1 /*
   2  * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/classLoaderData.hpp"
  27 #include "classfile/stringTable.hpp"
  28 #include "classfile/symbolTable.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/codeCache.hpp"
  31 #include "gc/cms/cmsCollectorPolicy.hpp"
  32 #include "gc/cms/cmsOopClosures.inline.hpp"
  33 #include "gc/cms/compactibleFreeListSpace.hpp"
  34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
  35 #include "gc/cms/concurrentMarkSweepThread.hpp"
  36 #include "gc/cms/parNewGeneration.hpp"
  37 #include "gc/cms/vmCMSOperations.hpp"
  38 #include "gc/serial/genMarkSweep.hpp"
  39 #include "gc/serial/tenuredGeneration.hpp"
  40 #include "gc/shared/adaptiveSizePolicy.hpp"
  41 #include "gc/shared/cardGeneration.inline.hpp"
  42 #include "gc/shared/cardTableRS.hpp"
  43 #include "gc/shared/collectedHeap.inline.hpp"
  44 #include "gc/shared/collectorCounters.hpp"
  45 #include "gc/shared/collectorPolicy.hpp"
  46 #include "gc/shared/gcLocker.inline.hpp"
  47 #include "gc/shared/gcPolicyCounters.hpp"
  48 #include "gc/shared/gcTimer.hpp"
  49 #include "gc/shared/gcTrace.hpp"
  50 #include "gc/shared/gcTraceTime.inline.hpp"
  51 #include "gc/shared/genCollectedHeap.hpp"
  52 #include "gc/shared/genOopClosures.inline.hpp"
  53 #include "gc/shared/isGCActiveMark.hpp"
  54 #include "gc/shared/referencePolicy.hpp"
  55 #include "gc/shared/strongRootsScope.hpp"
  56 #include "gc/shared/taskqueue.inline.hpp"
  57 #include "logging/log.hpp"
  58 #include "logging/logStream.hpp"
  59 #include "memory/allocation.hpp"
  60 #include "memory/iterator.inline.hpp"
  61 #include "memory/padded.hpp"
  62 #include "memory/resourceArea.hpp"
  63 #include "oops/oop.inline.hpp"
  64 #include "prims/jvmtiExport.hpp"
  65 #include "runtime/atomic.hpp"
  66 #include "runtime/globals_extension.hpp"
  67 #include "runtime/handles.inline.hpp"
  68 #include "runtime/java.hpp"
  69 #include "runtime/orderAccess.inline.hpp"
  70 #include "runtime/timer.hpp"
  71 #include "runtime/vmThread.hpp"
  72 #include "services/memoryService.hpp"
  73 #include "services/runtimeService.hpp"
  74 #include "utilities/align.hpp"
  75 #include "utilities/stack.inline.hpp"
  76 
  77 // statics
  78 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
  79 bool CMSCollector::_full_gc_requested = false;
  80 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
  81 
  82 //////////////////////////////////////////////////////////////////
  83 // In support of CMS/VM thread synchronization
  84 //////////////////////////////////////////////////////////////////
  85 // We split use of the CGC_lock into 2 "levels".
  86 // The low-level locking is of the usual CGC_lock monitor. We introduce
  87 // a higher level "token" (hereafter "CMS token") built on top of the
  88 // low level monitor (hereafter "CGC lock").
  89 // The token-passing protocol gives priority to the VM thread. The
  90 // CMS-lock doesn't provide any fairness guarantees, but clients
  91 // should ensure that it is only held for very short, bounded
  92 // durations.
  93 //
  94 // When either of the CMS thread or the VM thread is involved in
  95 // collection operations during which it does not want the other
  96 // thread to interfere, it obtains the CMS token.
  97 //
  98 // If either thread tries to get the token while the other has
  99 // it, that thread waits. However, if the VM thread and CMS thread
 100 // both want the token, then the VM thread gets priority while the
 101 // CMS thread waits. This ensures, for instance, that the "concurrent"
 102 // phases of the CMS thread's work do not block out the VM thread
 103 // for long periods of time as the CMS thread continues to hog
 104 // the token. (See bug 4616232).
 105 //
 106 // The baton-passing functions are, however, controlled by the
 107 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
 108 // and here the low-level CMS lock, not the high level token,
 109 // ensures mutual exclusion.
 110 //
 111 // Two important conditions that we have to satisfy:
 112 // 1. if a thread does a low-level wait on the CMS lock, then it
 113 //    relinquishes the CMS token if it were holding that token
 114 //    when it acquired the low-level CMS lock.
 115 // 2. any low-level notifications on the low-level lock
 116 //    should only be sent when a thread has relinquished the token.
 117 //
 118 // In the absence of either property, we'd have potential deadlock.
 119 //
 120 // We protect each of the CMS (concurrent and sequential) phases
 121 // with the CMS _token_, not the CMS _lock_.
 122 //
 123 // The only code protected by CMS lock is the token acquisition code
 124 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
 125 // baton-passing code.
 126 //
 127 // Unfortunately, i couldn't come up with a good abstraction to factor and
 128 // hide the naked CGC_lock manipulation in the baton-passing code
 129 // further below. That's something we should try to do. Also, the proof
 130 // of correctness of this 2-level locking scheme is far from obvious,
 131 // and potentially quite slippery. We have an uneasy suspicion, for instance,
 132 // that there may be a theoretical possibility of delay/starvation in the
 133 // low-level lock/wait/notify scheme used for the baton-passing because of
 134 // potential interference with the priority scheme embodied in the
 135 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
 136 // invocation further below and marked with "XXX 20011219YSR".
 137 // Indeed, as we note elsewhere, this may become yet more slippery
 138 // in the presence of multiple CMS and/or multiple VM threads. XXX
 139 
 140 class CMSTokenSync: public StackObj {
 141  private:
 142   bool _is_cms_thread;
 143  public:
 144   CMSTokenSync(bool is_cms_thread):
 145     _is_cms_thread(is_cms_thread) {
 146     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
 147            "Incorrect argument to constructor");
 148     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
 149   }
 150 
 151   ~CMSTokenSync() {
 152     assert(_is_cms_thread ?
 153              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
 154              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
 155           "Incorrect state");
 156     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
 157   }
 158 };
 159 
 160 // Convenience class that does a CMSTokenSync, and then acquires
 161 // upto three locks.
 162 class CMSTokenSyncWithLocks: public CMSTokenSync {
 163  private:
 164   // Note: locks are acquired in textual declaration order
 165   // and released in the opposite order
 166   MutexLockerEx _locker1, _locker2, _locker3;
 167  public:
 168   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
 169                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
 170     CMSTokenSync(is_cms_thread),
 171     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
 172     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
 173     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
 174   { }
 175 };
 176 
 177 
 178 //////////////////////////////////////////////////////////////////
 179 //  Concurrent Mark-Sweep Generation /////////////////////////////
 180 //////////////////////////////////////////////////////////////////
 181 
 182 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
 183 
 184 // This struct contains per-thread things necessary to support parallel
 185 // young-gen collection.
 186 class CMSParGCThreadState: public CHeapObj<mtGC> {
 187  public:
 188   CompactibleFreeListSpaceLAB lab;
 189   PromotionInfo promo;
 190 
 191   // Constructor.
 192   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
 193     promo.setSpace(cfls);
 194   }
 195 };
 196 
 197 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
 198      ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
 199   CardGeneration(rs, initial_byte_size, ct),
 200   _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
 201   _did_compact(false)
 202 {
 203   HeapWord* bottom = (HeapWord*) _virtual_space.low();
 204   HeapWord* end    = (HeapWord*) _virtual_space.high();
 205 
 206   _direct_allocated_words = 0;
 207   NOT_PRODUCT(
 208     _numObjectsPromoted = 0;
 209     _numWordsPromoted = 0;
 210     _numObjectsAllocated = 0;
 211     _numWordsAllocated = 0;
 212   )
 213 
 214   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
 215   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
 216   _cmsSpace->_old_gen = this;
 217 
 218   _gc_stats = new CMSGCStats();
 219 
 220   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
 221   // offsets match. The ability to tell free chunks from objects
 222   // depends on this property.
 223   debug_only(
 224     FreeChunk* junk = NULL;
 225     assert(UseCompressedClassPointers ||
 226            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
 227            "Offset of FreeChunk::_prev within FreeChunk must match"
 228            "  that of OopDesc::_klass within OopDesc");
 229   )
 230 
 231   _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
 232   for (uint i = 0; i < ParallelGCThreads; i++) {
 233     _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
 234   }
 235 
 236   _incremental_collection_failed = false;
 237   // The "dilatation_factor" is the expansion that can occur on
 238   // account of the fact that the minimum object size in the CMS
 239   // generation may be larger than that in, say, a contiguous young
 240   //  generation.
 241   // Ideally, in the calculation below, we'd compute the dilatation
 242   // factor as: MinChunkSize/(promoting_gen's min object size)
 243   // Since we do not have such a general query interface for the
 244   // promoting generation, we'll instead just use the minimum
 245   // object size (which today is a header's worth of space);
 246   // note that all arithmetic is in units of HeapWords.
 247   assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
 248   assert(_dilatation_factor >= 1.0, "from previous assert");
 249 }
 250 
 251 
 252 // The field "_initiating_occupancy" represents the occupancy percentage
 253 // at which we trigger a new collection cycle.  Unless explicitly specified
 254 // via CMSInitiatingOccupancyFraction (argument "io" below), it
 255 // is calculated by:
 256 //
 257 //   Let "f" be MinHeapFreeRatio in
 258 //
 259 //    _initiating_occupancy = 100-f +
 260 //                           f * (CMSTriggerRatio/100)
 261 //   where CMSTriggerRatio is the argument "tr" below.
 262 //
 263 // That is, if we assume the heap is at its desired maximum occupancy at the
 264 // end of a collection, we let CMSTriggerRatio of the (purported) free
 265 // space be allocated before initiating a new collection cycle.
 266 //
 267 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
 268   assert(io <= 100 && tr <= 100, "Check the arguments");
 269   if (io >= 0) {
 270     _initiating_occupancy = (double)io / 100.0;
 271   } else {
 272     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
 273                              (double)(tr * MinHeapFreeRatio) / 100.0)
 274                             / 100.0;
 275   }
 276 }
 277 
 278 void ConcurrentMarkSweepGeneration::ref_processor_init() {
 279   assert(collector() != NULL, "no collector");
 280   collector()->ref_processor_init();
 281 }
 282 
 283 void CMSCollector::ref_processor_init() {
 284   if (_ref_processor == NULL) {
 285     // Allocate and initialize a reference processor
 286     _ref_processor =
 287       new ReferenceProcessor(_span,                               // span
 288                              (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
 289                              ParallelGCThreads,                   // mt processing degree
 290                              _cmsGen->refs_discovery_is_mt(),     // mt discovery
 291                              MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
 292                              _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
 293                              &_is_alive_closure);                 // closure for liveness info
 294     // Initialize the _ref_processor field of CMSGen
 295     _cmsGen->set_ref_processor(_ref_processor);
 296 
 297   }
 298 }
 299 
 300 AdaptiveSizePolicy* CMSCollector::size_policy() {
 301   GenCollectedHeap* gch = GenCollectedHeap::heap();
 302   return gch->gen_policy()->size_policy();
 303 }
 304 
 305 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
 306 
 307   const char* gen_name = "old";
 308   GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy();
 309   // Generation Counters - generation 1, 1 subspace
 310   _gen_counters = new GenerationCounters(gen_name, 1, 1,
 311       gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
 312 
 313   _space_counters = new GSpaceCounters(gen_name, 0,
 314                                        _virtual_space.reserved_size(),
 315                                        this, _gen_counters);
 316 }
 317 
 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
 319   _cms_gen(cms_gen)
 320 {
 321   assert(alpha <= 100, "bad value");
 322   _saved_alpha = alpha;
 323 
 324   // Initialize the alphas to the bootstrap value of 100.
 325   _gc0_alpha = _cms_alpha = 100;
 326 
 327   _cms_begin_time.update();
 328   _cms_end_time.update();
 329 
 330   _gc0_duration = 0.0;
 331   _gc0_period = 0.0;
 332   _gc0_promoted = 0;
 333 
 334   _cms_duration = 0.0;
 335   _cms_period = 0.0;
 336   _cms_allocated = 0;
 337 
 338   _cms_used_at_gc0_begin = 0;
 339   _cms_used_at_gc0_end = 0;
 340   _allow_duty_cycle_reduction = false;
 341   _valid_bits = 0;
 342 }
 343 
 344 double CMSStats::cms_free_adjustment_factor(size_t free) const {
 345   // TBD: CR 6909490
 346   return 1.0;
 347 }
 348 
 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
 350 }
 351 
 352 // If promotion failure handling is on use
 353 // the padded average size of the promotion for each
 354 // young generation collection.
 355 double CMSStats::time_until_cms_gen_full() const {
 356   size_t cms_free = _cms_gen->cmsSpace()->free();
 357   GenCollectedHeap* gch = GenCollectedHeap::heap();
 358   size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
 359                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
 360   if (cms_free > expected_promotion) {
 361     // Start a cms collection if there isn't enough space to promote
 362     // for the next young collection.  Use the padded average as
 363     // a safety factor.
 364     cms_free -= expected_promotion;
 365 
 366     // Adjust by the safety factor.
 367     double cms_free_dbl = (double)cms_free;
 368     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
 369     // Apply a further correction factor which tries to adjust
 370     // for recent occurance of concurrent mode failures.
 371     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
 372     cms_free_dbl = cms_free_dbl * cms_adjustment;
 373 
 374     log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
 375                   cms_free, expected_promotion);
 376     log_trace(gc)("  cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
 377     // Add 1 in case the consumption rate goes to zero.
 378     return cms_free_dbl / (cms_consumption_rate() + 1.0);
 379   }
 380   return 0.0;
 381 }
 382 
 383 // Compare the duration of the cms collection to the
 384 // time remaining before the cms generation is empty.
 385 // Note that the time from the start of the cms collection
 386 // to the start of the cms sweep (less than the total
 387 // duration of the cms collection) can be used.  This
 388 // has been tried and some applications experienced
 389 // promotion failures early in execution.  This was
 390 // possibly because the averages were not accurate
 391 // enough at the beginning.
 392 double CMSStats::time_until_cms_start() const {
 393   // We add "gc0_period" to the "work" calculation
 394   // below because this query is done (mostly) at the
 395   // end of a scavenge, so we need to conservatively
 396   // account for that much possible delay
 397   // in the query so as to avoid concurrent mode failures
 398   // due to starting the collection just a wee bit too
 399   // late.
 400   double work = cms_duration() + gc0_period();
 401   double deadline = time_until_cms_gen_full();
 402   // If a concurrent mode failure occurred recently, we want to be
 403   // more conservative and halve our expected time_until_cms_gen_full()
 404   if (work > deadline) {
 405     log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
 406                           cms_duration(), gc0_period(), time_until_cms_gen_full());
 407     return 0.0;
 408   }
 409   return work - deadline;
 410 }
 411 
 412 #ifndef PRODUCT
 413 void CMSStats::print_on(outputStream *st) const {
 414   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
 415   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
 416                gc0_duration(), gc0_period(), gc0_promoted());
 417   st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
 418             cms_duration(), cms_period(), cms_allocated());
 419   st->print(",cms_since_beg=%g,cms_since_end=%g",
 420             cms_time_since_begin(), cms_time_since_end());
 421   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
 422             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
 423 
 424   if (valid()) {
 425     st->print(",promo_rate=%g,cms_alloc_rate=%g",
 426               promotion_rate(), cms_allocation_rate());
 427     st->print(",cms_consumption_rate=%g,time_until_full=%g",
 428               cms_consumption_rate(), time_until_cms_gen_full());
 429   }
 430   st->cr();
 431 }
 432 #endif // #ifndef PRODUCT
 433 
 434 CMSCollector::CollectorState CMSCollector::_collectorState =
 435                              CMSCollector::Idling;
 436 bool CMSCollector::_foregroundGCIsActive = false;
 437 bool CMSCollector::_foregroundGCShouldWait = false;
 438 
 439 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
 440                            CardTableRS*                   ct,
 441                            ConcurrentMarkSweepPolicy*     cp):
 442   _cmsGen(cmsGen),
 443   _ct(ct),
 444   _ref_processor(NULL),    // will be set later
 445   _conc_workers(NULL),     // may be set later
 446   _abort_preclean(false),
 447   _start_sampling(false),
 448   _between_prologue_and_epilogue(false),
 449   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
 450   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
 451                  -1 /* lock-free */, "No_lock" /* dummy */),
 452   _modUnionClosurePar(&_modUnionTable),
 453   // Adjust my span to cover old (cms) gen
 454   _span(cmsGen->reserved()),
 455   // Construct the is_alive_closure with _span & markBitMap
 456   _is_alive_closure(_span, &_markBitMap),
 457   _restart_addr(NULL),
 458   _overflow_list(NULL),
 459   _stats(cmsGen),
 460   _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
 461                              //verify that this lock should be acquired with safepoint check.
 462                              Monitor::_safepoint_check_sometimes)),
 463   _eden_chunk_array(NULL),     // may be set in ctor body
 464   _eden_chunk_capacity(0),     // -- ditto --
 465   _eden_chunk_index(0),        // -- ditto --
 466   _survivor_plab_array(NULL),  // -- ditto --
 467   _survivor_chunk_array(NULL), // -- ditto --
 468   _survivor_chunk_capacity(0), // -- ditto --
 469   _survivor_chunk_index(0),    // -- ditto --
 470   _ser_pmc_preclean_ovflw(0),
 471   _ser_kac_preclean_ovflw(0),
 472   _ser_pmc_remark_ovflw(0),
 473   _par_pmc_remark_ovflw(0),
 474   _ser_kac_ovflw(0),
 475   _par_kac_ovflw(0),
 476 #ifndef PRODUCT
 477   _num_par_pushes(0),
 478 #endif
 479   _collection_count_start(0),
 480   _verifying(false),
 481   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
 482   _completed_initialization(false),
 483   _collector_policy(cp),
 484   _should_unload_classes(CMSClassUnloadingEnabled),
 485   _concurrent_cycles_since_last_unload(0),
 486   _roots_scanning_options(GenCollectedHeap::SO_None),
 487   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 488   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 489   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
 490   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 491   _cms_start_registered(false)
 492 {
 493   // Now expand the span and allocate the collection support structures
 494   // (MUT, marking bit map etc.) to cover both generations subject to
 495   // collection.
 496 
 497   // For use by dirty card to oop closures.
 498   _cmsGen->cmsSpace()->set_collector(this);
 499 
 500   // Allocate MUT and marking bit map
 501   {
 502     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
 503     if (!_markBitMap.allocate(_span)) {
 504       log_warning(gc)("Failed to allocate CMS Bit Map");
 505       return;
 506     }
 507     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
 508   }
 509   {
 510     _modUnionTable.allocate(_span);
 511     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
 512   }
 513 
 514   if (!_markStack.allocate(MarkStackSize)) {
 515     log_warning(gc)("Failed to allocate CMS Marking Stack");
 516     return;
 517   }
 518 
 519   // Support for multi-threaded concurrent phases
 520   if (CMSConcurrentMTEnabled) {
 521     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
 522       // just for now
 523       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
 524     }
 525     if (ConcGCThreads > 1) {
 526       _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
 527                                  ConcGCThreads, true);
 528       if (_conc_workers == NULL) {
 529         log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
 530         CMSConcurrentMTEnabled = false;
 531       } else {
 532         _conc_workers->initialize_workers();
 533       }
 534     } else {
 535       CMSConcurrentMTEnabled = false;
 536     }
 537   }
 538   if (!CMSConcurrentMTEnabled) {
 539     ConcGCThreads = 0;
 540   } else {
 541     // Turn off CMSCleanOnEnter optimization temporarily for
 542     // the MT case where it's not fixed yet; see 6178663.
 543     CMSCleanOnEnter = false;
 544   }
 545   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
 546          "Inconsistency");
 547   log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
 548   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 549 
 550   // Parallel task queues; these are shared for the
 551   // concurrent and stop-world phases of CMS, but
 552   // are not shared with parallel scavenge (ParNew).
 553   {
 554     uint i;
 555     uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
 556 
 557     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
 558          || ParallelRefProcEnabled)
 559         && num_queues > 0) {
 560       _task_queues = new OopTaskQueueSet(num_queues);
 561       if (_task_queues == NULL) {
 562         log_warning(gc)("task_queues allocation failure.");
 563         return;
 564       }
 565       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
 566       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
 567       for (i = 0; i < num_queues; i++) {
 568         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
 569         if (q == NULL) {
 570           log_warning(gc)("work_queue allocation failure.");
 571           return;
 572         }
 573         _task_queues->register_queue(i, q);
 574       }
 575       for (i = 0; i < num_queues; i++) {
 576         _task_queues->queue(i)->initialize();
 577         _hash_seed[i] = 17;  // copied from ParNew
 578       }
 579     }
 580   }
 581 
 582   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
 583 
 584   // Clip CMSBootstrapOccupancy between 0 and 100.
 585   _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
 586 
 587   // Now tell CMS generations the identity of their collector
 588   ConcurrentMarkSweepGeneration::set_collector(this);
 589 
 590   // Create & start a CMS thread for this CMS collector
 591   _cmsThread = ConcurrentMarkSweepThread::start(this);
 592   assert(cmsThread() != NULL, "CMS Thread should have been created");
 593   assert(cmsThread()->collector() == this,
 594          "CMS Thread should refer to this gen");
 595   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 596 
 597   // Support for parallelizing young gen rescan
 598   GenCollectedHeap* gch = GenCollectedHeap::heap();
 599   assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
 600   _young_gen = (ParNewGeneration*)gch->young_gen();
 601   if (gch->supports_inline_contig_alloc()) {
 602     _top_addr = gch->top_addr();
 603     _end_addr = gch->end_addr();
 604     assert(_young_gen != NULL, "no _young_gen");
 605     _eden_chunk_index = 0;
 606     _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
 607     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
 608   }
 609 
 610   // Support for parallelizing survivor space rescan
 611   if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
 612     const size_t max_plab_samples =
 613       _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
 614 
 615     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
 616     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 617     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
 618     _survivor_chunk_capacity = max_plab_samples;
 619     for (uint i = 0; i < ParallelGCThreads; i++) {
 620       HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 621       ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
 622       assert(cur->end() == 0, "Should be 0");
 623       assert(cur->array() == vec, "Should be vec");
 624       assert(cur->capacity() == max_plab_samples, "Error");
 625     }
 626   }
 627 
 628   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
 629   _gc_counters = new CollectorCounters("CMS", 1);
 630   _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     LogStream ls(log.trace());
 700     cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
 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(oopDesc::is_oop(obj), "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(oopDesc::is_oop(old), "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(oopDesc::is_oop(obj) && 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       LogStream ls(log.error());
2252       oop(addr)->print_on(&ls);
2253       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2254       _failed = true;
2255     }
2256     return true;
2257   }
2258 
2259   bool failed() { return _failed; }
2260 };
2261 
2262 bool CMSCollector::verify_after_remark() {
2263   GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2264   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2265   static bool init = false;
2266 
2267   assert(SafepointSynchronize::is_at_safepoint(),
2268          "Else mutations in object graph will make answer suspect");
2269   assert(have_cms_token(),
2270          "Else there may be mutual interference in use of "
2271          " verification data structures");
2272   assert(_collectorState > Marking && _collectorState <= Sweeping,
2273          "Else marking info checked here may be obsolete");
2274   assert(haveFreelistLocks(), "must hold free list locks");
2275   assert_lock_strong(bitMapLock());
2276 
2277 
2278   // Allocate marking bit map if not already allocated
2279   if (!init) { // first time
2280     if (!verification_mark_bm()->allocate(_span)) {
2281       return false;
2282     }
2283     init = true;
2284   }
2285 
2286   assert(verification_mark_stack()->isEmpty(), "Should be empty");
2287 
2288   // Turn off refs discovery -- so we will be tracing through refs.
2289   // This is as intended, because by this time
2290   // GC must already have cleared any refs that need to be cleared,
2291   // and traced those that need to be marked; moreover,
2292   // the marking done here is not going to interfere in any
2293   // way with the marking information used by GC.
2294   NoRefDiscovery no_discovery(ref_processor());
2295 
2296 #if defined(COMPILER2) || INCLUDE_JVMCI
2297   DerivedPointerTableDeactivate dpt_deact;
2298 #endif
2299 
2300   // Clear any marks from a previous round
2301   verification_mark_bm()->clear_all();
2302   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2303   verify_work_stacks_empty();
2304 
2305   GenCollectedHeap* gch = GenCollectedHeap::heap();
2306   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2307   // Update the saved marks which may affect the root scans.
2308   gch->save_marks();
2309 
2310   if (CMSRemarkVerifyVariant == 1) {
2311     // In this first variant of verification, we complete
2312     // all marking, then check if the new marks-vector is
2313     // a subset of the CMS marks-vector.
2314     verify_after_remark_work_1();
2315   } else {
2316     guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2317     // In this second variant of verification, we flag an error
2318     // (i.e. an object reachable in the new marks-vector not reachable
2319     // in the CMS marks-vector) immediately, also indicating the
2320     // identify of an object (A) that references the unmarked object (B) --
2321     // presumably, a mutation to A failed to be picked up by preclean/remark?
2322     verify_after_remark_work_2();
2323   }
2324 
2325   return true;
2326 }
2327 
2328 void CMSCollector::verify_after_remark_work_1() {
2329   ResourceMark rm;
2330   HandleMark  hm;
2331   GenCollectedHeap* gch = GenCollectedHeap::heap();
2332 
2333   // Get a clear set of claim bits for the roots processing to work with.
2334   ClassLoaderDataGraph::clear_claimed_marks();
2335 
2336   // Mark from roots one level into CMS
2337   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2338   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2339 
2340   {
2341     StrongRootsScope srs(1);
2342 
2343     gch->cms_process_roots(&srs,
2344                            true,   // young gen as roots
2345                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2346                            should_unload_classes(),
2347                            &notOlder,
2348                            NULL);
2349   }
2350 
2351   // Now mark from the roots
2352   MarkFromRootsClosure markFromRootsClosure(this, _span,
2353     verification_mark_bm(), verification_mark_stack(),
2354     false /* don't yield */, true /* verifying */);
2355   assert(_restart_addr == NULL, "Expected pre-condition");
2356   verification_mark_bm()->iterate(&markFromRootsClosure);
2357   while (_restart_addr != NULL) {
2358     // Deal with stack overflow: by restarting at the indicated
2359     // address.
2360     HeapWord* ra = _restart_addr;
2361     markFromRootsClosure.reset(ra);
2362     _restart_addr = NULL;
2363     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2364   }
2365   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2366   verify_work_stacks_empty();
2367 
2368   // Marking completed -- now verify that each bit marked in
2369   // verification_mark_bm() is also marked in markBitMap(); flag all
2370   // errors by printing corresponding objects.
2371   VerifyMarkedClosure vcl(markBitMap());
2372   verification_mark_bm()->iterate(&vcl);
2373   if (vcl.failed()) {
2374     Log(gc, verify) log;
2375     log.error("Failed marking verification after remark");
2376     ResourceMark rm;
2377     LogStream ls(log.error());
2378     gch->print_on(&ls);
2379     fatal("CMS: failed marking verification after remark");
2380   }
2381 }
2382 
2383 class VerifyKlassOopsKlassClosure : public KlassClosure {
2384   class VerifyKlassOopsClosure : public OopClosure {
2385     CMSBitMap* _bitmap;
2386    public:
2387     VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2388     void do_oop(oop* p)       { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2389     void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2390   } _oop_closure;
2391  public:
2392   VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2393   void do_klass(Klass* k) {
2394     k->oops_do(&_oop_closure);
2395   }
2396 };
2397 
2398 void CMSCollector::verify_after_remark_work_2() {
2399   ResourceMark rm;
2400   HandleMark  hm;
2401   GenCollectedHeap* gch = GenCollectedHeap::heap();
2402 
2403   // Get a clear set of claim bits for the roots processing to work with.
2404   ClassLoaderDataGraph::clear_claimed_marks();
2405 
2406   // Mark from roots one level into CMS
2407   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2408                                      markBitMap());
2409   CLDToOopClosure cld_closure(&notOlder, true);
2410 
2411   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2412 
2413   {
2414     StrongRootsScope srs(1);
2415 
2416     gch->cms_process_roots(&srs,
2417                            true,   // young gen as roots
2418                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2419                            should_unload_classes(),
2420                            &notOlder,
2421                            &cld_closure);
2422   }
2423 
2424   // Now mark from the roots
2425   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2426     verification_mark_bm(), markBitMap(), verification_mark_stack());
2427   assert(_restart_addr == NULL, "Expected pre-condition");
2428   verification_mark_bm()->iterate(&markFromRootsClosure);
2429   while (_restart_addr != NULL) {
2430     // Deal with stack overflow: by restarting at the indicated
2431     // address.
2432     HeapWord* ra = _restart_addr;
2433     markFromRootsClosure.reset(ra);
2434     _restart_addr = NULL;
2435     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2436   }
2437   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2438   verify_work_stacks_empty();
2439 
2440   VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2441   ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2442 
2443   // Marking completed -- now verify that each bit marked in
2444   // verification_mark_bm() is also marked in markBitMap(); flag all
2445   // errors by printing corresponding objects.
2446   VerifyMarkedClosure vcl(markBitMap());
2447   verification_mark_bm()->iterate(&vcl);
2448   assert(!vcl.failed(), "Else verification above should not have succeeded");
2449 }
2450 
2451 void ConcurrentMarkSweepGeneration::save_marks() {
2452   // delegate to CMS space
2453   cmsSpace()->save_marks();
2454 }
2455 
2456 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2457   return cmsSpace()->no_allocs_since_save_marks();
2458 }
2459 
2460 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
2461                                                                 \
2462 void ConcurrentMarkSweepGeneration::                            \
2463 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
2464   cl->set_generation(this);                                     \
2465   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
2466   cl->reset_generation();                                       \
2467   save_marks();                                                 \
2468 }
2469 
2470 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2471 
2472 void
2473 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2474   if (freelistLock()->owned_by_self()) {
2475     Generation::oop_iterate(cl);
2476   } else {
2477     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2478     Generation::oop_iterate(cl);
2479   }
2480 }
2481 
2482 void
2483 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2484   if (freelistLock()->owned_by_self()) {
2485     Generation::object_iterate(cl);
2486   } else {
2487     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2488     Generation::object_iterate(cl);
2489   }
2490 }
2491 
2492 void
2493 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2494   if (freelistLock()->owned_by_self()) {
2495     Generation::safe_object_iterate(cl);
2496   } else {
2497     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2498     Generation::safe_object_iterate(cl);
2499   }
2500 }
2501 
2502 void
2503 ConcurrentMarkSweepGeneration::post_compact() {
2504 }
2505 
2506 void
2507 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2508   // Fix the linear allocation blocks to look like free blocks.
2509 
2510   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2511   // are not called when the heap is verified during universe initialization and
2512   // at vm shutdown.
2513   if (freelistLock()->owned_by_self()) {
2514     cmsSpace()->prepare_for_verify();
2515   } else {
2516     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2517     cmsSpace()->prepare_for_verify();
2518   }
2519 }
2520 
2521 void
2522 ConcurrentMarkSweepGeneration::verify() {
2523   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2524   // are not called when the heap is verified during universe initialization and
2525   // at vm shutdown.
2526   if (freelistLock()->owned_by_self()) {
2527     cmsSpace()->verify();
2528   } else {
2529     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2530     cmsSpace()->verify();
2531   }
2532 }
2533 
2534 void CMSCollector::verify() {
2535   _cmsGen->verify();
2536 }
2537 
2538 #ifndef PRODUCT
2539 bool CMSCollector::overflow_list_is_empty() const {
2540   assert(_num_par_pushes >= 0, "Inconsistency");
2541   if (_overflow_list == NULL) {
2542     assert(_num_par_pushes == 0, "Inconsistency");
2543   }
2544   return _overflow_list == NULL;
2545 }
2546 
2547 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2548 // merely consolidate assertion checks that appear to occur together frequently.
2549 void CMSCollector::verify_work_stacks_empty() const {
2550   assert(_markStack.isEmpty(), "Marking stack should be empty");
2551   assert(overflow_list_is_empty(), "Overflow list should be empty");
2552 }
2553 
2554 void CMSCollector::verify_overflow_empty() const {
2555   assert(overflow_list_is_empty(), "Overflow list should be empty");
2556   assert(no_preserved_marks(), "No preserved marks");
2557 }
2558 #endif // PRODUCT
2559 
2560 // Decide if we want to enable class unloading as part of the
2561 // ensuing concurrent GC cycle. We will collect and
2562 // unload classes if it's the case that:
2563 //  (a) class unloading is enabled at the command line, and
2564 //  (b) old gen is getting really full
2565 // NOTE: Provided there is no change in the state of the heap between
2566 // calls to this method, it should have idempotent results. Moreover,
2567 // its results should be monotonically increasing (i.e. going from 0 to 1,
2568 // but not 1 to 0) between successive calls between which the heap was
2569 // not collected. For the implementation below, it must thus rely on
2570 // the property that concurrent_cycles_since_last_unload()
2571 // will not decrease unless a collection cycle happened and that
2572 // _cmsGen->is_too_full() are
2573 // themselves also monotonic in that sense. See check_monotonicity()
2574 // below.
2575 void CMSCollector::update_should_unload_classes() {
2576   _should_unload_classes = false;
2577   if (CMSClassUnloadingEnabled) {
2578     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2579                               CMSClassUnloadingMaxInterval)
2580                            || _cmsGen->is_too_full();
2581   }
2582 }
2583 
2584 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2585   bool res = should_concurrent_collect();
2586   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2587   return res;
2588 }
2589 
2590 void CMSCollector::setup_cms_unloading_and_verification_state() {
2591   const  bool should_verify =   VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2592                              || VerifyBeforeExit;
2593   const  int  rso           =   GenCollectedHeap::SO_AllCodeCache;
2594 
2595   // We set the proper root for this CMS cycle here.
2596   if (should_unload_classes()) {   // Should unload classes this cycle
2597     remove_root_scanning_option(rso);  // Shrink the root set appropriately
2598     set_verifying(should_verify);    // Set verification state for this cycle
2599     return;                            // Nothing else needs to be done at this time
2600   }
2601 
2602   // Not unloading classes this cycle
2603   assert(!should_unload_classes(), "Inconsistency!");
2604 
2605   // If we are not unloading classes then add SO_AllCodeCache to root
2606   // scanning options.
2607   add_root_scanning_option(rso);
2608 
2609   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2610     set_verifying(true);
2611   } else if (verifying() && !should_verify) {
2612     // We were verifying, but some verification flags got disabled.
2613     set_verifying(false);
2614     // Exclude symbols, strings and code cache elements from root scanning to
2615     // reduce IM and RM pauses.
2616     remove_root_scanning_option(rso);
2617   }
2618 }
2619 
2620 
2621 #ifndef PRODUCT
2622 HeapWord* CMSCollector::block_start(const void* p) const {
2623   const HeapWord* addr = (HeapWord*)p;
2624   if (_span.contains(p)) {
2625     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2626       return _cmsGen->cmsSpace()->block_start(p);
2627     }
2628   }
2629   return NULL;
2630 }
2631 #endif
2632 
2633 HeapWord*
2634 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2635                                                    bool   tlab,
2636                                                    bool   parallel) {
2637   CMSSynchronousYieldRequest yr;
2638   assert(!tlab, "Can't deal with TLAB allocation");
2639   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2640   expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2641   if (GCExpandToAllocateDelayMillis > 0) {
2642     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2643   }
2644   return have_lock_and_allocate(word_size, tlab);
2645 }
2646 
2647 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2648     size_t bytes,
2649     size_t expand_bytes,
2650     CMSExpansionCause::Cause cause)
2651 {
2652 
2653   bool success = expand(bytes, expand_bytes);
2654 
2655   // remember why we expanded; this information is used
2656   // by shouldConcurrentCollect() when making decisions on whether to start
2657   // a new CMS cycle.
2658   if (success) {
2659     set_expansion_cause(cause);
2660     log_trace(gc)("Expanded CMS gen for %s",  CMSExpansionCause::to_string(cause));
2661   }
2662 }
2663 
2664 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2665   HeapWord* res = NULL;
2666   MutexLocker x(ParGCRareEvent_lock);
2667   while (true) {
2668     // Expansion by some other thread might make alloc OK now:
2669     res = ps->lab.alloc(word_sz);
2670     if (res != NULL) return res;
2671     // If there's not enough expansion space available, give up.
2672     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2673       return NULL;
2674     }
2675     // Otherwise, we try expansion.
2676     expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2677     // Now go around the loop and try alloc again;
2678     // A competing par_promote might beat us to the expansion space,
2679     // so we may go around the loop again if promotion fails again.
2680     if (GCExpandToAllocateDelayMillis > 0) {
2681       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2682     }
2683   }
2684 }
2685 
2686 
2687 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2688   PromotionInfo* promo) {
2689   MutexLocker x(ParGCRareEvent_lock);
2690   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2691   while (true) {
2692     // Expansion by some other thread might make alloc OK now:
2693     if (promo->ensure_spooling_space()) {
2694       assert(promo->has_spooling_space(),
2695              "Post-condition of successful ensure_spooling_space()");
2696       return true;
2697     }
2698     // If there's not enough expansion space available, give up.
2699     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2700       return false;
2701     }
2702     // Otherwise, we try expansion.
2703     expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2704     // Now go around the loop and try alloc again;
2705     // A competing allocation might beat us to the expansion space,
2706     // so we may go around the loop again if allocation fails again.
2707     if (GCExpandToAllocateDelayMillis > 0) {
2708       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2709     }
2710   }
2711 }
2712 
2713 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2714   // Only shrink if a compaction was done so that all the free space
2715   // in the generation is in a contiguous block at the end.
2716   if (did_compact()) {
2717     CardGeneration::shrink(bytes);
2718   }
2719 }
2720 
2721 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2722   assert_locked_or_safepoint(Heap_lock);
2723 }
2724 
2725 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2726   assert_locked_or_safepoint(Heap_lock);
2727   assert_lock_strong(freelistLock());
2728   log_trace(gc)("Shrinking of CMS not yet implemented");
2729   return;
2730 }
2731 
2732 
2733 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2734 // phases.
2735 class CMSPhaseAccounting: public StackObj {
2736  public:
2737   CMSPhaseAccounting(CMSCollector *collector,
2738                      const char *title);
2739   ~CMSPhaseAccounting();
2740 
2741  private:
2742   CMSCollector *_collector;
2743   const char *_title;
2744   GCTraceConcTime(Info, gc) _trace_time;
2745 
2746  public:
2747   // Not MT-safe; so do not pass around these StackObj's
2748   // where they may be accessed by other threads.
2749   double wallclock_millis() {
2750     return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2751   }
2752 };
2753 
2754 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2755                                        const char *title) :
2756   _collector(collector), _title(title), _trace_time(title) {
2757 
2758   _collector->resetYields();
2759   _collector->resetTimer();
2760   _collector->startTimer();
2761   _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2762 }
2763 
2764 CMSPhaseAccounting::~CMSPhaseAccounting() {
2765   _collector->gc_timer_cm()->register_gc_concurrent_end();
2766   _collector->stopTimer();
2767   log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2768   log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2769 }
2770 
2771 // CMS work
2772 
2773 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2774 class CMSParMarkTask : public AbstractGangTask {
2775  protected:
2776   CMSCollector*     _collector;
2777   uint              _n_workers;
2778   CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2779       AbstractGangTask(name),
2780       _collector(collector),
2781       _n_workers(n_workers) {}
2782   // Work method in support of parallel rescan ... of young gen spaces
2783   void do_young_space_rescan(OopsInGenClosure* cl,
2784                              ContiguousSpace* space,
2785                              HeapWord** chunk_array, size_t chunk_top);
2786   void work_on_young_gen_roots(OopsInGenClosure* cl);
2787 };
2788 
2789 // Parallel initial mark task
2790 class CMSParInitialMarkTask: public CMSParMarkTask {
2791   StrongRootsScope* _strong_roots_scope;
2792  public:
2793   CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2794       CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2795       _strong_roots_scope(strong_roots_scope) {}
2796   void work(uint worker_id);
2797 };
2798 
2799 // Checkpoint the roots into this generation from outside
2800 // this generation. [Note this initial checkpoint need only
2801 // be approximate -- we'll do a catch up phase subsequently.]
2802 void CMSCollector::checkpointRootsInitial() {
2803   assert(_collectorState == InitialMarking, "Wrong collector state");
2804   check_correct_thread_executing();
2805   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2806 
2807   save_heap_summary();
2808   report_heap_summary(GCWhen::BeforeGC);
2809 
2810   ReferenceProcessor* rp = ref_processor();
2811   assert(_restart_addr == NULL, "Control point invariant");
2812   {
2813     // acquire locks for subsequent manipulations
2814     MutexLockerEx x(bitMapLock(),
2815                     Mutex::_no_safepoint_check_flag);
2816     checkpointRootsInitialWork();
2817     // enable ("weak") refs discovery
2818     rp->enable_discovery();
2819     _collectorState = Marking;
2820   }
2821 }
2822 
2823 void CMSCollector::checkpointRootsInitialWork() {
2824   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2825   assert(_collectorState == InitialMarking, "just checking");
2826 
2827   // Already have locks.
2828   assert_lock_strong(bitMapLock());
2829   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2830 
2831   // Setup the verification and class unloading state for this
2832   // CMS collection cycle.
2833   setup_cms_unloading_and_verification_state();
2834 
2835   GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2836 
2837   // Reset all the PLAB chunk arrays if necessary.
2838   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2839     reset_survivor_plab_arrays();
2840   }
2841 
2842   ResourceMark rm;
2843   HandleMark  hm;
2844 
2845   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2846   GenCollectedHeap* gch = GenCollectedHeap::heap();
2847 
2848   verify_work_stacks_empty();
2849   verify_overflow_empty();
2850 
2851   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2852   // Update the saved marks which may affect the root scans.
2853   gch->save_marks();
2854 
2855   // weak reference processing has not started yet.
2856   ref_processor()->set_enqueuing_is_done(false);
2857 
2858   // Need to remember all newly created CLDs,
2859   // so that we can guarantee that the remark finds them.
2860   ClassLoaderDataGraph::remember_new_clds(true);
2861 
2862   // Whenever a CLD is found, it will be claimed before proceeding to mark
2863   // the klasses. The claimed marks need to be cleared before marking starts.
2864   ClassLoaderDataGraph::clear_claimed_marks();
2865 
2866   print_eden_and_survivor_chunk_arrays();
2867 
2868   {
2869 #if defined(COMPILER2) || INCLUDE_JVMCI
2870     DerivedPointerTableDeactivate dpt_deact;
2871 #endif
2872     if (CMSParallelInitialMarkEnabled) {
2873       // The parallel version.
2874       WorkGang* workers = gch->workers();
2875       assert(workers != NULL, "Need parallel worker threads.");
2876       uint n_workers = workers->active_workers();
2877 
2878       StrongRootsScope srs(n_workers);
2879 
2880       CMSParInitialMarkTask tsk(this, &srs, n_workers);
2881       initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2882       // If the total workers is greater than 1, then multiple workers
2883       // may be used at some time and the initialization has been set
2884       // such that the single threaded path cannot be used.
2885       if (workers->total_workers() > 1) {
2886         workers->run_task(&tsk);
2887       } else {
2888         tsk.work(0);
2889       }
2890     } else {
2891       // The serial version.
2892       CLDToOopClosure cld_closure(&notOlder, true);
2893       gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2894 
2895       StrongRootsScope srs(1);
2896 
2897       gch->cms_process_roots(&srs,
2898                              true,   // young gen as roots
2899                              GenCollectedHeap::ScanningOption(roots_scanning_options()),
2900                              should_unload_classes(),
2901                              &notOlder,
2902                              &cld_closure);
2903     }
2904   }
2905 
2906   // Clear mod-union table; it will be dirtied in the prologue of
2907   // CMS generation per each young generation collection.
2908 
2909   assert(_modUnionTable.isAllClear(),
2910        "Was cleared in most recent final checkpoint phase"
2911        " or no bits are set in the gc_prologue before the start of the next "
2912        "subsequent marking phase.");
2913 
2914   assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
2915 
2916   // Save the end of the used_region of the constituent generations
2917   // to be used to limit the extent of sweep in each generation.
2918   save_sweep_limits();
2919   verify_overflow_empty();
2920 }
2921 
2922 bool CMSCollector::markFromRoots() {
2923   // we might be tempted to assert that:
2924   // assert(!SafepointSynchronize::is_at_safepoint(),
2925   //        "inconsistent argument?");
2926   // However that wouldn't be right, because it's possible that
2927   // a safepoint is indeed in progress as a young generation
2928   // stop-the-world GC happens even as we mark in this generation.
2929   assert(_collectorState == Marking, "inconsistent state?");
2930   check_correct_thread_executing();
2931   verify_overflow_empty();
2932 
2933   // Weak ref discovery note: We may be discovering weak
2934   // refs in this generation concurrent (but interleaved) with
2935   // weak ref discovery by the young generation collector.
2936 
2937   CMSTokenSyncWithLocks ts(true, bitMapLock());
2938   GCTraceCPUTime tcpu;
2939   CMSPhaseAccounting pa(this, "Concurrent Mark");
2940   bool res = markFromRootsWork();
2941   if (res) {
2942     _collectorState = Precleaning;
2943   } else { // We failed and a foreground collection wants to take over
2944     assert(_foregroundGCIsActive, "internal state inconsistency");
2945     assert(_restart_addr == NULL,  "foreground will restart from scratch");
2946     log_debug(gc)("bailing out to foreground collection");
2947   }
2948   verify_overflow_empty();
2949   return res;
2950 }
2951 
2952 bool CMSCollector::markFromRootsWork() {
2953   // iterate over marked bits in bit map, doing a full scan and mark
2954   // from these roots using the following algorithm:
2955   // . if oop is to the right of the current scan pointer,
2956   //   mark corresponding bit (we'll process it later)
2957   // . else (oop is to left of current scan pointer)
2958   //   push oop on marking stack
2959   // . drain the marking stack
2960 
2961   // Note that when we do a marking step we need to hold the
2962   // bit map lock -- recall that direct allocation (by mutators)
2963   // and promotion (by the young generation collector) is also
2964   // marking the bit map. [the so-called allocate live policy.]
2965   // Because the implementation of bit map marking is not
2966   // robust wrt simultaneous marking of bits in the same word,
2967   // we need to make sure that there is no such interference
2968   // between concurrent such updates.
2969 
2970   // already have locks
2971   assert_lock_strong(bitMapLock());
2972 
2973   verify_work_stacks_empty();
2974   verify_overflow_empty();
2975   bool result = false;
2976   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2977     result = do_marking_mt();
2978   } else {
2979     result = do_marking_st();
2980   }
2981   return result;
2982 }
2983 
2984 // Forward decl
2985 class CMSConcMarkingTask;
2986 
2987 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2988   CMSCollector*       _collector;
2989   CMSConcMarkingTask* _task;
2990  public:
2991   virtual void yield();
2992 
2993   // "n_threads" is the number of threads to be terminated.
2994   // "queue_set" is a set of work queues of other threads.
2995   // "collector" is the CMS collector associated with this task terminator.
2996   // "yield" indicates whether we need the gang as a whole to yield.
2997   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
2998     ParallelTaskTerminator(n_threads, queue_set),
2999     _collector(collector) { }
3000 
3001   void set_task(CMSConcMarkingTask* task) {
3002     _task = task;
3003   }
3004 };
3005 
3006 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3007   CMSConcMarkingTask* _task;
3008  public:
3009   bool should_exit_termination();
3010   void set_task(CMSConcMarkingTask* task) {
3011     _task = task;
3012   }
3013 };
3014 
3015 // MT Concurrent Marking Task
3016 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3017   CMSCollector*             _collector;
3018   uint                      _n_workers;      // requested/desired # workers
3019   bool                      _result;
3020   CompactibleFreeListSpace* _cms_space;
3021   char                      _pad_front[64];   // padding to ...
3022   HeapWord* volatile        _global_finger;   // ... avoid sharing cache line
3023   char                      _pad_back[64];
3024   HeapWord*                 _restart_addr;
3025 
3026   //  Exposed here for yielding support
3027   Mutex* const _bit_map_lock;
3028 
3029   // The per thread work queues, available here for stealing
3030   OopTaskQueueSet*  _task_queues;
3031 
3032   // Termination (and yielding) support
3033   CMSConcMarkingTerminator _term;
3034   CMSConcMarkingTerminatorTerminator _term_term;
3035 
3036  public:
3037   CMSConcMarkingTask(CMSCollector* collector,
3038                  CompactibleFreeListSpace* cms_space,
3039                  YieldingFlexibleWorkGang* workers,
3040                  OopTaskQueueSet* task_queues):
3041     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3042     _collector(collector),
3043     _cms_space(cms_space),
3044     _n_workers(0), _result(true),
3045     _task_queues(task_queues),
3046     _term(_n_workers, task_queues, _collector),
3047     _bit_map_lock(collector->bitMapLock())
3048   {
3049     _requested_size = _n_workers;
3050     _term.set_task(this);
3051     _term_term.set_task(this);
3052     _restart_addr = _global_finger = _cms_space->bottom();
3053   }
3054 
3055 
3056   OopTaskQueueSet* task_queues()  { return _task_queues; }
3057 
3058   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3059 
3060   HeapWord* volatile* global_finger_addr() { return &_global_finger; }
3061 
3062   CMSConcMarkingTerminator* terminator() { return &_term; }
3063 
3064   virtual void set_for_termination(uint active_workers) {
3065     terminator()->reset_for_reuse(active_workers);
3066   }
3067 
3068   void work(uint worker_id);
3069   bool should_yield() {
3070     return    ConcurrentMarkSweepThread::should_yield()
3071            && !_collector->foregroundGCIsActive();
3072   }
3073 
3074   virtual void coordinator_yield();  // stuff done by coordinator
3075   bool result() { return _result; }
3076 
3077   void reset(HeapWord* ra) {
3078     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
3079     _restart_addr = _global_finger = ra;
3080     _term.reset_for_reuse();
3081   }
3082 
3083   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3084                                            OopTaskQueue* work_q);
3085 
3086  private:
3087   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3088   void do_work_steal(int i);
3089   void bump_global_finger(HeapWord* f);
3090 };
3091 
3092 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3093   assert(_task != NULL, "Error");
3094   return _task->yielding();
3095   // Note that we do not need the disjunct || _task->should_yield() above
3096   // because we want terminating threads to yield only if the task
3097   // is already in the midst of yielding, which happens only after at least one
3098   // thread has yielded.
3099 }
3100 
3101 void CMSConcMarkingTerminator::yield() {
3102   if (_task->should_yield()) {
3103     _task->yield();
3104   } else {
3105     ParallelTaskTerminator::yield();
3106   }
3107 }
3108 
3109 ////////////////////////////////////////////////////////////////
3110 // Concurrent Marking Algorithm Sketch
3111 ////////////////////////////////////////////////////////////////
3112 // Until all tasks exhausted (both spaces):
3113 // -- claim next available chunk
3114 // -- bump global finger via CAS
3115 // -- find first object that starts in this chunk
3116 //    and start scanning bitmap from that position
3117 // -- scan marked objects for oops
3118 // -- CAS-mark target, and if successful:
3119 //    . if target oop is above global finger (volatile read)
3120 //      nothing to do
3121 //    . if target oop is in chunk and above local finger
3122 //        then nothing to do
3123 //    . else push on work-queue
3124 // -- Deal with possible overflow issues:
3125 //    . local work-queue overflow causes stuff to be pushed on
3126 //      global (common) overflow queue
3127 //    . always first empty local work queue
3128 //    . then get a batch of oops from global work queue if any
3129 //    . then do work stealing
3130 // -- When all tasks claimed (both spaces)
3131 //    and local work queue empty,
3132 //    then in a loop do:
3133 //    . check global overflow stack; steal a batch of oops and trace
3134 //    . try to steal from other threads oif GOS is empty
3135 //    . if neither is available, offer termination
3136 // -- Terminate and return result
3137 //
3138 void CMSConcMarkingTask::work(uint worker_id) {
3139   elapsedTimer _timer;
3140   ResourceMark rm;
3141   HandleMark hm;
3142 
3143   DEBUG_ONLY(_collector->verify_overflow_empty();)
3144 
3145   // Before we begin work, our work queue should be empty
3146   assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3147   // Scan the bitmap covering _cms_space, tracing through grey objects.
3148   _timer.start();
3149   do_scan_and_mark(worker_id, _cms_space);
3150   _timer.stop();
3151   log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3152 
3153   // ... do work stealing
3154   _timer.reset();
3155   _timer.start();
3156   do_work_steal(worker_id);
3157   _timer.stop();
3158   log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3159   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3160   assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3161   // Note that under the current task protocol, the
3162   // following assertion is true even of the spaces
3163   // expanded since the completion of the concurrent
3164   // marking. XXX This will likely change under a strict
3165   // ABORT semantics.
3166   // After perm removal the comparison was changed to
3167   // greater than or equal to from strictly greater than.
3168   // Before perm removal the highest address sweep would
3169   // have been at the end of perm gen but now is at the
3170   // end of the tenured gen.
3171   assert(_global_finger >=  _cms_space->end(),
3172          "All tasks have been completed");
3173   DEBUG_ONLY(_collector->verify_overflow_empty();)
3174 }
3175 
3176 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3177   HeapWord* read = _global_finger;
3178   HeapWord* cur  = read;
3179   while (f > read) {
3180     cur = read;
3181     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3182     if (cur == read) {
3183       // our cas succeeded
3184       assert(_global_finger >= f, "protocol consistency");
3185       break;
3186     }
3187   }
3188 }
3189 
3190 // This is really inefficient, and should be redone by
3191 // using (not yet available) block-read and -write interfaces to the
3192 // stack and the work_queue. XXX FIX ME !!!
3193 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3194                                                       OopTaskQueue* work_q) {
3195   // Fast lock-free check
3196   if (ovflw_stk->length() == 0) {
3197     return false;
3198   }
3199   assert(work_q->size() == 0, "Shouldn't steal");
3200   MutexLockerEx ml(ovflw_stk->par_lock(),
3201                    Mutex::_no_safepoint_check_flag);
3202   // Grab up to 1/4 the size of the work queue
3203   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3204                     (size_t)ParGCDesiredObjsFromOverflowList);
3205   num = MIN2(num, ovflw_stk->length());
3206   for (int i = (int) num; i > 0; i--) {
3207     oop cur = ovflw_stk->pop();
3208     assert(cur != NULL, "Counted wrong?");
3209     work_q->push(cur);
3210   }
3211   return num > 0;
3212 }
3213 
3214 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3215   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3216   int n_tasks = pst->n_tasks();
3217   // We allow that there may be no tasks to do here because
3218   // we are restarting after a stack overflow.
3219   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3220   uint nth_task = 0;
3221 
3222   HeapWord* aligned_start = sp->bottom();
3223   if (sp->used_region().contains(_restart_addr)) {
3224     // Align down to a card boundary for the start of 0th task
3225     // for this space.
3226     aligned_start = align_down(_restart_addr, CardTableModRefBS::card_size);
3227   }
3228 
3229   size_t chunk_size = sp->marking_task_size();
3230   while (!pst->is_task_claimed(/* reference */ nth_task)) {
3231     // Having claimed the nth task in this space,
3232     // compute the chunk that it corresponds to:
3233     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3234                                aligned_start + (nth_task+1)*chunk_size);
3235     // Try and bump the global finger via a CAS;
3236     // note that we need to do the global finger bump
3237     // _before_ taking the intersection below, because
3238     // the task corresponding to that region will be
3239     // deemed done even if the used_region() expands
3240     // because of allocation -- as it almost certainly will
3241     // during start-up while the threads yield in the
3242     // closure below.
3243     HeapWord* finger = span.end();
3244     bump_global_finger(finger);   // atomically
3245     // There are null tasks here corresponding to chunks
3246     // beyond the "top" address of the space.
3247     span = span.intersection(sp->used_region());
3248     if (!span.is_empty()) {  // Non-null task
3249       HeapWord* prev_obj;
3250       assert(!span.contains(_restart_addr) || nth_task == 0,
3251              "Inconsistency");
3252       if (nth_task == 0) {
3253         // For the 0th task, we'll not need to compute a block_start.
3254         if (span.contains(_restart_addr)) {
3255           // In the case of a restart because of stack overflow,
3256           // we might additionally skip a chunk prefix.
3257           prev_obj = _restart_addr;
3258         } else {
3259           prev_obj = span.start();
3260         }
3261       } else {
3262         // We want to skip the first object because
3263         // the protocol is to scan any object in its entirety
3264         // that _starts_ in this span; a fortiori, any
3265         // object starting in an earlier span is scanned
3266         // as part of an earlier claimed task.
3267         // Below we use the "careful" version of block_start
3268         // so we do not try to navigate uninitialized objects.
3269         prev_obj = sp->block_start_careful(span.start());
3270         // Below we use a variant of block_size that uses the
3271         // Printezis bits to avoid waiting for allocated
3272         // objects to become initialized/parsable.
3273         while (prev_obj < span.start()) {
3274           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3275           if (sz > 0) {
3276             prev_obj += sz;
3277           } else {
3278             // In this case we may end up doing a bit of redundant
3279             // scanning, but that appears unavoidable, short of
3280             // locking the free list locks; see bug 6324141.
3281             break;
3282           }
3283         }
3284       }
3285       if (prev_obj < span.end()) {
3286         MemRegion my_span = MemRegion(prev_obj, span.end());
3287         // Do the marking work within a non-empty span --
3288         // the last argument to the constructor indicates whether the
3289         // iteration should be incremental with periodic yields.
3290         ParMarkFromRootsClosure cl(this, _collector, my_span,
3291                                    &_collector->_markBitMap,
3292                                    work_queue(i),
3293                                    &_collector->_markStack);
3294         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3295       } // else nothing to do for this task
3296     }   // else nothing to do for this task
3297   }
3298   // We'd be tempted to assert here that since there are no
3299   // more tasks left to claim in this space, the global_finger
3300   // must exceed space->top() and a fortiori space->end(). However,
3301   // that would not quite be correct because the bumping of
3302   // global_finger occurs strictly after the claiming of a task,
3303   // so by the time we reach here the global finger may not yet
3304   // have been bumped up by the thread that claimed the last
3305   // task.
3306   pst->all_tasks_completed();
3307 }
3308 
3309 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3310  private:
3311   CMSCollector* _collector;
3312   CMSConcMarkingTask* _task;
3313   MemRegion     _span;
3314   CMSBitMap*    _bit_map;
3315   CMSMarkStack* _overflow_stack;
3316   OopTaskQueue* _work_queue;
3317  protected:
3318   DO_OOP_WORK_DEFN
3319  public:
3320   ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3321                         CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3322     MetadataAwareOopClosure(collector->ref_processor()),
3323     _collector(collector),
3324     _task(task),
3325     _span(collector->_span),
3326     _work_queue(work_queue),
3327     _bit_map(bit_map),
3328     _overflow_stack(overflow_stack)
3329   { }
3330   virtual void do_oop(oop* p);
3331   virtual void do_oop(narrowOop* p);
3332 
3333   void trim_queue(size_t max);
3334   void handle_stack_overflow(HeapWord* lost);
3335   void do_yield_check() {
3336     if (_task->should_yield()) {
3337       _task->yield();
3338     }
3339   }
3340 };
3341 
3342 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3343 
3344 // Grey object scanning during work stealing phase --
3345 // the salient assumption here is that any references
3346 // that are in these stolen objects being scanned must
3347 // already have been initialized (else they would not have
3348 // been published), so we do not need to check for
3349 // uninitialized objects before pushing here.
3350 void ParConcMarkingClosure::do_oop(oop obj) {
3351   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3352   HeapWord* addr = (HeapWord*)obj;
3353   // Check if oop points into the CMS generation
3354   // and is not marked
3355   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3356     // a white object ...
3357     // If we manage to "claim" the object, by being the
3358     // first thread to mark it, then we push it on our
3359     // marking stack
3360     if (_bit_map->par_mark(addr)) {     // ... now grey
3361       // push on work queue (grey set)
3362       bool simulate_overflow = false;
3363       NOT_PRODUCT(
3364         if (CMSMarkStackOverflowALot &&
3365             _collector->simulate_overflow()) {
3366           // simulate a stack overflow
3367           simulate_overflow = true;
3368         }
3369       )
3370       if (simulate_overflow ||
3371           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3372         // stack overflow
3373         log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3374         // We cannot assert that the overflow stack is full because
3375         // it may have been emptied since.
3376         assert(simulate_overflow ||
3377                _work_queue->size() == _work_queue->max_elems(),
3378               "Else push should have succeeded");
3379         handle_stack_overflow(addr);
3380       }
3381     } // Else, some other thread got there first
3382     do_yield_check();
3383   }
3384 }
3385 
3386 void ParConcMarkingClosure::do_oop(oop* p)       { ParConcMarkingClosure::do_oop_work(p); }
3387 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3388 
3389 void ParConcMarkingClosure::trim_queue(size_t max) {
3390   while (_work_queue->size() > max) {
3391     oop new_oop;
3392     if (_work_queue->pop_local(new_oop)) {
3393       assert(oopDesc::is_oop(new_oop), "Should be an oop");
3394       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3395       assert(_span.contains((HeapWord*)new_oop), "Not in span");
3396       new_oop->oop_iterate(this);  // do_oop() above
3397       do_yield_check();
3398     }
3399   }
3400 }
3401 
3402 // Upon stack overflow, we discard (part of) the stack,
3403 // remembering the least address amongst those discarded
3404 // in CMSCollector's _restart_address.
3405 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3406   // We need to do this under a mutex to prevent other
3407   // workers from interfering with the work done below.
3408   MutexLockerEx ml(_overflow_stack->par_lock(),
3409                    Mutex::_no_safepoint_check_flag);
3410   // Remember the least grey address discarded
3411   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3412   _collector->lower_restart_addr(ra);
3413   _overflow_stack->reset();  // discard stack contents
3414   _overflow_stack->expand(); // expand the stack if possible
3415 }
3416 
3417 
3418 void CMSConcMarkingTask::do_work_steal(int i) {
3419   OopTaskQueue* work_q = work_queue(i);
3420   oop obj_to_scan;
3421   CMSBitMap* bm = &(_collector->_markBitMap);
3422   CMSMarkStack* ovflw = &(_collector->_markStack);
3423   int* seed = _collector->hash_seed(i);
3424   ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3425   while (true) {
3426     cl.trim_queue(0);
3427     assert(work_q->size() == 0, "Should have been emptied above");
3428     if (get_work_from_overflow_stack(ovflw, work_q)) {
3429       // Can't assert below because the work obtained from the
3430       // overflow stack may already have been stolen from us.
3431       // assert(work_q->size() > 0, "Work from overflow stack");
3432       continue;
3433     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3434       assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3435       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3436       obj_to_scan->oop_iterate(&cl);
3437     } else if (terminator()->offer_termination(&_term_term)) {
3438       assert(work_q->size() == 0, "Impossible!");
3439       break;
3440     } else if (yielding() || should_yield()) {
3441       yield();
3442     }
3443   }
3444 }
3445 
3446 // This is run by the CMS (coordinator) thread.
3447 void CMSConcMarkingTask::coordinator_yield() {
3448   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3449          "CMS thread should hold CMS token");
3450   // First give up the locks, then yield, then re-lock
3451   // We should probably use a constructor/destructor idiom to
3452   // do this unlock/lock or modify the MutexUnlocker class to
3453   // serve our purpose. XXX
3454   assert_lock_strong(_bit_map_lock);
3455   _bit_map_lock->unlock();
3456   ConcurrentMarkSweepThread::desynchronize(true);
3457   _collector->stopTimer();
3458   _collector->incrementYields();
3459 
3460   // It is possible for whichever thread initiated the yield request
3461   // not to get a chance to wake up and take the bitmap lock between
3462   // this thread releasing it and reacquiring it. So, while the
3463   // should_yield() flag is on, let's sleep for a bit to give the
3464   // other thread a chance to wake up. The limit imposed on the number
3465   // of iterations is defensive, to avoid any unforseen circumstances
3466   // putting us into an infinite loop. Since it's always been this
3467   // (coordinator_yield()) method that was observed to cause the
3468   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3469   // which is by default non-zero. For the other seven methods that
3470   // also perform the yield operation, as are using a different
3471   // parameter (CMSYieldSleepCount) which is by default zero. This way we
3472   // can enable the sleeping for those methods too, if necessary.
3473   // See 6442774.
3474   //
3475   // We really need to reconsider the synchronization between the GC
3476   // thread and the yield-requesting threads in the future and we
3477   // should really use wait/notify, which is the recommended
3478   // way of doing this type of interaction. Additionally, we should
3479   // consolidate the eight methods that do the yield operation and they
3480   // are almost identical into one for better maintainability and
3481   // readability. See 6445193.
3482   //
3483   // Tony 2006.06.29
3484   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3485                    ConcurrentMarkSweepThread::should_yield() &&
3486                    !CMSCollector::foregroundGCIsActive(); ++i) {
3487     os::sleep(Thread::current(), 1, false);
3488   }
3489 
3490   ConcurrentMarkSweepThread::synchronize(true);
3491   _bit_map_lock->lock_without_safepoint_check();
3492   _collector->startTimer();
3493 }
3494 
3495 bool CMSCollector::do_marking_mt() {
3496   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3497   uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3498                                                                   conc_workers()->active_workers(),
3499                                                                   Threads::number_of_non_daemon_threads());
3500   num_workers = conc_workers()->update_active_workers(num_workers);
3501   log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3502 
3503   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
3504 
3505   CMSConcMarkingTask tsk(this,
3506                          cms_space,
3507                          conc_workers(),
3508                          task_queues());
3509 
3510   // Since the actual number of workers we get may be different
3511   // from the number we requested above, do we need to do anything different
3512   // below? In particular, may be we need to subclass the SequantialSubTasksDone
3513   // class?? XXX
3514   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3515 
3516   // Refs discovery is already non-atomic.
3517   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3518   assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3519   conc_workers()->start_task(&tsk);
3520   while (tsk.yielded()) {
3521     tsk.coordinator_yield();
3522     conc_workers()->continue_task(&tsk);
3523   }
3524   // If the task was aborted, _restart_addr will be non-NULL
3525   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3526   while (_restart_addr != NULL) {
3527     // XXX For now we do not make use of ABORTED state and have not
3528     // yet implemented the right abort semantics (even in the original
3529     // single-threaded CMS case). That needs some more investigation
3530     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3531     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3532     // If _restart_addr is non-NULL, a marking stack overflow
3533     // occurred; we need to do a fresh marking iteration from the
3534     // indicated restart address.
3535     if (_foregroundGCIsActive) {
3536       // We may be running into repeated stack overflows, having
3537       // reached the limit of the stack size, while making very
3538       // slow forward progress. It may be best to bail out and
3539       // let the foreground collector do its job.
3540       // Clear _restart_addr, so that foreground GC
3541       // works from scratch. This avoids the headache of
3542       // a "rescan" which would otherwise be needed because
3543       // of the dirty mod union table & card table.
3544       _restart_addr = NULL;
3545       return false;
3546     }
3547     // Adjust the task to restart from _restart_addr
3548     tsk.reset(_restart_addr);
3549     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3550                   _restart_addr);
3551     _restart_addr = NULL;
3552     // Get the workers going again
3553     conc_workers()->start_task(&tsk);
3554     while (tsk.yielded()) {
3555       tsk.coordinator_yield();
3556       conc_workers()->continue_task(&tsk);
3557     }
3558   }
3559   assert(tsk.completed(), "Inconsistency");
3560   assert(tsk.result() == true, "Inconsistency");
3561   return true;
3562 }
3563 
3564 bool CMSCollector::do_marking_st() {
3565   ResourceMark rm;
3566   HandleMark   hm;
3567 
3568   // Temporarily make refs discovery single threaded (non-MT)
3569   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3570   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3571     &_markStack, CMSYield);
3572   // the last argument to iterate indicates whether the iteration
3573   // should be incremental with periodic yields.
3574   _markBitMap.iterate(&markFromRootsClosure);
3575   // If _restart_addr is non-NULL, a marking stack overflow
3576   // occurred; we need to do a fresh iteration from the
3577   // indicated restart address.
3578   while (_restart_addr != NULL) {
3579     if (_foregroundGCIsActive) {
3580       // We may be running into repeated stack overflows, having
3581       // reached the limit of the stack size, while making very
3582       // slow forward progress. It may be best to bail out and
3583       // let the foreground collector do its job.
3584       // Clear _restart_addr, so that foreground GC
3585       // works from scratch. This avoids the headache of
3586       // a "rescan" which would otherwise be needed because
3587       // of the dirty mod union table & card table.
3588       _restart_addr = NULL;
3589       return false;  // indicating failure to complete marking
3590     }
3591     // Deal with stack overflow:
3592     // we restart marking from _restart_addr
3593     HeapWord* ra = _restart_addr;
3594     markFromRootsClosure.reset(ra);
3595     _restart_addr = NULL;
3596     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3597   }
3598   return true;
3599 }
3600 
3601 void CMSCollector::preclean() {
3602   check_correct_thread_executing();
3603   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3604   verify_work_stacks_empty();
3605   verify_overflow_empty();
3606   _abort_preclean = false;
3607   if (CMSPrecleaningEnabled) {
3608     if (!CMSEdenChunksRecordAlways) {
3609       _eden_chunk_index = 0;
3610     }
3611     size_t used = get_eden_used();
3612     size_t capacity = get_eden_capacity();
3613     // Don't start sampling unless we will get sufficiently
3614     // many samples.
3615     if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3616                 * CMSScheduleRemarkEdenPenetration)) {
3617       _start_sampling = true;
3618     } else {
3619       _start_sampling = false;
3620     }
3621     GCTraceCPUTime tcpu;
3622     CMSPhaseAccounting pa(this, "Concurrent Preclean");
3623     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3624   }
3625   CMSTokenSync x(true); // is cms thread
3626   if (CMSPrecleaningEnabled) {
3627     sample_eden();
3628     _collectorState = AbortablePreclean;
3629   } else {
3630     _collectorState = FinalMarking;
3631   }
3632   verify_work_stacks_empty();
3633   verify_overflow_empty();
3634 }
3635 
3636 // Try and schedule the remark such that young gen
3637 // occupancy is CMSScheduleRemarkEdenPenetration %.
3638 void CMSCollector::abortable_preclean() {
3639   check_correct_thread_executing();
3640   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
3641   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3642 
3643   // If Eden's current occupancy is below this threshold,
3644   // immediately schedule the remark; else preclean
3645   // past the next scavenge in an effort to
3646   // schedule the pause as described above. By choosing
3647   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3648   // we will never do an actual abortable preclean cycle.
3649   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3650     GCTraceCPUTime tcpu;
3651     CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3652     // We need more smarts in the abortable preclean
3653     // loop below to deal with cases where allocation
3654     // in young gen is very very slow, and our precleaning
3655     // is running a losing race against a horde of
3656     // mutators intent on flooding us with CMS updates
3657     // (dirty cards).
3658     // One, admittedly dumb, strategy is to give up
3659     // after a certain number of abortable precleaning loops
3660     // or after a certain maximum time. We want to make
3661     // this smarter in the next iteration.
3662     // XXX FIX ME!!! YSR
3663     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3664     while (!(should_abort_preclean() ||
3665              ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3666       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3667       cumworkdone += workdone;
3668       loops++;
3669       // Voluntarily terminate abortable preclean phase if we have
3670       // been at it for too long.
3671       if ((CMSMaxAbortablePrecleanLoops != 0) &&
3672           loops >= CMSMaxAbortablePrecleanLoops) {
3673         log_debug(gc)(" CMS: abort preclean due to loops ");
3674         break;
3675       }
3676       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3677         log_debug(gc)(" CMS: abort preclean due to time ");
3678         break;
3679       }
3680       // If we are doing little work each iteration, we should
3681       // take a short break.
3682       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3683         // Sleep for some time, waiting for work to accumulate
3684         stopTimer();
3685         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3686         startTimer();
3687         waited++;
3688       }
3689     }
3690     log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3691                                loops, waited, cumworkdone);
3692   }
3693   CMSTokenSync x(true); // is cms thread
3694   if (_collectorState != Idling) {
3695     assert(_collectorState == AbortablePreclean,
3696            "Spontaneous state transition?");
3697     _collectorState = FinalMarking;
3698   } // Else, a foreground collection completed this CMS cycle.
3699   return;
3700 }
3701 
3702 // Respond to an Eden sampling opportunity
3703 void CMSCollector::sample_eden() {
3704   // Make sure a young gc cannot sneak in between our
3705   // reading and recording of a sample.
3706   assert(Thread::current()->is_ConcurrentGC_thread(),
3707          "Only the cms thread may collect Eden samples");
3708   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3709          "Should collect samples while holding CMS token");
3710   if (!_start_sampling) {
3711     return;
3712   }
3713   // When CMSEdenChunksRecordAlways is true, the eden chunk array
3714   // is populated by the young generation.
3715   if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3716     if (_eden_chunk_index < _eden_chunk_capacity) {
3717       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
3718       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3719              "Unexpected state of Eden");
3720       // We'd like to check that what we just sampled is an oop-start address;
3721       // however, we cannot do that here since the object may not yet have been
3722       // initialized. So we'll instead do the check when we _use_ this sample
3723       // later.
3724       if (_eden_chunk_index == 0 ||
3725           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3726                          _eden_chunk_array[_eden_chunk_index-1])
3727            >= CMSSamplingGrain)) {
3728         _eden_chunk_index++;  // commit sample
3729       }
3730     }
3731   }
3732   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3733     size_t used = get_eden_used();
3734     size_t capacity = get_eden_capacity();
3735     assert(used <= capacity, "Unexpected state of Eden");
3736     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3737       _abort_preclean = true;
3738     }
3739   }
3740 }
3741 
3742 
3743 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3744   assert(_collectorState == Precleaning ||
3745          _collectorState == AbortablePreclean, "incorrect state");
3746   ResourceMark rm;
3747   HandleMark   hm;
3748 
3749   // Precleaning is currently not MT but the reference processor
3750   // may be set for MT.  Disable it temporarily here.
3751   ReferenceProcessor* rp = ref_processor();
3752   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3753 
3754   // Do one pass of scrubbing the discovered reference lists
3755   // to remove any reference objects with strongly-reachable
3756   // referents.
3757   if (clean_refs) {
3758     CMSPrecleanRefsYieldClosure yield_cl(this);
3759     assert(rp->span().equals(_span), "Spans should be equal");
3760     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3761                                    &_markStack, true /* preclean */);
3762     CMSDrainMarkingStackClosure complete_trace(this,
3763                                    _span, &_markBitMap, &_markStack,
3764                                    &keep_alive, true /* preclean */);
3765 
3766     // We don't want this step to interfere with a young
3767     // collection because we don't want to take CPU
3768     // or memory bandwidth away from the young GC threads
3769     // (which may be as many as there are CPUs).
3770     // Note that we don't need to protect ourselves from
3771     // interference with mutators because they can't
3772     // manipulate the discovered reference lists nor affect
3773     // the computed reachability of the referents, the
3774     // only properties manipulated by the precleaning
3775     // of these reference lists.
3776     stopTimer();
3777     CMSTokenSyncWithLocks x(true /* is cms thread */,
3778                             bitMapLock());
3779     startTimer();
3780     sample_eden();
3781 
3782     // The following will yield to allow foreground
3783     // collection to proceed promptly. XXX YSR:
3784     // The code in this method may need further
3785     // tweaking for better performance and some restructuring
3786     // for cleaner interfaces.
3787     GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3788     rp->preclean_discovered_references(
3789           rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3790           gc_timer);
3791   }
3792 
3793   if (clean_survivor) {  // preclean the active survivor space(s)
3794     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3795                              &_markBitMap, &_modUnionTable,
3796                              &_markStack, true /* precleaning phase */);
3797     stopTimer();
3798     CMSTokenSyncWithLocks ts(true /* is cms thread */,
3799                              bitMapLock());
3800     startTimer();
3801     unsigned int before_count =
3802       GenCollectedHeap::heap()->total_collections();
3803     SurvivorSpacePrecleanClosure
3804       sss_cl(this, _span, &_markBitMap, &_markStack,
3805              &pam_cl, before_count, CMSYield);
3806     _young_gen->from()->object_iterate_careful(&sss_cl);
3807     _young_gen->to()->object_iterate_careful(&sss_cl);
3808   }
3809   MarkRefsIntoAndScanClosure
3810     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3811              &_markStack, this, CMSYield,
3812              true /* precleaning phase */);
3813   // CAUTION: The following closure has persistent state that may need to
3814   // be reset upon a decrease in the sequence of addresses it
3815   // processes.
3816   ScanMarkedObjectsAgainCarefullyClosure
3817     smoac_cl(this, _span,
3818       &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3819 
3820   // Preclean dirty cards in ModUnionTable and CardTable using
3821   // appropriate convergence criterion;
3822   // repeat CMSPrecleanIter times unless we find that
3823   // we are losing.
3824   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3825   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3826          "Bad convergence multiplier");
3827   assert(CMSPrecleanThreshold >= 100,
3828          "Unreasonably low CMSPrecleanThreshold");
3829 
3830   size_t numIter, cumNumCards, lastNumCards, curNumCards;
3831   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3832        numIter < CMSPrecleanIter;
3833        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3834     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
3835     log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3836     // Either there are very few dirty cards, so re-mark
3837     // pause will be small anyway, or our pre-cleaning isn't
3838     // that much faster than the rate at which cards are being
3839     // dirtied, so we might as well stop and re-mark since
3840     // precleaning won't improve our re-mark time by much.
3841     if (curNumCards <= CMSPrecleanThreshold ||
3842         (numIter > 0 &&
3843          (curNumCards * CMSPrecleanDenominator >
3844          lastNumCards * CMSPrecleanNumerator))) {
3845       numIter++;
3846       cumNumCards += curNumCards;
3847       break;
3848     }
3849   }
3850 
3851   preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
3852 
3853   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3854   cumNumCards += curNumCards;
3855   log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3856                              curNumCards, cumNumCards, numIter);
3857   return cumNumCards;   // as a measure of useful work done
3858 }
3859 
3860 // PRECLEANING NOTES:
3861 // Precleaning involves:
3862 // . reading the bits of the modUnionTable and clearing the set bits.
3863 // . For the cards corresponding to the set bits, we scan the
3864 //   objects on those cards. This means we need the free_list_lock
3865 //   so that we can safely iterate over the CMS space when scanning
3866 //   for oops.
3867 // . When we scan the objects, we'll be both reading and setting
3868 //   marks in the marking bit map, so we'll need the marking bit map.
3869 // . For protecting _collector_state transitions, we take the CGC_lock.
3870 //   Note that any races in the reading of of card table entries by the
3871 //   CMS thread on the one hand and the clearing of those entries by the
3872 //   VM thread or the setting of those entries by the mutator threads on the
3873 //   other are quite benign. However, for efficiency it makes sense to keep
3874 //   the VM thread from racing with the CMS thread while the latter is
3875 //   dirty card info to the modUnionTable. We therefore also use the
3876 //   CGC_lock to protect the reading of the card table and the mod union
3877 //   table by the CM thread.
3878 // . We run concurrently with mutator updates, so scanning
3879 //   needs to be done carefully  -- we should not try to scan
3880 //   potentially uninitialized objects.
3881 //
3882 // Locking strategy: While holding the CGC_lock, we scan over and
3883 // reset a maximal dirty range of the mod union / card tables, then lock
3884 // the free_list_lock and bitmap lock to do a full marking, then
3885 // release these locks; and repeat the cycle. This allows for a
3886 // certain amount of fairness in the sharing of these locks between
3887 // the CMS collector on the one hand, and the VM thread and the
3888 // mutators on the other.
3889 
3890 // NOTE: preclean_mod_union_table() and preclean_card_table()
3891 // further below are largely identical; if you need to modify
3892 // one of these methods, please check the other method too.
3893 
3894 size_t CMSCollector::preclean_mod_union_table(
3895   ConcurrentMarkSweepGeneration* old_gen,
3896   ScanMarkedObjectsAgainCarefullyClosure* cl) {
3897   verify_work_stacks_empty();
3898   verify_overflow_empty();
3899 
3900   // strategy: starting with the first card, accumulate contiguous
3901   // ranges of dirty cards; clear these cards, then scan the region
3902   // covered by these cards.
3903 
3904   // Since all of the MUT is committed ahead, we can just use
3905   // that, in case the generations expand while we are precleaning.
3906   // It might also be fine to just use the committed part of the
3907   // generation, but we might potentially miss cards when the
3908   // generation is rapidly expanding while we are in the midst
3909   // of precleaning.
3910   HeapWord* startAddr = old_gen->reserved().start();
3911   HeapWord* endAddr   = old_gen->reserved().end();
3912 
3913   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
3914 
3915   size_t numDirtyCards, cumNumDirtyCards;
3916   HeapWord *nextAddr, *lastAddr;
3917   for (cumNumDirtyCards = numDirtyCards = 0,
3918        nextAddr = lastAddr = startAddr;
3919        nextAddr < endAddr;
3920        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3921 
3922     ResourceMark rm;
3923     HandleMark   hm;
3924 
3925     MemRegion dirtyRegion;
3926     {
3927       stopTimer();
3928       // Potential yield point
3929       CMSTokenSync ts(true);
3930       startTimer();
3931       sample_eden();
3932       // Get dirty region starting at nextOffset (inclusive),
3933       // simultaneously clearing it.
3934       dirtyRegion =
3935         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3936       assert(dirtyRegion.start() >= nextAddr,
3937              "returned region inconsistent?");
3938     }
3939     // Remember where the next search should begin.
3940     // The returned region (if non-empty) is a right open interval,
3941     // so lastOffset is obtained from the right end of that
3942     // interval.
3943     lastAddr = dirtyRegion.end();
3944     // Should do something more transparent and less hacky XXX
3945     numDirtyCards =
3946       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3947 
3948     // We'll scan the cards in the dirty region (with periodic
3949     // yields for foreground GC as needed).
3950     if (!dirtyRegion.is_empty()) {
3951       assert(numDirtyCards > 0, "consistency check");
3952       HeapWord* stop_point = NULL;
3953       stopTimer();
3954       // Potential yield point
3955       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3956                                bitMapLock());
3957       startTimer();
3958       {
3959         verify_work_stacks_empty();
3960         verify_overflow_empty();
3961         sample_eden();
3962         stop_point =
3963           old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3964       }
3965       if (stop_point != NULL) {
3966         // The careful iteration stopped early either because it found an
3967         // uninitialized object, or because we were in the midst of an
3968         // "abortable preclean", which should now be aborted. Redirty
3969         // the bits corresponding to the partially-scanned or unscanned
3970         // cards. We'll either restart at the next block boundary or
3971         // abort the preclean.
3972         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3973                "Should only be AbortablePreclean.");
3974         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3975         if (should_abort_preclean()) {
3976           break; // out of preclean loop
3977         } else {
3978           // Compute the next address at which preclean should pick up;
3979           // might need bitMapLock in order to read P-bits.
3980           lastAddr = next_card_start_after_block(stop_point);
3981         }
3982       }
3983     } else {
3984       assert(lastAddr == endAddr, "consistency check");
3985       assert(numDirtyCards == 0, "consistency check");
3986       break;
3987     }
3988   }
3989   verify_work_stacks_empty();
3990   verify_overflow_empty();
3991   return cumNumDirtyCards;
3992 }
3993 
3994 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3995 // below are largely identical; if you need to modify
3996 // one of these methods, please check the other method too.
3997 
3998 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
3999   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4000   // strategy: it's similar to precleamModUnionTable above, in that
4001   // we accumulate contiguous ranges of dirty cards, mark these cards
4002   // precleaned, then scan the region covered by these cards.
4003   HeapWord* endAddr   = (HeapWord*)(old_gen->_virtual_space.high());
4004   HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4005 
4006   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
4007 
4008   size_t numDirtyCards, cumNumDirtyCards;
4009   HeapWord *lastAddr, *nextAddr;
4010 
4011   for (cumNumDirtyCards = numDirtyCards = 0,
4012        nextAddr = lastAddr = startAddr;
4013        nextAddr < endAddr;
4014        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4015 
4016     ResourceMark rm;
4017     HandleMark   hm;
4018 
4019     MemRegion dirtyRegion;
4020     {
4021       // See comments in "Precleaning notes" above on why we
4022       // do this locking. XXX Could the locking overheads be
4023       // too high when dirty cards are sparse? [I don't think so.]
4024       stopTimer();
4025       CMSTokenSync x(true); // is cms thread
4026       startTimer();
4027       sample_eden();
4028       // Get and clear dirty region from card table
4029       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4030                                     MemRegion(nextAddr, endAddr),
4031                                     true,
4032                                     CardTableModRefBS::precleaned_card_val());
4033 
4034       assert(dirtyRegion.start() >= nextAddr,
4035              "returned region inconsistent?");
4036     }
4037     lastAddr = dirtyRegion.end();
4038     numDirtyCards =
4039       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4040 
4041     if (!dirtyRegion.is_empty()) {
4042       stopTimer();
4043       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4044       startTimer();
4045       sample_eden();
4046       verify_work_stacks_empty();
4047       verify_overflow_empty();
4048       HeapWord* stop_point =
4049         old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4050       if (stop_point != NULL) {
4051         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4052                "Should only be AbortablePreclean.");
4053         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4054         if (should_abort_preclean()) {
4055           break; // out of preclean loop
4056         } else {
4057           // Compute the next address at which preclean should pick up.
4058           lastAddr = next_card_start_after_block(stop_point);
4059         }
4060       }
4061     } else {
4062       break;
4063     }
4064   }
4065   verify_work_stacks_empty();
4066   verify_overflow_empty();
4067   return cumNumDirtyCards;
4068 }
4069 
4070 class PrecleanKlassClosure : public KlassClosure {
4071   KlassToOopClosure _cm_klass_closure;
4072  public:
4073   PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4074   void do_klass(Klass* k) {
4075     if (k->has_accumulated_modified_oops()) {
4076       k->clear_accumulated_modified_oops();
4077 
4078       _cm_klass_closure.do_klass(k);
4079     }
4080   }
4081 };
4082 
4083 // The freelist lock is needed to prevent asserts, is it really needed?
4084 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4085 
4086   cl->set_freelistLock(freelistLock);
4087 
4088   CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4089 
4090   // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4091   // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4092   PrecleanKlassClosure preclean_klass_closure(cl);
4093   ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4094 
4095   verify_work_stacks_empty();
4096   verify_overflow_empty();
4097 }
4098 
4099 void CMSCollector::checkpointRootsFinal() {
4100   assert(_collectorState == FinalMarking, "incorrect state transition?");
4101   check_correct_thread_executing();
4102   // world is stopped at this checkpoint
4103   assert(SafepointSynchronize::is_at_safepoint(),
4104          "world should be stopped");
4105   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4106 
4107   verify_work_stacks_empty();
4108   verify_overflow_empty();
4109 
4110   log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4111                 _young_gen->used() / K, _young_gen->capacity() / K);
4112   {
4113     if (CMSScavengeBeforeRemark) {
4114       GenCollectedHeap* gch = GenCollectedHeap::heap();
4115       // Temporarily set flag to false, GCH->do_collection will
4116       // expect it to be false and set to true
4117       FlagSetting fl(gch->_is_gc_active, false);
4118 
4119       gch->do_collection(true,                      // full (i.e. force, see below)
4120                          false,                     // !clear_all_soft_refs
4121                          0,                         // size
4122                          false,                     // is_tlab
4123                          GenCollectedHeap::YoungGen // type
4124         );
4125     }
4126     FreelistLocker x(this);
4127     MutexLockerEx y(bitMapLock(),
4128                     Mutex::_no_safepoint_check_flag);
4129     checkpointRootsFinalWork();
4130   }
4131   verify_work_stacks_empty();
4132   verify_overflow_empty();
4133 }
4134 
4135 void CMSCollector::checkpointRootsFinalWork() {
4136   GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4137 
4138   assert(haveFreelistLocks(), "must have free list locks");
4139   assert_lock_strong(bitMapLock());
4140 
4141   ResourceMark rm;
4142   HandleMark   hm;
4143 
4144   GenCollectedHeap* gch = GenCollectedHeap::heap();
4145 
4146   if (should_unload_classes()) {
4147     CodeCache::gc_prologue();
4148   }
4149   assert(haveFreelistLocks(), "must have free list locks");
4150   assert_lock_strong(bitMapLock());
4151 
4152   // We might assume that we need not fill TLAB's when
4153   // CMSScavengeBeforeRemark is set, because we may have just done
4154   // a scavenge which would have filled all TLAB's -- and besides
4155   // Eden would be empty. This however may not always be the case --
4156   // for instance although we asked for a scavenge, it may not have
4157   // happened because of a JNI critical section. We probably need
4158   // a policy for deciding whether we can in that case wait until
4159   // the critical section releases and then do the remark following
4160   // the scavenge, and skip it here. In the absence of that policy,
4161   // or of an indication of whether the scavenge did indeed occur,
4162   // we cannot rely on TLAB's having been filled and must do
4163   // so here just in case a scavenge did not happen.
4164   gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
4165   // Update the saved marks which may affect the root scans.
4166   gch->save_marks();
4167 
4168   print_eden_and_survivor_chunk_arrays();
4169 
4170   {
4171 #if defined(COMPILER2) || INCLUDE_JVMCI
4172     DerivedPointerTableDeactivate dpt_deact;
4173 #endif
4174 
4175     // Note on the role of the mod union table:
4176     // Since the marker in "markFromRoots" marks concurrently with
4177     // mutators, it is possible for some reachable objects not to have been
4178     // scanned. For instance, an only reference to an object A was
4179     // placed in object B after the marker scanned B. Unless B is rescanned,
4180     // A would be collected. Such updates to references in marked objects
4181     // are detected via the mod union table which is the set of all cards
4182     // dirtied since the first checkpoint in this GC cycle and prior to
4183     // the most recent young generation GC, minus those cleaned up by the
4184     // concurrent precleaning.
4185     if (CMSParallelRemarkEnabled) {
4186       GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4187       do_remark_parallel();
4188     } else {
4189       GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4190       do_remark_non_parallel();
4191     }
4192   }
4193   verify_work_stacks_empty();
4194   verify_overflow_empty();
4195 
4196   {
4197     GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4198     refProcessingWork();
4199   }
4200   verify_work_stacks_empty();
4201   verify_overflow_empty();
4202 
4203   if (should_unload_classes()) {
4204     CodeCache::gc_epilogue();
4205   }
4206   JvmtiExport::gc_epilogue();
4207 
4208   // If we encountered any (marking stack / work queue) overflow
4209   // events during the current CMS cycle, take appropriate
4210   // remedial measures, where possible, so as to try and avoid
4211   // recurrence of that condition.
4212   assert(_markStack.isEmpty(), "No grey objects");
4213   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4214                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
4215   if (ser_ovflw > 0) {
4216     log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4217                          _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4218     _markStack.expand();
4219     _ser_pmc_remark_ovflw = 0;
4220     _ser_pmc_preclean_ovflw = 0;
4221     _ser_kac_preclean_ovflw = 0;
4222     _ser_kac_ovflw = 0;
4223   }
4224   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4225      log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4226                           _par_pmc_remark_ovflw, _par_kac_ovflw);
4227      _par_pmc_remark_ovflw = 0;
4228     _par_kac_ovflw = 0;
4229   }
4230    if (_markStack._hit_limit > 0) {
4231      log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4232                           _markStack._hit_limit);
4233    }
4234    if (_markStack._failed_double > 0) {
4235      log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4236                           _markStack._failed_double, _markStack.capacity());
4237    }
4238   _markStack._hit_limit = 0;
4239   _markStack._failed_double = 0;
4240 
4241   if ((VerifyAfterGC || VerifyDuringGC) &&
4242       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4243     verify_after_remark();
4244   }
4245 
4246   _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4247 
4248   // Change under the freelistLocks.
4249   _collectorState = Sweeping;
4250   // Call isAllClear() under bitMapLock
4251   assert(_modUnionTable.isAllClear(),
4252       "Should be clear by end of the final marking");
4253   assert(_ct->klass_rem_set()->mod_union_is_clear(),
4254       "Should be clear by end of the final marking");
4255 }
4256 
4257 void CMSParInitialMarkTask::work(uint worker_id) {
4258   elapsedTimer _timer;
4259   ResourceMark rm;
4260   HandleMark   hm;
4261 
4262   // ---------- scan from roots --------------
4263   _timer.start();
4264   GenCollectedHeap* gch = GenCollectedHeap::heap();
4265   ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4266 
4267   // ---------- young gen roots --------------
4268   {
4269     work_on_young_gen_roots(&par_mri_cl);
4270     _timer.stop();
4271     log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4272   }
4273 
4274   // ---------- remaining roots --------------
4275   _timer.reset();
4276   _timer.start();
4277 
4278   CLDToOopClosure cld_closure(&par_mri_cl, true);
4279 
4280   gch->cms_process_roots(_strong_roots_scope,
4281                          false,     // yg was scanned above
4282                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4283                          _collector->should_unload_classes(),
4284                          &par_mri_cl,
4285                          &cld_closure);
4286   assert(_collector->should_unload_classes()
4287          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4288          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4289   _timer.stop();
4290   log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4291 }
4292 
4293 // Parallel remark task
4294 class CMSParRemarkTask: public CMSParMarkTask {
4295   CompactibleFreeListSpace* _cms_space;
4296 
4297   // The per-thread work queues, available here for stealing.
4298   OopTaskQueueSet*       _task_queues;
4299   ParallelTaskTerminator _term;
4300   StrongRootsScope*      _strong_roots_scope;
4301 
4302  public:
4303   // A value of 0 passed to n_workers will cause the number of
4304   // workers to be taken from the active workers in the work gang.
4305   CMSParRemarkTask(CMSCollector* collector,
4306                    CompactibleFreeListSpace* cms_space,
4307                    uint n_workers, WorkGang* workers,
4308                    OopTaskQueueSet* task_queues,
4309                    StrongRootsScope* strong_roots_scope):
4310     CMSParMarkTask("Rescan roots and grey objects in parallel",
4311                    collector, n_workers),
4312     _cms_space(cms_space),
4313     _task_queues(task_queues),
4314     _term(n_workers, task_queues),
4315     _strong_roots_scope(strong_roots_scope) { }
4316 
4317   OopTaskQueueSet* task_queues() { return _task_queues; }
4318 
4319   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4320 
4321   ParallelTaskTerminator* terminator() { return &_term; }
4322   uint n_workers() { return _n_workers; }
4323 
4324   void work(uint worker_id);
4325 
4326  private:
4327   // ... of  dirty cards in old space
4328   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4329                                   ParMarkRefsIntoAndScanClosure* cl);
4330 
4331   // ... work stealing for the above
4332   void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4333 };
4334 
4335 class RemarkKlassClosure : public KlassClosure {
4336   KlassToOopClosure _cm_klass_closure;
4337  public:
4338   RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4339   void do_klass(Klass* k) {
4340     // Check if we have modified any oops in the Klass during the concurrent marking.
4341     if (k->has_accumulated_modified_oops()) {
4342       k->clear_accumulated_modified_oops();
4343 
4344       // We could have transfered the current modified marks to the accumulated marks,
4345       // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4346     } else if (k->has_modified_oops()) {
4347       // Don't clear anything, this info is needed by the next young collection.
4348     } else {
4349       // No modified oops in the Klass.
4350       return;
4351     }
4352 
4353     // The klass has modified fields, need to scan the klass.
4354     _cm_klass_closure.do_klass(k);
4355   }
4356 };
4357 
4358 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4359   ParNewGeneration* young_gen = _collector->_young_gen;
4360   ContiguousSpace* eden_space = young_gen->eden();
4361   ContiguousSpace* from_space = young_gen->from();
4362   ContiguousSpace* to_space   = young_gen->to();
4363 
4364   HeapWord** eca = _collector->_eden_chunk_array;
4365   size_t     ect = _collector->_eden_chunk_index;
4366   HeapWord** sca = _collector->_survivor_chunk_array;
4367   size_t     sct = _collector->_survivor_chunk_index;
4368 
4369   assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4370   assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4371 
4372   do_young_space_rescan(cl, to_space, NULL, 0);
4373   do_young_space_rescan(cl, from_space, sca, sct);
4374   do_young_space_rescan(cl, eden_space, eca, ect);
4375 }
4376 
4377 // work_queue(i) is passed to the closure
4378 // ParMarkRefsIntoAndScanClosure.  The "i" parameter
4379 // also is passed to do_dirty_card_rescan_tasks() and to
4380 // do_work_steal() to select the i-th task_queue.
4381 
4382 void CMSParRemarkTask::work(uint worker_id) {
4383   elapsedTimer _timer;
4384   ResourceMark rm;
4385   HandleMark   hm;
4386 
4387   // ---------- rescan from roots --------------
4388   _timer.start();
4389   GenCollectedHeap* gch = GenCollectedHeap::heap();
4390   ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4391     _collector->_span, _collector->ref_processor(),
4392     &(_collector->_markBitMap),
4393     work_queue(worker_id));
4394 
4395   // Rescan young gen roots first since these are likely
4396   // coarsely partitioned and may, on that account, constitute
4397   // the critical path; thus, it's best to start off that
4398   // work first.
4399   // ---------- young gen roots --------------
4400   {
4401     work_on_young_gen_roots(&par_mrias_cl);
4402     _timer.stop();
4403     log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4404   }
4405 
4406   // ---------- remaining roots --------------
4407   _timer.reset();
4408   _timer.start();
4409   gch->cms_process_roots(_strong_roots_scope,
4410                          false,     // yg was scanned above
4411                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4412                          _collector->should_unload_classes(),
4413                          &par_mrias_cl,
4414                          NULL);     // The dirty klasses will be handled below
4415 
4416   assert(_collector->should_unload_classes()
4417          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4418          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4419   _timer.stop();
4420   log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec",  worker_id, _timer.seconds());
4421 
4422   // ---------- unhandled CLD scanning ----------
4423   if (worker_id == 0) { // Single threaded at the moment.
4424     _timer.reset();
4425     _timer.start();
4426 
4427     // Scan all new class loader data objects and new dependencies that were
4428     // introduced during concurrent marking.
4429     ResourceMark rm;
4430     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4431     for (int i = 0; i < array->length(); i++) {
4432       par_mrias_cl.do_cld_nv(array->at(i));
4433     }
4434 
4435     // We don't need to keep track of new CLDs anymore.
4436     ClassLoaderDataGraph::remember_new_clds(false);
4437 
4438     _timer.stop();
4439     log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4440   }
4441 
4442   // ---------- dirty klass scanning ----------
4443   if (worker_id == 0) { // Single threaded at the moment.
4444     _timer.reset();
4445     _timer.start();
4446 
4447     // Scan all classes that was dirtied during the concurrent marking phase.
4448     RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4449     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4450 
4451     _timer.stop();
4452     log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4453   }
4454 
4455   // We might have added oops to ClassLoaderData::_handles during the
4456   // concurrent marking phase. These oops point to newly allocated objects
4457   // that are guaranteed to be kept alive. Either by the direct allocation
4458   // code, or when the young collector processes the roots. Hence,
4459   // we don't have to revisit the _handles block during the remark phase.
4460 
4461   // ---------- rescan dirty cards ------------
4462   _timer.reset();
4463   _timer.start();
4464 
4465   // Do the rescan tasks for each of the two spaces
4466   // (cms_space) in turn.
4467   // "worker_id" is passed to select the task_queue for "worker_id"
4468   do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4469   _timer.stop();
4470   log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4471 
4472   // ---------- steal work from other threads ...
4473   // ---------- ... and drain overflow list.
4474   _timer.reset();
4475   _timer.start();
4476   do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4477   _timer.stop();
4478   log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4479 }
4480 
4481 void
4482 CMSParMarkTask::do_young_space_rescan(
4483   OopsInGenClosure* cl, ContiguousSpace* space,
4484   HeapWord** chunk_array, size_t chunk_top) {
4485   // Until all tasks completed:
4486   // . claim an unclaimed task
4487   // . compute region boundaries corresponding to task claimed
4488   //   using chunk_array
4489   // . par_oop_iterate(cl) over that region
4490 
4491   ResourceMark rm;
4492   HandleMark   hm;
4493 
4494   SequentialSubTasksDone* pst = space->par_seq_tasks();
4495 
4496   uint nth_task = 0;
4497   uint n_tasks  = pst->n_tasks();
4498 
4499   if (n_tasks > 0) {
4500     assert(pst->valid(), "Uninitialized use?");
4501     HeapWord *start, *end;
4502     while (!pst->is_task_claimed(/* reference */ nth_task)) {
4503       // We claimed task # nth_task; compute its boundaries.
4504       if (chunk_top == 0) {  // no samples were taken
4505         assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4506         start = space->bottom();
4507         end   = space->top();
4508       } else if (nth_task == 0) {
4509         start = space->bottom();
4510         end   = chunk_array[nth_task];
4511       } else if (nth_task < (uint)chunk_top) {
4512         assert(nth_task >= 1, "Control point invariant");
4513         start = chunk_array[nth_task - 1];
4514         end   = chunk_array[nth_task];
4515       } else {
4516         assert(nth_task == (uint)chunk_top, "Control point invariant");
4517         start = chunk_array[chunk_top - 1];
4518         end   = space->top();
4519       }
4520       MemRegion mr(start, end);
4521       // Verify that mr is in space
4522       assert(mr.is_empty() || space->used_region().contains(mr),
4523              "Should be in space");
4524       // Verify that "start" is an object boundary
4525       assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4526              "Should be an oop");
4527       space->par_oop_iterate(mr, cl);
4528     }
4529     pst->all_tasks_completed();
4530   }
4531 }
4532 
4533 void
4534 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4535   CompactibleFreeListSpace* sp, int i,
4536   ParMarkRefsIntoAndScanClosure* cl) {
4537   // Until all tasks completed:
4538   // . claim an unclaimed task
4539   // . compute region boundaries corresponding to task claimed
4540   // . transfer dirty bits ct->mut for that region
4541   // . apply rescanclosure to dirty mut bits for that region
4542 
4543   ResourceMark rm;
4544   HandleMark   hm;
4545 
4546   OopTaskQueue* work_q = work_queue(i);
4547   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4548   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4549   // CAUTION: This closure has state that persists across calls to
4550   // the work method dirty_range_iterate_clear() in that it has
4551   // embedded in it a (subtype of) UpwardsObjectClosure. The
4552   // use of that state in the embedded UpwardsObjectClosure instance
4553   // assumes that the cards are always iterated (even if in parallel
4554   // by several threads) in monotonically increasing order per each
4555   // thread. This is true of the implementation below which picks
4556   // card ranges (chunks) in monotonically increasing order globally
4557   // and, a-fortiori, in monotonically increasing order per thread
4558   // (the latter order being a subsequence of the former).
4559   // If the work code below is ever reorganized into a more chaotic
4560   // work-partitioning form than the current "sequential tasks"
4561   // paradigm, the use of that persistent state will have to be
4562   // revisited and modified appropriately. See also related
4563   // bug 4756801 work on which should examine this code to make
4564   // sure that the changes there do not run counter to the
4565   // assumptions made here and necessary for correctness and
4566   // efficiency. Note also that this code might yield inefficient
4567   // behavior in the case of very large objects that span one or
4568   // more work chunks. Such objects would potentially be scanned
4569   // several times redundantly. Work on 4756801 should try and
4570   // address that performance anomaly if at all possible. XXX
4571   MemRegion  full_span  = _collector->_span;
4572   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
4573   MarkFromDirtyCardsClosure
4574     greyRescanClosure(_collector, full_span, // entire span of interest
4575                       sp, bm, work_q, cl);
4576 
4577   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4578   assert(pst->valid(), "Uninitialized use?");
4579   uint nth_task = 0;
4580   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4581   MemRegion span = sp->used_region();
4582   HeapWord* start_addr = span.start();
4583   HeapWord* end_addr = align_up(span.end(), alignment);
4584   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4585   assert(is_aligned(start_addr, alignment), "Check alignment");
4586   assert(is_aligned(chunk_size, alignment), "Check alignment");
4587 
4588   while (!pst->is_task_claimed(/* reference */ nth_task)) {
4589     // Having claimed the nth_task, compute corresponding mem-region,
4590     // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4591     // The alignment restriction ensures that we do not need any
4592     // synchronization with other gang-workers while setting or
4593     // clearing bits in thus chunk of the MUT.
4594     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4595                                     start_addr + (nth_task+1)*chunk_size);
4596     // The last chunk's end might be way beyond end of the
4597     // used region. In that case pull back appropriately.
4598     if (this_span.end() > end_addr) {
4599       this_span.set_end(end_addr);
4600       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4601     }
4602     // Iterate over the dirty cards covering this chunk, marking them
4603     // precleaned, and setting the corresponding bits in the mod union
4604     // table. Since we have been careful to partition at Card and MUT-word
4605     // boundaries no synchronization is needed between parallel threads.
4606     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4607                                                  &modUnionClosure);
4608 
4609     // Having transferred these marks into the modUnionTable,
4610     // rescan the marked objects on the dirty cards in the modUnionTable.
4611     // Even if this is at a synchronous collection, the initial marking
4612     // may have been done during an asynchronous collection so there
4613     // may be dirty bits in the mod-union table.
4614     _collector->_modUnionTable.dirty_range_iterate_clear(
4615                   this_span, &greyRescanClosure);
4616     _collector->_modUnionTable.verifyNoOneBitsInRange(
4617                                  this_span.start(),
4618                                  this_span.end());
4619   }
4620   pst->all_tasks_completed();  // declare that i am done
4621 }
4622 
4623 // . see if we can share work_queues with ParNew? XXX
4624 void
4625 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4626                                 int* seed) {
4627   OopTaskQueue* work_q = work_queue(i);
4628   NOT_PRODUCT(int num_steals = 0;)
4629   oop obj_to_scan;
4630   CMSBitMap* bm = &(_collector->_markBitMap);
4631 
4632   while (true) {
4633     // Completely finish any left over work from (an) earlier round(s)
4634     cl->trim_queue(0);
4635     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4636                                          (size_t)ParGCDesiredObjsFromOverflowList);
4637     // Now check if there's any work in the overflow list
4638     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4639     // only affects the number of attempts made to get work from the
4640     // overflow list and does not affect the number of workers.  Just
4641     // pass ParallelGCThreads so this behavior is unchanged.
4642     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4643                                                 work_q,
4644                                                 ParallelGCThreads)) {
4645       // found something in global overflow list;
4646       // not yet ready to go stealing work from others.
4647       // We'd like to assert(work_q->size() != 0, ...)
4648       // because we just took work from the overflow list,
4649       // but of course we can't since all of that could have
4650       // been already stolen from us.
4651       // "He giveth and He taketh away."
4652       continue;
4653     }
4654     // Verify that we have no work before we resort to stealing
4655     assert(work_q->size() == 0, "Have work, shouldn't steal");
4656     // Try to steal from other queues that have work
4657     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4658       NOT_PRODUCT(num_steals++;)
4659       assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4660       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4661       // Do scanning work
4662       obj_to_scan->oop_iterate(cl);
4663       // Loop around, finish this work, and try to steal some more
4664     } else if (terminator()->offer_termination()) {
4665         break;  // nirvana from the infinite cycle
4666     }
4667   }
4668   log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4669   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4670          "Else our work is not yet done");
4671 }
4672 
4673 // Record object boundaries in _eden_chunk_array by sampling the eden
4674 // top in the slow-path eden object allocation code path and record
4675 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4676 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4677 // sampling in sample_eden() that activates during the part of the
4678 // preclean phase.
4679 void CMSCollector::sample_eden_chunk() {
4680   if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4681     if (_eden_chunk_lock->try_lock()) {
4682       // Record a sample. This is the critical section. The contents
4683       // of the _eden_chunk_array have to be non-decreasing in the
4684       // address order.
4685       _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4686       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4687              "Unexpected state of Eden");
4688       if (_eden_chunk_index == 0 ||
4689           ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4690            (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4691                           _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4692         _eden_chunk_index++;  // commit sample
4693       }
4694       _eden_chunk_lock->unlock();
4695     }
4696   }
4697 }
4698 
4699 // Return a thread-local PLAB recording array, as appropriate.
4700 void* CMSCollector::get_data_recorder(int thr_num) {
4701   if (_survivor_plab_array != NULL &&
4702       (CMSPLABRecordAlways ||
4703        (_collectorState > Marking && _collectorState < FinalMarking))) {
4704     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4705     ChunkArray* ca = &_survivor_plab_array[thr_num];
4706     ca->reset();   // clear it so that fresh data is recorded
4707     return (void*) ca;
4708   } else {
4709     return NULL;
4710   }
4711 }
4712 
4713 // Reset all the thread-local PLAB recording arrays
4714 void CMSCollector::reset_survivor_plab_arrays() {
4715   for (uint i = 0; i < ParallelGCThreads; i++) {
4716     _survivor_plab_array[i].reset();
4717   }
4718 }
4719 
4720 // Merge the per-thread plab arrays into the global survivor chunk
4721 // array which will provide the partitioning of the survivor space
4722 // for CMS initial scan and rescan.
4723 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4724                                               int no_of_gc_threads) {
4725   assert(_survivor_plab_array  != NULL, "Error");
4726   assert(_survivor_chunk_array != NULL, "Error");
4727   assert(_collectorState == FinalMarking ||
4728          (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4729   for (int j = 0; j < no_of_gc_threads; j++) {
4730     _cursor[j] = 0;
4731   }
4732   HeapWord* top = surv->top();
4733   size_t i;
4734   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
4735     HeapWord* min_val = top;          // Higher than any PLAB address
4736     uint      min_tid = 0;            // position of min_val this round
4737     for (int j = 0; j < no_of_gc_threads; j++) {
4738       ChunkArray* cur_sca = &_survivor_plab_array[j];
4739       if (_cursor[j] == cur_sca->end()) {
4740         continue;
4741       }
4742       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4743       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4744       assert(surv->used_region().contains(cur_val), "Out of bounds value");
4745       if (cur_val < min_val) {
4746         min_tid = j;
4747         min_val = cur_val;
4748       } else {
4749         assert(cur_val < top, "All recorded addresses should be less");
4750       }
4751     }
4752     // At this point min_val and min_tid are respectively
4753     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4754     // and the thread (j) that witnesses that address.
4755     // We record this address in the _survivor_chunk_array[i]
4756     // and increment _cursor[min_tid] prior to the next round i.
4757     if (min_val == top) {
4758       break;
4759     }
4760     _survivor_chunk_array[i] = min_val;
4761     _cursor[min_tid]++;
4762   }
4763   // We are all done; record the size of the _survivor_chunk_array
4764   _survivor_chunk_index = i; // exclusive: [0, i)
4765   log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4766   // Verify that we used up all the recorded entries
4767   #ifdef ASSERT
4768     size_t total = 0;
4769     for (int j = 0; j < no_of_gc_threads; j++) {
4770       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4771       total += _cursor[j];
4772     }
4773     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4774     // Check that the merged array is in sorted order
4775     if (total > 0) {
4776       for (size_t i = 0; i < total - 1; i++) {
4777         log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4778                                      i, p2i(_survivor_chunk_array[i]));
4779         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4780                "Not sorted");
4781       }
4782     }
4783   #endif // ASSERT
4784 }
4785 
4786 // Set up the space's par_seq_tasks structure for work claiming
4787 // for parallel initial scan and rescan of young gen.
4788 // See ParRescanTask where this is currently used.
4789 void
4790 CMSCollector::
4791 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4792   assert(n_threads > 0, "Unexpected n_threads argument");
4793 
4794   // Eden space
4795   if (!_young_gen->eden()->is_empty()) {
4796     SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4797     assert(!pst->valid(), "Clobbering existing data?");
4798     // Each valid entry in [0, _eden_chunk_index) represents a task.
4799     size_t n_tasks = _eden_chunk_index + 1;
4800     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4801     // Sets the condition for completion of the subtask (how many threads
4802     // need to finish in order to be done).
4803     pst->set_n_threads(n_threads);
4804     pst->set_n_tasks((int)n_tasks);
4805   }
4806 
4807   // Merge the survivor plab arrays into _survivor_chunk_array
4808   if (_survivor_plab_array != NULL) {
4809     merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4810   } else {
4811     assert(_survivor_chunk_index == 0, "Error");
4812   }
4813 
4814   // To space
4815   {
4816     SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4817     assert(!pst->valid(), "Clobbering existing data?");
4818     // Sets the condition for completion of the subtask (how many threads
4819     // need to finish in order to be done).
4820     pst->set_n_threads(n_threads);
4821     pst->set_n_tasks(1);
4822     assert(pst->valid(), "Error");
4823   }
4824 
4825   // From space
4826   {
4827     SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4828     assert(!pst->valid(), "Clobbering existing data?");
4829     size_t n_tasks = _survivor_chunk_index + 1;
4830     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4831     // Sets the condition for completion of the subtask (how many threads
4832     // need to finish in order to be done).
4833     pst->set_n_threads(n_threads);
4834     pst->set_n_tasks((int)n_tasks);
4835     assert(pst->valid(), "Error");
4836   }
4837 }
4838 
4839 // Parallel version of remark
4840 void CMSCollector::do_remark_parallel() {
4841   GenCollectedHeap* gch = GenCollectedHeap::heap();
4842   WorkGang* workers = gch->workers();
4843   assert(workers != NULL, "Need parallel worker threads.");
4844   // Choose to use the number of GC workers most recently set
4845   // into "active_workers".
4846   uint n_workers = workers->active_workers();
4847 
4848   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
4849 
4850   StrongRootsScope srs(n_workers);
4851 
4852   CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4853 
4854   // We won't be iterating over the cards in the card table updating
4855   // the younger_gen cards, so we shouldn't call the following else
4856   // the verification code as well as subsequent younger_refs_iterate
4857   // code would get confused. XXX
4858   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4859 
4860   // The young gen rescan work will not be done as part of
4861   // process_roots (which currently doesn't know how to
4862   // parallelize such a scan), but rather will be broken up into
4863   // a set of parallel tasks (via the sampling that the [abortable]
4864   // preclean phase did of eden, plus the [two] tasks of
4865   // scanning the [two] survivor spaces. Further fine-grain
4866   // parallelization of the scanning of the survivor spaces
4867   // themselves, and of precleaning of the young gen itself
4868   // is deferred to the future.
4869   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4870 
4871   // The dirty card rescan work is broken up into a "sequence"
4872   // of parallel tasks (per constituent space) that are dynamically
4873   // claimed by the parallel threads.
4874   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4875 
4876   // It turns out that even when we're using 1 thread, doing the work in a
4877   // separate thread causes wide variance in run times.  We can't help this
4878   // in the multi-threaded case, but we special-case n=1 here to get
4879   // repeatable measurements of the 1-thread overhead of the parallel code.
4880   if (n_workers > 1) {
4881     // Make refs discovery MT-safe, if it isn't already: it may not
4882     // necessarily be so, since it's possible that we are doing
4883     // ST marking.
4884     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4885     workers->run_task(&tsk);
4886   } else {
4887     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4888     tsk.work(0);
4889   }
4890 
4891   // restore, single-threaded for now, any preserved marks
4892   // as a result of work_q overflow
4893   restore_preserved_marks_if_any();
4894 }
4895 
4896 // Non-parallel version of remark
4897 void CMSCollector::do_remark_non_parallel() {
4898   ResourceMark rm;
4899   HandleMark   hm;
4900   GenCollectedHeap* gch = GenCollectedHeap::heap();
4901   ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4902 
4903   MarkRefsIntoAndScanClosure
4904     mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4905              &_markStack, this,
4906              false /* should_yield */, false /* not precleaning */);
4907   MarkFromDirtyCardsClosure
4908     markFromDirtyCardsClosure(this, _span,
4909                               NULL,  // space is set further below
4910                               &_markBitMap, &_markStack, &mrias_cl);
4911   {
4912     GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4913     // Iterate over the dirty cards, setting the corresponding bits in the
4914     // mod union table.
4915     {
4916       ModUnionClosure modUnionClosure(&_modUnionTable);
4917       _ct->ct_bs()->dirty_card_iterate(
4918                       _cmsGen->used_region(),
4919                       &modUnionClosure);
4920     }
4921     // Having transferred these marks into the modUnionTable, we just need
4922     // to rescan the marked objects on the dirty cards in the modUnionTable.
4923     // The initial marking may have been done during an asynchronous
4924     // collection so there may be dirty bits in the mod-union table.
4925     const int alignment =
4926       CardTableModRefBS::card_size * BitsPerWord;
4927     {
4928       // ... First handle dirty cards in CMS gen
4929       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4930       MemRegion ur = _cmsGen->used_region();
4931       HeapWord* lb = ur.start();
4932       HeapWord* ub = align_up(ur.end(), alignment);
4933       MemRegion cms_span(lb, ub);
4934       _modUnionTable.dirty_range_iterate_clear(cms_span,
4935                                                &markFromDirtyCardsClosure);
4936       verify_work_stacks_empty();
4937       log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4938     }
4939   }
4940   if (VerifyDuringGC &&
4941       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4942     HandleMark hm;  // Discard invalid handles created during verification
4943     Universe::verify();
4944   }
4945   {
4946     GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4947 
4948     verify_work_stacks_empty();
4949 
4950     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4951     StrongRootsScope srs(1);
4952 
4953     gch->cms_process_roots(&srs,
4954                            true,  // young gen as roots
4955                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
4956                            should_unload_classes(),
4957                            &mrias_cl,
4958                            NULL); // The dirty klasses will be handled below
4959 
4960     assert(should_unload_classes()
4961            || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4962            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4963   }
4964 
4965   {
4966     GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4967 
4968     verify_work_stacks_empty();
4969 
4970     // Scan all class loader data objects that might have been introduced
4971     // during concurrent marking.
4972     ResourceMark rm;
4973     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4974     for (int i = 0; i < array->length(); i++) {
4975       mrias_cl.do_cld_nv(array->at(i));
4976     }
4977 
4978     // We don't need to keep track of new CLDs anymore.
4979     ClassLoaderDataGraph::remember_new_clds(false);
4980 
4981     verify_work_stacks_empty();
4982   }
4983 
4984   {
4985     GCTraceTime(Trace, gc, phases) t("Dirty Klass Scan", _gc_timer_cm);
4986 
4987     verify_work_stacks_empty();
4988 
4989     RemarkKlassClosure remark_klass_closure(&mrias_cl);
4990     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4991 
4992     verify_work_stacks_empty();
4993   }
4994 
4995   // We might have added oops to ClassLoaderData::_handles during the
4996   // concurrent marking phase. These oops point to newly allocated objects
4997   // that are guaranteed to be kept alive. Either by the direct allocation
4998   // code, or when the young collector processes the roots. Hence,
4999   // we don't have to revisit the _handles block during the remark phase.
5000 
5001   verify_work_stacks_empty();
5002   // Restore evacuated mark words, if any, used for overflow list links
5003   restore_preserved_marks_if_any();
5004 
5005   verify_overflow_empty();
5006 }
5007 
5008 ////////////////////////////////////////////////////////
5009 // Parallel Reference Processing Task Proxy Class
5010 ////////////////////////////////////////////////////////
5011 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5012   OopTaskQueueSet*       _queues;
5013   ParallelTaskTerminator _terminator;
5014  public:
5015   AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5016     AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5017   ParallelTaskTerminator* terminator() { return &_terminator; }
5018   OopTaskQueueSet* queues() { return _queues; }
5019 };
5020 
5021 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5022   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5023   CMSCollector*          _collector;
5024   CMSBitMap*             _mark_bit_map;
5025   const MemRegion        _span;
5026   ProcessTask&           _task;
5027 
5028 public:
5029   CMSRefProcTaskProxy(ProcessTask&     task,
5030                       CMSCollector*    collector,
5031                       const MemRegion& span,
5032                       CMSBitMap*       mark_bit_map,
5033                       AbstractWorkGang* workers,
5034                       OopTaskQueueSet* task_queues):
5035     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5036       task_queues,
5037       workers->active_workers()),
5038     _task(task),
5039     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5040   {
5041     assert(_collector->_span.equals(_span) && !_span.is_empty(),
5042            "Inconsistency in _span");
5043   }
5044 
5045   OopTaskQueueSet* task_queues() { return queues(); }
5046 
5047   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5048 
5049   void do_work_steal(int i,
5050                      CMSParDrainMarkingStackClosure* drain,
5051                      CMSParKeepAliveClosure* keep_alive,
5052                      int* seed);
5053 
5054   virtual void work(uint worker_id);
5055 };
5056 
5057 void CMSRefProcTaskProxy::work(uint worker_id) {
5058   ResourceMark rm;
5059   HandleMark hm;
5060   assert(_collector->_span.equals(_span), "Inconsistency in _span");
5061   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5062                                         _mark_bit_map,
5063                                         work_queue(worker_id));
5064   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5065                                                  _mark_bit_map,
5066                                                  work_queue(worker_id));
5067   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5068   _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5069   if (_task.marks_oops_alive()) {
5070     do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5071                   _collector->hash_seed(worker_id));
5072   }
5073   assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5074   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5075 }
5076 
5077 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5078   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5079   EnqueueTask& _task;
5080 
5081 public:
5082   CMSRefEnqueueTaskProxy(EnqueueTask& task)
5083     : AbstractGangTask("Enqueue reference objects in parallel"),
5084       _task(task)
5085   { }
5086 
5087   virtual void work(uint worker_id)
5088   {
5089     _task.work(worker_id);
5090   }
5091 };
5092 
5093 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5094   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5095    _span(span),
5096    _bit_map(bit_map),
5097    _work_queue(work_queue),
5098    _mark_and_push(collector, span, bit_map, work_queue),
5099    _low_water_mark(MIN2((work_queue->max_elems()/4),
5100                         ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5101 { }
5102 
5103 // . see if we can share work_queues with ParNew? XXX
5104 void CMSRefProcTaskProxy::do_work_steal(int i,
5105   CMSParDrainMarkingStackClosure* drain,
5106   CMSParKeepAliveClosure* keep_alive,
5107   int* seed) {
5108   OopTaskQueue* work_q = work_queue(i);
5109   NOT_PRODUCT(int num_steals = 0;)
5110   oop obj_to_scan;
5111 
5112   while (true) {
5113     // Completely finish any left over work from (an) earlier round(s)
5114     drain->trim_queue(0);
5115     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5116                                          (size_t)ParGCDesiredObjsFromOverflowList);
5117     // Now check if there's any work in the overflow list
5118     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5119     // only affects the number of attempts made to get work from the
5120     // overflow list and does not affect the number of workers.  Just
5121     // pass ParallelGCThreads so this behavior is unchanged.
5122     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5123                                                 work_q,
5124                                                 ParallelGCThreads)) {
5125       // Found something in global overflow list;
5126       // not yet ready to go stealing work from others.
5127       // We'd like to assert(work_q->size() != 0, ...)
5128       // because we just took work from the overflow list,
5129       // but of course we can't, since all of that might have
5130       // been already stolen from us.
5131       continue;
5132     }
5133     // Verify that we have no work before we resort to stealing
5134     assert(work_q->size() == 0, "Have work, shouldn't steal");
5135     // Try to steal from other queues that have work
5136     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5137       NOT_PRODUCT(num_steals++;)
5138       assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5139       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5140       // Do scanning work
5141       obj_to_scan->oop_iterate(keep_alive);
5142       // Loop around, finish this work, and try to steal some more
5143     } else if (terminator()->offer_termination()) {
5144       break;  // nirvana from the infinite cycle
5145     }
5146   }
5147   log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5148 }
5149 
5150 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5151 {
5152   GenCollectedHeap* gch = GenCollectedHeap::heap();
5153   WorkGang* workers = gch->workers();
5154   assert(workers != NULL, "Need parallel worker threads.");
5155   CMSRefProcTaskProxy rp_task(task, &_collector,
5156                               _collector.ref_processor()->span(),
5157                               _collector.markBitMap(),
5158                               workers, _collector.task_queues());
5159   workers->run_task(&rp_task);
5160 }
5161 
5162 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5163 {
5164 
5165   GenCollectedHeap* gch = GenCollectedHeap::heap();
5166   WorkGang* workers = gch->workers();
5167   assert(workers != NULL, "Need parallel worker threads.");
5168   CMSRefEnqueueTaskProxy enq_task(task);
5169   workers->run_task(&enq_task);
5170 }
5171 
5172 void CMSCollector::refProcessingWork() {
5173   ResourceMark rm;
5174   HandleMark   hm;
5175 
5176   ReferenceProcessor* rp = ref_processor();
5177   assert(rp->span().equals(_span), "Spans should be equal");
5178   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5179   // Process weak references.
5180   rp->setup_policy(false);
5181   verify_work_stacks_empty();
5182 
5183   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5184                                           &_markStack, false /* !preclean */);
5185   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5186                                 _span, &_markBitMap, &_markStack,
5187                                 &cmsKeepAliveClosure, false /* !preclean */);
5188   ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
5189   {
5190     GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5191 
5192     ReferenceProcessorStats stats;
5193     if (rp->processing_is_mt()) {
5194       // Set the degree of MT here.  If the discovery is done MT, there
5195       // may have been a different number of threads doing the discovery
5196       // and a different number of discovered lists may have Ref objects.
5197       // That is OK as long as the Reference lists are balanced (see
5198       // balance_all_queues() and balance_queues()).
5199       GenCollectedHeap* gch = GenCollectedHeap::heap();
5200       uint active_workers = ParallelGCThreads;
5201       WorkGang* workers = gch->workers();
5202       if (workers != NULL) {
5203         active_workers = workers->active_workers();
5204         // The expectation is that active_workers will have already
5205         // been set to a reasonable value.  If it has not been set,
5206         // investigate.
5207         assert(active_workers > 0, "Should have been set during scavenge");
5208       }
5209       rp->set_active_mt_degree(active_workers);
5210       CMSRefProcTaskExecutor task_executor(*this);
5211       stats = rp->process_discovered_references(&_is_alive_closure,
5212                                         &cmsKeepAliveClosure,
5213                                         &cmsDrainMarkingStackClosure,
5214                                         &task_executor,
5215                                         &pt);
5216     } else {
5217       stats = rp->process_discovered_references(&_is_alive_closure,
5218                                         &cmsKeepAliveClosure,
5219                                         &cmsDrainMarkingStackClosure,
5220                                         NULL,
5221                                         &pt);
5222     }
5223     _gc_tracer_cm->report_gc_reference_stats(stats);
5224     pt.print_all_references();
5225   }
5226 
5227   // This is the point where the entire marking should have completed.
5228   verify_work_stacks_empty();
5229 
5230   if (should_unload_classes()) {
5231     {
5232       GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5233 
5234       // Unload classes and purge the SystemDictionary.
5235       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm);
5236 
5237       // Unload nmethods.
5238       CodeCache::do_unloading(&_is_alive_closure, purged_class);
5239 
5240       // Prune dead klasses from subklass/sibling/implementor lists.
5241       Klass::clean_weak_klass_links(&_is_alive_closure);
5242     }
5243 
5244     {
5245       GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5246       // Clean up unreferenced symbols in symbol table.
5247       SymbolTable::unlink();
5248     }
5249 
5250     {
5251       GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5252       // Delete entries for dead interned strings.
5253       StringTable::unlink(&_is_alive_closure);
5254     }
5255   }
5256 
5257   // Restore any preserved marks as a result of mark stack or
5258   // work queue overflow
5259   restore_preserved_marks_if_any();  // done single-threaded for now
5260 
5261   rp->set_enqueuing_is_done(true);
5262   if (rp->processing_is_mt()) {
5263     rp->balance_all_queues();
5264     CMSRefProcTaskExecutor task_executor(*this);
5265     rp->enqueue_discovered_references(&task_executor, &pt);
5266   } else {
5267     rp->enqueue_discovered_references(NULL, &pt);
5268   }
5269   rp->verify_no_references_recorded();
5270   pt.print_enqueue_phase();
5271   assert(!rp->discovery_enabled(), "should have been disabled");
5272 }
5273 
5274 #ifndef PRODUCT
5275 void CMSCollector::check_correct_thread_executing() {
5276   Thread* t = Thread::current();
5277   // Only the VM thread or the CMS thread should be here.
5278   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5279          "Unexpected thread type");
5280   // If this is the vm thread, the foreground process
5281   // should not be waiting.  Note that _foregroundGCIsActive is
5282   // true while the foreground collector is waiting.
5283   if (_foregroundGCShouldWait) {
5284     // We cannot be the VM thread
5285     assert(t->is_ConcurrentGC_thread(),
5286            "Should be CMS thread");
5287   } else {
5288     // We can be the CMS thread only if we are in a stop-world
5289     // phase of CMS collection.
5290     if (t->is_ConcurrentGC_thread()) {
5291       assert(_collectorState == InitialMarking ||
5292              _collectorState == FinalMarking,
5293              "Should be a stop-world phase");
5294       // The CMS thread should be holding the CMS_token.
5295       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5296              "Potential interference with concurrently "
5297              "executing VM thread");
5298     }
5299   }
5300 }
5301 #endif
5302 
5303 void CMSCollector::sweep() {
5304   assert(_collectorState == Sweeping, "just checking");
5305   check_correct_thread_executing();
5306   verify_work_stacks_empty();
5307   verify_overflow_empty();
5308   increment_sweep_count();
5309   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5310 
5311   _inter_sweep_timer.stop();
5312   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5313 
5314   assert(!_intra_sweep_timer.is_active(), "Should not be active");
5315   _intra_sweep_timer.reset();
5316   _intra_sweep_timer.start();
5317   {
5318     GCTraceCPUTime tcpu;
5319     CMSPhaseAccounting pa(this, "Concurrent Sweep");
5320     // First sweep the old gen
5321     {
5322       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5323                                bitMapLock());
5324       sweepWork(_cmsGen);
5325     }
5326 
5327     // Update Universe::_heap_*_at_gc figures.
5328     // We need all the free list locks to make the abstract state
5329     // transition from Sweeping to Resetting. See detailed note
5330     // further below.
5331     {
5332       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5333       // Update heap occupancy information which is used as
5334       // input to soft ref clearing policy at the next gc.
5335       Universe::update_heap_info_at_gc();
5336       _collectorState = Resizing;
5337     }
5338   }
5339   verify_work_stacks_empty();
5340   verify_overflow_empty();
5341 
5342   if (should_unload_classes()) {
5343     // Delay purge to the beginning of the next safepoint.  Metaspace::contains
5344     // requires that the virtual spaces are stable and not deleted.
5345     ClassLoaderDataGraph::set_should_purge(true);
5346   }
5347 
5348   _intra_sweep_timer.stop();
5349   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5350 
5351   _inter_sweep_timer.reset();
5352   _inter_sweep_timer.start();
5353 
5354   // We need to use a monotonically non-decreasing time in ms
5355   // or we will see time-warp warnings and os::javaTimeMillis()
5356   // does not guarantee monotonicity.
5357   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5358   update_time_of_last_gc(now);
5359 
5360   // NOTE on abstract state transitions:
5361   // Mutators allocate-live and/or mark the mod-union table dirty
5362   // based on the state of the collection.  The former is done in
5363   // the interval [Marking, Sweeping] and the latter in the interval
5364   // [Marking, Sweeping).  Thus the transitions into the Marking state
5365   // and out of the Sweeping state must be synchronously visible
5366   // globally to the mutators.
5367   // The transition into the Marking state happens with the world
5368   // stopped so the mutators will globally see it.  Sweeping is
5369   // done asynchronously by the background collector so the transition
5370   // from the Sweeping state to the Resizing state must be done
5371   // under the freelistLock (as is the check for whether to
5372   // allocate-live and whether to dirty the mod-union table).
5373   assert(_collectorState == Resizing, "Change of collector state to"
5374     " Resizing must be done under the freelistLocks (plural)");
5375 
5376   // Now that sweeping has been completed, we clear
5377   // the incremental_collection_failed flag,
5378   // thus inviting a younger gen collection to promote into
5379   // this generation. If such a promotion may still fail,
5380   // the flag will be set again when a young collection is
5381   // attempted.
5382   GenCollectedHeap* gch = GenCollectedHeap::heap();
5383   gch->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
5384   gch->update_full_collections_completed(_collection_count_start);
5385 }
5386 
5387 // FIX ME!!! Looks like this belongs in CFLSpace, with
5388 // CMSGen merely delegating to it.
5389 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5390   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5391   HeapWord*  minAddr        = _cmsSpace->bottom();
5392   HeapWord*  largestAddr    =
5393     (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5394   if (largestAddr == NULL) {
5395     // The dictionary appears to be empty.  In this case
5396     // try to coalesce at the end of the heap.
5397     largestAddr = _cmsSpace->end();
5398   }
5399   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
5400   size_t nearLargestOffset =
5401     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5402   log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5403                           p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5404   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5405 }
5406 
5407 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5408   return addr >= _cmsSpace->nearLargestChunk();
5409 }
5410 
5411 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5412   return _cmsSpace->find_chunk_at_end();
5413 }
5414 
5415 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5416                                                     bool full) {
5417   // If the young generation has been collected, gather any statistics
5418   // that are of interest at this point.
5419   bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5420   if (!full && current_is_young) {
5421     // Gather statistics on the young generation collection.
5422     collector()->stats().record_gc0_end(used());
5423   }
5424 }
5425 
5426 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5427   // We iterate over the space(s) underlying this generation,
5428   // checking the mark bit map to see if the bits corresponding
5429   // to specific blocks are marked or not. Blocks that are
5430   // marked are live and are not swept up. All remaining blocks
5431   // are swept up, with coalescing on-the-fly as we sweep up
5432   // contiguous free and/or garbage blocks:
5433   // We need to ensure that the sweeper synchronizes with allocators
5434   // and stop-the-world collectors. In particular, the following
5435   // locks are used:
5436   // . CMS token: if this is held, a stop the world collection cannot occur
5437   // . freelistLock: if this is held no allocation can occur from this
5438   //                 generation by another thread
5439   // . bitMapLock: if this is held, no other thread can access or update
5440   //
5441 
5442   // Note that we need to hold the freelistLock if we use
5443   // block iterate below; else the iterator might go awry if
5444   // a mutator (or promotion) causes block contents to change
5445   // (for instance if the allocator divvies up a block).
5446   // If we hold the free list lock, for all practical purposes
5447   // young generation GC's can't occur (they'll usually need to
5448   // promote), so we might as well prevent all young generation
5449   // GC's while we do a sweeping step. For the same reason, we might
5450   // as well take the bit map lock for the entire duration
5451 
5452   // check that we hold the requisite locks
5453   assert(have_cms_token(), "Should hold cms token");
5454   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5455   assert_lock_strong(old_gen->freelistLock());
5456   assert_lock_strong(bitMapLock());
5457 
5458   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5459   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
5460   old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5461                                           _inter_sweep_estimate.padded_average(),
5462                                           _intra_sweep_estimate.padded_average());
5463   old_gen->setNearLargestChunk();
5464 
5465   {
5466     SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5467     old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5468     // We need to free-up/coalesce garbage/blocks from a
5469     // co-terminal free run. This is done in the SweepClosure
5470     // destructor; so, do not remove this scope, else the
5471     // end-of-sweep-census below will be off by a little bit.
5472   }
5473   old_gen->cmsSpace()->sweep_completed();
5474   old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5475   if (should_unload_classes()) {                // unloaded classes this cycle,
5476     _concurrent_cycles_since_last_unload = 0;   // ... reset count
5477   } else {                                      // did not unload classes,
5478     _concurrent_cycles_since_last_unload++;     // ... increment count
5479   }
5480 }
5481 
5482 // Reset CMS data structures (for now just the marking bit map)
5483 // preparatory for the next cycle.
5484 void CMSCollector::reset_concurrent() {
5485   CMSTokenSyncWithLocks ts(true, bitMapLock());
5486 
5487   // If the state is not "Resetting", the foreground  thread
5488   // has done a collection and the resetting.
5489   if (_collectorState != Resetting) {
5490     assert(_collectorState == Idling, "The state should only change"
5491       " because the foreground collector has finished the collection");
5492     return;
5493   }
5494 
5495   {
5496     // Clear the mark bitmap (no grey objects to start with)
5497     // for the next cycle.
5498     GCTraceCPUTime tcpu;
5499     CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5500 
5501     HeapWord* curAddr = _markBitMap.startWord();
5502     while (curAddr < _markBitMap.endWord()) {
5503       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
5504       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5505       _markBitMap.clear_large_range(chunk);
5506       if (ConcurrentMarkSweepThread::should_yield() &&
5507           !foregroundGCIsActive() &&
5508           CMSYield) {
5509         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5510                "CMS thread should hold CMS token");
5511         assert_lock_strong(bitMapLock());
5512         bitMapLock()->unlock();
5513         ConcurrentMarkSweepThread::desynchronize(true);
5514         stopTimer();
5515         incrementYields();
5516 
5517         // See the comment in coordinator_yield()
5518         for (unsigned i = 0; i < CMSYieldSleepCount &&
5519                          ConcurrentMarkSweepThread::should_yield() &&
5520                          !CMSCollector::foregroundGCIsActive(); ++i) {
5521           os::sleep(Thread::current(), 1, false);
5522         }
5523 
5524         ConcurrentMarkSweepThread::synchronize(true);
5525         bitMapLock()->lock_without_safepoint_check();
5526         startTimer();
5527       }
5528       curAddr = chunk.end();
5529     }
5530     // A successful mostly concurrent collection has been done.
5531     // Because only the full (i.e., concurrent mode failure) collections
5532     // are being measured for gc overhead limits, clean the "near" flag
5533     // and count.
5534     size_policy()->reset_gc_overhead_limit_count();
5535     _collectorState = Idling;
5536   }
5537 
5538   register_gc_end();
5539 }
5540 
5541 // Same as above but for STW paths
5542 void CMSCollector::reset_stw() {
5543   // already have the lock
5544   assert(_collectorState == Resetting, "just checking");
5545   assert_lock_strong(bitMapLock());
5546   GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id());
5547   _markBitMap.clear_all();
5548   _collectorState = Idling;
5549   register_gc_end();
5550 }
5551 
5552 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5553   GCTraceCPUTime tcpu;
5554   TraceCollectorStats tcs(counters());
5555 
5556   switch (op) {
5557     case CMS_op_checkpointRootsInitial: {
5558       GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5559       SvcGCMarker sgcm(SvcGCMarker::OTHER);
5560       checkpointRootsInitial();
5561       break;
5562     }
5563     case CMS_op_checkpointRootsFinal: {
5564       GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5565       SvcGCMarker sgcm(SvcGCMarker::OTHER);
5566       checkpointRootsFinal();
5567       break;
5568     }
5569     default:
5570       fatal("No such CMS_op");
5571   }
5572 }
5573 
5574 #ifndef PRODUCT
5575 size_t const CMSCollector::skip_header_HeapWords() {
5576   return FreeChunk::header_size();
5577 }
5578 
5579 // Try and collect here conditions that should hold when
5580 // CMS thread is exiting. The idea is that the foreground GC
5581 // thread should not be blocked if it wants to terminate
5582 // the CMS thread and yet continue to run the VM for a while
5583 // after that.
5584 void CMSCollector::verify_ok_to_terminate() const {
5585   assert(Thread::current()->is_ConcurrentGC_thread(),
5586          "should be called by CMS thread");
5587   assert(!_foregroundGCShouldWait, "should be false");
5588   // We could check here that all the various low-level locks
5589   // are not held by the CMS thread, but that is overkill; see
5590   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5591   // is checked.
5592 }
5593 #endif
5594 
5595 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5596    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5597           "missing Printezis mark?");
5598   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5599   size_t size = pointer_delta(nextOneAddr + 1, addr);
5600   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5601          "alignment problem");
5602   assert(size >= 3, "Necessary for Printezis marks to work");
5603   return size;
5604 }
5605 
5606 // A variant of the above (block_size_using_printezis_bits()) except
5607 // that we return 0 if the P-bits are not yet set.
5608 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5609   if (_markBitMap.isMarked(addr + 1)) {
5610     assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5611     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5612     size_t size = pointer_delta(nextOneAddr + 1, addr);
5613     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5614            "alignment problem");
5615     assert(size >= 3, "Necessary for Printezis marks to work");
5616     return size;
5617   }
5618   return 0;
5619 }
5620 
5621 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5622   size_t sz = 0;
5623   oop p = (oop)addr;
5624   if (p->klass_or_null_acquire() != NULL) {
5625     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5626   } else {
5627     sz = block_size_using_printezis_bits(addr);
5628   }
5629   assert(sz > 0, "size must be nonzero");
5630   HeapWord* next_block = addr + sz;
5631   HeapWord* next_card  = align_up(next_block, CardTableModRefBS::card_size);
5632   assert(align_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
5633          align_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5634          "must be different cards");
5635   return next_card;
5636 }
5637 
5638 
5639 // CMS Bit Map Wrapper /////////////////////////////////////////
5640 
5641 // Construct a CMS bit map infrastructure, but don't create the
5642 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5643 // further below.
5644 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5645   _bm(),
5646   _shifter(shifter),
5647   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5648                                     Monitor::_safepoint_check_sometimes) : NULL)
5649 {
5650   _bmStartWord = 0;
5651   _bmWordSize  = 0;
5652 }
5653 
5654 bool CMSBitMap::allocate(MemRegion mr) {
5655   _bmStartWord = mr.start();
5656   _bmWordSize  = mr.word_size();
5657   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5658                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5659   if (!brs.is_reserved()) {
5660     log_warning(gc)("CMS bit map allocation failure");
5661     return false;
5662   }
5663   // For now we'll just commit all of the bit map up front.
5664   // Later on we'll try to be more parsimonious with swap.
5665   if (!_virtual_space.initialize(brs, brs.size())) {
5666     log_warning(gc)("CMS bit map backing store failure");
5667     return false;
5668   }
5669   assert(_virtual_space.committed_size() == brs.size(),
5670          "didn't reserve backing store for all of CMS bit map?");
5671   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5672          _bmWordSize, "inconsistency in bit map sizing");
5673   _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5674 
5675   // bm.clear(); // can we rely on getting zero'd memory? verify below
5676   assert(isAllClear(),
5677          "Expected zero'd memory from ReservedSpace constructor");
5678   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5679          "consistency check");
5680   return true;
5681 }
5682 
5683 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5684   HeapWord *next_addr, *end_addr, *last_addr;
5685   assert_locked();
5686   assert(covers(mr), "out-of-range error");
5687   // XXX assert that start and end are appropriately aligned
5688   for (next_addr = mr.start(), end_addr = mr.end();
5689        next_addr < end_addr; next_addr = last_addr) {
5690     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5691     last_addr = dirty_region.end();
5692     if (!dirty_region.is_empty()) {
5693       cl->do_MemRegion(dirty_region);
5694     } else {
5695       assert(last_addr == end_addr, "program logic");
5696       return;
5697     }
5698   }
5699 }
5700 
5701 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5702   _bm.print_on_error(st, prefix);
5703 }
5704 
5705 #ifndef PRODUCT
5706 void CMSBitMap::assert_locked() const {
5707   CMSLockVerifier::assert_locked(lock());
5708 }
5709 
5710 bool CMSBitMap::covers(MemRegion mr) const {
5711   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5712   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5713          "size inconsistency");
5714   return (mr.start() >= _bmStartWord) &&
5715          (mr.end()   <= endWord());
5716 }
5717 
5718 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5719     return (start >= _bmStartWord && (start + size) <= endWord());
5720 }
5721 
5722 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5723   // verify that there are no 1 bits in the interval [left, right)
5724   FalseBitMapClosure falseBitMapClosure;
5725   iterate(&falseBitMapClosure, left, right);
5726 }
5727 
5728 void CMSBitMap::region_invariant(MemRegion mr)
5729 {
5730   assert_locked();
5731   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5732   assert(!mr.is_empty(), "unexpected empty region");
5733   assert(covers(mr), "mr should be covered by bit map");
5734   // convert address range into offset range
5735   size_t start_ofs = heapWordToOffset(mr.start());
5736   // Make sure that end() is appropriately aligned
5737   assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))),
5738          "Misaligned mr.end()");
5739   size_t end_ofs   = heapWordToOffset(mr.end());
5740   assert(end_ofs > start_ofs, "Should mark at least one bit");
5741 }
5742 
5743 #endif
5744 
5745 bool CMSMarkStack::allocate(size_t size) {
5746   // allocate a stack of the requisite depth
5747   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5748                    size * sizeof(oop)));
5749   if (!rs.is_reserved()) {
5750     log_warning(gc)("CMSMarkStack allocation failure");
5751     return false;
5752   }
5753   if (!_virtual_space.initialize(rs, rs.size())) {
5754     log_warning(gc)("CMSMarkStack backing store failure");
5755     return false;
5756   }
5757   assert(_virtual_space.committed_size() == rs.size(),
5758          "didn't reserve backing store for all of CMS stack?");
5759   _base = (oop*)(_virtual_space.low());
5760   _index = 0;
5761   _capacity = size;
5762   NOT_PRODUCT(_max_depth = 0);
5763   return true;
5764 }
5765 
5766 // XXX FIX ME !!! In the MT case we come in here holding a
5767 // leaf lock. For printing we need to take a further lock
5768 // which has lower rank. We need to recalibrate the two
5769 // lock-ranks involved in order to be able to print the
5770 // messages below. (Or defer the printing to the caller.
5771 // For now we take the expedient path of just disabling the
5772 // messages for the problematic case.)
5773 void CMSMarkStack::expand() {
5774   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5775   if (_capacity == MarkStackSizeMax) {
5776     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5777       // We print a warning message only once per CMS cycle.
5778       log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5779     }
5780     return;
5781   }
5782   // Double capacity if possible
5783   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5784   // Do not give up existing stack until we have managed to
5785   // get the double capacity that we desired.
5786   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5787                    new_capacity * sizeof(oop)));
5788   if (rs.is_reserved()) {
5789     // Release the backing store associated with old stack
5790     _virtual_space.release();
5791     // Reinitialize virtual space for new stack
5792     if (!_virtual_space.initialize(rs, rs.size())) {
5793       fatal("Not enough swap for expanded marking stack");
5794     }
5795     _base = (oop*)(_virtual_space.low());
5796     _index = 0;
5797     _capacity = new_capacity;
5798   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5799     // Failed to double capacity, continue;
5800     // we print a detail message only once per CMS cycle.
5801     log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5802                         _capacity / K, new_capacity / K);
5803   }
5804 }
5805 
5806 
5807 // Closures
5808 // XXX: there seems to be a lot of code  duplication here;
5809 // should refactor and consolidate common code.
5810 
5811 // This closure is used to mark refs into the CMS generation in
5812 // the CMS bit map. Called at the first checkpoint. This closure
5813 // assumes that we do not need to re-mark dirty cards; if the CMS
5814 // generation on which this is used is not an oldest
5815 // generation then this will lose younger_gen cards!
5816 
5817 MarkRefsIntoClosure::MarkRefsIntoClosure(
5818   MemRegion span, CMSBitMap* bitMap):
5819     _span(span),
5820     _bitMap(bitMap)
5821 {
5822   assert(ref_processor() == NULL, "deliberately left NULL");
5823   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5824 }
5825 
5826 void MarkRefsIntoClosure::do_oop(oop obj) {
5827   // if p points into _span, then mark corresponding bit in _markBitMap
5828   assert(oopDesc::is_oop(obj), "expected an oop");
5829   HeapWord* addr = (HeapWord*)obj;
5830   if (_span.contains(addr)) {
5831     // this should be made more efficient
5832     _bitMap->mark(addr);
5833   }
5834 }
5835 
5836 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
5837 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5838 
5839 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5840   MemRegion span, CMSBitMap* bitMap):
5841     _span(span),
5842     _bitMap(bitMap)
5843 {
5844   assert(ref_processor() == NULL, "deliberately left NULL");
5845   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5846 }
5847 
5848 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5849   // if p points into _span, then mark corresponding bit in _markBitMap
5850   assert(oopDesc::is_oop(obj), "expected an oop");
5851   HeapWord* addr = (HeapWord*)obj;
5852   if (_span.contains(addr)) {
5853     // this should be made more efficient
5854     _bitMap->par_mark(addr);
5855   }
5856 }
5857 
5858 void ParMarkRefsIntoClosure::do_oop(oop* p)       { ParMarkRefsIntoClosure::do_oop_work(p); }
5859 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5860 
5861 // A variant of the above, used for CMS marking verification.
5862 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5863   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5864     _span(span),
5865     _verification_bm(verification_bm),
5866     _cms_bm(cms_bm)
5867 {
5868   assert(ref_processor() == NULL, "deliberately left NULL");
5869   assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5870 }
5871 
5872 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5873   // if p points into _span, then mark corresponding bit in _markBitMap
5874   assert(oopDesc::is_oop(obj), "expected an oop");
5875   HeapWord* addr = (HeapWord*)obj;
5876   if (_span.contains(addr)) {
5877     _verification_bm->mark(addr);
5878     if (!_cms_bm->isMarked(addr)) {
5879       Log(gc, verify) log;
5880       ResourceMark rm;
5881       LogStream ls(log.error());
5882       oop(addr)->print_on(&ls);
5883       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5884       fatal("... aborting");
5885     }
5886   }
5887 }
5888 
5889 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5890 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5891 
5892 //////////////////////////////////////////////////
5893 // MarkRefsIntoAndScanClosure
5894 //////////////////////////////////////////////////
5895 
5896 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5897                                                        ReferenceProcessor* rp,
5898                                                        CMSBitMap* bit_map,
5899                                                        CMSBitMap* mod_union_table,
5900                                                        CMSMarkStack*  mark_stack,
5901                                                        CMSCollector* collector,
5902                                                        bool should_yield,
5903                                                        bool concurrent_precleaning):
5904   _collector(collector),
5905   _span(span),
5906   _bit_map(bit_map),
5907   _mark_stack(mark_stack),
5908   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
5909                       mark_stack, concurrent_precleaning),
5910   _yield(should_yield),
5911   _concurrent_precleaning(concurrent_precleaning),
5912   _freelistLock(NULL)
5913 {
5914   // FIXME: Should initialize in base class constructor.
5915   assert(rp != NULL, "ref_processor shouldn't be NULL");
5916   set_ref_processor_internal(rp);
5917 }
5918 
5919 // This closure is used to mark refs into the CMS generation at the
5920 // second (final) checkpoint, and to scan and transitively follow
5921 // the unmarked oops. It is also used during the concurrent precleaning
5922 // phase while scanning objects on dirty cards in the CMS generation.
5923 // The marks are made in the marking bit map and the marking stack is
5924 // used for keeping the (newly) grey objects during the scan.
5925 // The parallel version (Par_...) appears further below.
5926 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5927   if (obj != NULL) {
5928     assert(oopDesc::is_oop(obj), "expected an oop");
5929     HeapWord* addr = (HeapWord*)obj;
5930     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5931     assert(_collector->overflow_list_is_empty(),
5932            "overflow list should be empty");
5933     if (_span.contains(addr) &&
5934         !_bit_map->isMarked(addr)) {
5935       // mark bit map (object is now grey)
5936       _bit_map->mark(addr);
5937       // push on marking stack (stack should be empty), and drain the
5938       // stack by applying this closure to the oops in the oops popped
5939       // from the stack (i.e. blacken the grey objects)
5940       bool res = _mark_stack->push(obj);
5941       assert(res, "Should have space to push on empty stack");
5942       do {
5943         oop new_oop = _mark_stack->pop();
5944         assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5945         assert(_bit_map->isMarked((HeapWord*)new_oop),
5946                "only grey objects on this stack");
5947         // iterate over the oops in this oop, marking and pushing
5948         // the ones in CMS heap (i.e. in _span).
5949         new_oop->oop_iterate(&_pushAndMarkClosure);
5950         // check if it's time to yield
5951         do_yield_check();
5952       } while (!_mark_stack->isEmpty() ||
5953                (!_concurrent_precleaning && take_from_overflow_list()));
5954         // if marking stack is empty, and we are not doing this
5955         // during precleaning, then check the overflow list
5956     }
5957     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5958     assert(_collector->overflow_list_is_empty(),
5959            "overflow list was drained above");
5960 
5961     assert(_collector->no_preserved_marks(),
5962            "All preserved marks should have been restored above");
5963   }
5964 }
5965 
5966 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5967 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5968 
5969 void MarkRefsIntoAndScanClosure::do_yield_work() {
5970   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5971          "CMS thread should hold CMS token");
5972   assert_lock_strong(_freelistLock);
5973   assert_lock_strong(_bit_map->lock());
5974   // relinquish the free_list_lock and bitMaplock()
5975   _bit_map->lock()->unlock();
5976   _freelistLock->unlock();
5977   ConcurrentMarkSweepThread::desynchronize(true);
5978   _collector->stopTimer();
5979   _collector->incrementYields();
5980 
5981   // See the comment in coordinator_yield()
5982   for (unsigned i = 0;
5983        i < CMSYieldSleepCount &&
5984        ConcurrentMarkSweepThread::should_yield() &&
5985        !CMSCollector::foregroundGCIsActive();
5986        ++i) {
5987     os::sleep(Thread::current(), 1, false);
5988   }
5989 
5990   ConcurrentMarkSweepThread::synchronize(true);
5991   _freelistLock->lock_without_safepoint_check();
5992   _bit_map->lock()->lock_without_safepoint_check();
5993   _collector->startTimer();
5994 }
5995 
5996 ///////////////////////////////////////////////////////////
5997 // ParMarkRefsIntoAndScanClosure: a parallel version of
5998 //                                MarkRefsIntoAndScanClosure
5999 ///////////////////////////////////////////////////////////
6000 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6001   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6002   CMSBitMap* bit_map, OopTaskQueue* work_queue):
6003   _span(span),
6004   _bit_map(bit_map),
6005   _work_queue(work_queue),
6006   _low_water_mark(MIN2((work_queue->max_elems()/4),
6007                        ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6008   _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6009 {
6010   // FIXME: Should initialize in base class constructor.
6011   assert(rp != NULL, "ref_processor shouldn't be NULL");
6012   set_ref_processor_internal(rp);
6013 }
6014 
6015 // This closure is used to mark refs into the CMS generation at the
6016 // second (final) checkpoint, and to scan and transitively follow
6017 // the unmarked oops. The marks are made in the marking bit map and
6018 // the work_queue is used for keeping the (newly) grey objects during
6019 // the scan phase whence they are also available for stealing by parallel
6020 // threads. Since the marking bit map is shared, updates are
6021 // synchronized (via CAS).
6022 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6023   if (obj != NULL) {
6024     // Ignore mark word because this could be an already marked oop
6025     // that may be chained at the end of the overflow list.
6026     assert(oopDesc::is_oop(obj, true), "expected an oop");
6027     HeapWord* addr = (HeapWord*)obj;
6028     if (_span.contains(addr) &&
6029         !_bit_map->isMarked(addr)) {
6030       // mark bit map (object will become grey):
6031       // It is possible for several threads to be
6032       // trying to "claim" this object concurrently;
6033       // the unique thread that succeeds in marking the
6034       // object first will do the subsequent push on
6035       // to the work queue (or overflow list).
6036       if (_bit_map->par_mark(addr)) {
6037         // push on work_queue (which may not be empty), and trim the
6038         // queue to an appropriate length by applying this closure to
6039         // the oops in the oops popped from the stack (i.e. blacken the
6040         // grey objects)
6041         bool res = _work_queue->push(obj);
6042         assert(res, "Low water mark should be less than capacity?");
6043         trim_queue(_low_water_mark);
6044       } // Else, another thread claimed the object
6045     }
6046   }
6047 }
6048 
6049 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p)       { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6050 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6051 
6052 // This closure is used to rescan the marked objects on the dirty cards
6053 // in the mod union table and the card table proper.
6054 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6055   oop p, MemRegion mr) {
6056 
6057   size_t size = 0;
6058   HeapWord* addr = (HeapWord*)p;
6059   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6060   assert(_span.contains(addr), "we are scanning the CMS generation");
6061   // check if it's time to yield
6062   if (do_yield_check()) {
6063     // We yielded for some foreground stop-world work,
6064     // and we have been asked to abort this ongoing preclean cycle.
6065     return 0;
6066   }
6067   if (_bitMap->isMarked(addr)) {
6068     // it's marked; is it potentially uninitialized?
6069     if (p->klass_or_null_acquire() != NULL) {
6070         // an initialized object; ignore mark word in verification below
6071         // since we are running concurrent with mutators
6072         assert(oopDesc::is_oop(p, true), "should be an oop");
6073         if (p->is_objArray()) {
6074           // objArrays are precisely marked; restrict scanning
6075           // to dirty cards only.
6076           size = CompactibleFreeListSpace::adjustObjectSize(
6077                    p->oop_iterate_size(_scanningClosure, mr));
6078         } else {
6079           // A non-array may have been imprecisely marked; we need
6080           // to scan object in its entirety.
6081           size = CompactibleFreeListSpace::adjustObjectSize(
6082                    p->oop_iterate_size(_scanningClosure));
6083         }
6084       #ifdef ASSERT
6085         size_t direct_size =
6086           CompactibleFreeListSpace::adjustObjectSize(p->size());
6087         assert(size == direct_size, "Inconsistency in size");
6088         assert(size >= 3, "Necessary for Printezis marks to work");
6089         HeapWord* start_pbit = addr + 1;
6090         HeapWord* end_pbit = addr + size - 1;
6091         assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6092                "inconsistent Printezis mark");
6093         // Verify inner mark bits (between Printezis bits) are clear,
6094         // but don't repeat if there are multiple dirty regions for
6095         // the same object, to avoid potential O(N^2) performance.
6096         if (addr != _last_scanned_object) {
6097           _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6098           _last_scanned_object = addr;
6099         }
6100       #endif // ASSERT
6101     } else {
6102       // An uninitialized object.
6103       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6104       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6105       size = pointer_delta(nextOneAddr + 1, addr);
6106       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6107              "alignment problem");
6108       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6109       // will dirty the card when the klass pointer is installed in the
6110       // object (signaling the completion of initialization).
6111     }
6112   } else {
6113     // Either a not yet marked object or an uninitialized object
6114     if (p->klass_or_null_acquire() == NULL) {
6115       // An uninitialized object, skip to the next card, since
6116       // we may not be able to read its P-bits yet.
6117       assert(size == 0, "Initial value");
6118     } else {
6119       // An object not (yet) reached by marking: we merely need to
6120       // compute its size so as to go look at the next block.
6121       assert(oopDesc::is_oop(p, true), "should be an oop");
6122       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6123     }
6124   }
6125   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6126   return size;
6127 }
6128 
6129 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6130   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6131          "CMS thread should hold CMS token");
6132   assert_lock_strong(_freelistLock);
6133   assert_lock_strong(_bitMap->lock());
6134   // relinquish the free_list_lock and bitMaplock()
6135   _bitMap->lock()->unlock();
6136   _freelistLock->unlock();
6137   ConcurrentMarkSweepThread::desynchronize(true);
6138   _collector->stopTimer();
6139   _collector->incrementYields();
6140 
6141   // See the comment in coordinator_yield()
6142   for (unsigned i = 0; i < CMSYieldSleepCount &&
6143                    ConcurrentMarkSweepThread::should_yield() &&
6144                    !CMSCollector::foregroundGCIsActive(); ++i) {
6145     os::sleep(Thread::current(), 1, false);
6146   }
6147 
6148   ConcurrentMarkSweepThread::synchronize(true);
6149   _freelistLock->lock_without_safepoint_check();
6150   _bitMap->lock()->lock_without_safepoint_check();
6151   _collector->startTimer();
6152 }
6153 
6154 
6155 //////////////////////////////////////////////////////////////////
6156 // SurvivorSpacePrecleanClosure
6157 //////////////////////////////////////////////////////////////////
6158 // This (single-threaded) closure is used to preclean the oops in
6159 // the survivor spaces.
6160 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6161 
6162   HeapWord* addr = (HeapWord*)p;
6163   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6164   assert(!_span.contains(addr), "we are scanning the survivor spaces");
6165   assert(p->klass_or_null() != NULL, "object should be initialized");
6166   // an initialized object; ignore mark word in verification below
6167   // since we are running concurrent with mutators
6168   assert(oopDesc::is_oop(p, true), "should be an oop");
6169   // Note that we do not yield while we iterate over
6170   // the interior oops of p, pushing the relevant ones
6171   // on our marking stack.
6172   size_t size = p->oop_iterate_size(_scanning_closure);
6173   do_yield_check();
6174   // Observe that below, we do not abandon the preclean
6175   // phase as soon as we should; rather we empty the
6176   // marking stack before returning. This is to satisfy
6177   // some existing assertions. In general, it may be a
6178   // good idea to abort immediately and complete the marking
6179   // from the grey objects at a later time.
6180   while (!_mark_stack->isEmpty()) {
6181     oop new_oop = _mark_stack->pop();
6182     assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6183     assert(_bit_map->isMarked((HeapWord*)new_oop),
6184            "only grey objects on this stack");
6185     // iterate over the oops in this oop, marking and pushing
6186     // the ones in CMS heap (i.e. in _span).
6187     new_oop->oop_iterate(_scanning_closure);
6188     // check if it's time to yield
6189     do_yield_check();
6190   }
6191   unsigned int after_count =
6192     GenCollectedHeap::heap()->total_collections();
6193   bool abort = (_before_count != after_count) ||
6194                _collector->should_abort_preclean();
6195   return abort ? 0 : size;
6196 }
6197 
6198 void SurvivorSpacePrecleanClosure::do_yield_work() {
6199   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6200          "CMS thread should hold CMS token");
6201   assert_lock_strong(_bit_map->lock());
6202   // Relinquish the bit map lock
6203   _bit_map->lock()->unlock();
6204   ConcurrentMarkSweepThread::desynchronize(true);
6205   _collector->stopTimer();
6206   _collector->incrementYields();
6207 
6208   // See the comment in coordinator_yield()
6209   for (unsigned i = 0; i < CMSYieldSleepCount &&
6210                        ConcurrentMarkSweepThread::should_yield() &&
6211                        !CMSCollector::foregroundGCIsActive(); ++i) {
6212     os::sleep(Thread::current(), 1, false);
6213   }
6214 
6215   ConcurrentMarkSweepThread::synchronize(true);
6216   _bit_map->lock()->lock_without_safepoint_check();
6217   _collector->startTimer();
6218 }
6219 
6220 // This closure is used to rescan the marked objects on the dirty cards
6221 // in the mod union table and the card table proper. In the parallel
6222 // case, although the bitMap is shared, we do a single read so the
6223 // isMarked() query is "safe".
6224 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6225   // Ignore mark word because we are running concurrent with mutators
6226   assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6227   HeapWord* addr = (HeapWord*)p;
6228   assert(_span.contains(addr), "we are scanning the CMS generation");
6229   bool is_obj_array = false;
6230   #ifdef ASSERT
6231     if (!_parallel) {
6232       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6233       assert(_collector->overflow_list_is_empty(),
6234              "overflow list should be empty");
6235 
6236     }
6237   #endif // ASSERT
6238   if (_bit_map->isMarked(addr)) {
6239     // Obj arrays are precisely marked, non-arrays are not;
6240     // so we scan objArrays precisely and non-arrays in their
6241     // entirety.
6242     if (p->is_objArray()) {
6243       is_obj_array = true;
6244       if (_parallel) {
6245         p->oop_iterate(_par_scan_closure, mr);
6246       } else {
6247         p->oop_iterate(_scan_closure, mr);
6248       }
6249     } else {
6250       if (_parallel) {
6251         p->oop_iterate(_par_scan_closure);
6252       } else {
6253         p->oop_iterate(_scan_closure);
6254       }
6255     }
6256   }
6257   #ifdef ASSERT
6258     if (!_parallel) {
6259       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6260       assert(_collector->overflow_list_is_empty(),
6261              "overflow list should be empty");
6262 
6263     }
6264   #endif // ASSERT
6265   return is_obj_array;
6266 }
6267 
6268 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6269                         MemRegion span,
6270                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
6271                         bool should_yield, bool verifying):
6272   _collector(collector),
6273   _span(span),
6274   _bitMap(bitMap),
6275   _mut(&collector->_modUnionTable),
6276   _markStack(markStack),
6277   _yield(should_yield),
6278   _skipBits(0)
6279 {
6280   assert(_markStack->isEmpty(), "stack should be empty");
6281   _finger = _bitMap->startWord();
6282   _threshold = _finger;
6283   assert(_collector->_restart_addr == NULL, "Sanity check");
6284   assert(_span.contains(_finger), "Out of bounds _finger?");
6285   DEBUG_ONLY(_verifying = verifying;)
6286 }
6287 
6288 void MarkFromRootsClosure::reset(HeapWord* addr) {
6289   assert(_markStack->isEmpty(), "would cause duplicates on stack");
6290   assert(_span.contains(addr), "Out of bounds _finger?");
6291   _finger = addr;
6292   _threshold = align_up(_finger, CardTableModRefBS::card_size);
6293 }
6294 
6295 // Should revisit to see if this should be restructured for
6296 // greater efficiency.
6297 bool MarkFromRootsClosure::do_bit(size_t offset) {
6298   if (_skipBits > 0) {
6299     _skipBits--;
6300     return true;
6301   }
6302   // convert offset into a HeapWord*
6303   HeapWord* addr = _bitMap->startWord() + offset;
6304   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6305          "address out of range");
6306   assert(_bitMap->isMarked(addr), "tautology");
6307   if (_bitMap->isMarked(addr+1)) {
6308     // this is an allocated but not yet initialized object
6309     assert(_skipBits == 0, "tautology");
6310     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
6311     oop p = oop(addr);
6312     if (p->klass_or_null_acquire() == NULL) {
6313       DEBUG_ONLY(if (!_verifying) {)
6314         // We re-dirty the cards on which this object lies and increase
6315         // the _threshold so that we'll come back to scan this object
6316         // during the preclean or remark phase. (CMSCleanOnEnter)
6317         if (CMSCleanOnEnter) {
6318           size_t sz = _collector->block_size_using_printezis_bits(addr);
6319           HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
6320           MemRegion redirty_range = MemRegion(addr, end_card_addr);
6321           assert(!redirty_range.is_empty(), "Arithmetical tautology");
6322           // Bump _threshold to end_card_addr; note that
6323           // _threshold cannot possibly exceed end_card_addr, anyhow.
6324           // This prevents future clearing of the card as the scan proceeds
6325           // to the right.
6326           assert(_threshold <= end_card_addr,
6327                  "Because we are just scanning into this object");
6328           if (_threshold < end_card_addr) {
6329             _threshold = end_card_addr;
6330           }
6331           if (p->klass_or_null_acquire() != NULL) {
6332             // Redirty the range of cards...
6333             _mut->mark_range(redirty_range);
6334           } // ...else the setting of klass will dirty the card anyway.
6335         }
6336       DEBUG_ONLY(})
6337       return true;
6338     }
6339   }
6340   scanOopsInOop(addr);
6341   return true;
6342 }
6343 
6344 // We take a break if we've been at this for a while,
6345 // so as to avoid monopolizing the locks involved.
6346 void MarkFromRootsClosure::do_yield_work() {
6347   // First give up the locks, then yield, then re-lock
6348   // We should probably use a constructor/destructor idiom to
6349   // do this unlock/lock or modify the MutexUnlocker class to
6350   // serve our purpose. XXX
6351   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6352          "CMS thread should hold CMS token");
6353   assert_lock_strong(_bitMap->lock());
6354   _bitMap->lock()->unlock();
6355   ConcurrentMarkSweepThread::desynchronize(true);
6356   _collector->stopTimer();
6357   _collector->incrementYields();
6358 
6359   // See the comment in coordinator_yield()
6360   for (unsigned i = 0; i < CMSYieldSleepCount &&
6361                        ConcurrentMarkSweepThread::should_yield() &&
6362                        !CMSCollector::foregroundGCIsActive(); ++i) {
6363     os::sleep(Thread::current(), 1, false);
6364   }
6365 
6366   ConcurrentMarkSweepThread::synchronize(true);
6367   _bitMap->lock()->lock_without_safepoint_check();
6368   _collector->startTimer();
6369 }
6370 
6371 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6372   assert(_bitMap->isMarked(ptr), "expected bit to be set");
6373   assert(_markStack->isEmpty(),
6374          "should drain stack to limit stack usage");
6375   // convert ptr to an oop preparatory to scanning
6376   oop obj = oop(ptr);
6377   // Ignore mark word in verification below, since we
6378   // may be running concurrent with mutators.
6379   assert(oopDesc::is_oop(obj, true), "should be an oop");
6380   assert(_finger <= ptr, "_finger runneth ahead");
6381   // advance the finger to right end of this object
6382   _finger = ptr + obj->size();
6383   assert(_finger > ptr, "we just incremented it above");
6384   // On large heaps, it may take us some time to get through
6385   // the marking phase. During
6386   // this time it's possible that a lot of mutations have
6387   // accumulated in the card table and the mod union table --
6388   // these mutation records are redundant until we have
6389   // actually traced into the corresponding card.
6390   // Here, we check whether advancing the finger would make
6391   // us cross into a new card, and if so clear corresponding
6392   // cards in the MUT (preclean them in the card-table in the
6393   // future).
6394 
6395   DEBUG_ONLY(if (!_verifying) {)
6396     // The clean-on-enter optimization is disabled by default,
6397     // until we fix 6178663.
6398     if (CMSCleanOnEnter && (_finger > _threshold)) {
6399       // [_threshold, _finger) represents the interval
6400       // of cards to be cleared  in MUT (or precleaned in card table).
6401       // The set of cards to be cleared is all those that overlap
6402       // with the interval [_threshold, _finger); note that
6403       // _threshold is always kept card-aligned but _finger isn't
6404       // always card-aligned.
6405       HeapWord* old_threshold = _threshold;
6406       assert(is_aligned(old_threshold, CardTableModRefBS::card_size),
6407              "_threshold should always be card-aligned");
6408       _threshold = align_up(_finger, CardTableModRefBS::card_size);
6409       MemRegion mr(old_threshold, _threshold);
6410       assert(!mr.is_empty(), "Control point invariant");
6411       assert(_span.contains(mr), "Should clear within span");
6412       _mut->clear_range(mr);
6413     }
6414   DEBUG_ONLY(})
6415   // Note: the finger doesn't advance while we drain
6416   // the stack below.
6417   PushOrMarkClosure pushOrMarkClosure(_collector,
6418                                       _span, _bitMap, _markStack,
6419                                       _finger, this);
6420   bool res = _markStack->push(obj);
6421   assert(res, "Empty non-zero size stack should have space for single push");
6422   while (!_markStack->isEmpty()) {
6423     oop new_oop = _markStack->pop();
6424     // Skip verifying header mark word below because we are
6425     // running concurrent with mutators.
6426     assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6427     // now scan this oop's oops
6428     new_oop->oop_iterate(&pushOrMarkClosure);
6429     do_yield_check();
6430   }
6431   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6432 }
6433 
6434 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6435                        CMSCollector* collector, MemRegion span,
6436                        CMSBitMap* bit_map,
6437                        OopTaskQueue* work_queue,
6438                        CMSMarkStack*  overflow_stack):
6439   _collector(collector),
6440   _whole_span(collector->_span),
6441   _span(span),
6442   _bit_map(bit_map),
6443   _mut(&collector->_modUnionTable),
6444   _work_queue(work_queue),
6445   _overflow_stack(overflow_stack),
6446   _skip_bits(0),
6447   _task(task)
6448 {
6449   assert(_work_queue->size() == 0, "work_queue should be empty");
6450   _finger = span.start();
6451   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
6452   assert(_span.contains(_finger), "Out of bounds _finger?");
6453 }
6454 
6455 // Should revisit to see if this should be restructured for
6456 // greater efficiency.
6457 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6458   if (_skip_bits > 0) {
6459     _skip_bits--;
6460     return true;
6461   }
6462   // convert offset into a HeapWord*
6463   HeapWord* addr = _bit_map->startWord() + offset;
6464   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6465          "address out of range");
6466   assert(_bit_map->isMarked(addr), "tautology");
6467   if (_bit_map->isMarked(addr+1)) {
6468     // this is an allocated object that might not yet be initialized
6469     assert(_skip_bits == 0, "tautology");
6470     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
6471     oop p = oop(addr);
6472     if (p->klass_or_null_acquire() == NULL) {
6473       // in the case of Clean-on-Enter optimization, redirty card
6474       // and avoid clearing card by increasing  the threshold.
6475       return true;
6476     }
6477   }
6478   scan_oops_in_oop(addr);
6479   return true;
6480 }
6481 
6482 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6483   assert(_bit_map->isMarked(ptr), "expected bit to be set");
6484   // Should we assert that our work queue is empty or
6485   // below some drain limit?
6486   assert(_work_queue->size() == 0,
6487          "should drain stack to limit stack usage");
6488   // convert ptr to an oop preparatory to scanning
6489   oop obj = oop(ptr);
6490   // Ignore mark word in verification below, since we
6491   // may be running concurrent with mutators.
6492   assert(oopDesc::is_oop(obj, true), "should be an oop");
6493   assert(_finger <= ptr, "_finger runneth ahead");
6494   // advance the finger to right end of this object
6495   _finger = ptr + obj->size();
6496   assert(_finger > ptr, "we just incremented it above");
6497   // On large heaps, it may take us some time to get through
6498   // the marking phase. During
6499   // this time it's possible that a lot of mutations have
6500   // accumulated in the card table and the mod union table --
6501   // these mutation records are redundant until we have
6502   // actually traced into the corresponding card.
6503   // Here, we check whether advancing the finger would make
6504   // us cross into a new card, and if so clear corresponding
6505   // cards in the MUT (preclean them in the card-table in the
6506   // future).
6507 
6508   // The clean-on-enter optimization is disabled by default,
6509   // until we fix 6178663.
6510   if (CMSCleanOnEnter && (_finger > _threshold)) {
6511     // [_threshold, _finger) represents the interval
6512     // of cards to be cleared  in MUT (or precleaned in card table).
6513     // The set of cards to be cleared is all those that overlap
6514     // with the interval [_threshold, _finger); note that
6515     // _threshold is always kept card-aligned but _finger isn't
6516     // always card-aligned.
6517     HeapWord* old_threshold = _threshold;
6518     assert(is_aligned(old_threshold, CardTableModRefBS::card_size),
6519            "_threshold should always be card-aligned");
6520     _threshold = align_up(_finger, CardTableModRefBS::card_size);
6521     MemRegion mr(old_threshold, _threshold);
6522     assert(!mr.is_empty(), "Control point invariant");
6523     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6524     _mut->clear_range(mr);
6525   }
6526 
6527   // Note: the local finger doesn't advance while we drain
6528   // the stack below, but the global finger sure can and will.
6529   HeapWord* volatile* gfa = _task->global_finger_addr();
6530   ParPushOrMarkClosure pushOrMarkClosure(_collector,
6531                                          _span, _bit_map,
6532                                          _work_queue,
6533                                          _overflow_stack,
6534                                          _finger,
6535                                          gfa, this);
6536   bool res = _work_queue->push(obj);   // overflow could occur here
6537   assert(res, "Will hold once we use workqueues");
6538   while (true) {
6539     oop new_oop;
6540     if (!_work_queue->pop_local(new_oop)) {
6541       // We emptied our work_queue; check if there's stuff that can
6542       // be gotten from the overflow stack.
6543       if (CMSConcMarkingTask::get_work_from_overflow_stack(
6544             _overflow_stack, _work_queue)) {
6545         do_yield_check();
6546         continue;
6547       } else {  // done
6548         break;
6549       }
6550     }
6551     // Skip verifying header mark word below because we are
6552     // running concurrent with mutators.
6553     assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6554     // now scan this oop's oops
6555     new_oop->oop_iterate(&pushOrMarkClosure);
6556     do_yield_check();
6557   }
6558   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6559 }
6560 
6561 // Yield in response to a request from VM Thread or
6562 // from mutators.
6563 void ParMarkFromRootsClosure::do_yield_work() {
6564   assert(_task != NULL, "sanity");
6565   _task->yield();
6566 }
6567 
6568 // A variant of the above used for verifying CMS marking work.
6569 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6570                         MemRegion span,
6571                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6572                         CMSMarkStack*  mark_stack):
6573   _collector(collector),
6574   _span(span),
6575   _verification_bm(verification_bm),
6576   _cms_bm(cms_bm),
6577   _mark_stack(mark_stack),
6578   _pam_verify_closure(collector, span, verification_bm, cms_bm,
6579                       mark_stack)
6580 {
6581   assert(_mark_stack->isEmpty(), "stack should be empty");
6582   _finger = _verification_bm->startWord();
6583   assert(_collector->_restart_addr == NULL, "Sanity check");
6584   assert(_span.contains(_finger), "Out of bounds _finger?");
6585 }
6586 
6587 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6588   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6589   assert(_span.contains(addr), "Out of bounds _finger?");
6590   _finger = addr;
6591 }
6592 
6593 // Should revisit to see if this should be restructured for
6594 // greater efficiency.
6595 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6596   // convert offset into a HeapWord*
6597   HeapWord* addr = _verification_bm->startWord() + offset;
6598   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6599          "address out of range");
6600   assert(_verification_bm->isMarked(addr), "tautology");
6601   assert(_cms_bm->isMarked(addr), "tautology");
6602 
6603   assert(_mark_stack->isEmpty(),
6604          "should drain stack to limit stack usage");
6605   // convert addr to an oop preparatory to scanning
6606   oop obj = oop(addr);
6607   assert(oopDesc::is_oop(obj), "should be an oop");
6608   assert(_finger <= addr, "_finger runneth ahead");
6609   // advance the finger to right end of this object
6610   _finger = addr + obj->size();
6611   assert(_finger > addr, "we just incremented it above");
6612   // Note: the finger doesn't advance while we drain
6613   // the stack below.
6614   bool res = _mark_stack->push(obj);
6615   assert(res, "Empty non-zero size stack should have space for single push");
6616   while (!_mark_stack->isEmpty()) {
6617     oop new_oop = _mark_stack->pop();
6618     assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6619     // now scan this oop's oops
6620     new_oop->oop_iterate(&_pam_verify_closure);
6621   }
6622   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6623   return true;
6624 }
6625 
6626 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6627   CMSCollector* collector, MemRegion span,
6628   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6629   CMSMarkStack*  mark_stack):
6630   MetadataAwareOopClosure(collector->ref_processor()),
6631   _collector(collector),
6632   _span(span),
6633   _verification_bm(verification_bm),
6634   _cms_bm(cms_bm),
6635   _mark_stack(mark_stack)
6636 { }
6637 
6638 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
6639 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6640 
6641 // Upon stack overflow, we discard (part of) the stack,
6642 // remembering the least address amongst those discarded
6643 // in CMSCollector's _restart_address.
6644 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6645   // Remember the least grey address discarded
6646   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6647   _collector->lower_restart_addr(ra);
6648   _mark_stack->reset();  // discard stack contents
6649   _mark_stack->expand(); // expand the stack if possible
6650 }
6651 
6652 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6653   assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6654   HeapWord* addr = (HeapWord*)obj;
6655   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6656     // Oop lies in _span and isn't yet grey or black
6657     _verification_bm->mark(addr);            // now grey
6658     if (!_cms_bm->isMarked(addr)) {
6659       Log(gc, verify) log;
6660       ResourceMark rm;
6661       LogStream ls(log.error());
6662       oop(addr)->print_on(&ls);
6663       log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6664       fatal("... aborting");
6665     }
6666 
6667     if (!_mark_stack->push(obj)) { // stack overflow
6668       log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6669       assert(_mark_stack->isFull(), "Else push should have succeeded");
6670       handle_stack_overflow(addr);
6671     }
6672     // anything including and to the right of _finger
6673     // will be scanned as we iterate over the remainder of the
6674     // bit map
6675   }
6676 }
6677 
6678 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6679                      MemRegion span,
6680                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
6681                      HeapWord* finger, MarkFromRootsClosure* parent) :
6682   MetadataAwareOopClosure(collector->ref_processor()),
6683   _collector(collector),
6684   _span(span),
6685   _bitMap(bitMap),
6686   _markStack(markStack),
6687   _finger(finger),
6688   _parent(parent)
6689 { }
6690 
6691 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6692                                            MemRegion span,
6693                                            CMSBitMap* bit_map,
6694                                            OopTaskQueue* work_queue,
6695                                            CMSMarkStack*  overflow_stack,
6696                                            HeapWord* finger,
6697                                            HeapWord* volatile* global_finger_addr,
6698                                            ParMarkFromRootsClosure* parent) :
6699   MetadataAwareOopClosure(collector->ref_processor()),
6700   _collector(collector),
6701   _whole_span(collector->_span),
6702   _span(span),
6703   _bit_map(bit_map),
6704   _work_queue(work_queue),
6705   _overflow_stack(overflow_stack),
6706   _finger(finger),
6707   _global_finger_addr(global_finger_addr),
6708   _parent(parent)
6709 { }
6710 
6711 // Assumes thread-safe access by callers, who are
6712 // responsible for mutual exclusion.
6713 void CMSCollector::lower_restart_addr(HeapWord* low) {
6714   assert(_span.contains(low), "Out of bounds addr");
6715   if (_restart_addr == NULL) {
6716     _restart_addr = low;
6717   } else {
6718     _restart_addr = MIN2(_restart_addr, low);
6719   }
6720 }
6721 
6722 // Upon stack overflow, we discard (part of) the stack,
6723 // remembering the least address amongst those discarded
6724 // in CMSCollector's _restart_address.
6725 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6726   // Remember the least grey address discarded
6727   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6728   _collector->lower_restart_addr(ra);
6729   _markStack->reset();  // discard stack contents
6730   _markStack->expand(); // expand the stack if possible
6731 }
6732 
6733 // Upon stack overflow, we discard (part of) the stack,
6734 // remembering the least address amongst those discarded
6735 // in CMSCollector's _restart_address.
6736 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6737   // We need to do this under a mutex to prevent other
6738   // workers from interfering with the work done below.
6739   MutexLockerEx ml(_overflow_stack->par_lock(),
6740                    Mutex::_no_safepoint_check_flag);
6741   // Remember the least grey address discarded
6742   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6743   _collector->lower_restart_addr(ra);
6744   _overflow_stack->reset();  // discard stack contents
6745   _overflow_stack->expand(); // expand the stack if possible
6746 }
6747 
6748 void PushOrMarkClosure::do_oop(oop obj) {
6749   // Ignore mark word because we are running concurrent with mutators.
6750   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6751   HeapWord* addr = (HeapWord*)obj;
6752   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6753     // Oop lies in _span and isn't yet grey or black
6754     _bitMap->mark(addr);            // now grey
6755     if (addr < _finger) {
6756       // the bit map iteration has already either passed, or
6757       // sampled, this bit in the bit map; we'll need to
6758       // use the marking stack to scan this oop's oops.
6759       bool simulate_overflow = false;
6760       NOT_PRODUCT(
6761         if (CMSMarkStackOverflowALot &&
6762             _collector->simulate_overflow()) {
6763           // simulate a stack overflow
6764           simulate_overflow = true;
6765         }
6766       )
6767       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6768         log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6769         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6770         handle_stack_overflow(addr);
6771       }
6772     }
6773     // anything including and to the right of _finger
6774     // will be scanned as we iterate over the remainder of the
6775     // bit map
6776     do_yield_check();
6777   }
6778 }
6779 
6780 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
6781 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6782 
6783 void ParPushOrMarkClosure::do_oop(oop obj) {
6784   // Ignore mark word because we are running concurrent with mutators.
6785   assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6786   HeapWord* addr = (HeapWord*)obj;
6787   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6788     // Oop lies in _span and isn't yet grey or black
6789     // We read the global_finger (volatile read) strictly after marking oop
6790     bool res = _bit_map->par_mark(addr);    // now grey
6791     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6792     // Should we push this marked oop on our stack?
6793     // -- if someone else marked it, nothing to do
6794     // -- if target oop is above global finger nothing to do
6795     // -- if target oop is in chunk and above local finger
6796     //      then nothing to do
6797     // -- else push on work queue
6798     if (   !res       // someone else marked it, they will deal with it
6799         || (addr >= *gfa)  // will be scanned in a later task
6800         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6801       return;
6802     }
6803     // the bit map iteration has already either passed, or
6804     // sampled, this bit in the bit map; we'll need to
6805     // use the marking stack to scan this oop's oops.
6806     bool simulate_overflow = false;
6807     NOT_PRODUCT(
6808       if (CMSMarkStackOverflowALot &&
6809           _collector->simulate_overflow()) {
6810         // simulate a stack overflow
6811         simulate_overflow = true;
6812       }
6813     )
6814     if (simulate_overflow ||
6815         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6816       // stack overflow
6817       log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6818       // We cannot assert that the overflow stack is full because
6819       // it may have been emptied since.
6820       assert(simulate_overflow ||
6821              _work_queue->size() == _work_queue->max_elems(),
6822             "Else push should have succeeded");
6823       handle_stack_overflow(addr);
6824     }
6825     do_yield_check();
6826   }
6827 }
6828 
6829 void ParPushOrMarkClosure::do_oop(oop* p)       { ParPushOrMarkClosure::do_oop_work(p); }
6830 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6831 
6832 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6833                                        MemRegion span,
6834                                        ReferenceProcessor* rp,
6835                                        CMSBitMap* bit_map,
6836                                        CMSBitMap* mod_union_table,
6837                                        CMSMarkStack*  mark_stack,
6838                                        bool           concurrent_precleaning):
6839   MetadataAwareOopClosure(rp),
6840   _collector(collector),
6841   _span(span),
6842   _bit_map(bit_map),
6843   _mod_union_table(mod_union_table),
6844   _mark_stack(mark_stack),
6845   _concurrent_precleaning(concurrent_precleaning)
6846 {
6847   assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6848 }
6849 
6850 // Grey object rescan during pre-cleaning and second checkpoint phases --
6851 // the non-parallel version (the parallel version appears further below.)
6852 void PushAndMarkClosure::do_oop(oop obj) {
6853   // Ignore mark word verification. If during concurrent precleaning,
6854   // the object monitor may be locked. If during the checkpoint
6855   // phases, the object may already have been reached by a  different
6856   // path and may be at the end of the global overflow list (so
6857   // the mark word may be NULL).
6858   assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6859          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6860   HeapWord* addr = (HeapWord*)obj;
6861   // Check if oop points into the CMS generation
6862   // and is not marked
6863   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6864     // a white object ...
6865     _bit_map->mark(addr);         // ... now grey
6866     // push on the marking stack (grey set)
6867     bool simulate_overflow = false;
6868     NOT_PRODUCT(
6869       if (CMSMarkStackOverflowALot &&
6870           _collector->simulate_overflow()) {
6871         // simulate a stack overflow
6872         simulate_overflow = true;
6873       }
6874     )
6875     if (simulate_overflow || !_mark_stack->push(obj)) {
6876       if (_concurrent_precleaning) {
6877          // During precleaning we can just dirty the appropriate card(s)
6878          // in the mod union table, thus ensuring that the object remains
6879          // in the grey set  and continue. In the case of object arrays
6880          // we need to dirty all of the cards that the object spans,
6881          // since the rescan of object arrays will be limited to the
6882          // dirty cards.
6883          // Note that no one can be interfering with us in this action
6884          // of dirtying the mod union table, so no locking or atomics
6885          // are required.
6886          if (obj->is_objArray()) {
6887            size_t sz = obj->size();
6888            HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
6889            MemRegion redirty_range = MemRegion(addr, end_card_addr);
6890            assert(!redirty_range.is_empty(), "Arithmetical tautology");
6891            _mod_union_table->mark_range(redirty_range);
6892          } else {
6893            _mod_union_table->mark(addr);
6894          }
6895          _collector->_ser_pmc_preclean_ovflw++;
6896       } else {
6897          // During the remark phase, we need to remember this oop
6898          // in the overflow list.
6899          _collector->push_on_overflow_list(obj);
6900          _collector->_ser_pmc_remark_ovflw++;
6901       }
6902     }
6903   }
6904 }
6905 
6906 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6907                                              MemRegion span,
6908                                              ReferenceProcessor* rp,
6909                                              CMSBitMap* bit_map,
6910                                              OopTaskQueue* work_queue):
6911   MetadataAwareOopClosure(rp),
6912   _collector(collector),
6913   _span(span),
6914   _bit_map(bit_map),
6915   _work_queue(work_queue)
6916 {
6917   assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6918 }
6919 
6920 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
6921 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6922 
6923 // Grey object rescan during second checkpoint phase --
6924 // the parallel version.
6925 void ParPushAndMarkClosure::do_oop(oop obj) {
6926   // In the assert below, we ignore the mark word because
6927   // this oop may point to an already visited object that is
6928   // on the overflow stack (in which case the mark word has
6929   // been hijacked for chaining into the overflow stack --
6930   // if this is the last object in the overflow stack then
6931   // its mark word will be NULL). Because this object may
6932   // have been subsequently popped off the global overflow
6933   // stack, and the mark word possibly restored to the prototypical
6934   // value, by the time we get to examined this failing assert in
6935   // the debugger, is_oop_or_null(false) may subsequently start
6936   // to hold.
6937   assert(oopDesc::is_oop_or_null(obj, true),
6938          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6939   HeapWord* addr = (HeapWord*)obj;
6940   // Check if oop points into the CMS generation
6941   // and is not marked
6942   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6943     // a white object ...
6944     // If we manage to "claim" the object, by being the
6945     // first thread to mark it, then we push it on our
6946     // marking stack
6947     if (_bit_map->par_mark(addr)) {     // ... now grey
6948       // push on work queue (grey set)
6949       bool simulate_overflow = false;
6950       NOT_PRODUCT(
6951         if (CMSMarkStackOverflowALot &&
6952             _collector->par_simulate_overflow()) {
6953           // simulate a stack overflow
6954           simulate_overflow = true;
6955         }
6956       )
6957       if (simulate_overflow || !_work_queue->push(obj)) {
6958         _collector->par_push_on_overflow_list(obj);
6959         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
6960       }
6961     } // Else, some other thread got there first
6962   }
6963 }
6964 
6965 void ParPushAndMarkClosure::do_oop(oop* p)       { ParPushAndMarkClosure::do_oop_work(p); }
6966 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6967 
6968 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6969   Mutex* bml = _collector->bitMapLock();
6970   assert_lock_strong(bml);
6971   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6972          "CMS thread should hold CMS token");
6973 
6974   bml->unlock();
6975   ConcurrentMarkSweepThread::desynchronize(true);
6976 
6977   _collector->stopTimer();
6978   _collector->incrementYields();
6979 
6980   // See the comment in coordinator_yield()
6981   for (unsigned i = 0; i < CMSYieldSleepCount &&
6982                        ConcurrentMarkSweepThread::should_yield() &&
6983                        !CMSCollector::foregroundGCIsActive(); ++i) {
6984     os::sleep(Thread::current(), 1, false);
6985   }
6986 
6987   ConcurrentMarkSweepThread::synchronize(true);
6988   bml->lock();
6989 
6990   _collector->startTimer();
6991 }
6992 
6993 bool CMSPrecleanRefsYieldClosure::should_return() {
6994   if (ConcurrentMarkSweepThread::should_yield()) {
6995     do_yield_work();
6996   }
6997   return _collector->foregroundGCIsActive();
6998 }
6999 
7000 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7001   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7002          "mr should be aligned to start at a card boundary");
7003   // We'd like to assert:
7004   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7005   //        "mr should be a range of cards");
7006   // However, that would be too strong in one case -- the last
7007   // partition ends at _unallocated_block which, in general, can be
7008   // an arbitrary boundary, not necessarily card aligned.
7009   _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words;
7010   _space->object_iterate_mem(mr, &_scan_cl);
7011 }
7012 
7013 SweepClosure::SweepClosure(CMSCollector* collector,
7014                            ConcurrentMarkSweepGeneration* g,
7015                            CMSBitMap* bitMap, bool should_yield) :
7016   _collector(collector),
7017   _g(g),
7018   _sp(g->cmsSpace()),
7019   _limit(_sp->sweep_limit()),
7020   _freelistLock(_sp->freelistLock()),
7021   _bitMap(bitMap),
7022   _yield(should_yield),
7023   _inFreeRange(false),           // No free range at beginning of sweep
7024   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
7025   _lastFreeRangeCoalesced(false),
7026   _freeFinger(g->used_region().start())
7027 {
7028   NOT_PRODUCT(
7029     _numObjectsFreed = 0;
7030     _numWordsFreed   = 0;
7031     _numObjectsLive = 0;
7032     _numWordsLive = 0;
7033     _numObjectsAlreadyFree = 0;
7034     _numWordsAlreadyFree = 0;
7035     _last_fc = NULL;
7036 
7037     _sp->initializeIndexedFreeListArrayReturnedBytes();
7038     _sp->dictionary()->initialize_dict_returned_bytes();
7039   )
7040   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7041          "sweep _limit out of bounds");
7042   log_develop_trace(gc, sweep)("====================");
7043   log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7044 }
7045 
7046 void SweepClosure::print_on(outputStream* st) const {
7047   st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7048                p2i(_sp->bottom()), p2i(_sp->end()));
7049   st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7050   st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7051   NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7052   st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7053                _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7054 }
7055 
7056 #ifndef PRODUCT
7057 // Assertion checking only:  no useful work in product mode --
7058 // however, if any of the flags below become product flags,
7059 // you may need to review this code to see if it needs to be
7060 // enabled in product mode.
7061 SweepClosure::~SweepClosure() {
7062   assert_lock_strong(_freelistLock);
7063   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7064          "sweep _limit out of bounds");
7065   if (inFreeRange()) {
7066     Log(gc, sweep) log;
7067     log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7068     ResourceMark rm;
7069     LogStream ls(log.error());
7070     print_on(&ls);
7071     ShouldNotReachHere();
7072   }
7073 
7074   if (log_is_enabled(Debug, gc, sweep)) {
7075     log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7076                          _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7077     log_debug(gc, sweep)("Live " SIZE_FORMAT " objects,  " SIZE_FORMAT " bytes  Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7078                          _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7079     size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7080     log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7081   }
7082 
7083   if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7084     size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7085     size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7086     size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7087     log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes   Indexed List Returned " SIZE_FORMAT " bytes        Dictionary Returned " SIZE_FORMAT " bytes",
7088                          returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7089   }
7090   log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7091   log_develop_trace(gc, sweep)("================");
7092 }
7093 #endif  // PRODUCT
7094 
7095 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7096     bool freeRangeInFreeLists) {
7097   log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7098                                p2i(freeFinger), freeRangeInFreeLists);
7099   assert(!inFreeRange(), "Trampling existing free range");
7100   set_inFreeRange(true);
7101   set_lastFreeRangeCoalesced(false);
7102 
7103   set_freeFinger(freeFinger);
7104   set_freeRangeInFreeLists(freeRangeInFreeLists);
7105   if (CMSTestInFreeList) {
7106     if (freeRangeInFreeLists) {
7107       FreeChunk* fc = (FreeChunk*) freeFinger;
7108       assert(fc->is_free(), "A chunk on the free list should be free.");
7109       assert(fc->size() > 0, "Free range should have a size");
7110       assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7111     }
7112   }
7113 }
7114 
7115 // Note that the sweeper runs concurrently with mutators. Thus,
7116 // it is possible for direct allocation in this generation to happen
7117 // in the middle of the sweep. Note that the sweeper also coalesces
7118 // contiguous free blocks. Thus, unless the sweeper and the allocator
7119 // synchronize appropriately freshly allocated blocks may get swept up.
7120 // This is accomplished by the sweeper locking the free lists while
7121 // it is sweeping. Thus blocks that are determined to be free are
7122 // indeed free. There is however one additional complication:
7123 // blocks that have been allocated since the final checkpoint and
7124 // mark, will not have been marked and so would be treated as
7125 // unreachable and swept up. To prevent this, the allocator marks
7126 // the bit map when allocating during the sweep phase. This leads,
7127 // however, to a further complication -- objects may have been allocated
7128 // but not yet initialized -- in the sense that the header isn't yet
7129 // installed. The sweeper can not then determine the size of the block
7130 // in order to skip over it. To deal with this case, we use a technique
7131 // (due to Printezis) to encode such uninitialized block sizes in the
7132 // bit map. Since the bit map uses a bit per every HeapWord, but the
7133 // CMS generation has a minimum object size of 3 HeapWords, it follows
7134 // that "normal marks" won't be adjacent in the bit map (there will
7135 // always be at least two 0 bits between successive 1 bits). We make use
7136 // of these "unused" bits to represent uninitialized blocks -- the bit
7137 // corresponding to the start of the uninitialized object and the next
7138 // bit are both set. Finally, a 1 bit marks the end of the object that
7139 // started with the two consecutive 1 bits to indicate its potentially
7140 // uninitialized state.
7141 
7142 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7143   FreeChunk* fc = (FreeChunk*)addr;
7144   size_t res;
7145 
7146   // Check if we are done sweeping. Below we check "addr >= _limit" rather
7147   // than "addr == _limit" because although _limit was a block boundary when
7148   // we started the sweep, it may no longer be one because heap expansion
7149   // may have caused us to coalesce the block ending at the address _limit
7150   // with a newly expanded chunk (this happens when _limit was set to the
7151   // previous _end of the space), so we may have stepped past _limit:
7152   // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7153   if (addr >= _limit) { // we have swept up to or past the limit: finish up
7154     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7155            "sweep _limit out of bounds");
7156     assert(addr < _sp->end(), "addr out of bounds");
7157     // Flush any free range we might be holding as a single
7158     // coalesced chunk to the appropriate free list.
7159     if (inFreeRange()) {
7160       assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7161              "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7162       flush_cur_free_chunk(freeFinger(),
7163                            pointer_delta(addr, freeFinger()));
7164       log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7165                                    p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7166                                    lastFreeRangeCoalesced() ? 1 : 0);
7167     }
7168 
7169     // help the iterator loop finish
7170     return pointer_delta(_sp->end(), addr);
7171   }
7172 
7173   assert(addr < _limit, "sweep invariant");
7174   // check if we should yield
7175   do_yield_check(addr);
7176   if (fc->is_free()) {
7177     // Chunk that is already free
7178     res = fc->size();
7179     do_already_free_chunk(fc);
7180     debug_only(_sp->verifyFreeLists());
7181     // If we flush the chunk at hand in lookahead_and_flush()
7182     // and it's coalesced with a preceding chunk, then the
7183     // process of "mangling" the payload of the coalesced block
7184     // will cause erasure of the size information from the
7185     // (erstwhile) header of all the coalesced blocks but the
7186     // first, so the first disjunct in the assert will not hold
7187     // in that specific case (in which case the second disjunct
7188     // will hold).
7189     assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7190            "Otherwise the size info doesn't change at this step");
7191     NOT_PRODUCT(
7192       _numObjectsAlreadyFree++;
7193       _numWordsAlreadyFree += res;
7194     )
7195     NOT_PRODUCT(_last_fc = fc;)
7196   } else if (!_bitMap->isMarked(addr)) {
7197     // Chunk is fresh garbage
7198     res = do_garbage_chunk(fc);
7199     debug_only(_sp->verifyFreeLists());
7200     NOT_PRODUCT(
7201       _numObjectsFreed++;
7202       _numWordsFreed += res;
7203     )
7204   } else {
7205     // Chunk that is alive.
7206     res = do_live_chunk(fc);
7207     debug_only(_sp->verifyFreeLists());
7208     NOT_PRODUCT(
7209         _numObjectsLive++;
7210         _numWordsLive += res;
7211     )
7212   }
7213   return res;
7214 }
7215 
7216 // For the smart allocation, record following
7217 //  split deaths - a free chunk is removed from its free list because
7218 //      it is being split into two or more chunks.
7219 //  split birth - a free chunk is being added to its free list because
7220 //      a larger free chunk has been split and resulted in this free chunk.
7221 //  coal death - a free chunk is being removed from its free list because
7222 //      it is being coalesced into a large free chunk.
7223 //  coal birth - a free chunk is being added to its free list because
7224 //      it was created when two or more free chunks where coalesced into
7225 //      this free chunk.
7226 //
7227 // These statistics are used to determine the desired number of free
7228 // chunks of a given size.  The desired number is chosen to be relative
7229 // to the end of a CMS sweep.  The desired number at the end of a sweep
7230 // is the
7231 //      count-at-end-of-previous-sweep (an amount that was enough)
7232 //              - count-at-beginning-of-current-sweep  (the excess)
7233 //              + split-births  (gains in this size during interval)
7234 //              - split-deaths  (demands on this size during interval)
7235 // where the interval is from the end of one sweep to the end of the
7236 // next.
7237 //
7238 // When sweeping the sweeper maintains an accumulated chunk which is
7239 // the chunk that is made up of chunks that have been coalesced.  That
7240 // will be termed the left-hand chunk.  A new chunk of garbage that
7241 // is being considered for coalescing will be referred to as the
7242 // right-hand chunk.
7243 //
7244 // When making a decision on whether to coalesce a right-hand chunk with
7245 // the current left-hand chunk, the current count vs. the desired count
7246 // of the left-hand chunk is considered.  Also if the right-hand chunk
7247 // is near the large chunk at the end of the heap (see
7248 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7249 // left-hand chunk is coalesced.
7250 //
7251 // When making a decision about whether to split a chunk, the desired count
7252 // vs. the current count of the candidate to be split is also considered.
7253 // If the candidate is underpopulated (currently fewer chunks than desired)
7254 // a chunk of an overpopulated (currently more chunks than desired) size may
7255 // be chosen.  The "hint" associated with a free list, if non-null, points
7256 // to a free list which may be overpopulated.
7257 //
7258 
7259 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7260   const size_t size = fc->size();
7261   // Chunks that cannot be coalesced are not in the
7262   // free lists.
7263   if (CMSTestInFreeList && !fc->cantCoalesce()) {
7264     assert(_sp->verify_chunk_in_free_list(fc),
7265            "free chunk should be in free lists");
7266   }
7267   // a chunk that is already free, should not have been
7268   // marked in the bit map
7269   HeapWord* const addr = (HeapWord*) fc;
7270   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7271   // Verify that the bit map has no bits marked between
7272   // addr and purported end of this block.
7273   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7274 
7275   // Some chunks cannot be coalesced under any circumstances.
7276   // See the definition of cantCoalesce().
7277   if (!fc->cantCoalesce()) {
7278     // This chunk can potentially be coalesced.
7279     // All the work is done in
7280     do_post_free_or_garbage_chunk(fc, size);
7281     // Note that if the chunk is not coalescable (the else arm
7282     // below), we unconditionally flush, without needing to do
7283     // a "lookahead," as we do below.
7284     if (inFreeRange()) lookahead_and_flush(fc, size);
7285   } else {
7286     // Code path common to both original and adaptive free lists.
7287 
7288     // cant coalesce with previous block; this should be treated
7289     // as the end of a free run if any
7290     if (inFreeRange()) {
7291       // we kicked some butt; time to pick up the garbage
7292       assert(freeFinger() < addr, "freeFinger points too high");
7293       flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7294     }
7295     // else, nothing to do, just continue
7296   }
7297 }
7298 
7299 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7300   // This is a chunk of garbage.  It is not in any free list.
7301   // Add it to a free list or let it possibly be coalesced into
7302   // a larger chunk.
7303   HeapWord* const addr = (HeapWord*) fc;
7304   const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7305 
7306   // Verify that the bit map has no bits marked between
7307   // addr and purported end of just dead object.
7308   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7309   do_post_free_or_garbage_chunk(fc, size);
7310 
7311   assert(_limit >= addr + size,
7312          "A freshly garbage chunk can't possibly straddle over _limit");
7313   if (inFreeRange()) lookahead_and_flush(fc, size);
7314   return size;
7315 }
7316 
7317 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7318   HeapWord* addr = (HeapWord*) fc;
7319   // The sweeper has just found a live object. Return any accumulated
7320   // left hand chunk to the free lists.
7321   if (inFreeRange()) {
7322     assert(freeFinger() < addr, "freeFinger points too high");
7323     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7324   }
7325 
7326   // This object is live: we'd normally expect this to be
7327   // an oop, and like to assert the following:
7328   // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7329   // However, as we commented above, this may be an object whose
7330   // header hasn't yet been initialized.
7331   size_t size;
7332   assert(_bitMap->isMarked(addr), "Tautology for this control point");
7333   if (_bitMap->isMarked(addr + 1)) {
7334     // Determine the size from the bit map, rather than trying to
7335     // compute it from the object header.
7336     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7337     size = pointer_delta(nextOneAddr + 1, addr);
7338     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7339            "alignment problem");
7340 
7341 #ifdef ASSERT
7342       if (oop(addr)->klass_or_null_acquire() != NULL) {
7343         // Ignore mark word because we are running concurrent with mutators
7344         assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7345         assert(size ==
7346                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7347                "P-mark and computed size do not agree");
7348       }
7349 #endif
7350 
7351   } else {
7352     // This should be an initialized object that's alive.
7353     assert(oop(addr)->klass_or_null_acquire() != NULL,
7354            "Should be an initialized object");
7355     // Ignore mark word because we are running concurrent with mutators
7356     assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7357     // Verify that the bit map has no bits marked between
7358     // addr and purported end of this block.
7359     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7360     assert(size >= 3, "Necessary for Printezis marks to work");
7361     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7362     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7363   }
7364   return size;
7365 }
7366 
7367 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7368                                                  size_t chunkSize) {
7369   // do_post_free_or_garbage_chunk() should only be called in the case
7370   // of the adaptive free list allocator.
7371   const bool fcInFreeLists = fc->is_free();
7372   assert((HeapWord*)fc <= _limit, "sweep invariant");
7373   if (CMSTestInFreeList && fcInFreeLists) {
7374     assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7375   }
7376 
7377   log_develop_trace(gc, sweep)("  -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7378 
7379   HeapWord* const fc_addr = (HeapWord*) fc;
7380 
7381   bool coalesce = false;
7382   const size_t left  = pointer_delta(fc_addr, freeFinger());
7383   const size_t right = chunkSize;
7384   switch (FLSCoalescePolicy) {
7385     // numeric value forms a coalition aggressiveness metric
7386     case 0:  { // never coalesce
7387       coalesce = false;
7388       break;
7389     }
7390     case 1: { // coalesce if left & right chunks on overpopulated lists
7391       coalesce = _sp->coalOverPopulated(left) &&
7392                  _sp->coalOverPopulated(right);
7393       break;
7394     }
7395     case 2: { // coalesce if left chunk on overpopulated list (default)
7396       coalesce = _sp->coalOverPopulated(left);
7397       break;
7398     }
7399     case 3: { // coalesce if left OR right chunk on overpopulated list
7400       coalesce = _sp->coalOverPopulated(left) ||
7401                  _sp->coalOverPopulated(right);
7402       break;
7403     }
7404     case 4: { // always coalesce
7405       coalesce = true;
7406       break;
7407     }
7408     default:
7409      ShouldNotReachHere();
7410   }
7411 
7412   // Should the current free range be coalesced?
7413   // If the chunk is in a free range and either we decided to coalesce above
7414   // or the chunk is near the large block at the end of the heap
7415   // (isNearLargestChunk() returns true), then coalesce this chunk.
7416   const bool doCoalesce = inFreeRange()
7417                           && (coalesce || _g->isNearLargestChunk(fc_addr));
7418   if (doCoalesce) {
7419     // Coalesce the current free range on the left with the new
7420     // chunk on the right.  If either is on a free list,
7421     // it must be removed from the list and stashed in the closure.
7422     if (freeRangeInFreeLists()) {
7423       FreeChunk* const ffc = (FreeChunk*)freeFinger();
7424       assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7425              "Size of free range is inconsistent with chunk size.");
7426       if (CMSTestInFreeList) {
7427         assert(_sp->verify_chunk_in_free_list(ffc),
7428                "Chunk is not in free lists");
7429       }
7430       _sp->coalDeath(ffc->size());
7431       _sp->removeFreeChunkFromFreeLists(ffc);
7432       set_freeRangeInFreeLists(false);
7433     }
7434     if (fcInFreeLists) {
7435       _sp->coalDeath(chunkSize);
7436       assert(fc->size() == chunkSize,
7437         "The chunk has the wrong size or is not in the free lists");
7438       _sp->removeFreeChunkFromFreeLists(fc);
7439     }
7440     set_lastFreeRangeCoalesced(true);
7441     print_free_block_coalesced(fc);
7442   } else {  // not in a free range and/or should not coalesce
7443     // Return the current free range and start a new one.
7444     if (inFreeRange()) {
7445       // In a free range but cannot coalesce with the right hand chunk.
7446       // Put the current free range into the free lists.
7447       flush_cur_free_chunk(freeFinger(),
7448                            pointer_delta(fc_addr, freeFinger()));
7449     }
7450     // Set up for new free range.  Pass along whether the right hand
7451     // chunk is in the free lists.
7452     initialize_free_range((HeapWord*)fc, fcInFreeLists);
7453   }
7454 }
7455 
7456 // Lookahead flush:
7457 // If we are tracking a free range, and this is the last chunk that
7458 // we'll look at because its end crosses past _limit, we'll preemptively
7459 // flush it along with any free range we may be holding on to. Note that
7460 // this can be the case only for an already free or freshly garbage
7461 // chunk. If this block is an object, it can never straddle
7462 // over _limit. The "straddling" occurs when _limit is set at
7463 // the previous end of the space when this cycle started, and
7464 // a subsequent heap expansion caused the previously co-terminal
7465 // free block to be coalesced with the newly expanded portion,
7466 // thus rendering _limit a non-block-boundary making it dangerous
7467 // for the sweeper to step over and examine.
7468 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7469   assert(inFreeRange(), "Should only be called if currently in a free range.");
7470   HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7471   assert(_sp->used_region().contains(eob - 1),
7472          "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7473          " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7474          " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7475          p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7476   if (eob >= _limit) {
7477     assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7478     log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7479                                  "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7480                                  "[" PTR_FORMAT "," PTR_FORMAT ")",
7481                                  p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7482     // Return the storage we are tracking back into the free lists.
7483     log_develop_trace(gc, sweep)("Flushing ... ");
7484     assert(freeFinger() < eob, "Error");
7485     flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7486   }
7487 }
7488 
7489 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7490   assert(inFreeRange(), "Should only be called if currently in a free range.");
7491   assert(size > 0,
7492     "A zero sized chunk cannot be added to the free lists.");
7493   if (!freeRangeInFreeLists()) {
7494     if (CMSTestInFreeList) {
7495       FreeChunk* fc = (FreeChunk*) chunk;
7496       fc->set_size(size);
7497       assert(!_sp->verify_chunk_in_free_list(fc),
7498              "chunk should not be in free lists yet");
7499     }
7500     log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7501     // A new free range is going to be starting.  The current
7502     // free range has not been added to the free lists yet or
7503     // was removed so add it back.
7504     // If the current free range was coalesced, then the death
7505     // of the free range was recorded.  Record a birth now.
7506     if (lastFreeRangeCoalesced()) {
7507       _sp->coalBirth(size);
7508     }
7509     _sp->addChunkAndRepairOffsetTable(chunk, size,
7510             lastFreeRangeCoalesced());
7511   } else {
7512     log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7513   }
7514   set_inFreeRange(false);
7515   set_freeRangeInFreeLists(false);
7516 }
7517 
7518 // We take a break if we've been at this for a while,
7519 // so as to avoid monopolizing the locks involved.
7520 void SweepClosure::do_yield_work(HeapWord* addr) {
7521   // Return current free chunk being used for coalescing (if any)
7522   // to the appropriate freelist.  After yielding, the next
7523   // free block encountered will start a coalescing range of
7524   // free blocks.  If the next free block is adjacent to the
7525   // chunk just flushed, they will need to wait for the next
7526   // sweep to be coalesced.
7527   if (inFreeRange()) {
7528     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7529   }
7530 
7531   // First give up the locks, then yield, then re-lock.
7532   // We should probably use a constructor/destructor idiom to
7533   // do this unlock/lock or modify the MutexUnlocker class to
7534   // serve our purpose. XXX
7535   assert_lock_strong(_bitMap->lock());
7536   assert_lock_strong(_freelistLock);
7537   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7538          "CMS thread should hold CMS token");
7539   _bitMap->lock()->unlock();
7540   _freelistLock->unlock();
7541   ConcurrentMarkSweepThread::desynchronize(true);
7542   _collector->stopTimer();
7543   _collector->incrementYields();
7544 
7545   // See the comment in coordinator_yield()
7546   for (unsigned i = 0; i < CMSYieldSleepCount &&
7547                        ConcurrentMarkSweepThread::should_yield() &&
7548                        !CMSCollector::foregroundGCIsActive(); ++i) {
7549     os::sleep(Thread::current(), 1, false);
7550   }
7551 
7552   ConcurrentMarkSweepThread::synchronize(true);
7553   _freelistLock->lock();
7554   _bitMap->lock()->lock_without_safepoint_check();
7555   _collector->startTimer();
7556 }
7557 
7558 #ifndef PRODUCT
7559 // This is actually very useful in a product build if it can
7560 // be called from the debugger.  Compile it into the product
7561 // as needed.
7562 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7563   return debug_cms_space->verify_chunk_in_free_list(fc);
7564 }
7565 #endif
7566 
7567 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7568   log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7569                                p2i(fc), fc->size());
7570 }
7571 
7572 // CMSIsAliveClosure
7573 bool CMSIsAliveClosure::do_object_b(oop obj) {
7574   HeapWord* addr = (HeapWord*)obj;
7575   return addr != NULL &&
7576          (!_span.contains(addr) || _bit_map->isMarked(addr));
7577 }
7578 
7579 
7580 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7581                       MemRegion span,
7582                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7583                       bool cpc):
7584   _collector(collector),
7585   _span(span),
7586   _bit_map(bit_map),
7587   _mark_stack(mark_stack),
7588   _concurrent_precleaning(cpc) {
7589   assert(!_span.is_empty(), "Empty span could spell trouble");
7590 }
7591 
7592 
7593 // CMSKeepAliveClosure: the serial version
7594 void CMSKeepAliveClosure::do_oop(oop obj) {
7595   HeapWord* addr = (HeapWord*)obj;
7596   if (_span.contains(addr) &&
7597       !_bit_map->isMarked(addr)) {
7598     _bit_map->mark(addr);
7599     bool simulate_overflow = false;
7600     NOT_PRODUCT(
7601       if (CMSMarkStackOverflowALot &&
7602           _collector->simulate_overflow()) {
7603         // simulate a stack overflow
7604         simulate_overflow = true;
7605       }
7606     )
7607     if (simulate_overflow || !_mark_stack->push(obj)) {
7608       if (_concurrent_precleaning) {
7609         // We dirty the overflown object and let the remark
7610         // phase deal with it.
7611         assert(_collector->overflow_list_is_empty(), "Error");
7612         // In the case of object arrays, we need to dirty all of
7613         // the cards that the object spans. No locking or atomics
7614         // are needed since no one else can be mutating the mod union
7615         // table.
7616         if (obj->is_objArray()) {
7617           size_t sz = obj->size();
7618           HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size);
7619           MemRegion redirty_range = MemRegion(addr, end_card_addr);
7620           assert(!redirty_range.is_empty(), "Arithmetical tautology");
7621           _collector->_modUnionTable.mark_range(redirty_range);
7622         } else {
7623           _collector->_modUnionTable.mark(addr);
7624         }
7625         _collector->_ser_kac_preclean_ovflw++;
7626       } else {
7627         _collector->push_on_overflow_list(obj);
7628         _collector->_ser_kac_ovflw++;
7629       }
7630     }
7631   }
7632 }
7633 
7634 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
7635 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7636 
7637 // CMSParKeepAliveClosure: a parallel version of the above.
7638 // The work queues are private to each closure (thread),
7639 // but (may be) available for stealing by other threads.
7640 void CMSParKeepAliveClosure::do_oop(oop obj) {
7641   HeapWord* addr = (HeapWord*)obj;
7642   if (_span.contains(addr) &&
7643       !_bit_map->isMarked(addr)) {
7644     // In general, during recursive tracing, several threads
7645     // may be concurrently getting here; the first one to
7646     // "tag" it, claims it.
7647     if (_bit_map->par_mark(addr)) {
7648       bool res = _work_queue->push(obj);
7649       assert(res, "Low water mark should be much less than capacity");
7650       // Do a recursive trim in the hope that this will keep
7651       // stack usage lower, but leave some oops for potential stealers
7652       trim_queue(_low_water_mark);
7653     } // Else, another thread got there first
7654   }
7655 }
7656 
7657 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
7658 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7659 
7660 void CMSParKeepAliveClosure::trim_queue(uint max) {
7661   while (_work_queue->size() > max) {
7662     oop new_oop;
7663     if (_work_queue->pop_local(new_oop)) {
7664       assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7665       assert(_bit_map->isMarked((HeapWord*)new_oop),
7666              "no white objects on this stack!");
7667       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7668       // iterate over the oops in this oop, marking and pushing
7669       // the ones in CMS heap (i.e. in _span).
7670       new_oop->oop_iterate(&_mark_and_push);
7671     }
7672   }
7673 }
7674 
7675 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7676                                 CMSCollector* collector,
7677                                 MemRegion span, CMSBitMap* bit_map,
7678                                 OopTaskQueue* work_queue):
7679   _collector(collector),
7680   _span(span),
7681   _bit_map(bit_map),
7682   _work_queue(work_queue) { }
7683 
7684 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7685   HeapWord* addr = (HeapWord*)obj;
7686   if (_span.contains(addr) &&
7687       !_bit_map->isMarked(addr)) {
7688     if (_bit_map->par_mark(addr)) {
7689       bool simulate_overflow = false;
7690       NOT_PRODUCT(
7691         if (CMSMarkStackOverflowALot &&
7692             _collector->par_simulate_overflow()) {
7693           // simulate a stack overflow
7694           simulate_overflow = true;
7695         }
7696       )
7697       if (simulate_overflow || !_work_queue->push(obj)) {
7698         _collector->par_push_on_overflow_list(obj);
7699         _collector->_par_kac_ovflw++;
7700       }
7701     } // Else another thread got there already
7702   }
7703 }
7704 
7705 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7706 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7707 
7708 //////////////////////////////////////////////////////////////////
7709 //  CMSExpansionCause                /////////////////////////////
7710 //////////////////////////////////////////////////////////////////
7711 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7712   switch (cause) {
7713     case _no_expansion:
7714       return "No expansion";
7715     case _satisfy_free_ratio:
7716       return "Free ratio";
7717     case _satisfy_promotion:
7718       return "Satisfy promotion";
7719     case _satisfy_allocation:
7720       return "allocation";
7721     case _allocate_par_lab:
7722       return "Par LAB";
7723     case _allocate_par_spooling_space:
7724       return "Par Spooling Space";
7725     case _adaptive_size_policy:
7726       return "Ergonomics";
7727     default:
7728       return "unknown";
7729   }
7730 }
7731 
7732 void CMSDrainMarkingStackClosure::do_void() {
7733   // the max number to take from overflow list at a time
7734   const size_t num = _mark_stack->capacity()/4;
7735   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7736          "Overflow list should be NULL during concurrent phases");
7737   while (!_mark_stack->isEmpty() ||
7738          // if stack is empty, check the overflow list
7739          _collector->take_from_overflow_list(num, _mark_stack)) {
7740     oop obj = _mark_stack->pop();
7741     HeapWord* addr = (HeapWord*)obj;
7742     assert(_span.contains(addr), "Should be within span");
7743     assert(_bit_map->isMarked(addr), "Should be marked");
7744     assert(oopDesc::is_oop(obj), "Should be an oop");
7745     obj->oop_iterate(_keep_alive);
7746   }
7747 }
7748 
7749 void CMSParDrainMarkingStackClosure::do_void() {
7750   // drain queue
7751   trim_queue(0);
7752 }
7753 
7754 // Trim our work_queue so its length is below max at return
7755 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7756   while (_work_queue->size() > max) {
7757     oop new_oop;
7758     if (_work_queue->pop_local(new_oop)) {
7759       assert(oopDesc::is_oop(new_oop), "Expected an oop");
7760       assert(_bit_map->isMarked((HeapWord*)new_oop),
7761              "no white objects on this stack!");
7762       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7763       // iterate over the oops in this oop, marking and pushing
7764       // the ones in CMS heap (i.e. in _span).
7765       new_oop->oop_iterate(&_mark_and_push);
7766     }
7767   }
7768 }
7769 
7770 ////////////////////////////////////////////////////////////////////
7771 // Support for Marking Stack Overflow list handling and related code
7772 ////////////////////////////////////////////////////////////////////
7773 // Much of the following code is similar in shape and spirit to the
7774 // code used in ParNewGC. We should try and share that code
7775 // as much as possible in the future.
7776 
7777 #ifndef PRODUCT
7778 // Debugging support for CMSStackOverflowALot
7779 
7780 // It's OK to call this multi-threaded;  the worst thing
7781 // that can happen is that we'll get a bunch of closely
7782 // spaced simulated overflows, but that's OK, in fact
7783 // probably good as it would exercise the overflow code
7784 // under contention.
7785 bool CMSCollector::simulate_overflow() {
7786   if (_overflow_counter-- <= 0) { // just being defensive
7787     _overflow_counter = CMSMarkStackOverflowInterval;
7788     return true;
7789   } else {
7790     return false;
7791   }
7792 }
7793 
7794 bool CMSCollector::par_simulate_overflow() {
7795   return simulate_overflow();
7796 }
7797 #endif
7798 
7799 // Single-threaded
7800 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7801   assert(stack->isEmpty(), "Expected precondition");
7802   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7803   size_t i = num;
7804   oop  cur = _overflow_list;
7805   const markOop proto = markOopDesc::prototype();
7806   NOT_PRODUCT(ssize_t n = 0;)
7807   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7808     next = oop(cur->mark());
7809     cur->set_mark(proto);   // until proven otherwise
7810     assert(oopDesc::is_oop(cur), "Should be an oop");
7811     bool res = stack->push(cur);
7812     assert(res, "Bit off more than can chew?");
7813     NOT_PRODUCT(n++;)
7814   }
7815   _overflow_list = cur;
7816 #ifndef PRODUCT
7817   assert(_num_par_pushes >= n, "Too many pops?");
7818   _num_par_pushes -=n;
7819 #endif
7820   return !stack->isEmpty();
7821 }
7822 
7823 #define BUSY  (cast_to_oop<intptr_t>(0x1aff1aff))
7824 // (MT-safe) Get a prefix of at most "num" from the list.
7825 // The overflow list is chained through the mark word of
7826 // each object in the list. We fetch the entire list,
7827 // break off a prefix of the right size and return the
7828 // remainder. If other threads try to take objects from
7829 // the overflow list at that time, they will wait for
7830 // some time to see if data becomes available. If (and
7831 // only if) another thread places one or more object(s)
7832 // on the global list before we have returned the suffix
7833 // to the global list, we will walk down our local list
7834 // to find its end and append the global list to
7835 // our suffix before returning it. This suffix walk can
7836 // prove to be expensive (quadratic in the amount of traffic)
7837 // when there are many objects in the overflow list and
7838 // there is much producer-consumer contention on the list.
7839 // *NOTE*: The overflow list manipulation code here and
7840 // in ParNewGeneration:: are very similar in shape,
7841 // except that in the ParNew case we use the old (from/eden)
7842 // copy of the object to thread the list via its klass word.
7843 // Because of the common code, if you make any changes in
7844 // the code below, please check the ParNew version to see if
7845 // similar changes might be needed.
7846 // CR 6797058 has been filed to consolidate the common code.
7847 bool CMSCollector::par_take_from_overflow_list(size_t num,
7848                                                OopTaskQueue* work_q,
7849                                                int no_of_gc_threads) {
7850   assert(work_q->size() == 0, "First empty local work queue");
7851   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7852   if (_overflow_list == NULL) {
7853     return false;
7854   }
7855   // Grab the entire list; we'll put back a suffix
7856   oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
7857   Thread* tid = Thread::current();
7858   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7859   // set to ParallelGCThreads.
7860   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7861   size_t sleep_time_millis = MAX2((size_t)1, num/100);
7862   // If the list is busy, we spin for a short while,
7863   // sleeping between attempts to get the list.
7864   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7865     os::sleep(tid, sleep_time_millis, false);
7866     if (_overflow_list == NULL) {
7867       // Nothing left to take
7868       return false;
7869     } else if (_overflow_list != BUSY) {
7870       // Try and grab the prefix
7871       prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
7872     }
7873   }
7874   // If the list was found to be empty, or we spun long
7875   // enough, we give up and return empty-handed. If we leave
7876   // the list in the BUSY state below, it must be the case that
7877   // some other thread holds the overflow list and will set it
7878   // to a non-BUSY state in the future.
7879   if (prefix == NULL || prefix == BUSY) {
7880      // Nothing to take or waited long enough
7881      if (prefix == NULL) {
7882        // Write back the NULL in case we overwrote it with BUSY above
7883        // and it is still the same value.
7884        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
7885      }
7886      return false;
7887   }
7888   assert(prefix != NULL && prefix != BUSY, "Error");
7889   size_t i = num;
7890   oop cur = prefix;
7891   // Walk down the first "num" objects, unless we reach the end.
7892   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
7893   if (cur->mark() == NULL) {
7894     // We have "num" or fewer elements in the list, so there
7895     // is nothing to return to the global list.
7896     // Write back the NULL in lieu of the BUSY we wrote
7897     // above, if it is still the same value.
7898     if (_overflow_list == BUSY) {
7899       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
7900     }
7901   } else {
7902     // Chop off the suffix and return it to the global list.
7903     assert(cur->mark() != BUSY, "Error");
7904     oop suffix_head = cur->mark(); // suffix will be put back on global list
7905     cur->set_mark(NULL);           // break off suffix
7906     // It's possible that the list is still in the empty(busy) state
7907     // we left it in a short while ago; in that case we may be
7908     // able to place back the suffix without incurring the cost
7909     // of a walk down the list.
7910     oop observed_overflow_list = _overflow_list;
7911     oop cur_overflow_list = observed_overflow_list;
7912     bool attached = false;
7913     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7914       observed_overflow_list =
7915         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
7916       if (cur_overflow_list == observed_overflow_list) {
7917         attached = true;
7918         break;
7919       } else cur_overflow_list = observed_overflow_list;
7920     }
7921     if (!attached) {
7922       // Too bad, someone else sneaked in (at least) an element; we'll need
7923       // to do a splice. Find tail of suffix so we can prepend suffix to global
7924       // list.
7925       for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
7926       oop suffix_tail = cur;
7927       assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
7928              "Tautology");
7929       observed_overflow_list = _overflow_list;
7930       do {
7931         cur_overflow_list = observed_overflow_list;
7932         if (cur_overflow_list != BUSY) {
7933           // Do the splice ...
7934           suffix_tail->set_mark(markOop(cur_overflow_list));
7935         } else { // cur_overflow_list == BUSY
7936           suffix_tail->set_mark(NULL);
7937         }
7938         // ... and try to place spliced list back on overflow_list ...
7939         observed_overflow_list =
7940           (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
7941       } while (cur_overflow_list != observed_overflow_list);
7942       // ... until we have succeeded in doing so.
7943     }
7944   }
7945 
7946   // Push the prefix elements on work_q
7947   assert(prefix != NULL, "control point invariant");
7948   const markOop proto = markOopDesc::prototype();
7949   oop next;
7950   NOT_PRODUCT(ssize_t n = 0;)
7951   for (cur = prefix; cur != NULL; cur = next) {
7952     next = oop(cur->mark());
7953     cur->set_mark(proto);   // until proven otherwise
7954     assert(oopDesc::is_oop(cur), "Should be an oop");
7955     bool res = work_q->push(cur);
7956     assert(res, "Bit off more than we can chew?");
7957     NOT_PRODUCT(n++;)
7958   }
7959 #ifndef PRODUCT
7960   assert(_num_par_pushes >= n, "Too many pops?");
7961   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
7962 #endif
7963   return true;
7964 }
7965 
7966 // Single-threaded
7967 void CMSCollector::push_on_overflow_list(oop p) {
7968   NOT_PRODUCT(_num_par_pushes++;)
7969   assert(oopDesc::is_oop(p), "Not an oop");
7970   preserve_mark_if_necessary(p);
7971   p->set_mark((markOop)_overflow_list);
7972   _overflow_list = p;
7973 }
7974 
7975 // Multi-threaded; use CAS to prepend to overflow list
7976 void CMSCollector::par_push_on_overflow_list(oop p) {
7977   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
7978   assert(oopDesc::is_oop(p), "Not an oop");
7979   par_preserve_mark_if_necessary(p);
7980   oop observed_overflow_list = _overflow_list;
7981   oop cur_overflow_list;
7982   do {
7983     cur_overflow_list = observed_overflow_list;
7984     if (cur_overflow_list != BUSY) {
7985       p->set_mark(markOop(cur_overflow_list));
7986     } else {
7987       p->set_mark(NULL);
7988     }
7989     observed_overflow_list =
7990       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
7991   } while (cur_overflow_list != observed_overflow_list);
7992 }
7993 #undef BUSY
7994 
7995 // Single threaded
7996 // General Note on GrowableArray: pushes may silently fail
7997 // because we are (temporarily) out of C-heap for expanding
7998 // the stack. The problem is quite ubiquitous and affects
7999 // a lot of code in the JVM. The prudent thing for GrowableArray
8000 // to do (for now) is to exit with an error. However, that may
8001 // be too draconian in some cases because the caller may be
8002 // able to recover without much harm. For such cases, we
8003 // should probably introduce a "soft_push" method which returns
8004 // an indication of success or failure with the assumption that
8005 // the caller may be able to recover from a failure; code in
8006 // the VM can then be changed, incrementally, to deal with such
8007 // failures where possible, thus, incrementally hardening the VM
8008 // in such low resource situations.
8009 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8010   _preserved_oop_stack.push(p);
8011   _preserved_mark_stack.push(m);
8012   assert(m == p->mark(), "Mark word changed");
8013   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8014          "bijection");
8015 }
8016 
8017 // Single threaded
8018 void CMSCollector::preserve_mark_if_necessary(oop p) {
8019   markOop m = p->mark();
8020   if (m->must_be_preserved(p)) {
8021     preserve_mark_work(p, m);
8022   }
8023 }
8024 
8025 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8026   markOop m = p->mark();
8027   if (m->must_be_preserved(p)) {
8028     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8029     // Even though we read the mark word without holding
8030     // the lock, we are assured that it will not change
8031     // because we "own" this oop, so no other thread can
8032     // be trying to push it on the overflow list; see
8033     // the assertion in preserve_mark_work() that checks
8034     // that m == p->mark().
8035     preserve_mark_work(p, m);
8036   }
8037 }
8038 
8039 // We should be able to do this multi-threaded,
8040 // a chunk of stack being a task (this is
8041 // correct because each oop only ever appears
8042 // once in the overflow list. However, it's
8043 // not very easy to completely overlap this with
8044 // other operations, so will generally not be done
8045 // until all work's been completed. Because we
8046 // expect the preserved oop stack (set) to be small,
8047 // it's probably fine to do this single-threaded.
8048 // We can explore cleverer concurrent/overlapped/parallel
8049 // processing of preserved marks if we feel the
8050 // need for this in the future. Stack overflow should
8051 // be so rare in practice and, when it happens, its
8052 // effect on performance so great that this will
8053 // likely just be in the noise anyway.
8054 void CMSCollector::restore_preserved_marks_if_any() {
8055   assert(SafepointSynchronize::is_at_safepoint(),
8056          "world should be stopped");
8057   assert(Thread::current()->is_ConcurrentGC_thread() ||
8058          Thread::current()->is_VM_thread(),
8059          "should be single-threaded");
8060   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8061          "bijection");
8062 
8063   while (!_preserved_oop_stack.is_empty()) {
8064     oop p = _preserved_oop_stack.pop();
8065     assert(oopDesc::is_oop(p), "Should be an oop");
8066     assert(_span.contains(p), "oop should be in _span");
8067     assert(p->mark() == markOopDesc::prototype(),
8068            "Set when taken from overflow list");
8069     markOop m = _preserved_mark_stack.pop();
8070     p->set_mark(m);
8071   }
8072   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8073          "stacks were cleared above");
8074 }
8075 
8076 #ifndef PRODUCT
8077 bool CMSCollector::no_preserved_marks() const {
8078   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8079 }
8080 #endif
8081 
8082 // Transfer some number of overflown objects to usual marking
8083 // stack. Return true if some objects were transferred.
8084 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8085   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8086                     (size_t)ParGCDesiredObjsFromOverflowList);
8087 
8088   bool res = _collector->take_from_overflow_list(num, _mark_stack);
8089   assert(_collector->overflow_list_is_empty() || res,
8090          "If list is not empty, we should have taken something");
8091   assert(!res || !_mark_stack->isEmpty(),
8092          "If we took something, it should now be on our stack");
8093   return res;
8094 }
8095 
8096 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8097   size_t res = _sp->block_size_no_stall(addr, _collector);
8098   if (_sp->block_is_obj(addr)) {
8099     if (_live_bit_map->isMarked(addr)) {
8100       // It can't have been dead in a previous cycle
8101       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8102     } else {
8103       _dead_bit_map->mark(addr);      // mark the dead object
8104     }
8105   }
8106   // Could be 0, if the block size could not be computed without stalling.
8107   return res;
8108 }
8109 
8110 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8111 
8112   switch (phase) {
8113     case CMSCollector::InitialMarking:
8114       initialize(true  /* fullGC */ ,
8115                  cause /* cause of the GC */,
8116                  true  /* recordGCBeginTime */,
8117                  true  /* recordPreGCUsage */,
8118                  false /* recordPeakUsage */,
8119                  false /* recordPostGCusage */,
8120                  true  /* recordAccumulatedGCTime */,
8121                  false /* recordGCEndTime */,
8122                  false /* countCollection */  );
8123       break;
8124 
8125     case CMSCollector::FinalMarking:
8126       initialize(true  /* fullGC */ ,
8127                  cause /* cause of the GC */,
8128                  false /* recordGCBeginTime */,
8129                  false /* recordPreGCUsage */,
8130                  false /* recordPeakUsage */,
8131                  false /* recordPostGCusage */,
8132                  true  /* recordAccumulatedGCTime */,
8133                  false /* recordGCEndTime */,
8134                  false /* countCollection */  );
8135       break;
8136 
8137     case CMSCollector::Sweeping:
8138       initialize(true  /* fullGC */ ,
8139                  cause /* cause of the GC */,
8140                  false /* recordGCBeginTime */,
8141                  false /* recordPreGCUsage */,
8142                  true  /* recordPeakUsage */,
8143                  true  /* recordPostGCusage */,
8144                  false /* recordAccumulatedGCTime */,
8145                  true  /* recordGCEndTime */,
8146                  true  /* countCollection */  );
8147       break;
8148 
8149     default:
8150       ShouldNotReachHere();
8151   }
8152 }