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