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