the free list lock, for all practical purposes
5448   // young generation GC's can't occur (they'll usually need to
5449   // promote), so we might as well prevent all young generation
5450   // GC's while we do a sweeping step. For the same reason, we might
5451   // as well take the bit map lock for the entire duration
5452 
5453   // check that we hold the requisite locks
5454   assert(have_cms_token(), "Should hold cms token");
5455   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5456   assert_lock_strong(old_gen->freelistLock());
5457   assert_lock_strong(bitMapLock());
5458 
5459   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5460   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
5461   old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5462                                           _inter_sweep_estimate.padded_average(),
5463                                           _intra_sweep_estimate.padded_average());
5464   old_gen->setNearLargestChunk();
5465 
5466   {
5467     SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5468     old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5469     // We need to free-up/coalesce garbage/blocks from a
5470     // co-terminal free run. This is done in the SweepClosure
5471     // destructor; so, do not remove this scope, else the
5472     // end-of-sweep-census below will be off by a little bit.
5473   }
5474   old_gen->cmsSpace()->sweep_completed();
5475   old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5476   if (should_unload_classes()) {                // unloaded classes this cycle,
5477     _concurrent_cycles_since_last_unload = 0;   // ... reset count
5478   } else {                                      // did not unload classes,
5479     _concurrent_cycles_since_last_unload++;     // ... increment count
5480   }
5481 }
5482 
5483 // Reset CMS data structures (for now just the marking bit map)
5484 // preparatory for the next cycle.
5485 void CMSCollector::reset_concurrent() {
5486   CMSTokenSyncWithLocks ts(true, bitMapLock());
5487 
5488   // If the state is not "Resetting", the foreground  thread
5489   // has done a collection and the resetting.
5490   if (_collectorState != Resetting) {
5491     assert(_collectorState == Idling, "The state should only change"
5492       " because the foreground collector has finished the collection");
5493     return;
5494   }
5495 
5496   {
5497     // Clear the mark bitmap (no grey objects to start with)
5498     // for the next cycle.
5499     GCTraceCPUTime tcpu;
5500     CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5501 
5502     HeapWord* curAddr = _markBitMap.startWord();
5503     while (curAddr < _markBitMap.endWord()) {
5504       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
5505       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5506       _markBitMap.clear_large_range(chunk);
5507       if (ConcurrentMarkSweepThread::should_yield() &&
5508           !foregroundGCIsActive() &&
5509           CMSYield) {
5510         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5511                "CMS thread should hold CMS token");
5512         assert_lock_strong(bitMapLock());
5513         bitMapLock()->unlock();
5514         ConcurrentMarkSweepThread::desynchronize(true);
5515         stopTimer();
5516         incrementYields();
5517 
5518         // See the comment in coordinator_yield()
5519         for (unsigned i = 0; i < CMSYieldSleepCount &&
5520                          ConcurrentMarkSweepThread::should_yield() &&
5521                          !CMSCollector::foregroundGCIsActive(); ++i) {
5522           os::sleep(Thread::current(), 1, false);
5523         }
5524 
5525         ConcurrentMarkSweepThread::synchronize(true);
5526         bitMapLock()->lock_without_safepoint_check();
5527         startTimer();
5528       }
5529       curAddr = chunk.end();
5530     }
5531     // A successful mostly concurrent collection has been done.
5532     // Because only the full (i.e., concurrent mode failure) collections
5533     // are being measured for gc overhead limits, clean the "near" flag
5534     // and count.
5535     size_policy()->reset_gc_overhead_limit_count();
5536     _collectorState = Idling;
5537   }
5538 
5539   register_gc_end();
5540 }
5541 
5542 // Same as above but for STW paths
5543 void CMSCollector::reset_stw() {
5544   // already have the lock
5545   assert(_collectorState == Resetting, "just checking");
5546   assert_lock_strong(bitMapLock());
5547   GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id());
5548   _markBitMap.clear_all();
5549   _collectorState = Idling;
5550   register_gc_end();
5551 }
5552 
5553 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5554   GCTraceCPUTime tcpu;
5555   TraceCollectorStats tcs(EnableConcGCPerfCounter ? NULL : counters());
5556   TraceCollectorStats tcs_cgc(cgc_counters());
5557 
5558   switch (op) {
5559     case CMS_op_checkpointRootsInitial: {
5560       GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5561       SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5562       checkpointRootsInitial();
5563       break;
5564     }
5565     case CMS_op_checkpointRootsFinal: {
5566       GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5567       SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5568       checkpointRootsFinal();
5569       break;
5570     }
5571     default:
5572       fatal("No such CMS_op");
5573   }
5574 }
5575 
5576 #ifndef PRODUCT
5577 size_t const CMSCollector::skip_header_HeapWords() {
5578   return FreeChunk::header_size();
5579 }
5580 
5581 // Try and collect here conditions that should hold when
5582 // CMS thread is exiting. The idea is that the foreground GC
5583 // thread should not be blocked if it wants to terminate
5584 // the CMS thread and yet continue to run the VM for a while
5585 // after that.
5586 void CMSCollector::verify_ok_to_terminate() const {
5587   assert(Thread::current()->is_ConcurrentGC_thread(),
5588          "should be called by CMS thread");
5589   assert(!_foregroundGCShouldWait, "should be false");
5590   // We could check here that all the various low-level locks
5591   // are not held by the CMS thread, but that is overkill; see
5592   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5593   // is checked.
5594 }
5595 #endif
5596 
5597 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5598    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5599           "missing Printezis mark?");
5600   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5601   size_t size = pointer_delta(nextOneAddr + 1, addr);
5602   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5603          "alignment problem");
5604   assert(size >= 3, "Necessary for Printezis marks to work");
5605   return size;
5606 }
5607 
5608 // A variant of the above (block_size_using_printezis_bits()) except
5609 // that we return 0 if the P-bits are not yet set.
5610 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5611   if (_markBitMap.isMarked(addr + 1)) {
5612     assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5613     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5614     size_t size = pointer_delta(nextOneAddr + 1, addr);
5615     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5616            "alignment problem");
5617     assert(size >= 3, "Necessary for Printezis marks to work");
5618     return size;
5619   }
5620   return 0;
5621 }
5622 
5623 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5624   size_t sz = 0;
5625   oop p = (oop)addr;
5626   if (p->klass_or_null_acquire() != NULL) {
5627     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5628   } else {
5629     sz = block_size_using_printezis_bits(addr);
5630   }
5631   assert(sz > 0, "size must be nonzero");
5632   HeapWord* next_block = addr + sz;
5633   HeapWord* next_card  = align_up(next_block, CardTableModRefBS::card_size);
5634   assert(align_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
5635          align_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5636          "must be different cards");
5637   return next_card;
5638 }
5639 
5640 
5641 // CMS Bit Map Wrapper /////////////////////////////////////////
5642 
5643 // Construct a CMS bit map infrastructure, but don't create the
5644 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5645 // further below.
5646 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5647   _bm(),
5648   _shifter(shifter),
5649   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5650                                     Monitor::_safepoint_check_sometimes) : NULL)
5651 {
5652   _bmStartWord = 0;
5653   _bmWordSize  = 0;
5654 }
5655 
5656 bool CMSBitMap::allocate(MemRegion mr) {
5657   _bmStartWord = mr.start();
5658   _bmWordSize  = mr.word_size();
5659   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5660                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5661   if (!brs.is_reserved()) {
5662     log_warning(gc)("CMS bit map allocation failure");
5663     return false;
5664   }
5665   // For now we'll just commit all of the bit map up front.
5666   // Later on we'll try to be more parsimonious with swap.
5667   if (!_virtual_space.initialize(brs, brs.size())) {
5668     log_warning(gc)("CMS bit map backing store failure");
5669     return false;
5670   }
5671   assert(_virtual_space.committed_size() == brs.size(),
5672          "didn't reserve backing store for all of CMS bit map?");
5673   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5674          _bmWordSize, "inconsistency in bit map sizing");
5675   _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5676 
5677   // bm.clear(); // can we rely on getting zero'd memory? verify below
5678   assert(isAllClear(),
5679          "Expected zero'd memory from ReservedSpace constructor");
5680   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5681          "consistency check");
5682   return true;
5683 }
5684 
5685 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5686   HeapWord *next_addr, *end_addr, *last_addr;
5687   assert_locked();
5688   assert(covers(mr), "out-of-range error");
5689   // XXX assert that start and end are appropriately aligned
5690   for (next_addr = mr.start(), end_addr = mr.end();
5691        next_addr < end_addr; next_addr = last_addr) {
5692     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5693     last_addr = dirty_region.end();
5694     if (!dirty_region.is_empty()) {
5695       cl->do_MemRegion(dirty_region);
5696     } else {
5697       assert(last_addr == end_addr, "program logic");
5698       return;
5699     }
5700   }
5701 }
5702 
5703 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5704   _bm.print_on_error(st, prefix);
5705 }
5706 
5707 #ifndef PRODUCT
5708 void CMSBitMap::assert_locked() const {
5709   CMSLockVerifier::assert_locked(lock());
5710 }
5711 
5712 bool CMSBitMap::covers(MemRegion mr) const {
5713   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5714   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5715          "size inconsistency");
5716   return (mr.start() >= _bmStartWord) &&
5717          (mr.end()   <= endWord());
5718 }
5719 
5720 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5721     return (start >= _bmStartWord && (start + size) <= endWord());
5722 }
5723 
5724 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5725   // verify that there are no 1 bits in the interval [left, right)
5726   FalseBitMapClosure falseBitMapClosure;
5727   iterate(&falseBitMapClosure, left, right);
5728 }
5729 
5730 void CMSBitMap::region_invariant(MemRegion mr)
5731 {
5732   assert_locked();
5733   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5734   assert(!mr.is_empty(), "unexpected empty region");
5735   assert(covers(mr), "mr should be covered by bit map");
5736   // convert address range into offset range
5737   size_t start_ofs = heapWordToOffset(mr.start());
5738   // Make sure that end() is appropriately aligned
5739   assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))),
5740          "Misaligned mr.end()");
5741   size_t end_ofs   = heapWordToOffset(mr.end());
5742   assert(end_ofs > start_ofs, "Should mark at least one bit");
5743 }
5744 
5745 #endif
5746 
5747 bool CMSMarkStack::allocate(size_t size) {
5748   // allocate a stack of the requisite depth
5749   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5750                    size * sizeof(oop)));
5751   if (!rs.is_reserved()) {
5752     log_warning(gc)("CMSMarkStack allocation failure");
5753     return false;
5754   }
5755   if (!_virtual_space.initialize(rs, rs.size())) {
5756     log_warning(gc)("CMSMarkStack backing store failure");
5757     return false;
5758   }
5759   assert(_virtual_space.committed_size() == rs.size(),
5760          "didn't reserve backing store for all of CMS stack?");
5761   _base = (oop*)(_virtual_space.low());
5762   _index = 0;
5763   _capacity = size;
5764   NOT_PRODUCT(_max_depth = 0);
5765   return true;
5766 }
5767 
5768 // XXX FIX ME !!! In the MT case we come in here holding a
5769 // leaf lock. For printing we need to take a further lock
5770 // which has lower rank. We need to recalibrate the two
5771 // lock-ranks involved in order to be able to print the
5772 // messages below. (Or defer the printing to the caller.
5773 // For now we take the expedient path of just disabling the
5774 // messages for the problematic case.)
5775 void CMSMarkStack::expand() {
5776   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5777   if (_capacity == MarkStackSizeMax) {
5778     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5779       // We print a warning message only once per CMS cycle.
5780       log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5781     }
5782     return;
5783   }
5784   // Double capacity if possible
5785   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5786   // Do not give up existing stack until we have managed to
5787   // get the double capacity that we desired.
5788   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5789                    new_capacity * sizeof(oop)));
5790   if (rs.is_reserved()) {
5791     // Release the backing store associated with old stack
5792     _virtual_space.release();
5793     // Reinitialize virtual space for new stack
5794     if (!_virtual_space.initialize(rs, rs.size())) {
5795       fatal("Not enough swap for expanded marking stack");
5796     }
5797     _base = (oop*)(_virtual_space.low());
5798     _index = 0;
5799     _capacity = new_capacity;
5800   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5801     // Failed to double capacity, continue;
5802     // we print a detail message only once per CMS cycle.
5803     log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5804                         _capacity / K, new_capacity / K);
5805   }
5806 }
5807 
5808 
5809 // Closures
5810 // XXX: there seems to be a lot of code  duplication here;
5811 // should refactor and consolidate common code.
5812 
5813 // This closure is used to mark refs into the CMS generation in
5814 // the CMS bit map. Called at the first checkpoint. This closure
5815 // assumes that we do not need to re-mark dirty cards; if the CMS
5816 // generation on which this is used is not an oldest
5817 // generation then this will lose younger_gen cards!
5818 
5819 MarkRefsIntoClosure::MarkRefsIntoClosure(
5820   MemRegion span, CMSBitMap* bitMap):
5821     _span(span),
5822     _bitMap(bitMap)
5823 {
5824   assert(ref_processor() == NULL, "deliberately left NULL");
5825   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5826 }
5827 
5828 void MarkRefsIntoClosure::do_oop(oop obj) {
5829   // if p points into _span, then mark corresponding bit in _markBitMap
5830   assert(oopDesc::is_oop(obj), "expected an oop");
5831   HeapWord* addr = (HeapWord*)obj;
5832   if (_span.contains(addr)) {
5833     // this should be made more efficient
5834     _bitMap->mark(addr);
5835   }
5836 }
5837 
5838 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
5839 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5840 
5841 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5842   MemRegion span, CMSBitMap* bitMap):
5843     _span(span),
5844     _bitMap(bitMap)
5845 {
5846   assert(ref_processor() == NULL, "deliberately left NULL");
5847   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5848 }
5849 
5850 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5851   // if p points into _span, then mark corresponding bit in _markBitMap
5852   assert(oopDesc::is_oop(obj), "expected an oop");
5853   HeapWord* addr = (HeapWord*)obj;
5854   if (_span.contains(addr)) {
5855     // this should be made more efficient
5856     _bitMap->par_mark(addr);
5857   }
5858 }
5859 
5860 void ParMarkRefsIntoClosure::do_oop(oop* p)       { ParMarkRefsIntoClosure::do_oop_work(p); }
5861 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5862 
5863 // A variant of the above, used for CMS marking verification.
5864 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5865   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5866     _span(span),
5867     _verification_bm(verification_bm),
5868     _cms_bm(cms_bm)
5869 {
5870   assert(ref_processor() == NULL, "deliberately left NULL");
5871   assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5872 }
5873 
5874 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5875   // if p points into _span, then mark corresponding bit in _markBitMap
5876   assert(oopDesc::is_oop(obj), "expected an oop");
5877   HeapWord* addr = (HeapWord*)obj;
5878   if (_span.contains(addr)) {
5879     _verification_bm->mark(addr);
5880     if (!_cms_bm->isMarked(addr)) {
5881       Log(gc, verify) log;
5882       ResourceMark rm;
5883       LogStream ls(log.error());
5884       oop(addr)->print_on(&ls);
5885       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5886       fatal("... aborting");
5887     }
5888   }
5889 }
5890 
5891 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5892 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5893 
5894 //////////////////////////////////////////////////
5895 // MarkRefsIntoAndScanClosure
5896 //////////////////////////////////////////////////
5897 
5898 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5899                                                        ReferenceProcessor* rp,
5900                                                        CMSBitMap* bit_map,
5901                                                        CMSBitMap* mod_union_table,
5902                                                        CMSMarkStack*  mark_stack,
5903                                                        CMSCollector* collector,
5904                                                        bool should_yield,
5905                                                        bool concurrent_precleaning):
5906   _collector(collector),
5907   _span(span),
5908   _bit_map(bit_map),
5909   _mark_stack(mark_stack),
5910   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
5911                       mark_stack, concurrent_precleaning),
5912   _yield(should_yield),
5913   _concurrent_precleaning(concurrent_precleaning),
5914   _freelistLock(NULL)
5915 {
5916   // FIXME: Should initialize in base class constructor.
5917   assert(rp != NULL, "ref_processor shouldn't be NULL");
5918   set_ref_processor_internal(rp);
5919 }
5920 
5921 // This closure is used to mark refs into the CMS generation at the
5922 // second (final) checkpoint, and to scan and transitively follow
5923 // the unmarked oops. It is also used during the concurrent precleaning
5924 // phase while scanning objects on dirty cards in the CMS generation.
5925 // The marks are made in the marking bit map and the marking stack is
5926 // used for keeping the (newly) grey objects during the scan.
5927 // The parallel version (Par_...) appears further below.
5928 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5929   if (obj != NULL) {
5930     assert(oopDesc::is_oop(obj), "expected an oop");
5931     HeapWord* addr = (HeapWord*)obj;
5932     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5933     assert(_collector->overflow_list_is_empty(),
5934            "overflow list should be empty");
5935     if (_span.contains(addr) &&
5936         !_bit_map->isMarked(addr)) {
5937       // mark bit map (object is now grey)
5938       _bit_map->mark(addr);
5939       // push on marking stack (stack should be empty), and drain the
5940       // stack by applying this closure to the oops in the oops popped
5941       // from the stack (i.e. blacken the grey objects)
5942       bool res = _mark_stack->push(obj);
5943       assert(res, "Should have space to push on empty stack");
5944       do {
5945         oop new_oop = _mark_stack->pop();
5946         assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5947         assert(_bit_map->isMarked((HeapWord*)new_oop),
5948                "only grey objects on this stack");
5949         // iterate over the oops in this oop, marking and pushing
5950         // the ones in CMS heap (i.e. in _span).
5951         new_oop->oop_iterate(&_pushAndMarkClosure);
5952         // check if it's time to yield
5953         do_yield_check();
5954       } while (!_mark_stack->isEmpty() ||
5955                (!_concurrent_precleaning && take_from_overflow_list()));
5956         // if marking stack is empty, and we are not doing this
5957         // during precleaning, then check the overflow list
5958     }
5959     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5960     assert(_collector->overflow_list_is_empty(),
5961            "overflow list was drained above");
5962 
5963     assert(_collector->no_preserved_marks(),
5964            "All preserved marks should have been restored above");
5965   }
5966 }
5967 
5968 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5969 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5970 
5971 void MarkRefsIntoAndScanClosure::do_yield_work() {
5972   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5973          "CMS thread should hold CMS token");
5974   assert_lock_strong(_freelistLock);
5975   assert_lock_strong(_bit_map->lock());
5976   // relinquish the free_list_lock and bitMaplock()
5977   _bit_map->lock()->unlock();
5978   _freelistLock->unlock();
5979   ConcurrentMarkSweepThread::desynchronize(true);
5980   _collector->stopTimer();
5981   _collector->incrementYields();
5982 
5983   // See the comment in coordinator_yield()
5984   for (unsigned i = 0;
5985        i < CMSYieldSleepCount &&
5986        ConcurrentMarkSweepThread::should_yield() &&
5987        !CMSCollector::foregroundGCIsActive();
5988        ++i) {
5989     os::sleep(Thread::current(), 1, false);
5990   }
5991 
5992   ConcurrentMarkSweepThread::synchronize(true);
5993   _freelistLock->lock_without_safepoint_check();
5994   _bit_map->lock()->lock_without_safepoint_check();
5995   _collector->startTimer();
5996 }
5997 
5998 ///////////////////////////////////////////////////////////
5999 // ParMarkRefsIntoAndScanClosure: a parallel version of
6000 //                                MarkRefsIntoAndScanClosure
6001 ///////////////////////////////////////////////////////////
6002 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6003   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6004   CMSBitMap* bit_map, OopTaskQueue* work_queue):
6005   _span(span),
6006   _bit_map(bit_map),
6007   _work_queue(work_queue),
6008   _low_water_mark(MIN2((work_queue->max_elems()/4),
6009                        ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6010   _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6011 {
6012   // FIXME: Should initialize in base class constructor.
6013   assert(rp != NULL, "ref_processor shouldn't be NULL");
6014   set_ref_processor_internal(rp);
6015 }
6016 
6017 // This closure is used to mark refs into the CMS generation at the
6018 // second (final) checkpoint, and to scan and transitively follow
6019 // the unmarked oops. The marks are made in the marking bit map and
6020 // the work_queue is used for keeping the (newly) grey objects during
6021 // the scan phase whence they are also available for stealing by parallel
6022 // threads. Since the marking bit map is shared, updates are
6023 // synchronized (via CAS).
6024 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6025   if (obj != NULL) {
6026     // Ignore mark word because this could be an already marked oop
6027     // that may be chained at the end of the overflow list.
6028     assert(oopDesc::is_oop(obj, true), "expected an oop");
6029     HeapWord* addr = (HeapWord*)obj;
6030     if (_span.contains(addr) &&
6031         !_bit_map->isMarked(addr)) {
6032       // mark bit map (object will become grey):
6033       // It is possible for several threads to be
6034       // trying to "claim" this object concurrently;
6035       // the unique thread that succeeds in marking the
6036       // object first will do the subsequent push on
6037       // to the work queue (or overflow list).
6038       if (_bit_map->par_mark(addr)) {
6039         // push on work_queue (which may not be empty), and trim the
6040         // queue to an appropriate length by applying this closure to
6041         // the oops in the oops popped from the stack (i.e. blacken the
6042         // grey objects)
6043         bool res = _work_queue->push(obj);
6044         assert(res, "Low water mark should be less than capacity?");
6045         trim_queue(_low_water_mark);
6046       } // Else, another thread claimed the object
6047     }
6048   }
6049 }
6050 
6051 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p)       { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6052 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6053 
6054 // This closure is used to rescan the marked objects on the dirty cards
6055 // in the mod union table and the card table proper.
6056 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6057   oop p, MemRegion mr) {
6058 
6059   size_t size = 0;
6060   HeapWord* addr = (HeapWord*)p;
6061   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6062   assert(_span.contains(addr), "we are scanning the CMS generation");
6063   // check if it's time to yield
6064   if (do_yield_check()) {
6065     // We yielded for some foreground stop-world work,
6066     // and we have been asked to abort this ongoing preclean cycle.
6067     return 0;
6068   }
6069   if (_bitMap->isMarked(addr)) {
6070     // it's marked; is it potentially uninitialized?
6071     if (p->klass_or_null_acquire() != NULL) {
6072         // an initialized object; ignore mark word in verification below
6073         // since we are running concurrent with mutators
6074         assert(oopDesc::is_oop(p, true), "should be an oop");
6075         if (p->is_objArray()) {
6076           // objArrays are precisely marked; restrict scanning
6077           // to dirty cards only.
6078           size = CompactibleFreeListSpace::adjustObjectSize(
6079                    p->oop_iterate_size(_scanningClosure, mr));
6080         } else {
6081           // A non-array may have been imprecisely marked; we need
6082           // to scan object in its entirety.
6083           size = CompactibleFreeListSpace::adjustObjectSize(
6084                    p->oop_iterate_size(_scanningClosure));
6085         }
6086       #ifdef ASSERT
6087         size_t direct_size =
6088           CompactibleFreeListSpace::adjustObjectSize(p->size());
6089         assert(size == direct_size, "Inconsistency in size");
6090         assert(size >= 3, "Necessary for Printezis marks to work");
6091         HeapWord* start_pbit = addr + 1;
6092         HeapWord* end_pbit = addr + size - 1;
6093         assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6094                "inconsistent Printezis mark");
6095         // Verify inner mark bits (between Printezis bits) are clear,
6096         // but don't repeat if there are multiple dirty regions for
6097         // the same object, to avoid potential O(N^2) performance.
6098         if (addr != _last_scanned_object) {
6099           _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6100           _last_scanned_object = addr;
6101         }
6102       #endif // ASSERT
6103     } else {
6104       // An uninitialized object.
6105       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6106       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6107       size = pointer_delta(nextOneAddr + 1, addr);
6108       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6109              "alignment problem");
6110       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6111       // will dirty the card when the klass pointer is installed in the
6112       // object (signaling the completion of initialization).
6113     }
6114   } else {
6115     // Either a not yet marked object or an uninitialized object
6116     if (p->klass_or_null_acquire() == NULL) {
6117       // An uninitialized object, skip to the next card, since
6118       // we may not be able to read its P-bits yet.
6119       assert(size == 0, "Initial value");
6120     } else {
6121       // An object not (yet) reached by marking: we merely need to
6122       // compute its size so as to go look at the next block.
6123       assert(oopDesc::is_oop(p, true), "should be an oop");
6124       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6125     }
6126   }
6127   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6128   return size;
6129 }
6130 
6131 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6132   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6133          "CMS thread should hold CMS token");
6134   assert_lock_strong(_freelistLock);
6135   assert_lock_strong(_bitMap->lock());
6136   // relinquish the free_list_lock and bitMaplock()
6137   _bitMap->lock()->unlock();
6138   _freelistLock->unlock();
6139   ConcurrentMarkSweepThread::desynchronize(true);
6140   _collector->stopTimer();
6141   _collector->incrementYields();
6142 
6143   // See the comment in coordinator_yield()
6144   for (unsigned i = 0; i < CMSYieldSleepCount &&
6145                    ConcurrentMarkSweepThread::should_yield() &&
6146                    !CMSCollector::foregroundGCIsActive(); ++i) {
6147     os::sleep(Thread::current(), 1, false);
6148   }
6149 
6150   ConcurrentMarkSweepThread::synchronize(true);
6151   _freelistLock->lock_without_safepoint_check();
6152   _bitMap->lock()->lock_without_safepoint_check();
6153   _collector->startTimer();
6154 }
6155 
6156 
6157 //////////////////////////////////////////////////////////////////
6158 // SurvivorSpacePrecleanClosure
6159 //////////////////////////////////////////////////////////////////
6160 // This (single-threaded) closure is used to preclean the oops in
6161 // the survivor spaces.
6162 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6163 
6164   HeapWord* addr = (HeapWord*)p;
6165   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6166   assert(!_span.contains(addr), "we are scanning the survivor spaces");
6167   assert(p->klass_or_null() != NULL, "object should be initialized");
6168   // an initialized object; ignore mark word in verification below
6169   // since we are running concurrent with mutators
6170   assert(oopDesc::is_oop(p, true), "should be an oop");
6171   // Note that we do not yield while we iterate over
6172   // the interior oops of p, pushing the relevant ones
6173   // on our marking stack.
6174   size_t size = p->oop_iterate_size(_scanning_closure);
6175   do_yield_check();
6176   // Observe that below, we do not abandon the preclean
6177   // phase as soon as we should; rather we empty the
6178   // marking stack before returning. This is to satisfy
6179   // some existing assertions. In general, it may be a
6180   // good idea to abort immediately and complete the marking
6181   // from the grey objects at a later time.
6182   while (!_mark_stack->isEmpty()) {
6183     oop new_oop = _mark_stack->pop();
6184     assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6185     assert(_bit_map->isMarked((HeapWord*)new_oop),
6186            "only grey objects on this stack");
6187     // iterate over the oops in this oop, marking and pushing
6188     // the ones in CMS heap (i.e. in _span).
6189     new_oop->oop_iterate(_scanning_closure);
6190     // check if it's time to yield
6191     do_yield_check();
6192   }
6193   unsigned int after_count =
6194     GenCollectedHeap::heap()->total_collections();
6195   bool abort = (_before_count != after_count) ||
6196                _collector->should_abort_preclean();
6197   return abort ? 0 : size;
6198 }
6199 
6200 void SurvivorSpacePrecleanClosure::do_yield_work() {
6201   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6202          "CMS thread should hold CMS token");
6203   assert_lock_strong(_bit_map->lock());
6204   // Relinquish the bit map lock
6205   _bit_map->lock()->unlock();
6206   ConcurrentMarkSweepThread::desynchronize(true);
6207   _collector->stopTimer();
6208   _collector->incrementYields();
6209 
6210   // See the comment in coordinator_yield()
6211   for (unsigned i = 0; i < CMSYieldSleepCount &&
6212                        ConcurrentMarkSweepThread::should_yield() &&
6213                        !CMSCollector::foregroundGCIsActive(); ++i) {
6214     os::sleep(Thread::current(), 1, false);
6215   }
6216 
6217   ConcurrentMarkSweepThread::synchronize(true);
6218   _bit_map->lock()->lock_without_safepoint_check();
6219   _collector->startTimer();
6220 }
6221 
6222 // This closure is used to rescan the marked objects on the dirty cards
6223 // in the mod union table and the card table proper. In the parallel
6224 // case, although the bitMap is shared, we do a single read so the
6225 // isMarked() query is "safe".
6226 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6227   // Ignore mark word because we are running concurrent with mutators
6228   assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6229   HeapWord* addr = (HeapWord*)p;
6230   assert(_span.contains(addr), "we are scanning the CMS generation");
6231   bool is_obj_array = false;
6232   #ifdef ASSERT
6233     if (!_parallel) {
6234       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6235       assert(_collector->overflow_list_is_empty(),
6236              "overflow list should be empty");
6237 
6238     }
6239   #endif // ASSERT
6240   if (_bit_map->isMarked(addr)) {
6241     // Obj arrays are precisely marked, non-arrays are not;
6242     // so we scan objArrays precisely and non-arrays in their
6243     // entirety.
6244     if (p->is_objArray()) {
6245       is_obj_array = true;
6246       if (_parallel) {
6247         p->oop_iterate(_par_scan_closure, mr);
6248       } else {
6249         p->oop_iterate(_scan_closure, mr);
6250       }
6251     } else {
6252       if (_parallel) {
6253         p->oop_iterate(_par_scan_closure);
6254       } else {
6255         p->oop_iterate(_scan_closure);
6256       }
6257     }
6258   }
6259   #ifdef ASSERT
6260     if (!_parallel) {
6261       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6262       assert(_collector->overflow_list_is_empty(),
6263              "overflow list should be empty");
6264 
6265     }
6266   #endif // ASSERT
6267   return is_obj_array;
6268 }
6269 
6270 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6271                         MemRegion span,
6272                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
6273                         bool should_yield, bool verifying):
6274   _collector(collector),
6275   _span(span),
6276   _bitMap(bitMap),
6277   _mut(&collector->_modUnionTable),
6278   _markStack(markStack),
6279   _yield(should_yield),
6280   _skipBits(0)
6281 {
6282   assert(_markStack->isEmpty(), "stack should be empty");
6283   _finger = _bitMap->startWord();
6284   _threshold = _finger;
6285   assert(_collector->_restart_addr == NULL, "Sanity check");
6286   assert(_span.contains(_finger), "Out of bounds _finger?");
6287   DEBUG_ONLY(_verifying = verifying;)
6288 }
6289 
6290 void MarkFromRootsClosure::reset(HeapWord* addr) {
6291   assert(_markStack->isEmpty(), "would cause duplicates on stack");
6292   assert(_span.contains(addr), "Out of bounds _finger?");
6293   _finger = addr;
6294   _threshold = align_up(_finger, CardTableModRefBS::card_size);
6295 }
6296 
6297 // Should revisit to see if this should be restructured for
6298 // greater efficiency.
6299 bool MarkFromRootsClosure::do_bit(size_t offset) {
6300   if (_skipBits > 0) {
6301     _skipBits--;
6302     return true;
6303   }
6304   // convert offset into a HeapWord*
6305   HeapWord* addr = _bitMap->startWord() + offset;
6306   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6307          "address out of range");
6308   assert(_bitMap->isMarked(addr), "tautology");
6309   if (_bitMap->isMarked(addr+1)) {
6310     // this is an allocated but not yet initialized object
6311     assert(_skipBits == 0, "tautology");
6312     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
6313     oop p = oop(addr);
6314     if (p->klass_or_null_acquire() == NULL) {
6315       DEBUG_ONLY(if (!_verifying) {)
6316         // We re-dirty the cards on which this object lies and increase
6317         // the _threshold so that we'll come back to scan this object
6318         // during the preclean or remark phase. (CMSCleanOnEnter)
6319         if (CMSCleanOnEnter) {
6320           size_t sz = _collector->block_size_using_printezis_bits(addr);
6321           HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
6322           MemRegion redirty_range = MemRegion(addr, end_card_addr);
6323           assert(!redirty_range.is_empty(), "Arithmetical tautology");
6324           // Bump _threshold to end_card_addr; note that
6325           // _threshold cannot possibly exceed end_card_addr, anyhow.
6326           // This prevents future clearing of the card as the scan proceeds
6327           // to the right.
6328           assert(_threshold <= end_card_addr,
6329                  "Because we are just scanning into this object");
6330           if (_threshold < end_card_addr) {
6331             _threshold = end_card_addr;
6332           }
6333           if (p->klass_or_null_acquire() != NULL) {
6334             // Redirty the range of cards...
6335             _mut->mark_range(redirty_range);
6336           } // ...else the setting of klass will dirty the card anyway.
6337         }
6338       DEBUG_ONLY(})
6339       return true;
6340     }
6341   }
6342   scanOopsInOop(addr);
6343   return true;
6344 }
6345 
6346 // We take a break if we've been at this for a while,
6347 // so as to avoid monopolizing the locks involved.
6348 void MarkFromRootsClosure::do_yield_work() {
6349   // First give up the locks, then yield, then re-lock
6350   // We should probably use a constructor/destructor idiom to
6351   // do this unlock/lock or modify the MutexUnlocker class to
6352   // serve our purpose. XXX
6353   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6354          "CMS thread should hold CMS token");
6355   assert_lock_strong(_bitMap->lock());
6356   _bitMap->lock()->unlock();
6357   ConcurrentMarkSweepThread::desynchronize(true);
6358   _collector->stopTimer();
6359   _collector->incrementYields();
6360 
6361   // See the comment in coordinator_yield()
6362   for (unsigned i = 0; i < CMSYieldSleepCount &&
6363                        ConcurrentMarkSweepThread::should_yield() &&
6364                        !CMSCollector::foregroundGCIsActive(); ++i) {
6365     os::sleep(Thread::current(), 1, false);
6366   }
6367 
6368   ConcurrentMarkSweepThread::synchronize(true);
6369   _bitMap->lock()->lock_without_safepoint_check();
6370   _collector->startTimer();
6371 }
6372 
6373 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6374   assert(_bitMap->isMarked(ptr), "expected bit to be set");
6375   assert(_markStack->isEmpty(),
6376          "should drain stack to limit stack usage");
6377   // convert ptr to an oop preparatory to scanning
6378   oop obj = oop(ptr);
6379   // Ignore mark word in verification below, since we
6380   // may be running concurrent with mutators.
6381   assert(oopDesc::is_oop(obj, true), "should be an oop");
6382   assert(_finger <= ptr, "_finger runneth ahead");
6383   // advance the finger to right end of this object
6384   _finger = ptr + obj->size();
6385   assert(_finger > ptr, "we just incremented it above");
6386   // On large heaps, it may take us some time to get through
6387   // the marking phase. During
6388   // this time it's possible that a lot of mutations have
6389   // accumulated in the card table and the mod union table --
6390   // these mutation records are redundant until we have
6391   // actually traced into the corresponding card.
6392   // Here, we check whether advancing the finger would make
6393   // us cross into a new card, and if so clear corresponding
6394   // cards in the MUT (preclean them in the card-table in the
6395   // future).
6396 
6397   DEBUG_ONLY(if (!_verifying) {)
6398     // The clean-on-enter optimization is disabled by default,
6399     // until we fix 6178663.
6400     if (CMSCleanOnEnter && (_finger > _threshold)) {
6401       // [_threshold, _finger) represents the interval
6402       // of cards to be cleared  in MUT (or precleaned in card table).
6403       // The set of cards to be cleared is all those that overlap
6404       // with the interval [_threshold, _finger); note that
6405       // _threshold is always kept card-aligned but _finger isn't
6406       // always card-aligned.
6407       HeapWord* old_threshold = _threshold;
6408       assert(is_aligned(old_threshold, CardTableModRefBS::card_size),
6409              "_threshold should always be card-aligned");
6410       _threshold = align_up(_finger, CardTableModRefBS::card_size);
6411       MemRegion mr(old_threshold, _threshold);
6412       assert(!mr.is_empty(), "Control point invariant");
6413       assert(_span.contains(mr), "Should clear within span");
6414       _mut->clear_range(mr);
6415     }
6416   DEBUG_ONLY(})
6417   // Note: the finger doesn't advance while we drain
6418   // the stack below.
6419   PushOrMarkClosure pushOrMarkClosure(_collector,
6420                                       _span, _bitMap, _markStack,
6421                                       _finger, this);
6422   bool res = _markStack->push(obj);
6423   assert(res, "Empty non-zero size stack should have space for single push");
6424   while (!_markStack->isEmpty()) {
6425     oop new_oop = _markStack->pop();
6426     // Skip verifying header mark word below because we are
6427     // running concurrent with mutators.
6428     assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6429     // now scan this oop's oops
6430     new_oop->oop_iterate(&pushOrMarkClosure);
6431     do_yield_check();
6432   }
6433   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6434 }
6435 
6436 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6437                        CMSCollector* collector, MemRegion span,
6438                        CMSBitMap* bit_map,
6439                        OopTaskQueue* work_queue,
6440                        CMSMarkStack*  overflow_stack):
6441   _collector(collector),
6442   _whole_span(collector->_span),
6443   _span(span),
6444   _bit_map(bit_map),
6445   _mut(&collector->_modUnionTable),
6446   _work_queue(work_queue),
6447   _overflow_stack(overflow_stack),
6448   _skip_bits(0),
6449   _task(task)
6450 {
6451   assert(_work_queue->size() == 0, "work_queue should be empty");
6452   _finger = span.start();
6453   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
6454   assert(_span.contains(_finger), "Out of bounds _finger?");
6455 }
6456 
6457 // Should revisit to see if this should be restructured for
6458 // greater efficiency.
6459 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6460   if (_skip_bits > 0) {
6461     _skip_bits--;
6462     return true;
6463   }
6464   // convert offset into a HeapWord*
6465   HeapWord* addr = _bit_map->startWord() + offset;
6466   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6467          "address out of range");
6468   assert(_bit_map->isMarked(addr), "tautology");
6469   if (_bit_map->isMarked(addr+1)) {
6470     // this is an allocated object that might not yet be initialized
6471     assert(_skip_bits == 0, "tautology");
6472     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
6473     oop p = oop(addr);
6474     if (p->klass_or_null_acquire() == NULL) {
6475       // in the case of Clean-on-Enter optimization, redirty card
6476       // and avoid clearing card by increasing  the threshold.
6477       return true;
6478     }
6479   }
6480   scan_oops_in_oop(addr);
6481   return true;
6482 }
6483 
6484 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6485   assert(_bit_map->isMarked(ptr), "expected bit to be set");
6486   // Should we assert that our work queue is empty or
6487   // below some drain limit?
6488   assert(_work_queue->size() == 0,
6489          "should drain stack to limit stack usage");
6490   // convert ptr to an oop preparatory to scanning
6491   oop obj = oop(ptr);
6492   // Ignore mark word in verification below, since we
6493   // may be running concurrent with mutators.
6494   assert(oopDesc::is_oop(obj, true), "should be an oop");
6495   assert(_finger <= ptr, "_finger runneth ahead");
6496   // advance the finger to right end of this object
6497   _finger = ptr + obj->size();
6498   assert(_finger > ptr, "we just incremented it above");
6499   // On large heaps, it may take us some time to get through
6500   // the marking phase. During
6501   // this time it's possible that a lot of mutations have
6502   // accumulated in the card table and the mod union table --
6503   // these mutation records are redundant until we have
6504   // actually traced into the corresponding card.
6505   // Here, we check whether advancing the finger would make
6506   // us cross into a new card, and if so clear corresponding
6507   // cards in the MUT (preclean them in the card-table in the
6508   // future).
6509 
6510   // The clean-on-enter optimization is disabled by default,
6511   // until we fix 6178663.
6512   if (CMSCleanOnEnter && (_finger > _threshold)) {
6513     // [_threshold, _finger) represents the interval
6514     // of cards to be cleared  in MUT (or precleaned in card table).
6515     // The set of cards to be cleared is all those that overlap
6516     // with the interval [_threshold, _finger); note that
6517     // _threshold is always kept card-aligned but _finger isn't
6518     // always card-aligned.
6519     HeapWord* old_threshold = _threshold;
6520     assert(is_aligned(old_threshold, CardTableModRefBS::card_size),
6521            "_threshold should always be card-aligned");
6522     _threshold = align_up(_finger, CardTableModRefBS::card_size);
6523     MemRegion mr(old_threshold, _threshold);
6524     assert(!mr.is_empty(), "Control point invariant");
6525     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6526     _mut->clear_range(mr);
6527   }
6528 
6529   // Note: the local finger doesn't advance while we drain
6530   // the stack below, but the global finger sure can and will.
6531   HeapWord* volatile* gfa = _task->global_finger_addr();
6532   ParPushOrMarkClosure pushOrMarkClosure(_collector,
6533                                          _span, _bit_map,
6534                                          _work_queue,
6535                                          _overflow_stack,
6536                                          _finger,
6537                                          gfa, this);
6538   bool res = _work_queue->push(obj);   // overflow could occur here
6539   assert(res, "Will hold once we use workqueues");
6540   while (true) {
6541     oop new_oop;
6542     if (!_work_queue->pop_local(new_oop)) {
6543       // We emptied our work_queue; check if there's stuff that can
6544       // be gotten from the overflow stack.
6545       if (CMSConcMarkingTask::get_work_from_overflow_stack(
6546             _overflow_stack, _work_queue)) {
6547         do_yield_check();
6548         continue;
6549       } else {  // done
6550         break;
6551       }
6552     }
6553     // Skip verifying header mark word below because we are
6554     // running concurrent with mutators.
6555     assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6556     // now scan this oop's oops
6557     new_oop->oop_iterate(&pushOrMarkClosure);
6558     do_yield_check();
6559   }
6560   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6561 }
6562 
6563 // Yield in response to a request from VM Thread or
6564 // from mutators.
6565 void ParMarkFromRootsClosure::do_yield_work() {
6566   assert(_task != NULL, "sanity");
6567   _task->yield();
6568 }
6569 
6570 // A variant of the above used for verifying CMS marking work.
6571 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6572                         MemRegion span,
6573                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6574                         CMSMarkStack*  mark_stack):
6575   _collector(collector),
6576   _span(span),
6577   _verification_bm(verification_bm),
6578   _cms_bm(cms_bm),
6579   _mark_stack(mark_stack),
6580   _pam_verify_closure(collector, span, verification_bm, cms_bm,
6581                       mark_stack)
6582 {
6583   assert(_mark_stack->isEmpty(), "stack should be empty");
6584   _finger = _verification_bm->startWord();
6585   assert(_collector->_restart_addr == NULL, "Sanity check");
6586   assert(_span.contains(_finger), "Out of bounds _finger?");
6587 }
6588 
6589 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6590   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6591   assert(_span.contains(addr), "Out of bounds _finger?");
6592   _finger = addr;
6593 }
6594 
6595 // Should revisit to see if this should be restructured for
6596 // greater efficiency.
6597 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6598   // convert offset into a HeapWord*
6599   HeapWord* addr = _verification_bm->startWord() + offset;
6600   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6601          "address out of range");
6602   assert(_verification_bm->isMarked(addr), "tautology");
6603   assert(_cms_bm->isMarked(addr), "tautology");
6604 
6605   assert(_mark_stack->isEmpty(),
6606          "should drain stack to limit stack usage");
6607   // convert addr to an oop preparatory to scanning
6608   oop obj = oop(addr);
6609   assert(oopDesc::is_oop(obj), "should be an oop");
6610   assert(_finger <= addr, "_finger runneth ahead");
6611   // advance the finger to right end of this object
6612   _finger = addr + obj->size();
6613   assert(_finger > addr, "we just incremented it above");
6614   // Note: the finger doesn't advance while we drain
6615   // the stack below.
6616   bool res = _mark_stack->push(obj);
6617   assert(res, "Empty non-zero size stack should have space for single push");
6618   while (!_mark_stack->isEmpty()) {
6619     oop new_oop = _mark_stack->pop();
6620     assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6621     // now scan this oop's oops
6622     new_oop->oop_iterate(&_pam_verify_closure);
6623   }
6624   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6625   return true;
6626 }
6627 
6628 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6629   CMSCollector* collector, MemRegion span,
6630   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6631   CMSMarkStack*  mark_stack):
6632   MetadataAwareOopClosure(collector->ref_processor()),
6633   _collector(collector),
6634   _span(span),
6635   _verification_bm(verification_bm),
6636   _cms_bm(cms_bm),
6637   _mark_stack(mark_stack)
6638 { }
6639 
6640 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
6641 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6642 
6643 // Upon stack overflow, we discard (part of) the stack,
6644 // remembering the least address amongst those discarded
6645 // in CMSCollector's _restart_address.
6646 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6647   // Remember the least grey address discarded
6648   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6649   _collector->lower_restart_addr(ra);
6650   _mark_stack->reset();  // discard stack contents
6651   _mark_stack->expand(); // expand the stack if possible
6652 }
6653 
6654 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6655   assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6656   HeapWord* addr = (HeapWord*)obj;
6657   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6658     // Oop lies in _span and isn't yet grey or black
6659     _verification_bm->mark(addr);            // now grey
6660     if (!_cms_bm->isMarked(addr)) {
6661       Log(gc, verify) log;
6662       ResourceMark rm;
6663       LogStream ls(log.error());
6664       oop(addr)->print_on(&ls);
6665       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6666       fatal("... aborting");
6667     }
6668 
6669     if (!_mark_stack->push(obj)) { // stack overflow
6670       log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6671       assert(_mark_stack->isFull(), "Else push should have succeeded");
6672       handle_stack_overflow(addr);
6673     }
6674     // anything including and to the right of _finger
6675     // will be scanned as we iterate over the remainder of the
6676     // bit map
6677   }
6678 }
6679 
6680 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6681                      MemRegion span,
6682                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
6683                      HeapWord* finger, MarkFromRootsClosure* parent) :
6684   MetadataAwareOopClosure(collector->ref_processor()),
6685   _collector(collector),
6686   _span(span),
6687   _bitMap(bitMap),
6688   _markStack(markStack),
6689   _finger(finger),
6690   _parent(parent)
6691 { }
6692 
6693 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6694                                            MemRegion span,
6695                                            CMSBitMap* bit_map,
6696                                            OopTaskQueue* work_queue,
6697                                            CMSMarkStack*  overflow_stack,
6698                                            HeapWord* finger,
6699                                            HeapWord* volatile* global_finger_addr,
6700                                            ParMarkFromRootsClosure* parent) :
6701   MetadataAwareOopClosure(collector->ref_processor()),
6702   _collector(collector),
6703   _whole_span(collector->_span),
6704   _span(span),
6705   _bit_map(bit_map),
6706   _work_queue(work_queue),
6707   _overflow_stack(overflow_stack),
6708   _finger(finger),
6709   _global_finger_addr(global_finger_addr),
6710   _parent(parent)
6711 { }
6712 
6713 // Assumes thread-safe access by callers, who are
6714 // responsible for mutual exclusion.
6715 void CMSCollector::lower_restart_addr(HeapWord* low) {
6716   assert(_span.contains(low), "Out of bounds addr");
6717   if (_restart_addr == NULL) {
6718     _restart_addr = low;
6719   } else {
6720     _restart_addr = MIN2(_restart_addr, low);
6721   }
6722 }
6723 
6724 // Upon stack overflow, we discard (part of) the stack,
6725 // remembering the least address amongst those discarded
6726 // in CMSCollector's _restart_address.
6727 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6728   // Remember the least grey address discarded
6729   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6730   _collector->lower_restart_addr(ra);
6731   _markStack->reset();  // discard stack contents
6732   _markStack->expand(); // expand the stack if possible
6733 }
6734 
6735 // Upon stack overflow, we discard (part of) the stack,
6736 // remembering the least address amongst those discarded
6737 // in CMSCollector's _restart_address.
6738 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6739   // We need to do this under a mutex to prevent other
6740   // workers from interfering with the work done below.
6741   MutexLockerEx ml(_overflow_stack->par_lock(),
6742                    Mutex::_no_safepoint_check_flag);
6743   // Remember the least grey address discarded
6744   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6745   _collector->lower_restart_addr(ra);
6746   _overflow_stack->reset();  // discard stack contents
6747   _overflow_stack->expand(); // expand the stack if possible
6748 }
6749 
6750 void PushOrMarkClosure::do_oop(oop obj) {
6751   // Ignore mark word because we are running concurrent with mutators.
6752   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6753   HeapWord* addr = (HeapWord*)obj;
6754   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6755     // Oop lies in _span and isn't yet grey or black
6756     _bitMap->mark(addr);            // now grey
6757     if (addr < _finger) {
6758       // the bit map iteration has already either passed, or
6759       // sampled, this bit in the bit map; we'll need to
6760       // use the marking stack to scan this oop's oops.
6761       bool simulate_overflow = false;
6762       NOT_PRODUCT(
6763         if (CMSMarkStackOverflowALot &&
6764             _collector->simulate_overflow()) {
6765           // simulate a stack overflow
6766           simulate_overflow = true;
6767         }
6768       )
6769       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6770         log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6771         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6772         handle_stack_overflow(addr);
6773       }
6774     }
6775     // anything including and to the right of _finger
6776     // will be scanned as we iterate over the remainder of the
6777     // bit map
6778     do_yield_check();
6779   }
6780 }
6781 
6782 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
6783 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6784 
6785 void ParPushOrMarkClosure::do_oop(oop obj) {
6786   // Ignore mark word because we are running concurrent with mutators.
6787   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6788   HeapWord* addr = (HeapWord*)obj;
6789   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6790     // Oop lies in _span and isn't yet grey or black
6791     // We read the global_finger (volatile read) strictly after marking oop
6792     bool res = _bit_map->par_mark(addr);    // now grey
6793     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6794     // Should we push this marked oop on our stack?
6795     // -- if someone else marked it, nothing to do
6796     // -- if target oop is above global finger nothing to do
6797     // -- if target oop is in chunk and above local finger
6798     //      then nothing to do
6799     // -- else push on work queue
6800     if (   !res       // someone else marked it, they will deal with it
6801         || (addr >= *gfa)  // will be scanned in a later task
6802         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6803       return;
6804     }
6805     // the bit map iteration has already either passed, or
6806     // sampled, this bit in the bit map; we'll need to
6807     // use the marking stack to scan this oop's oops.
6808     bool simulate_overflow = false;
6809     NOT_PRODUCT(
6810       if (CMSMarkStackOverflowALot &&
6811           _collector->simulate_overflow()) {
6812         // simulate a stack overflow
6813         simulate_overflow = true;
6814       }
6815     )
6816     if (simulate_overflow ||
6817         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6818       // stack overflow
6819       log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6820       // We cannot assert that the overflow stack is full because
6821       // it may have been emptied since.
6822       assert(simulate_overflow ||
6823              _work_queue->size() == _work_queue->max_elems(),
6824             "Else push should have succeeded");
6825       handle_stack_overflow(addr);
6826     }
6827     do_yield_check();
6828   }
6829 }
6830 
6831 void ParPushOrMarkClosure::do_oop(oop* p)       { ParPushOrMarkClosure::do_oop_work(p); }
6832 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6833 
6834 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6835                                        MemRegion span,
6836                                        ReferenceProcessor* rp,
6837                                        CMSBitMap* bit_map,
6838                                        CMSBitMap* mod_union_table,
6839                                        CMSMarkStack*  mark_stack,
6840                                        bool           concurrent_precleaning):
6841   MetadataAwareOopClosure(rp),
6842   _collector(collector),
6843   _span(span),
6844   _bit_map(bit_map),
6845   _mod_union_table(mod_union_table),
6846   _mark_stack(mark_stack),
6847   _concurrent_precleaning(concurrent_precleaning)
6848 {
6849   assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6850 }
6851 
6852 // Grey object rescan during pre-cleaning and second checkpoint phases --
6853 // the non-parallel version (the parallel version appears further below.)
6854 void PushAndMarkClosure::do_oop(oop obj) {
6855   // Ignore mark word verification. If during concurrent precleaning,
6856   // the object monitor may be locked. If during the checkpoint
6857   // phases, the object may already have been reached by a  different
6858   // path and may be at the end of the global overflow list (so
6859   // the mark word may be NULL).
6860   assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6861          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6862   HeapWord* addr = (HeapWord*)obj;
6863   // Check if oop points into the CMS generation
6864   // and is not marked
6865   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6866     // a white object ...
6867     _bit_map->mark(addr);         // ... now grey
6868     // push on the marking stack (grey set)
6869     bool simulate_overflow = false;
6870     NOT_PRODUCT(
6871       if (CMSMarkStackOverflowALot &&
6872           _collector->simulate_overflow()) {
6873         // simulate a stack overflow
6874         simulate_overflow = true;
6875       }
6876     )
6877     if (simulate_overflow || !_mark_stack->push(obj)) {
6878       if (_concurrent_precleaning) {
6879          // During precleaning we can just dirty the appropriate card(s)
6880          // in the mod union table, thus ensuring that the object remains
6881          // in the grey set  and continue. In the case of object arrays
6882          // we need to dirty all of the cards that the object spans,
6883          // since the rescan of object arrays will be limited to the
6884          // dirty cards.
6885          // Note that no one can be interfering with us in this action
6886          // of dirtying the mod union table, so no locking or atomics
6887          // are required.
6888          if (obj->is_objArray()) {
6889            size_t sz = obj->size();
6890            HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
6891            MemRegion redirty_range = MemRegion(addr, end_card_addr);
6892            assert(!redirty_range.is_empty(), "Arithmetical tautology");
6893            _mod_union_table->mark_range(redirty_range);
6894          } else {
6895            _mod_union_table->mark(addr);
6896          }
6897          _collector->_ser_pmc_preclean_ovflw++;
6898       } else {
6899          // During the remark phase, we need to remember this oop
6900          // in the overflow list.
6901          _collector->push_on_overflow_list(obj);
6902          _collector->_ser_pmc_remark_ovflw++;
6903       }
6904     }
6905   }
6906 }
6907 
6908 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6909                                              MemRegion span,
6910                                              ReferenceProcessor* rp,
6911                                              CMSBitMap* bit_map,
6912                                              OopTaskQueue* work_queue):
6913   MetadataAwareOopClosure(rp),
6914   _collector(collector),
6915   _span(span),
6916   _bit_map(bit_map),
6917   _work_queue(work_queue)
6918 {
6919   assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6920 }
6921 
6922 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
6923 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6924 
6925 // Grey object rescan during second checkpoint phase --
6926 // the parallel version.
6927 void ParPushAndMarkClosure::do_oop(oop obj) {
6928   // In the assert below, we ignore the mark word because
6929   // this oop may point to an already visited object that is
6930   // on the overflow stack (in which case the mark word has
6931   // been hijacked for chaining into the overflow stack --
6932   // if this is the last object in the overflow stack then
6933   // its mark word will be NULL). Because this object may
6934   // have been subsequently popped off the global overflow
6935   // stack, and the mark word possibly restored to the prototypical
6936   // value, by the time we get to examined this failing assert in
6937   // the debugger, is_oop_or_null(false) may subsequently start
6938   // to hold.
6939   assert(oopDesc::is_oop_or_null(obj, true),
6940          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6941   HeapWord* addr = (HeapWord*)obj;
6942   // Check if oop points into the CMS generation
6943   // and is not marked
6944   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6945     // a white object ...
6946     // If we manage to "claim" the object, by being the
6947     // first thread to mark it, then we push it on our
6948     // marking stack
6949     if (_bit_map->par_mark(addr)) {     // ... now grey
6950       // push on work queue (grey set)
6951       bool simulate_overflow = false;
6952       NOT_PRODUCT(
6953         if (CMSMarkStackOverflowALot &&
6954             _collector->par_simulate_overflow()) {
6955           // simulate a stack overflow
6956           simulate_overflow = true;
6957         }
6958       )
6959       if (simulate_overflow || !_work_queue->push(obj)) {
6960         _collector->par_push_on_overflow_list(obj);
6961         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
6962       }
6963     } // Else, some other thread got there first
6964   }
6965 }
6966 
6967 void ParPushAndMarkClosure::do_oop(oop* p)       { ParPushAndMarkClosure::do_oop_work(p); }
6968 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6969 
6970 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6971   Mutex* bml = _collector->bitMapLock();
6972   assert_lock_strong(bml);
6973   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6974          "CMS thread should hold CMS token");
6975 
6976   bml->unlock();
6977   ConcurrentMarkSweepThread::desynchronize(true);
6978 
6979   _collector->stopTimer();
6980   _collector->incrementYields();
6981 
6982   // See the comment in coordinator_yield()
6983   for (unsigned i = 0; i < CMSYieldSleepCount &&
6984                        ConcurrentMarkSweepThread::should_yield() &&
6985                        !CMSCollector::foregroundGCIsActive(); ++i) {
6986     os::sleep(Thread::current(), 1, false);
6987   }
6988 
6989   ConcurrentMarkSweepThread::synchronize(true);
6990   bml->lock();
6991 
6992   _collector->startTimer();
6993 }
6994 
6995 bool CMSPrecleanRefsYieldClosure::should_return() {
6996   if (ConcurrentMarkSweepThread::should_yield()) {
6997     do_yield_work();
6998   }
6999   return _collector->foregroundGCIsActive();
7000 }
7001 
7002 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7003   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7004          "mr should be aligned to start at a card boundary");
7005   // We'd like to assert:
7006   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7007   //        "mr should be a range of cards");
7008   // However, that would be too strong in one case -- the last
7009   // partition ends at _unallocated_block which, in general, can be
7010   // an arbitrary boundary, not necessarily card aligned.
7011   _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words;
7012   _space->object_iterate_mem(mr, &_scan_cl);
7013 }
7014 
7015 SweepClosure::SweepClosure(CMSCollector* collector,
7016                            ConcurrentMarkSweepGeneration* g,
7017                            CMSBitMap* bitMap, bool should_yield) :
7018   _collector(collector),
7019   _g(g),
7020   _sp(g->cmsSpace()),
7021   _limit(_sp->sweep_limit()),
7022   _freelistLock(_sp->freelistLock()),
7023   _bitMap(bitMap),
7024   _yield(should_yield),
7025   _inFreeRange(false),           // No free range at beginning of sweep
7026   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
7027   _lastFreeRangeCoalesced(false),
7028   _freeFinger(g->used_region().start())
7029 {
7030   NOT_PRODUCT(
7031     _numObjectsFreed = 0;
7032     _numWordsFreed   = 0;
7033     _numObjectsLive = 0;
7034     _numWordsLive = 0;
7035     _numObjectsAlreadyFree = 0;
7036     _numWordsAlreadyFree = 0;
7037     _last_fc = NULL;
7038 
7039     _sp->initializeIndexedFreeListArrayReturnedBytes();
7040     _sp->dictionary()->initialize_dict_returned_bytes();
7041   )
7042   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7043          "sweep _limit out of bounds");
7044   log_develop_trace(gc, sweep)("====================");
7045   log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7046 }
7047 
7048 void SweepClosure::print_on(outputStream* st) const {
7049   st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7050                p2i(_sp->bottom()), p2i(_sp->end()));
7051   st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7052   st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7053   NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7054   st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7055                _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7056 }
7057 
7058 #ifndef PRODUCT
7059 // Assertion checking only:  no useful work in product mode --
7060 // however, if any of the flags below become product flags,
7061 // you may need to review this code to see if it needs to be
7062 // enabled in product mode.
7063 SweepClosure::~SweepClosure() {
7064   assert_lock_strong(_freelistLock);
7065   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7066          "sweep _limit out of bounds");
7067   if (inFreeRange()) {
7068     Log(gc, sweep) log;
7069     log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7070     ResourceMark rm;
7071     LogStream ls(log.error());
7072     print_on(&ls);
7073     ShouldNotReachHere();
7074   }
7075 
7076   if (log_is_enabled(Debug, gc, sweep)) {
7077     log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7078                          _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7079     log_debug(gc, sweep)("Live " SIZE_FORMAT " objects,  " SIZE_FORMAT " bytes  Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7080                          _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7081     size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7082     log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7083   }
7084 
7085   if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7086     size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7087     size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7088     size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7089     log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes   Indexed List Returned " SIZE_FORMAT " bytes        Dictionary Returned " SIZE_FORMAT " bytes",
7090                          returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7091   }
7092   log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7093   log_develop_trace(gc, sweep)("================");
7094 }
7095 #endif  // PRODUCT
7096 
7097 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7098     bool freeRangeInFreeLists) {
7099   log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7100                                p2i(freeFinger), freeRangeInFreeLists);
7101   assert(!inFreeRange(), "Trampling existing free range");
7102   set_inFreeRange(true);
7103   set_lastFreeRangeCoalesced(false);
7104 
7105   set_freeFinger(freeFinger);
7106   set_freeRangeInFreeLists(freeRangeInFreeLists);
7107   if (CMSTestInFreeList) {
7108     if (freeRangeInFreeLists) {
7109       FreeChunk* fc = (FreeChunk*) freeFinger;
7110       assert(fc->is_free(), "A chunk on the free list should be free.");
7111       assert(fc->size() > 0, "Free range should have a size");
7112       assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7113     }
7114   }
7115 }
7116 
7117 // Note that the sweeper runs concurrently with mutators. Thus,
7118 // it is possible for direct allocation in this generation to happen
7119 // in the middle of the sweep. Note that the sweeper also coalesces
7120 // contiguous free blocks. Thus, unless the sweeper and the allocator
7121 // synchronize appropriately freshly allocated blocks may get swept up.
7122 // This is accomplished by the sweeper locking the free lists while
7123 // it is sweeping. Thus blocks that are determined to be free are
7124 // indeed free. There is however one additional complication:
7125 // blocks that have been allocated since the final checkpoint and
7126 // mark, will not have been marked and so would be treated as
7127 // unreachable and swept up. To prevent this, the allocator marks
7128 // the bit map when allocating during the sweep phase. This leads,
7129 // however, to a further complication -- objects may have been allocated
7130 // but not yet initialized -- in the sense that the header isn't yet
7131 // installed. The sweeper can not then determine the size of the block
7132 // in order to skip over it. To deal with this case, we use a technique
7133 // (due to Printezis) to encode such uninitialized block sizes in the
7134 // bit map. Since the bit map uses a bit per every HeapWord, but the
7135 // CMS generation has a minimum object size of 3 HeapWords, it follows
7136 // that "normal marks" won't be adjacent in the bit map (there will
7137 // always be at least two 0 bits between successive 1 bits). We make use
7138 // of these "unused" bits to represent uninitialized blocks -- the bit
7139 // corresponding to the start of the uninitialized object and the next
7140 // bit are both set. Finally, a 1 bit marks the end of the object that
7141 // started with the two consecutive 1 bits to indicate its potentially
7142 // uninitialized state.
7143 
7144 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7145   FreeChunk* fc = (FreeChunk*)addr;
7146   size_t res;
7147 
7148   // Check if we are done sweeping. Below we check "addr >= _limit" rather
7149   // than "addr == _limit" because although _limit was a block boundary when
7150   // we started the sweep, it may no longer be one because heap expansion
7151   // may have caused us to coalesce the block ending at the address _limit
7152   // with a newly expanded chunk (this happens when _limit was set to the
7153   // previous _end of the space), so we may have stepped past _limit:
7154   // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7155   if (addr >= _limit) { // we have swept up to or past the limit: finish up
7156     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7157            "sweep _limit out of bounds");
7158     assert(addr < _sp->end(), "addr out of bounds");
7159     // Flush any free range we might be holding as a single
7160     // coalesced chunk to the appropriate free list.
7161     if (inFreeRange()) {
7162       assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7163              "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7164       flush_cur_free_chunk(freeFinger(),
7165                            pointer_delta(addr, freeFinger()));
7166       log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7167                                    p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7168                                    lastFreeRangeCoalesced() ? 1 : 0);
7169     }
7170 
7171     // help the iterator loop finish
7172     return pointer_delta(_sp->end(), addr);
7173   }
7174 
7175   assert(addr < _limit, "sweep invariant");
7176   // check if we should yield
7177   do_yield_check(addr);
7178   if (fc->is_free()) {
7179     // Chunk that is already free
7180     res = fc->size();
7181     do_already_free_chunk(fc);
7182     debug_only(_sp->verifyFreeLists());
7183     // If we flush the chunk at hand in lookahead_and_flush()
7184     // and it's coalesced with a preceding chunk, then the
7185     // process of "mangling" the payload of the coalesced block
7186     // will cause erasure of the size information from the
7187     // (erstwhile) header of all the coalesced blocks but the
7188     // first, so the first disjunct in the assert will not hold
7189     // in that specific case (in which case the second disjunct
7190     // will hold).
7191     assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7192            "Otherwise the size info doesn't change at this step");
7193     NOT_PRODUCT(
7194       _numObjectsAlreadyFree++;
7195       _numWordsAlreadyFree += res;
7196     )
7197     NOT_PRODUCT(_last_fc = fc;)
7198   } else if (!_bitMap->isMarked(addr)) {
7199     // Chunk is fresh garbage
7200     res = do_garbage_chunk(fc);
7201     debug_only(_sp->verifyFreeLists());
7202     NOT_PRODUCT(
7203       _numObjectsFreed++;
7204       _numWordsFreed += res;
7205     )
7206   } else {
7207     // Chunk that is alive.
7208     res = do_live_chunk(fc);
7209     debug_only(_sp->verifyFreeLists());
7210     NOT_PRODUCT(
7211         _numObjectsLive++;
7212         _numWordsLive += res;
7213     )
7214   }
7215   return res;
7216 }
7217 
7218 // For the smart allocation, record following
7219 //  split deaths - a free chunk is removed from its free list because
7220 //      it is being split into two or more chunks.
7221 //  split birth - a free chunk is being added to its free list because
7222 //      a larger free chunk has been split and resulted in this free chunk.
7223 //  coal death - a free chunk is being removed from its free list because
7224 //      it is being coalesced into a large free chunk.
7225 //  coal birth - a free chunk is being added to its free list because
7226 //      it was created when two or more free chunks where coalesced into
7227 //      this free chunk.
7228 //
7229 // These statistics are used to determine the desired number of free
7230 // chunks of a given size.  The desired number is chosen to be relative
7231 // to the end of a CMS sweep.  The desired number at the end of a sweep
7232 // is the
7233 //      count-at-end-of-previous-sweep (an amount that was enough)
7234 //              - count-at-beginning-of-current-sweep  (the excess)
7235 //              + split-births  (gains in this size during interval)
7236 //              - split-deaths  (demands on this size during interval)
7237 // where the interval is from the end of one sweep to the end of the
7238 // next.
7239 //
7240 // When sweeping the sweeper maintains an accumulated chunk which is
7241 // the chunk that is made up of chunks that have been coalesced.  That
7242 // will be termed the left-hand chunk.  A new chunk of garbage that
7243 // is being considered for coalescing will be referred to as the
7244 // right-hand chunk.
7245 //
7246 // When making a decision on whether to coalesce a right-hand chunk with
7247 // the current left-hand chunk, the current count vs. the desired count
7248 // of the left-hand chunk is considered.  Also if the right-hand chunk
7249 // is near the large chunk at the end of the heap (see
7250 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7251 // left-hand chunk is coalesced.
7252 //
7253 // When making a decision about whether to split a chunk, the desired count
7254 // vs. the current count of the candidate to be split is also considered.
7255 // If the candidate is underpopulated (currently fewer chunks than desired)
7256 // a chunk of an overpopulated (currently more chunks than desired) size may
7257 // be chosen.  The "hint" associated with a free list, if non-null, points
7258 // to a free list which may be overpopulated.
7259 //
7260 
7261 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7262   const size_t size = fc->size();
7263   // Chunks that cannot be coalesced are not in the
7264   // free lists.
7265   if (CMSTestInFreeList && !fc->cantCoalesce()) {
7266     assert(_sp->verify_chunk_in_free_list(fc),
7267            "free chunk should be in free lists");
7268   }
7269   // a chunk that is already free, should not have been
7270   // marked in the bit map
7271   HeapWord* const addr = (HeapWord*) fc;
7272   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7273   // Verify that the bit map has no bits marked between
7274   // addr and purported end of this block.
7275   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7276 
7277   // Some chunks cannot be coalesced under any circumstances.
7278   // See the definition of cantCoalesce().
7279   if (!fc->cantCoalesce()) {
7280     // This chunk can potentially be coalesced.
7281     // All the work is done in
7282     do_post_free_or_garbage_chunk(fc, size);
7283     // Note that if the chunk is not coalescable (the else arm
7284     // below), we unconditionally flush, without needing to do
7285     // a "lookahead," as we do below.
7286     if (inFreeRange()) lookahead_and_flush(fc, size);
7287   } else {
7288     // Code path common to both original and adaptive free lists.
7289 
7290     // cant coalesce with previous block; this should be treated
7291     // as the end of a free run if any
7292     if (inFreeRange()) {
7293       // we kicked some butt; time to pick up the garbage
7294       assert(freeFinger() < addr, "freeFinger points too high");
7295       flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7296     }
7297     // else, nothing to do, just continue
7298   }
7299 }
7300 
7301 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7302   // This is a chunk of garbage.  It is not in any free list.
7303   // Add it to a free list or let it possibly be coalesced into
7304   // a larger chunk.
7305   HeapWord* const addr = (HeapWord*) fc;
7306   const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7307 
7308   // Verify that the bit map has no bits marked between
7309   // addr and purported end of just dead object.
7310   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7311   do_post_free_or_garbage_chunk(fc, size);
7312 
7313   assert(_limit >= addr + size,
7314          "A freshly garbage chunk can't possibly straddle over _limit");
7315   if (inFreeRange()) lookahead_and_flush(fc, size);
7316   return size;
7317 }
7318 
7319 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7320   HeapWord* addr = (HeapWord*) fc;
7321   // The sweeper has just found a live object. Return any accumulated
7322   // left hand chunk to the free lists.
7323   if (inFreeRange()) {
7324     assert(freeFinger() < addr, "freeFinger points too high");
7325     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7326   }
7327 
7328   // This object is live: we'd normally expect this to be
7329   // an oop, and like to assert the following:
7330   // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7331   // However, as we commented above, this may be an object whose
7332   // header hasn't yet been initialized.
7333   size_t size;
7334   assert(_bitMap->isMarked(addr), "Tautology for this control point");
7335   if (_bitMap->isMarked(addr + 1)) {
7336     // Determine the size from the bit map, rather than trying to
7337     // compute it from the object header.
7338     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7339     size = pointer_delta(nextOneAddr + 1, addr);
7340     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7341            "alignment problem");
7342 
7343 #ifdef ASSERT
7344       if (oop(addr)->klass_or_null_acquire() != NULL) {
7345         // Ignore mark word because we are running concurrent with mutators
7346         assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7347         assert(size ==
7348                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7349                "P-mark and computed size do not agree");
7350       }
7351 #endif
7352 
7353   } else {
7354     // This should be an initialized object that's alive.
7355     assert(oop(addr)->klass_or_null_acquire() != NULL,
7356            "Should be an initialized object");
7357     // Ignore mark word because we are running concurrent with mutators
7358     assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7359     // Verify that the bit map has no bits marked between
7360     // addr and purported end of this block.
7361     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7362     assert(size >= 3, "Necessary for Printezis marks to work");
7363     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7364     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7365   }
7366   return size;
7367 }
7368 
7369 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7370                                                  size_t chunkSize) {
7371   // do_post_free_or_garbage_chunk() should only be called in the case
7372   // of the adaptive free list allocator.
7373   const bool fcInFreeLists = fc->is_free();
7374   assert((HeapWord*)fc <= _limit, "sweep invariant");
7375   if (CMSTestInFreeList && fcInFreeLists) {
7376     assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7377   }
7378 
7379   log_develop_trace(gc, sweep)("  -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7380 
7381   HeapWord* const fc_addr = (HeapWord*) fc;
7382 
7383   bool coalesce = false;
7384   const size_t left  = pointer_delta(fc_addr, freeFinger());
7385   const size_t right = chunkSize;
7386   switch (FLSCoalescePolicy) {
7387     // numeric value forms a coalition aggressiveness metric
7388     case 0:  { // never coalesce
7389       coalesce = false;
7390       break;
7391     }
7392     case 1: { // coalesce if left & right chunks on overpopulated lists
7393       coalesce = _sp->coalOverPopulated(left) &&
7394                  _sp->coalOverPopulated(right);
7395       break;
7396     }
7397     case 2: { // coalesce if left chunk on overpopulated list (default)
7398       coalesce = _sp->coalOverPopulated(left);
7399       break;
7400     }
7401     case 3: { // coalesce if left OR right chunk on overpopulated list
7402       coalesce = _sp->coalOverPopulated(left) ||
7403                  _sp->coalOverPopulated(right);
7404       break;
7405     }
7406     case 4: { // always coalesce
7407       coalesce = true;
7408       break;
7409     }
7410     default:
7411      ShouldNotReachHere();
7412   }
7413 
7414   // Should the current free range be coalesced?
7415   // If the chunk is in a free range and either we decided to coalesce above
7416   // or the chunk is near the large block at the end of the heap
7417   // (isNearLargestChunk() returns true), then coalesce this chunk.
7418   const bool doCoalesce = inFreeRange()
7419                           && (coalesce || _g->isNearLargestChunk(fc_addr));
7420   if (doCoalesce) {
7421     // Coalesce the current free range on the left with the new
7422     // chunk on the right.  If either is on a free list,
7423     // it must be removed from the list and stashed in the closure.
7424     if (freeRangeInFreeLists()) {
7425       FreeChunk* const ffc = (FreeChunk*)freeFinger();
7426       assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7427              "Size of free range is inconsistent with chunk size.");
7428       if (CMSTestInFreeList) {
7429         assert(_sp->verify_chunk_in_free_list(ffc),
7430                "Chunk is not in free lists");
7431       }
7432       _sp->coalDeath(ffc->size());
7433       _sp->removeFreeChunkFromFreeLists(ffc);
7434       set_freeRangeInFreeLists(false);
7435     }
7436     if (fcInFreeLists) {
7437       _sp->coalDeath(chunkSize);
7438       assert(fc->size() == chunkSize,
7439         "The chunk has the wrong size or is not in the free lists");
7440       _sp->removeFreeChunkFromFreeLists(fc);
7441     }
7442     set_lastFreeRangeCoalesced(true);
7443     print_free_block_coalesced(fc);
7444   } else {  // not in a free range and/or should not coalesce
7445     // Return the current free range and start a new one.
7446     if (inFreeRange()) {
7447       // In a free range but cannot coalesce with the right hand chunk.
7448       // Put the current free range into the free lists.
7449       flush_cur_free_chunk(freeFinger(),
7450                            pointer_delta(fc_addr, freeFinger()));
7451     }
7452     // Set up for new free range.  Pass along whether the right hand
7453     // chunk is in the free lists.
7454     initialize_free_range((HeapWord*)fc, fcInFreeLists);
7455   }
7456 }
7457 
7458 // Lookahead flush:
7459 // If we are tracking a free range, and this is the last chunk that
7460 // we'll look at because its end crosses past _limit, we'll preemptively
7461 // flush it along with any free range we may be holding on to. Note that
7462 // this can be the case only for an already free or freshly garbage
7463 // chunk. If this block is an object, it can never straddle
7464 // over _limit. The "straddling" occurs when _limit is set at
7465 // the previous end of the space when this cycle started, and
7466 // a subsequent heap expansion caused the previously co-terminal
7467 // free block to be coalesced with the newly expanded portion,
7468 // thus rendering _limit a non-block-boundary making it dangerous
7469 // for the sweeper to step over and examine.
7470 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7471   assert(inFreeRange(), "Should only be called if currently in a free range.");
7472   HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7473   assert(_sp->used_region().contains(eob - 1),
7474          "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7475          " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7476          " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7477          p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7478   if (eob >= _limit) {
7479     assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7480     log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7481                                  "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7482                                  "[" PTR_FORMAT "," PTR_FORMAT ")",
7483                                  p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7484     // Return the storage we are tracking back into the free lists.
7485     log_develop_trace(gc, sweep)("Flushing ... ");
7486     assert(freeFinger() < eob, "Error");
7487     flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7488   }
7489 }
7490 
7491 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7492   assert(inFreeRange(), "Should only be called if currently in a free range.");
7493   assert(size > 0,
7494     "A zero sized chunk cannot be added to the free lists.");
7495   if (!freeRangeInFreeLists()) {
7496     if (CMSTestInFreeList) {
7497       FreeChunk* fc = (FreeChunk*) chunk;
7498       fc->set_size(size);
7499       assert(!_sp->verify_chunk_in_free_list(fc),
7500              "chunk should not be in free lists yet");
7501     }
7502     log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7503     // A new free range is going to be starting.  The current
7504     // free range has not been added to the free lists yet or
7505     // was removed so add it back.
7506     // If the current free range was coalesced, then the death
7507     // of the free range was recorded.  Record a birth now.
7508     if (lastFreeRangeCoalesced()) {
7509       _sp->coalBirth(size);
7510     }
7511     _sp->addChunkAndRepairOffsetTable(chunk, size,
7512             lastFreeRangeCoalesced());
7513   } else {
7514     log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7515   }
7516   set_inFreeRange(false);
7517   set_freeRangeInFreeLists(false);
7518 }
7519 
7520 // We take a break if we've been at this for a while,
7521 // so as to avoid monopolizing the locks involved.
7522 void SweepClosure::do_yield_work(HeapWord* addr) {
7523   // Return current free chunk being used for coalescing (if any)
7524   // to the appropriate freelist.  After yielding, the next
7525   // free block encountered will start a coalescing range of
7526   // free blocks.  If the next free block is adjacent to the
7527   // chunk just flushed, they will need to wait for the next
7528   // sweep to be coalesced.
7529   if (inFreeRange()) {
7530     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7531   }
7532 
7533   // First give up the locks, then yield, then re-lock.
7534   // We should probably use a constructor/destructor idiom to
7535   // do this unlock/lock or modify the MutexUnlocker class to
7536   // serve our purpose. XXX
7537   assert_lock_strong(_bitMap->lock());
7538   assert_lock_strong(_freelistLock);
7539   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7540          "CMS thread should hold CMS token");
7541   _bitMap->lock()->unlock();
7542   _freelistLock->unlock();
7543   ConcurrentMarkSweepThread::desynchronize(true);
7544   _collector->stopTimer();
7545   _collector->incrementYields();
7546 
7547   // See the comment in coordinator_yield()
7548   for (unsigned i = 0; i < CMSYieldSleepCount &&
7549                        ConcurrentMarkSweepThread::should_yield() &&
7550                        !CMSCollector::foregroundGCIsActive(); ++i) {
7551     os::sleep(Thread::current(), 1, false);
7552   }
7553 
7554   ConcurrentMarkSweepThread::synchronize(true);
7555   _freelistLock->lock();
7556   _bitMap->lock()->lock_without_safepoint_check();
7557   _collector->startTimer();
7558 }
7559 
7560 #ifndef PRODUCT
7561 // This is actually very useful in a product build if it can
7562 // be called from the debugger.  Compile it into the product
7563 // as needed.
7564 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7565   return debug_cms_space->verify_chunk_in_free_list(fc);
7566 }
7567 #endif
7568 
7569 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7570   log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7571                                p2i(fc), fc->size());
7572 }
7573 
7574 // CMSIsAliveClosure
7575 bool CMSIsAliveClosure::do_object_b(oop obj) {
7576   HeapWord* addr = (HeapWord*)obj;
7577   return addr != NULL &&
7578          (!_span.contains(addr) || _bit_map->isMarked(addr));
7579 }
7580 
7581 
7582 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7583                       MemRegion span,
7584                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7585                       bool cpc):
7586   _collector(collector),
7587   _span(span),
7588   _bit_map(bit_map),
7589   _mark_stack(mark_stack),
7590   _concurrent_precleaning(cpc) {
7591   assert(!_span.is_empty(), "Empty span could spell trouble");
7592 }
7593 
7594 
7595 // CMSKeepAliveClosure: the serial version
7596 void CMSKeepAliveClosure::do_oop(oop obj) {
7597   HeapWord* addr = (HeapWord*)obj;
7598   if (_span.contains(addr) &&
7599       !_bit_map->isMarked(addr)) {
7600     _bit_map->mark(addr);
7601     bool simulate_overflow = false;
7602     NOT_PRODUCT(
7603       if (CMSMarkStackOverflowALot &&
7604           _collector->simulate_overflow()) {
7605         // simulate a stack overflow
7606         simulate_overflow = true;
7607       }
7608     )
7609     if (simulate_overflow || !_mark_stack->push(obj)) {
7610       if (_concurrent_precleaning) {
7611         // We dirty the overflown object and let the remark
7612         // phase deal with it.
7613         assert(_collector->overflow_list_is_empty(), "Error");
7614         // In the case of object arrays, we need to dirty all of
7615         // the cards that the object spans. No locking or atomics
7616         // are needed since no one else can be mutating the mod union
7617         // table.
7618         if (obj->is_objArray()) {
7619           size_t sz = obj->size();
7620           HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
7621           MemRegion redirty_range = MemRegion(addr, end_card_addr);
7622           assert(!redirty_range.is_empty(), "Arithmetical tautology");
7623           _collector->_modUnionTable.mark_range(redirty_range);
7624         } else {
7625           _collector->_modUnionTable.mark(addr);
7626         }
7627         _collector->_ser_kac_preclean_ovflw++;
7628       } else {
7629         _collector->push_on_overflow_list(obj);
7630         _collector->_ser_kac_ovflw++;
7631       }
7632     }
7633   }
7634 }
7635 
7636 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
7637 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7638 
7639 // CMSParKeepAliveClosure: a parallel version of the above.
7640 // The work queues are private to each closure (thread),
7641 // but (may be) available for stealing by other threads.
7642 void CMSParKeepAliveClosure::do_oop(oop obj) {
7643   HeapWord* addr = (HeapWord*)obj;
7644   if (_span.contains(addr) &&
7645       !_bit_map->isMarked(addr)) {
7646     // In general, during recursive tracing, several threads
7647     // may be concurrently getting here; the first one to
7648     // "tag" it, claims it.
7649     if (_bit_map->par_mark(addr)) {
7650       bool res = _work_queue->push(obj);
7651       assert(res, "Low water mark should be much less than capacity");
7652       // Do a recursive trim in the hope that this will keep
7653       // stack usage lower, but leave some oops for potential stealers
7654       trim_queue(_low_water_mark);
7655     } // Else, another thread got there first
7656   }
7657 }
7658 
7659 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
7660 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7661 
7662 void CMSParKeepAliveClosure::trim_queue(uint max) {
7663   while (_work_queue->size() > max) {
7664     oop new_oop;
7665     if (_work_queue->pop_local(new_oop)) {
7666       assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7667       assert(_bit_map->isMarked((HeapWord*)new_oop),
7668              "no white objects on this stack!");
7669       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7670       // iterate over the oops in this oop, marking and pushing
7671       // the ones in CMS heap (i.e. in _span).
7672       new_oop->oop_iterate(&_mark_and_push);
7673     }
7674   }
7675 }
7676 
7677 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7678                                 CMSCollector* collector,
7679                                 MemRegion span, CMSBitMap* bit_map,
7680                                 OopTaskQueue* work_queue):
7681   _collector(collector),
7682   _span(span),
7683   _bit_map(bit_map),
7684   _work_queue(work_queue) { }
7685 
7686 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7687   HeapWord* addr = (HeapWord*)obj;
7688   if (_span.contains(addr) &&
7689       !_bit_map->isMarked(addr)) {
7690     if (_bit_map->par_mark(addr)) {
7691       bool simulate_overflow = false;
7692       NOT_PRODUCT(
7693         if (CMSMarkStackOverflowALot &&
7694             _collector->par_simulate_overflow()) {
7695           // simulate a stack overflow
7696           simulate_overflow = true;
7697         }
7698       )
7699       if (simulate_overflow || !_work_queue->push(obj)) {
7700         _collector->par_push_on_overflow_list(obj);
7701         _collector->_par_kac_ovflw++;
7702       }
7703     } // Else another thread got there already
7704   }
7705 }
7706 
7707 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7708 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7709 
7710 //////////////////////////////////////////////////////////////////
7711 //  CMSExpansionCause                /////////////////////////////
7712 //////////////////////////////////////////////////////////////////
7713 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7714   switch (cause) {
7715     case _no_expansion:
7716       return "No expansion";
7717     case _satisfy_free_ratio:
7718       return "Free ratio";
7719     case _satisfy_promotion:
7720       return "Satisfy promotion";
7721     case _satisfy_allocation:
7722       return "allocation";
7723     case _allocate_par_lab:
7724       return "Par LAB";
7725     case _allocate_par_spooling_space:
7726       return "Par Spooling Space";
7727     case _adaptive_size_policy:
7728       return "Ergonomics";
7729     default:
7730       return "unknown";
7731   }
7732 }
7733 
7734 void CMSDrainMarkingStackClosure::do_void() {
7735   // the max number to take from overflow list at a time
7736   const size_t num = _mark_stack->capacity()/4;
7737   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7738          "Overflow list should be NULL during concurrent phases");
7739   while (!_mark_stack->isEmpty() ||
7740          // if stack is empty, check the overflow list
7741          _collector->take_from_overflow_list(num, _mark_stack)) {
7742     oop obj = _mark_stack->pop();
7743     HeapWord* addr = (HeapWord*)obj;
7744     assert(_span.contains(addr), "Should be within span");
7745     assert(_bit_map->isMarked(addr), "Should be marked");
7746     assert(oopDesc::is_oop(obj), "Should be an oop");
7747     obj->oop_iterate(_keep_alive);
7748   }
7749 }
7750 
7751 void CMSParDrainMarkingStackClosure::do_void() {
7752   // drain queue
7753   trim_queue(0);
7754 }
7755 
7756 // Trim our work_queue so its length is below max at return
7757 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7758   while (_work_queue->size() > max) {
7759     oop new_oop;
7760     if (_work_queue->pop_local(new_oop)) {
7761       assert(oopDesc::is_oop(new_oop), "Expected an oop");
7762       assert(_bit_map->isMarked((HeapWord*)new_oop),
7763              "no white objects on this stack!");
7764       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7765       // iterate over the oops in this oop, marking and pushing
7766       // the ones in CMS heap (i.e. in _span).
7767       new_oop->oop_iterate(&_mark_and_push);
7768     }
7769   }
7770 }
7771 
7772 ////////////////////////////////////////////////////////////////////
7773 // Support for Marking Stack Overflow list handling and related code
7774 ////////////////////////////////////////////////////////////////////
7775 // Much of the following code is similar in shape and spirit to the
7776 // code used in ParNewGC. We should try and share that code
7777 // as much as possible in the future.
7778 
7779 #ifndef PRODUCT
7780 // Debugging support for CMSStackOverflowALot
7781 
7782 // It's OK to call this multi-threaded;  the worst thing
7783 // that can happen is that we'll get a bunch of closely
7784 // spaced simulated overflows, but that's OK, in fact
7785 // probably good as it would exercise the overflow code
7786 // under contention.
7787 bool CMSCollector::simulate_overflow() {
7788   if (_overflow_counter-- <= 0) { // just being defensive
7789     _overflow_counter = CMSMarkStackOverflowInterval;
7790     return true;
7791   } else {
7792     return false;
7793   }
7794 }
7795 
7796 bool CMSCollector::par_simulate_overflow() {
7797   return simulate_overflow();
7798 }
7799 #endif
7800 
7801 // Single-threaded
7802 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7803   assert(stack->isEmpty(), "Expected precondition");
7804   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7805   size_t i = num;
7806   oop  cur = _overflow_list;
7807   const markOop proto = markOopDesc::prototype();
7808   NOT_PRODUCT(ssize_t n = 0;)
7809   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7810     next = oop(cur->mark());
7811     cur->set_mark(proto);   // until proven otherwise
7812     assert(oopDesc::is_oop(cur), "Should be an oop");
7813     bool res = stack->push(cur);
7814     assert(res, "Bit off more than can chew?");
7815     NOT_PRODUCT(n++;)
7816   }
7817   _overflow_list = cur;
7818 #ifndef PRODUCT
7819   assert(_num_par_pushes >= n, "Too many pops?");
7820   _num_par_pushes -=n;
7821 #endif
7822   return !stack->isEmpty();
7823 }
7824 
7825 #define BUSY  (cast_to_oop<intptr_t>(0x1aff1aff))
7826 // (MT-safe) Get a prefix of at most "num" from the list.
7827 // The overflow list is chained through the mark word of
7828 // each object in the list. We fetch the entire list,
7829 // break off a prefix of the right size and return the
7830 // remainder. If other threads try to take objects from
7831 // the overflow list at that time, they will wait for
7832 // some time to see if data becomes available. If (and
7833 // only if) another thread places one or more object(s)
7834 // on the global list before we have returned the suffix
7835 // to the global list, we will walk down our local list
7836 // to find its end and append the global list to
7837 // our suffix before returning it. This suffix walk can
7838 // prove to be expensive (quadratic in the amount of traffic)
7839 // when there are many objects in the overflow list and
7840 // there is much producer-consumer contention on the list.
7841 // *NOTE*: The overflow list manipulation code here and
7842 // in ParNewGeneration:: are very similar in shape,
7843 // except that in the ParNew case we use the old (from/eden)
7844 // copy of the object to thread the list via its klass word.
7845 // Because of the common code, if you make any changes in
7846 // the code below, please check the ParNew version to see if
7847 // similar changes might be needed.
7848 // CR 6797058 has been filed to consolidate the common code.
7849 bool CMSCollector::par_take_from_overflow_list(size_t num,
7850                                                OopTaskQueue* work_q,
7851                                                int no_of_gc_threads) {
7852   assert(work_q->size() == 0, "First empty local work queue");
7853   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7854   if (_overflow_list == NULL) {
7855     return false;
7856   }
7857   // Grab the entire list; we'll put back a suffix
7858   oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
7859   Thread* tid = Thread::current();
7860   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7861   // set to ParallelGCThreads.
7862   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7863   size_t sleep_time_millis = MAX2((size_t)1, num/100);
7864   // If the list is busy, we spin for a short while,
7865   // sleeping between attempts to get the list.
7866   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7867     os::sleep(tid, sleep_time_millis, false);
7868     if (_overflow_list == NULL) {
7869       // Nothing left to take
7870       return false;
7871     } else if (_overflow_list != BUSY) {
7872       // Try and grab the prefix
7873       prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
7874     }
7875   }
7876   // If the list was found to be empty, or we spun long
7877   // enough, we give up and return empty-handed. If we leave
7878   // the list in the BUSY state below, it must be the case that
7879   // some other thread holds the overflow list and will set it
7880   // to a non-BUSY state in the future.
7881   if (prefix == NULL || prefix == BUSY) {
7882      // Nothing to take or waited long enough
7883      if (prefix == NULL) {
7884        // Write back the NULL in case we overwrote it with BUSY above
7885        // and it is still the same value.
7886        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
7887      }
7888      return false;
7889   }
7890   assert(prefix != NULL && prefix != BUSY, "Error");
7891   size_t i = num;
7892   oop cur = prefix;
7893   // Walk down the first "num" objects, unless we reach the end.
7894   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
7895   if (cur->mark() == NULL) {
7896     // We have "num" or fewer elements in the list, so there
7897     // is nothing to return to the global list.
7898     // Write back the NULL in lieu of the BUSY we wrote
7899     // above, if it is still the same value.
7900     if (_overflow_list == BUSY) {
7901       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
7902     }
7903   } else {
7904     // Chop off the suffix and return it to the global list.
7905     assert(cur->mark() != BUSY, "Error");
7906     oop suffix_head = cur->mark(); // suffix will be put back on global list
7907     cur->set_mark(NULL);           // break off suffix
7908     // It's possible that the list is still in the empty(busy) state
7909     // we left it in a short while ago; in that case we may be
7910     // able to place back the suffix without incurring the cost
7911     // of a walk down the list.
7912     oop observed_overflow_list = _overflow_list;
7913     oop cur_overflow_list = observed_overflow_list;
7914     bool attached = false;
7915     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7916       observed_overflow_list =
7917         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
7918       if (cur_overflow_list == observed_overflow_list) {
7919         attached = true;
7920         break;
7921       } else cur_overflow_list = observed_overflow_list;
7922     }
7923     if (!attached) {
7924       // Too bad, someone else sneaked in (at least) an element; we'll need
7925       // to do a splice. Find tail of suffix so we can prepend suffix to global
7926       // list.
7927       for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
7928       oop suffix_tail = cur;
7929       assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
7930              "Tautology");
7931       observed_overflow_list = _overflow_list;
7932       do {
7933         cur_overflow_list = observed_overflow_list;
7934         if (cur_overflow_list != BUSY) {
7935           // Do the splice ...
7936           suffix_tail->set_mark(markOop(cur_overflow_list));
7937         } else { // cur_overflow_list == BUSY
7938           suffix_tail->set_mark(NULL);
7939         }
7940         // ... and try to place spliced list back on overflow_list ...
7941         observed_overflow_list =
7942           (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
7943       } while (cur_overflow_list != observed_overflow_list);
7944       // ... until we have succeeded in doing so.
7945     }
7946   }
7947 
7948   // Push the prefix elements on work_q
7949   assert(prefix != NULL, "control point invariant");
7950   const markOop proto = markOopDesc::prototype();
7951   oop next;
7952   NOT_PRODUCT(ssize_t n = 0;)
7953   for (cur = prefix; cur != NULL; cur = next) {
7954     next = oop(cur->mark());
7955     cur->set_mark(proto);   // until proven otherwise
7956     assert(oopDesc::is_oop(cur), "Should be an oop");
7957     bool res = work_q->push(cur);
7958     assert(res, "Bit off more than we can chew?");
7959     NOT_PRODUCT(n++;)
7960   }
7961 #ifndef PRODUCT
7962   assert(_num_par_pushes >= n, "Too many pops?");
7963   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
7964 #endif
7965   return true;
7966 }
7967 
7968 // Single-threaded
7969 void CMSCollector::push_on_overflow_list(oop p) {
7970   NOT_PRODUCT(_num_par_pushes++;)
7971   assert(oopDesc::is_oop(p), "Not an oop");
7972   preserve_mark_if_necessary(p);
7973   p->set_mark((markOop)_overflow_list);
7974   _overflow_list = p;
7975 }
7976 
7977 // Multi-threaded; use CAS to prepend to overflow list
7978 void CMSCollector::par_push_on_overflow_list(oop p) {
7979   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
7980   assert(oopDesc::is_oop(p), "Not an oop");
7981   par_preserve_mark_if_necessary(p);
7982   oop observed_overflow_list = _overflow_list;
7983   oop cur_overflow_list;
7984   do {
7985     cur_overflow_list = observed_overflow_list;
7986     if (cur_overflow_list != BUSY) {
7987       p->set_mark(markOop(cur_overflow_list));
7988     } else {
7989       p->set_mark(NULL);
7990     }
7991     observed_overflow_list =
7992       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
7993   } while (cur_overflow_list != observed_overflow_list);
7994 }
7995 #undef BUSY
7996 
7997 // Single threaded
7998 // General Note on GrowableArray: pushes may silently fail
7999 // because we are (temporarily) out of C-heap for expanding
8000 // the stack. The problem is quite ubiquitous and affects
8001 // a lot of code in the JVM. The prudent thing for GrowableArray
8002 // to do (for now) is to exit with an error. However, that may
8003 // be too draconian in some cases because the caller may be
8004 // able to recover without much harm. For such cases, we
8005 // should probably introduce a "soft_push" method which returns
8006 // an indication of success or failure with the assumption that
8007 // the caller may be able to recover from a failure; code in
8008 // the VM can then be changed, incrementally, to deal with such
8009 // failures where possible, thus, incrementally hardening the VM
8010 // in such low resource situations.
8011 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8012   _preserved_oop_stack.push(p);
8013   _preserved_mark_stack.push(m);
8014   assert(m == p->mark(), "Mark word changed");
8015   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8016          "bijection");
8017 }
8018 
8019 // Single threaded
8020 void CMSCollector::preserve_mark_if_necessary(oop p) {
8021   markOop m = p->mark();
8022   if (m->must_be_preserved(p)) {
8023     preserve_mark_work(p, m);
8024   }
8025 }
8026 
8027 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8028   markOop m = p->mark();
8029   if (m->must_be_preserved(p)) {
8030     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8031     // Even though we read the mark word without holding
8032     // the lock, we are assured that it will not change
8033     // because we "own" this oop, so no other thread can
8034     // be trying to push it on the overflow list; see
8035     // the assertion in preserve_mark_work() that checks
8036     // that m == p->mark().
8037     preserve_mark_work(p, m);
8038   }
8039 }
8040 
8041 // We should be able to do this multi-threaded,
8042 // a chunk of stack being a task (this is
8043 // correct because each oop only ever appears
8044 // once in the overflow list. However, it's
8045 // not very easy to completely overlap this with
8046 // other operations, so will generally not be done
8047 // until all work's been completed. Because we
8048 // expect the preserved oop stack (set) to be small,
8049 // it's probably fine to do this single-threaded.
8050 // We can explore cleverer concurrent/overlapped/parallel
8051 // processing of preserved marks if we feel the
8052 // need for this in the future. Stack overflow should
8053 // be so rare in practice and, when it happens, its
8054 // effect on performance so great that this will
8055 // likely just be in the noise anyway.
8056 void CMSCollector::restore_preserved_marks_if_any() {
8057   assert(SafepointSynchronize::is_at_safepoint(),
8058          "world should be stopped");
8059   assert(Thread::current()->is_ConcurrentGC_thread() ||
8060          Thread::current()->is_VM_thread(),
8061          "should be single-threaded");
8062   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8063          "bijection");
8064 
8065   while (!_preserved_oop_stack.is_empty()) {
8066     oop p = _preserved_oop_stack.pop();
8067     assert(oopDesc::is_oop(p), "Should be an oop");
8068     assert(_span.contains(p), "oop should be in _span");
8069     assert(p->mark() == markOopDesc::prototype(),
8070            "Set when taken from overflow list");
8071     markOop m = _preserved_mark_stack.pop();
8072     p->set_mark(m);
8073   }
8074   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8075          "stacks were cleared above");
8076 }
8077 
8078 #ifndef PRODUCT
8079 bool CMSCollector::no_preserved_marks() const {
8080   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8081 }
8082 #endif
8083 
8084 // Transfer some number of overflown objects to usual marking
8085 // stack. Return true if some objects were transferred.
8086 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8087   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8088                     (size_t)ParGCDesiredObjsFromOverflowList);
8089 
8090   bool res = _collector->take_from_overflow_list(num, _mark_stack);
8091   assert(_collector->overflow_list_is_empty() || res,
8092          "If list is not empty, we should have taken something");
8093   assert(!res || !_mark_stack->isEmpty(),
8094          "If we took something, it should now be on our stack");
8095   return res;
8096 }
8097 
8098 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8099   size_t res = _sp->block_size_no_stall(addr, _collector);
8100   if (_sp->block_is_obj(addr)) {
8101     if (_live_bit_map->isMarked(addr)) {
8102       // It can't have been dead in a previous cycle
8103       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8104     } else {
8105       _dead_bit_map->mark(addr);      // mark the dead object
8106     }
8107   }
8108   // Could be 0, if the block size could not be computed without stalling.
8109   return res;
8110 }
8111 
8112 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8113 
8114   switch (phase) {
8115     case CMSCollector::InitialMarking:
8116       initialize(true  /* fullGC */ ,
8117                  cause /* cause of the GC */,
8118                  true  /* recordGCBeginTime */,
8119                  true  /* recordPreGCUsage */,
8120                  false /* recordPeakUsage */,
8121                  false /* recordPostGCusage */,
8122                  true  /* recordAccumulatedGCTime */,
8123                  false /* recordGCEndTime */,
8124                  false /* countCollection */  );
8125       break;
8126 
8127     case CMSCollector::FinalMarking:
8128       initialize(true  /* fullGC */ ,
8129                  cause /* cause of the GC */,
8130                  false /* recordGCBeginTime */,
8131                  false /* recordPreGCUsage */,
8132                  false /* recordPeakUsage */,
8133                  false /* recordPostGCusage */,
8134                  true  /* recordAccumulatedGCTime */,
8135                  false /* recordGCEndTime */,
8136                  false /* countCollection */  );
8137       break;
8138 
8139     case CMSCollector::Sweeping:
8140       initialize(true  /* fullGC */ ,
8141                  cause /* cause of the GC */,
8142                  false /* recordGCBeginTime */,
8143                  false /* recordPreGCUsage */,
8144                  true  /* recordPeakUsage */,
8145                  true  /* recordPostGCusage */,
8146                  false /* recordAccumulatedGCTime */,
8147                  true  /* recordGCEndTime */,
8148                  true  /* countCollection */  );
8149       break;
8150 
8151     default:
8152       ShouldNotReachHere();
8153   }
8154 }