1 /*
   2  * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/classLoaderData.hpp"
  27 #include "classfile/stringTable.hpp"
  28 #include "classfile/systemDictionary.hpp"
  29 #include "code/codeCache.hpp"
  30 #include "gc/cms/cmsCollectorPolicy.hpp"
  31 #include "gc/cms/cmsOopClosures.inline.hpp"
  32 #include "gc/cms/compactibleFreeListSpace.hpp"
  33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
  34 #include "gc/cms/concurrentMarkSweepThread.hpp"
  35 #include "gc/cms/parNewGeneration.hpp"
  36 #include "gc/cms/vmCMSOperations.hpp"
  37 #include "gc/serial/genMarkSweep.hpp"
  38 #include "gc/serial/tenuredGeneration.hpp"
  39 #include "gc/shared/adaptiveSizePolicy.hpp"
  40 #include "gc/shared/cardGeneration.inline.hpp"
  41 #include "gc/shared/cardTableRS.hpp"
  42 #include "gc/shared/collectedHeap.inline.hpp"
  43 #include "gc/shared/collectorCounters.hpp"
  44 #include "gc/shared/collectorPolicy.hpp"
  45 #include "gc/shared/gcLocker.inline.hpp"
  46 #include "gc/shared/gcPolicyCounters.hpp"
  47 #include "gc/shared/gcTimer.hpp"
  48 #include "gc/shared/gcTrace.hpp"
  49 #include "gc/shared/gcTraceTime.hpp"
  50 #include "gc/shared/genCollectedHeap.hpp"
  51 #include "gc/shared/genOopClosures.inline.hpp"
  52 #include "gc/shared/isGCActiveMark.hpp"
  53 #include "gc/shared/referencePolicy.hpp"
  54 #include "gc/shared/strongRootsScope.hpp"
  55 #include "gc/shared/taskqueue.inline.hpp"
  56 #include "logging/log.hpp"
  57 #include "memory/allocation.hpp"
  58 #include "memory/iterator.inline.hpp"
  59 #include "memory/padded.hpp"
  60 #include "memory/resourceArea.hpp"
  61 #include "oops/oop.inline.hpp"
  62 #include "prims/jvmtiExport.hpp"
  63 #include "runtime/atomic.inline.hpp"
  64 #include "runtime/globals_extension.hpp"
  65 #include "runtime/handles.inline.hpp"
  66 #include "runtime/java.hpp"
  67 #include "runtime/orderAccess.inline.hpp"
  68 #include "runtime/vmThread.hpp"
  69 #include "services/memoryService.hpp"
  70 #include "services/runtimeService.hpp"
  71 #include "utilities/stack.inline.hpp"
  72 
  73 // statics
  74 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
  75 bool CMSCollector::_full_gc_requested = false;
  76 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
  77 
  78 //////////////////////////////////////////////////////////////////
  79 // In support of CMS/VM thread synchronization
  80 //////////////////////////////////////////////////////////////////
  81 // We split use of the CGC_lock into 2 "levels".
  82 // The low-level locking is of the usual CGC_lock monitor. We introduce
  83 // a higher level "token" (hereafter "CMS token") built on top of the
  84 // low level monitor (hereafter "CGC lock").
  85 // The token-passing protocol gives priority to the VM thread. The
  86 // CMS-lock doesn't provide any fairness guarantees, but clients
  87 // should ensure that it is only held for very short, bounded
  88 // durations.
  89 //
  90 // When either of the CMS thread or the VM thread is involved in
  91 // collection operations during which it does not want the other
  92 // thread to interfere, it obtains the CMS token.
  93 //
  94 // If either thread tries to get the token while the other has
  95 // it, that thread waits. However, if the VM thread and CMS thread
  96 // both want the token, then the VM thread gets priority while the
  97 // CMS thread waits. This ensures, for instance, that the "concurrent"
  98 // phases of the CMS thread's work do not block out the VM thread
  99 // for long periods of time as the CMS thread continues to hog
 100 // the token. (See bug 4616232).
 101 //
 102 // The baton-passing functions are, however, controlled by the
 103 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
 104 // and here the low-level CMS lock, not the high level token,
 105 // ensures mutual exclusion.
 106 //
 107 // Two important conditions that we have to satisfy:
 108 // 1. if a thread does a low-level wait on the CMS lock, then it
 109 //    relinquishes the CMS token if it were holding that token
 110 //    when it acquired the low-level CMS lock.
 111 // 2. any low-level notifications on the low-level lock
 112 //    should only be sent when a thread has relinquished the token.
 113 //
 114 // In the absence of either property, we'd have potential deadlock.
 115 //
 116 // We protect each of the CMS (concurrent and sequential) phases
 117 // with the CMS _token_, not the CMS _lock_.
 118 //
 119 // The only code protected by CMS lock is the token acquisition code
 120 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
 121 // baton-passing code.
 122 //
 123 // Unfortunately, i couldn't come up with a good abstraction to factor and
 124 // hide the naked CGC_lock manipulation in the baton-passing code
 125 // further below. That's something we should try to do. Also, the proof
 126 // of correctness of this 2-level locking scheme is far from obvious,
 127 // and potentially quite slippery. We have an uneasy suspicion, for instance,
 128 // that there may be a theoretical possibility of delay/starvation in the
 129 // low-level lock/wait/notify scheme used for the baton-passing because of
 130 // potential interference with the priority scheme embodied in the
 131 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
 132 // invocation further below and marked with "XXX 20011219YSR".
 133 // Indeed, as we note elsewhere, this may become yet more slippery
 134 // in the presence of multiple CMS and/or multiple VM threads. XXX
 135 
 136 class CMSTokenSync: public StackObj {
 137  private:
 138   bool _is_cms_thread;
 139  public:
 140   CMSTokenSync(bool is_cms_thread):
 141     _is_cms_thread(is_cms_thread) {
 142     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
 143            "Incorrect argument to constructor");
 144     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
 145   }
 146 
 147   ~CMSTokenSync() {
 148     assert(_is_cms_thread ?
 149              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
 150              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
 151           "Incorrect state");
 152     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
 153   }
 154 };
 155 
 156 // Convenience class that does a CMSTokenSync, and then acquires
 157 // upto three locks.
 158 class CMSTokenSyncWithLocks: public CMSTokenSync {
 159  private:
 160   // Note: locks are acquired in textual declaration order
 161   // and released in the opposite order
 162   MutexLockerEx _locker1, _locker2, _locker3;
 163  public:
 164   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
 165                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
 166     CMSTokenSync(is_cms_thread),
 167     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
 168     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
 169     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
 170   { }
 171 };
 172 
 173 
 174 //////////////////////////////////////////////////////////////////
 175 //  Concurrent Mark-Sweep Generation /////////////////////////////
 176 //////////////////////////////////////////////////////////////////
 177 
 178 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
 179 
 180 // This struct contains per-thread things necessary to support parallel
 181 // young-gen collection.
 182 class CMSParGCThreadState: public CHeapObj<mtGC> {
 183  public:
 184   CFLS_LAB lab;
 185   PromotionInfo promo;
 186 
 187   // Constructor.
 188   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
 189     promo.setSpace(cfls);
 190   }
 191 };
 192 
 193 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
 194      ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
 195   CardGeneration(rs, initial_byte_size, ct),
 196   _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
 197   _did_compact(false)
 198 {
 199   HeapWord* bottom = (HeapWord*) _virtual_space.low();
 200   HeapWord* end    = (HeapWord*) _virtual_space.high();
 201 
 202   _direct_allocated_words = 0;
 203   NOT_PRODUCT(
 204     _numObjectsPromoted = 0;
 205     _numWordsPromoted = 0;
 206     _numObjectsAllocated = 0;
 207     _numWordsAllocated = 0;
 208   )
 209 
 210   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
 211   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
 212   _cmsSpace->_old_gen = this;
 213 
 214   _gc_stats = new CMSGCStats();
 215 
 216   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
 217   // offsets match. The ability to tell free chunks from objects
 218   // depends on this property.
 219   debug_only(
 220     FreeChunk* junk = NULL;
 221     assert(UseCompressedClassPointers ||
 222            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
 223            "Offset of FreeChunk::_prev within FreeChunk must match"
 224            "  that of OopDesc::_klass within OopDesc");
 225   )
 226 
 227   _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
 228   for (uint i = 0; i < ParallelGCThreads; i++) {
 229     _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
 230   }
 231 
 232   _incremental_collection_failed = false;
 233   // The "dilatation_factor" is the expansion that can occur on
 234   // account of the fact that the minimum object size in the CMS
 235   // generation may be larger than that in, say, a contiguous young
 236   //  generation.
 237   // Ideally, in the calculation below, we'd compute the dilatation
 238   // factor as: MinChunkSize/(promoting_gen's min object size)
 239   // Since we do not have such a general query interface for the
 240   // promoting generation, we'll instead just use the minimum
 241   // object size (which today is a header's worth of space);
 242   // note that all arithmetic is in units of HeapWords.
 243   assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
 244   assert(_dilatation_factor >= 1.0, "from previous assert");
 245 }
 246 
 247 
 248 // The field "_initiating_occupancy" represents the occupancy percentage
 249 // at which we trigger a new collection cycle.  Unless explicitly specified
 250 // via CMSInitiatingOccupancyFraction (argument "io" below), it
 251 // is calculated by:
 252 //
 253 //   Let "f" be MinHeapFreeRatio in
 254 //
 255 //    _initiating_occupancy = 100-f +
 256 //                           f * (CMSTriggerRatio/100)
 257 //   where CMSTriggerRatio is the argument "tr" below.
 258 //
 259 // That is, if we assume the heap is at its desired maximum occupancy at the
 260 // end of a collection, we let CMSTriggerRatio of the (purported) free
 261 // space be allocated before initiating a new collection cycle.
 262 //
 263 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
 264   assert(io <= 100 && tr <= 100, "Check the arguments");
 265   if (io >= 0) {
 266     _initiating_occupancy = (double)io / 100.0;
 267   } else {
 268     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
 269                              (double)(tr * MinHeapFreeRatio) / 100.0)
 270                             / 100.0;
 271   }
 272 }
 273 
 274 void ConcurrentMarkSweepGeneration::ref_processor_init() {
 275   assert(collector() != NULL, "no collector");
 276   collector()->ref_processor_init();
 277 }
 278 
 279 void CMSCollector::ref_processor_init() {
 280   if (_ref_processor == NULL) {
 281     // Allocate and initialize a reference processor
 282     _ref_processor =
 283       new ReferenceProcessor(_span,                               // span
 284                              (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
 285                              ParallelGCThreads,                   // mt processing degree
 286                              _cmsGen->refs_discovery_is_mt(),     // mt discovery
 287                              MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
 288                              _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
 289                              &_is_alive_closure);                 // closure for liveness info
 290     // Initialize the _ref_processor field of CMSGen
 291     _cmsGen->set_ref_processor(_ref_processor);
 292 
 293   }
 294 }
 295 
 296 AdaptiveSizePolicy* CMSCollector::size_policy() {
 297   GenCollectedHeap* gch = GenCollectedHeap::heap();
 298   return gch->gen_policy()->size_policy();
 299 }
 300 
 301 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
 302 
 303   const char* gen_name = "old";
 304   GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy();
 305   // Generation Counters - generation 1, 1 subspace
 306   _gen_counters = new GenerationCounters(gen_name, 1, 1,
 307       gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
 308 
 309   _space_counters = new GSpaceCounters(gen_name, 0,
 310                                        _virtual_space.reserved_size(),
 311                                        this, _gen_counters);
 312 }
 313 
 314 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
 315   _cms_gen(cms_gen)
 316 {
 317   assert(alpha <= 100, "bad value");
 318   _saved_alpha = alpha;
 319 
 320   // Initialize the alphas to the bootstrap value of 100.
 321   _gc0_alpha = _cms_alpha = 100;
 322 
 323   _cms_begin_time.update();
 324   _cms_end_time.update();
 325 
 326   _gc0_duration = 0.0;
 327   _gc0_period = 0.0;
 328   _gc0_promoted = 0;
 329 
 330   _cms_duration = 0.0;
 331   _cms_period = 0.0;
 332   _cms_allocated = 0;
 333 
 334   _cms_used_at_gc0_begin = 0;
 335   _cms_used_at_gc0_end = 0;
 336   _allow_duty_cycle_reduction = false;
 337   _valid_bits = 0;
 338 }
 339 
 340 double CMSStats::cms_free_adjustment_factor(size_t free) const {
 341   // TBD: CR 6909490
 342   return 1.0;
 343 }
 344 
 345 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
 346 }
 347 
 348 // If promotion failure handling is on use
 349 // the padded average size of the promotion for each
 350 // young generation collection.
 351 double CMSStats::time_until_cms_gen_full() const {
 352   size_t cms_free = _cms_gen->cmsSpace()->free();
 353   GenCollectedHeap* gch = GenCollectedHeap::heap();
 354   size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
 355                                    (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
 356   if (cms_free > expected_promotion) {
 357     // Start a cms collection if there isn't enough space to promote
 358     // for the next young collection.  Use the padded average as
 359     // a safety factor.
 360     cms_free -= expected_promotion;
 361 
 362     // Adjust by the safety factor.
 363     double cms_free_dbl = (double)cms_free;
 364     double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
 365     // Apply a further correction factor which tries to adjust
 366     // for recent occurance of concurrent mode failures.
 367     cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
 368     cms_free_dbl = cms_free_dbl * cms_adjustment;
 369 
 370     log_trace(gc, stats)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
 371                          cms_free, expected_promotion);
 372     log_trace(gc, stats)("  cms_free_dbl %f cms_consumption_rate %f",
 373                          cms_free_dbl, cms_consumption_rate() + 1.0);
 374     // Add 1 in case the consumption rate goes to zero.
 375     return cms_free_dbl / (cms_consumption_rate() + 1.0);
 376   }
 377   return 0.0;
 378 }
 379 
 380 // Compare the duration of the cms collection to the
 381 // time remaining before the cms generation is empty.
 382 // Note that the time from the start of the cms collection
 383 // to the start of the cms sweep (less than the total
 384 // duration of the cms collection) can be used.  This
 385 // has been tried and some applications experienced
 386 // promotion failures early in execution.  This was
 387 // possibly because the averages were not accurate
 388 // enough at the beginning.
 389 double CMSStats::time_until_cms_start() const {
 390   // We add "gc0_period" to the "work" calculation
 391   // below because this query is done (mostly) at the
 392   // end of a scavenge, so we need to conservatively
 393   // account for that much possible delay
 394   // in the query so as to avoid concurrent mode failures
 395   // due to starting the collection just a wee bit too
 396   // late.
 397   double work = cms_duration() + gc0_period();
 398   double deadline = time_until_cms_gen_full();
 399   // If a concurrent mode failure occurred recently, we want to be
 400   // more conservative and halve our expected time_until_cms_gen_full()
 401   if (work > deadline) {
 402     log_develop(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
 403                     cms_duration(), gc0_period(), time_until_cms_gen_full());
 404     return 0.0;
 405   }
 406   return work - deadline;
 407 }
 408 
 409 #ifndef PRODUCT
 410 void CMSStats::print_on(outputStream *st) const {
 411   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
 412   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
 413                gc0_duration(), gc0_period(), gc0_promoted());
 414   st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
 415             cms_duration(), cms_period(), cms_allocated());
 416   st->print(",cms_since_beg=%g,cms_since_end=%g",
 417             cms_time_since_begin(), cms_time_since_end());
 418   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
 419             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
 420 
 421   if (valid()) {
 422     st->print(",promo_rate=%g,cms_alloc_rate=%g",
 423               promotion_rate(), cms_allocation_rate());
 424     st->print(",cms_consumption_rate=%g,time_until_full=%g",
 425               cms_consumption_rate(), time_until_cms_gen_full());
 426   }
 427   st->print(" ");
 428 }
 429 #endif // #ifndef PRODUCT
 430 
 431 CMSCollector::CollectorState CMSCollector::_collectorState =
 432                              CMSCollector::Idling;
 433 bool CMSCollector::_foregroundGCIsActive = false;
 434 bool CMSCollector::_foregroundGCShouldWait = false;
 435 
 436 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
 437                            CardTableRS*                   ct,
 438                            ConcurrentMarkSweepPolicy*     cp):
 439   _cmsGen(cmsGen),
 440   _ct(ct),
 441   _ref_processor(NULL),    // will be set later
 442   _conc_workers(NULL),     // may be set later
 443   _abort_preclean(false),
 444   _start_sampling(false),
 445   _between_prologue_and_epilogue(false),
 446   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
 447   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
 448                  -1 /* lock-free */, "No_lock" /* dummy */),
 449   _modUnionClosurePar(&_modUnionTable),
 450   // Adjust my span to cover old (cms) gen
 451   _span(cmsGen->reserved()),
 452   // Construct the is_alive_closure with _span & markBitMap
 453   _is_alive_closure(_span, &_markBitMap),
 454   _restart_addr(NULL),
 455   _overflow_list(NULL),
 456   _stats(cmsGen),
 457   _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
 458                              //verify that this lock should be acquired with safepoint check.
 459                              Monitor::_safepoint_check_sometimes)),
 460   _eden_chunk_array(NULL),     // may be set in ctor body
 461   _eden_chunk_capacity(0),     // -- ditto --
 462   _eden_chunk_index(0),        // -- ditto --
 463   _survivor_plab_array(NULL),  // -- ditto --
 464   _survivor_chunk_array(NULL), // -- ditto --
 465   _survivor_chunk_capacity(0), // -- ditto --
 466   _survivor_chunk_index(0),    // -- ditto --
 467   _ser_pmc_preclean_ovflw(0),
 468   _ser_kac_preclean_ovflw(0),
 469   _ser_pmc_remark_ovflw(0),
 470   _par_pmc_remark_ovflw(0),
 471   _ser_kac_ovflw(0),
 472   _par_kac_ovflw(0),
 473 #ifndef PRODUCT
 474   _num_par_pushes(0),
 475 #endif
 476   _collection_count_start(0),
 477   _verifying(false),
 478   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
 479   _completed_initialization(false),
 480   _collector_policy(cp),
 481   _should_unload_classes(CMSClassUnloadingEnabled),
 482   _concurrent_cycles_since_last_unload(0),
 483   _roots_scanning_options(GenCollectedHeap::SO_None),
 484   _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 485   _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
 486   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
 487   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 488   _cms_start_registered(false)
 489 {
 490   if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
 491     ExplicitGCInvokesConcurrent = true;
 492   }
 493   // Now expand the span and allocate the collection support structures
 494   // (MUT, marking bit map etc.) to cover both generations subject to
 495   // collection.
 496 
 497   // For use by dirty card to oop closures.
 498   _cmsGen->cmsSpace()->set_collector(this);
 499 
 500   // Allocate MUT and marking bit map
 501   {
 502     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
 503     if (!_markBitMap.allocate(_span)) {
 504       warning("Failed to allocate CMS Bit Map");
 505       return;
 506     }
 507     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
 508   }
 509   {
 510     _modUnionTable.allocate(_span);
 511     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
 512   }
 513 
 514   if (!_markStack.allocate(MarkStackSize)) {
 515     warning("Failed to allocate CMS Marking Stack");
 516     return;
 517   }
 518 
 519   // Support for multi-threaded concurrent phases
 520   if (CMSConcurrentMTEnabled) {
 521     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
 522       // just for now
 523       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
 524     }
 525     if (ConcGCThreads > 1) {
 526       _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
 527                                  ConcGCThreads, true);
 528       if (_conc_workers == NULL) {
 529         warning("GC/CMS: _conc_workers allocation failure: "
 530               "forcing -CMSConcurrentMTEnabled");
 531         CMSConcurrentMTEnabled = false;
 532       } else {
 533         _conc_workers->initialize_workers();
 534       }
 535     } else {
 536       CMSConcurrentMTEnabled = false;
 537     }
 538   }
 539   if (!CMSConcurrentMTEnabled) {
 540     ConcGCThreads = 0;
 541   } else {
 542     // Turn off CMSCleanOnEnter optimization temporarily for
 543     // the MT case where it's not fixed yet; see 6178663.
 544     CMSCleanOnEnter = false;
 545   }
 546   assert((_conc_workers != NULL) == (ConcGCThreads > 1),
 547          "Inconsistency");
 548 
 549   // Parallel task queues; these are shared for the
 550   // concurrent and stop-world phases of CMS, but
 551   // are not shared with parallel scavenge (ParNew).
 552   {
 553     uint i;
 554     uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
 555 
 556     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
 557          || ParallelRefProcEnabled)
 558         && num_queues > 0) {
 559       _task_queues = new OopTaskQueueSet(num_queues);
 560       if (_task_queues == NULL) {
 561         warning("task_queues allocation failure.");
 562         return;
 563       }
 564       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
 565       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
 566       for (i = 0; i < num_queues; i++) {
 567         PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
 568         if (q == NULL) {
 569           warning("work_queue allocation failure.");
 570           return;
 571         }
 572         _task_queues->register_queue(i, q);
 573       }
 574       for (i = 0; i < num_queues; i++) {
 575         _task_queues->queue(i)->initialize();
 576         _hash_seed[i] = 17;  // copied from ParNew
 577       }
 578     }
 579   }
 580 
 581   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
 582 
 583   // Clip CMSBootstrapOccupancy between 0 and 100.
 584   _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
 585 
 586   // Now tell CMS generations the identity of their collector
 587   ConcurrentMarkSweepGeneration::set_collector(this);
 588 
 589   // Create & start a CMS thread for this CMS collector
 590   _cmsThread = ConcurrentMarkSweepThread::start(this);
 591   assert(cmsThread() != NULL, "CMS Thread should have been created");
 592   assert(cmsThread()->collector() == this,
 593          "CMS Thread should refer to this gen");
 594   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 595 
 596   // Support for parallelizing young gen rescan
 597   GenCollectedHeap* gch = GenCollectedHeap::heap();
 598   assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
 599   _young_gen = (ParNewGeneration*)gch->young_gen();
 600   if (gch->supports_inline_contig_alloc()) {
 601     _top_addr = gch->top_addr();
 602     _end_addr = gch->end_addr();
 603     assert(_young_gen != NULL, "no _young_gen");
 604     _eden_chunk_index = 0;
 605     _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
 606     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
 607   }
 608 
 609   // Support for parallelizing survivor space rescan
 610   if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
 611     const size_t max_plab_samples =
 612       _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
 613 
 614     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
 615     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 616     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
 617     _survivor_chunk_capacity = max_plab_samples;
 618     for (uint i = 0; i < ParallelGCThreads; i++) {
 619       HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
 620       ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
 621       assert(cur->end() == 0, "Should be 0");
 622       assert(cur->array() == vec, "Should be vec");
 623       assert(cur->capacity() == max_plab_samples, "Error");
 624     }
 625   }
 626 
 627   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
 628   _gc_counters = new CollectorCounters("CMS", 1);
 629   _completed_initialization = true;
 630   _inter_sweep_timer.start();  // start of time
 631 }
 632 
 633 const char* ConcurrentMarkSweepGeneration::name() const {
 634   return "concurrent mark-sweep generation";
 635 }
 636 void ConcurrentMarkSweepGeneration::update_counters() {
 637   if (UsePerfData) {
 638     _space_counters->update_all();
 639     _gen_counters->update_all();
 640   }
 641 }
 642 
 643 // this is an optimized version of update_counters(). it takes the
 644 // used value as a parameter rather than computing it.
 645 //
 646 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
 647   if (UsePerfData) {
 648     _space_counters->update_used(used);
 649     _space_counters->update_capacity();
 650     _gen_counters->update_all();
 651   }
 652 }
 653 
 654 void ConcurrentMarkSweepGeneration::print() const {
 655   Generation::print();
 656   cmsSpace()->print();
 657 }
 658 
 659 #ifndef PRODUCT
 660 void ConcurrentMarkSweepGeneration::print_statistics() {
 661   cmsSpace()->printFLCensus(0);
 662 }
 663 #endif
 664 
 665 size_t
 666 ConcurrentMarkSweepGeneration::contiguous_available() const {
 667   // dld proposes an improvement in precision here. If the committed
 668   // part of the space ends in a free block we should add that to
 669   // uncommitted size in the calculation below. Will make this
 670   // change later, staying with the approximation below for the
 671   // time being. -- ysr.
 672   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
 673 }
 674 
 675 size_t
 676 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
 677   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
 678 }
 679 
 680 size_t ConcurrentMarkSweepGeneration::max_available() const {
 681   return free() + _virtual_space.uncommitted_size();
 682 }
 683 
 684 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
 685   size_t available = max_available();
 686   size_t av_promo  = (size_t)gc_stats()->avg_promoted()->padded_average();
 687   bool   res = (available >= av_promo) || (available >= max_promotion_in_bytes);
 688   log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
 689                            res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
 690   return res;
 691 }
 692 
 693 // At a promotion failure dump information on block layout in heap
 694 // (cms old generation).
 695 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
 696   LogHandle(gc, promotion) log;
 697   if (log.is_trace()) {
 698     ResourceMark rm;
 699     cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream());
 700   }
 701 }
 702 
 703 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
 704   // Clear the promotion information.  These pointers can be adjusted
 705   // along with all the other pointers into the heap but
 706   // compaction is expected to be a rare event with
 707   // a heap using cms so don't do it without seeing the need.
 708   for (uint i = 0; i < ParallelGCThreads; i++) {
 709     _par_gc_thread_states[i]->promo.reset();
 710   }
 711 }
 712 
 713 void ConcurrentMarkSweepGeneration::compute_new_size() {
 714   assert_locked_or_safepoint(Heap_lock);
 715 
 716   // If incremental collection failed, we just want to expand
 717   // to the limit.
 718   if (incremental_collection_failed()) {
 719     clear_incremental_collection_failed();
 720     grow_to_reserved();
 721     return;
 722   }
 723 
 724   // The heap has been compacted but not reset yet.
 725   // Any metric such as free() or used() will be incorrect.
 726 
 727   CardGeneration::compute_new_size();
 728 
 729   // Reset again after a possible resizing
 730   if (did_compact()) {
 731     cmsSpace()->reset_after_compaction();
 732   }
 733 }
 734 
 735 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
 736   assert_locked_or_safepoint(Heap_lock);
 737 
 738   // If incremental collection failed, we just want to expand
 739   // to the limit.
 740   if (incremental_collection_failed()) {
 741     clear_incremental_collection_failed();
 742     grow_to_reserved();
 743     return;
 744   }
 745 
 746   double free_percentage = ((double) free()) / capacity();
 747   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
 748   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
 749 
 750   // compute expansion delta needed for reaching desired free percentage
 751   if (free_percentage < desired_free_percentage) {
 752     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 753     assert(desired_capacity >= capacity(), "invalid expansion size");
 754     size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
 755     LogHandle(gc) log;
 756     if (log.is_trace()) {
 757       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 758       log.trace("From compute_new_size: ");
 759       log.trace("  Free fraction %f", free_percentage);
 760       log.trace("  Desired free fraction %f", desired_free_percentage);
 761       log.trace("  Maximum free fraction %f", maximum_free_percentage);
 762       log.trace("  Capacity " SIZE_FORMAT, capacity() / 1000);
 763       log.trace("  Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
 764       GenCollectedHeap* gch = GenCollectedHeap::heap();
 765       assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
 766       size_t young_size = gch->young_gen()->capacity();
 767       log.trace("  Young gen size " SIZE_FORMAT, young_size / 1000);
 768       log.trace("  unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
 769       log.trace("  contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
 770       log.trace("  Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
 771     }
 772     // safe if expansion fails
 773     expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
 774     log.trace("  Expanded free fraction %f", ((double) free()) / capacity());
 775   } else {
 776     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 777     assert(desired_capacity <= capacity(), "invalid expansion size");
 778     size_t shrink_bytes = capacity() - desired_capacity;
 779     // Don't shrink unless the delta is greater than the minimum shrink we want
 780     if (shrink_bytes >= MinHeapDeltaBytes) {
 781       shrink_free_list_by(shrink_bytes);
 782     }
 783   }
 784 }
 785 
 786 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
 787   return cmsSpace()->freelistLock();
 788 }
 789 
 790 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
 791   CMSSynchronousYieldRequest yr;
 792   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
 793   return have_lock_and_allocate(size, tlab);
 794 }
 795 
 796 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
 797                                                                 bool   tlab /* ignored */) {
 798   assert_lock_strong(freelistLock());
 799   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
 800   HeapWord* res = cmsSpace()->allocate(adjustedSize);
 801   // Allocate the object live (grey) if the background collector has
 802   // started marking. This is necessary because the marker may
 803   // have passed this address and consequently this object will
 804   // not otherwise be greyed and would be incorrectly swept up.
 805   // Note that if this object contains references, the writing
 806   // of those references will dirty the card containing this object
 807   // allowing the object to be blackened (and its references scanned)
 808   // either during a preclean phase or at the final checkpoint.
 809   if (res != NULL) {
 810     // We may block here with an uninitialized object with
 811     // its mark-bit or P-bits not yet set. Such objects need
 812     // to be safely navigable by block_start().
 813     assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
 814     assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
 815     collector()->direct_allocated(res, adjustedSize);
 816     _direct_allocated_words += adjustedSize;
 817     // allocation counters
 818     NOT_PRODUCT(
 819       _numObjectsAllocated++;
 820       _numWordsAllocated += (int)adjustedSize;
 821     )
 822   }
 823   return res;
 824 }
 825 
 826 // In the case of direct allocation by mutators in a generation that
 827 // is being concurrently collected, the object must be allocated
 828 // live (grey) if the background collector has started marking.
 829 // This is necessary because the marker may
 830 // have passed this address and consequently this object will
 831 // not otherwise be greyed and would be incorrectly swept up.
 832 // Note that if this object contains references, the writing
 833 // of those references will dirty the card containing this object
 834 // allowing the object to be blackened (and its references scanned)
 835 // either during a preclean phase or at the final checkpoint.
 836 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
 837   assert(_markBitMap.covers(start, size), "Out of bounds");
 838   if (_collectorState >= Marking) {
 839     MutexLockerEx y(_markBitMap.lock(),
 840                     Mutex::_no_safepoint_check_flag);
 841     // [see comments preceding SweepClosure::do_blk() below for details]
 842     //
 843     // Can the P-bits be deleted now?  JJJ
 844     //
 845     // 1. need to mark the object as live so it isn't collected
 846     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
 847     // 3. need to mark the end of the object so marking, precleaning or sweeping
 848     //    can skip over uninitialized or unparsable objects. An allocated
 849     //    object is considered uninitialized for our purposes as long as
 850     //    its klass word is NULL.  All old gen objects are parsable
 851     //    as soon as they are initialized.)
 852     _markBitMap.mark(start);          // object is live
 853     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
 854     _markBitMap.mark(start + size - 1);
 855                                       // mark end of object
 856   }
 857   // check that oop looks uninitialized
 858   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
 859 }
 860 
 861 void CMSCollector::promoted(bool par, HeapWord* start,
 862                             bool is_obj_array, size_t obj_size) {
 863   assert(_markBitMap.covers(start), "Out of bounds");
 864   // See comment in direct_allocated() about when objects should
 865   // be allocated live.
 866   if (_collectorState >= Marking) {
 867     // we already hold the marking bit map lock, taken in
 868     // the prologue
 869     if (par) {
 870       _markBitMap.par_mark(start);
 871     } else {
 872       _markBitMap.mark(start);
 873     }
 874     // We don't need to mark the object as uninitialized (as
 875     // in direct_allocated above) because this is being done with the
 876     // world stopped and the object will be initialized by the
 877     // time the marking, precleaning or sweeping get to look at it.
 878     // But see the code for copying objects into the CMS generation,
 879     // where we need to ensure that concurrent readers of the
 880     // block offset table are able to safely navigate a block that
 881     // is in flux from being free to being allocated (and in
 882     // transition while being copied into) and subsequently
 883     // becoming a bona-fide object when the copy/promotion is complete.
 884     assert(SafepointSynchronize::is_at_safepoint(),
 885            "expect promotion only at safepoints");
 886 
 887     if (_collectorState < Sweeping) {
 888       // Mark the appropriate cards in the modUnionTable, so that
 889       // this object gets scanned before the sweep. If this is
 890       // not done, CMS generation references in the object might
 891       // not get marked.
 892       // For the case of arrays, which are otherwise precisely
 893       // marked, we need to dirty the entire array, not just its head.
 894       if (is_obj_array) {
 895         // The [par_]mark_range() method expects mr.end() below to
 896         // be aligned to the granularity of a bit's representation
 897         // in the heap. In the case of the MUT below, that's a
 898         // card size.
 899         MemRegion mr(start,
 900                      (HeapWord*)round_to((intptr_t)(start + obj_size),
 901                         CardTableModRefBS::card_size /* bytes */));
 902         if (par) {
 903           _modUnionTable.par_mark_range(mr);
 904         } else {
 905           _modUnionTable.mark_range(mr);
 906         }
 907       } else {  // not an obj array; we can just mark the head
 908         if (par) {
 909           _modUnionTable.par_mark(start);
 910         } else {
 911           _modUnionTable.mark(start);
 912         }
 913       }
 914     }
 915   }
 916 }
 917 
 918 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
 919   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
 920   // allocate, copy and if necessary update promoinfo --
 921   // delegate to underlying space.
 922   assert_lock_strong(freelistLock());
 923 
 924 #ifndef PRODUCT
 925   if (GenCollectedHeap::heap()->promotion_should_fail()) {
 926     return NULL;
 927   }
 928 #endif  // #ifndef PRODUCT
 929 
 930   oop res = _cmsSpace->promote(obj, obj_size);
 931   if (res == NULL) {
 932     // expand and retry
 933     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
 934     expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
 935     // Since this is the old generation, we don't try to promote
 936     // into a more senior generation.
 937     res = _cmsSpace->promote(obj, obj_size);
 938   }
 939   if (res != NULL) {
 940     // See comment in allocate() about when objects should
 941     // be allocated live.
 942     assert(obj->is_oop(), "Will dereference klass pointer below");
 943     collector()->promoted(false,           // Not parallel
 944                           (HeapWord*)res, obj->is_objArray(), obj_size);
 945     // promotion counters
 946     NOT_PRODUCT(
 947       _numObjectsPromoted++;
 948       _numWordsPromoted +=
 949         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
 950     )
 951   }
 952   return res;
 953 }
 954 
 955 
 956 // IMPORTANT: Notes on object size recognition in CMS.
 957 // ---------------------------------------------------
 958 // A block of storage in the CMS generation is always in
 959 // one of three states. A free block (FREE), an allocated
 960 // object (OBJECT) whose size() method reports the correct size,
 961 // and an intermediate state (TRANSIENT) in which its size cannot
 962 // be accurately determined.
 963 // STATE IDENTIFICATION:   (32 bit and 64 bit w/o COOPS)
 964 // -----------------------------------------------------
 965 // FREE:      klass_word & 1 == 1; mark_word holds block size
 966 //
 967 // OBJECT:    klass_word installed; klass_word != 0 && klass_word & 1 == 0;
 968 //            obj->size() computes correct size
 969 //
 970 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
 971 //
 972 // STATE IDENTIFICATION: (64 bit+COOPS)
 973 // ------------------------------------
 974 // FREE:      mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
 975 //
 976 // OBJECT:    klass_word installed; klass_word != 0;
 977 //            obj->size() computes correct size
 978 //
 979 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
 980 //
 981 //
 982 // STATE TRANSITION DIAGRAM
 983 //
 984 //        mut / parnew                     mut  /  parnew
 985 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
 986 //  ^                                                                   |
 987 //  |------------------------ DEAD <------------------------------------|
 988 //         sweep                            mut
 989 //
 990 // While a block is in TRANSIENT state its size cannot be determined
 991 // so readers will either need to come back later or stall until
 992 // the size can be determined. Note that for the case of direct
 993 // allocation, P-bits, when available, may be used to determine the
 994 // size of an object that may not yet have been initialized.
 995 
 996 // Things to support parallel young-gen collection.
 997 oop
 998 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
 999                                            oop old, markOop m,
1000                                            size_t word_sz) {
1001 #ifndef PRODUCT
1002   if (GenCollectedHeap::heap()->promotion_should_fail()) {
1003     return NULL;
1004   }
1005 #endif  // #ifndef PRODUCT
1006 
1007   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1008   PromotionInfo* promoInfo = &ps->promo;
1009   // if we are tracking promotions, then first ensure space for
1010   // promotion (including spooling space for saving header if necessary).
1011   // then allocate and copy, then track promoted info if needed.
1012   // When tracking (see PromotionInfo::track()), the mark word may
1013   // be displaced and in this case restoration of the mark word
1014   // occurs in the (oop_since_save_marks_)iterate phase.
1015   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1016     // Out of space for allocating spooling buffers;
1017     // try expanding and allocating spooling buffers.
1018     if (!expand_and_ensure_spooling_space(promoInfo)) {
1019       return NULL;
1020     }
1021   }
1022   assert(promoInfo->has_spooling_space(), "Control point invariant");
1023   const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1024   HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1025   if (obj_ptr == NULL) {
1026      obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1027      if (obj_ptr == NULL) {
1028        return NULL;
1029      }
1030   }
1031   oop obj = oop(obj_ptr);
1032   OrderAccess::storestore();
1033   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1034   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1035   // IMPORTANT: See note on object initialization for CMS above.
1036   // Otherwise, copy the object.  Here we must be careful to insert the
1037   // klass pointer last, since this marks the block as an allocated object.
1038   // Except with compressed oops it's the mark word.
1039   HeapWord* old_ptr = (HeapWord*)old;
1040   // Restore the mark word copied above.
1041   obj->set_mark(m);
1042   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1043   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1044   OrderAccess::storestore();
1045 
1046   if (UseCompressedClassPointers) {
1047     // Copy gap missed by (aligned) header size calculation below
1048     obj->set_klass_gap(old->klass_gap());
1049   }
1050   if (word_sz > (size_t)oopDesc::header_size()) {
1051     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1052                                  obj_ptr + oopDesc::header_size(),
1053                                  word_sz - oopDesc::header_size());
1054   }
1055 
1056   // Now we can track the promoted object, if necessary.  We take care
1057   // to delay the transition from uninitialized to full object
1058   // (i.e., insertion of klass pointer) until after, so that it
1059   // atomically becomes a promoted object.
1060   if (promoInfo->tracking()) {
1061     promoInfo->track((PromotedObject*)obj, old->klass());
1062   }
1063   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1064   assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1065   assert(old->is_oop(), "Will use and dereference old klass ptr below");
1066 
1067   // Finally, install the klass pointer (this should be volatile).
1068   OrderAccess::storestore();
1069   obj->set_klass(old->klass());
1070   // We should now be able to calculate the right size for this object
1071   assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1072 
1073   collector()->promoted(true,          // parallel
1074                         obj_ptr, old->is_objArray(), word_sz);
1075 
1076   NOT_PRODUCT(
1077     Atomic::inc_ptr(&_numObjectsPromoted);
1078     Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1079   )
1080 
1081   return obj;
1082 }
1083 
1084 void
1085 ConcurrentMarkSweepGeneration::
1086 par_promote_alloc_done(int thread_num) {
1087   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1088   ps->lab.retire(thread_num);
1089 }
1090 
1091 void
1092 ConcurrentMarkSweepGeneration::
1093 par_oop_since_save_marks_iterate_done(int thread_num) {
1094   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1095   ParScanWithoutBarrierClosure* dummy_cl = NULL;
1096   ps->promo.promoted_oops_iterate_nv(dummy_cl);
1097 }
1098 
1099 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
1100                                                    size_t size,
1101                                                    bool   tlab)
1102 {
1103   // We allow a STW collection only if a full
1104   // collection was requested.
1105   return full || should_allocate(size, tlab); // FIX ME !!!
1106   // This and promotion failure handling are connected at the
1107   // hip and should be fixed by untying them.
1108 }
1109 
1110 bool CMSCollector::shouldConcurrentCollect() {
1111   if (_full_gc_requested) {
1112     log_trace(gc)("CMSCollector: collect because of explicit  gc request (or gc_locker)");
1113     return true;
1114   }
1115 
1116   FreelistLocker x(this);
1117   // ------------------------------------------------------------------
1118   // Print out lots of information which affects the initiation of
1119   // a collection.
1120   LogHandle(gc, stats) log;
1121   if (log.is_debug() && stats().valid()) {
1122     log.debug("CMSCollector shouldConcurrentCollect: ");
1123     ResourceMark rm;
1124     stats().print_on(log.debug_stream());
1125     log.debug("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1126     log.debug("free=" SIZE_FORMAT, _cmsGen->free());
1127     log.debug("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1128     log.debug("promotion_rate=%g", stats().promotion_rate());
1129     log.debug("cms_allocation_rate=%g", stats().cms_allocation_rate());
1130     log.debug("occupancy=%3.7f", _cmsGen->occupancy());
1131     log.debug("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1132     log.debug("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1133     log.debug("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1134     log.debug("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1135   }
1136   // ------------------------------------------------------------------
1137 
1138   // If the estimated time to complete a cms collection (cms_duration())
1139   // is less than the estimated time remaining until the cms generation
1140   // is full, start a collection.
1141   if (!UseCMSInitiatingOccupancyOnly) {
1142     if (stats().valid()) {
1143       if (stats().time_until_cms_start() == 0.0) {
1144         return true;
1145       }
1146     } else {
1147       // We want to conservatively collect somewhat early in order
1148       // to try and "bootstrap" our CMS/promotion statistics;
1149       // this branch will not fire after the first successful CMS
1150       // collection because the stats should then be valid.
1151       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1152         log_trace(gc)(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1153                       _cmsGen->occupancy(), _bootstrap_occupancy);
1154         return true;
1155       }
1156     }
1157   }
1158 
1159   // Otherwise, we start a collection cycle if
1160   // old gen want a collection cycle started. Each may use
1161   // an appropriate criterion for making this decision.
1162   // XXX We need to make sure that the gen expansion
1163   // criterion dovetails well with this. XXX NEED TO FIX THIS
1164   if (_cmsGen->should_concurrent_collect()) {
1165     log_trace(gc)("CMS old gen initiated");
1166     return true;
1167   }
1168 
1169   // We start a collection if we believe an incremental collection may fail;
1170   // this is not likely to be productive in practice because it's probably too
1171   // late anyway.
1172   GenCollectedHeap* gch = GenCollectedHeap::heap();
1173   assert(gch->collector_policy()->is_generation_policy(),
1174          "You may want to check the correctness of the following");
1175   if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1176     log_trace(gc)("CMSCollector: collect because incremental collection will fail ");
1177     return true;
1178   }
1179 
1180   if (MetaspaceGC::should_concurrent_collect()) {
1181     log_trace(gc)("CMSCollector: collect for metadata allocation ");
1182     return true;
1183   }
1184 
1185   // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1186   if (CMSTriggerInterval >= 0) {
1187     if (CMSTriggerInterval == 0) {
1188       // Trigger always
1189       return true;
1190     }
1191 
1192     // Check the CMS time since begin (we do not check the stats validity
1193     // as we want to be able to trigger the first CMS cycle as well)
1194     if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1195       if (stats().valid()) {
1196         log_trace(gc)("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1197                       stats().cms_time_since_begin());
1198       } else {
1199         log_trace(gc)("CMSCollector: collect because of trigger interval (first collection)");
1200       }
1201       return true;
1202     }
1203   }
1204 
1205   return false;
1206 }
1207 
1208 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1209 
1210 // Clear _expansion_cause fields of constituent generations
1211 void CMSCollector::clear_expansion_cause() {
1212   _cmsGen->clear_expansion_cause();
1213 }
1214 
1215 // We should be conservative in starting a collection cycle.  To
1216 // start too eagerly runs the risk of collecting too often in the
1217 // extreme.  To collect too rarely falls back on full collections,
1218 // which works, even if not optimum in terms of concurrent work.
1219 // As a work around for too eagerly collecting, use the flag
1220 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
1221 // giving the user an easily understandable way of controlling the
1222 // collections.
1223 // We want to start a new collection cycle if any of the following
1224 // conditions hold:
1225 // . our current occupancy exceeds the configured initiating occupancy
1226 //   for this generation, or
1227 // . we recently needed to expand this space and have not, since that
1228 //   expansion, done a collection of this generation, or
1229 // . the underlying space believes that it may be a good idea to initiate
1230 //   a concurrent collection (this may be based on criteria such as the
1231 //   following: the space uses linear allocation and linear allocation is
1232 //   going to fail, or there is believed to be excessive fragmentation in
1233 //   the generation, etc... or ...
1234 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1235 //   the case of the old generation; see CR 6543076):
1236 //   we may be approaching a point at which allocation requests may fail because
1237 //   we will be out of sufficient free space given allocation rate estimates.]
1238 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1239 
1240   assert_lock_strong(freelistLock());
1241   if (occupancy() > initiating_occupancy()) {
1242     log_trace(gc)(" %s: collect because of occupancy %f / %f  ",
1243                   short_name(), occupancy(), initiating_occupancy());
1244     return true;
1245   }
1246   if (UseCMSInitiatingOccupancyOnly) {
1247     return false;
1248   }
1249   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1250     log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1251     return true;
1252   }
1253   return false;
1254 }
1255 
1256 void ConcurrentMarkSweepGeneration::collect(bool   full,
1257                                             bool   clear_all_soft_refs,
1258                                             size_t size,
1259                                             bool   tlab)
1260 {
1261   collector()->collect(full, clear_all_soft_refs, size, tlab);
1262 }
1263 
1264 void CMSCollector::collect(bool   full,
1265                            bool   clear_all_soft_refs,
1266                            size_t size,
1267                            bool   tlab)
1268 {
1269   // The following "if" branch is present for defensive reasons.
1270   // In the current uses of this interface, it can be replaced with:
1271   // assert(!GC_locker.is_active(), "Can't be called otherwise");
1272   // But I am not placing that assert here to allow future
1273   // generality in invoking this interface.
1274   if (GC_locker::is_active()) {
1275     // A consistency test for GC_locker
1276     assert(GC_locker::needs_gc(), "Should have been set already");
1277     // Skip this foreground collection, instead
1278     // expanding the heap if necessary.
1279     // Need the free list locks for the call to free() in compute_new_size()
1280     compute_new_size();
1281     return;
1282   }
1283   acquire_control_and_collect(full, clear_all_soft_refs);
1284 }
1285 
1286 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1287   GenCollectedHeap* gch = GenCollectedHeap::heap();
1288   unsigned int gc_count = gch->total_full_collections();
1289   if (gc_count == full_gc_count) {
1290     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1291     _full_gc_requested = true;
1292     _full_gc_cause = cause;
1293     CGC_lock->notify();   // nudge CMS thread
1294   } else {
1295     assert(gc_count > full_gc_count, "Error: causal loop");
1296   }
1297 }
1298 
1299 bool CMSCollector::is_external_interruption() {
1300   GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1301   return GCCause::is_user_requested_gc(cause) ||
1302          GCCause::is_serviceability_requested_gc(cause);
1303 }
1304 
1305 void CMSCollector::report_concurrent_mode_interruption() {
1306   if (is_external_interruption()) {
1307     log_debug(gc)("Concurrent mode interrupted");
1308   } else {
1309     log_debug(gc)("Concurrent mode failure");
1310     _gc_tracer_cm->report_concurrent_mode_failure();
1311   }
1312 }
1313 
1314 
1315 // The foreground and background collectors need to coordinate in order
1316 // to make sure that they do not mutually interfere with CMS collections.
1317 // When a background collection is active,
1318 // the foreground collector may need to take over (preempt) and
1319 // synchronously complete an ongoing collection. Depending on the
1320 // frequency of the background collections and the heap usage
1321 // of the application, this preemption can be seldom or frequent.
1322 // There are only certain
1323 // points in the background collection that the "collection-baton"
1324 // can be passed to the foreground collector.
1325 //
1326 // The foreground collector will wait for the baton before
1327 // starting any part of the collection.  The foreground collector
1328 // will only wait at one location.
1329 //
1330 // The background collector will yield the baton before starting a new
1331 // phase of the collection (e.g., before initial marking, marking from roots,
1332 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
1333 // of the loop which switches the phases. The background collector does some
1334 // of the phases (initial mark, final re-mark) with the world stopped.
1335 // Because of locking involved in stopping the world,
1336 // the foreground collector should not block waiting for the background
1337 // collector when it is doing a stop-the-world phase.  The background
1338 // collector will yield the baton at an additional point just before
1339 // it enters a stop-the-world phase.  Once the world is stopped, the
1340 // background collector checks the phase of the collection.  If the
1341 // phase has not changed, it proceeds with the collection.  If the
1342 // phase has changed, it skips that phase of the collection.  See
1343 // the comments on the use of the Heap_lock in collect_in_background().
1344 //
1345 // Variable used in baton passing.
1346 //   _foregroundGCIsActive - Set to true by the foreground collector when
1347 //      it wants the baton.  The foreground clears it when it has finished
1348 //      the collection.
1349 //   _foregroundGCShouldWait - Set to true by the background collector
1350 //        when it is running.  The foreground collector waits while
1351 //      _foregroundGCShouldWait is true.
1352 //  CGC_lock - monitor used to protect access to the above variables
1353 //      and to notify the foreground and background collectors.
1354 //  _collectorState - current state of the CMS collection.
1355 //
1356 // The foreground collector
1357 //   acquires the CGC_lock
1358 //   sets _foregroundGCIsActive
1359 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
1360 //     various locks acquired in preparation for the collection
1361 //     are released so as not to block the background collector
1362 //     that is in the midst of a collection
1363 //   proceeds with the collection
1364 //   clears _foregroundGCIsActive
1365 //   returns
1366 //
1367 // The background collector in a loop iterating on the phases of the
1368 //      collection
1369 //   acquires the CGC_lock
1370 //   sets _foregroundGCShouldWait
1371 //   if _foregroundGCIsActive is set
1372 //     clears _foregroundGCShouldWait, notifies _CGC_lock
1373 //     waits on _CGC_lock for _foregroundGCIsActive to become false
1374 //     and exits the loop.
1375 //   otherwise
1376 //     proceed with that phase of the collection
1377 //     if the phase is a stop-the-world phase,
1378 //       yield the baton once more just before enqueueing
1379 //       the stop-world CMS operation (executed by the VM thread).
1380 //   returns after all phases of the collection are done
1381 //
1382 
1383 void CMSCollector::acquire_control_and_collect(bool full,
1384         bool clear_all_soft_refs) {
1385   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1386   assert(!Thread::current()->is_ConcurrentGC_thread(),
1387          "shouldn't try to acquire control from self!");
1388 
1389   // Start the protocol for acquiring control of the
1390   // collection from the background collector (aka CMS thread).
1391   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1392          "VM thread should have CMS token");
1393   // Remember the possibly interrupted state of an ongoing
1394   // concurrent collection
1395   CollectorState first_state = _collectorState;
1396 
1397   // Signal to a possibly ongoing concurrent collection that
1398   // we want to do a foreground collection.
1399   _foregroundGCIsActive = true;
1400 
1401   // release locks and wait for a notify from the background collector
1402   // releasing the locks in only necessary for phases which
1403   // do yields to improve the granularity of the collection.
1404   assert_lock_strong(bitMapLock());
1405   // We need to lock the Free list lock for the space that we are
1406   // currently collecting.
1407   assert(haveFreelistLocks(), "Must be holding free list locks");
1408   bitMapLock()->unlock();
1409   releaseFreelistLocks();
1410   {
1411     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1412     if (_foregroundGCShouldWait) {
1413       // We are going to be waiting for action for the CMS thread;
1414       // it had better not be gone (for instance at shutdown)!
1415       assert(ConcurrentMarkSweepThread::cmst() != NULL,
1416              "CMS thread must be running");
1417       // Wait here until the background collector gives us the go-ahead
1418       ConcurrentMarkSweepThread::clear_CMS_flag(
1419         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
1420       // Get a possibly blocked CMS thread going:
1421       //   Note that we set _foregroundGCIsActive true above,
1422       //   without protection of the CGC_lock.
1423       CGC_lock->notify();
1424       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1425              "Possible deadlock");
1426       while (_foregroundGCShouldWait) {
1427         // wait for notification
1428         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1429         // Possibility of delay/starvation here, since CMS token does
1430         // not know to give priority to VM thread? Actually, i think
1431         // there wouldn't be any delay/starvation, but the proof of
1432         // that "fact" (?) appears non-trivial. XXX 20011219YSR
1433       }
1434       ConcurrentMarkSweepThread::set_CMS_flag(
1435         ConcurrentMarkSweepThread::CMS_vm_has_token);
1436     }
1437   }
1438   // The CMS_token is already held.  Get back the other locks.
1439   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1440          "VM thread should have CMS token");
1441   getFreelistLocks();
1442   bitMapLock()->lock_without_safepoint_check();
1443   log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1444                        p2i(Thread::current()), first_state);
1445   log_debug(gc, state)("    gets control with state %d", _collectorState);
1446 
1447   // Inform cms gen if this was due to partial collection failing.
1448   // The CMS gen may use this fact to determine its expansion policy.
1449   GenCollectedHeap* gch = GenCollectedHeap::heap();
1450   if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1451     assert(!_cmsGen->incremental_collection_failed(),
1452            "Should have been noticed, reacted to and cleared");
1453     _cmsGen->set_incremental_collection_failed();
1454   }
1455 
1456   if (first_state > Idling) {
1457     report_concurrent_mode_interruption();
1458   }
1459 
1460   set_did_compact(true);
1461 
1462   // If the collection is being acquired from the background
1463   // collector, there may be references on the discovered
1464   // references lists.  Abandon those references, since some
1465   // of them may have become unreachable after concurrent
1466   // discovery; the STW compacting collector will redo discovery
1467   // more precisely, without being subject to floating garbage.
1468   // Leaving otherwise unreachable references in the discovered
1469   // lists would require special handling.
1470   ref_processor()->disable_discovery();
1471   ref_processor()->abandon_partial_discovery();
1472   ref_processor()->verify_no_references_recorded();
1473 
1474   if (first_state > Idling) {
1475     save_heap_summary();
1476   }
1477 
1478   do_compaction_work(clear_all_soft_refs);
1479 
1480   // Has the GC time limit been exceeded?
1481   size_t max_eden_size = _young_gen->max_eden_size();
1482   GCCause::Cause gc_cause = gch->gc_cause();
1483   size_policy()->check_gc_overhead_limit(_young_gen->used(),
1484                                          _young_gen->eden()->used(),
1485                                          _cmsGen->max_capacity(),
1486                                          max_eden_size,
1487                                          full,
1488                                          gc_cause,
1489                                          gch->collector_policy());
1490 
1491   // Reset the expansion cause, now that we just completed
1492   // a collection cycle.
1493   clear_expansion_cause();
1494   _foregroundGCIsActive = false;
1495   return;
1496 }
1497 
1498 // Resize the tenured generation
1499 // after obtaining the free list locks for the
1500 // two generations.
1501 void CMSCollector::compute_new_size() {
1502   assert_locked_or_safepoint(Heap_lock);
1503   FreelistLocker z(this);
1504   MetaspaceGC::compute_new_size();
1505   _cmsGen->compute_new_size_free_list();
1506 }
1507 
1508 // A work method used by the foreground collector to do
1509 // a mark-sweep-compact.
1510 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1511   GenCollectedHeap* gch = GenCollectedHeap::heap();
1512 
1513   STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1514   gc_timer->register_gc_start();
1515 
1516   SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1517   gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1518 
1519   GCTraceTime(Trace, gc) t("CMS:MSC");
1520 
1521   // Temporarily widen the span of the weak reference processing to
1522   // the entire heap.
1523   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1524   ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1525   // Temporarily, clear the "is_alive_non_header" field of the
1526   // reference processor.
1527   ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1528   // Temporarily make reference _processing_ single threaded (non-MT).
1529   ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1530   // Temporarily make refs discovery atomic
1531   ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1532   // Temporarily make reference _discovery_ single threaded (non-MT)
1533   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1534 
1535   ref_processor()->set_enqueuing_is_done(false);
1536   ref_processor()->enable_discovery();
1537   ref_processor()->setup_policy(clear_all_soft_refs);
1538   // If an asynchronous collection finishes, the _modUnionTable is
1539   // all clear.  If we are assuming the collection from an asynchronous
1540   // collection, clear the _modUnionTable.
1541   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1542     "_modUnionTable should be clear if the baton was not passed");
1543   _modUnionTable.clear_all();
1544   assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1545     "mod union for klasses should be clear if the baton was passed");
1546   _ct->klass_rem_set()->clear_mod_union();
1547 
1548   // We must adjust the allocation statistics being maintained
1549   // in the free list space. We do so by reading and clearing
1550   // the sweep timer and updating the block flux rate estimates below.
1551   assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1552   if (_inter_sweep_timer.is_active()) {
1553     _inter_sweep_timer.stop();
1554     // Note that we do not use this sample to update the _inter_sweep_estimate.
1555     _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1556                                             _inter_sweep_estimate.padded_average(),
1557                                             _intra_sweep_estimate.padded_average());
1558   }
1559 
1560   GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1561   #ifdef ASSERT
1562     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1563     size_t free_size = cms_space->free();
1564     assert(free_size ==
1565            pointer_delta(cms_space->end(), cms_space->compaction_top())
1566            * HeapWordSize,
1567       "All the free space should be compacted into one chunk at top");
1568     assert(cms_space->dictionary()->total_chunk_size(
1569                                       debug_only(cms_space->freelistLock())) == 0 ||
1570            cms_space->totalSizeInIndexedFreeLists() == 0,
1571       "All the free space should be in a single chunk");
1572     size_t num = cms_space->totalCount();
1573     assert((free_size == 0 && num == 0) ||
1574            (free_size > 0  && (num == 1 || num == 2)),
1575          "There should be at most 2 free chunks after compaction");
1576   #endif // ASSERT
1577   _collectorState = Resetting;
1578   assert(_restart_addr == NULL,
1579          "Should have been NULL'd before baton was passed");
1580   reset_stw();
1581   _cmsGen->reset_after_compaction();
1582   _concurrent_cycles_since_last_unload = 0;
1583 
1584   // Clear any data recorded in the PLAB chunk arrays.
1585   if (_survivor_plab_array != NULL) {
1586     reset_survivor_plab_arrays();
1587   }
1588 
1589   // Adjust the per-size allocation stats for the next epoch.
1590   _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1591   // Restart the "inter sweep timer" for the next epoch.
1592   _inter_sweep_timer.reset();
1593   _inter_sweep_timer.start();
1594 
1595   gc_timer->register_gc_end();
1596 
1597   gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1598 
1599   // For a mark-sweep-compact, compute_new_size() will be called
1600   // in the heap's do_collection() method.
1601 }
1602 
1603 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1604   LogHandle(gc, heap) log;
1605   if (!log.is_trace()) {
1606     return;
1607   }
1608 
1609   ContiguousSpace* eden_space = _young_gen->eden();
1610   ContiguousSpace* from_space = _young_gen->from();
1611   ContiguousSpace* to_space   = _young_gen->to();
1612   // Eden
1613   if (_eden_chunk_array != NULL) {
1614     log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1615               p2i(eden_space->bottom()), p2i(eden_space->top()),
1616               p2i(eden_space->end()), eden_space->capacity());
1617     log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1618               _eden_chunk_index, _eden_chunk_capacity);
1619     for (size_t i = 0; i < _eden_chunk_index; i++) {
1620       log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1621     }
1622   }
1623   // Survivor
1624   if (_survivor_chunk_array != NULL) {
1625     log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1626               p2i(from_space->bottom()), p2i(from_space->top()),
1627               p2i(from_space->end()), from_space->capacity());
1628     log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1629               _survivor_chunk_index, _survivor_chunk_capacity);
1630     for (size_t i = 0; i < _survivor_chunk_index; i++) {
1631       log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1632     }
1633   }
1634 }
1635 
1636 void CMSCollector::getFreelistLocks() const {
1637   // Get locks for all free lists in all generations that this
1638   // collector is responsible for
1639   _cmsGen->freelistLock()->lock_without_safepoint_check();
1640 }
1641 
1642 void CMSCollector::releaseFreelistLocks() const {
1643   // Release locks for all free lists in all generations that this
1644   // collector is responsible for
1645   _cmsGen->freelistLock()->unlock();
1646 }
1647 
1648 bool CMSCollector::haveFreelistLocks() const {
1649   // Check locks for all free lists in all generations that this
1650   // collector is responsible for
1651   assert_lock_strong(_cmsGen->freelistLock());
1652   PRODUCT_ONLY(ShouldNotReachHere());
1653   return true;
1654 }
1655 
1656 // A utility class that is used by the CMS collector to
1657 // temporarily "release" the foreground collector from its
1658 // usual obligation to wait for the background collector to
1659 // complete an ongoing phase before proceeding.
1660 class ReleaseForegroundGC: public StackObj {
1661  private:
1662   CMSCollector* _c;
1663  public:
1664   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1665     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1666     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1667     // allow a potentially blocked foreground collector to proceed
1668     _c->_foregroundGCShouldWait = false;
1669     if (_c->_foregroundGCIsActive) {
1670       CGC_lock->notify();
1671     }
1672     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1673            "Possible deadlock");
1674   }
1675 
1676   ~ReleaseForegroundGC() {
1677     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1678     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1679     _c->_foregroundGCShouldWait = true;
1680   }
1681 };
1682 
1683 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1684   assert(Thread::current()->is_ConcurrentGC_thread(),
1685     "A CMS asynchronous collection is only allowed on a CMS thread.");
1686 
1687   GenCollectedHeap* gch = GenCollectedHeap::heap();
1688   {
1689     bool safepoint_check = Mutex::_no_safepoint_check_flag;
1690     MutexLockerEx hl(Heap_lock, safepoint_check);
1691     FreelistLocker fll(this);
1692     MutexLockerEx x(CGC_lock, safepoint_check);
1693     if (_foregroundGCIsActive) {
1694       // The foreground collector is. Skip this
1695       // background collection.
1696       assert(!_foregroundGCShouldWait, "Should be clear");
1697       return;
1698     } else {
1699       assert(_collectorState == Idling, "Should be idling before start.");
1700       _collectorState = InitialMarking;
1701       register_gc_start(cause);
1702       // Reset the expansion cause, now that we are about to begin
1703       // a new cycle.
1704       clear_expansion_cause();
1705 
1706       // Clear the MetaspaceGC flag since a concurrent collection
1707       // is starting but also clear it after the collection.
1708       MetaspaceGC::set_should_concurrent_collect(false);
1709     }
1710     // Decide if we want to enable class unloading as part of the
1711     // ensuing concurrent GC cycle.
1712     update_should_unload_classes();
1713     _full_gc_requested = false;           // acks all outstanding full gc requests
1714     _full_gc_cause = GCCause::_no_gc;
1715     // Signal that we are about to start a collection
1716     gch->increment_total_full_collections();  // ... starting a collection cycle
1717     _collection_count_start = gch->total_full_collections();
1718   }
1719 
1720   size_t prev_used = _cmsGen->used();
1721 
1722   // The change of the collection state is normally done at this level;
1723   // the exceptions are phases that are executed while the world is
1724   // stopped.  For those phases the change of state is done while the
1725   // world is stopped.  For baton passing purposes this allows the
1726   // background collector to finish the phase and change state atomically.
1727   // The foreground collector cannot wait on a phase that is done
1728   // while the world is stopped because the foreground collector already
1729   // has the world stopped and would deadlock.
1730   while (_collectorState != Idling) {
1731     log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1732                          p2i(Thread::current()), _collectorState);
1733     // The foreground collector
1734     //   holds the Heap_lock throughout its collection.
1735     //   holds the CMS token (but not the lock)
1736     //     except while it is waiting for the background collector to yield.
1737     //
1738     // The foreground collector should be blocked (not for long)
1739     //   if the background collector is about to start a phase
1740     //   executed with world stopped.  If the background
1741     //   collector has already started such a phase, the
1742     //   foreground collector is blocked waiting for the
1743     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
1744     //   are executed in the VM thread.
1745     //
1746     // The locking order is
1747     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
1748     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
1749     //   CMS token  (claimed in
1750     //                stop_world_and_do() -->
1751     //                  safepoint_synchronize() -->
1752     //                    CMSThread::synchronize())
1753 
1754     {
1755       // Check if the FG collector wants us to yield.
1756       CMSTokenSync x(true); // is cms thread
1757       if (waitForForegroundGC()) {
1758         // We yielded to a foreground GC, nothing more to be
1759         // done this round.
1760         assert(_foregroundGCShouldWait == false, "We set it to false in "
1761                "waitForForegroundGC()");
1762         log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1763                              p2i(Thread::current()), _collectorState);
1764         return;
1765       } else {
1766         // The background collector can run but check to see if the
1767         // foreground collector has done a collection while the
1768         // background collector was waiting to get the CGC_lock
1769         // above.  If yes, break so that _foregroundGCShouldWait
1770         // is cleared before returning.
1771         if (_collectorState == Idling) {
1772           break;
1773         }
1774       }
1775     }
1776 
1777     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1778       "should be waiting");
1779 
1780     switch (_collectorState) {
1781       case InitialMarking:
1782         {
1783           ReleaseForegroundGC x(this);
1784           stats().record_cms_begin();
1785           VM_CMS_Initial_Mark initial_mark_op(this);
1786           VMThread::execute(&initial_mark_op);
1787         }
1788         // The collector state may be any legal state at this point
1789         // since the background collector may have yielded to the
1790         // foreground collector.
1791         break;
1792       case Marking:
1793         // initial marking in checkpointRootsInitialWork has been completed
1794         if (markFromRoots()) { // we were successful
1795           assert(_collectorState == Precleaning, "Collector state should "
1796             "have changed");
1797         } else {
1798           assert(_foregroundGCIsActive, "Internal state inconsistency");
1799         }
1800         break;
1801       case Precleaning:
1802         // marking from roots in markFromRoots has been completed
1803         preclean();
1804         assert(_collectorState == AbortablePreclean ||
1805                _collectorState == FinalMarking,
1806                "Collector state should have changed");
1807         break;
1808       case AbortablePreclean:
1809         abortable_preclean();
1810         assert(_collectorState == FinalMarking, "Collector state should "
1811           "have changed");
1812         break;
1813       case FinalMarking:
1814         {
1815           ReleaseForegroundGC x(this);
1816 
1817           VM_CMS_Final_Remark final_remark_op(this);
1818           VMThread::execute(&final_remark_op);
1819         }
1820         assert(_foregroundGCShouldWait, "block post-condition");
1821         break;
1822       case Sweeping:
1823         // final marking in checkpointRootsFinal has been completed
1824         sweep();
1825         assert(_collectorState == Resizing, "Collector state change "
1826           "to Resizing must be done under the free_list_lock");
1827 
1828       case Resizing: {
1829         // Sweeping has been completed...
1830         // At this point the background collection has completed.
1831         // Don't move the call to compute_new_size() down
1832         // into code that might be executed if the background
1833         // collection was preempted.
1834         {
1835           ReleaseForegroundGC x(this);   // unblock FG collection
1836           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
1837           CMSTokenSync        z(true);   // not strictly needed.
1838           if (_collectorState == Resizing) {
1839             compute_new_size();
1840             save_heap_summary();
1841             _collectorState = Resetting;
1842           } else {
1843             assert(_collectorState == Idling, "The state should only change"
1844                    " because the foreground collector has finished the collection");
1845           }
1846         }
1847         break;
1848       }
1849       case Resetting:
1850         // CMS heap resizing has been completed
1851         reset_concurrent();
1852         assert(_collectorState == Idling, "Collector state should "
1853           "have changed");
1854 
1855         MetaspaceGC::set_should_concurrent_collect(false);
1856 
1857         stats().record_cms_end();
1858         // Don't move the concurrent_phases_end() and compute_new_size()
1859         // calls to here because a preempted background collection
1860         // has it's state set to "Resetting".
1861         break;
1862       case Idling:
1863       default:
1864         ShouldNotReachHere();
1865         break;
1866     }
1867     log_debug(gc, state)("  Thread " INTPTR_FORMAT " done - next CMS state %d",
1868                          p2i(Thread::current()), _collectorState);
1869     assert(_foregroundGCShouldWait, "block post-condition");
1870   }
1871 
1872   // Should this be in gc_epilogue?
1873   collector_policy()->counters()->update_counters();
1874 
1875   {
1876     // Clear _foregroundGCShouldWait and, in the event that the
1877     // foreground collector is waiting, notify it, before
1878     // returning.
1879     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1880     _foregroundGCShouldWait = false;
1881     if (_foregroundGCIsActive) {
1882       CGC_lock->notify();
1883     }
1884     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1885            "Possible deadlock");
1886   }
1887   log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1888                        p2i(Thread::current()), _collectorState);
1889   log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K("  SIZE_FORMAT "K)",
1890                      prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1891 }
1892 
1893 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1894   _cms_start_registered = true;
1895   _gc_timer_cm->register_gc_start();
1896   _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1897 }
1898 
1899 void CMSCollector::register_gc_end() {
1900   if (_cms_start_registered) {
1901     report_heap_summary(GCWhen::AfterGC);
1902 
1903     _gc_timer_cm->register_gc_end();
1904     _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1905     _cms_start_registered = false;
1906   }
1907 }
1908 
1909 void CMSCollector::save_heap_summary() {
1910   GenCollectedHeap* gch = GenCollectedHeap::heap();
1911   _last_heap_summary = gch->create_heap_summary();
1912   _last_metaspace_summary = gch->create_metaspace_summary();
1913 }
1914 
1915 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1916   _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1917   _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1918 }
1919 
1920 bool CMSCollector::waitForForegroundGC() {
1921   bool res = false;
1922   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1923          "CMS thread should have CMS token");
1924   // Block the foreground collector until the
1925   // background collectors decides whether to
1926   // yield.
1927   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1928   _foregroundGCShouldWait = true;
1929   if (_foregroundGCIsActive) {
1930     // The background collector yields to the
1931     // foreground collector and returns a value
1932     // indicating that it has yielded.  The foreground
1933     // collector can proceed.
1934     res = true;
1935     _foregroundGCShouldWait = false;
1936     ConcurrentMarkSweepThread::clear_CMS_flag(
1937       ConcurrentMarkSweepThread::CMS_cms_has_token);
1938     ConcurrentMarkSweepThread::set_CMS_flag(
1939       ConcurrentMarkSweepThread::CMS_cms_wants_token);
1940     // Get a possibly blocked foreground thread going
1941     CGC_lock->notify();
1942     log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1943                          p2i(Thread::current()), _collectorState);
1944     while (_foregroundGCIsActive) {
1945       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1946     }
1947     ConcurrentMarkSweepThread::set_CMS_flag(
1948       ConcurrentMarkSweepThread::CMS_cms_has_token);
1949     ConcurrentMarkSweepThread::clear_CMS_flag(
1950       ConcurrentMarkSweepThread::CMS_cms_wants_token);
1951   }
1952   log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1953                        p2i(Thread::current()), _collectorState);
1954   return res;
1955 }
1956 
1957 // Because of the need to lock the free lists and other structures in
1958 // the collector, common to all the generations that the collector is
1959 // collecting, we need the gc_prologues of individual CMS generations
1960 // delegate to their collector. It may have been simpler had the
1961 // current infrastructure allowed one to call a prologue on a
1962 // collector. In the absence of that we have the generation's
1963 // prologue delegate to the collector, which delegates back
1964 // some "local" work to a worker method in the individual generations
1965 // that it's responsible for collecting, while itself doing any
1966 // work common to all generations it's responsible for. A similar
1967 // comment applies to the  gc_epilogue()'s.
1968 // The role of the variable _between_prologue_and_epilogue is to
1969 // enforce the invocation protocol.
1970 void CMSCollector::gc_prologue(bool full) {
1971   // Call gc_prologue_work() for the CMSGen
1972   // we are responsible for.
1973 
1974   // The following locking discipline assumes that we are only called
1975   // when the world is stopped.
1976   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
1977 
1978   // The CMSCollector prologue must call the gc_prologues for the
1979   // "generations" that it's responsible
1980   // for.
1981 
1982   assert(   Thread::current()->is_VM_thread()
1983          || (   CMSScavengeBeforeRemark
1984              && Thread::current()->is_ConcurrentGC_thread()),
1985          "Incorrect thread type for prologue execution");
1986 
1987   if (_between_prologue_and_epilogue) {
1988     // We have already been invoked; this is a gc_prologue delegation
1989     // from yet another CMS generation that we are responsible for, just
1990     // ignore it since all relevant work has already been done.
1991     return;
1992   }
1993 
1994   // set a bit saying prologue has been called; cleared in epilogue
1995   _between_prologue_and_epilogue = true;
1996   // Claim locks for common data structures, then call gc_prologue_work()
1997   // for each CMSGen.
1998 
1999   getFreelistLocks();   // gets free list locks on constituent spaces
2000   bitMapLock()->lock_without_safepoint_check();
2001 
2002   // Should call gc_prologue_work() for all cms gens we are responsible for
2003   bool duringMarking =    _collectorState >= Marking
2004                          && _collectorState < Sweeping;
2005 
2006   // The young collections clear the modified oops state, which tells if
2007   // there are any modified oops in the class. The remark phase also needs
2008   // that information. Tell the young collection to save the union of all
2009   // modified klasses.
2010   if (duringMarking) {
2011     _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2012   }
2013 
2014   bool registerClosure = duringMarking;
2015 
2016   _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2017 
2018   if (!full) {
2019     stats().record_gc0_begin();
2020   }
2021 }
2022 
2023 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2024 
2025   _capacity_at_prologue = capacity();
2026   _used_at_prologue = used();
2027 
2028   // Delegate to CMScollector which knows how to coordinate between
2029   // this and any other CMS generations that it is responsible for
2030   // collecting.
2031   collector()->gc_prologue(full);
2032 }
2033 
2034 // This is a "private" interface for use by this generation's CMSCollector.
2035 // Not to be called directly by any other entity (for instance,
2036 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2037 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2038   bool registerClosure, ModUnionClosure* modUnionClosure) {
2039   assert(!incremental_collection_failed(), "Shouldn't be set yet");
2040   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2041     "Should be NULL");
2042   if (registerClosure) {
2043     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2044   }
2045   cmsSpace()->gc_prologue();
2046   // Clear stat counters
2047   NOT_PRODUCT(
2048     assert(_numObjectsPromoted == 0, "check");
2049     assert(_numWordsPromoted   == 0, "check");
2050     log_develop(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2051                                  _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2052     _numObjectsAllocated = 0;
2053     _numWordsAllocated   = 0;
2054   )
2055 }
2056 
2057 void CMSCollector::gc_epilogue(bool full) {
2058   // The following locking discipline assumes that we are only called
2059   // when the world is stopped.
2060   assert(SafepointSynchronize::is_at_safepoint(),
2061          "world is stopped assumption");
2062 
2063   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2064   // if linear allocation blocks need to be appropriately marked to allow the
2065   // the blocks to be parsable. We also check here whether we need to nudge the
2066   // CMS collector thread to start a new cycle (if it's not already active).
2067   assert(   Thread::current()->is_VM_thread()
2068          || (   CMSScavengeBeforeRemark
2069              && Thread::current()->is_ConcurrentGC_thread()),
2070          "Incorrect thread type for epilogue execution");
2071 
2072   if (!_between_prologue_and_epilogue) {
2073     // We have already been invoked; this is a gc_epilogue delegation
2074     // from yet another CMS generation that we are responsible for, just
2075     // ignore it since all relevant work has already been done.
2076     return;
2077   }
2078   assert(haveFreelistLocks(), "must have freelist locks");
2079   assert_lock_strong(bitMapLock());
2080 
2081   _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2082 
2083   _cmsGen->gc_epilogue_work(full);
2084 
2085   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2086     // in case sampling was not already enabled, enable it
2087     _start_sampling = true;
2088   }
2089   // reset _eden_chunk_array so sampling starts afresh
2090   _eden_chunk_index = 0;
2091 
2092   size_t cms_used   = _cmsGen->cmsSpace()->used();
2093 
2094   // update performance counters - this uses a special version of
2095   // update_counters() that allows the utilization to be passed as a
2096   // parameter, avoiding multiple calls to used().
2097   //
2098   _cmsGen->update_counters(cms_used);
2099 
2100   bitMapLock()->unlock();
2101   releaseFreelistLocks();
2102 
2103   if (!CleanChunkPoolAsync) {
2104     Chunk::clean_chunk_pool();
2105   }
2106 
2107   set_did_compact(false);
2108   _between_prologue_and_epilogue = false;  // ready for next cycle
2109 }
2110 
2111 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2112   collector()->gc_epilogue(full);
2113 
2114   // Also reset promotion tracking in par gc thread states.
2115   for (uint i = 0; i < ParallelGCThreads; i++) {
2116     _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2117   }
2118 }
2119 
2120 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2121   assert(!incremental_collection_failed(), "Should have been cleared");
2122   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2123   cmsSpace()->gc_epilogue();
2124     // Print stat counters
2125   NOT_PRODUCT(
2126     assert(_numObjectsAllocated == 0, "check");
2127     assert(_numWordsAllocated == 0, "check");
2128     log_develop(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2129                                      _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2130     _numObjectsPromoted = 0;
2131     _numWordsPromoted   = 0;
2132   )
2133 
2134   // Call down the chain in contiguous_available needs the freelistLock
2135   // so print this out before releasing the freeListLock.
2136   log_develop(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2137 }
2138 
2139 #ifndef PRODUCT
2140 bool CMSCollector::have_cms_token() {
2141   Thread* thr = Thread::current();
2142   if (thr->is_VM_thread()) {
2143     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2144   } else if (thr->is_ConcurrentGC_thread()) {
2145     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2146   } else if (thr->is_GC_task_thread()) {
2147     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2148            ParGCRareEvent_lock->owned_by_self();
2149   }
2150   return false;
2151 }
2152 #endif
2153 
2154 void
2155 CMSCollector::print_on_error(outputStream* st) {
2156   CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2157   if (collector != NULL) {
2158     CMSBitMap* bitmap = &collector->_markBitMap;
2159     st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2160     bitmap->print_on_error(st, " Bits: ");
2161 
2162     st->cr();
2163 
2164     CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2165     st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2166     mut_bitmap->print_on_error(st, " Bits: ");
2167   }
2168 }
2169 
2170 ////////////////////////////////////////////////////////
2171 // CMS Verification Support
2172 ////////////////////////////////////////////////////////
2173 // Following the remark phase, the following invariant
2174 // should hold -- each object in the CMS heap which is
2175 // marked in markBitMap() should be marked in the verification_mark_bm().
2176 
2177 class VerifyMarkedClosure: public BitMapClosure {
2178   CMSBitMap* _marks;
2179   bool       _failed;
2180 
2181  public:
2182   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2183 
2184   bool do_bit(size_t offset) {
2185     HeapWord* addr = _marks->offsetToHeapWord(offset);
2186     if (!_marks->isMarked(addr)) {
2187       LogHandle(gc, verify) log;
2188       ResourceMark rm;
2189       oop(addr)->print_on(log.info_stream());
2190       log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2191       _failed = true;
2192     }
2193     return true;
2194   }
2195 
2196   bool failed() { return _failed; }
2197 };
2198 
2199 bool CMSCollector::verify_after_remark() {
2200   GCTraceTime(Info, gc, verify) tm("Verifying CMS Marking.");
2201   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2202   static bool init = false;
2203 
2204   assert(SafepointSynchronize::is_at_safepoint(),
2205          "Else mutations in object graph will make answer suspect");
2206   assert(have_cms_token(),
2207          "Else there may be mutual interference in use of "
2208          " verification data structures");
2209   assert(_collectorState > Marking && _collectorState <= Sweeping,
2210          "Else marking info checked here may be obsolete");
2211   assert(haveFreelistLocks(), "must hold free list locks");
2212   assert_lock_strong(bitMapLock());
2213 
2214 
2215   // Allocate marking bit map if not already allocated
2216   if (!init) { // first time
2217     if (!verification_mark_bm()->allocate(_span)) {
2218       return false;
2219     }
2220     init = true;
2221   }
2222 
2223   assert(verification_mark_stack()->isEmpty(), "Should be empty");
2224 
2225   // Turn off refs discovery -- so we will be tracing through refs.
2226   // This is as intended, because by this time
2227   // GC must already have cleared any refs that need to be cleared,
2228   // and traced those that need to be marked; moreover,
2229   // the marking done here is not going to interfere in any
2230   // way with the marking information used by GC.
2231   NoRefDiscovery no_discovery(ref_processor());
2232 
2233 #if defined(COMPILER2) || INCLUDE_JVMCI
2234   DerivedPointerTableDeactivate dpt_deact;
2235 #endif
2236 
2237   // Clear any marks from a previous round
2238   verification_mark_bm()->clear_all();
2239   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2240   verify_work_stacks_empty();
2241 
2242   GenCollectedHeap* gch = GenCollectedHeap::heap();
2243   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2244   // Update the saved marks which may affect the root scans.
2245   gch->save_marks();
2246 
2247   if (CMSRemarkVerifyVariant == 1) {
2248     // In this first variant of verification, we complete
2249     // all marking, then check if the new marks-vector is
2250     // a subset of the CMS marks-vector.
2251     verify_after_remark_work_1();
2252   } else if (CMSRemarkVerifyVariant == 2) {
2253     // In this second variant of verification, we flag an error
2254     // (i.e. an object reachable in the new marks-vector not reachable
2255     // in the CMS marks-vector) immediately, also indicating the
2256     // identify of an object (A) that references the unmarked object (B) --
2257     // presumably, a mutation to A failed to be picked up by preclean/remark?
2258     verify_after_remark_work_2();
2259   } else {
2260     warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2261             CMSRemarkVerifyVariant);
2262   }
2263   return true;
2264 }
2265 
2266 void CMSCollector::verify_after_remark_work_1() {
2267   ResourceMark rm;
2268   HandleMark  hm;
2269   GenCollectedHeap* gch = GenCollectedHeap::heap();
2270 
2271   // Get a clear set of claim bits for the roots processing to work with.
2272   ClassLoaderDataGraph::clear_claimed_marks();
2273 
2274   // Mark from roots one level into CMS
2275   MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2276   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2277 
2278   {
2279     StrongRootsScope srs(1);
2280 
2281     gch->gen_process_roots(&srs,
2282                            GenCollectedHeap::OldGen,
2283                            true,   // young gen as roots
2284                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2285                            should_unload_classes(),
2286                            &notOlder,
2287                            NULL,
2288                            NULL);
2289   }
2290 
2291   // Now mark from the roots
2292   MarkFromRootsClosure markFromRootsClosure(this, _span,
2293     verification_mark_bm(), verification_mark_stack(),
2294     false /* don't yield */, true /* verifying */);
2295   assert(_restart_addr == NULL, "Expected pre-condition");
2296   verification_mark_bm()->iterate(&markFromRootsClosure);
2297   while (_restart_addr != NULL) {
2298     // Deal with stack overflow: by restarting at the indicated
2299     // address.
2300     HeapWord* ra = _restart_addr;
2301     markFromRootsClosure.reset(ra);
2302     _restart_addr = NULL;
2303     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2304   }
2305   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2306   verify_work_stacks_empty();
2307 
2308   // Marking completed -- now verify that each bit marked in
2309   // verification_mark_bm() is also marked in markBitMap(); flag all
2310   // errors by printing corresponding objects.
2311   VerifyMarkedClosure vcl(markBitMap());
2312   verification_mark_bm()->iterate(&vcl);
2313   if (vcl.failed()) {
2314     LogHandle(gc, verify) log;
2315     log.info("Verification failed");
2316     ResourceMark rm;
2317     gch->print_on(log.info_stream());
2318     fatal("CMS: failed marking verification after remark");
2319   }
2320 }
2321 
2322 class VerifyKlassOopsKlassClosure : public KlassClosure {
2323   class VerifyKlassOopsClosure : public OopClosure {
2324     CMSBitMap* _bitmap;
2325    public:
2326     VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2327     void do_oop(oop* p)       { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2328     void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2329   } _oop_closure;
2330  public:
2331   VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2332   void do_klass(Klass* k) {
2333     k->oops_do(&_oop_closure);
2334   }
2335 };
2336 
2337 void CMSCollector::verify_after_remark_work_2() {
2338   ResourceMark rm;
2339   HandleMark  hm;
2340   GenCollectedHeap* gch = GenCollectedHeap::heap();
2341 
2342   // Get a clear set of claim bits for the roots processing to work with.
2343   ClassLoaderDataGraph::clear_claimed_marks();
2344 
2345   // Mark from roots one level into CMS
2346   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2347                                      markBitMap());
2348   CLDToOopClosure cld_closure(&notOlder, true);
2349 
2350   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2351 
2352   {
2353     StrongRootsScope srs(1);
2354 
2355     gch->gen_process_roots(&srs,
2356                            GenCollectedHeap::OldGen,
2357                            true,   // young gen as roots
2358                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
2359                            should_unload_classes(),
2360                            &notOlder,
2361                            NULL,
2362                            &cld_closure);
2363   }
2364 
2365   // Now mark from the roots
2366   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2367     verification_mark_bm(), markBitMap(), verification_mark_stack());
2368   assert(_restart_addr == NULL, "Expected pre-condition");
2369   verification_mark_bm()->iterate(&markFromRootsClosure);
2370   while (_restart_addr != NULL) {
2371     // Deal with stack overflow: by restarting at the indicated
2372     // address.
2373     HeapWord* ra = _restart_addr;
2374     markFromRootsClosure.reset(ra);
2375     _restart_addr = NULL;
2376     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2377   }
2378   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2379   verify_work_stacks_empty();
2380 
2381   VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2382   ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2383 
2384   // Marking completed -- now verify that each bit marked in
2385   // verification_mark_bm() is also marked in markBitMap(); flag all
2386   // errors by printing corresponding objects.
2387   VerifyMarkedClosure vcl(markBitMap());
2388   verification_mark_bm()->iterate(&vcl);
2389   assert(!vcl.failed(), "Else verification above should not have succeeded");
2390 }
2391 
2392 void ConcurrentMarkSweepGeneration::save_marks() {
2393   // delegate to CMS space
2394   cmsSpace()->save_marks();
2395   for (uint i = 0; i < ParallelGCThreads; i++) {
2396     _par_gc_thread_states[i]->promo.startTrackingPromotions();
2397   }
2398 }
2399 
2400 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2401   return cmsSpace()->no_allocs_since_save_marks();
2402 }
2403 
2404 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
2405                                                                 \
2406 void ConcurrentMarkSweepGeneration::                            \
2407 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
2408   cl->set_generation(this);                                     \
2409   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
2410   cl->reset_generation();                                       \
2411   save_marks();                                                 \
2412 }
2413 
2414 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2415 
2416 void
2417 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2418   if (freelistLock()->owned_by_self()) {
2419     Generation::oop_iterate(cl);
2420   } else {
2421     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2422     Generation::oop_iterate(cl);
2423   }
2424 }
2425 
2426 void
2427 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2428   if (freelistLock()->owned_by_self()) {
2429     Generation::object_iterate(cl);
2430   } else {
2431     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2432     Generation::object_iterate(cl);
2433   }
2434 }
2435 
2436 void
2437 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2438   if (freelistLock()->owned_by_self()) {
2439     Generation::safe_object_iterate(cl);
2440   } else {
2441     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2442     Generation::safe_object_iterate(cl);
2443   }
2444 }
2445 
2446 void
2447 ConcurrentMarkSweepGeneration::post_compact() {
2448 }
2449 
2450 void
2451 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2452   // Fix the linear allocation blocks to look like free blocks.
2453 
2454   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2455   // are not called when the heap is verified during universe initialization and
2456   // at vm shutdown.
2457   if (freelistLock()->owned_by_self()) {
2458     cmsSpace()->prepare_for_verify();
2459   } else {
2460     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2461     cmsSpace()->prepare_for_verify();
2462   }
2463 }
2464 
2465 void
2466 ConcurrentMarkSweepGeneration::verify() {
2467   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2468   // are not called when the heap is verified during universe initialization and
2469   // at vm shutdown.
2470   if (freelistLock()->owned_by_self()) {
2471     cmsSpace()->verify();
2472   } else {
2473     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2474     cmsSpace()->verify();
2475   }
2476 }
2477 
2478 void CMSCollector::verify() {
2479   _cmsGen->verify();
2480 }
2481 
2482 #ifndef PRODUCT
2483 bool CMSCollector::overflow_list_is_empty() const {
2484   assert(_num_par_pushes >= 0, "Inconsistency");
2485   if (_overflow_list == NULL) {
2486     assert(_num_par_pushes == 0, "Inconsistency");
2487   }
2488   return _overflow_list == NULL;
2489 }
2490 
2491 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2492 // merely consolidate assertion checks that appear to occur together frequently.
2493 void CMSCollector::verify_work_stacks_empty() const {
2494   assert(_markStack.isEmpty(), "Marking stack should be empty");
2495   assert(overflow_list_is_empty(), "Overflow list should be empty");
2496 }
2497 
2498 void CMSCollector::verify_overflow_empty() const {
2499   assert(overflow_list_is_empty(), "Overflow list should be empty");
2500   assert(no_preserved_marks(), "No preserved marks");
2501 }
2502 #endif // PRODUCT
2503 
2504 // Decide if we want to enable class unloading as part of the
2505 // ensuing concurrent GC cycle. We will collect and
2506 // unload classes if it's the case that:
2507 // (1) an explicit gc request has been made and the flag
2508 //     ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2509 // (2) (a) class unloading is enabled at the command line, and
2510 //     (b) old gen is getting really full
2511 // NOTE: Provided there is no change in the state of the heap between
2512 // calls to this method, it should have idempotent results. Moreover,
2513 // its results should be monotonically increasing (i.e. going from 0 to 1,
2514 // but not 1 to 0) between successive calls between which the heap was
2515 // not collected. For the implementation below, it must thus rely on
2516 // the property that concurrent_cycles_since_last_unload()
2517 // will not decrease unless a collection cycle happened and that
2518 // _cmsGen->is_too_full() are
2519 // themselves also monotonic in that sense. See check_monotonicity()
2520 // below.
2521 void CMSCollector::update_should_unload_classes() {
2522   _should_unload_classes = false;
2523   // Condition 1 above
2524   if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2525     _should_unload_classes = true;
2526   } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2527     // Disjuncts 2.b.(i,ii,iii) above
2528     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2529                               CMSClassUnloadingMaxInterval)
2530                            || _cmsGen->is_too_full();
2531   }
2532 }
2533 
2534 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2535   bool res = should_concurrent_collect();
2536   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2537   return res;
2538 }
2539 
2540 void CMSCollector::setup_cms_unloading_and_verification_state() {
2541   const  bool should_verify =   VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2542                              || VerifyBeforeExit;
2543   const  int  rso           =   GenCollectedHeap::SO_AllCodeCache;
2544 
2545   // We set the proper root for this CMS cycle here.
2546   if (should_unload_classes()) {   // Should unload classes this cycle
2547     remove_root_scanning_option(rso);  // Shrink the root set appropriately
2548     set_verifying(should_verify);    // Set verification state for this cycle
2549     return;                            // Nothing else needs to be done at this time
2550   }
2551 
2552   // Not unloading classes this cycle
2553   assert(!should_unload_classes(), "Inconsistency!");
2554 
2555   // If we are not unloading classes then add SO_AllCodeCache to root
2556   // scanning options.
2557   add_root_scanning_option(rso);
2558 
2559   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2560     set_verifying(true);
2561   } else if (verifying() && !should_verify) {
2562     // We were verifying, but some verification flags got disabled.
2563     set_verifying(false);
2564     // Exclude symbols, strings and code cache elements from root scanning to
2565     // reduce IM and RM pauses.
2566     remove_root_scanning_option(rso);
2567   }
2568 }
2569 
2570 
2571 #ifndef PRODUCT
2572 HeapWord* CMSCollector::block_start(const void* p) const {
2573   const HeapWord* addr = (HeapWord*)p;
2574   if (_span.contains(p)) {
2575     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2576       return _cmsGen->cmsSpace()->block_start(p);
2577     }
2578   }
2579   return NULL;
2580 }
2581 #endif
2582 
2583 HeapWord*
2584 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2585                                                    bool   tlab,
2586                                                    bool   parallel) {
2587   CMSSynchronousYieldRequest yr;
2588   assert(!tlab, "Can't deal with TLAB allocation");
2589   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2590   expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2591   if (GCExpandToAllocateDelayMillis > 0) {
2592     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2593   }
2594   return have_lock_and_allocate(word_size, tlab);
2595 }
2596 
2597 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2598     size_t bytes,
2599     size_t expand_bytes,
2600     CMSExpansionCause::Cause cause)
2601 {
2602 
2603   bool success = expand(bytes, expand_bytes);
2604 
2605   // remember why we expanded; this information is used
2606   // by shouldConcurrentCollect() when making decisions on whether to start
2607   // a new CMS cycle.
2608   if (success) {
2609     set_expansion_cause(cause);
2610     log_trace(gc)("Expanded CMS gen for %s",  CMSExpansionCause::to_string(cause));
2611   }
2612 }
2613 
2614 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2615   HeapWord* res = NULL;
2616   MutexLocker x(ParGCRareEvent_lock);
2617   while (true) {
2618     // Expansion by some other thread might make alloc OK now:
2619     res = ps->lab.alloc(word_sz);
2620     if (res != NULL) return res;
2621     // If there's not enough expansion space available, give up.
2622     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2623       return NULL;
2624     }
2625     // Otherwise, we try expansion.
2626     expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2627     // Now go around the loop and try alloc again;
2628     // A competing par_promote might beat us to the expansion space,
2629     // so we may go around the loop again if promotion fails again.
2630     if (GCExpandToAllocateDelayMillis > 0) {
2631       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2632     }
2633   }
2634 }
2635 
2636 
2637 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2638   PromotionInfo* promo) {
2639   MutexLocker x(ParGCRareEvent_lock);
2640   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2641   while (true) {
2642     // Expansion by some other thread might make alloc OK now:
2643     if (promo->ensure_spooling_space()) {
2644       assert(promo->has_spooling_space(),
2645              "Post-condition of successful ensure_spooling_space()");
2646       return true;
2647     }
2648     // If there's not enough expansion space available, give up.
2649     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2650       return false;
2651     }
2652     // Otherwise, we try expansion.
2653     expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2654     // Now go around the loop and try alloc again;
2655     // A competing allocation might beat us to the expansion space,
2656     // so we may go around the loop again if allocation fails again.
2657     if (GCExpandToAllocateDelayMillis > 0) {
2658       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2659     }
2660   }
2661 }
2662 
2663 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2664   // Only shrink if a compaction was done so that all the free space
2665   // in the generation is in a contiguous block at the end.
2666   if (did_compact()) {
2667     CardGeneration::shrink(bytes);
2668   }
2669 }
2670 
2671 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2672   assert_locked_or_safepoint(Heap_lock);
2673 }
2674 
2675 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2676   assert_locked_or_safepoint(Heap_lock);
2677   assert_lock_strong(freelistLock());
2678   log_trace(gc)("Shrinking of CMS not yet implemented");
2679   return;
2680 }
2681 
2682 
2683 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2684 // phases.
2685 class CMSPhaseAccounting: public StackObj {
2686  public:
2687   CMSPhaseAccounting(CMSCollector *collector,
2688                      const char *title);
2689   ~CMSPhaseAccounting();
2690 
2691  private:
2692   CMSCollector *_collector;
2693   const char *_title;
2694   GCTraceConcTime(Info, gc) _trace_time;
2695 
2696  public:
2697   // Not MT-safe; so do not pass around these StackObj's
2698   // where they may be accessed by other threads.
2699   jlong wallclock_millis() {
2700     return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2701   }
2702 };
2703 
2704 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2705                                        const char *title) :
2706   _collector(collector), _title(title), _trace_time(title) {
2707 
2708   _collector->resetYields();
2709   _collector->resetTimer();
2710   _collector->startTimer();
2711 }
2712 
2713 CMSPhaseAccounting::~CMSPhaseAccounting() {
2714   _collector->stopTimer();
2715   jlong end_counter = os::elapsed_counter();
2716   log_debug(gc)("Concurrent active time: %.3fms", _collector->timerValue());
2717   log_debug(gc, stats)(" (CMS %s yielded %d times)", _title, _collector->yields());
2718 }
2719 
2720 // CMS work
2721 
2722 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2723 class CMSParMarkTask : public AbstractGangTask {
2724  protected:
2725   CMSCollector*     _collector;
2726   uint              _n_workers;
2727   CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2728       AbstractGangTask(name),
2729       _collector(collector),
2730       _n_workers(n_workers) {}
2731   // Work method in support of parallel rescan ... of young gen spaces
2732   void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2733                              ContiguousSpace* space,
2734                              HeapWord** chunk_array, size_t chunk_top);
2735   void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2736 };
2737 
2738 // Parallel initial mark task
2739 class CMSParInitialMarkTask: public CMSParMarkTask {
2740   StrongRootsScope* _strong_roots_scope;
2741  public:
2742   CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2743       CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2744       _strong_roots_scope(strong_roots_scope) {}
2745   void work(uint worker_id);
2746 };
2747 
2748 // Checkpoint the roots into this generation from outside
2749 // this generation. [Note this initial checkpoint need only
2750 // be approximate -- we'll do a catch up phase subsequently.]
2751 void CMSCollector::checkpointRootsInitial() {
2752   assert(_collectorState == InitialMarking, "Wrong collector state");
2753   check_correct_thread_executing();
2754   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2755 
2756   save_heap_summary();
2757   report_heap_summary(GCWhen::BeforeGC);
2758 
2759   ReferenceProcessor* rp = ref_processor();
2760   assert(_restart_addr == NULL, "Control point invariant");
2761   {
2762     // acquire locks for subsequent manipulations
2763     MutexLockerEx x(bitMapLock(),
2764                     Mutex::_no_safepoint_check_flag);
2765     checkpointRootsInitialWork();
2766     // enable ("weak") refs discovery
2767     rp->enable_discovery();
2768     _collectorState = Marking;
2769   }
2770 }
2771 
2772 void CMSCollector::checkpointRootsInitialWork() {
2773   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2774   assert(_collectorState == InitialMarking, "just checking");
2775 
2776   // Already have locks.
2777   assert_lock_strong(bitMapLock());
2778   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2779 
2780   // Setup the verification and class unloading state for this
2781   // CMS collection cycle.
2782   setup_cms_unloading_and_verification_state();
2783 
2784   GCTraceTime(Trace, gc) ts("checkpointRootsInitialWork", _gc_timer_cm);
2785 
2786   // Reset all the PLAB chunk arrays if necessary.
2787   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2788     reset_survivor_plab_arrays();
2789   }
2790 
2791   ResourceMark rm;
2792   HandleMark  hm;
2793 
2794   MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2795   GenCollectedHeap* gch = GenCollectedHeap::heap();
2796 
2797   verify_work_stacks_empty();
2798   verify_overflow_empty();
2799 
2800   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2801   // Update the saved marks which may affect the root scans.
2802   gch->save_marks();
2803 
2804   // weak reference processing has not started yet.
2805   ref_processor()->set_enqueuing_is_done(false);
2806 
2807   // Need to remember all newly created CLDs,
2808   // so that we can guarantee that the remark finds them.
2809   ClassLoaderDataGraph::remember_new_clds(true);
2810 
2811   // Whenever a CLD is found, it will be claimed before proceeding to mark
2812   // the klasses. The claimed marks need to be cleared before marking starts.
2813   ClassLoaderDataGraph::clear_claimed_marks();
2814 
2815   print_eden_and_survivor_chunk_arrays();
2816 
2817   {
2818 #if defined(COMPILER2) || INCLUDE_JVMCI
2819     DerivedPointerTableDeactivate dpt_deact;
2820 #endif
2821     if (CMSParallelInitialMarkEnabled) {
2822       // The parallel version.
2823       WorkGang* workers = gch->workers();
2824       assert(workers != NULL, "Need parallel worker threads.");
2825       uint n_workers = workers->active_workers();
2826 
2827       StrongRootsScope srs(n_workers);
2828 
2829       CMSParInitialMarkTask tsk(this, &srs, n_workers);
2830       initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2831       if (n_workers > 1) {
2832         workers->run_task(&tsk);
2833       } else {
2834         tsk.work(0);
2835       }
2836     } else {
2837       // The serial version.
2838       CLDToOopClosure cld_closure(&notOlder, true);
2839       gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2840 
2841       StrongRootsScope srs(1);
2842 
2843       gch->gen_process_roots(&srs,
2844                              GenCollectedHeap::OldGen,
2845                              true,   // young gen as roots
2846                              GenCollectedHeap::ScanningOption(roots_scanning_options()),
2847                              should_unload_classes(),
2848                              &notOlder,
2849                              NULL,
2850                              &cld_closure);
2851     }
2852   }
2853 
2854   // Clear mod-union table; it will be dirtied in the prologue of
2855   // CMS generation per each young generation collection.
2856 
2857   assert(_modUnionTable.isAllClear(),
2858        "Was cleared in most recent final checkpoint phase"
2859        " or no bits are set in the gc_prologue before the start of the next "
2860        "subsequent marking phase.");
2861 
2862   assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
2863 
2864   // Save the end of the used_region of the constituent generations
2865   // to be used to limit the extent of sweep in each generation.
2866   save_sweep_limits();
2867   verify_overflow_empty();
2868 }
2869 
2870 bool CMSCollector::markFromRoots() {
2871   // we might be tempted to assert that:
2872   // assert(!SafepointSynchronize::is_at_safepoint(),
2873   //        "inconsistent argument?");
2874   // However that wouldn't be right, because it's possible that
2875   // a safepoint is indeed in progress as a young generation
2876   // stop-the-world GC happens even as we mark in this generation.
2877   assert(_collectorState == Marking, "inconsistent state?");
2878   check_correct_thread_executing();
2879   verify_overflow_empty();
2880 
2881   // Weak ref discovery note: We may be discovering weak
2882   // refs in this generation concurrent (but interleaved) with
2883   // weak ref discovery by the young generation collector.
2884 
2885   CMSTokenSyncWithLocks ts(true, bitMapLock());
2886   GCTraceCPUTime tcpu;
2887   CMSPhaseAccounting pa(this, "Concrurrent Mark");
2888   bool res = markFromRootsWork();
2889   if (res) {
2890     _collectorState = Precleaning;
2891   } else { // We failed and a foreground collection wants to take over
2892     assert(_foregroundGCIsActive, "internal state inconsistency");
2893     assert(_restart_addr == NULL,  "foreground will restart from scratch");
2894     log_debug(gc)("bailing out to foreground collection");
2895   }
2896   verify_overflow_empty();
2897   return res;
2898 }
2899 
2900 bool CMSCollector::markFromRootsWork() {
2901   // iterate over marked bits in bit map, doing a full scan and mark
2902   // from these roots using the following algorithm:
2903   // . if oop is to the right of the current scan pointer,
2904   //   mark corresponding bit (we'll process it later)
2905   // . else (oop is to left of current scan pointer)
2906   //   push oop on marking stack
2907   // . drain the marking stack
2908 
2909   // Note that when we do a marking step we need to hold the
2910   // bit map lock -- recall that direct allocation (by mutators)
2911   // and promotion (by the young generation collector) is also
2912   // marking the bit map. [the so-called allocate live policy.]
2913   // Because the implementation of bit map marking is not
2914   // robust wrt simultaneous marking of bits in the same word,
2915   // we need to make sure that there is no such interference
2916   // between concurrent such updates.
2917 
2918   // already have locks
2919   assert_lock_strong(bitMapLock());
2920 
2921   verify_work_stacks_empty();
2922   verify_overflow_empty();
2923   bool result = false;
2924   if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2925     result = do_marking_mt();
2926   } else {
2927     result = do_marking_st();
2928   }
2929   return result;
2930 }
2931 
2932 // Forward decl
2933 class CMSConcMarkingTask;
2934 
2935 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2936   CMSCollector*       _collector;
2937   CMSConcMarkingTask* _task;
2938  public:
2939   virtual void yield();
2940 
2941   // "n_threads" is the number of threads to be terminated.
2942   // "queue_set" is a set of work queues of other threads.
2943   // "collector" is the CMS collector associated with this task terminator.
2944   // "yield" indicates whether we need the gang as a whole to yield.
2945   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
2946     ParallelTaskTerminator(n_threads, queue_set),
2947     _collector(collector) { }
2948 
2949   void set_task(CMSConcMarkingTask* task) {
2950     _task = task;
2951   }
2952 };
2953 
2954 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
2955   CMSConcMarkingTask* _task;
2956  public:
2957   bool should_exit_termination();
2958   void set_task(CMSConcMarkingTask* task) {
2959     _task = task;
2960   }
2961 };
2962 
2963 // MT Concurrent Marking Task
2964 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
2965   CMSCollector* _collector;
2966   uint          _n_workers;       // requested/desired # workers
2967   bool          _result;
2968   CompactibleFreeListSpace*  _cms_space;
2969   char          _pad_front[64];   // padding to ...
2970   HeapWord*     _global_finger;   // ... avoid sharing cache line
2971   char          _pad_back[64];
2972   HeapWord*     _restart_addr;
2973 
2974   //  Exposed here for yielding support
2975   Mutex* const _bit_map_lock;
2976 
2977   // The per thread work queues, available here for stealing
2978   OopTaskQueueSet*  _task_queues;
2979 
2980   // Termination (and yielding) support
2981   CMSConcMarkingTerminator _term;
2982   CMSConcMarkingTerminatorTerminator _term_term;
2983 
2984  public:
2985   CMSConcMarkingTask(CMSCollector* collector,
2986                  CompactibleFreeListSpace* cms_space,
2987                  YieldingFlexibleWorkGang* workers,
2988                  OopTaskQueueSet* task_queues):
2989     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
2990     _collector(collector),
2991     _cms_space(cms_space),
2992     _n_workers(0), _result(true),
2993     _task_queues(task_queues),
2994     _term(_n_workers, task_queues, _collector),
2995     _bit_map_lock(collector->bitMapLock())
2996   {
2997     _requested_size = _n_workers;
2998     _term.set_task(this);
2999     _term_term.set_task(this);
3000     _restart_addr = _global_finger = _cms_space->bottom();
3001   }
3002 
3003 
3004   OopTaskQueueSet* task_queues()  { return _task_queues; }
3005 
3006   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3007 
3008   HeapWord** global_finger_addr() { return &_global_finger; }
3009 
3010   CMSConcMarkingTerminator* terminator() { return &_term; }
3011 
3012   virtual void set_for_termination(uint active_workers) {
3013     terminator()->reset_for_reuse(active_workers);
3014   }
3015 
3016   void work(uint worker_id);
3017   bool should_yield() {
3018     return    ConcurrentMarkSweepThread::should_yield()
3019            && !_collector->foregroundGCIsActive();
3020   }
3021 
3022   virtual void coordinator_yield();  // stuff done by coordinator
3023   bool result() { return _result; }
3024 
3025   void reset(HeapWord* ra) {
3026     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
3027     _restart_addr = _global_finger = ra;
3028     _term.reset_for_reuse();
3029   }
3030 
3031   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3032                                            OopTaskQueue* work_q);
3033 
3034  private:
3035   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3036   void do_work_steal(int i);
3037   void bump_global_finger(HeapWord* f);
3038 };
3039 
3040 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3041   assert(_task != NULL, "Error");
3042   return _task->yielding();
3043   // Note that we do not need the disjunct || _task->should_yield() above
3044   // because we want terminating threads to yield only if the task
3045   // is already in the midst of yielding, which happens only after at least one
3046   // thread has yielded.
3047 }
3048 
3049 void CMSConcMarkingTerminator::yield() {
3050   if (_task->should_yield()) {
3051     _task->yield();
3052   } else {
3053     ParallelTaskTerminator::yield();
3054   }
3055 }
3056 
3057 ////////////////////////////////////////////////////////////////
3058 // Concurrent Marking Algorithm Sketch
3059 ////////////////////////////////////////////////////////////////
3060 // Until all tasks exhausted (both spaces):
3061 // -- claim next available chunk
3062 // -- bump global finger via CAS
3063 // -- find first object that starts in this chunk
3064 //    and start scanning bitmap from that position
3065 // -- scan marked objects for oops
3066 // -- CAS-mark target, and if successful:
3067 //    . if target oop is above global finger (volatile read)
3068 //      nothing to do
3069 //    . if target oop is in chunk and above local finger
3070 //        then nothing to do
3071 //    . else push on work-queue
3072 // -- Deal with possible overflow issues:
3073 //    . local work-queue overflow causes stuff to be pushed on
3074 //      global (common) overflow queue
3075 //    . always first empty local work queue
3076 //    . then get a batch of oops from global work queue if any
3077 //    . then do work stealing
3078 // -- When all tasks claimed (both spaces)
3079 //    and local work queue empty,
3080 //    then in a loop do:
3081 //    . check global overflow stack; steal a batch of oops and trace
3082 //    . try to steal from other threads oif GOS is empty
3083 //    . if neither is available, offer termination
3084 // -- Terminate and return result
3085 //
3086 void CMSConcMarkingTask::work(uint worker_id) {
3087   elapsedTimer _timer;
3088   ResourceMark rm;
3089   HandleMark hm;
3090 
3091   DEBUG_ONLY(_collector->verify_overflow_empty();)
3092 
3093   // Before we begin work, our work queue should be empty
3094   assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3095   // Scan the bitmap covering _cms_space, tracing through grey objects.
3096   _timer.start();
3097   do_scan_and_mark(worker_id, _cms_space);
3098   _timer.stop();
3099   log_debug(gc, stats)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3100 
3101   // ... do work stealing
3102   _timer.reset();
3103   _timer.start();
3104   do_work_steal(worker_id);
3105   _timer.stop();
3106   log_debug(gc, stats)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3107   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3108   assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3109   // Note that under the current task protocol, the
3110   // following assertion is true even of the spaces
3111   // expanded since the completion of the concurrent
3112   // marking. XXX This will likely change under a strict
3113   // ABORT semantics.
3114   // After perm removal the comparison was changed to
3115   // greater than or equal to from strictly greater than.
3116   // Before perm removal the highest address sweep would
3117   // have been at the end of perm gen but now is at the
3118   // end of the tenured gen.
3119   assert(_global_finger >=  _cms_space->end(),
3120          "All tasks have been completed");
3121   DEBUG_ONLY(_collector->verify_overflow_empty();)
3122 }
3123 
3124 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3125   HeapWord* read = _global_finger;
3126   HeapWord* cur  = read;
3127   while (f > read) {
3128     cur = read;
3129     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3130     if (cur == read) {
3131       // our cas succeeded
3132       assert(_global_finger >= f, "protocol consistency");
3133       break;
3134     }
3135   }
3136 }
3137 
3138 // This is really inefficient, and should be redone by
3139 // using (not yet available) block-read and -write interfaces to the
3140 // stack and the work_queue. XXX FIX ME !!!
3141 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3142                                                       OopTaskQueue* work_q) {
3143   // Fast lock-free check
3144   if (ovflw_stk->length() == 0) {
3145     return false;
3146   }
3147   assert(work_q->size() == 0, "Shouldn't steal");
3148   MutexLockerEx ml(ovflw_stk->par_lock(),
3149                    Mutex::_no_safepoint_check_flag);
3150   // Grab up to 1/4 the size of the work queue
3151   size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3152                     (size_t)ParGCDesiredObjsFromOverflowList);
3153   num = MIN2(num, ovflw_stk->length());
3154   for (int i = (int) num; i > 0; i--) {
3155     oop cur = ovflw_stk->pop();
3156     assert(cur != NULL, "Counted wrong?");
3157     work_q->push(cur);
3158   }
3159   return num > 0;
3160 }
3161 
3162 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3163   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3164   int n_tasks = pst->n_tasks();
3165   // We allow that there may be no tasks to do here because
3166   // we are restarting after a stack overflow.
3167   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3168   uint nth_task = 0;
3169 
3170   HeapWord* aligned_start = sp->bottom();
3171   if (sp->used_region().contains(_restart_addr)) {
3172     // Align down to a card boundary for the start of 0th task
3173     // for this space.
3174     aligned_start =
3175       (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3176                                  CardTableModRefBS::card_size);
3177   }
3178 
3179   size_t chunk_size = sp->marking_task_size();
3180   while (!pst->is_task_claimed(/* reference */ nth_task)) {
3181     // Having claimed the nth task in this space,
3182     // compute the chunk that it corresponds to:
3183     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3184                                aligned_start + (nth_task+1)*chunk_size);
3185     // Try and bump the global finger via a CAS;
3186     // note that we need to do the global finger bump
3187     // _before_ taking the intersection below, because
3188     // the task corresponding to that region will be
3189     // deemed done even if the used_region() expands
3190     // because of allocation -- as it almost certainly will
3191     // during start-up while the threads yield in the
3192     // closure below.
3193     HeapWord* finger = span.end();
3194     bump_global_finger(finger);   // atomically
3195     // There are null tasks here corresponding to chunks
3196     // beyond the "top" address of the space.
3197     span = span.intersection(sp->used_region());
3198     if (!span.is_empty()) {  // Non-null task
3199       HeapWord* prev_obj;
3200       assert(!span.contains(_restart_addr) || nth_task == 0,
3201              "Inconsistency");
3202       if (nth_task == 0) {
3203         // For the 0th task, we'll not need to compute a block_start.
3204         if (span.contains(_restart_addr)) {
3205           // In the case of a restart because of stack overflow,
3206           // we might additionally skip a chunk prefix.
3207           prev_obj = _restart_addr;
3208         } else {
3209           prev_obj = span.start();
3210         }
3211       } else {
3212         // We want to skip the first object because
3213         // the protocol is to scan any object in its entirety
3214         // that _starts_ in this span; a fortiori, any
3215         // object starting in an earlier span is scanned
3216         // as part of an earlier claimed task.
3217         // Below we use the "careful" version of block_start
3218         // so we do not try to navigate uninitialized objects.
3219         prev_obj = sp->block_start_careful(span.start());
3220         // Below we use a variant of block_size that uses the
3221         // Printezis bits to avoid waiting for allocated
3222         // objects to become initialized/parsable.
3223         while (prev_obj < span.start()) {
3224           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3225           if (sz > 0) {
3226             prev_obj += sz;
3227           } else {
3228             // In this case we may end up doing a bit of redundant
3229             // scanning, but that appears unavoidable, short of
3230             // locking the free list locks; see bug 6324141.
3231             break;
3232           }
3233         }
3234       }
3235       if (prev_obj < span.end()) {
3236         MemRegion my_span = MemRegion(prev_obj, span.end());
3237         // Do the marking work within a non-empty span --
3238         // the last argument to the constructor indicates whether the
3239         // iteration should be incremental with periodic yields.
3240         Par_MarkFromRootsClosure cl(this, _collector, my_span,
3241                                     &_collector->_markBitMap,
3242                                     work_queue(i),
3243                                     &_collector->_markStack);
3244         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3245       } // else nothing to do for this task
3246     }   // else nothing to do for this task
3247   }
3248   // We'd be tempted to assert here that since there are no
3249   // more tasks left to claim in this space, the global_finger
3250   // must exceed space->top() and a fortiori space->end(). However,
3251   // that would not quite be correct because the bumping of
3252   // global_finger occurs strictly after the claiming of a task,
3253   // so by the time we reach here the global finger may not yet
3254   // have been bumped up by the thread that claimed the last
3255   // task.
3256   pst->all_tasks_completed();
3257 }
3258 
3259 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3260  private:
3261   CMSCollector* _collector;
3262   CMSConcMarkingTask* _task;
3263   MemRegion     _span;
3264   CMSBitMap*    _bit_map;
3265   CMSMarkStack* _overflow_stack;
3266   OopTaskQueue* _work_queue;
3267  protected:
3268   DO_OOP_WORK_DEFN
3269  public:
3270   Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3271                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3272     MetadataAwareOopClosure(collector->ref_processor()),
3273     _collector(collector),
3274     _task(task),
3275     _span(collector->_span),
3276     _work_queue(work_queue),
3277     _bit_map(bit_map),
3278     _overflow_stack(overflow_stack)
3279   { }
3280   virtual void do_oop(oop* p);
3281   virtual void do_oop(narrowOop* p);
3282 
3283   void trim_queue(size_t max);
3284   void handle_stack_overflow(HeapWord* lost);
3285   void do_yield_check() {
3286     if (_task->should_yield()) {
3287       _task->yield();
3288     }
3289   }
3290 };
3291 
3292 // Grey object scanning during work stealing phase --
3293 // the salient assumption here is that any references
3294 // that are in these stolen objects being scanned must
3295 // already have been initialized (else they would not have
3296 // been published), so we do not need to check for
3297 // uninitialized objects before pushing here.
3298 void Par_ConcMarkingClosure::do_oop(oop obj) {
3299   assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3300   HeapWord* addr = (HeapWord*)obj;
3301   // Check if oop points into the CMS generation
3302   // and is not marked
3303   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3304     // a white object ...
3305     // If we manage to "claim" the object, by being the
3306     // first thread to mark it, then we push it on our
3307     // marking stack
3308     if (_bit_map->par_mark(addr)) {     // ... now grey
3309       // push on work queue (grey set)
3310       bool simulate_overflow = false;
3311       NOT_PRODUCT(
3312         if (CMSMarkStackOverflowALot &&
3313             _collector->simulate_overflow()) {
3314           // simulate a stack overflow
3315           simulate_overflow = true;
3316         }
3317       )
3318       if (simulate_overflow ||
3319           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3320         // stack overflow
3321         log_debug(gc, stats)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3322         // We cannot assert that the overflow stack is full because
3323         // it may have been emptied since.
3324         assert(simulate_overflow ||
3325                _work_queue->size() == _work_queue->max_elems(),
3326               "Else push should have succeeded");
3327         handle_stack_overflow(addr);
3328       }
3329     } // Else, some other thread got there first
3330     do_yield_check();
3331   }
3332 }
3333 
3334 void Par_ConcMarkingClosure::do_oop(oop* p)       { Par_ConcMarkingClosure::do_oop_work(p); }
3335 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3336 
3337 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3338   while (_work_queue->size() > max) {
3339     oop new_oop;
3340     if (_work_queue->pop_local(new_oop)) {
3341       assert(new_oop->is_oop(), "Should be an oop");
3342       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3343       assert(_span.contains((HeapWord*)new_oop), "Not in span");
3344       new_oop->oop_iterate(this);  // do_oop() above
3345       do_yield_check();
3346     }
3347   }
3348 }
3349 
3350 // Upon stack overflow, we discard (part of) the stack,
3351 // remembering the least address amongst those discarded
3352 // in CMSCollector's _restart_address.
3353 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3354   // We need to do this under a mutex to prevent other
3355   // workers from interfering with the work done below.
3356   MutexLockerEx ml(_overflow_stack->par_lock(),
3357                    Mutex::_no_safepoint_check_flag);
3358   // Remember the least grey address discarded
3359   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3360   _collector->lower_restart_addr(ra);
3361   _overflow_stack->reset();  // discard stack contents
3362   _overflow_stack->expand(); // expand the stack if possible
3363 }
3364 
3365 
3366 void CMSConcMarkingTask::do_work_steal(int i) {
3367   OopTaskQueue* work_q = work_queue(i);
3368   oop obj_to_scan;
3369   CMSBitMap* bm = &(_collector->_markBitMap);
3370   CMSMarkStack* ovflw = &(_collector->_markStack);
3371   int* seed = _collector->hash_seed(i);
3372   Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3373   while (true) {
3374     cl.trim_queue(0);
3375     assert(work_q->size() == 0, "Should have been emptied above");
3376     if (get_work_from_overflow_stack(ovflw, work_q)) {
3377       // Can't assert below because the work obtained from the
3378       // overflow stack may already have been stolen from us.
3379       // assert(work_q->size() > 0, "Work from overflow stack");
3380       continue;
3381     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3382       assert(obj_to_scan->is_oop(), "Should be an oop");
3383       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3384       obj_to_scan->oop_iterate(&cl);
3385     } else if (terminator()->offer_termination(&_term_term)) {
3386       assert(work_q->size() == 0, "Impossible!");
3387       break;
3388     } else if (yielding() || should_yield()) {
3389       yield();
3390     }
3391   }
3392 }
3393 
3394 // This is run by the CMS (coordinator) thread.
3395 void CMSConcMarkingTask::coordinator_yield() {
3396   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3397          "CMS thread should hold CMS token");
3398   // First give up the locks, then yield, then re-lock
3399   // We should probably use a constructor/destructor idiom to
3400   // do this unlock/lock or modify the MutexUnlocker class to
3401   // serve our purpose. XXX
3402   assert_lock_strong(_bit_map_lock);
3403   _bit_map_lock->unlock();
3404   ConcurrentMarkSweepThread::desynchronize(true);
3405   _collector->stopTimer();
3406   _collector->incrementYields();
3407 
3408   // It is possible for whichever thread initiated the yield request
3409   // not to get a chance to wake up and take the bitmap lock between
3410   // this thread releasing it and reacquiring it. So, while the
3411   // should_yield() flag is on, let's sleep for a bit to give the
3412   // other thread a chance to wake up. The limit imposed on the number
3413   // of iterations is defensive, to avoid any unforseen circumstances
3414   // putting us into an infinite loop. Since it's always been this
3415   // (coordinator_yield()) method that was observed to cause the
3416   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3417   // which is by default non-zero. For the other seven methods that
3418   // also perform the yield operation, as are using a different
3419   // parameter (CMSYieldSleepCount) which is by default zero. This way we
3420   // can enable the sleeping for those methods too, if necessary.
3421   // See 6442774.
3422   //
3423   // We really need to reconsider the synchronization between the GC
3424   // thread and the yield-requesting threads in the future and we
3425   // should really use wait/notify, which is the recommended
3426   // way of doing this type of interaction. Additionally, we should
3427   // consolidate the eight methods that do the yield operation and they
3428   // are almost identical into one for better maintainability and
3429   // readability. See 6445193.
3430   //
3431   // Tony 2006.06.29
3432   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3433                    ConcurrentMarkSweepThread::should_yield() &&
3434                    !CMSCollector::foregroundGCIsActive(); ++i) {
3435     os::sleep(Thread::current(), 1, false);
3436   }
3437 
3438   ConcurrentMarkSweepThread::synchronize(true);
3439   _bit_map_lock->lock_without_safepoint_check();
3440   _collector->startTimer();
3441 }
3442 
3443 bool CMSCollector::do_marking_mt() {
3444   assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3445   uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3446                                                                   conc_workers()->active_workers(),
3447                                                                   Threads::number_of_non_daemon_threads());
3448   conc_workers()->set_active_workers(num_workers);
3449 
3450   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
3451 
3452   CMSConcMarkingTask tsk(this,
3453                          cms_space,
3454                          conc_workers(),
3455                          task_queues());
3456 
3457   // Since the actual number of workers we get may be different
3458   // from the number we requested above, do we need to do anything different
3459   // below? In particular, may be we need to subclass the SequantialSubTasksDone
3460   // class?? XXX
3461   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3462 
3463   // Refs discovery is already non-atomic.
3464   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3465   assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3466   conc_workers()->start_task(&tsk);
3467   while (tsk.yielded()) {
3468     tsk.coordinator_yield();
3469     conc_workers()->continue_task(&tsk);
3470   }
3471   // If the task was aborted, _restart_addr will be non-NULL
3472   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3473   while (_restart_addr != NULL) {
3474     // XXX For now we do not make use of ABORTED state and have not
3475     // yet implemented the right abort semantics (even in the original
3476     // single-threaded CMS case). That needs some more investigation
3477     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3478     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3479     // If _restart_addr is non-NULL, a marking stack overflow
3480     // occurred; we need to do a fresh marking iteration from the
3481     // indicated restart address.
3482     if (_foregroundGCIsActive) {
3483       // We may be running into repeated stack overflows, having
3484       // reached the limit of the stack size, while making very
3485       // slow forward progress. It may be best to bail out and
3486       // let the foreground collector do its job.
3487       // Clear _restart_addr, so that foreground GC
3488       // works from scratch. This avoids the headache of
3489       // a "rescan" which would otherwise be needed because
3490       // of the dirty mod union table & card table.
3491       _restart_addr = NULL;
3492       return false;
3493     }
3494     // Adjust the task to restart from _restart_addr
3495     tsk.reset(_restart_addr);
3496     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3497                   _restart_addr);
3498     _restart_addr = NULL;
3499     // Get the workers going again
3500     conc_workers()->start_task(&tsk);
3501     while (tsk.yielded()) {
3502       tsk.coordinator_yield();
3503       conc_workers()->continue_task(&tsk);
3504     }
3505   }
3506   assert(tsk.completed(), "Inconsistency");
3507   assert(tsk.result() == true, "Inconsistency");
3508   return true;
3509 }
3510 
3511 bool CMSCollector::do_marking_st() {
3512   ResourceMark rm;
3513   HandleMark   hm;
3514 
3515   // Temporarily make refs discovery single threaded (non-MT)
3516   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3517   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3518     &_markStack, CMSYield);
3519   // the last argument to iterate indicates whether the iteration
3520   // should be incremental with periodic yields.
3521   _markBitMap.iterate(&markFromRootsClosure);
3522   // If _restart_addr is non-NULL, a marking stack overflow
3523   // occurred; we need to do a fresh iteration from the
3524   // indicated restart address.
3525   while (_restart_addr != NULL) {
3526     if (_foregroundGCIsActive) {
3527       // We may be running into repeated stack overflows, having
3528       // reached the limit of the stack size, while making very
3529       // slow forward progress. It may be best to bail out and
3530       // let the foreground collector do its job.
3531       // Clear _restart_addr, so that foreground GC
3532       // works from scratch. This avoids the headache of
3533       // a "rescan" which would otherwise be needed because
3534       // of the dirty mod union table & card table.
3535       _restart_addr = NULL;
3536       return false;  // indicating failure to complete marking
3537     }
3538     // Deal with stack overflow:
3539     // we restart marking from _restart_addr
3540     HeapWord* ra = _restart_addr;
3541     markFromRootsClosure.reset(ra);
3542     _restart_addr = NULL;
3543     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3544   }
3545   return true;
3546 }
3547 
3548 void CMSCollector::preclean() {
3549   check_correct_thread_executing();
3550   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3551   verify_work_stacks_empty();
3552   verify_overflow_empty();
3553   _abort_preclean = false;
3554   if (CMSPrecleaningEnabled) {
3555     if (!CMSEdenChunksRecordAlways) {
3556       _eden_chunk_index = 0;
3557     }
3558     size_t used = get_eden_used();
3559     size_t capacity = get_eden_capacity();
3560     // Don't start sampling unless we will get sufficiently
3561     // many samples.
3562     if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3563                 * CMSScheduleRemarkEdenPenetration)) {
3564       _start_sampling = true;
3565     } else {
3566       _start_sampling = false;
3567     }
3568     GCTraceCPUTime tcpu;
3569     CMSPhaseAccounting pa(this, "Concurrent Preclean");
3570     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3571   }
3572   CMSTokenSync x(true); // is cms thread
3573   if (CMSPrecleaningEnabled) {
3574     sample_eden();
3575     _collectorState = AbortablePreclean;
3576   } else {
3577     _collectorState = FinalMarking;
3578   }
3579   verify_work_stacks_empty();
3580   verify_overflow_empty();
3581 }
3582 
3583 // Try and schedule the remark such that young gen
3584 // occupancy is CMSScheduleRemarkEdenPenetration %.
3585 void CMSCollector::abortable_preclean() {
3586   check_correct_thread_executing();
3587   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
3588   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3589 
3590   // If Eden's current occupancy is below this threshold,
3591   // immediately schedule the remark; else preclean
3592   // past the next scavenge in an effort to
3593   // schedule the pause as described above. By choosing
3594   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3595   // we will never do an actual abortable preclean cycle.
3596   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3597     GCTraceCPUTime tcpu;
3598     CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3599     // We need more smarts in the abortable preclean
3600     // loop below to deal with cases where allocation
3601     // in young gen is very very slow, and our precleaning
3602     // is running a losing race against a horde of
3603     // mutators intent on flooding us with CMS updates
3604     // (dirty cards).
3605     // One, admittedly dumb, strategy is to give up
3606     // after a certain number of abortable precleaning loops
3607     // or after a certain maximum time. We want to make
3608     // this smarter in the next iteration.
3609     // XXX FIX ME!!! YSR
3610     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3611     while (!(should_abort_preclean() ||
3612              ConcurrentMarkSweepThread::should_terminate())) {
3613       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3614       cumworkdone += workdone;
3615       loops++;
3616       // Voluntarily terminate abortable preclean phase if we have
3617       // been at it for too long.
3618       if ((CMSMaxAbortablePrecleanLoops != 0) &&
3619           loops >= CMSMaxAbortablePrecleanLoops) {
3620         log_debug(gc)(" CMS: abort preclean due to loops ");
3621         break;
3622       }
3623       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3624         log_debug(gc)(" CMS: abort preclean due to time ");
3625         break;
3626       }
3627       // If we are doing little work each iteration, we should
3628       // take a short break.
3629       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3630         // Sleep for some time, waiting for work to accumulate
3631         stopTimer();
3632         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3633         startTimer();
3634         waited++;
3635       }
3636     }
3637     log_debug(gc, stats)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3638                                loops, waited, cumworkdone);
3639   }
3640   CMSTokenSync x(true); // is cms thread
3641   if (_collectorState != Idling) {
3642     assert(_collectorState == AbortablePreclean,
3643            "Spontaneous state transition?");
3644     _collectorState = FinalMarking;
3645   } // Else, a foreground collection completed this CMS cycle.
3646   return;
3647 }
3648 
3649 // Respond to an Eden sampling opportunity
3650 void CMSCollector::sample_eden() {
3651   // Make sure a young gc cannot sneak in between our
3652   // reading and recording of a sample.
3653   assert(Thread::current()->is_ConcurrentGC_thread(),
3654          "Only the cms thread may collect Eden samples");
3655   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3656          "Should collect samples while holding CMS token");
3657   if (!_start_sampling) {
3658     return;
3659   }
3660   // When CMSEdenChunksRecordAlways is true, the eden chunk array
3661   // is populated by the young generation.
3662   if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3663     if (_eden_chunk_index < _eden_chunk_capacity) {
3664       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
3665       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3666              "Unexpected state of Eden");
3667       // We'd like to check that what we just sampled is an oop-start address;
3668       // however, we cannot do that here since the object may not yet have been
3669       // initialized. So we'll instead do the check when we _use_ this sample
3670       // later.
3671       if (_eden_chunk_index == 0 ||
3672           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3673                          _eden_chunk_array[_eden_chunk_index-1])
3674            >= CMSSamplingGrain)) {
3675         _eden_chunk_index++;  // commit sample
3676       }
3677     }
3678   }
3679   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3680     size_t used = get_eden_used();
3681     size_t capacity = get_eden_capacity();
3682     assert(used <= capacity, "Unexpected state of Eden");
3683     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3684       _abort_preclean = true;
3685     }
3686   }
3687 }
3688 
3689 
3690 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3691   assert(_collectorState == Precleaning ||
3692          _collectorState == AbortablePreclean, "incorrect state");
3693   ResourceMark rm;
3694   HandleMark   hm;
3695 
3696   // Precleaning is currently not MT but the reference processor
3697   // may be set for MT.  Disable it temporarily here.
3698   ReferenceProcessor* rp = ref_processor();
3699   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3700 
3701   // Do one pass of scrubbing the discovered reference lists
3702   // to remove any reference objects with strongly-reachable
3703   // referents.
3704   if (clean_refs) {
3705     CMSPrecleanRefsYieldClosure yield_cl(this);
3706     assert(rp->span().equals(_span), "Spans should be equal");
3707     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3708                                    &_markStack, true /* preclean */);
3709     CMSDrainMarkingStackClosure complete_trace(this,
3710                                    _span, &_markBitMap, &_markStack,
3711                                    &keep_alive, true /* preclean */);
3712 
3713     // We don't want this step to interfere with a young
3714     // collection because we don't want to take CPU
3715     // or memory bandwidth away from the young GC threads
3716     // (which may be as many as there are CPUs).
3717     // Note that we don't need to protect ourselves from
3718     // interference with mutators because they can't
3719     // manipulate the discovered reference lists nor affect
3720     // the computed reachability of the referents, the
3721     // only properties manipulated by the precleaning
3722     // of these reference lists.
3723     stopTimer();
3724     CMSTokenSyncWithLocks x(true /* is cms thread */,
3725                             bitMapLock());
3726     startTimer();
3727     sample_eden();
3728 
3729     // The following will yield to allow foreground
3730     // collection to proceed promptly. XXX YSR:
3731     // The code in this method may need further
3732     // tweaking for better performance and some restructuring
3733     // for cleaner interfaces.
3734     GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3735     rp->preclean_discovered_references(
3736           rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3737           gc_timer);
3738   }
3739 
3740   if (clean_survivor) {  // preclean the active survivor space(s)
3741     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3742                              &_markBitMap, &_modUnionTable,
3743                              &_markStack, true /* precleaning phase */);
3744     stopTimer();
3745     CMSTokenSyncWithLocks ts(true /* is cms thread */,
3746                              bitMapLock());
3747     startTimer();
3748     unsigned int before_count =
3749       GenCollectedHeap::heap()->total_collections();
3750     SurvivorSpacePrecleanClosure
3751       sss_cl(this, _span, &_markBitMap, &_markStack,
3752              &pam_cl, before_count, CMSYield);
3753     _young_gen->from()->object_iterate_careful(&sss_cl);
3754     _young_gen->to()->object_iterate_careful(&sss_cl);
3755   }
3756   MarkRefsIntoAndScanClosure
3757     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3758              &_markStack, this, CMSYield,
3759              true /* precleaning phase */);
3760   // CAUTION: The following closure has persistent state that may need to
3761   // be reset upon a decrease in the sequence of addresses it
3762   // processes.
3763   ScanMarkedObjectsAgainCarefullyClosure
3764     smoac_cl(this, _span,
3765       &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3766 
3767   // Preclean dirty cards in ModUnionTable and CardTable using
3768   // appropriate convergence criterion;
3769   // repeat CMSPrecleanIter times unless we find that
3770   // we are losing.
3771   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3772   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3773          "Bad convergence multiplier");
3774   assert(CMSPrecleanThreshold >= 100,
3775          "Unreasonably low CMSPrecleanThreshold");
3776 
3777   size_t numIter, cumNumCards, lastNumCards, curNumCards;
3778   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3779        numIter < CMSPrecleanIter;
3780        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3781     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
3782     log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3783     // Either there are very few dirty cards, so re-mark
3784     // pause will be small anyway, or our pre-cleaning isn't
3785     // that much faster than the rate at which cards are being
3786     // dirtied, so we might as well stop and re-mark since
3787     // precleaning won't improve our re-mark time by much.
3788     if (curNumCards <= CMSPrecleanThreshold ||
3789         (numIter > 0 &&
3790          (curNumCards * CMSPrecleanDenominator >
3791          lastNumCards * CMSPrecleanNumerator))) {
3792       numIter++;
3793       cumNumCards += curNumCards;
3794       break;
3795     }
3796   }
3797 
3798   preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
3799 
3800   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3801   cumNumCards += curNumCards;
3802   log_debug(gc, stats)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3803                              curNumCards, cumNumCards, numIter);
3804   return cumNumCards;   // as a measure of useful work done
3805 }
3806 
3807 // PRECLEANING NOTES:
3808 // Precleaning involves:
3809 // . reading the bits of the modUnionTable and clearing the set bits.
3810 // . For the cards corresponding to the set bits, we scan the
3811 //   objects on those cards. This means we need the free_list_lock
3812 //   so that we can safely iterate over the CMS space when scanning
3813 //   for oops.
3814 // . When we scan the objects, we'll be both reading and setting
3815 //   marks in the marking bit map, so we'll need the marking bit map.
3816 // . For protecting _collector_state transitions, we take the CGC_lock.
3817 //   Note that any races in the reading of of card table entries by the
3818 //   CMS thread on the one hand and the clearing of those entries by the
3819 //   VM thread or the setting of those entries by the mutator threads on the
3820 //   other are quite benign. However, for efficiency it makes sense to keep
3821 //   the VM thread from racing with the CMS thread while the latter is
3822 //   dirty card info to the modUnionTable. We therefore also use the
3823 //   CGC_lock to protect the reading of the card table and the mod union
3824 //   table by the CM thread.
3825 // . We run concurrently with mutator updates, so scanning
3826 //   needs to be done carefully  -- we should not try to scan
3827 //   potentially uninitialized objects.
3828 //
3829 // Locking strategy: While holding the CGC_lock, we scan over and
3830 // reset a maximal dirty range of the mod union / card tables, then lock
3831 // the free_list_lock and bitmap lock to do a full marking, then
3832 // release these locks; and repeat the cycle. This allows for a
3833 // certain amount of fairness in the sharing of these locks between
3834 // the CMS collector on the one hand, and the VM thread and the
3835 // mutators on the other.
3836 
3837 // NOTE: preclean_mod_union_table() and preclean_card_table()
3838 // further below are largely identical; if you need to modify
3839 // one of these methods, please check the other method too.
3840 
3841 size_t CMSCollector::preclean_mod_union_table(
3842   ConcurrentMarkSweepGeneration* old_gen,
3843   ScanMarkedObjectsAgainCarefullyClosure* cl) {
3844   verify_work_stacks_empty();
3845   verify_overflow_empty();
3846 
3847   // strategy: starting with the first card, accumulate contiguous
3848   // ranges of dirty cards; clear these cards, then scan the region
3849   // covered by these cards.
3850 
3851   // Since all of the MUT is committed ahead, we can just use
3852   // that, in case the generations expand while we are precleaning.
3853   // It might also be fine to just use the committed part of the
3854   // generation, but we might potentially miss cards when the
3855   // generation is rapidly expanding while we are in the midst
3856   // of precleaning.
3857   HeapWord* startAddr = old_gen->reserved().start();
3858   HeapWord* endAddr   = old_gen->reserved().end();
3859 
3860   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
3861 
3862   size_t numDirtyCards, cumNumDirtyCards;
3863   HeapWord *nextAddr, *lastAddr;
3864   for (cumNumDirtyCards = numDirtyCards = 0,
3865        nextAddr = lastAddr = startAddr;
3866        nextAddr < endAddr;
3867        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3868 
3869     ResourceMark rm;
3870     HandleMark   hm;
3871 
3872     MemRegion dirtyRegion;
3873     {
3874       stopTimer();
3875       // Potential yield point
3876       CMSTokenSync ts(true);
3877       startTimer();
3878       sample_eden();
3879       // Get dirty region starting at nextOffset (inclusive),
3880       // simultaneously clearing it.
3881       dirtyRegion =
3882         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3883       assert(dirtyRegion.start() >= nextAddr,
3884              "returned region inconsistent?");
3885     }
3886     // Remember where the next search should begin.
3887     // The returned region (if non-empty) is a right open interval,
3888     // so lastOffset is obtained from the right end of that
3889     // interval.
3890     lastAddr = dirtyRegion.end();
3891     // Should do something more transparent and less hacky XXX
3892     numDirtyCards =
3893       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3894 
3895     // We'll scan the cards in the dirty region (with periodic
3896     // yields for foreground GC as needed).
3897     if (!dirtyRegion.is_empty()) {
3898       assert(numDirtyCards > 0, "consistency check");
3899       HeapWord* stop_point = NULL;
3900       stopTimer();
3901       // Potential yield point
3902       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3903                                bitMapLock());
3904       startTimer();
3905       {
3906         verify_work_stacks_empty();
3907         verify_overflow_empty();
3908         sample_eden();
3909         stop_point =
3910           old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3911       }
3912       if (stop_point != NULL) {
3913         // The careful iteration stopped early either because it found an
3914         // uninitialized object, or because we were in the midst of an
3915         // "abortable preclean", which should now be aborted. Redirty
3916         // the bits corresponding to the partially-scanned or unscanned
3917         // cards. We'll either restart at the next block boundary or
3918         // abort the preclean.
3919         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3920                "Should only be AbortablePreclean.");
3921         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3922         if (should_abort_preclean()) {
3923           break; // out of preclean loop
3924         } else {
3925           // Compute the next address at which preclean should pick up;
3926           // might need bitMapLock in order to read P-bits.
3927           lastAddr = next_card_start_after_block(stop_point);
3928         }
3929       }
3930     } else {
3931       assert(lastAddr == endAddr, "consistency check");
3932       assert(numDirtyCards == 0, "consistency check");
3933       break;
3934     }
3935   }
3936   verify_work_stacks_empty();
3937   verify_overflow_empty();
3938   return cumNumDirtyCards;
3939 }
3940 
3941 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3942 // below are largely identical; if you need to modify
3943 // one of these methods, please check the other method too.
3944 
3945 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
3946   ScanMarkedObjectsAgainCarefullyClosure* cl) {
3947   // strategy: it's similar to precleamModUnionTable above, in that
3948   // we accumulate contiguous ranges of dirty cards, mark these cards
3949   // precleaned, then scan the region covered by these cards.
3950   HeapWord* endAddr   = (HeapWord*)(old_gen->_virtual_space.high());
3951   HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
3952 
3953   cl->setFreelistLock(old_gen->freelistLock());   // needed for yielding
3954 
3955   size_t numDirtyCards, cumNumDirtyCards;
3956   HeapWord *lastAddr, *nextAddr;
3957 
3958   for (cumNumDirtyCards = numDirtyCards = 0,
3959        nextAddr = lastAddr = startAddr;
3960        nextAddr < endAddr;
3961        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3962 
3963     ResourceMark rm;
3964     HandleMark   hm;
3965 
3966     MemRegion dirtyRegion;
3967     {
3968       // See comments in "Precleaning notes" above on why we
3969       // do this locking. XXX Could the locking overheads be
3970       // too high when dirty cards are sparse? [I don't think so.]
3971       stopTimer();
3972       CMSTokenSync x(true); // is cms thread
3973       startTimer();
3974       sample_eden();
3975       // Get and clear dirty region from card table
3976       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
3977                                     MemRegion(nextAddr, endAddr),
3978                                     true,
3979                                     CardTableModRefBS::precleaned_card_val());
3980 
3981       assert(dirtyRegion.start() >= nextAddr,
3982              "returned region inconsistent?");
3983     }
3984     lastAddr = dirtyRegion.end();
3985     numDirtyCards =
3986       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
3987 
3988     if (!dirtyRegion.is_empty()) {
3989       stopTimer();
3990       CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
3991       startTimer();
3992       sample_eden();
3993       verify_work_stacks_empty();
3994       verify_overflow_empty();
3995       HeapWord* stop_point =
3996         old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3997       if (stop_point != NULL) {
3998         assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3999                "Should only be AbortablePreclean.");
4000         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4001         if (should_abort_preclean()) {
4002           break; // out of preclean loop
4003         } else {
4004           // Compute the next address at which preclean should pick up.
4005           lastAddr = next_card_start_after_block(stop_point);
4006         }
4007       }
4008     } else {
4009       break;
4010     }
4011   }
4012   verify_work_stacks_empty();
4013   verify_overflow_empty();
4014   return cumNumDirtyCards;
4015 }
4016 
4017 class PrecleanKlassClosure : public KlassClosure {
4018   KlassToOopClosure _cm_klass_closure;
4019  public:
4020   PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4021   void do_klass(Klass* k) {
4022     if (k->has_accumulated_modified_oops()) {
4023       k->clear_accumulated_modified_oops();
4024 
4025       _cm_klass_closure.do_klass(k);
4026     }
4027   }
4028 };
4029 
4030 // The freelist lock is needed to prevent asserts, is it really needed?
4031 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4032 
4033   cl->set_freelistLock(freelistLock);
4034 
4035   CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4036 
4037   // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4038   // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4039   PrecleanKlassClosure preclean_klass_closure(cl);
4040   ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4041 
4042   verify_work_stacks_empty();
4043   verify_overflow_empty();
4044 }
4045 
4046 void CMSCollector::checkpointRootsFinal() {
4047   assert(_collectorState == FinalMarking, "incorrect state transition?");
4048   check_correct_thread_executing();
4049   // world is stopped at this checkpoint
4050   assert(SafepointSynchronize::is_at_safepoint(),
4051          "world should be stopped");
4052   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4053 
4054   verify_work_stacks_empty();
4055   verify_overflow_empty();
4056 
4057   log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4058                       _young_gen->used() / K, _young_gen->capacity() / K);
4059   {
4060     if (CMSScavengeBeforeRemark) {
4061       GenCollectedHeap* gch = GenCollectedHeap::heap();
4062       // Temporarily set flag to false, GCH->do_collection will
4063       // expect it to be false and set to true
4064       FlagSetting fl(gch->_is_gc_active, false);
4065 
4066       GCTraceTime(Trace, gc) tm("Scavenge Before Remark", _gc_timer_cm);
4067 
4068       gch->do_collection(true,                      // full (i.e. force, see below)
4069                          false,                     // !clear_all_soft_refs
4070                          0,                         // size
4071                          false,                     // is_tlab
4072                          GenCollectedHeap::YoungGen // type
4073         );
4074     }
4075     FreelistLocker x(this);
4076     MutexLockerEx y(bitMapLock(),
4077                     Mutex::_no_safepoint_check_flag);
4078     checkpointRootsFinalWork();
4079   }
4080   verify_work_stacks_empty();
4081   verify_overflow_empty();
4082 }
4083 
4084 void CMSCollector::checkpointRootsFinalWork() {
4085   GCTraceTime(Trace, gc) tm("checkpointRootsFinalWork", _gc_timer_cm);
4086 
4087   assert(haveFreelistLocks(), "must have free list locks");
4088   assert_lock_strong(bitMapLock());
4089 
4090   ResourceMark rm;
4091   HandleMark   hm;
4092 
4093   GenCollectedHeap* gch = GenCollectedHeap::heap();
4094 
4095   if (should_unload_classes()) {
4096     CodeCache::gc_prologue();
4097   }
4098   assert(haveFreelistLocks(), "must have free list locks");
4099   assert_lock_strong(bitMapLock());
4100 
4101   // We might assume that we need not fill TLAB's when
4102   // CMSScavengeBeforeRemark is set, because we may have just done
4103   // a scavenge which would have filled all TLAB's -- and besides
4104   // Eden would be empty. This however may not always be the case --
4105   // for instance although we asked for a scavenge, it may not have
4106   // happened because of a JNI critical section. We probably need
4107   // a policy for deciding whether we can in that case wait until
4108   // the critical section releases and then do the remark following
4109   // the scavenge, and skip it here. In the absence of that policy,
4110   // or of an indication of whether the scavenge did indeed occur,
4111   // we cannot rely on TLAB's having been filled and must do
4112   // so here just in case a scavenge did not happen.
4113   gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
4114   // Update the saved marks which may affect the root scans.
4115   gch->save_marks();
4116 
4117   print_eden_and_survivor_chunk_arrays();
4118 
4119   {
4120 #if defined(COMPILER2) || INCLUDE_JVMCI
4121     DerivedPointerTableDeactivate dpt_deact;
4122 #endif
4123 
4124     // Note on the role of the mod union table:
4125     // Since the marker in "markFromRoots" marks concurrently with
4126     // mutators, it is possible for some reachable objects not to have been
4127     // scanned. For instance, an only reference to an object A was
4128     // placed in object B after the marker scanned B. Unless B is rescanned,
4129     // A would be collected. Such updates to references in marked objects
4130     // are detected via the mod union table which is the set of all cards
4131     // dirtied since the first checkpoint in this GC cycle and prior to
4132     // the most recent young generation GC, minus those cleaned up by the
4133     // concurrent precleaning.
4134     if (CMSParallelRemarkEnabled) {
4135       GCTraceTime(Debug, gc) t("Rescan (parallel)", _gc_timer_cm);
4136       do_remark_parallel();
4137     } else {
4138       GCTraceTime(Debug, gc) t("Rescan (non-parallel)", _gc_timer_cm);
4139       do_remark_non_parallel();
4140     }
4141   }
4142   verify_work_stacks_empty();
4143   verify_overflow_empty();
4144 
4145   {
4146     GCTraceTime(Trace, gc) ts("refProcessingWork", _gc_timer_cm);
4147     refProcessingWork();
4148   }
4149   verify_work_stacks_empty();
4150   verify_overflow_empty();
4151 
4152   if (should_unload_classes()) {
4153     CodeCache::gc_epilogue();
4154   }
4155   JvmtiExport::gc_epilogue();
4156 
4157   // If we encountered any (marking stack / work queue) overflow
4158   // events during the current CMS cycle, take appropriate
4159   // remedial measures, where possible, so as to try and avoid
4160   // recurrence of that condition.
4161   assert(_markStack.isEmpty(), "No grey objects");
4162   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4163                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
4164   if (ser_ovflw > 0) {
4165     log_debug(gc, stats)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4166                                _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4167     _markStack.expand();
4168     _ser_pmc_remark_ovflw = 0;
4169     _ser_pmc_preclean_ovflw = 0;
4170     _ser_kac_preclean_ovflw = 0;
4171     _ser_kac_ovflw = 0;
4172   }
4173   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4174      log_debug(gc, stats)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4175                                  _par_pmc_remark_ovflw, _par_kac_ovflw);
4176      _par_pmc_remark_ovflw = 0;
4177     _par_kac_ovflw = 0;
4178   }
4179    if (_markStack._hit_limit > 0) {
4180      log_debug(gc, stats)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4181                                 _markStack._hit_limit);
4182    }
4183    if (_markStack._failed_double > 0) {
4184      log_debug(gc, stats)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4185                                 _markStack._failed_double, _markStack.capacity());
4186    }
4187   _markStack._hit_limit = 0;
4188   _markStack._failed_double = 0;
4189 
4190   if ((VerifyAfterGC || VerifyDuringGC) &&
4191       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4192     verify_after_remark();
4193   }
4194 
4195   _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4196 
4197   // Change under the freelistLocks.
4198   _collectorState = Sweeping;
4199   // Call isAllClear() under bitMapLock
4200   assert(_modUnionTable.isAllClear(),
4201       "Should be clear by end of the final marking");
4202   assert(_ct->klass_rem_set()->mod_union_is_clear(),
4203       "Should be clear by end of the final marking");
4204 }
4205 
4206 void CMSParInitialMarkTask::work(uint worker_id) {
4207   elapsedTimer _timer;
4208   ResourceMark rm;
4209   HandleMark   hm;
4210 
4211   // ---------- scan from roots --------------
4212   _timer.start();
4213   GenCollectedHeap* gch = GenCollectedHeap::heap();
4214   Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4215 
4216   // ---------- young gen roots --------------
4217   {
4218     work_on_young_gen_roots(worker_id, &par_mri_cl);
4219     _timer.stop();
4220     log_debug(gc, stats)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4221   }
4222 
4223   // ---------- remaining roots --------------
4224   _timer.reset();
4225   _timer.start();
4226 
4227   CLDToOopClosure cld_closure(&par_mri_cl, true);
4228 
4229   gch->gen_process_roots(_strong_roots_scope,
4230                          GenCollectedHeap::OldGen,
4231                          false,     // yg was scanned above
4232                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4233                          _collector->should_unload_classes(),
4234                          &par_mri_cl,
4235                          NULL,
4236                          &cld_closure);
4237   assert(_collector->should_unload_classes()
4238          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4239          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4240   _timer.stop();
4241   log_debug(gc, stats)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4242 }
4243 
4244 // Parallel remark task
4245 class CMSParRemarkTask: public CMSParMarkTask {
4246   CompactibleFreeListSpace* _cms_space;
4247 
4248   // The per-thread work queues, available here for stealing.
4249   OopTaskQueueSet*       _task_queues;
4250   ParallelTaskTerminator _term;
4251   StrongRootsScope*      _strong_roots_scope;
4252 
4253  public:
4254   // A value of 0 passed to n_workers will cause the number of
4255   // workers to be taken from the active workers in the work gang.
4256   CMSParRemarkTask(CMSCollector* collector,
4257                    CompactibleFreeListSpace* cms_space,
4258                    uint n_workers, WorkGang* workers,
4259                    OopTaskQueueSet* task_queues,
4260                    StrongRootsScope* strong_roots_scope):
4261     CMSParMarkTask("Rescan roots and grey objects in parallel",
4262                    collector, n_workers),
4263     _cms_space(cms_space),
4264     _task_queues(task_queues),
4265     _term(n_workers, task_queues),
4266     _strong_roots_scope(strong_roots_scope) { }
4267 
4268   OopTaskQueueSet* task_queues() { return _task_queues; }
4269 
4270   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4271 
4272   ParallelTaskTerminator* terminator() { return &_term; }
4273   uint n_workers() { return _n_workers; }
4274 
4275   void work(uint worker_id);
4276 
4277  private:
4278   // ... of  dirty cards in old space
4279   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4280                                   Par_MarkRefsIntoAndScanClosure* cl);
4281 
4282   // ... work stealing for the above
4283   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4284 };
4285 
4286 class RemarkKlassClosure : public KlassClosure {
4287   KlassToOopClosure _cm_klass_closure;
4288  public:
4289   RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4290   void do_klass(Klass* k) {
4291     // Check if we have modified any oops in the Klass during the concurrent marking.
4292     if (k->has_accumulated_modified_oops()) {
4293       k->clear_accumulated_modified_oops();
4294 
4295       // We could have transfered the current modified marks to the accumulated marks,
4296       // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4297     } else if (k->has_modified_oops()) {
4298       // Don't clear anything, this info is needed by the next young collection.
4299     } else {
4300       // No modified oops in the Klass.
4301       return;
4302     }
4303 
4304     // The klass has modified fields, need to scan the klass.
4305     _cm_klass_closure.do_klass(k);
4306   }
4307 };
4308 
4309 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4310   ParNewGeneration* young_gen = _collector->_young_gen;
4311   ContiguousSpace* eden_space = young_gen->eden();
4312   ContiguousSpace* from_space = young_gen->from();
4313   ContiguousSpace* to_space   = young_gen->to();
4314 
4315   HeapWord** eca = _collector->_eden_chunk_array;
4316   size_t     ect = _collector->_eden_chunk_index;
4317   HeapWord** sca = _collector->_survivor_chunk_array;
4318   size_t     sct = _collector->_survivor_chunk_index;
4319 
4320   assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4321   assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4322 
4323   do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4324   do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4325   do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4326 }
4327 
4328 // work_queue(i) is passed to the closure
4329 // Par_MarkRefsIntoAndScanClosure.  The "i" parameter
4330 // also is passed to do_dirty_card_rescan_tasks() and to
4331 // do_work_steal() to select the i-th task_queue.
4332 
4333 void CMSParRemarkTask::work(uint worker_id) {
4334   elapsedTimer _timer;
4335   ResourceMark rm;
4336   HandleMark   hm;
4337 
4338   // ---------- rescan from roots --------------
4339   _timer.start();
4340   GenCollectedHeap* gch = GenCollectedHeap::heap();
4341   Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4342     _collector->_span, _collector->ref_processor(),
4343     &(_collector->_markBitMap),
4344     work_queue(worker_id));
4345 
4346   // Rescan young gen roots first since these are likely
4347   // coarsely partitioned and may, on that account, constitute
4348   // the critical path; thus, it's best to start off that
4349   // work first.
4350   // ---------- young gen roots --------------
4351   {
4352     work_on_young_gen_roots(worker_id, &par_mrias_cl);
4353     _timer.stop();
4354     log_debug(gc, stats)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4355   }
4356 
4357   // ---------- remaining roots --------------
4358   _timer.reset();
4359   _timer.start();
4360   gch->gen_process_roots(_strong_roots_scope,
4361                          GenCollectedHeap::OldGen,
4362                          false,     // yg was scanned above
4363                          GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4364                          _collector->should_unload_classes(),
4365                          &par_mrias_cl,
4366                          NULL,
4367                          NULL);     // The dirty klasses will be handled below
4368 
4369   assert(_collector->should_unload_classes()
4370          || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4371          "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4372   _timer.stop();
4373   log_debug(gc, stats)("Finished remaining root rescan work in %dth thread: %3.3f sec",  worker_id, _timer.seconds());
4374 
4375   // ---------- unhandled CLD scanning ----------
4376   if (worker_id == 0) { // Single threaded at the moment.
4377     _timer.reset();
4378     _timer.start();
4379 
4380     // Scan all new class loader data objects and new dependencies that were
4381     // introduced during concurrent marking.
4382     ResourceMark rm;
4383     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4384     for (int i = 0; i < array->length(); i++) {
4385       par_mrias_cl.do_cld_nv(array->at(i));
4386     }
4387 
4388     // We don't need to keep track of new CLDs anymore.
4389     ClassLoaderDataGraph::remember_new_clds(false);
4390 
4391     _timer.stop();
4392     log_debug(gc, stats)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4393   }
4394 
4395   // ---------- dirty klass scanning ----------
4396   if (worker_id == 0) { // Single threaded at the moment.
4397     _timer.reset();
4398     _timer.start();
4399 
4400     // Scan all classes that was dirtied during the concurrent marking phase.
4401     RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4402     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4403 
4404     _timer.stop();
4405     log_debug(gc, stats)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4406   }
4407 
4408   // We might have added oops to ClassLoaderData::_handles during the
4409   // concurrent marking phase. These oops point to newly allocated objects
4410   // that are guaranteed to be kept alive. Either by the direct allocation
4411   // code, or when the young collector processes the roots. Hence,
4412   // we don't have to revisit the _handles block during the remark phase.
4413 
4414   // ---------- rescan dirty cards ------------
4415   _timer.reset();
4416   _timer.start();
4417 
4418   // Do the rescan tasks for each of the two spaces
4419   // (cms_space) in turn.
4420   // "worker_id" is passed to select the task_queue for "worker_id"
4421   do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4422   _timer.stop();
4423   log_debug(gc, stats)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4424 
4425   // ---------- steal work from other threads ...
4426   // ---------- ... and drain overflow list.
4427   _timer.reset();
4428   _timer.start();
4429   do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4430   _timer.stop();
4431   log_debug(gc, stats)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4432 }
4433 
4434 // Note that parameter "i" is not used.
4435 void
4436 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4437   OopsInGenClosure* cl, ContiguousSpace* space,
4438   HeapWord** chunk_array, size_t chunk_top) {
4439   // Until all tasks completed:
4440   // . claim an unclaimed task
4441   // . compute region boundaries corresponding to task claimed
4442   //   using chunk_array
4443   // . par_oop_iterate(cl) over that region
4444 
4445   ResourceMark rm;
4446   HandleMark   hm;
4447 
4448   SequentialSubTasksDone* pst = space->par_seq_tasks();
4449 
4450   uint nth_task = 0;
4451   uint n_tasks  = pst->n_tasks();
4452 
4453   if (n_tasks > 0) {
4454     assert(pst->valid(), "Uninitialized use?");
4455     HeapWord *start, *end;
4456     while (!pst->is_task_claimed(/* reference */ nth_task)) {
4457       // We claimed task # nth_task; compute its boundaries.
4458       if (chunk_top == 0) {  // no samples were taken
4459         assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4460         start = space->bottom();
4461         end   = space->top();
4462       } else if (nth_task == 0) {
4463         start = space->bottom();
4464         end   = chunk_array[nth_task];
4465       } else if (nth_task < (uint)chunk_top) {
4466         assert(nth_task >= 1, "Control point invariant");
4467         start = chunk_array[nth_task - 1];
4468         end   = chunk_array[nth_task];
4469       } else {
4470         assert(nth_task == (uint)chunk_top, "Control point invariant");
4471         start = chunk_array[chunk_top - 1];
4472         end   = space->top();
4473       }
4474       MemRegion mr(start, end);
4475       // Verify that mr is in space
4476       assert(mr.is_empty() || space->used_region().contains(mr),
4477              "Should be in space");
4478       // Verify that "start" is an object boundary
4479       assert(mr.is_empty() || oop(mr.start())->is_oop(),
4480              "Should be an oop");
4481       space->par_oop_iterate(mr, cl);
4482     }
4483     pst->all_tasks_completed();
4484   }
4485 }
4486 
4487 void
4488 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4489   CompactibleFreeListSpace* sp, int i,
4490   Par_MarkRefsIntoAndScanClosure* cl) {
4491   // Until all tasks completed:
4492   // . claim an unclaimed task
4493   // . compute region boundaries corresponding to task claimed
4494   // . transfer dirty bits ct->mut for that region
4495   // . apply rescanclosure to dirty mut bits for that region
4496 
4497   ResourceMark rm;
4498   HandleMark   hm;
4499 
4500   OopTaskQueue* work_q = work_queue(i);
4501   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4502   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4503   // CAUTION: This closure has state that persists across calls to
4504   // the work method dirty_range_iterate_clear() in that it has
4505   // embedded in it a (subtype of) UpwardsObjectClosure. The
4506   // use of that state in the embedded UpwardsObjectClosure instance
4507   // assumes that the cards are always iterated (even if in parallel
4508   // by several threads) in monotonically increasing order per each
4509   // thread. This is true of the implementation below which picks
4510   // card ranges (chunks) in monotonically increasing order globally
4511   // and, a-fortiori, in monotonically increasing order per thread
4512   // (the latter order being a subsequence of the former).
4513   // If the work code below is ever reorganized into a more chaotic
4514   // work-partitioning form than the current "sequential tasks"
4515   // paradigm, the use of that persistent state will have to be
4516   // revisited and modified appropriately. See also related
4517   // bug 4756801 work on which should examine this code to make
4518   // sure that the changes there do not run counter to the
4519   // assumptions made here and necessary for correctness and
4520   // efficiency. Note also that this code might yield inefficient
4521   // behavior in the case of very large objects that span one or
4522   // more work chunks. Such objects would potentially be scanned
4523   // several times redundantly. Work on 4756801 should try and
4524   // address that performance anomaly if at all possible. XXX
4525   MemRegion  full_span  = _collector->_span;
4526   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
4527   MarkFromDirtyCardsClosure
4528     greyRescanClosure(_collector, full_span, // entire span of interest
4529                       sp, bm, work_q, cl);
4530 
4531   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4532   assert(pst->valid(), "Uninitialized use?");
4533   uint nth_task = 0;
4534   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4535   MemRegion span = sp->used_region();
4536   HeapWord* start_addr = span.start();
4537   HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4538                                            alignment);
4539   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4540   assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4541          start_addr, "Check alignment");
4542   assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4543          chunk_size, "Check alignment");
4544 
4545   while (!pst->is_task_claimed(/* reference */ nth_task)) {
4546     // Having claimed the nth_task, compute corresponding mem-region,
4547     // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4548     // The alignment restriction ensures that we do not need any
4549     // synchronization with other gang-workers while setting or
4550     // clearing bits in thus chunk of the MUT.
4551     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4552                                     start_addr + (nth_task+1)*chunk_size);
4553     // The last chunk's end might be way beyond end of the
4554     // used region. In that case pull back appropriately.
4555     if (this_span.end() > end_addr) {
4556       this_span.set_end(end_addr);
4557       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4558     }
4559     // Iterate over the dirty cards covering this chunk, marking them
4560     // precleaned, and setting the corresponding bits in the mod union
4561     // table. Since we have been careful to partition at Card and MUT-word
4562     // boundaries no synchronization is needed between parallel threads.
4563     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4564                                                  &modUnionClosure);
4565 
4566     // Having transferred these marks into the modUnionTable,
4567     // rescan the marked objects on the dirty cards in the modUnionTable.
4568     // Even if this is at a synchronous collection, the initial marking
4569     // may have been done during an asynchronous collection so there
4570     // may be dirty bits in the mod-union table.
4571     _collector->_modUnionTable.dirty_range_iterate_clear(
4572                   this_span, &greyRescanClosure);
4573     _collector->_modUnionTable.verifyNoOneBitsInRange(
4574                                  this_span.start(),
4575                                  this_span.end());
4576   }
4577   pst->all_tasks_completed();  // declare that i am done
4578 }
4579 
4580 // . see if we can share work_queues with ParNew? XXX
4581 void
4582 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4583                                 int* seed) {
4584   OopTaskQueue* work_q = work_queue(i);
4585   NOT_PRODUCT(int num_steals = 0;)
4586   oop obj_to_scan;
4587   CMSBitMap* bm = &(_collector->_markBitMap);
4588 
4589   while (true) {
4590     // Completely finish any left over work from (an) earlier round(s)
4591     cl->trim_queue(0);
4592     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4593                                          (size_t)ParGCDesiredObjsFromOverflowList);
4594     // Now check if there's any work in the overflow list
4595     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4596     // only affects the number of attempts made to get work from the
4597     // overflow list and does not affect the number of workers.  Just
4598     // pass ParallelGCThreads so this behavior is unchanged.
4599     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4600                                                 work_q,
4601                                                 ParallelGCThreads)) {
4602       // found something in global overflow list;
4603       // not yet ready to go stealing work from others.
4604       // We'd like to assert(work_q->size() != 0, ...)
4605       // because we just took work from the overflow list,
4606       // but of course we can't since all of that could have
4607       // been already stolen from us.
4608       // "He giveth and He taketh away."
4609       continue;
4610     }
4611     // Verify that we have no work before we resort to stealing
4612     assert(work_q->size() == 0, "Have work, shouldn't steal");
4613     // Try to steal from other queues that have work
4614     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4615       NOT_PRODUCT(num_steals++;)
4616       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4617       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4618       // Do scanning work
4619       obj_to_scan->oop_iterate(cl);
4620       // Loop around, finish this work, and try to steal some more
4621     } else if (terminator()->offer_termination()) {
4622         break;  // nirvana from the infinite cycle
4623     }
4624   }
4625   log_develop(gc, stats)("\t(%d: stole %d oops)", i, num_steals);
4626   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4627          "Else our work is not yet done");
4628 }
4629 
4630 // Record object boundaries in _eden_chunk_array by sampling the eden
4631 // top in the slow-path eden object allocation code path and record
4632 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4633 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4634 // sampling in sample_eden() that activates during the part of the
4635 // preclean phase.
4636 void CMSCollector::sample_eden_chunk() {
4637   if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4638     if (_eden_chunk_lock->try_lock()) {
4639       // Record a sample. This is the critical section. The contents
4640       // of the _eden_chunk_array have to be non-decreasing in the
4641       // address order.
4642       _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4643       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4644              "Unexpected state of Eden");
4645       if (_eden_chunk_index == 0 ||
4646           ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4647            (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4648                           _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4649         _eden_chunk_index++;  // commit sample
4650       }
4651       _eden_chunk_lock->unlock();
4652     }
4653   }
4654 }
4655 
4656 // Return a thread-local PLAB recording array, as appropriate.
4657 void* CMSCollector::get_data_recorder(int thr_num) {
4658   if (_survivor_plab_array != NULL &&
4659       (CMSPLABRecordAlways ||
4660        (_collectorState > Marking && _collectorState < FinalMarking))) {
4661     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4662     ChunkArray* ca = &_survivor_plab_array[thr_num];
4663     ca->reset();   // clear it so that fresh data is recorded
4664     return (void*) ca;
4665   } else {
4666     return NULL;
4667   }
4668 }
4669 
4670 // Reset all the thread-local PLAB recording arrays
4671 void CMSCollector::reset_survivor_plab_arrays() {
4672   for (uint i = 0; i < ParallelGCThreads; i++) {
4673     _survivor_plab_array[i].reset();
4674   }
4675 }
4676 
4677 // Merge the per-thread plab arrays into the global survivor chunk
4678 // array which will provide the partitioning of the survivor space
4679 // for CMS initial scan and rescan.
4680 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4681                                               int no_of_gc_threads) {
4682   assert(_survivor_plab_array  != NULL, "Error");
4683   assert(_survivor_chunk_array != NULL, "Error");
4684   assert(_collectorState == FinalMarking ||
4685          (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4686   for (int j = 0; j < no_of_gc_threads; j++) {
4687     _cursor[j] = 0;
4688   }
4689   HeapWord* top = surv->top();
4690   size_t i;
4691   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
4692     HeapWord* min_val = top;          // Higher than any PLAB address
4693     uint      min_tid = 0;            // position of min_val this round
4694     for (int j = 0; j < no_of_gc_threads; j++) {
4695       ChunkArray* cur_sca = &_survivor_plab_array[j];
4696       if (_cursor[j] == cur_sca->end()) {
4697         continue;
4698       }
4699       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4700       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4701       assert(surv->used_region().contains(cur_val), "Out of bounds value");
4702       if (cur_val < min_val) {
4703         min_tid = j;
4704         min_val = cur_val;
4705       } else {
4706         assert(cur_val < top, "All recorded addresses should be less");
4707       }
4708     }
4709     // At this point min_val and min_tid are respectively
4710     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4711     // and the thread (j) that witnesses that address.
4712     // We record this address in the _survivor_chunk_array[i]
4713     // and increment _cursor[min_tid] prior to the next round i.
4714     if (min_val == top) {
4715       break;
4716     }
4717     _survivor_chunk_array[i] = min_val;
4718     _cursor[min_tid]++;
4719   }
4720   // We are all done; record the size of the _survivor_chunk_array
4721   _survivor_chunk_index = i; // exclusive: [0, i)
4722   log_debug(gc, stats)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4723   // Verify that we used up all the recorded entries
4724   #ifdef ASSERT
4725     size_t total = 0;
4726     for (int j = 0; j < no_of_gc_threads; j++) {
4727       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4728       total += _cursor[j];
4729     }
4730     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4731     // Check that the merged array is in sorted order
4732     if (total > 0) {
4733       for (size_t i = 0; i < total - 1; i++) {
4734         log_develop(gc, stats)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4735                                     i, p2i(_survivor_chunk_array[i]));
4736         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4737                "Not sorted");
4738       }
4739     }
4740   #endif // ASSERT
4741 }
4742 
4743 // Set up the space's par_seq_tasks structure for work claiming
4744 // for parallel initial scan and rescan of young gen.
4745 // See ParRescanTask where this is currently used.
4746 void
4747 CMSCollector::
4748 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4749   assert(n_threads > 0, "Unexpected n_threads argument");
4750 
4751   // Eden space
4752   if (!_young_gen->eden()->is_empty()) {
4753     SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4754     assert(!pst->valid(), "Clobbering existing data?");
4755     // Each valid entry in [0, _eden_chunk_index) represents a task.
4756     size_t n_tasks = _eden_chunk_index + 1;
4757     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4758     // Sets the condition for completion of the subtask (how many threads
4759     // need to finish in order to be done).
4760     pst->set_n_threads(n_threads);
4761     pst->set_n_tasks((int)n_tasks);
4762   }
4763 
4764   // Merge the survivor plab arrays into _survivor_chunk_array
4765   if (_survivor_plab_array != NULL) {
4766     merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4767   } else {
4768     assert(_survivor_chunk_index == 0, "Error");
4769   }
4770 
4771   // To space
4772   {
4773     SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4774     assert(!pst->valid(), "Clobbering existing data?");
4775     // Sets the condition for completion of the subtask (how many threads
4776     // need to finish in order to be done).
4777     pst->set_n_threads(n_threads);
4778     pst->set_n_tasks(1);
4779     assert(pst->valid(), "Error");
4780   }
4781 
4782   // From space
4783   {
4784     SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4785     assert(!pst->valid(), "Clobbering existing data?");
4786     size_t n_tasks = _survivor_chunk_index + 1;
4787     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4788     // Sets the condition for completion of the subtask (how many threads
4789     // need to finish in order to be done).
4790     pst->set_n_threads(n_threads);
4791     pst->set_n_tasks((int)n_tasks);
4792     assert(pst->valid(), "Error");
4793   }
4794 }
4795 
4796 // Parallel version of remark
4797 void CMSCollector::do_remark_parallel() {
4798   GenCollectedHeap* gch = GenCollectedHeap::heap();
4799   WorkGang* workers = gch->workers();
4800   assert(workers != NULL, "Need parallel worker threads.");
4801   // Choose to use the number of GC workers most recently set
4802   // into "active_workers".
4803   uint n_workers = workers->active_workers();
4804 
4805   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
4806 
4807   StrongRootsScope srs(n_workers);
4808 
4809   CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4810 
4811   // We won't be iterating over the cards in the card table updating
4812   // the younger_gen cards, so we shouldn't call the following else
4813   // the verification code as well as subsequent younger_refs_iterate
4814   // code would get confused. XXX
4815   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4816 
4817   // The young gen rescan work will not be done as part of
4818   // process_roots (which currently doesn't know how to
4819   // parallelize such a scan), but rather will be broken up into
4820   // a set of parallel tasks (via the sampling that the [abortable]
4821   // preclean phase did of eden, plus the [two] tasks of
4822   // scanning the [two] survivor spaces. Further fine-grain
4823   // parallelization of the scanning of the survivor spaces
4824   // themselves, and of precleaning of the young gen itself
4825   // is deferred to the future.
4826   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4827 
4828   // The dirty card rescan work is broken up into a "sequence"
4829   // of parallel tasks (per constituent space) that are dynamically
4830   // claimed by the parallel threads.
4831   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4832 
4833   // It turns out that even when we're using 1 thread, doing the work in a
4834   // separate thread causes wide variance in run times.  We can't help this
4835   // in the multi-threaded case, but we special-case n=1 here to get
4836   // repeatable measurements of the 1-thread overhead of the parallel code.
4837   if (n_workers > 1) {
4838     // Make refs discovery MT-safe, if it isn't already: it may not
4839     // necessarily be so, since it's possible that we are doing
4840     // ST marking.
4841     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4842     workers->run_task(&tsk);
4843   } else {
4844     ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4845     tsk.work(0);
4846   }
4847 
4848   // restore, single-threaded for now, any preserved marks
4849   // as a result of work_q overflow
4850   restore_preserved_marks_if_any();
4851 }
4852 
4853 // Non-parallel version of remark
4854 void CMSCollector::do_remark_non_parallel() {
4855   ResourceMark rm;
4856   HandleMark   hm;
4857   GenCollectedHeap* gch = GenCollectedHeap::heap();
4858   ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4859 
4860   MarkRefsIntoAndScanClosure
4861     mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4862              &_markStack, this,
4863              false /* should_yield */, false /* not precleaning */);
4864   MarkFromDirtyCardsClosure
4865     markFromDirtyCardsClosure(this, _span,
4866                               NULL,  // space is set further below
4867                               &_markBitMap, &_markStack, &mrias_cl);
4868   {
4869     GCTraceTime(Trace, gc) t("Grey Object Rescan", _gc_timer_cm);
4870     // Iterate over the dirty cards, setting the corresponding bits in the
4871     // mod union table.
4872     {
4873       ModUnionClosure modUnionClosure(&_modUnionTable);
4874       _ct->ct_bs()->dirty_card_iterate(
4875                       _cmsGen->used_region(),
4876                       &modUnionClosure);
4877     }
4878     // Having transferred these marks into the modUnionTable, we just need
4879     // to rescan the marked objects on the dirty cards in the modUnionTable.
4880     // The initial marking may have been done during an asynchronous
4881     // collection so there may be dirty bits in the mod-union table.
4882     const int alignment =
4883       CardTableModRefBS::card_size * BitsPerWord;
4884     {
4885       // ... First handle dirty cards in CMS gen
4886       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4887       MemRegion ur = _cmsGen->used_region();
4888       HeapWord* lb = ur.start();
4889       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
4890       MemRegion cms_span(lb, ub);
4891       _modUnionTable.dirty_range_iterate_clear(cms_span,
4892                                                &markFromDirtyCardsClosure);
4893       verify_work_stacks_empty();
4894       log_debug(gc, stats)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4895     }
4896   }
4897   if (VerifyDuringGC &&
4898       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4899     HandleMark hm;  // Discard invalid handles created during verification
4900     Universe::verify();
4901   }
4902   {
4903     GCTraceTime(Trace, gc) t("Root Rescan", _gc_timer_cm);
4904 
4905     verify_work_stacks_empty();
4906 
4907     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4908     StrongRootsScope srs(1);
4909 
4910     gch->gen_process_roots(&srs,
4911                            GenCollectedHeap::OldGen,
4912                            true,  // young gen as roots
4913                            GenCollectedHeap::ScanningOption(roots_scanning_options()),
4914                            should_unload_classes(),
4915                            &mrias_cl,
4916                            NULL,
4917                            NULL); // The dirty klasses will be handled below
4918 
4919     assert(should_unload_classes()
4920            || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4921            "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4922   }
4923 
4924   {
4925     GCTraceTime(Trace, gc) t("Visit Unhandled CLDs", _gc_timer_cm);
4926 
4927     verify_work_stacks_empty();
4928 
4929     // Scan all class loader data objects that might have been introduced
4930     // during concurrent marking.
4931     ResourceMark rm;
4932     GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4933     for (int i = 0; i < array->length(); i++) {
4934       mrias_cl.do_cld_nv(array->at(i));
4935     }
4936 
4937     // We don't need to keep track of new CLDs anymore.
4938     ClassLoaderDataGraph::remember_new_clds(false);
4939 
4940     verify_work_stacks_empty();
4941   }
4942 
4943   {
4944     GCTraceTime(Trace, gc) t("Dirty Klass Scan", _gc_timer_cm);
4945 
4946     verify_work_stacks_empty();
4947 
4948     RemarkKlassClosure remark_klass_closure(&mrias_cl);
4949     ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4950 
4951     verify_work_stacks_empty();
4952   }
4953 
4954   // We might have added oops to ClassLoaderData::_handles during the
4955   // concurrent marking phase. These oops point to newly allocated objects
4956   // that are guaranteed to be kept alive. Either by the direct allocation
4957   // code, or when the young collector processes the roots. Hence,
4958   // we don't have to revisit the _handles block during the remark phase.
4959 
4960   verify_work_stacks_empty();
4961   // Restore evacuated mark words, if any, used for overflow list links
4962   restore_preserved_marks_if_any();
4963 
4964   verify_overflow_empty();
4965 }
4966 
4967 ////////////////////////////////////////////////////////
4968 // Parallel Reference Processing Task Proxy Class
4969 ////////////////////////////////////////////////////////
4970 class AbstractGangTaskWOopQueues : public AbstractGangTask {
4971   OopTaskQueueSet*       _queues;
4972   ParallelTaskTerminator _terminator;
4973  public:
4974   AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
4975     AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
4976   ParallelTaskTerminator* terminator() { return &_terminator; }
4977   OopTaskQueueSet* queues() { return _queues; }
4978 };
4979 
4980 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
4981   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
4982   CMSCollector*          _collector;
4983   CMSBitMap*             _mark_bit_map;
4984   const MemRegion        _span;
4985   ProcessTask&           _task;
4986 
4987 public:
4988   CMSRefProcTaskProxy(ProcessTask&     task,
4989                       CMSCollector*    collector,
4990                       const MemRegion& span,
4991                       CMSBitMap*       mark_bit_map,
4992                       AbstractWorkGang* workers,
4993                       OopTaskQueueSet* task_queues):
4994     AbstractGangTaskWOopQueues("Process referents by policy in parallel",
4995       task_queues,
4996       workers->active_workers()),
4997     _task(task),
4998     _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
4999   {
5000     assert(_collector->_span.equals(_span) && !_span.is_empty(),
5001            "Inconsistency in _span");
5002   }
5003 
5004   OopTaskQueueSet* task_queues() { return queues(); }
5005 
5006   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5007 
5008   void do_work_steal(int i,
5009                      CMSParDrainMarkingStackClosure* drain,
5010                      CMSParKeepAliveClosure* keep_alive,
5011                      int* seed);
5012 
5013   virtual void work(uint worker_id);
5014 };
5015 
5016 void CMSRefProcTaskProxy::work(uint worker_id) {
5017   ResourceMark rm;
5018   HandleMark hm;
5019   assert(_collector->_span.equals(_span), "Inconsistency in _span");
5020   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5021                                         _mark_bit_map,
5022                                         work_queue(worker_id));
5023   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5024                                                  _mark_bit_map,
5025                                                  work_queue(worker_id));
5026   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5027   _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5028   if (_task.marks_oops_alive()) {
5029     do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5030                   _collector->hash_seed(worker_id));
5031   }
5032   assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5033   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5034 }
5035 
5036 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5037   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5038   EnqueueTask& _task;
5039 
5040 public:
5041   CMSRefEnqueueTaskProxy(EnqueueTask& task)
5042     : AbstractGangTask("Enqueue reference objects in parallel"),
5043       _task(task)
5044   { }
5045 
5046   virtual void work(uint worker_id)
5047   {
5048     _task.work(worker_id);
5049   }
5050 };
5051 
5052 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5053   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5054    _span(span),
5055    _bit_map(bit_map),
5056    _work_queue(work_queue),
5057    _mark_and_push(collector, span, bit_map, work_queue),
5058    _low_water_mark(MIN2((work_queue->max_elems()/4),
5059                         ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5060 { }
5061 
5062 // . see if we can share work_queues with ParNew? XXX
5063 void CMSRefProcTaskProxy::do_work_steal(int i,
5064   CMSParDrainMarkingStackClosure* drain,
5065   CMSParKeepAliveClosure* keep_alive,
5066   int* seed) {
5067   OopTaskQueue* work_q = work_queue(i);
5068   NOT_PRODUCT(int num_steals = 0;)
5069   oop obj_to_scan;
5070 
5071   while (true) {
5072     // Completely finish any left over work from (an) earlier round(s)
5073     drain->trim_queue(0);
5074     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5075                                          (size_t)ParGCDesiredObjsFromOverflowList);
5076     // Now check if there's any work in the overflow list
5077     // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5078     // only affects the number of attempts made to get work from the
5079     // overflow list and does not affect the number of workers.  Just
5080     // pass ParallelGCThreads so this behavior is unchanged.
5081     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5082                                                 work_q,
5083                                                 ParallelGCThreads)) {
5084       // Found something in global overflow list;
5085       // not yet ready to go stealing work from others.
5086       // We'd like to assert(work_q->size() != 0, ...)
5087       // because we just took work from the overflow list,
5088       // but of course we can't, since all of that might have
5089       // been already stolen from us.
5090       continue;
5091     }
5092     // Verify that we have no work before we resort to stealing
5093     assert(work_q->size() == 0, "Have work, shouldn't steal");
5094     // Try to steal from other queues that have work
5095     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5096       NOT_PRODUCT(num_steals++;)
5097       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5098       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5099       // Do scanning work
5100       obj_to_scan->oop_iterate(keep_alive);
5101       // Loop around, finish this work, and try to steal some more
5102     } else if (terminator()->offer_termination()) {
5103       break;  // nirvana from the infinite cycle
5104     }
5105   }
5106   log_develop(gc, stats)("\t(%d: stole %d oops)", i, num_steals);
5107 }
5108 
5109 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5110 {
5111   GenCollectedHeap* gch = GenCollectedHeap::heap();
5112   WorkGang* workers = gch->workers();
5113   assert(workers != NULL, "Need parallel worker threads.");
5114   CMSRefProcTaskProxy rp_task(task, &_collector,
5115                               _collector.ref_processor()->span(),
5116                               _collector.markBitMap(),
5117                               workers, _collector.task_queues());
5118   workers->run_task(&rp_task);
5119 }
5120 
5121 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5122 {
5123 
5124   GenCollectedHeap* gch = GenCollectedHeap::heap();
5125   WorkGang* workers = gch->workers();
5126   assert(workers != NULL, "Need parallel worker threads.");
5127   CMSRefEnqueueTaskProxy enq_task(task);
5128   workers->run_task(&enq_task);
5129 }
5130 
5131 void CMSCollector::refProcessingWork() {
5132   ResourceMark rm;
5133   HandleMark   hm;
5134 
5135   ReferenceProcessor* rp = ref_processor();
5136   assert(rp->span().equals(_span), "Spans should be equal");
5137   assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5138   // Process weak references.
5139   rp->setup_policy(false);
5140   verify_work_stacks_empty();
5141 
5142   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5143                                           &_markStack, false /* !preclean */);
5144   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5145                                 _span, &_markBitMap, &_markStack,
5146                                 &cmsKeepAliveClosure, false /* !preclean */);
5147   {
5148     GCTraceTime(Debug, gc) t("Weak Refs Processing", _gc_timer_cm);
5149 
5150     ReferenceProcessorStats stats;
5151     if (rp->processing_is_mt()) {
5152       // Set the degree of MT here.  If the discovery is done MT, there
5153       // may have been a different number of threads doing the discovery
5154       // and a different number of discovered lists may have Ref objects.
5155       // That is OK as long as the Reference lists are balanced (see
5156       // balance_all_queues() and balance_queues()).
5157       GenCollectedHeap* gch = GenCollectedHeap::heap();
5158       uint active_workers = ParallelGCThreads;
5159       WorkGang* workers = gch->workers();
5160       if (workers != NULL) {
5161         active_workers = workers->active_workers();
5162         // The expectation is that active_workers will have already
5163         // been set to a reasonable value.  If it has not been set,
5164         // investigate.
5165         assert(active_workers > 0, "Should have been set during scavenge");
5166       }
5167       rp->set_active_mt_degree(active_workers);
5168       CMSRefProcTaskExecutor task_executor(*this);
5169       stats = rp->process_discovered_references(&_is_alive_closure,
5170                                         &cmsKeepAliveClosure,
5171                                         &cmsDrainMarkingStackClosure,
5172                                         &task_executor,
5173                                         _gc_timer_cm);
5174     } else {
5175       stats = rp->process_discovered_references(&_is_alive_closure,
5176                                         &cmsKeepAliveClosure,
5177                                         &cmsDrainMarkingStackClosure,
5178                                         NULL,
5179                                         _gc_timer_cm);
5180     }
5181     _gc_tracer_cm->report_gc_reference_stats(stats);
5182 
5183   }
5184 
5185   // This is the point where the entire marking should have completed.
5186   verify_work_stacks_empty();
5187 
5188   if (should_unload_classes()) {
5189     {
5190       GCTraceTime(Debug, gc) t("Class Unloading", _gc_timer_cm);
5191 
5192       // Unload classes and purge the SystemDictionary.
5193       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5194 
5195       // Unload nmethods.
5196       CodeCache::do_unloading(&_is_alive_closure, purged_class);
5197 
5198       // Prune dead klasses from subklass/sibling/implementor lists.
5199       Klass::clean_weak_klass_links(&_is_alive_closure);
5200     }
5201 
5202     {
5203       GCTraceTime(Debug, gc) t("Scrub Symbol Table", _gc_timer_cm);
5204       // Clean up unreferenced symbols in symbol table.
5205       SymbolTable::unlink();
5206     }
5207 
5208     {
5209       GCTraceTime(Debug, gc) t("Scrub String Table", _gc_timer_cm);
5210       // Delete entries for dead interned strings.
5211       StringTable::unlink(&_is_alive_closure);
5212     }
5213   }
5214 
5215 
5216   // Restore any preserved marks as a result of mark stack or
5217   // work queue overflow
5218   restore_preserved_marks_if_any();  // done single-threaded for now
5219 
5220   rp->set_enqueuing_is_done(true);
5221   if (rp->processing_is_mt()) {
5222     rp->balance_all_queues();
5223     CMSRefProcTaskExecutor task_executor(*this);
5224     rp->enqueue_discovered_references(&task_executor);
5225   } else {
5226     rp->enqueue_discovered_references(NULL);
5227   }
5228   rp->verify_no_references_recorded();
5229   assert(!rp->discovery_enabled(), "should have been disabled");
5230 }
5231 
5232 #ifndef PRODUCT
5233 void CMSCollector::check_correct_thread_executing() {
5234   Thread* t = Thread::current();
5235   // Only the VM thread or the CMS thread should be here.
5236   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5237          "Unexpected thread type");
5238   // If this is the vm thread, the foreground process
5239   // should not be waiting.  Note that _foregroundGCIsActive is
5240   // true while the foreground collector is waiting.
5241   if (_foregroundGCShouldWait) {
5242     // We cannot be the VM thread
5243     assert(t->is_ConcurrentGC_thread(),
5244            "Should be CMS thread");
5245   } else {
5246     // We can be the CMS thread only if we are in a stop-world
5247     // phase of CMS collection.
5248     if (t->is_ConcurrentGC_thread()) {
5249       assert(_collectorState == InitialMarking ||
5250              _collectorState == FinalMarking,
5251              "Should be a stop-world phase");
5252       // The CMS thread should be holding the CMS_token.
5253       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5254              "Potential interference with concurrently "
5255              "executing VM thread");
5256     }
5257   }
5258 }
5259 #endif
5260 
5261 void CMSCollector::sweep() {
5262   assert(_collectorState == Sweeping, "just checking");
5263   check_correct_thread_executing();
5264   verify_work_stacks_empty();
5265   verify_overflow_empty();
5266   increment_sweep_count();
5267   TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5268 
5269   _inter_sweep_timer.stop();
5270   _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5271 
5272   assert(!_intra_sweep_timer.is_active(), "Should not be active");
5273   _intra_sweep_timer.reset();
5274   _intra_sweep_timer.start();
5275   {
5276     GCTraceCPUTime tcpu;
5277     CMSPhaseAccounting pa(this, "Concurrent Sweep");
5278     // First sweep the old gen
5279     {
5280       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5281                                bitMapLock());
5282       sweepWork(_cmsGen);
5283     }
5284 
5285     // Update Universe::_heap_*_at_gc figures.
5286     // We need all the free list locks to make the abstract state
5287     // transition from Sweeping to Resetting. See detailed note
5288     // further below.
5289     {
5290       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5291       // Update heap occupancy information which is used as
5292       // input to soft ref clearing policy at the next gc.
5293       Universe::update_heap_info_at_gc();
5294       _collectorState = Resizing;
5295     }
5296   }
5297   verify_work_stacks_empty();
5298   verify_overflow_empty();
5299 
5300   if (should_unload_classes()) {
5301     // Delay purge to the beginning of the next safepoint.  Metaspace::contains
5302     // requires that the virtual spaces are stable and not deleted.
5303     ClassLoaderDataGraph::set_should_purge(true);
5304   }
5305 
5306   _intra_sweep_timer.stop();
5307   _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5308 
5309   _inter_sweep_timer.reset();
5310   _inter_sweep_timer.start();
5311 
5312   // We need to use a monotonically non-decreasing time in ms
5313   // or we will see time-warp warnings and os::javaTimeMillis()
5314   // does not guarantee monotonicity.
5315   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5316   update_time_of_last_gc(now);
5317 
5318   // NOTE on abstract state transitions:
5319   // Mutators allocate-live and/or mark the mod-union table dirty
5320   // based on the state of the collection.  The former is done in
5321   // the interval [Marking, Sweeping] and the latter in the interval
5322   // [Marking, Sweeping).  Thus the transitions into the Marking state
5323   // and out of the Sweeping state must be synchronously visible
5324   // globally to the mutators.
5325   // The transition into the Marking state happens with the world
5326   // stopped so the mutators will globally see it.  Sweeping is
5327   // done asynchronously by the background collector so the transition
5328   // from the Sweeping state to the Resizing state must be done
5329   // under the freelistLock (as is the check for whether to
5330   // allocate-live and whether to dirty the mod-union table).
5331   assert(_collectorState == Resizing, "Change of collector state to"
5332     " Resizing must be done under the freelistLocks (plural)");
5333 
5334   // Now that sweeping has been completed, we clear
5335   // the incremental_collection_failed flag,
5336   // thus inviting a younger gen collection to promote into
5337   // this generation. If such a promotion may still fail,
5338   // the flag will be set again when a young collection is
5339   // attempted.
5340   GenCollectedHeap* gch = GenCollectedHeap::heap();
5341   gch->clear_incremental_collection_failed();  // Worth retrying as fresh space may have been freed up
5342   gch->update_full_collections_completed(_collection_count_start);
5343 }
5344 
5345 // FIX ME!!! Looks like this belongs in CFLSpace, with
5346 // CMSGen merely delegating to it.
5347 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5348   double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5349   HeapWord*  minAddr        = _cmsSpace->bottom();
5350   HeapWord*  largestAddr    =
5351     (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5352   if (largestAddr == NULL) {
5353     // The dictionary appears to be empty.  In this case
5354     // try to coalesce at the end of the heap.
5355     largestAddr = _cmsSpace->end();
5356   }
5357   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
5358   size_t nearLargestOffset =
5359     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5360   log_debug(gc, freelist, stats)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5361                                  p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5362   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5363 }
5364 
5365 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5366   return addr >= _cmsSpace->nearLargestChunk();
5367 }
5368 
5369 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5370   return _cmsSpace->find_chunk_at_end();
5371 }
5372 
5373 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5374                                                     bool full) {
5375   // If the young generation has been collected, gather any statistics
5376   // that are of interest at this point.
5377   bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5378   if (!full && current_is_young) {
5379     // Gather statistics on the young generation collection.
5380     collector()->stats().record_gc0_end(used());
5381   }
5382 }
5383 
5384 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5385   // We iterate over the space(s) underlying this generation,
5386   // checking the mark bit map to see if the bits corresponding
5387   // to specific blocks are marked or not. Blocks that are
5388   // marked are live and are not swept up. All remaining blocks
5389   // are swept up, with coalescing on-the-fly as we sweep up
5390   // contiguous free and/or garbage blocks:
5391   // We need to ensure that the sweeper synchronizes with allocators
5392   // and stop-the-world collectors. In particular, the following
5393   // locks are used:
5394   // . CMS token: if this is held, a stop the world collection cannot occur
5395   // . freelistLock: if this is held no allocation can occur from this
5396   //                 generation by another thread
5397   // . bitMapLock: if this is held, no other thread can access or update
5398   //
5399 
5400   // Note that we need to hold the freelistLock if we use
5401   // block iterate below; else the iterator might go awry if
5402   // a mutator (or promotion) causes block contents to change
5403   // (for instance if the allocator divvies up a block).
5404   // If we hold the free list lock, for all practical purposes
5405   // young generation GC's can't occur (they'll usually need to
5406   // promote), so we might as well prevent all young generation
5407   // GC's while we do a sweeping step. For the same reason, we might
5408   // as well take the bit map lock for the entire duration
5409 
5410   // check that we hold the requisite locks
5411   assert(have_cms_token(), "Should hold cms token");
5412   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5413   assert_lock_strong(old_gen->freelistLock());
5414   assert_lock_strong(bitMapLock());
5415 
5416   assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5417   assert(_intra_sweep_timer.is_active(),  "Was switched on  in an outer context");
5418   old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5419                                           _inter_sweep_estimate.padded_average(),
5420                                           _intra_sweep_estimate.padded_average());
5421   old_gen->setNearLargestChunk();
5422 
5423   {
5424     SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5425     old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5426     // We need to free-up/coalesce garbage/blocks from a
5427     // co-terminal free run. This is done in the SweepClosure
5428     // destructor; so, do not remove this scope, else the
5429     // end-of-sweep-census below will be off by a little bit.
5430   }
5431   old_gen->cmsSpace()->sweep_completed();
5432   old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5433   if (should_unload_classes()) {                // unloaded classes this cycle,
5434     _concurrent_cycles_since_last_unload = 0;   // ... reset count
5435   } else {                                      // did not unload classes,
5436     _concurrent_cycles_since_last_unload++;     // ... increment count
5437   }
5438 }
5439 
5440 // Reset CMS data structures (for now just the marking bit map)
5441 // preparatory for the next cycle.
5442 void CMSCollector::reset_concurrent() {
5443   CMSTokenSyncWithLocks ts(true, bitMapLock());
5444 
5445   // If the state is not "Resetting", the foreground  thread
5446   // has done a collection and the resetting.
5447   if (_collectorState != Resetting) {
5448     assert(_collectorState == Idling, "The state should only change"
5449       " because the foreground collector has finished the collection");
5450     return;
5451   }
5452 
5453   // Clear the mark bitmap (no grey objects to start with)
5454   // for the next cycle.
5455   GCTraceCPUTime tcpu;
5456   CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5457 
5458   HeapWord* curAddr = _markBitMap.startWord();
5459   while (curAddr < _markBitMap.endWord()) {
5460     size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
5461     MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5462     _markBitMap.clear_large_range(chunk);
5463     if (ConcurrentMarkSweepThread::should_yield() &&
5464         !foregroundGCIsActive() &&
5465         CMSYield) {
5466       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5467              "CMS thread should hold CMS token");
5468       assert_lock_strong(bitMapLock());
5469       bitMapLock()->unlock();
5470       ConcurrentMarkSweepThread::desynchronize(true);
5471       stopTimer();
5472       incrementYields();
5473 
5474       // See the comment in coordinator_yield()
5475       for (unsigned i = 0; i < CMSYieldSleepCount &&
5476                        ConcurrentMarkSweepThread::should_yield() &&
5477                        !CMSCollector::foregroundGCIsActive(); ++i) {
5478         os::sleep(Thread::current(), 1, false);
5479       }
5480 
5481       ConcurrentMarkSweepThread::synchronize(true);
5482       bitMapLock()->lock_without_safepoint_check();
5483       startTimer();
5484     }
5485     curAddr = chunk.end();
5486   }
5487   // A successful mostly concurrent collection has been done.
5488   // Because only the full (i.e., concurrent mode failure) collections
5489   // are being measured for gc overhead limits, clean the "near" flag
5490   // and count.
5491   size_policy()->reset_gc_overhead_limit_count();
5492   _collectorState = Idling;
5493 
5494   register_gc_end();
5495 }
5496 
5497 // Same as above but for STW paths
5498 void CMSCollector::reset_stw() {
5499   // already have the lock
5500   assert(_collectorState == Resetting, "just checking");
5501   assert_lock_strong(bitMapLock());
5502   GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id());
5503   _markBitMap.clear_all();
5504   _collectorState = Idling;
5505   register_gc_end();
5506 }
5507 
5508 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5509   GCTraceCPUTime tcpu;
5510   TraceCollectorStats tcs(counters());
5511 
5512   switch (op) {
5513     case CMS_op_checkpointRootsInitial: {
5514       GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5515       SvcGCMarker sgcm(SvcGCMarker::OTHER);
5516       checkpointRootsInitial();
5517       break;
5518     }
5519     case CMS_op_checkpointRootsFinal: {
5520       GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5521       SvcGCMarker sgcm(SvcGCMarker::OTHER);
5522       checkpointRootsFinal();
5523       break;
5524     }
5525     default:
5526       fatal("No such CMS_op");
5527   }
5528 }
5529 
5530 #ifndef PRODUCT
5531 size_t const CMSCollector::skip_header_HeapWords() {
5532   return FreeChunk::header_size();
5533 }
5534 
5535 // Try and collect here conditions that should hold when
5536 // CMS thread is exiting. The idea is that the foreground GC
5537 // thread should not be blocked if it wants to terminate
5538 // the CMS thread and yet continue to run the VM for a while
5539 // after that.
5540 void CMSCollector::verify_ok_to_terminate() const {
5541   assert(Thread::current()->is_ConcurrentGC_thread(),
5542          "should be called by CMS thread");
5543   assert(!_foregroundGCShouldWait, "should be false");
5544   // We could check here that all the various low-level locks
5545   // are not held by the CMS thread, but that is overkill; see
5546   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5547   // is checked.
5548 }
5549 #endif
5550 
5551 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5552    assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5553           "missing Printezis mark?");
5554   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5555   size_t size = pointer_delta(nextOneAddr + 1, addr);
5556   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5557          "alignment problem");
5558   assert(size >= 3, "Necessary for Printezis marks to work");
5559   return size;
5560 }
5561 
5562 // A variant of the above (block_size_using_printezis_bits()) except
5563 // that we return 0 if the P-bits are not yet set.
5564 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5565   if (_markBitMap.isMarked(addr + 1)) {
5566     assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5567     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5568     size_t size = pointer_delta(nextOneAddr + 1, addr);
5569     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5570            "alignment problem");
5571     assert(size >= 3, "Necessary for Printezis marks to work");
5572     return size;
5573   }
5574   return 0;
5575 }
5576 
5577 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5578   size_t sz = 0;
5579   oop p = (oop)addr;
5580   if (p->klass_or_null() != NULL) {
5581     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5582   } else {
5583     sz = block_size_using_printezis_bits(addr);
5584   }
5585   assert(sz > 0, "size must be nonzero");
5586   HeapWord* next_block = addr + sz;
5587   HeapWord* next_card  = (HeapWord*)round_to((uintptr_t)next_block,
5588                                              CardTableModRefBS::card_size);
5589   assert(round_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
5590          round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5591          "must be different cards");
5592   return next_card;
5593 }
5594 
5595 
5596 // CMS Bit Map Wrapper /////////////////////////////////////////
5597 
5598 // Construct a CMS bit map infrastructure, but don't create the
5599 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5600 // further below.
5601 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5602   _bm(),
5603   _shifter(shifter),
5604   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5605                                     Monitor::_safepoint_check_sometimes) : NULL)
5606 {
5607   _bmStartWord = 0;
5608   _bmWordSize  = 0;
5609 }
5610 
5611 bool CMSBitMap::allocate(MemRegion mr) {
5612   _bmStartWord = mr.start();
5613   _bmWordSize  = mr.word_size();
5614   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5615                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5616   if (!brs.is_reserved()) {
5617     warning("CMS bit map allocation failure");
5618     return false;
5619   }
5620   // For now we'll just commit all of the bit map up front.
5621   // Later on we'll try to be more parsimonious with swap.
5622   if (!_virtual_space.initialize(brs, brs.size())) {
5623     warning("CMS bit map backing store failure");
5624     return false;
5625   }
5626   assert(_virtual_space.committed_size() == brs.size(),
5627          "didn't reserve backing store for all of CMS bit map?");
5628   _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5629   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5630          _bmWordSize, "inconsistency in bit map sizing");
5631   _bm.set_size(_bmWordSize >> _shifter);
5632 
5633   // bm.clear(); // can we rely on getting zero'd memory? verify below
5634   assert(isAllClear(),
5635          "Expected zero'd memory from ReservedSpace constructor");
5636   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5637          "consistency check");
5638   return true;
5639 }
5640 
5641 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5642   HeapWord *next_addr, *end_addr, *last_addr;
5643   assert_locked();
5644   assert(covers(mr), "out-of-range error");
5645   // XXX assert that start and end are appropriately aligned
5646   for (next_addr = mr.start(), end_addr = mr.end();
5647        next_addr < end_addr; next_addr = last_addr) {
5648     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5649     last_addr = dirty_region.end();
5650     if (!dirty_region.is_empty()) {
5651       cl->do_MemRegion(dirty_region);
5652     } else {
5653       assert(last_addr == end_addr, "program logic");
5654       return;
5655     }
5656   }
5657 }
5658 
5659 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5660   _bm.print_on_error(st, prefix);
5661 }
5662 
5663 #ifndef PRODUCT
5664 void CMSBitMap::assert_locked() const {
5665   CMSLockVerifier::assert_locked(lock());
5666 }
5667 
5668 bool CMSBitMap::covers(MemRegion mr) const {
5669   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5670   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5671          "size inconsistency");
5672   return (mr.start() >= _bmStartWord) &&
5673          (mr.end()   <= endWord());
5674 }
5675 
5676 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5677     return (start >= _bmStartWord && (start + size) <= endWord());
5678 }
5679 
5680 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5681   // verify that there are no 1 bits in the interval [left, right)
5682   FalseBitMapClosure falseBitMapClosure;
5683   iterate(&falseBitMapClosure, left, right);
5684 }
5685 
5686 void CMSBitMap::region_invariant(MemRegion mr)
5687 {
5688   assert_locked();
5689   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5690   assert(!mr.is_empty(), "unexpected empty region");
5691   assert(covers(mr), "mr should be covered by bit map");
5692   // convert address range into offset range
5693   size_t start_ofs = heapWordToOffset(mr.start());
5694   // Make sure that end() is appropriately aligned
5695   assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5696                         (1 << (_shifter+LogHeapWordSize))),
5697          "Misaligned mr.end()");
5698   size_t end_ofs   = heapWordToOffset(mr.end());
5699   assert(end_ofs > start_ofs, "Should mark at least one bit");
5700 }
5701 
5702 #endif
5703 
5704 bool CMSMarkStack::allocate(size_t size) {
5705   // allocate a stack of the requisite depth
5706   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5707                    size * sizeof(oop)));
5708   if (!rs.is_reserved()) {
5709     warning("CMSMarkStack allocation failure");
5710     return false;
5711   }
5712   if (!_virtual_space.initialize(rs, rs.size())) {
5713     warning("CMSMarkStack backing store failure");
5714     return false;
5715   }
5716   assert(_virtual_space.committed_size() == rs.size(),
5717          "didn't reserve backing store for all of CMS stack?");
5718   _base = (oop*)(_virtual_space.low());
5719   _index = 0;
5720   _capacity = size;
5721   NOT_PRODUCT(_max_depth = 0);
5722   return true;
5723 }
5724 
5725 // XXX FIX ME !!! In the MT case we come in here holding a
5726 // leaf lock. For printing we need to take a further lock
5727 // which has lower rank. We need to recalibrate the two
5728 // lock-ranks involved in order to be able to print the
5729 // messages below. (Or defer the printing to the caller.
5730 // For now we take the expedient path of just disabling the
5731 // messages for the problematic case.)
5732 void CMSMarkStack::expand() {
5733   assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5734   if (_capacity == MarkStackSizeMax) {
5735     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5736       // We print a warning message only once per CMS cycle.
5737       log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5738     }
5739     return;
5740   }
5741   // Double capacity if possible
5742   size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5743   // Do not give up existing stack until we have managed to
5744   // get the double capacity that we desired.
5745   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5746                    new_capacity * sizeof(oop)));
5747   if (rs.is_reserved()) {
5748     // Release the backing store associated with old stack
5749     _virtual_space.release();
5750     // Reinitialize virtual space for new stack
5751     if (!_virtual_space.initialize(rs, rs.size())) {
5752       fatal("Not enough swap for expanded marking stack");
5753     }
5754     _base = (oop*)(_virtual_space.low());
5755     _index = 0;
5756     _capacity = new_capacity;
5757   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5758     // Failed to double capacity, continue;
5759     // we print a detail message only once per CMS cycle.
5760     log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5761                         _capacity / K, new_capacity / K);
5762   }
5763 }
5764 
5765 
5766 // Closures
5767 // XXX: there seems to be a lot of code  duplication here;
5768 // should refactor and consolidate common code.
5769 
5770 // This closure is used to mark refs into the CMS generation in
5771 // the CMS bit map. Called at the first checkpoint. This closure
5772 // assumes that we do not need to re-mark dirty cards; if the CMS
5773 // generation on which this is used is not an oldest
5774 // generation then this will lose younger_gen cards!
5775 
5776 MarkRefsIntoClosure::MarkRefsIntoClosure(
5777   MemRegion span, CMSBitMap* bitMap):
5778     _span(span),
5779     _bitMap(bitMap)
5780 {
5781   assert(ref_processor() == NULL, "deliberately left NULL");
5782   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5783 }
5784 
5785 void MarkRefsIntoClosure::do_oop(oop obj) {
5786   // if p points into _span, then mark corresponding bit in _markBitMap
5787   assert(obj->is_oop(), "expected an oop");
5788   HeapWord* addr = (HeapWord*)obj;
5789   if (_span.contains(addr)) {
5790     // this should be made more efficient
5791     _bitMap->mark(addr);
5792   }
5793 }
5794 
5795 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
5796 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5797 
5798 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
5799   MemRegion span, CMSBitMap* bitMap):
5800     _span(span),
5801     _bitMap(bitMap)
5802 {
5803   assert(ref_processor() == NULL, "deliberately left NULL");
5804   assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5805 }
5806 
5807 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
5808   // if p points into _span, then mark corresponding bit in _markBitMap
5809   assert(obj->is_oop(), "expected an oop");
5810   HeapWord* addr = (HeapWord*)obj;
5811   if (_span.contains(addr)) {
5812     // this should be made more efficient
5813     _bitMap->par_mark(addr);
5814   }
5815 }
5816 
5817 void Par_MarkRefsIntoClosure::do_oop(oop* p)       { Par_MarkRefsIntoClosure::do_oop_work(p); }
5818 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
5819 
5820 // A variant of the above, used for CMS marking verification.
5821 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5822   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5823     _span(span),
5824     _verification_bm(verification_bm),
5825     _cms_bm(cms_bm)
5826 {
5827   assert(ref_processor() == NULL, "deliberately left NULL");
5828   assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5829 }
5830 
5831 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5832   // if p points into _span, then mark corresponding bit in _markBitMap
5833   assert(obj->is_oop(), "expected an oop");
5834   HeapWord* addr = (HeapWord*)obj;
5835   if (_span.contains(addr)) {
5836     _verification_bm->mark(addr);
5837     if (!_cms_bm->isMarked(addr)) {
5838       LogHandle(gc, verify) log;
5839       ResourceMark rm;
5840       oop(addr)->print_on(log.info_stream());
5841       log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5842       fatal("... aborting");
5843     }
5844   }
5845 }
5846 
5847 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5848 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5849 
5850 //////////////////////////////////////////////////
5851 // MarkRefsIntoAndScanClosure
5852 //////////////////////////////////////////////////
5853 
5854 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5855                                                        ReferenceProcessor* rp,
5856                                                        CMSBitMap* bit_map,
5857                                                        CMSBitMap* mod_union_table,
5858                                                        CMSMarkStack*  mark_stack,
5859                                                        CMSCollector* collector,
5860                                                        bool should_yield,
5861                                                        bool concurrent_precleaning):
5862   _collector(collector),
5863   _span(span),
5864   _bit_map(bit_map),
5865   _mark_stack(mark_stack),
5866   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
5867                       mark_stack, concurrent_precleaning),
5868   _yield(should_yield),
5869   _concurrent_precleaning(concurrent_precleaning),
5870   _freelistLock(NULL)
5871 {
5872   // FIXME: Should initialize in base class constructor.
5873   assert(rp != NULL, "ref_processor shouldn't be NULL");
5874   set_ref_processor_internal(rp);
5875 }
5876 
5877 // This closure is used to mark refs into the CMS generation at the
5878 // second (final) checkpoint, and to scan and transitively follow
5879 // the unmarked oops. It is also used during the concurrent precleaning
5880 // phase while scanning objects on dirty cards in the CMS generation.
5881 // The marks are made in the marking bit map and the marking stack is
5882 // used for keeping the (newly) grey objects during the scan.
5883 // The parallel version (Par_...) appears further below.
5884 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5885   if (obj != NULL) {
5886     assert(obj->is_oop(), "expected an oop");
5887     HeapWord* addr = (HeapWord*)obj;
5888     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5889     assert(_collector->overflow_list_is_empty(),
5890            "overflow list should be empty");
5891     if (_span.contains(addr) &&
5892         !_bit_map->isMarked(addr)) {
5893       // mark bit map (object is now grey)
5894       _bit_map->mark(addr);
5895       // push on marking stack (stack should be empty), and drain the
5896       // stack by applying this closure to the oops in the oops popped
5897       // from the stack (i.e. blacken the grey objects)
5898       bool res = _mark_stack->push(obj);
5899       assert(res, "Should have space to push on empty stack");
5900       do {
5901         oop new_oop = _mark_stack->pop();
5902         assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
5903         assert(_bit_map->isMarked((HeapWord*)new_oop),
5904                "only grey objects on this stack");
5905         // iterate over the oops in this oop, marking and pushing
5906         // the ones in CMS heap (i.e. in _span).
5907         new_oop->oop_iterate(&_pushAndMarkClosure);
5908         // check if it's time to yield
5909         do_yield_check();
5910       } while (!_mark_stack->isEmpty() ||
5911                (!_concurrent_precleaning && take_from_overflow_list()));
5912         // if marking stack is empty, and we are not doing this
5913         // during precleaning, then check the overflow list
5914     }
5915     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5916     assert(_collector->overflow_list_is_empty(),
5917            "overflow list was drained above");
5918 
5919     assert(_collector->no_preserved_marks(),
5920            "All preserved marks should have been restored above");
5921   }
5922 }
5923 
5924 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5925 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5926 
5927 void MarkRefsIntoAndScanClosure::do_yield_work() {
5928   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5929          "CMS thread should hold CMS token");
5930   assert_lock_strong(_freelistLock);
5931   assert_lock_strong(_bit_map->lock());
5932   // relinquish the free_list_lock and bitMaplock()
5933   _bit_map->lock()->unlock();
5934   _freelistLock->unlock();
5935   ConcurrentMarkSweepThread::desynchronize(true);
5936   _collector->stopTimer();
5937   _collector->incrementYields();
5938 
5939   // See the comment in coordinator_yield()
5940   for (unsigned i = 0;
5941        i < CMSYieldSleepCount &&
5942        ConcurrentMarkSweepThread::should_yield() &&
5943        !CMSCollector::foregroundGCIsActive();
5944        ++i) {
5945     os::sleep(Thread::current(), 1, false);
5946   }
5947 
5948   ConcurrentMarkSweepThread::synchronize(true);
5949   _freelistLock->lock_without_safepoint_check();
5950   _bit_map->lock()->lock_without_safepoint_check();
5951   _collector->startTimer();
5952 }
5953 
5954 ///////////////////////////////////////////////////////////
5955 // Par_MarkRefsIntoAndScanClosure: a parallel version of
5956 //                                 MarkRefsIntoAndScanClosure
5957 ///////////////////////////////////////////////////////////
5958 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
5959   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
5960   CMSBitMap* bit_map, OopTaskQueue* work_queue):
5961   _span(span),
5962   _bit_map(bit_map),
5963   _work_queue(work_queue),
5964   _low_water_mark(MIN2((work_queue->max_elems()/4),
5965                        ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
5966   _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
5967 {
5968   // FIXME: Should initialize in base class constructor.
5969   assert(rp != NULL, "ref_processor shouldn't be NULL");
5970   set_ref_processor_internal(rp);
5971 }
5972 
5973 // This closure is used to mark refs into the CMS generation at the
5974 // second (final) checkpoint, and to scan and transitively follow
5975 // the unmarked oops. The marks are made in the marking bit map and
5976 // the work_queue is used for keeping the (newly) grey objects during
5977 // the scan phase whence they are also available for stealing by parallel
5978 // threads. Since the marking bit map is shared, updates are
5979 // synchronized (via CAS).
5980 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5981   if (obj != NULL) {
5982     // Ignore mark word because this could be an already marked oop
5983     // that may be chained at the end of the overflow list.
5984     assert(obj->is_oop(true), "expected an oop");
5985     HeapWord* addr = (HeapWord*)obj;
5986     if (_span.contains(addr) &&
5987         !_bit_map->isMarked(addr)) {
5988       // mark bit map (object will become grey):
5989       // It is possible for several threads to be
5990       // trying to "claim" this object concurrently;
5991       // the unique thread that succeeds in marking the
5992       // object first will do the subsequent push on
5993       // to the work queue (or overflow list).
5994       if (_bit_map->par_mark(addr)) {
5995         // push on work_queue (which may not be empty), and trim the
5996         // queue to an appropriate length by applying this closure to
5997         // the oops in the oops popped from the stack (i.e. blacken the
5998         // grey objects)
5999         bool res = _work_queue->push(obj);
6000         assert(res, "Low water mark should be less than capacity?");
6001         trim_queue(_low_water_mark);
6002       } // Else, another thread claimed the object
6003     }
6004   }
6005 }
6006 
6007 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p)       { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6008 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6009 
6010 // This closure is used to rescan the marked objects on the dirty cards
6011 // in the mod union table and the card table proper.
6012 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6013   oop p, MemRegion mr) {
6014 
6015   size_t size = 0;
6016   HeapWord* addr = (HeapWord*)p;
6017   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6018   assert(_span.contains(addr), "we are scanning the CMS generation");
6019   // check if it's time to yield
6020   if (do_yield_check()) {
6021     // We yielded for some foreground stop-world work,
6022     // and we have been asked to abort this ongoing preclean cycle.
6023     return 0;
6024   }
6025   if (_bitMap->isMarked(addr)) {
6026     // it's marked; is it potentially uninitialized?
6027     if (p->klass_or_null() != NULL) {
6028         // an initialized object; ignore mark word in verification below
6029         // since we are running concurrent with mutators
6030         assert(p->is_oop(true), "should be an oop");
6031         if (p->is_objArray()) {
6032           // objArrays are precisely marked; restrict scanning
6033           // to dirty cards only.
6034           size = CompactibleFreeListSpace::adjustObjectSize(
6035                    p->oop_iterate_size(_scanningClosure, mr));
6036         } else {
6037           // A non-array may have been imprecisely marked; we need
6038           // to scan object in its entirety.
6039           size = CompactibleFreeListSpace::adjustObjectSize(
6040                    p->oop_iterate_size(_scanningClosure));
6041         }
6042         #ifdef ASSERT
6043           size_t direct_size =
6044             CompactibleFreeListSpace::adjustObjectSize(p->size());
6045           assert(size == direct_size, "Inconsistency in size");
6046           assert(size >= 3, "Necessary for Printezis marks to work");
6047           if (!_bitMap->isMarked(addr+1)) {
6048             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6049           } else {
6050             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6051             assert(_bitMap->isMarked(addr+size-1),
6052                    "inconsistent Printezis mark");
6053           }
6054         #endif // ASSERT
6055     } else {
6056       // An uninitialized object.
6057       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6058       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6059       size = pointer_delta(nextOneAddr + 1, addr);
6060       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6061              "alignment problem");
6062       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6063       // will dirty the card when the klass pointer is installed in the
6064       // object (signaling the completion of initialization).
6065     }
6066   } else {
6067     // Either a not yet marked object or an uninitialized object
6068     if (p->klass_or_null() == NULL) {
6069       // An uninitialized object, skip to the next card, since
6070       // we may not be able to read its P-bits yet.
6071       assert(size == 0, "Initial value");
6072     } else {
6073       // An object not (yet) reached by marking: we merely need to
6074       // compute its size so as to go look at the next block.
6075       assert(p->is_oop(true), "should be an oop");
6076       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6077     }
6078   }
6079   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6080   return size;
6081 }
6082 
6083 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6084   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6085          "CMS thread should hold CMS token");
6086   assert_lock_strong(_freelistLock);
6087   assert_lock_strong(_bitMap->lock());
6088   // relinquish the free_list_lock and bitMaplock()
6089   _bitMap->lock()->unlock();
6090   _freelistLock->unlock();
6091   ConcurrentMarkSweepThread::desynchronize(true);
6092   _collector->stopTimer();
6093   _collector->incrementYields();
6094 
6095   // See the comment in coordinator_yield()
6096   for (unsigned i = 0; i < CMSYieldSleepCount &&
6097                    ConcurrentMarkSweepThread::should_yield() &&
6098                    !CMSCollector::foregroundGCIsActive(); ++i) {
6099     os::sleep(Thread::current(), 1, false);
6100   }
6101 
6102   ConcurrentMarkSweepThread::synchronize(true);
6103   _freelistLock->lock_without_safepoint_check();
6104   _bitMap->lock()->lock_without_safepoint_check();
6105   _collector->startTimer();
6106 }
6107 
6108 
6109 //////////////////////////////////////////////////////////////////
6110 // SurvivorSpacePrecleanClosure
6111 //////////////////////////////////////////////////////////////////
6112 // This (single-threaded) closure is used to preclean the oops in
6113 // the survivor spaces.
6114 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6115 
6116   HeapWord* addr = (HeapWord*)p;
6117   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6118   assert(!_span.contains(addr), "we are scanning the survivor spaces");
6119   assert(p->klass_or_null() != NULL, "object should be initialized");
6120   // an initialized object; ignore mark word in verification below
6121   // since we are running concurrent with mutators
6122   assert(p->is_oop(true), "should be an oop");
6123   // Note that we do not yield while we iterate over
6124   // the interior oops of p, pushing the relevant ones
6125   // on our marking stack.
6126   size_t size = p->oop_iterate_size(_scanning_closure);
6127   do_yield_check();
6128   // Observe that below, we do not abandon the preclean
6129   // phase as soon as we should; rather we empty the
6130   // marking stack before returning. This is to satisfy
6131   // some existing assertions. In general, it may be a
6132   // good idea to abort immediately and complete the marking
6133   // from the grey objects at a later time.
6134   while (!_mark_stack->isEmpty()) {
6135     oop new_oop = _mark_stack->pop();
6136     assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6137     assert(_bit_map->isMarked((HeapWord*)new_oop),
6138            "only grey objects on this stack");
6139     // iterate over the oops in this oop, marking and pushing
6140     // the ones in CMS heap (i.e. in _span).
6141     new_oop->oop_iterate(_scanning_closure);
6142     // check if it's time to yield
6143     do_yield_check();
6144   }
6145   unsigned int after_count =
6146     GenCollectedHeap::heap()->total_collections();
6147   bool abort = (_before_count != after_count) ||
6148                _collector->should_abort_preclean();
6149   return abort ? 0 : size;
6150 }
6151 
6152 void SurvivorSpacePrecleanClosure::do_yield_work() {
6153   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6154          "CMS thread should hold CMS token");
6155   assert_lock_strong(_bit_map->lock());
6156   // Relinquish the bit map lock
6157   _bit_map->lock()->unlock();
6158   ConcurrentMarkSweepThread::desynchronize(true);
6159   _collector->stopTimer();
6160   _collector->incrementYields();
6161 
6162   // See the comment in coordinator_yield()
6163   for (unsigned i = 0; i < CMSYieldSleepCount &&
6164                        ConcurrentMarkSweepThread::should_yield() &&
6165                        !CMSCollector::foregroundGCIsActive(); ++i) {
6166     os::sleep(Thread::current(), 1, false);
6167   }
6168 
6169   ConcurrentMarkSweepThread::synchronize(true);
6170   _bit_map->lock()->lock_without_safepoint_check();
6171   _collector->startTimer();
6172 }
6173 
6174 // This closure is used to rescan the marked objects on the dirty cards
6175 // in the mod union table and the card table proper. In the parallel
6176 // case, although the bitMap is shared, we do a single read so the
6177 // isMarked() query is "safe".
6178 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6179   // Ignore mark word because we are running concurrent with mutators
6180   assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6181   HeapWord* addr = (HeapWord*)p;
6182   assert(_span.contains(addr), "we are scanning the CMS generation");
6183   bool is_obj_array = false;
6184   #ifdef ASSERT
6185     if (!_parallel) {
6186       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6187       assert(_collector->overflow_list_is_empty(),
6188              "overflow list should be empty");
6189 
6190     }
6191   #endif // ASSERT
6192   if (_bit_map->isMarked(addr)) {
6193     // Obj arrays are precisely marked, non-arrays are not;
6194     // so we scan objArrays precisely and non-arrays in their
6195     // entirety.
6196     if (p->is_objArray()) {
6197       is_obj_array = true;
6198       if (_parallel) {
6199         p->oop_iterate(_par_scan_closure, mr);
6200       } else {
6201         p->oop_iterate(_scan_closure, mr);
6202       }
6203     } else {
6204       if (_parallel) {
6205         p->oop_iterate(_par_scan_closure);
6206       } else {
6207         p->oop_iterate(_scan_closure);
6208       }
6209     }
6210   }
6211   #ifdef ASSERT
6212     if (!_parallel) {
6213       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6214       assert(_collector->overflow_list_is_empty(),
6215              "overflow list should be empty");
6216 
6217     }
6218   #endif // ASSERT
6219   return is_obj_array;
6220 }
6221 
6222 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6223                         MemRegion span,
6224                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
6225                         bool should_yield, bool verifying):
6226   _collector(collector),
6227   _span(span),
6228   _bitMap(bitMap),
6229   _mut(&collector->_modUnionTable),
6230   _markStack(markStack),
6231   _yield(should_yield),
6232   _skipBits(0)
6233 {
6234   assert(_markStack->isEmpty(), "stack should be empty");
6235   _finger = _bitMap->startWord();
6236   _threshold = _finger;
6237   assert(_collector->_restart_addr == NULL, "Sanity check");
6238   assert(_span.contains(_finger), "Out of bounds _finger?");
6239   DEBUG_ONLY(_verifying = verifying;)
6240 }
6241 
6242 void MarkFromRootsClosure::reset(HeapWord* addr) {
6243   assert(_markStack->isEmpty(), "would cause duplicates on stack");
6244   assert(_span.contains(addr), "Out of bounds _finger?");
6245   _finger = addr;
6246   _threshold = (HeapWord*)round_to(
6247                  (intptr_t)_finger, CardTableModRefBS::card_size);
6248 }
6249 
6250 // Should revisit to see if this should be restructured for
6251 // greater efficiency.
6252 bool MarkFromRootsClosure::do_bit(size_t offset) {
6253   if (_skipBits > 0) {
6254     _skipBits--;
6255     return true;
6256   }
6257   // convert offset into a HeapWord*
6258   HeapWord* addr = _bitMap->startWord() + offset;
6259   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6260          "address out of range");
6261   assert(_bitMap->isMarked(addr), "tautology");
6262   if (_bitMap->isMarked(addr+1)) {
6263     // this is an allocated but not yet initialized object
6264     assert(_skipBits == 0, "tautology");
6265     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
6266     oop p = oop(addr);
6267     if (p->klass_or_null() == NULL) {
6268       DEBUG_ONLY(if (!_verifying) {)
6269         // We re-dirty the cards on which this object lies and increase
6270         // the _threshold so that we'll come back to scan this object
6271         // during the preclean or remark phase. (CMSCleanOnEnter)
6272         if (CMSCleanOnEnter) {
6273           size_t sz = _collector->block_size_using_printezis_bits(addr);
6274           HeapWord* end_card_addr   = (HeapWord*)round_to(
6275                                          (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6276           MemRegion redirty_range = MemRegion(addr, end_card_addr);
6277           assert(!redirty_range.is_empty(), "Arithmetical tautology");
6278           // Bump _threshold to end_card_addr; note that
6279           // _threshold cannot possibly exceed end_card_addr, anyhow.
6280           // This prevents future clearing of the card as the scan proceeds
6281           // to the right.
6282           assert(_threshold <= end_card_addr,
6283                  "Because we are just scanning into this object");
6284           if (_threshold < end_card_addr) {
6285             _threshold = end_card_addr;
6286           }
6287           if (p->klass_or_null() != NULL) {
6288             // Redirty the range of cards...
6289             _mut->mark_range(redirty_range);
6290           } // ...else the setting of klass will dirty the card anyway.
6291         }
6292       DEBUG_ONLY(})
6293       return true;
6294     }
6295   }
6296   scanOopsInOop(addr);
6297   return true;
6298 }
6299 
6300 // We take a break if we've been at this for a while,
6301 // so as to avoid monopolizing the locks involved.
6302 void MarkFromRootsClosure::do_yield_work() {
6303   // First give up the locks, then yield, then re-lock
6304   // We should probably use a constructor/destructor idiom to
6305   // do this unlock/lock or modify the MutexUnlocker class to
6306   // serve our purpose. XXX
6307   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6308          "CMS thread should hold CMS token");
6309   assert_lock_strong(_bitMap->lock());
6310   _bitMap->lock()->unlock();
6311   ConcurrentMarkSweepThread::desynchronize(true);
6312   _collector->stopTimer();
6313   _collector->incrementYields();
6314 
6315   // See the comment in coordinator_yield()
6316   for (unsigned i = 0; i < CMSYieldSleepCount &&
6317                        ConcurrentMarkSweepThread::should_yield() &&
6318                        !CMSCollector::foregroundGCIsActive(); ++i) {
6319     os::sleep(Thread::current(), 1, false);
6320   }
6321 
6322   ConcurrentMarkSweepThread::synchronize(true);
6323   _bitMap->lock()->lock_without_safepoint_check();
6324   _collector->startTimer();
6325 }
6326 
6327 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6328   assert(_bitMap->isMarked(ptr), "expected bit to be set");
6329   assert(_markStack->isEmpty(),
6330          "should drain stack to limit stack usage");
6331   // convert ptr to an oop preparatory to scanning
6332   oop obj = oop(ptr);
6333   // Ignore mark word in verification below, since we
6334   // may be running concurrent with mutators.
6335   assert(obj->is_oop(true), "should be an oop");
6336   assert(_finger <= ptr, "_finger runneth ahead");
6337   // advance the finger to right end of this object
6338   _finger = ptr + obj->size();
6339   assert(_finger > ptr, "we just incremented it above");
6340   // On large heaps, it may take us some time to get through
6341   // the marking phase. During
6342   // this time it's possible that a lot of mutations have
6343   // accumulated in the card table and the mod union table --
6344   // these mutation records are redundant until we have
6345   // actually traced into the corresponding card.
6346   // Here, we check whether advancing the finger would make
6347   // us cross into a new card, and if so clear corresponding
6348   // cards in the MUT (preclean them in the card-table in the
6349   // future).
6350 
6351   DEBUG_ONLY(if (!_verifying) {)
6352     // The clean-on-enter optimization is disabled by default,
6353     // until we fix 6178663.
6354     if (CMSCleanOnEnter && (_finger > _threshold)) {
6355       // [_threshold, _finger) represents the interval
6356       // of cards to be cleared  in MUT (or precleaned in card table).
6357       // The set of cards to be cleared is all those that overlap
6358       // with the interval [_threshold, _finger); note that
6359       // _threshold is always kept card-aligned but _finger isn't
6360       // always card-aligned.
6361       HeapWord* old_threshold = _threshold;
6362       assert(old_threshold == (HeapWord*)round_to(
6363               (intptr_t)old_threshold, CardTableModRefBS::card_size),
6364              "_threshold should always be card-aligned");
6365       _threshold = (HeapWord*)round_to(
6366                      (intptr_t)_finger, CardTableModRefBS::card_size);
6367       MemRegion mr(old_threshold, _threshold);
6368       assert(!mr.is_empty(), "Control point invariant");
6369       assert(_span.contains(mr), "Should clear within span");
6370       _mut->clear_range(mr);
6371     }
6372   DEBUG_ONLY(})
6373   // Note: the finger doesn't advance while we drain
6374   // the stack below.
6375   PushOrMarkClosure pushOrMarkClosure(_collector,
6376                                       _span, _bitMap, _markStack,
6377                                       _finger, this);
6378   bool res = _markStack->push(obj);
6379   assert(res, "Empty non-zero size stack should have space for single push");
6380   while (!_markStack->isEmpty()) {
6381     oop new_oop = _markStack->pop();
6382     // Skip verifying header mark word below because we are
6383     // running concurrent with mutators.
6384     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6385     // now scan this oop's oops
6386     new_oop->oop_iterate(&pushOrMarkClosure);
6387     do_yield_check();
6388   }
6389   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6390 }
6391 
6392 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6393                        CMSCollector* collector, MemRegion span,
6394                        CMSBitMap* bit_map,
6395                        OopTaskQueue* work_queue,
6396                        CMSMarkStack*  overflow_stack):
6397   _collector(collector),
6398   _whole_span(collector->_span),
6399   _span(span),
6400   _bit_map(bit_map),
6401   _mut(&collector->_modUnionTable),
6402   _work_queue(work_queue),
6403   _overflow_stack(overflow_stack),
6404   _skip_bits(0),
6405   _task(task)
6406 {
6407   assert(_work_queue->size() == 0, "work_queue should be empty");
6408   _finger = span.start();
6409   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
6410   assert(_span.contains(_finger), "Out of bounds _finger?");
6411 }
6412 
6413 // Should revisit to see if this should be restructured for
6414 // greater efficiency.
6415 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6416   if (_skip_bits > 0) {
6417     _skip_bits--;
6418     return true;
6419   }
6420   // convert offset into a HeapWord*
6421   HeapWord* addr = _bit_map->startWord() + offset;
6422   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6423          "address out of range");
6424   assert(_bit_map->isMarked(addr), "tautology");
6425   if (_bit_map->isMarked(addr+1)) {
6426     // this is an allocated object that might not yet be initialized
6427     assert(_skip_bits == 0, "tautology");
6428     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
6429     oop p = oop(addr);
6430     if (p->klass_or_null() == NULL) {
6431       // in the case of Clean-on-Enter optimization, redirty card
6432       // and avoid clearing card by increasing  the threshold.
6433       return true;
6434     }
6435   }
6436   scan_oops_in_oop(addr);
6437   return true;
6438 }
6439 
6440 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6441   assert(_bit_map->isMarked(ptr), "expected bit to be set");
6442   // Should we assert that our work queue is empty or
6443   // below some drain limit?
6444   assert(_work_queue->size() == 0,
6445          "should drain stack to limit stack usage");
6446   // convert ptr to an oop preparatory to scanning
6447   oop obj = oop(ptr);
6448   // Ignore mark word in verification below, since we
6449   // may be running concurrent with mutators.
6450   assert(obj->is_oop(true), "should be an oop");
6451   assert(_finger <= ptr, "_finger runneth ahead");
6452   // advance the finger to right end of this object
6453   _finger = ptr + obj->size();
6454   assert(_finger > ptr, "we just incremented it above");
6455   // On large heaps, it may take us some time to get through
6456   // the marking phase. During
6457   // this time it's possible that a lot of mutations have
6458   // accumulated in the card table and the mod union table --
6459   // these mutation records are redundant until we have
6460   // actually traced into the corresponding card.
6461   // Here, we check whether advancing the finger would make
6462   // us cross into a new card, and if so clear corresponding
6463   // cards in the MUT (preclean them in the card-table in the
6464   // future).
6465 
6466   // The clean-on-enter optimization is disabled by default,
6467   // until we fix 6178663.
6468   if (CMSCleanOnEnter && (_finger > _threshold)) {
6469     // [_threshold, _finger) represents the interval
6470     // of cards to be cleared  in MUT (or precleaned in card table).
6471     // The set of cards to be cleared is all those that overlap
6472     // with the interval [_threshold, _finger); note that
6473     // _threshold is always kept card-aligned but _finger isn't
6474     // always card-aligned.
6475     HeapWord* old_threshold = _threshold;
6476     assert(old_threshold == (HeapWord*)round_to(
6477             (intptr_t)old_threshold, CardTableModRefBS::card_size),
6478            "_threshold should always be card-aligned");
6479     _threshold = (HeapWord*)round_to(
6480                    (intptr_t)_finger, CardTableModRefBS::card_size);
6481     MemRegion mr(old_threshold, _threshold);
6482     assert(!mr.is_empty(), "Control point invariant");
6483     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6484     _mut->clear_range(mr);
6485   }
6486 
6487   // Note: the local finger doesn't advance while we drain
6488   // the stack below, but the global finger sure can and will.
6489   HeapWord** gfa = _task->global_finger_addr();
6490   Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6491                                       _span, _bit_map,
6492                                       _work_queue,
6493                                       _overflow_stack,
6494                                       _finger,
6495                                       gfa, this);
6496   bool res = _work_queue->push(obj);   // overflow could occur here
6497   assert(res, "Will hold once we use workqueues");
6498   while (true) {
6499     oop new_oop;
6500     if (!_work_queue->pop_local(new_oop)) {
6501       // We emptied our work_queue; check if there's stuff that can
6502       // be gotten from the overflow stack.
6503       if (CMSConcMarkingTask::get_work_from_overflow_stack(
6504             _overflow_stack, _work_queue)) {
6505         do_yield_check();
6506         continue;
6507       } else {  // done
6508         break;
6509       }
6510     }
6511     // Skip verifying header mark word below because we are
6512     // running concurrent with mutators.
6513     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6514     // now scan this oop's oops
6515     new_oop->oop_iterate(&pushOrMarkClosure);
6516     do_yield_check();
6517   }
6518   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6519 }
6520 
6521 // Yield in response to a request from VM Thread or
6522 // from mutators.
6523 void Par_MarkFromRootsClosure::do_yield_work() {
6524   assert(_task != NULL, "sanity");
6525   _task->yield();
6526 }
6527 
6528 // A variant of the above used for verifying CMS marking work.
6529 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6530                         MemRegion span,
6531                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6532                         CMSMarkStack*  mark_stack):
6533   _collector(collector),
6534   _span(span),
6535   _verification_bm(verification_bm),
6536   _cms_bm(cms_bm),
6537   _mark_stack(mark_stack),
6538   _pam_verify_closure(collector, span, verification_bm, cms_bm,
6539                       mark_stack)
6540 {
6541   assert(_mark_stack->isEmpty(), "stack should be empty");
6542   _finger = _verification_bm->startWord();
6543   assert(_collector->_restart_addr == NULL, "Sanity check");
6544   assert(_span.contains(_finger), "Out of bounds _finger?");
6545 }
6546 
6547 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6548   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6549   assert(_span.contains(addr), "Out of bounds _finger?");
6550   _finger = addr;
6551 }
6552 
6553 // Should revisit to see if this should be restructured for
6554 // greater efficiency.
6555 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6556   // convert offset into a HeapWord*
6557   HeapWord* addr = _verification_bm->startWord() + offset;
6558   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6559          "address out of range");
6560   assert(_verification_bm->isMarked(addr), "tautology");
6561   assert(_cms_bm->isMarked(addr), "tautology");
6562 
6563   assert(_mark_stack->isEmpty(),
6564          "should drain stack to limit stack usage");
6565   // convert addr to an oop preparatory to scanning
6566   oop obj = oop(addr);
6567   assert(obj->is_oop(), "should be an oop");
6568   assert(_finger <= addr, "_finger runneth ahead");
6569   // advance the finger to right end of this object
6570   _finger = addr + obj->size();
6571   assert(_finger > addr, "we just incremented it above");
6572   // Note: the finger doesn't advance while we drain
6573   // the stack below.
6574   bool res = _mark_stack->push(obj);
6575   assert(res, "Empty non-zero size stack should have space for single push");
6576   while (!_mark_stack->isEmpty()) {
6577     oop new_oop = _mark_stack->pop();
6578     assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6579     // now scan this oop's oops
6580     new_oop->oop_iterate(&_pam_verify_closure);
6581   }
6582   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6583   return true;
6584 }
6585 
6586 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6587   CMSCollector* collector, MemRegion span,
6588   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6589   CMSMarkStack*  mark_stack):
6590   MetadataAwareOopClosure(collector->ref_processor()),
6591   _collector(collector),
6592   _span(span),
6593   _verification_bm(verification_bm),
6594   _cms_bm(cms_bm),
6595   _mark_stack(mark_stack)
6596 { }
6597 
6598 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
6599 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6600 
6601 // Upon stack overflow, we discard (part of) the stack,
6602 // remembering the least address amongst those discarded
6603 // in CMSCollector's _restart_address.
6604 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6605   // Remember the least grey address discarded
6606   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6607   _collector->lower_restart_addr(ra);
6608   _mark_stack->reset();  // discard stack contents
6609   _mark_stack->expand(); // expand the stack if possible
6610 }
6611 
6612 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6613   assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6614   HeapWord* addr = (HeapWord*)obj;
6615   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6616     // Oop lies in _span and isn't yet grey or black
6617     _verification_bm->mark(addr);            // now grey
6618     if (!_cms_bm->isMarked(addr)) {
6619       LogHandle(gc, verify) log;
6620       ResourceMark rm;
6621       oop(addr)->print_on(log.info_stream());
6622       log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6623       fatal("... aborting");
6624     }
6625 
6626     if (!_mark_stack->push(obj)) { // stack overflow
6627       log_debug(gc, stats)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6628       assert(_mark_stack->isFull(), "Else push should have succeeded");
6629       handle_stack_overflow(addr);
6630     }
6631     // anything including and to the right of _finger
6632     // will be scanned as we iterate over the remainder of the
6633     // bit map
6634   }
6635 }
6636 
6637 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6638                      MemRegion span,
6639                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
6640                      HeapWord* finger, MarkFromRootsClosure* parent) :
6641   MetadataAwareOopClosure(collector->ref_processor()),
6642   _collector(collector),
6643   _span(span),
6644   _bitMap(bitMap),
6645   _markStack(markStack),
6646   _finger(finger),
6647   _parent(parent)
6648 { }
6649 
6650 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6651                      MemRegion span,
6652                      CMSBitMap* bit_map,
6653                      OopTaskQueue* work_queue,
6654                      CMSMarkStack*  overflow_stack,
6655                      HeapWord* finger,
6656                      HeapWord** global_finger_addr,
6657                      Par_MarkFromRootsClosure* parent) :
6658   MetadataAwareOopClosure(collector->ref_processor()),
6659   _collector(collector),
6660   _whole_span(collector->_span),
6661   _span(span),
6662   _bit_map(bit_map),
6663   _work_queue(work_queue),
6664   _overflow_stack(overflow_stack),
6665   _finger(finger),
6666   _global_finger_addr(global_finger_addr),
6667   _parent(parent)
6668 { }
6669 
6670 // Assumes thread-safe access by callers, who are
6671 // responsible for mutual exclusion.
6672 void CMSCollector::lower_restart_addr(HeapWord* low) {
6673   assert(_span.contains(low), "Out of bounds addr");
6674   if (_restart_addr == NULL) {
6675     _restart_addr = low;
6676   } else {
6677     _restart_addr = MIN2(_restart_addr, low);
6678   }
6679 }
6680 
6681 // Upon stack overflow, we discard (part of) the stack,
6682 // remembering the least address amongst those discarded
6683 // in CMSCollector's _restart_address.
6684 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6685   // Remember the least grey address discarded
6686   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6687   _collector->lower_restart_addr(ra);
6688   _markStack->reset();  // discard stack contents
6689   _markStack->expand(); // expand the stack if possible
6690 }
6691 
6692 // Upon stack overflow, we discard (part of) the stack,
6693 // remembering the least address amongst those discarded
6694 // in CMSCollector's _restart_address.
6695 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6696   // We need to do this under a mutex to prevent other
6697   // workers from interfering with the work done below.
6698   MutexLockerEx ml(_overflow_stack->par_lock(),
6699                    Mutex::_no_safepoint_check_flag);
6700   // Remember the least grey address discarded
6701   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6702   _collector->lower_restart_addr(ra);
6703   _overflow_stack->reset();  // discard stack contents
6704   _overflow_stack->expand(); // expand the stack if possible
6705 }
6706 
6707 void PushOrMarkClosure::do_oop(oop obj) {
6708   // Ignore mark word because we are running concurrent with mutators.
6709   assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6710   HeapWord* addr = (HeapWord*)obj;
6711   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6712     // Oop lies in _span and isn't yet grey or black
6713     _bitMap->mark(addr);            // now grey
6714     if (addr < _finger) {
6715       // the bit map iteration has already either passed, or
6716       // sampled, this bit in the bit map; we'll need to
6717       // use the marking stack to scan this oop's oops.
6718       bool simulate_overflow = false;
6719       NOT_PRODUCT(
6720         if (CMSMarkStackOverflowALot &&
6721             _collector->simulate_overflow()) {
6722           // simulate a stack overflow
6723           simulate_overflow = true;
6724         }
6725       )
6726       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6727         log_debug(gc, stats)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6728         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6729         handle_stack_overflow(addr);
6730       }
6731     }
6732     // anything including and to the right of _finger
6733     // will be scanned as we iterate over the remainder of the
6734     // bit map
6735     do_yield_check();
6736   }
6737 }
6738 
6739 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
6740 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6741 
6742 void Par_PushOrMarkClosure::do_oop(oop obj) {
6743   // Ignore mark word because we are running concurrent with mutators.
6744   assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6745   HeapWord* addr = (HeapWord*)obj;
6746   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6747     // Oop lies in _span and isn't yet grey or black
6748     // We read the global_finger (volatile read) strictly after marking oop
6749     bool res = _bit_map->par_mark(addr);    // now grey
6750     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6751     // Should we push this marked oop on our stack?
6752     // -- if someone else marked it, nothing to do
6753     // -- if target oop is above global finger nothing to do
6754     // -- if target oop is in chunk and above local finger
6755     //      then nothing to do
6756     // -- else push on work queue
6757     if (   !res       // someone else marked it, they will deal with it
6758         || (addr >= *gfa)  // will be scanned in a later task
6759         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6760       return;
6761     }
6762     // the bit map iteration has already either passed, or
6763     // sampled, this bit in the bit map; we'll need to
6764     // use the marking stack to scan this oop's oops.
6765     bool simulate_overflow = false;
6766     NOT_PRODUCT(
6767       if (CMSMarkStackOverflowALot &&
6768           _collector->simulate_overflow()) {
6769         // simulate a stack overflow
6770         simulate_overflow = true;
6771       }
6772     )
6773     if (simulate_overflow ||
6774         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6775       // stack overflow
6776       log_debug(gc, stats)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6777       // We cannot assert that the overflow stack is full because
6778       // it may have been emptied since.
6779       assert(simulate_overflow ||
6780              _work_queue->size() == _work_queue->max_elems(),
6781             "Else push should have succeeded");
6782       handle_stack_overflow(addr);
6783     }
6784     do_yield_check();
6785   }
6786 }
6787 
6788 void Par_PushOrMarkClosure::do_oop(oop* p)       { Par_PushOrMarkClosure::do_oop_work(p); }
6789 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
6790 
6791 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6792                                        MemRegion span,
6793                                        ReferenceProcessor* rp,
6794                                        CMSBitMap* bit_map,
6795                                        CMSBitMap* mod_union_table,
6796                                        CMSMarkStack*  mark_stack,
6797                                        bool           concurrent_precleaning):
6798   MetadataAwareOopClosure(rp),
6799   _collector(collector),
6800   _span(span),
6801   _bit_map(bit_map),
6802   _mod_union_table(mod_union_table),
6803   _mark_stack(mark_stack),
6804   _concurrent_precleaning(concurrent_precleaning)
6805 {
6806   assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6807 }
6808 
6809 // Grey object rescan during pre-cleaning and second checkpoint phases --
6810 // the non-parallel version (the parallel version appears further below.)
6811 void PushAndMarkClosure::do_oop(oop obj) {
6812   // Ignore mark word verification. If during concurrent precleaning,
6813   // the object monitor may be locked. If during the checkpoint
6814   // phases, the object may already have been reached by a  different
6815   // path and may be at the end of the global overflow list (so
6816   // the mark word may be NULL).
6817   assert(obj->is_oop_or_null(true /* ignore mark word */),
6818          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6819   HeapWord* addr = (HeapWord*)obj;
6820   // Check if oop points into the CMS generation
6821   // and is not marked
6822   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6823     // a white object ...
6824     _bit_map->mark(addr);         // ... now grey
6825     // push on the marking stack (grey set)
6826     bool simulate_overflow = false;
6827     NOT_PRODUCT(
6828       if (CMSMarkStackOverflowALot &&
6829           _collector->simulate_overflow()) {
6830         // simulate a stack overflow
6831         simulate_overflow = true;
6832       }
6833     )
6834     if (simulate_overflow || !_mark_stack->push(obj)) {
6835       if (_concurrent_precleaning) {
6836          // During precleaning we can just dirty the appropriate card(s)
6837          // in the mod union table, thus ensuring that the object remains
6838          // in the grey set  and continue. In the case of object arrays
6839          // we need to dirty all of the cards that the object spans,
6840          // since the rescan of object arrays will be limited to the
6841          // dirty cards.
6842          // Note that no one can be interfering with us in this action
6843          // of dirtying the mod union table, so no locking or atomics
6844          // are required.
6845          if (obj->is_objArray()) {
6846            size_t sz = obj->size();
6847            HeapWord* end_card_addr = (HeapWord*)round_to(
6848                                         (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6849            MemRegion redirty_range = MemRegion(addr, end_card_addr);
6850            assert(!redirty_range.is_empty(), "Arithmetical tautology");
6851            _mod_union_table->mark_range(redirty_range);
6852          } else {
6853            _mod_union_table->mark(addr);
6854          }
6855          _collector->_ser_pmc_preclean_ovflw++;
6856       } else {
6857          // During the remark phase, we need to remember this oop
6858          // in the overflow list.
6859          _collector->push_on_overflow_list(obj);
6860          _collector->_ser_pmc_remark_ovflw++;
6861       }
6862     }
6863   }
6864 }
6865 
6866 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
6867                                                MemRegion span,
6868                                                ReferenceProcessor* rp,
6869                                                CMSBitMap* bit_map,
6870                                                OopTaskQueue* work_queue):
6871   MetadataAwareOopClosure(rp),
6872   _collector(collector),
6873   _span(span),
6874   _bit_map(bit_map),
6875   _work_queue(work_queue)
6876 {
6877   assert(ref_processor() != NULL, "ref_processor shouldn't be NULL");
6878 }
6879 
6880 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
6881 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6882 
6883 // Grey object rescan during second checkpoint phase --
6884 // the parallel version.
6885 void Par_PushAndMarkClosure::do_oop(oop obj) {
6886   // In the assert below, we ignore the mark word because
6887   // this oop may point to an already visited object that is
6888   // on the overflow stack (in which case the mark word has
6889   // been hijacked for chaining into the overflow stack --
6890   // if this is the last object in the overflow stack then
6891   // its mark word will be NULL). Because this object may
6892   // have been subsequently popped off the global overflow
6893   // stack, and the mark word possibly restored to the prototypical
6894   // value, by the time we get to examined this failing assert in
6895   // the debugger, is_oop_or_null(false) may subsequently start
6896   // to hold.
6897   assert(obj->is_oop_or_null(true),
6898          "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6899   HeapWord* addr = (HeapWord*)obj;
6900   // Check if oop points into the CMS generation
6901   // and is not marked
6902   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6903     // a white object ...
6904     // If we manage to "claim" the object, by being the
6905     // first thread to mark it, then we push it on our
6906     // marking stack
6907     if (_bit_map->par_mark(addr)) {     // ... now grey
6908       // push on work queue (grey set)
6909       bool simulate_overflow = false;
6910       NOT_PRODUCT(
6911         if (CMSMarkStackOverflowALot &&
6912             _collector->par_simulate_overflow()) {
6913           // simulate a stack overflow
6914           simulate_overflow = true;
6915         }
6916       )
6917       if (simulate_overflow || !_work_queue->push(obj)) {
6918         _collector->par_push_on_overflow_list(obj);
6919         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
6920       }
6921     } // Else, some other thread got there first
6922   }
6923 }
6924 
6925 void Par_PushAndMarkClosure::do_oop(oop* p)       { Par_PushAndMarkClosure::do_oop_work(p); }
6926 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
6927 
6928 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6929   Mutex* bml = _collector->bitMapLock();
6930   assert_lock_strong(bml);
6931   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6932          "CMS thread should hold CMS token");
6933 
6934   bml->unlock();
6935   ConcurrentMarkSweepThread::desynchronize(true);
6936 
6937   _collector->stopTimer();
6938   _collector->incrementYields();
6939 
6940   // See the comment in coordinator_yield()
6941   for (unsigned i = 0; i < CMSYieldSleepCount &&
6942                        ConcurrentMarkSweepThread::should_yield() &&
6943                        !CMSCollector::foregroundGCIsActive(); ++i) {
6944     os::sleep(Thread::current(), 1, false);
6945   }
6946 
6947   ConcurrentMarkSweepThread::synchronize(true);
6948   bml->lock();
6949 
6950   _collector->startTimer();
6951 }
6952 
6953 bool CMSPrecleanRefsYieldClosure::should_return() {
6954   if (ConcurrentMarkSweepThread::should_yield()) {
6955     do_yield_work();
6956   }
6957   return _collector->foregroundGCIsActive();
6958 }
6959 
6960 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
6961   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
6962          "mr should be aligned to start at a card boundary");
6963   // We'd like to assert:
6964   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
6965   //        "mr should be a range of cards");
6966   // However, that would be too strong in one case -- the last
6967   // partition ends at _unallocated_block which, in general, can be
6968   // an arbitrary boundary, not necessarily card aligned.
6969   _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words;
6970   _space->object_iterate_mem(mr, &_scan_cl);
6971 }
6972 
6973 SweepClosure::SweepClosure(CMSCollector* collector,
6974                            ConcurrentMarkSweepGeneration* g,
6975                            CMSBitMap* bitMap, bool should_yield) :
6976   _collector(collector),
6977   _g(g),
6978   _sp(g->cmsSpace()),
6979   _limit(_sp->sweep_limit()),
6980   _freelistLock(_sp->freelistLock()),
6981   _bitMap(bitMap),
6982   _yield(should_yield),
6983   _inFreeRange(false),           // No free range at beginning of sweep
6984   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
6985   _lastFreeRangeCoalesced(false),
6986   _freeFinger(g->used_region().start())
6987 {
6988   NOT_PRODUCT(
6989     _numObjectsFreed = 0;
6990     _numWordsFreed   = 0;
6991     _numObjectsLive = 0;
6992     _numWordsLive = 0;
6993     _numObjectsAlreadyFree = 0;
6994     _numWordsAlreadyFree = 0;
6995     _last_fc = NULL;
6996 
6997     _sp->initializeIndexedFreeListArrayReturnedBytes();
6998     _sp->dictionary()->initialize_dict_returned_bytes();
6999   )
7000   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7001          "sweep _limit out of bounds");
7002   log_develop(gc, sweep)("====================");
7003   log_develop(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7004 }
7005 
7006 void SweepClosure::print_on(outputStream* st) const {
7007   tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7008                 p2i(_sp->bottom()), p2i(_sp->end()));
7009   tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7010   tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7011   NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7012   tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7013                 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7014 }
7015 
7016 #ifndef PRODUCT
7017 // Assertion checking only:  no useful work in product mode --
7018 // however, if any of the flags below become product flags,
7019 // you may need to review this code to see if it needs to be
7020 // enabled in product mode.
7021 SweepClosure::~SweepClosure() {
7022   assert_lock_strong(_freelistLock);
7023   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7024          "sweep _limit out of bounds");
7025   if (inFreeRange()) {
7026     warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7027     print();
7028     ShouldNotReachHere();
7029   }
7030 
7031   if (log_is_enabled(Debug, gc, sweep)) {
7032     log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7033                          _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7034     log_debug(gc, sweep)("Live " SIZE_FORMAT " objects,  " SIZE_FORMAT " bytes  Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7035                          _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7036     size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7037     log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7038   }
7039 
7040   if (log_is_enabled(Debug, gc, stats) && CMSVerifyReturnedBytes) {
7041     size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7042     size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7043     size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7044     log_debug(gc, stats)("Returned " SIZE_FORMAT " bytes   Indexed List Returned " SIZE_FORMAT " bytes        Dictionary Returned " SIZE_FORMAT " bytes",
7045                                returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7046   }
7047   log_develop(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7048   log_develop(gc, sweep)("================");
7049 }
7050 #endif  // PRODUCT
7051 
7052 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7053     bool freeRangeInFreeLists) {
7054   log_develop(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7055                          p2i(freeFinger), freeRangeInFreeLists);
7056   assert(!inFreeRange(), "Trampling existing free range");
7057   set_inFreeRange(true);
7058   set_lastFreeRangeCoalesced(false);
7059 
7060   set_freeFinger(freeFinger);
7061   set_freeRangeInFreeLists(freeRangeInFreeLists);
7062   if (CMSTestInFreeList) {
7063     if (freeRangeInFreeLists) {
7064       FreeChunk* fc = (FreeChunk*) freeFinger;
7065       assert(fc->is_free(), "A chunk on the free list should be free.");
7066       assert(fc->size() > 0, "Free range should have a size");
7067       assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7068     }
7069   }
7070 }
7071 
7072 // Note that the sweeper runs concurrently with mutators. Thus,
7073 // it is possible for direct allocation in this generation to happen
7074 // in the middle of the sweep. Note that the sweeper also coalesces
7075 // contiguous free blocks. Thus, unless the sweeper and the allocator
7076 // synchronize appropriately freshly allocated blocks may get swept up.
7077 // This is accomplished by the sweeper locking the free lists while
7078 // it is sweeping. Thus blocks that are determined to be free are
7079 // indeed free. There is however one additional complication:
7080 // blocks that have been allocated since the final checkpoint and
7081 // mark, will not have been marked and so would be treated as
7082 // unreachable and swept up. To prevent this, the allocator marks
7083 // the bit map when allocating during the sweep phase. This leads,
7084 // however, to a further complication -- objects may have been allocated
7085 // but not yet initialized -- in the sense that the header isn't yet
7086 // installed. The sweeper can not then determine the size of the block
7087 // in order to skip over it. To deal with this case, we use a technique
7088 // (due to Printezis) to encode such uninitialized block sizes in the
7089 // bit map. Since the bit map uses a bit per every HeapWord, but the
7090 // CMS generation has a minimum object size of 3 HeapWords, it follows
7091 // that "normal marks" won't be adjacent in the bit map (there will
7092 // always be at least two 0 bits between successive 1 bits). We make use
7093 // of these "unused" bits to represent uninitialized blocks -- the bit
7094 // corresponding to the start of the uninitialized object and the next
7095 // bit are both set. Finally, a 1 bit marks the end of the object that
7096 // started with the two consecutive 1 bits to indicate its potentially
7097 // uninitialized state.
7098 
7099 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7100   FreeChunk* fc = (FreeChunk*)addr;
7101   size_t res;
7102 
7103   // Check if we are done sweeping. Below we check "addr >= _limit" rather
7104   // than "addr == _limit" because although _limit was a block boundary when
7105   // we started the sweep, it may no longer be one because heap expansion
7106   // may have caused us to coalesce the block ending at the address _limit
7107   // with a newly expanded chunk (this happens when _limit was set to the
7108   // previous _end of the space), so we may have stepped past _limit:
7109   // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7110   if (addr >= _limit) { // we have swept up to or past the limit: finish up
7111     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7112            "sweep _limit out of bounds");
7113     assert(addr < _sp->end(), "addr out of bounds");
7114     // Flush any free range we might be holding as a single
7115     // coalesced chunk to the appropriate free list.
7116     if (inFreeRange()) {
7117       assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7118              "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7119       flush_cur_free_chunk(freeFinger(),
7120                            pointer_delta(addr, freeFinger()));
7121       log_develop(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7122                              p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7123                              lastFreeRangeCoalesced() ? 1 : 0);
7124     }
7125 
7126     // help the iterator loop finish
7127     return pointer_delta(_sp->end(), addr);
7128   }
7129 
7130   assert(addr < _limit, "sweep invariant");
7131   // check if we should yield
7132   do_yield_check(addr);
7133   if (fc->is_free()) {
7134     // Chunk that is already free
7135     res = fc->size();
7136     do_already_free_chunk(fc);
7137     debug_only(_sp->verifyFreeLists());
7138     // If we flush the chunk at hand in lookahead_and_flush()
7139     // and it's coalesced with a preceding chunk, then the
7140     // process of "mangling" the payload of the coalesced block
7141     // will cause erasure of the size information from the
7142     // (erstwhile) header of all the coalesced blocks but the
7143     // first, so the first disjunct in the assert will not hold
7144     // in that specific case (in which case the second disjunct
7145     // will hold).
7146     assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7147            "Otherwise the size info doesn't change at this step");
7148     NOT_PRODUCT(
7149       _numObjectsAlreadyFree++;
7150       _numWordsAlreadyFree += res;
7151     )
7152     NOT_PRODUCT(_last_fc = fc;)
7153   } else if (!_bitMap->isMarked(addr)) {
7154     // Chunk is fresh garbage
7155     res = do_garbage_chunk(fc);
7156     debug_only(_sp->verifyFreeLists());
7157     NOT_PRODUCT(
7158       _numObjectsFreed++;
7159       _numWordsFreed += res;
7160     )
7161   } else {
7162     // Chunk that is alive.
7163     res = do_live_chunk(fc);
7164     debug_only(_sp->verifyFreeLists());
7165     NOT_PRODUCT(
7166         _numObjectsLive++;
7167         _numWordsLive += res;
7168     )
7169   }
7170   return res;
7171 }
7172 
7173 // For the smart allocation, record following
7174 //  split deaths - a free chunk is removed from its free list because
7175 //      it is being split into two or more chunks.
7176 //  split birth - a free chunk is being added to its free list because
7177 //      a larger free chunk has been split and resulted in this free chunk.
7178 //  coal death - a free chunk is being removed from its free list because
7179 //      it is being coalesced into a large free chunk.
7180 //  coal birth - a free chunk is being added to its free list because
7181 //      it was created when two or more free chunks where coalesced into
7182 //      this free chunk.
7183 //
7184 // These statistics are used to determine the desired number of free
7185 // chunks of a given size.  The desired number is chosen to be relative
7186 // to the end of a CMS sweep.  The desired number at the end of a sweep
7187 // is the
7188 //      count-at-end-of-previous-sweep (an amount that was enough)
7189 //              - count-at-beginning-of-current-sweep  (the excess)
7190 //              + split-births  (gains in this size during interval)
7191 //              - split-deaths  (demands on this size during interval)
7192 // where the interval is from the end of one sweep to the end of the
7193 // next.
7194 //
7195 // When sweeping the sweeper maintains an accumulated chunk which is
7196 // the chunk that is made up of chunks that have been coalesced.  That
7197 // will be termed the left-hand chunk.  A new chunk of garbage that
7198 // is being considered for coalescing will be referred to as the
7199 // right-hand chunk.
7200 //
7201 // When making a decision on whether to coalesce a right-hand chunk with
7202 // the current left-hand chunk, the current count vs. the desired count
7203 // of the left-hand chunk is considered.  Also if the right-hand chunk
7204 // is near the large chunk at the end of the heap (see
7205 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7206 // left-hand chunk is coalesced.
7207 //
7208 // When making a decision about whether to split a chunk, the desired count
7209 // vs. the current count of the candidate to be split is also considered.
7210 // If the candidate is underpopulated (currently fewer chunks than desired)
7211 // a chunk of an overpopulated (currently more chunks than desired) size may
7212 // be chosen.  The "hint" associated with a free list, if non-null, points
7213 // to a free list which may be overpopulated.
7214 //
7215 
7216 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7217   const size_t size = fc->size();
7218   // Chunks that cannot be coalesced are not in the
7219   // free lists.
7220   if (CMSTestInFreeList && !fc->cantCoalesce()) {
7221     assert(_sp->verify_chunk_in_free_list(fc),
7222            "free chunk should be in free lists");
7223   }
7224   // a chunk that is already free, should not have been
7225   // marked in the bit map
7226   HeapWord* const addr = (HeapWord*) fc;
7227   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7228   // Verify that the bit map has no bits marked between
7229   // addr and purported end of this block.
7230   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7231 
7232   // Some chunks cannot be coalesced under any circumstances.
7233   // See the definition of cantCoalesce().
7234   if (!fc->cantCoalesce()) {
7235     // This chunk can potentially be coalesced.
7236     // All the work is done in
7237     do_post_free_or_garbage_chunk(fc, size);
7238     // Note that if the chunk is not coalescable (the else arm
7239     // below), we unconditionally flush, without needing to do
7240     // a "lookahead," as we do below.
7241     if (inFreeRange()) lookahead_and_flush(fc, size);
7242   } else {
7243     // Code path common to both original and adaptive free lists.
7244 
7245     // cant coalesce with previous block; this should be treated
7246     // as the end of a free run if any
7247     if (inFreeRange()) {
7248       // we kicked some butt; time to pick up the garbage
7249       assert(freeFinger() < addr, "freeFinger points too high");
7250       flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7251     }
7252     // else, nothing to do, just continue
7253   }
7254 }
7255 
7256 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7257   // This is a chunk of garbage.  It is not in any free list.
7258   // Add it to a free list or let it possibly be coalesced into
7259   // a larger chunk.
7260   HeapWord* const addr = (HeapWord*) fc;
7261   const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7262 
7263   // Verify that the bit map has no bits marked between
7264   // addr and purported end of just dead object.
7265   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7266   do_post_free_or_garbage_chunk(fc, size);
7267 
7268   assert(_limit >= addr + size,
7269          "A freshly garbage chunk can't possibly straddle over _limit");
7270   if (inFreeRange()) lookahead_and_flush(fc, size);
7271   return size;
7272 }
7273 
7274 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7275   HeapWord* addr = (HeapWord*) fc;
7276   // The sweeper has just found a live object. Return any accumulated
7277   // left hand chunk to the free lists.
7278   if (inFreeRange()) {
7279     assert(freeFinger() < addr, "freeFinger points too high");
7280     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7281   }
7282 
7283   // This object is live: we'd normally expect this to be
7284   // an oop, and like to assert the following:
7285   // assert(oop(addr)->is_oop(), "live block should be an oop");
7286   // However, as we commented above, this may be an object whose
7287   // header hasn't yet been initialized.
7288   size_t size;
7289   assert(_bitMap->isMarked(addr), "Tautology for this control point");
7290   if (_bitMap->isMarked(addr + 1)) {
7291     // Determine the size from the bit map, rather than trying to
7292     // compute it from the object header.
7293     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7294     size = pointer_delta(nextOneAddr + 1, addr);
7295     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7296            "alignment problem");
7297 
7298 #ifdef ASSERT
7299       if (oop(addr)->klass_or_null() != NULL) {
7300         // Ignore mark word because we are running concurrent with mutators
7301         assert(oop(addr)->is_oop(true), "live block should be an oop");
7302         assert(size ==
7303                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7304                "P-mark and computed size do not agree");
7305       }
7306 #endif
7307 
7308   } else {
7309     // This should be an initialized object that's alive.
7310     assert(oop(addr)->klass_or_null() != NULL,
7311            "Should be an initialized object");
7312     // Ignore mark word because we are running concurrent with mutators
7313     assert(oop(addr)->is_oop(true), "live block should be an oop");
7314     // Verify that the bit map has no bits marked between
7315     // addr and purported end of this block.
7316     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7317     assert(size >= 3, "Necessary for Printezis marks to work");
7318     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7319     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7320   }
7321   return size;
7322 }
7323 
7324 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7325                                                  size_t chunkSize) {
7326   // do_post_free_or_garbage_chunk() should only be called in the case
7327   // of the adaptive free list allocator.
7328   const bool fcInFreeLists = fc->is_free();
7329   assert((HeapWord*)fc <= _limit, "sweep invariant");
7330   if (CMSTestInFreeList && fcInFreeLists) {
7331     assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7332   }
7333 
7334   log_develop(gc, sweep)("  -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7335 
7336   HeapWord* const fc_addr = (HeapWord*) fc;
7337 
7338   bool coalesce = false;
7339   const size_t left  = pointer_delta(fc_addr, freeFinger());
7340   const size_t right = chunkSize;
7341   switch (FLSCoalescePolicy) {
7342     // numeric value forms a coalition aggressiveness metric
7343     case 0:  { // never coalesce
7344       coalesce = false;
7345       break;
7346     }
7347     case 1: { // coalesce if left & right chunks on overpopulated lists
7348       coalesce = _sp->coalOverPopulated(left) &&
7349                  _sp->coalOverPopulated(right);
7350       break;
7351     }
7352     case 2: { // coalesce if left chunk on overpopulated list (default)
7353       coalesce = _sp->coalOverPopulated(left);
7354       break;
7355     }
7356     case 3: { // coalesce if left OR right chunk on overpopulated list
7357       coalesce = _sp->coalOverPopulated(left) ||
7358                  _sp->coalOverPopulated(right);
7359       break;
7360     }
7361     case 4: { // always coalesce
7362       coalesce = true;
7363       break;
7364     }
7365     default:
7366      ShouldNotReachHere();
7367   }
7368 
7369   // Should the current free range be coalesced?
7370   // If the chunk is in a free range and either we decided to coalesce above
7371   // or the chunk is near the large block at the end of the heap
7372   // (isNearLargestChunk() returns true), then coalesce this chunk.
7373   const bool doCoalesce = inFreeRange()
7374                           && (coalesce || _g->isNearLargestChunk(fc_addr));
7375   if (doCoalesce) {
7376     // Coalesce the current free range on the left with the new
7377     // chunk on the right.  If either is on a free list,
7378     // it must be removed from the list and stashed in the closure.
7379     if (freeRangeInFreeLists()) {
7380       FreeChunk* const ffc = (FreeChunk*)freeFinger();
7381       assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7382              "Size of free range is inconsistent with chunk size.");
7383       if (CMSTestInFreeList) {
7384         assert(_sp->verify_chunk_in_free_list(ffc),
7385                "Chunk is not in free lists");
7386       }
7387       _sp->coalDeath(ffc->size());
7388       _sp->removeFreeChunkFromFreeLists(ffc);
7389       set_freeRangeInFreeLists(false);
7390     }
7391     if (fcInFreeLists) {
7392       _sp->coalDeath(chunkSize);
7393       assert(fc->size() == chunkSize,
7394         "The chunk has the wrong size or is not in the free lists");
7395       _sp->removeFreeChunkFromFreeLists(fc);
7396     }
7397     set_lastFreeRangeCoalesced(true);
7398     print_free_block_coalesced(fc);
7399   } else {  // not in a free range and/or should not coalesce
7400     // Return the current free range and start a new one.
7401     if (inFreeRange()) {
7402       // In a free range but cannot coalesce with the right hand chunk.
7403       // Put the current free range into the free lists.
7404       flush_cur_free_chunk(freeFinger(),
7405                            pointer_delta(fc_addr, freeFinger()));
7406     }
7407     // Set up for new free range.  Pass along whether the right hand
7408     // chunk is in the free lists.
7409     initialize_free_range((HeapWord*)fc, fcInFreeLists);
7410   }
7411 }
7412 
7413 // Lookahead flush:
7414 // If we are tracking a free range, and this is the last chunk that
7415 // we'll look at because its end crosses past _limit, we'll preemptively
7416 // flush it along with any free range we may be holding on to. Note that
7417 // this can be the case only for an already free or freshly garbage
7418 // chunk. If this block is an object, it can never straddle
7419 // over _limit. The "straddling" occurs when _limit is set at
7420 // the previous end of the space when this cycle started, and
7421 // a subsequent heap expansion caused the previously co-terminal
7422 // free block to be coalesced with the newly expanded portion,
7423 // thus rendering _limit a non-block-boundary making it dangerous
7424 // for the sweeper to step over and examine.
7425 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7426   assert(inFreeRange(), "Should only be called if currently in a free range.");
7427   HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7428   assert(_sp->used_region().contains(eob - 1),
7429          "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7430          " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7431          " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7432          p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7433   if (eob >= _limit) {
7434     assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7435     log_develop(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7436                            "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7437                            "[" PTR_FORMAT "," PTR_FORMAT ")",
7438                            p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7439     // Return the storage we are tracking back into the free lists.
7440     log_develop(gc, sweep)("Flushing ... ");
7441     assert(freeFinger() < eob, "Error");
7442     flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7443   }
7444 }
7445 
7446 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7447   assert(inFreeRange(), "Should only be called if currently in a free range.");
7448   assert(size > 0,
7449     "A zero sized chunk cannot be added to the free lists.");
7450   if (!freeRangeInFreeLists()) {
7451     if (CMSTestInFreeList) {
7452       FreeChunk* fc = (FreeChunk*) chunk;
7453       fc->set_size(size);
7454       assert(!_sp->verify_chunk_in_free_list(fc),
7455              "chunk should not be in free lists yet");
7456     }
7457     log_develop(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7458     // A new free range is going to be starting.  The current
7459     // free range has not been added to the free lists yet or
7460     // was removed so add it back.
7461     // If the current free range was coalesced, then the death
7462     // of the free range was recorded.  Record a birth now.
7463     if (lastFreeRangeCoalesced()) {
7464       _sp->coalBirth(size);
7465     }
7466     _sp->addChunkAndRepairOffsetTable(chunk, size,
7467             lastFreeRangeCoalesced());
7468   } else {
7469     log_develop(gc, sweep)("Already in free list: nothing to flush");
7470   }
7471   set_inFreeRange(false);
7472   set_freeRangeInFreeLists(false);
7473 }
7474 
7475 // We take a break if we've been at this for a while,
7476 // so as to avoid monopolizing the locks involved.
7477 void SweepClosure::do_yield_work(HeapWord* addr) {
7478   // Return current free chunk being used for coalescing (if any)
7479   // to the appropriate freelist.  After yielding, the next
7480   // free block encountered will start a coalescing range of
7481   // free blocks.  If the next free block is adjacent to the
7482   // chunk just flushed, they will need to wait for the next
7483   // sweep to be coalesced.
7484   if (inFreeRange()) {
7485     flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7486   }
7487 
7488   // First give up the locks, then yield, then re-lock.
7489   // We should probably use a constructor/destructor idiom to
7490   // do this unlock/lock or modify the MutexUnlocker class to
7491   // serve our purpose. XXX
7492   assert_lock_strong(_bitMap->lock());
7493   assert_lock_strong(_freelistLock);
7494   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7495          "CMS thread should hold CMS token");
7496   _bitMap->lock()->unlock();
7497   _freelistLock->unlock();
7498   ConcurrentMarkSweepThread::desynchronize(true);
7499   _collector->stopTimer();
7500   _collector->incrementYields();
7501 
7502   // See the comment in coordinator_yield()
7503   for (unsigned i = 0; i < CMSYieldSleepCount &&
7504                        ConcurrentMarkSweepThread::should_yield() &&
7505                        !CMSCollector::foregroundGCIsActive(); ++i) {
7506     os::sleep(Thread::current(), 1, false);
7507   }
7508 
7509   ConcurrentMarkSweepThread::synchronize(true);
7510   _freelistLock->lock();
7511   _bitMap->lock()->lock_without_safepoint_check();
7512   _collector->startTimer();
7513 }
7514 
7515 #ifndef PRODUCT
7516 // This is actually very useful in a product build if it can
7517 // be called from the debugger.  Compile it into the product
7518 // as needed.
7519 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7520   return debug_cms_space->verify_chunk_in_free_list(fc);
7521 }
7522 #endif
7523 
7524 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7525   log_develop(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7526                          p2i(fc), fc->size());
7527 }
7528 
7529 // CMSIsAliveClosure
7530 bool CMSIsAliveClosure::do_object_b(oop obj) {
7531   HeapWord* addr = (HeapWord*)obj;
7532   return addr != NULL &&
7533          (!_span.contains(addr) || _bit_map->isMarked(addr));
7534 }
7535 
7536 
7537 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7538                       MemRegion span,
7539                       CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7540                       bool cpc):
7541   _collector(collector),
7542   _span(span),
7543   _bit_map(bit_map),
7544   _mark_stack(mark_stack),
7545   _concurrent_precleaning(cpc) {
7546   assert(!_span.is_empty(), "Empty span could spell trouble");
7547 }
7548 
7549 
7550 // CMSKeepAliveClosure: the serial version
7551 void CMSKeepAliveClosure::do_oop(oop obj) {
7552   HeapWord* addr = (HeapWord*)obj;
7553   if (_span.contains(addr) &&
7554       !_bit_map->isMarked(addr)) {
7555     _bit_map->mark(addr);
7556     bool simulate_overflow = false;
7557     NOT_PRODUCT(
7558       if (CMSMarkStackOverflowALot &&
7559           _collector->simulate_overflow()) {
7560         // simulate a stack overflow
7561         simulate_overflow = true;
7562       }
7563     )
7564     if (simulate_overflow || !_mark_stack->push(obj)) {
7565       if (_concurrent_precleaning) {
7566         // We dirty the overflown object and let the remark
7567         // phase deal with it.
7568         assert(_collector->overflow_list_is_empty(), "Error");
7569         // In the case of object arrays, we need to dirty all of
7570         // the cards that the object spans. No locking or atomics
7571         // are needed since no one else can be mutating the mod union
7572         // table.
7573         if (obj->is_objArray()) {
7574           size_t sz = obj->size();
7575           HeapWord* end_card_addr =
7576             (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7577           MemRegion redirty_range = MemRegion(addr, end_card_addr);
7578           assert(!redirty_range.is_empty(), "Arithmetical tautology");
7579           _collector->_modUnionTable.mark_range(redirty_range);
7580         } else {
7581           _collector->_modUnionTable.mark(addr);
7582         }
7583         _collector->_ser_kac_preclean_ovflw++;
7584       } else {
7585         _collector->push_on_overflow_list(obj);
7586         _collector->_ser_kac_ovflw++;
7587       }
7588     }
7589   }
7590 }
7591 
7592 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
7593 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7594 
7595 // CMSParKeepAliveClosure: a parallel version of the above.
7596 // The work queues are private to each closure (thread),
7597 // but (may be) available for stealing by other threads.
7598 void CMSParKeepAliveClosure::do_oop(oop obj) {
7599   HeapWord* addr = (HeapWord*)obj;
7600   if (_span.contains(addr) &&
7601       !_bit_map->isMarked(addr)) {
7602     // In general, during recursive tracing, several threads
7603     // may be concurrently getting here; the first one to
7604     // "tag" it, claims it.
7605     if (_bit_map->par_mark(addr)) {
7606       bool res = _work_queue->push(obj);
7607       assert(res, "Low water mark should be much less than capacity");
7608       // Do a recursive trim in the hope that this will keep
7609       // stack usage lower, but leave some oops for potential stealers
7610       trim_queue(_low_water_mark);
7611     } // Else, another thread got there first
7612   }
7613 }
7614 
7615 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
7616 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7617 
7618 void CMSParKeepAliveClosure::trim_queue(uint max) {
7619   while (_work_queue->size() > max) {
7620     oop new_oop;
7621     if (_work_queue->pop_local(new_oop)) {
7622       assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7623       assert(_bit_map->isMarked((HeapWord*)new_oop),
7624              "no white objects on this stack!");
7625       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7626       // iterate over the oops in this oop, marking and pushing
7627       // the ones in CMS heap (i.e. in _span).
7628       new_oop->oop_iterate(&_mark_and_push);
7629     }
7630   }
7631 }
7632 
7633 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7634                                 CMSCollector* collector,
7635                                 MemRegion span, CMSBitMap* bit_map,
7636                                 OopTaskQueue* work_queue):
7637   _collector(collector),
7638   _span(span),
7639   _bit_map(bit_map),
7640   _work_queue(work_queue) { }
7641 
7642 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7643   HeapWord* addr = (HeapWord*)obj;
7644   if (_span.contains(addr) &&
7645       !_bit_map->isMarked(addr)) {
7646     if (_bit_map->par_mark(addr)) {
7647       bool simulate_overflow = false;
7648       NOT_PRODUCT(
7649         if (CMSMarkStackOverflowALot &&
7650             _collector->par_simulate_overflow()) {
7651           // simulate a stack overflow
7652           simulate_overflow = true;
7653         }
7654       )
7655       if (simulate_overflow || !_work_queue->push(obj)) {
7656         _collector->par_push_on_overflow_list(obj);
7657         _collector->_par_kac_ovflw++;
7658       }
7659     } // Else another thread got there already
7660   }
7661 }
7662 
7663 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7664 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7665 
7666 //////////////////////////////////////////////////////////////////
7667 //  CMSExpansionCause                /////////////////////////////
7668 //////////////////////////////////////////////////////////////////
7669 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7670   switch (cause) {
7671     case _no_expansion:
7672       return "No expansion";
7673     case _satisfy_free_ratio:
7674       return "Free ratio";
7675     case _satisfy_promotion:
7676       return "Satisfy promotion";
7677     case _satisfy_allocation:
7678       return "allocation";
7679     case _allocate_par_lab:
7680       return "Par LAB";
7681     case _allocate_par_spooling_space:
7682       return "Par Spooling Space";
7683     case _adaptive_size_policy:
7684       return "Ergonomics";
7685     default:
7686       return "unknown";
7687   }
7688 }
7689 
7690 void CMSDrainMarkingStackClosure::do_void() {
7691   // the max number to take from overflow list at a time
7692   const size_t num = _mark_stack->capacity()/4;
7693   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7694          "Overflow list should be NULL during concurrent phases");
7695   while (!_mark_stack->isEmpty() ||
7696          // if stack is empty, check the overflow list
7697          _collector->take_from_overflow_list(num, _mark_stack)) {
7698     oop obj = _mark_stack->pop();
7699     HeapWord* addr = (HeapWord*)obj;
7700     assert(_span.contains(addr), "Should be within span");
7701     assert(_bit_map->isMarked(addr), "Should be marked");
7702     assert(obj->is_oop(), "Should be an oop");
7703     obj->oop_iterate(_keep_alive);
7704   }
7705 }
7706 
7707 void CMSParDrainMarkingStackClosure::do_void() {
7708   // drain queue
7709   trim_queue(0);
7710 }
7711 
7712 // Trim our work_queue so its length is below max at return
7713 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7714   while (_work_queue->size() > max) {
7715     oop new_oop;
7716     if (_work_queue->pop_local(new_oop)) {
7717       assert(new_oop->is_oop(), "Expected an oop");
7718       assert(_bit_map->isMarked((HeapWord*)new_oop),
7719              "no white objects on this stack!");
7720       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7721       // iterate over the oops in this oop, marking and pushing
7722       // the ones in CMS heap (i.e. in _span).
7723       new_oop->oop_iterate(&_mark_and_push);
7724     }
7725   }
7726 }
7727 
7728 ////////////////////////////////////////////////////////////////////
7729 // Support for Marking Stack Overflow list handling and related code
7730 ////////////////////////////////////////////////////////////////////
7731 // Much of the following code is similar in shape and spirit to the
7732 // code used in ParNewGC. We should try and share that code
7733 // as much as possible in the future.
7734 
7735 #ifndef PRODUCT
7736 // Debugging support for CMSStackOverflowALot
7737 
7738 // It's OK to call this multi-threaded;  the worst thing
7739 // that can happen is that we'll get a bunch of closely
7740 // spaced simulated overflows, but that's OK, in fact
7741 // probably good as it would exercise the overflow code
7742 // under contention.
7743 bool CMSCollector::simulate_overflow() {
7744   if (_overflow_counter-- <= 0) { // just being defensive
7745     _overflow_counter = CMSMarkStackOverflowInterval;
7746     return true;
7747   } else {
7748     return false;
7749   }
7750 }
7751 
7752 bool CMSCollector::par_simulate_overflow() {
7753   return simulate_overflow();
7754 }
7755 #endif
7756 
7757 // Single-threaded
7758 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7759   assert(stack->isEmpty(), "Expected precondition");
7760   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7761   size_t i = num;
7762   oop  cur = _overflow_list;
7763   const markOop proto = markOopDesc::prototype();
7764   NOT_PRODUCT(ssize_t n = 0;)
7765   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7766     next = oop(cur->mark());
7767     cur->set_mark(proto);   // until proven otherwise
7768     assert(cur->is_oop(), "Should be an oop");
7769     bool res = stack->push(cur);
7770     assert(res, "Bit off more than can chew?");
7771     NOT_PRODUCT(n++;)
7772   }
7773   _overflow_list = cur;
7774 #ifndef PRODUCT
7775   assert(_num_par_pushes >= n, "Too many pops?");
7776   _num_par_pushes -=n;
7777 #endif
7778   return !stack->isEmpty();
7779 }
7780 
7781 #define BUSY  (cast_to_oop<intptr_t>(0x1aff1aff))
7782 // (MT-safe) Get a prefix of at most "num" from the list.
7783 // The overflow list is chained through the mark word of
7784 // each object in the list. We fetch the entire list,
7785 // break off a prefix of the right size and return the
7786 // remainder. If other threads try to take objects from
7787 // the overflow list at that time, they will wait for
7788 // some time to see if data becomes available. If (and
7789 // only if) another thread places one or more object(s)
7790 // on the global list before we have returned the suffix
7791 // to the global list, we will walk down our local list
7792 // to find its end and append the global list to
7793 // our suffix before returning it. This suffix walk can
7794 // prove to be expensive (quadratic in the amount of traffic)
7795 // when there are many objects in the overflow list and
7796 // there is much producer-consumer contention on the list.
7797 // *NOTE*: The overflow list manipulation code here and
7798 // in ParNewGeneration:: are very similar in shape,
7799 // except that in the ParNew case we use the old (from/eden)
7800 // copy of the object to thread the list via its klass word.
7801 // Because of the common code, if you make any changes in
7802 // the code below, please check the ParNew version to see if
7803 // similar changes might be needed.
7804 // CR 6797058 has been filed to consolidate the common code.
7805 bool CMSCollector::par_take_from_overflow_list(size_t num,
7806                                                OopTaskQueue* work_q,
7807                                                int no_of_gc_threads) {
7808   assert(work_q->size() == 0, "First empty local work queue");
7809   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7810   if (_overflow_list == NULL) {
7811     return false;
7812   }
7813   // Grab the entire list; we'll put back a suffix
7814   oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
7815   Thread* tid = Thread::current();
7816   // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7817   // set to ParallelGCThreads.
7818   size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7819   size_t sleep_time_millis = MAX2((size_t)1, num/100);
7820   // If the list is busy, we spin for a short while,
7821   // sleeping between attempts to get the list.
7822   for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7823     os::sleep(tid, sleep_time_millis, false);
7824     if (_overflow_list == NULL) {
7825       // Nothing left to take
7826       return false;
7827     } else if (_overflow_list != BUSY) {
7828       // Try and grab the prefix
7829       prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
7830     }
7831   }
7832   // If the list was found to be empty, or we spun long
7833   // enough, we give up and return empty-handed. If we leave
7834   // the list in the BUSY state below, it must be the case that
7835   // some other thread holds the overflow list and will set it
7836   // to a non-BUSY state in the future.
7837   if (prefix == NULL || prefix == BUSY) {
7838      // Nothing to take or waited long enough
7839      if (prefix == NULL) {
7840        // Write back the NULL in case we overwrote it with BUSY above
7841        // and it is still the same value.
7842        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
7843      }
7844      return false;
7845   }
7846   assert(prefix != NULL && prefix != BUSY, "Error");
7847   size_t i = num;
7848   oop cur = prefix;
7849   // Walk down the first "num" objects, unless we reach the end.
7850   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
7851   if (cur->mark() == NULL) {
7852     // We have "num" or fewer elements in the list, so there
7853     // is nothing to return to the global list.
7854     // Write back the NULL in lieu of the BUSY we wrote
7855     // above, if it is still the same value.
7856     if (_overflow_list == BUSY) {
7857       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
7858     }
7859   } else {
7860     // Chop off the suffix and return it to the global list.
7861     assert(cur->mark() != BUSY, "Error");
7862     oop suffix_head = cur->mark(); // suffix will be put back on global list
7863     cur->set_mark(NULL);           // break off suffix
7864     // It's possible that the list is still in the empty(busy) state
7865     // we left it in a short while ago; in that case we may be
7866     // able to place back the suffix without incurring the cost
7867     // of a walk down the list.
7868     oop observed_overflow_list = _overflow_list;
7869     oop cur_overflow_list = observed_overflow_list;
7870     bool attached = false;
7871     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7872       observed_overflow_list =
7873         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
7874       if (cur_overflow_list == observed_overflow_list) {
7875         attached = true;
7876         break;
7877       } else cur_overflow_list = observed_overflow_list;
7878     }
7879     if (!attached) {
7880       // Too bad, someone else sneaked in (at least) an element; we'll need
7881       // to do a splice. Find tail of suffix so we can prepend suffix to global
7882       // list.
7883       for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
7884       oop suffix_tail = cur;
7885       assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
7886              "Tautology");
7887       observed_overflow_list = _overflow_list;
7888       do {
7889         cur_overflow_list = observed_overflow_list;
7890         if (cur_overflow_list != BUSY) {
7891           // Do the splice ...
7892           suffix_tail->set_mark(markOop(cur_overflow_list));
7893         } else { // cur_overflow_list == BUSY
7894           suffix_tail->set_mark(NULL);
7895         }
7896         // ... and try to place spliced list back on overflow_list ...
7897         observed_overflow_list =
7898           (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
7899       } while (cur_overflow_list != observed_overflow_list);
7900       // ... until we have succeeded in doing so.
7901     }
7902   }
7903 
7904   // Push the prefix elements on work_q
7905   assert(prefix != NULL, "control point invariant");
7906   const markOop proto = markOopDesc::prototype();
7907   oop next;
7908   NOT_PRODUCT(ssize_t n = 0;)
7909   for (cur = prefix; cur != NULL; cur = next) {
7910     next = oop(cur->mark());
7911     cur->set_mark(proto);   // until proven otherwise
7912     assert(cur->is_oop(), "Should be an oop");
7913     bool res = work_q->push(cur);
7914     assert(res, "Bit off more than we can chew?");
7915     NOT_PRODUCT(n++;)
7916   }
7917 #ifndef PRODUCT
7918   assert(_num_par_pushes >= n, "Too many pops?");
7919   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
7920 #endif
7921   return true;
7922 }
7923 
7924 // Single-threaded
7925 void CMSCollector::push_on_overflow_list(oop p) {
7926   NOT_PRODUCT(_num_par_pushes++;)
7927   assert(p->is_oop(), "Not an oop");
7928   preserve_mark_if_necessary(p);
7929   p->set_mark((markOop)_overflow_list);
7930   _overflow_list = p;
7931 }
7932 
7933 // Multi-threaded; use CAS to prepend to overflow list
7934 void CMSCollector::par_push_on_overflow_list(oop p) {
7935   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
7936   assert(p->is_oop(), "Not an oop");
7937   par_preserve_mark_if_necessary(p);
7938   oop observed_overflow_list = _overflow_list;
7939   oop cur_overflow_list;
7940   do {
7941     cur_overflow_list = observed_overflow_list;
7942     if (cur_overflow_list != BUSY) {
7943       p->set_mark(markOop(cur_overflow_list));
7944     } else {
7945       p->set_mark(NULL);
7946     }
7947     observed_overflow_list =
7948       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
7949   } while (cur_overflow_list != observed_overflow_list);
7950 }
7951 #undef BUSY
7952 
7953 // Single threaded
7954 // General Note on GrowableArray: pushes may silently fail
7955 // because we are (temporarily) out of C-heap for expanding
7956 // the stack. The problem is quite ubiquitous and affects
7957 // a lot of code in the JVM. The prudent thing for GrowableArray
7958 // to do (for now) is to exit with an error. However, that may
7959 // be too draconian in some cases because the caller may be
7960 // able to recover without much harm. For such cases, we
7961 // should probably introduce a "soft_push" method which returns
7962 // an indication of success or failure with the assumption that
7963 // the caller may be able to recover from a failure; code in
7964 // the VM can then be changed, incrementally, to deal with such
7965 // failures where possible, thus, incrementally hardening the VM
7966 // in such low resource situations.
7967 void CMSCollector::preserve_mark_work(oop p, markOop m) {
7968   _preserved_oop_stack.push(p);
7969   _preserved_mark_stack.push(m);
7970   assert(m == p->mark(), "Mark word changed");
7971   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
7972          "bijection");
7973 }
7974 
7975 // Single threaded
7976 void CMSCollector::preserve_mark_if_necessary(oop p) {
7977   markOop m = p->mark();
7978   if (m->must_be_preserved(p)) {
7979     preserve_mark_work(p, m);
7980   }
7981 }
7982 
7983 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
7984   markOop m = p->mark();
7985   if (m->must_be_preserved(p)) {
7986     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
7987     // Even though we read the mark word without holding
7988     // the lock, we are assured that it will not change
7989     // because we "own" this oop, so no other thread can
7990     // be trying to push it on the overflow list; see
7991     // the assertion in preserve_mark_work() that checks
7992     // that m == p->mark().
7993     preserve_mark_work(p, m);
7994   }
7995 }
7996 
7997 // We should be able to do this multi-threaded,
7998 // a chunk of stack being a task (this is
7999 // correct because each oop only ever appears
8000 // once in the overflow list. However, it's
8001 // not very easy to completely overlap this with
8002 // other operations, so will generally not be done
8003 // until all work's been completed. Because we
8004 // expect the preserved oop stack (set) to be small,
8005 // it's probably fine to do this single-threaded.
8006 // We can explore cleverer concurrent/overlapped/parallel
8007 // processing of preserved marks if we feel the
8008 // need for this in the future. Stack overflow should
8009 // be so rare in practice and, when it happens, its
8010 // effect on performance so great that this will
8011 // likely just be in the noise anyway.
8012 void CMSCollector::restore_preserved_marks_if_any() {
8013   assert(SafepointSynchronize::is_at_safepoint(),
8014          "world should be stopped");
8015   assert(Thread::current()->is_ConcurrentGC_thread() ||
8016          Thread::current()->is_VM_thread(),
8017          "should be single-threaded");
8018   assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8019          "bijection");
8020 
8021   while (!_preserved_oop_stack.is_empty()) {
8022     oop p = _preserved_oop_stack.pop();
8023     assert(p->is_oop(), "Should be an oop");
8024     assert(_span.contains(p), "oop should be in _span");
8025     assert(p->mark() == markOopDesc::prototype(),
8026            "Set when taken from overflow list");
8027     markOop m = _preserved_mark_stack.pop();
8028     p->set_mark(m);
8029   }
8030   assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8031          "stacks were cleared above");
8032 }
8033 
8034 #ifndef PRODUCT
8035 bool CMSCollector::no_preserved_marks() const {
8036   return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8037 }
8038 #endif
8039 
8040 // Transfer some number of overflown objects to usual marking
8041 // stack. Return true if some objects were transferred.
8042 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8043   size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8044                     (size_t)ParGCDesiredObjsFromOverflowList);
8045 
8046   bool res = _collector->take_from_overflow_list(num, _mark_stack);
8047   assert(_collector->overflow_list_is_empty() || res,
8048          "If list is not empty, we should have taken something");
8049   assert(!res || !_mark_stack->isEmpty(),
8050          "If we took something, it should now be on our stack");
8051   return res;
8052 }
8053 
8054 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8055   size_t res = _sp->block_size_no_stall(addr, _collector);
8056   if (_sp->block_is_obj(addr)) {
8057     if (_live_bit_map->isMarked(addr)) {
8058       // It can't have been dead in a previous cycle
8059       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8060     } else {
8061       _dead_bit_map->mark(addr);      // mark the dead object
8062     }
8063   }
8064   // Could be 0, if the block size could not be computed without stalling.
8065   return res;
8066 }
8067 
8068 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8069 
8070   switch (phase) {
8071     case CMSCollector::InitialMarking:
8072       initialize(true  /* fullGC */ ,
8073                  cause /* cause of the GC */,
8074                  true  /* recordGCBeginTime */,
8075                  true  /* recordPreGCUsage */,
8076                  false /* recordPeakUsage */,
8077                  false /* recordPostGCusage */,
8078                  true  /* recordAccumulatedGCTime */,
8079                  false /* recordGCEndTime */,
8080                  false /* countCollection */  );
8081       break;
8082 
8083     case CMSCollector::FinalMarking:
8084       initialize(true  /* fullGC */ ,
8085                  cause /* cause of the GC */,
8086                  false /* recordGCBeginTime */,
8087                  false /* recordPreGCUsage */,
8088                  false /* recordPeakUsage */,
8089                  false /* recordPostGCusage */,
8090                  true  /* recordAccumulatedGCTime */,
8091                  false /* recordGCEndTime */,
8092                  false /* countCollection */  );
8093       break;
8094 
8095     case CMSCollector::Sweeping:
8096       initialize(true  /* fullGC */ ,
8097                  cause /* cause of the GC */,
8098                  false /* recordGCBeginTime */,
8099                  false /* recordPreGCUsage */,
8100                  true  /* recordPeakUsage */,
8101                  true  /* recordPostGCusage */,
8102                  false /* recordAccumulatedGCTime */,
8103                  true  /* recordGCEndTime */,
8104                  true  /* countCollection */  );
8105       break;
8106 
8107     default:
8108       ShouldNotReachHere();
8109   }
8110 }