1 /*
   2  * Copyright 2001-2008 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  20  * CA 95054 USA or visit www.sun.com if you need additional information or
  21  * have any questions.
  22  *
  23  */
  24 
  25 # include "incls/_precompiled.incl"
  26 # include "incls/_concurrentMarkSweepGeneration.cpp.incl"
  27 
  28 // statics
  29 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
  30 bool          CMSCollector::_full_gc_requested          = false;
  31 
  32 //////////////////////////////////////////////////////////////////
  33 // In support of CMS/VM thread synchronization
  34 //////////////////////////////////////////////////////////////////
  35 // We split use of the CGC_lock into 2 "levels".
  36 // The low-level locking is of the usual CGC_lock monitor. We introduce
  37 // a higher level "token" (hereafter "CMS token") built on top of the
  38 // low level monitor (hereafter "CGC lock").
  39 // The token-passing protocol gives priority to the VM thread. The
  40 // CMS-lock doesn't provide any fairness guarantees, but clients
  41 // should ensure that it is only held for very short, bounded
  42 // durations.
  43 //
  44 // When either of the CMS thread or the VM thread is involved in
  45 // collection operations during which it does not want the other
  46 // thread to interfere, it obtains the CMS token.
  47 //
  48 // If either thread tries to get the token while the other has
  49 // it, that thread waits. However, if the VM thread and CMS thread
  50 // both want the token, then the VM thread gets priority while the
  51 // CMS thread waits. This ensures, for instance, that the "concurrent"
  52 // phases of the CMS thread's work do not block out the VM thread
  53 // for long periods of time as the CMS thread continues to hog
  54 // the token. (See bug 4616232).
  55 //
  56 // The baton-passing functions are, however, controlled by the
  57 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
  58 // and here the low-level CMS lock, not the high level token,
  59 // ensures mutual exclusion.
  60 //
  61 // Two important conditions that we have to satisfy:
  62 // 1. if a thread does a low-level wait on the CMS lock, then it
  63 //    relinquishes the CMS token if it were holding that token
  64 //    when it acquired the low-level CMS lock.
  65 // 2. any low-level notifications on the low-level lock
  66 //    should only be sent when a thread has relinquished the token.
  67 //
  68 // In the absence of either property, we'd have potential deadlock.
  69 //
  70 // We protect each of the CMS (concurrent and sequential) phases
  71 // with the CMS _token_, not the CMS _lock_.
  72 //
  73 // The only code protected by CMS lock is the token acquisition code
  74 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
  75 // baton-passing code.
  76 //
  77 // Unfortunately, i couldn't come up with a good abstraction to factor and
  78 // hide the naked CGC_lock manipulation in the baton-passing code
  79 // further below. That's something we should try to do. Also, the proof
  80 // of correctness of this 2-level locking scheme is far from obvious,
  81 // and potentially quite slippery. We have an uneasy supsicion, for instance,
  82 // that there may be a theoretical possibility of delay/starvation in the
  83 // low-level lock/wait/notify scheme used for the baton-passing because of
  84 // potential intereference with the priority scheme embodied in the
  85 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
  86 // invocation further below and marked with "XXX 20011219YSR".
  87 // Indeed, as we note elsewhere, this may become yet more slippery
  88 // in the presence of multiple CMS and/or multiple VM threads. XXX
  89 
  90 class CMSTokenSync: public StackObj {
  91  private:
  92   bool _is_cms_thread;
  93  public:
  94   CMSTokenSync(bool is_cms_thread):
  95     _is_cms_thread(is_cms_thread) {
  96     assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
  97            "Incorrect argument to constructor");
  98     ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
  99   }
 100 
 101   ~CMSTokenSync() {
 102     assert(_is_cms_thread ?
 103              ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
 104              ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
 105           "Incorrect state");
 106     ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
 107   }
 108 };
 109 
 110 // Convenience class that does a CMSTokenSync, and then acquires
 111 // upto three locks.
 112 class CMSTokenSyncWithLocks: public CMSTokenSync {
 113  private:
 114   // Note: locks are acquired in textual declaration order
 115   // and released in the opposite order
 116   MutexLockerEx _locker1, _locker2, _locker3;
 117  public:
 118   CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
 119                         Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
 120     CMSTokenSync(is_cms_thread),
 121     _locker1(mutex1, Mutex::_no_safepoint_check_flag),
 122     _locker2(mutex2, Mutex::_no_safepoint_check_flag),
 123     _locker3(mutex3, Mutex::_no_safepoint_check_flag)
 124   { }
 125 };
 126 
 127 
 128 // Wrapper class to temporarily disable icms during a foreground cms collection.
 129 class ICMSDisabler: public StackObj {
 130  public:
 131   // The ctor disables icms and wakes up the thread so it notices the change;
 132   // the dtor re-enables icms.  Note that the CMSCollector methods will check
 133   // CMSIncrementalMode.
 134   ICMSDisabler()  { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
 135   ~ICMSDisabler() { CMSCollector::enable_icms(); }
 136 };
 137 
 138 //////////////////////////////////////////////////////////////////
 139 //  Concurrent Mark-Sweep Generation /////////////////////////////
 140 //////////////////////////////////////////////////////////////////
 141 
 142 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
 143 
 144 // This struct contains per-thread things necessary to support parallel
 145 // young-gen collection.
 146 class CMSParGCThreadState: public CHeapObj {
 147  public:
 148   CFLS_LAB lab;
 149   PromotionInfo promo;
 150 
 151   // Constructor.
 152   CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
 153     promo.setSpace(cfls);
 154   }
 155 };
 156 
 157 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
 158      ReservedSpace rs, size_t initial_byte_size, int level,
 159      CardTableRS* ct, bool use_adaptive_freelists,
 160      FreeBlockDictionary::DictionaryChoice dictionaryChoice) :
 161   CardGeneration(rs, initial_byte_size, level, ct),
 162   _dilatation_factor(((double)MinChunkSize)/((double)(oopDesc::header_size()))),
 163   _debug_collection_type(Concurrent_collection_type)
 164 {
 165   HeapWord* bottom = (HeapWord*) _virtual_space.low();
 166   HeapWord* end    = (HeapWord*) _virtual_space.high();
 167 
 168   _direct_allocated_words = 0;
 169   NOT_PRODUCT(
 170     _numObjectsPromoted = 0;
 171     _numWordsPromoted = 0;
 172     _numObjectsAllocated = 0;
 173     _numWordsAllocated = 0;
 174   )
 175 
 176   _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
 177                                            use_adaptive_freelists,
 178                                            dictionaryChoice);
 179   NOT_PRODUCT(debug_cms_space = _cmsSpace;)
 180   if (_cmsSpace == NULL) {
 181     vm_exit_during_initialization(
 182       "CompactibleFreeListSpace allocation failure");
 183   }
 184   _cmsSpace->_gen = this;
 185 
 186   _gc_stats = new CMSGCStats();
 187 
 188   // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
 189   // offsets match. The ability to tell free chunks from objects
 190   // depends on this property.
 191   debug_only(
 192     FreeChunk* junk = NULL;
 193     assert(UseCompressedOops ||
 194            junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
 195            "Offset of FreeChunk::_prev within FreeChunk must match"
 196            "  that of OopDesc::_klass within OopDesc");
 197   )
 198   if (ParallelGCThreads > 0) {
 199     typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
 200     _par_gc_thread_states =
 201       NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
 202     if (_par_gc_thread_states == NULL) {
 203       vm_exit_during_initialization("Could not allocate par gc structs");
 204     }
 205     for (uint i = 0; i < ParallelGCThreads; i++) {
 206       _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
 207       if (_par_gc_thread_states[i] == NULL) {
 208         vm_exit_during_initialization("Could not allocate par gc structs");
 209       }
 210     }
 211   } else {
 212     _par_gc_thread_states = NULL;
 213   }
 214   _incremental_collection_failed = false;
 215   // The "dilatation_factor" is the expansion that can occur on
 216   // account of the fact that the minimum object size in the CMS
 217   // generation may be larger than that in, say, a contiguous young
 218   //  generation.
 219   // Ideally, in the calculation below, we'd compute the dilatation
 220   // factor as: MinChunkSize/(promoting_gen's min object size)
 221   // Since we do not have such a general query interface for the
 222   // promoting generation, we'll instead just use the mimimum
 223   // object size (which today is a header's worth of space);
 224   // note that all arithmetic is in units of HeapWords.
 225   assert(MinChunkSize >= oopDesc::header_size(), "just checking");
 226   assert(_dilatation_factor >= 1.0, "from previous assert");
 227 }
 228 
 229 
 230 // The field "_initiating_occupancy" represents the occupancy percentage
 231 // at which we trigger a new collection cycle.  Unless explicitly specified
 232 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
 233 // is calculated by:
 234 //
 235 //   Let "f" be MinHeapFreeRatio in
 236 //
 237 //    _intiating_occupancy = 100-f +
 238 //                           f * (CMSTrigger[Perm]Ratio/100)
 239 //   where CMSTrigger[Perm]Ratio is the argument "tr" below.
 240 //
 241 // That is, if we assume the heap is at its desired maximum occupancy at the
 242 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
 243 // space be allocated before initiating a new collection cycle.
 244 //
 245 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
 246   assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
 247   if (io >= 0) {
 248     _initiating_occupancy = (double)io / 100.0;
 249   } else {
 250     _initiating_occupancy = ((100 - MinHeapFreeRatio) +
 251                              (double)(tr * MinHeapFreeRatio) / 100.0)
 252                             / 100.0;
 253   }
 254 }
 255 
 256 
 257 void ConcurrentMarkSweepGeneration::ref_processor_init() {
 258   assert(collector() != NULL, "no collector");
 259   collector()->ref_processor_init();
 260 }
 261 
 262 void CMSCollector::ref_processor_init() {
 263   if (_ref_processor == NULL) {
 264     // Allocate and initialize a reference processor
 265     _ref_processor = ReferenceProcessor::create_ref_processor(
 266         _span,                               // span
 267         _cmsGen->refs_discovery_is_atomic(), // atomic_discovery
 268         _cmsGen->refs_discovery_is_mt(),     // mt_discovery
 269         &_is_alive_closure,
 270         ParallelGCThreads,
 271         ParallelRefProcEnabled);
 272     // Initialize the _ref_processor field of CMSGen
 273     _cmsGen->set_ref_processor(_ref_processor);
 274 
 275     // Allocate a dummy ref processor for perm gen.
 276     ReferenceProcessor* rp2 = new ReferenceProcessor();
 277     if (rp2 == NULL) {
 278       vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
 279     }
 280     _permGen->set_ref_processor(rp2);
 281   }
 282 }
 283 
 284 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
 285   GenCollectedHeap* gch = GenCollectedHeap::heap();
 286   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
 287     "Wrong type of heap");
 288   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
 289     gch->gen_policy()->size_policy();
 290   assert(sp->is_gc_cms_adaptive_size_policy(),
 291     "Wrong type of size policy");
 292   return sp;
 293 }
 294 
 295 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
 296   CMSGCAdaptivePolicyCounters* results =
 297     (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
 298   assert(
 299     results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
 300     "Wrong gc policy counter kind");
 301   return results;
 302 }
 303 
 304 
 305 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
 306 
 307   const char* gen_name = "old";
 308 
 309   // Generation Counters - generation 1, 1 subspace
 310   _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
 311 
 312   _space_counters = new GSpaceCounters(gen_name, 0,
 313                                        _virtual_space.reserved_size(),
 314                                        this, _gen_counters);
 315 }
 316 
 317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
 318   _cms_gen(cms_gen)
 319 {
 320   assert(alpha <= 100, "bad value");
 321   _saved_alpha = alpha;
 322 
 323   // Initialize the alphas to the bootstrap value of 100.
 324   _gc0_alpha = _cms_alpha = 100;
 325 
 326   _cms_begin_time.update();
 327   _cms_end_time.update();
 328 
 329   _gc0_duration = 0.0;
 330   _gc0_period = 0.0;
 331   _gc0_promoted = 0;
 332 
 333   _cms_duration = 0.0;
 334   _cms_period = 0.0;
 335   _cms_allocated = 0;
 336 
 337   _cms_used_at_gc0_begin = 0;
 338   _cms_used_at_gc0_end = 0;
 339   _allow_duty_cycle_reduction = false;
 340   _valid_bits = 0;
 341   _icms_duty_cycle = CMSIncrementalDutyCycle;
 342 }
 343 
 344 // If promotion failure handling is on use
 345 // the padded average size of the promotion for each
 346 // young generation collection.
 347 double CMSStats::time_until_cms_gen_full() const {
 348   size_t cms_free = _cms_gen->cmsSpace()->free();
 349   GenCollectedHeap* gch = GenCollectedHeap::heap();
 350   size_t expected_promotion = gch->get_gen(0)->capacity();
 351   if (HandlePromotionFailure) {
 352     expected_promotion = MIN2(
 353         (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(),
 354         expected_promotion);
 355   }
 356   if (cms_free > expected_promotion) {
 357     // Start a cms collection if there isn't enough space to promote
 358     // for the next minor 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     cms_free_dbl = cms_free_dbl * (100.0 - CMSIncrementalSafetyFactor) / 100.0;
 365 
 366     if (PrintGCDetails && Verbose) {
 367       gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
 368         SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
 369         cms_free, expected_promotion);
 370       gclog_or_tty->print_cr("  cms_free_dbl %f cms_consumption_rate %f",
 371         cms_free_dbl, cms_consumption_rate() + 1.0);
 372     }
 373     // Add 1 in case the consumption rate goes to zero.
 374     return cms_free_dbl / (cms_consumption_rate() + 1.0);
 375   }
 376   return 0.0;
 377 }
 378 
 379 // Compare the duration of the cms collection to the
 380 // time remaining before the cms generation is empty.
 381 // Note that the time from the start of the cms collection
 382 // to the start of the cms sweep (less than the total
 383 // duration of the cms collection) can be used.  This
 384 // has been tried and some applications experienced
 385 // promotion failures early in execution.  This was
 386 // possibly because the averages were not accurate
 387 // enough at the beginning.
 388 double CMSStats::time_until_cms_start() const {
 389   // We add "gc0_period" to the "work" calculation
 390   // below because this query is done (mostly) at the
 391   // end of a scavenge, so we need to conservatively
 392   // account for that much possible delay
 393   // in the query so as to avoid concurrent mode failures
 394   // due to starting the collection just a wee bit too
 395   // late.
 396   double work = cms_duration() + gc0_period();
 397   double deadline = time_until_cms_gen_full();
 398   if (work > deadline) {
 399     if (Verbose && PrintGCDetails) {
 400       gclog_or_tty->print(
 401         " CMSCollector: collect because of anticipated promotion "
 402         "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
 403         gc0_period(), time_until_cms_gen_full());
 404     }
 405     return 0.0;
 406   }
 407   return work - deadline;
 408 }
 409 
 410 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
 411 // amount of change to prevent wild oscillation.
 412 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
 413                                               unsigned int new_duty_cycle) {
 414   assert(old_duty_cycle <= 100, "bad input value");
 415   assert(new_duty_cycle <= 100, "bad input value");
 416 
 417   // Note:  use subtraction with caution since it may underflow (values are
 418   // unsigned).  Addition is safe since we're in the range 0-100.
 419   unsigned int damped_duty_cycle = new_duty_cycle;
 420   if (new_duty_cycle < old_duty_cycle) {
 421     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
 422     if (new_duty_cycle + largest_delta < old_duty_cycle) {
 423       damped_duty_cycle = old_duty_cycle - largest_delta;
 424     }
 425   } else if (new_duty_cycle > old_duty_cycle) {
 426     const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
 427     if (new_duty_cycle > old_duty_cycle + largest_delta) {
 428       damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
 429     }
 430   }
 431   assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
 432 
 433   if (CMSTraceIncrementalPacing) {
 434     gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
 435                            old_duty_cycle, new_duty_cycle, damped_duty_cycle);
 436   }
 437   return damped_duty_cycle;
 438 }
 439 
 440 unsigned int CMSStats::icms_update_duty_cycle_impl() {
 441   assert(CMSIncrementalPacing && valid(),
 442          "should be handled in icms_update_duty_cycle()");
 443 
 444   double cms_time_so_far = cms_timer().seconds();
 445   double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
 446   double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
 447 
 448   // Avoid division by 0.
 449   double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
 450   double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
 451 
 452   unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
 453   if (new_duty_cycle > _icms_duty_cycle) {
 454     // Avoid very small duty cycles (1 or 2); 0 is allowed.
 455     if (new_duty_cycle > 2) {
 456       _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
 457                                                 new_duty_cycle);
 458     }
 459   } else if (_allow_duty_cycle_reduction) {
 460     // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
 461     new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
 462     // Respect the minimum duty cycle.
 463     unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
 464     _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
 465   }
 466 
 467   if (PrintGCDetails || CMSTraceIncrementalPacing) {
 468     gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
 469   }
 470 
 471   _allow_duty_cycle_reduction = false;
 472   return _icms_duty_cycle;
 473 }
 474 
 475 #ifndef PRODUCT
 476 void CMSStats::print_on(outputStream *st) const {
 477   st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
 478   st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
 479                gc0_duration(), gc0_period(), gc0_promoted());
 480   st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
 481             cms_duration(), cms_duration_per_mb(),
 482             cms_period(), cms_allocated());
 483   st->print(",cms_since_beg=%g,cms_since_end=%g",
 484             cms_time_since_begin(), cms_time_since_end());
 485   st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
 486             _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
 487   if (CMSIncrementalMode) {
 488     st->print(",dc=%d", icms_duty_cycle());
 489   }
 490 
 491   if (valid()) {
 492     st->print(",promo_rate=%g,cms_alloc_rate=%g",
 493               promotion_rate(), cms_allocation_rate());
 494     st->print(",cms_consumption_rate=%g,time_until_full=%g",
 495               cms_consumption_rate(), time_until_cms_gen_full());
 496   }
 497   st->print(" ");
 498 }
 499 #endif // #ifndef PRODUCT
 500 
 501 CMSCollector::CollectorState CMSCollector::_collectorState =
 502                              CMSCollector::Idling;
 503 bool CMSCollector::_foregroundGCIsActive = false;
 504 bool CMSCollector::_foregroundGCShouldWait = false;
 505 
 506 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
 507                            ConcurrentMarkSweepGeneration* permGen,
 508                            CardTableRS*                   ct,
 509                            ConcurrentMarkSweepPolicy*     cp):
 510   _cmsGen(cmsGen),
 511   _permGen(permGen),
 512   _ct(ct),
 513   _ref_processor(NULL),    // will be set later
 514   _conc_workers(NULL),     // may be set later
 515   _abort_preclean(false),
 516   _start_sampling(false),
 517   _between_prologue_and_epilogue(false),
 518   _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
 519   _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
 520   _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
 521                  -1 /* lock-free */, "No_lock" /* dummy */),
 522   _modUnionClosure(&_modUnionTable),
 523   _modUnionClosurePar(&_modUnionTable),
 524   // Adjust my span to cover old (cms) gen and perm gen
 525   _span(cmsGen->reserved()._union(permGen->reserved())),
 526   // Construct the is_alive_closure with _span & markBitMap
 527   _is_alive_closure(_span, &_markBitMap),
 528   _restart_addr(NULL),
 529   _overflow_list(NULL),
 530   _preserved_oop_stack(NULL),
 531   _preserved_mark_stack(NULL),
 532   _stats(cmsGen),
 533   _eden_chunk_array(NULL),     // may be set in ctor body
 534   _eden_chunk_capacity(0),     // -- ditto --
 535   _eden_chunk_index(0),        // -- ditto --
 536   _survivor_plab_array(NULL),  // -- ditto --
 537   _survivor_chunk_array(NULL), // -- ditto --
 538   _survivor_chunk_capacity(0), // -- ditto --
 539   _survivor_chunk_index(0),    // -- ditto --
 540   _ser_pmc_preclean_ovflw(0),
 541   _ser_kac_preclean_ovflw(0),
 542   _ser_pmc_remark_ovflw(0),
 543   _par_pmc_remark_ovflw(0),
 544   _ser_kac_ovflw(0),
 545   _par_kac_ovflw(0),
 546 #ifndef PRODUCT
 547   _num_par_pushes(0),
 548 #endif
 549   _collection_count_start(0),
 550   _verifying(false),
 551   _icms_start_limit(NULL),
 552   _icms_stop_limit(NULL),
 553   _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
 554   _completed_initialization(false),
 555   _collector_policy(cp),
 556   _should_unload_classes(false),
 557   _concurrent_cycles_since_last_unload(0),
 558   _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
 559 {
 560   if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
 561     ExplicitGCInvokesConcurrent = true;
 562   }
 563   // Now expand the span and allocate the collection support structures
 564   // (MUT, marking bit map etc.) to cover both generations subject to
 565   // collection.
 566 
 567   // First check that _permGen is adjacent to _cmsGen and above it.
 568   assert(   _cmsGen->reserved().word_size()  > 0
 569          && _permGen->reserved().word_size() > 0,
 570          "generations should not be of zero size");
 571   assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
 572          "_cmsGen and _permGen should not overlap");
 573   assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
 574          "_cmsGen->end() different from _permGen->start()");
 575 
 576   // For use by dirty card to oop closures.
 577   _cmsGen->cmsSpace()->set_collector(this);
 578   _permGen->cmsSpace()->set_collector(this);
 579 
 580   // Allocate MUT and marking bit map
 581   {
 582     MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
 583     if (!_markBitMap.allocate(_span)) {
 584       warning("Failed to allocate CMS Bit Map");
 585       return;
 586     }
 587     assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
 588   }
 589   {
 590     _modUnionTable.allocate(_span);
 591     assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
 592   }
 593 
 594   if (!_markStack.allocate(CMSMarkStackSize)) {
 595     warning("Failed to allocate CMS Marking Stack");
 596     return;
 597   }
 598   if (!_revisitStack.allocate(CMSRevisitStackSize)) {
 599     warning("Failed to allocate CMS Revisit Stack");
 600     return;
 601   }
 602 
 603   // Support for multi-threaded concurrent phases
 604   if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) {
 605     if (FLAG_IS_DEFAULT(ParallelCMSThreads)) {
 606       // just for now
 607       FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4);
 608     }
 609     if (ParallelCMSThreads > 1) {
 610       _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
 611                                  ParallelCMSThreads, true);
 612       if (_conc_workers == NULL) {
 613         warning("GC/CMS: _conc_workers allocation failure: "
 614               "forcing -CMSConcurrentMTEnabled");
 615         CMSConcurrentMTEnabled = false;
 616       }
 617     } else {
 618       CMSConcurrentMTEnabled = false;
 619     }
 620   }
 621   if (!CMSConcurrentMTEnabled) {
 622     ParallelCMSThreads = 0;
 623   } else {
 624     // Turn off CMSCleanOnEnter optimization temporarily for
 625     // the MT case where it's not fixed yet; see 6178663.
 626     CMSCleanOnEnter = false;
 627   }
 628   assert((_conc_workers != NULL) == (ParallelCMSThreads > 1),
 629          "Inconsistency");
 630 
 631   // Parallel task queues; these are shared for the
 632   // concurrent and stop-world phases of CMS, but
 633   // are not shared with parallel scavenge (ParNew).
 634   {
 635     uint i;
 636     uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads);
 637 
 638     if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
 639          || ParallelRefProcEnabled)
 640         && num_queues > 0) {
 641       _task_queues = new OopTaskQueueSet(num_queues);
 642       if (_task_queues == NULL) {
 643         warning("task_queues allocation failure.");
 644         return;
 645       }
 646       _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
 647       if (_hash_seed == NULL) {
 648         warning("_hash_seed array allocation failure");
 649         return;
 650       }
 651 
 652       // XXX use a global constant instead of 64!
 653       typedef struct OopTaskQueuePadded {
 654         OopTaskQueue work_queue;
 655         char pad[64 - sizeof(OopTaskQueue)];  // prevent false sharing
 656       } OopTaskQueuePadded;
 657 
 658       for (i = 0; i < num_queues; i++) {
 659         OopTaskQueuePadded *q_padded = new OopTaskQueuePadded();
 660         if (q_padded == NULL) {
 661           warning("work_queue allocation failure.");
 662           return;
 663         }
 664         _task_queues->register_queue(i, &q_padded->work_queue);
 665       }
 666       for (i = 0; i < num_queues; i++) {
 667         _task_queues->queue(i)->initialize();
 668         _hash_seed[i] = 17;  // copied from ParNew
 669       }
 670     }
 671   }
 672 
 673   _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
 674   _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
 675 
 676   // Clip CMSBootstrapOccupancy between 0 and 100.
 677   _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
 678                          /(double)100;
 679 
 680   _full_gcs_since_conc_gc = 0;
 681 
 682   // Now tell CMS generations the identity of their collector
 683   ConcurrentMarkSweepGeneration::set_collector(this);
 684 
 685   // Create & start a CMS thread for this CMS collector
 686   _cmsThread = ConcurrentMarkSweepThread::start(this);
 687   assert(cmsThread() != NULL, "CMS Thread should have been created");
 688   assert(cmsThread()->collector() == this,
 689          "CMS Thread should refer to this gen");
 690   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 691 
 692   // Support for parallelizing young gen rescan
 693   GenCollectedHeap* gch = GenCollectedHeap::heap();
 694   _young_gen = gch->prev_gen(_cmsGen);
 695   if (gch->supports_inline_contig_alloc()) {
 696     _top_addr = gch->top_addr();
 697     _end_addr = gch->end_addr();
 698     assert(_young_gen != NULL, "no _young_gen");
 699     _eden_chunk_index = 0;
 700     _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
 701     _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
 702     if (_eden_chunk_array == NULL) {
 703       _eden_chunk_capacity = 0;
 704       warning("GC/CMS: _eden_chunk_array allocation failure");
 705     }
 706   }
 707   assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
 708 
 709   // Support for parallelizing survivor space rescan
 710   if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
 711     size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize);
 712     _survivor_plab_array  = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
 713     _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
 714     _cursor               = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
 715     if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
 716         || _cursor == NULL) {
 717       warning("Failed to allocate survivor plab/chunk array");
 718       if (_survivor_plab_array  != NULL) {
 719         FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
 720         _survivor_plab_array = NULL;
 721       }
 722       if (_survivor_chunk_array != NULL) {
 723         FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
 724         _survivor_chunk_array = NULL;
 725       }
 726       if (_cursor != NULL) {
 727         FREE_C_HEAP_ARRAY(size_t, _cursor);
 728         _cursor = NULL;
 729       }
 730     } else {
 731       _survivor_chunk_capacity = 2*max_plab_samples;
 732       for (uint i = 0; i < ParallelGCThreads; i++) {
 733         HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
 734         if (vec == NULL) {
 735           warning("Failed to allocate survivor plab array");
 736           for (int j = i; j > 0; j--) {
 737             FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
 738           }
 739           FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
 740           FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
 741           _survivor_plab_array = NULL;
 742           _survivor_chunk_array = NULL;
 743           _survivor_chunk_capacity = 0;
 744           break;
 745         } else {
 746           ChunkArray* cur =
 747             ::new (&_survivor_plab_array[i]) ChunkArray(vec,
 748                                                         max_plab_samples);
 749           assert(cur->end() == 0, "Should be 0");
 750           assert(cur->array() == vec, "Should be vec");
 751           assert(cur->capacity() == max_plab_samples, "Error");
 752         }
 753       }
 754     }
 755   }
 756   assert(   (   _survivor_plab_array  != NULL
 757              && _survivor_chunk_array != NULL)
 758          || (   _survivor_chunk_capacity == 0
 759              && _survivor_chunk_index == 0),
 760          "Error");
 761 
 762   // Choose what strong roots should be scanned depending on verification options
 763   // and perm gen collection mode.
 764   if (!CMSClassUnloadingEnabled) {
 765     // If class unloading is disabled we want to include all classes into the root set.
 766     add_root_scanning_option(SharedHeap::SO_AllClasses);
 767   } else {
 768     add_root_scanning_option(SharedHeap::SO_SystemClasses);
 769   }
 770 
 771   NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
 772   _gc_counters = new CollectorCounters("CMS", 1);
 773   _completed_initialization = true;
 774   _sweep_timer.start();  // start of time
 775 }
 776 
 777 const char* ConcurrentMarkSweepGeneration::name() const {
 778   return "concurrent mark-sweep generation";
 779 }
 780 void ConcurrentMarkSweepGeneration::update_counters() {
 781   if (UsePerfData) {
 782     _space_counters->update_all();
 783     _gen_counters->update_all();
 784   }
 785 }
 786 
 787 // this is an optimized version of update_counters(). it takes the
 788 // used value as a parameter rather than computing it.
 789 //
 790 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
 791   if (UsePerfData) {
 792     _space_counters->update_used(used);
 793     _space_counters->update_capacity();
 794     _gen_counters->update_all();
 795   }
 796 }
 797 
 798 void ConcurrentMarkSweepGeneration::print() const {
 799   Generation::print();
 800   cmsSpace()->print();
 801 }
 802 
 803 #ifndef PRODUCT
 804 void ConcurrentMarkSweepGeneration::print_statistics() {
 805   cmsSpace()->printFLCensus(0);
 806 }
 807 #endif
 808 
 809 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
 810   GenCollectedHeap* gch = GenCollectedHeap::heap();
 811   if (PrintGCDetails) {
 812     if (Verbose) {
 813       gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
 814         level(), short_name(), s, used(), capacity());
 815     } else {
 816       gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
 817         level(), short_name(), s, used() / K, capacity() / K);
 818     }
 819   }
 820   if (Verbose) {
 821     gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
 822               gch->used(), gch->capacity());
 823   } else {
 824     gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
 825               gch->used() / K, gch->capacity() / K);
 826   }
 827 }
 828 
 829 size_t
 830 ConcurrentMarkSweepGeneration::contiguous_available() const {
 831   // dld proposes an improvement in precision here. If the committed
 832   // part of the space ends in a free block we should add that to
 833   // uncommitted size in the calculation below. Will make this
 834   // change later, staying with the approximation below for the
 835   // time being. -- ysr.
 836   return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
 837 }
 838 
 839 size_t
 840 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
 841   return _cmsSpace->max_alloc_in_words() * HeapWordSize;
 842 }
 843 
 844 size_t ConcurrentMarkSweepGeneration::max_available() const {
 845   return free() + _virtual_space.uncommitted_size();
 846 }
 847 
 848 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(
 849     size_t max_promotion_in_bytes,
 850     bool younger_handles_promotion_failure) const {
 851 
 852   // This is the most conservative test.  Full promotion is
 853   // guaranteed if this is used. The multiplicative factor is to
 854   // account for the worst case "dilatation".
 855   double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes;
 856   if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t
 857     adjusted_max_promo_bytes = (double)max_uintx;
 858   }
 859   bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes);
 860 
 861   if (younger_handles_promotion_failure && !result) {
 862     // Full promotion is not guaranteed because fragmentation
 863     // of the cms generation can prevent the full promotion.
 864     result = (max_available() >= (size_t)adjusted_max_promo_bytes);
 865 
 866     if (!result) {
 867       // With promotion failure handling the test for the ability
 868       // to support the promotion does not have to be guaranteed.
 869       // Use an average of the amount promoted.
 870       result = max_available() >= (size_t)
 871         gc_stats()->avg_promoted()->padded_average();
 872       if (PrintGC && Verbose && result) {
 873         gclog_or_tty->print_cr(
 874           "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
 875           " max_available: " SIZE_FORMAT
 876           " avg_promoted: " SIZE_FORMAT,
 877           max_available(), (size_t)
 878           gc_stats()->avg_promoted()->padded_average());
 879       }
 880     } else {
 881       if (PrintGC && Verbose) {
 882         gclog_or_tty->print_cr(
 883           "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
 884           " max_available: " SIZE_FORMAT
 885           " adj_max_promo_bytes: " SIZE_FORMAT,
 886           max_available(), (size_t)adjusted_max_promo_bytes);
 887       }
 888     }
 889   } else {
 890     if (PrintGC && Verbose) {
 891       gclog_or_tty->print_cr(
 892         "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
 893         " contiguous_available: " SIZE_FORMAT
 894         " adj_max_promo_bytes: " SIZE_FORMAT,
 895         max_contiguous_available(), (size_t)adjusted_max_promo_bytes);
 896     }
 897   }
 898   return result;
 899 }
 900 
 901 CompactibleSpace*
 902 ConcurrentMarkSweepGeneration::first_compaction_space() const {
 903   return _cmsSpace;
 904 }
 905 
 906 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
 907   // Clear the promotion information.  These pointers can be adjusted
 908   // along with all the other pointers into the heap but
 909   // compaction is expected to be a rare event with
 910   // a heap using cms so don't do it without seeing the need.
 911   if (ParallelGCThreads > 0) {
 912     for (uint i = 0; i < ParallelGCThreads; i++) {
 913       _par_gc_thread_states[i]->promo.reset();
 914     }
 915   }
 916 }
 917 
 918 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
 919   blk->do_space(_cmsSpace);
 920 }
 921 
 922 void ConcurrentMarkSweepGeneration::compute_new_size() {
 923   assert_locked_or_safepoint(Heap_lock);
 924 
 925   // If incremental collection failed, we just want to expand
 926   // to the limit.
 927   if (incremental_collection_failed()) {
 928     clear_incremental_collection_failed();
 929     grow_to_reserved();
 930     return;
 931   }
 932 
 933   size_t expand_bytes = 0;
 934   double free_percentage = ((double) free()) / capacity();
 935   double desired_free_percentage = (double) MinHeapFreeRatio / 100;
 936   double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
 937 
 938   // compute expansion delta needed for reaching desired free percentage
 939   if (free_percentage < desired_free_percentage) {
 940     size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 941     assert(desired_capacity >= capacity(), "invalid expansion size");
 942     expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
 943   }
 944   if (expand_bytes > 0) {
 945     if (PrintGCDetails && Verbose) {
 946       size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
 947       gclog_or_tty->print_cr("\nFrom compute_new_size: ");
 948       gclog_or_tty->print_cr("  Free fraction %f", free_percentage);
 949       gclog_or_tty->print_cr("  Desired free fraction %f",
 950         desired_free_percentage);
 951       gclog_or_tty->print_cr("  Maximum free fraction %f",
 952         maximum_free_percentage);
 953       gclog_or_tty->print_cr("  Capactiy "SIZE_FORMAT, capacity()/1000);
 954       gclog_or_tty->print_cr("  Desired capacity "SIZE_FORMAT,
 955         desired_capacity/1000);
 956       int prev_level = level() - 1;
 957       if (prev_level >= 0) {
 958         size_t prev_size = 0;
 959         GenCollectedHeap* gch = GenCollectedHeap::heap();
 960         Generation* prev_gen = gch->_gens[prev_level];
 961         prev_size = prev_gen->capacity();
 962           gclog_or_tty->print_cr("  Younger gen size "SIZE_FORMAT,
 963                                  prev_size/1000);
 964       }
 965       gclog_or_tty->print_cr("  unsafe_max_alloc_nogc "SIZE_FORMAT,
 966         unsafe_max_alloc_nogc()/1000);
 967       gclog_or_tty->print_cr("  contiguous available "SIZE_FORMAT,
 968         contiguous_available()/1000);
 969       gclog_or_tty->print_cr("  Expand by "SIZE_FORMAT" (bytes)",
 970         expand_bytes);
 971     }
 972     // safe if expansion fails
 973     expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
 974     if (PrintGCDetails && Verbose) {
 975       gclog_or_tty->print_cr("  Expanded free fraction %f",
 976         ((double) free()) / capacity());
 977     }
 978   }
 979 }
 980 
 981 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
 982   return cmsSpace()->freelistLock();
 983 }
 984 
 985 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
 986                                                   bool   tlab) {
 987   CMSSynchronousYieldRequest yr;
 988   MutexLockerEx x(freelistLock(),
 989                   Mutex::_no_safepoint_check_flag);
 990   return have_lock_and_allocate(size, tlab);
 991 }
 992 
 993 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
 994                                                   bool   tlab) {
 995   assert_lock_strong(freelistLock());
 996   size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
 997   HeapWord* res = cmsSpace()->allocate(adjustedSize);
 998   // Allocate the object live (grey) if the background collector has
 999   // started marking. This is necessary because the marker may
1000   // have passed this address and consequently this object will
1001   // not otherwise be greyed and would be incorrectly swept up.
1002   // Note that if this object contains references, the writing
1003   // of those references will dirty the card containing this object
1004   // allowing the object to be blackened (and its references scanned)
1005   // either during a preclean phase or at the final checkpoint.
1006   if (res != NULL) {
1007     collector()->direct_allocated(res, adjustedSize);
1008     _direct_allocated_words += adjustedSize;
1009     // allocation counters
1010     NOT_PRODUCT(
1011       _numObjectsAllocated++;
1012       _numWordsAllocated += (int)adjustedSize;
1013     )
1014   }
1015   return res;
1016 }
1017 
1018 // In the case of direct allocation by mutators in a generation that
1019 // is being concurrently collected, the object must be allocated
1020 // live (grey) if the background collector has started marking.
1021 // This is necessary because the marker may
1022 // have passed this address and consequently this object will
1023 // not otherwise be greyed and would be incorrectly swept up.
1024 // Note that if this object contains references, the writing
1025 // of those references will dirty the card containing this object
1026 // allowing the object to be blackened (and its references scanned)
1027 // either during a preclean phase or at the final checkpoint.
1028 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1029   assert(_markBitMap.covers(start, size), "Out of bounds");
1030   if (_collectorState >= Marking) {
1031     MutexLockerEx y(_markBitMap.lock(),
1032                     Mutex::_no_safepoint_check_flag);
1033     // [see comments preceding SweepClosure::do_blk() below for details]
1034     // 1. need to mark the object as live so it isn't collected
1035     // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1036     // 3. need to mark the end of the object so sweeper can skip over it
1037     //    if it's uninitialized when the sweeper reaches it.
1038     _markBitMap.mark(start);          // object is live
1039     _markBitMap.mark(start + 1);      // object is potentially uninitialized?
1040     _markBitMap.mark(start + size - 1);
1041                                       // mark end of object
1042   }
1043   // check that oop looks uninitialized
1044   assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1045 }
1046 
1047 void CMSCollector::promoted(bool par, HeapWord* start,
1048                             bool is_obj_array, size_t obj_size) {
1049   assert(_markBitMap.covers(start), "Out of bounds");
1050   // See comment in direct_allocated() about when objects should
1051   // be allocated live.
1052   if (_collectorState >= Marking) {
1053     // we already hold the marking bit map lock, taken in
1054     // the prologue
1055     if (par) {
1056       _markBitMap.par_mark(start);
1057     } else {
1058       _markBitMap.mark(start);
1059     }
1060     // We don't need to mark the object as uninitialized (as
1061     // in direct_allocated above) because this is being done with the
1062     // world stopped and the object will be initialized by the
1063     // time the sweeper gets to look at it.
1064     assert(SafepointSynchronize::is_at_safepoint(),
1065            "expect promotion only at safepoints");
1066 
1067     if (_collectorState < Sweeping) {
1068       // Mark the appropriate cards in the modUnionTable, so that
1069       // this object gets scanned before the sweep. If this is
1070       // not done, CMS generation references in the object might
1071       // not get marked.
1072       // For the case of arrays, which are otherwise precisely
1073       // marked, we need to dirty the entire array, not just its head.
1074       if (is_obj_array) {
1075         // The [par_]mark_range() method expects mr.end() below to
1076         // be aligned to the granularity of a bit's representation
1077         // in the heap. In the case of the MUT below, that's a
1078         // card size.
1079         MemRegion mr(start,
1080                      (HeapWord*)round_to((intptr_t)(start + obj_size),
1081                         CardTableModRefBS::card_size /* bytes */));
1082         if (par) {
1083           _modUnionTable.par_mark_range(mr);
1084         } else {
1085           _modUnionTable.mark_range(mr);
1086         }
1087       } else {  // not an obj array; we can just mark the head
1088         if (par) {
1089           _modUnionTable.par_mark(start);
1090         } else {
1091           _modUnionTable.mark(start);
1092         }
1093       }
1094     }
1095   }
1096 }
1097 
1098 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1099 {
1100   size_t delta = pointer_delta(addr, space->bottom());
1101   return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1102 }
1103 
1104 void CMSCollector::icms_update_allocation_limits()
1105 {
1106   Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1107   EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1108 
1109   const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1110   if (CMSTraceIncrementalPacing) {
1111     stats().print();
1112   }
1113 
1114   assert(duty_cycle <= 100, "invalid duty cycle");
1115   if (duty_cycle != 0) {
1116     // The duty_cycle is a percentage between 0 and 100; convert to words and
1117     // then compute the offset from the endpoints of the space.
1118     size_t free_words = eden->free() / HeapWordSize;
1119     double free_words_dbl = (double)free_words;
1120     size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1121     size_t offset_words = (free_words - duty_cycle_words) / 2;
1122 
1123     _icms_start_limit = eden->top() + offset_words;
1124     _icms_stop_limit = eden->end() - offset_words;
1125 
1126     // The limits may be adjusted (shifted to the right) by
1127     // CMSIncrementalOffset, to allow the application more mutator time after a
1128     // young gen gc (when all mutators were stopped) and before CMS starts and
1129     // takes away one or more cpus.
1130     if (CMSIncrementalOffset != 0) {
1131       double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1132       size_t adjustment = (size_t)adjustment_dbl;
1133       HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1134       if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1135         _icms_start_limit += adjustment;
1136         _icms_stop_limit = tmp_stop;
1137       }
1138     }
1139   }
1140   if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1141     _icms_start_limit = _icms_stop_limit = eden->end();
1142   }
1143 
1144   // Install the new start limit.
1145   eden->set_soft_end(_icms_start_limit);
1146 
1147   if (CMSTraceIncrementalMode) {
1148     gclog_or_tty->print(" icms alloc limits:  "
1149                            PTR_FORMAT "," PTR_FORMAT
1150                            " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1151                            _icms_start_limit, _icms_stop_limit,
1152                            percent_of_space(eden, _icms_start_limit),
1153                            percent_of_space(eden, _icms_stop_limit));
1154     if (Verbose) {
1155       gclog_or_tty->print("eden:  ");
1156       eden->print_on(gclog_or_tty);
1157     }
1158   }
1159 }
1160 
1161 // Any changes here should try to maintain the invariant
1162 // that if this method is called with _icms_start_limit
1163 // and _icms_stop_limit both NULL, then it should return NULL
1164 // and not notify the icms thread.
1165 HeapWord*
1166 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1167                                        size_t word_size)
1168 {
1169   // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1170   // nop.
1171   if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1172     if (top <= _icms_start_limit) {
1173       if (CMSTraceIncrementalMode) {
1174         space->print_on(gclog_or_tty);
1175         gclog_or_tty->stamp();
1176         gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1177                                ", new limit=" PTR_FORMAT
1178                                " (" SIZE_FORMAT "%%)",
1179                                top, _icms_stop_limit,
1180                                percent_of_space(space, _icms_stop_limit));
1181       }
1182       ConcurrentMarkSweepThread::start_icms();
1183       assert(top < _icms_stop_limit, "Tautology");
1184       if (word_size < pointer_delta(_icms_stop_limit, top)) {
1185         return _icms_stop_limit;
1186       }
1187 
1188       // The allocation will cross both the _start and _stop limits, so do the
1189       // stop notification also and return end().
1190       if (CMSTraceIncrementalMode) {
1191         space->print_on(gclog_or_tty);
1192         gclog_or_tty->stamp();
1193         gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1194                                ", new limit=" PTR_FORMAT
1195                                " (" SIZE_FORMAT "%%)",
1196                                top, space->end(),
1197                                percent_of_space(space, space->end()));
1198       }
1199       ConcurrentMarkSweepThread::stop_icms();
1200       return space->end();
1201     }
1202 
1203     if (top <= _icms_stop_limit) {
1204       if (CMSTraceIncrementalMode) {
1205         space->print_on(gclog_or_tty);
1206         gclog_or_tty->stamp();
1207         gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1208                                ", new limit=" PTR_FORMAT
1209                                " (" SIZE_FORMAT "%%)",
1210                                top, space->end(),
1211                                percent_of_space(space, space->end()));
1212       }
1213       ConcurrentMarkSweepThread::stop_icms();
1214       return space->end();
1215     }
1216 
1217     if (CMSTraceIncrementalMode) {
1218       space->print_on(gclog_or_tty);
1219       gclog_or_tty->stamp();
1220       gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1221                              ", new limit=" PTR_FORMAT,
1222                              top, NULL);
1223     }
1224   }
1225 
1226   return NULL;
1227 }
1228 
1229 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1230   assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1231   // allocate, copy and if necessary update promoinfo --
1232   // delegate to underlying space.
1233   assert_lock_strong(freelistLock());
1234 
1235 #ifndef PRODUCT
1236   if (Universe::heap()->promotion_should_fail()) {
1237     return NULL;
1238   }
1239 #endif  // #ifndef PRODUCT
1240 
1241   oop res = _cmsSpace->promote(obj, obj_size);
1242   if (res == NULL) {
1243     // expand and retry
1244     size_t s = _cmsSpace->expansionSpaceRequired(obj_size);  // HeapWords
1245     expand(s*HeapWordSize, MinHeapDeltaBytes,
1246       CMSExpansionCause::_satisfy_promotion);
1247     // Since there's currently no next generation, we don't try to promote
1248     // into a more senior generation.
1249     assert(next_gen() == NULL, "assumption, based upon which no attempt "
1250                                "is made to pass on a possibly failing "
1251                                "promotion to next generation");
1252     res = _cmsSpace->promote(obj, obj_size);
1253   }
1254   if (res != NULL) {
1255     // See comment in allocate() about when objects should
1256     // be allocated live.
1257     assert(obj->is_oop(), "Will dereference klass pointer below");
1258     collector()->promoted(false,           // Not parallel
1259                           (HeapWord*)res, obj->is_objArray(), obj_size);
1260     // promotion counters
1261     NOT_PRODUCT(
1262       _numObjectsPromoted++;
1263       _numWordsPromoted +=
1264         (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1265     )
1266   }
1267   return res;
1268 }
1269 
1270 
1271 HeapWord*
1272 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1273                                              HeapWord* top,
1274                                              size_t word_sz)
1275 {
1276   return collector()->allocation_limit_reached(space, top, word_sz);
1277 }
1278 
1279 // Things to support parallel young-gen collection.
1280 oop
1281 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1282                                            oop old, markOop m,
1283                                            size_t word_sz) {
1284 #ifndef PRODUCT
1285   if (Universe::heap()->promotion_should_fail()) {
1286     return NULL;
1287   }
1288 #endif  // #ifndef PRODUCT
1289 
1290   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1291   PromotionInfo* promoInfo = &ps->promo;
1292   // if we are tracking promotions, then first ensure space for
1293   // promotion (including spooling space for saving header if necessary).
1294   // then allocate and copy, then track promoted info if needed.
1295   // When tracking (see PromotionInfo::track()), the mark word may
1296   // be displaced and in this case restoration of the mark word
1297   // occurs in the (oop_since_save_marks_)iterate phase.
1298   if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1299     // Out of space for allocating spooling buffers;
1300     // try expanding and allocating spooling buffers.
1301     if (!expand_and_ensure_spooling_space(promoInfo)) {
1302       return NULL;
1303     }
1304   }
1305   assert(promoInfo->has_spooling_space(), "Control point invariant");
1306   HeapWord* obj_ptr = ps->lab.alloc(word_sz);
1307   if (obj_ptr == NULL) {
1308      obj_ptr = expand_and_par_lab_allocate(ps, word_sz);
1309      if (obj_ptr == NULL) {
1310        return NULL;
1311      }
1312   }
1313   oop obj = oop(obj_ptr);
1314   assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1315   // Otherwise, copy the object.  Here we must be careful to insert the
1316   // klass pointer last, since this marks the block as an allocated object.
1317   // Except with compressed oops it's the mark word.
1318   HeapWord* old_ptr = (HeapWord*)old;
1319   if (word_sz > (size_t)oopDesc::header_size()) {
1320     Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1321                                  obj_ptr + oopDesc::header_size(),
1322                                  word_sz - oopDesc::header_size());
1323   }
1324 
1325   if (UseCompressedOops) {
1326     // Copy gap missed by (aligned) header size calculation above
1327     obj->set_klass_gap(old->klass_gap());
1328   }
1329 
1330   // Restore the mark word copied above.
1331   obj->set_mark(m);
1332 
1333   // Now we can track the promoted object, if necessary.  We take care
1334   // To delay the transition from uninitialized to full object
1335   // (i.e., insertion of klass pointer) until after, so that it
1336   // atomically becomes a promoted object.
1337   if (promoInfo->tracking()) {
1338     promoInfo->track((PromotedObject*)obj, old->klass());
1339   }
1340 
1341   // Finally, install the klass pointer (this should be volatile).
1342   obj->set_klass(old->klass());
1343 
1344   assert(old->is_oop(), "Will dereference klass ptr below");
1345   collector()->promoted(true,          // parallel
1346                         obj_ptr, old->is_objArray(), word_sz);
1347 
1348   NOT_PRODUCT(
1349     Atomic::inc(&_numObjectsPromoted);
1350     Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()),
1351                 &_numWordsPromoted);
1352   )
1353 
1354   return obj;
1355 }
1356 
1357 void
1358 ConcurrentMarkSweepGeneration::
1359 par_promote_alloc_undo(int thread_num,
1360                        HeapWord* obj, size_t word_sz) {
1361   // CMS does not support promotion undo.
1362   ShouldNotReachHere();
1363 }
1364 
1365 void
1366 ConcurrentMarkSweepGeneration::
1367 par_promote_alloc_done(int thread_num) {
1368   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1369   ps->lab.retire();
1370 #if CFLS_LAB_REFILL_STATS
1371   if (thread_num == 0) {
1372     _cmsSpace->print_par_alloc_stats();
1373   }
1374 #endif
1375 }
1376 
1377 void
1378 ConcurrentMarkSweepGeneration::
1379 par_oop_since_save_marks_iterate_done(int thread_num) {
1380   CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1381   ParScanWithoutBarrierClosure* dummy_cl = NULL;
1382   ps->promo.promoted_oops_iterate_nv(dummy_cl);
1383 }
1384 
1385 // XXXPERM
1386 bool ConcurrentMarkSweepGeneration::should_collect(bool   full,
1387                                                    size_t size,
1388                                                    bool   tlab)
1389 {
1390   // We allow a STW collection only if a full
1391   // collection was requested.
1392   return full || should_allocate(size, tlab); // FIX ME !!!
1393   // This and promotion failure handling are connected at the
1394   // hip and should be fixed by untying them.
1395 }
1396 
1397 bool CMSCollector::shouldConcurrentCollect() {
1398   if (_full_gc_requested) {
1399     assert(ExplicitGCInvokesConcurrent, "Unexpected state");
1400     if (Verbose && PrintGCDetails) {
1401       gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1402                              " gc request");
1403     }
1404     return true;
1405   }
1406 
1407   // For debugging purposes, change the type of collection.
1408   // If the rotation is not on the concurrent collection
1409   // type, don't start a concurrent collection.
1410   NOT_PRODUCT(
1411     if (RotateCMSCollectionTypes &&
1412         (_cmsGen->debug_collection_type() !=
1413           ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1414       assert(_cmsGen->debug_collection_type() !=
1415         ConcurrentMarkSweepGeneration::Unknown_collection_type,
1416         "Bad cms collection type");
1417       return false;
1418     }
1419   )
1420 
1421   FreelistLocker x(this);
1422   // ------------------------------------------------------------------
1423   // Print out lots of information which affects the initiation of
1424   // a collection.
1425   if (PrintCMSInitiationStatistics && stats().valid()) {
1426     gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1427     gclog_or_tty->stamp();
1428     gclog_or_tty->print_cr("");
1429     stats().print_on(gclog_or_tty);
1430     gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1431       stats().time_until_cms_gen_full());
1432     gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1433     gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1434                            _cmsGen->contiguous_available());
1435     gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1436     gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1437     gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1438     gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1439     gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1440   }
1441   // ------------------------------------------------------------------
1442 
1443   // If the estimated time to complete a cms collection (cms_duration())
1444   // is less than the estimated time remaining until the cms generation
1445   // is full, start a collection.
1446   if (!UseCMSInitiatingOccupancyOnly) {
1447     if (stats().valid()) {
1448       if (stats().time_until_cms_start() == 0.0) {
1449         return true;
1450       }
1451     } else {
1452       // We want to conservatively collect somewhat early in order
1453       // to try and "bootstrap" our CMS/promotion statistics;
1454       // this branch will not fire after the first successful CMS
1455       // collection because the stats should then be valid.
1456       if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1457         if (Verbose && PrintGCDetails) {
1458           gclog_or_tty->print_cr(
1459             " CMSCollector: collect for bootstrapping statistics:"
1460             " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1461             _bootstrap_occupancy);
1462         }
1463         return true;
1464       }
1465     }
1466   }
1467 
1468   // Otherwise, we start a collection cycle if either the perm gen or
1469   // old gen want a collection cycle started. Each may use
1470   // an appropriate criterion for making this decision.
1471   // XXX We need to make sure that the gen expansion
1472   // criterion dovetails well with this. XXX NEED TO FIX THIS
1473   if (_cmsGen->should_concurrent_collect()) {
1474     if (Verbose && PrintGCDetails) {
1475       gclog_or_tty->print_cr("CMS old gen initiated");
1476     }
1477     return true;
1478   }
1479 
1480   // We start a collection if we believe an incremental collection may fail;
1481   // this is not likely to be productive in practice because it's probably too
1482   // late anyway.
1483   GenCollectedHeap* gch = GenCollectedHeap::heap();
1484   assert(gch->collector_policy()->is_two_generation_policy(),
1485          "You may want to check the correctness of the following");
1486   if (gch->incremental_collection_will_fail()) {
1487     if (PrintGCDetails && Verbose) {
1488       gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1489     }
1490     return true;
1491   }
1492 
1493   if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1494     bool res = update_should_unload_classes();
1495     if (res) {
1496       if (Verbose && PrintGCDetails) {
1497         gclog_or_tty->print_cr("CMS perm gen initiated");
1498       }
1499       return true;
1500     }
1501   }
1502   return false;
1503 }
1504 
1505 // Clear _expansion_cause fields of constituent generations
1506 void CMSCollector::clear_expansion_cause() {
1507   _cmsGen->clear_expansion_cause();
1508   _permGen->clear_expansion_cause();
1509 }
1510 
1511 // We should be conservative in starting a collection cycle.  To
1512 // start too eagerly runs the risk of collecting too often in the
1513 // extreme.  To collect too rarely falls back on full collections,
1514 // which works, even if not optimum in terms of concurrent work.
1515 // As a work around for too eagerly collecting, use the flag
1516 // UseCMSInitiatingOccupancyOnly.  This also has the advantage of
1517 // giving the user an easily understandable way of controlling the
1518 // collections.
1519 // We want to start a new collection cycle if any of the following
1520 // conditions hold:
1521 // . our current occupancy exceeds the configured initiating occupancy
1522 //   for this generation, or
1523 // . we recently needed to expand this space and have not, since that
1524 //   expansion, done a collection of this generation, or
1525 // . the underlying space believes that it may be a good idea to initiate
1526 //   a concurrent collection (this may be based on criteria such as the
1527 //   following: the space uses linear allocation and linear allocation is
1528 //   going to fail, or there is believed to be excessive fragmentation in
1529 //   the generation, etc... or ...
1530 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1531 //   the case of the old generation, not the perm generation; see CR 6543076):
1532 //   we may be approaching a point at which allocation requests may fail because
1533 //   we will be out of sufficient free space given allocation rate estimates.]
1534 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1535 
1536   assert_lock_strong(freelistLock());
1537   if (occupancy() > initiating_occupancy()) {
1538     if (PrintGCDetails && Verbose) {
1539       gclog_or_tty->print(" %s: collect because of occupancy %f / %f  ",
1540         short_name(), occupancy(), initiating_occupancy());
1541     }
1542     return true;
1543   }
1544   if (UseCMSInitiatingOccupancyOnly) {
1545     return false;
1546   }
1547   if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1548     if (PrintGCDetails && Verbose) {
1549       gclog_or_tty->print(" %s: collect because expanded for allocation ",
1550         short_name());
1551     }
1552     return true;
1553   }
1554   if (_cmsSpace->should_concurrent_collect()) {
1555     if (PrintGCDetails && Verbose) {
1556       gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1557         short_name());
1558     }
1559     return true;
1560   }
1561   return false;
1562 }
1563 
1564 void ConcurrentMarkSweepGeneration::collect(bool   full,
1565                                             bool   clear_all_soft_refs,
1566                                             size_t size,
1567                                             bool   tlab)
1568 {
1569   collector()->collect(full, clear_all_soft_refs, size, tlab);
1570 }
1571 
1572 void CMSCollector::collect(bool   full,
1573                            bool   clear_all_soft_refs,
1574                            size_t size,
1575                            bool   tlab)
1576 {
1577   if (!UseCMSCollectionPassing && _collectorState > Idling) {
1578     // For debugging purposes skip the collection if the state
1579     // is not currently idle
1580     if (TraceCMSState) {
1581       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1582         Thread::current(), full, _collectorState);
1583     }
1584     return;
1585   }
1586 
1587   // The following "if" branch is present for defensive reasons.
1588   // In the current uses of this interface, it can be replaced with:
1589   // assert(!GC_locker.is_active(), "Can't be called otherwise");
1590   // But I am not placing that assert here to allow future
1591   // generality in invoking this interface.
1592   if (GC_locker::is_active()) {
1593     // A consistency test for GC_locker
1594     assert(GC_locker::needs_gc(), "Should have been set already");
1595     // Skip this foreground collection, instead
1596     // expanding the heap if necessary.
1597     // Need the free list locks for the call to free() in compute_new_size()
1598     compute_new_size();
1599     return;
1600   }
1601   acquire_control_and_collect(full, clear_all_soft_refs);
1602   _full_gcs_since_conc_gc++;
1603 
1604 }
1605 
1606 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1607   GenCollectedHeap* gch = GenCollectedHeap::heap();
1608   unsigned int gc_count = gch->total_full_collections();
1609   if (gc_count == full_gc_count) {
1610     MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1611     _full_gc_requested = true;
1612     CGC_lock->notify();   // nudge CMS thread
1613   }
1614 }
1615 
1616 
1617 // The foreground and background collectors need to coordinate in order
1618 // to make sure that they do not mutually interfere with CMS collections.
1619 // When a background collection is active,
1620 // the foreground collector may need to take over (preempt) and
1621 // synchronously complete an ongoing collection. Depending on the
1622 // frequency of the background collections and the heap usage
1623 // of the application, this preemption can be seldom or frequent.
1624 // There are only certain
1625 // points in the background collection that the "collection-baton"
1626 // can be passed to the foreground collector.
1627 //
1628 // The foreground collector will wait for the baton before
1629 // starting any part of the collection.  The foreground collector
1630 // will only wait at one location.
1631 //
1632 // The background collector will yield the baton before starting a new
1633 // phase of the collection (e.g., before initial marking, marking from roots,
1634 // precleaning, final re-mark, sweep etc.)  This is normally done at the head
1635 // of the loop which switches the phases. The background collector does some
1636 // of the phases (initial mark, final re-mark) with the world stopped.
1637 // Because of locking involved in stopping the world,
1638 // the foreground collector should not block waiting for the background
1639 // collector when it is doing a stop-the-world phase.  The background
1640 // collector will yield the baton at an additional point just before
1641 // it enters a stop-the-world phase.  Once the world is stopped, the
1642 // background collector checks the phase of the collection.  If the
1643 // phase has not changed, it proceeds with the collection.  If the
1644 // phase has changed, it skips that phase of the collection.  See
1645 // the comments on the use of the Heap_lock in collect_in_background().
1646 //
1647 // Variable used in baton passing.
1648 //   _foregroundGCIsActive - Set to true by the foreground collector when
1649 //      it wants the baton.  The foreground clears it when it has finished
1650 //      the collection.
1651 //   _foregroundGCShouldWait - Set to true by the background collector
1652 //        when it is running.  The foreground collector waits while
1653 //      _foregroundGCShouldWait is true.
1654 //  CGC_lock - monitor used to protect access to the above variables
1655 //      and to notify the foreground and background collectors.
1656 //  _collectorState - current state of the CMS collection.
1657 //
1658 // The foreground collector
1659 //   acquires the CGC_lock
1660 //   sets _foregroundGCIsActive
1661 //   waits on the CGC_lock for _foregroundGCShouldWait to be false
1662 //     various locks acquired in preparation for the collection
1663 //     are released so as not to block the background collector
1664 //     that is in the midst of a collection
1665 //   proceeds with the collection
1666 //   clears _foregroundGCIsActive
1667 //   returns
1668 //
1669 // The background collector in a loop iterating on the phases of the
1670 //      collection
1671 //   acquires the CGC_lock
1672 //   sets _foregroundGCShouldWait
1673 //   if _foregroundGCIsActive is set
1674 //     clears _foregroundGCShouldWait, notifies _CGC_lock
1675 //     waits on _CGC_lock for _foregroundGCIsActive to become false
1676 //     and exits the loop.
1677 //   otherwise
1678 //     proceed with that phase of the collection
1679 //     if the phase is a stop-the-world phase,
1680 //       yield the baton once more just before enqueueing
1681 //       the stop-world CMS operation (executed by the VM thread).
1682 //   returns after all phases of the collection are done
1683 //
1684 
1685 void CMSCollector::acquire_control_and_collect(bool full,
1686         bool clear_all_soft_refs) {
1687   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1688   assert(!Thread::current()->is_ConcurrentGC_thread(),
1689          "shouldn't try to acquire control from self!");
1690 
1691   // Start the protocol for acquiring control of the
1692   // collection from the background collector (aka CMS thread).
1693   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1694          "VM thread should have CMS token");
1695   // Remember the possibly interrupted state of an ongoing
1696   // concurrent collection
1697   CollectorState first_state = _collectorState;
1698 
1699   // Signal to a possibly ongoing concurrent collection that
1700   // we want to do a foreground collection.
1701   _foregroundGCIsActive = true;
1702 
1703   // Disable incremental mode during a foreground collection.
1704   ICMSDisabler icms_disabler;
1705 
1706   // release locks and wait for a notify from the background collector
1707   // releasing the locks in only necessary for phases which
1708   // do yields to improve the granularity of the collection.
1709   assert_lock_strong(bitMapLock());
1710   // We need to lock the Free list lock for the space that we are
1711   // currently collecting.
1712   assert(haveFreelistLocks(), "Must be holding free list locks");
1713   bitMapLock()->unlock();
1714   releaseFreelistLocks();
1715   {
1716     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1717     if (_foregroundGCShouldWait) {
1718       // We are going to be waiting for action for the CMS thread;
1719       // it had better not be gone (for instance at shutdown)!
1720       assert(ConcurrentMarkSweepThread::cmst() != NULL,
1721              "CMS thread must be running");
1722       // Wait here until the background collector gives us the go-ahead
1723       ConcurrentMarkSweepThread::clear_CMS_flag(
1724         ConcurrentMarkSweepThread::CMS_vm_has_token);  // release token
1725       // Get a possibly blocked CMS thread going:
1726       //   Note that we set _foregroundGCIsActive true above,
1727       //   without protection of the CGC_lock.
1728       CGC_lock->notify();
1729       assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1730              "Possible deadlock");
1731       while (_foregroundGCShouldWait) {
1732         // wait for notification
1733         CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1734         // Possibility of delay/starvation here, since CMS token does
1735         // not know to give priority to VM thread? Actually, i think
1736         // there wouldn't be any delay/starvation, but the proof of
1737         // that "fact" (?) appears non-trivial. XXX 20011219YSR
1738       }
1739       ConcurrentMarkSweepThread::set_CMS_flag(
1740         ConcurrentMarkSweepThread::CMS_vm_has_token);
1741     }
1742   }
1743   // The CMS_token is already held.  Get back the other locks.
1744   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1745          "VM thread should have CMS token");
1746   getFreelistLocks();
1747   bitMapLock()->lock_without_safepoint_check();
1748   if (TraceCMSState) {
1749     gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1750       INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1751     gclog_or_tty->print_cr("    gets control with state %d", _collectorState);
1752   }
1753 
1754   // Check if we need to do a compaction, or if not, whether
1755   // we need to start the mark-sweep from scratch.
1756   bool should_compact    = false;
1757   bool should_start_over = false;
1758   decide_foreground_collection_type(clear_all_soft_refs,
1759     &should_compact, &should_start_over);
1760 
1761 NOT_PRODUCT(
1762   if (RotateCMSCollectionTypes) {
1763     if (_cmsGen->debug_collection_type() ==
1764         ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1765       should_compact = true;
1766     } else if (_cmsGen->debug_collection_type() ==
1767                ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1768       should_compact = false;
1769     }
1770   }
1771 )
1772 
1773   if (PrintGCDetails && first_state > Idling) {
1774     GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1775     if (GCCause::is_user_requested_gc(cause) ||
1776         GCCause::is_serviceability_requested_gc(cause)) {
1777       gclog_or_tty->print(" (concurrent mode interrupted)");
1778     } else {
1779       gclog_or_tty->print(" (concurrent mode failure)");
1780     }
1781   }
1782 
1783   if (should_compact) {
1784     // If the collection is being acquired from the background
1785     // collector, there may be references on the discovered
1786     // references lists that have NULL referents (being those
1787     // that were concurrently cleared by a mutator) or
1788     // that are no longer active (having been enqueued concurrently
1789     // by the mutator).
1790     // Scrub the list of those references because Mark-Sweep-Compact
1791     // code assumes referents are not NULL and that all discovered
1792     // Reference objects are active.
1793     ref_processor()->clean_up_discovered_references();
1794 
1795     do_compaction_work(clear_all_soft_refs);
1796 
1797     // Has the GC time limit been exceeded?
1798     check_gc_time_limit();
1799 
1800   } else {
1801     do_mark_sweep_work(clear_all_soft_refs, first_state,
1802       should_start_over);
1803   }
1804   // Reset the expansion cause, now that we just completed
1805   // a collection cycle.
1806   clear_expansion_cause();
1807   _foregroundGCIsActive = false;
1808   return;
1809 }
1810 
1811 void CMSCollector::check_gc_time_limit() {
1812 
1813   // Ignore explicit GC's.  Exiting here does not set the flag and
1814   // does not reset the count.  Updating of the averages for system
1815   // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC.
1816   GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause();
1817   if (GCCause::is_user_requested_gc(gc_cause) ||
1818       GCCause::is_serviceability_requested_gc(gc_cause)) {
1819     return;
1820   }
1821 
1822   // Calculate the fraction of the CMS generation was freed during
1823   // the last collection.
1824   // Only consider the STW compacting cost for now.
1825   //
1826   // Note that the gc time limit test only works for the collections
1827   // of the young gen + tenured gen and not for collections of the
1828   // permanent gen.  That is because the calculation of the space
1829   // freed by the collection is the free space in the young gen +
1830   // tenured gen.
1831 
1832   double fraction_free =
1833     ((double)_cmsGen->free())/((double)_cmsGen->max_capacity());
1834   if ((100.0 * size_policy()->compacting_gc_cost()) >
1835          ((double) GCTimeLimit) &&
1836         ((fraction_free * 100) < GCHeapFreeLimit)) {
1837     size_policy()->inc_gc_time_limit_count();
1838     if (UseGCOverheadLimit &&
1839         (size_policy()->gc_time_limit_count() >
1840          AdaptiveSizePolicyGCTimeLimitThreshold)) {
1841       size_policy()->set_gc_time_limit_exceeded(true);
1842       // Avoid consecutive OOM due to the gc time limit by resetting
1843       // the counter.
1844       size_policy()->reset_gc_time_limit_count();
1845       if (PrintGCDetails) {
1846         gclog_or_tty->print_cr("      GC is exceeding overhead limit "
1847           "of %d%%", GCTimeLimit);
1848       }
1849     } else {
1850       if (PrintGCDetails) {
1851         gclog_or_tty->print_cr("      GC would exceed overhead limit "
1852           "of %d%%", GCTimeLimit);
1853       }
1854     }
1855   } else {
1856     size_policy()->reset_gc_time_limit_count();
1857   }
1858 }
1859 
1860 // Resize the perm generation and the tenured generation
1861 // after obtaining the free list locks for the
1862 // two generations.
1863 void CMSCollector::compute_new_size() {
1864   assert_locked_or_safepoint(Heap_lock);
1865   FreelistLocker z(this);
1866   _permGen->compute_new_size();
1867   _cmsGen->compute_new_size();
1868 }
1869 
1870 // A work method used by foreground collection to determine
1871 // what type of collection (compacting or not, continuing or fresh)
1872 // it should do.
1873 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1874 // and CMSCompactWhenClearAllSoftRefs the default in the future
1875 // and do away with the flags after a suitable period.
1876 void CMSCollector::decide_foreground_collection_type(
1877   bool clear_all_soft_refs, bool* should_compact,
1878   bool* should_start_over) {
1879   // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1880   // flag is set, and we have either requested a System.gc() or
1881   // the number of full gc's since the last concurrent cycle
1882   // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1883   // or if an incremental collection has failed
1884   GenCollectedHeap* gch = GenCollectedHeap::heap();
1885   assert(gch->collector_policy()->is_two_generation_policy(),
1886          "You may want to check the correctness of the following");
1887   // Inform cms gen if this was due to partial collection failing.
1888   // The CMS gen may use this fact to determine its expansion policy.
1889   if (gch->incremental_collection_will_fail()) {
1890     assert(!_cmsGen->incremental_collection_failed(),
1891            "Should have been noticed, reacted to and cleared");
1892     _cmsGen->set_incremental_collection_failed();
1893   }
1894   *should_compact =
1895     UseCMSCompactAtFullCollection &&
1896     ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1897      GCCause::is_user_requested_gc(gch->gc_cause()) ||
1898      gch->incremental_collection_will_fail());
1899   *should_start_over = false;
1900   if (clear_all_soft_refs && !*should_compact) {
1901     // We are about to do a last ditch collection attempt
1902     // so it would normally make sense to do a compaction
1903     // to reclaim as much space as possible.
1904     if (CMSCompactWhenClearAllSoftRefs) {
1905       // Default: The rationale is that in this case either
1906       // we are past the final marking phase, in which case
1907       // we'd have to start over, or so little has been done
1908       // that there's little point in saving that work. Compaction
1909       // appears to be the sensible choice in either case.
1910       *should_compact = true;
1911     } else {
1912       // We have been asked to clear all soft refs, but not to
1913       // compact. Make sure that we aren't past the final checkpoint
1914       // phase, for that is where we process soft refs. If we are already
1915       // past that phase, we'll need to redo the refs discovery phase and
1916       // if necessary clear soft refs that weren't previously
1917       // cleared. We do so by remembering the phase in which
1918       // we came in, and if we are past the refs processing
1919       // phase, we'll choose to just redo the mark-sweep
1920       // collection from scratch.
1921       if (_collectorState > FinalMarking) {
1922         // We are past the refs processing phase;
1923         // start over and do a fresh synchronous CMS cycle
1924         _collectorState = Resetting; // skip to reset to start new cycle
1925         reset(false /* == !asynch */);
1926         *should_start_over = true;
1927       } // else we can continue a possibly ongoing current cycle
1928     }
1929   }
1930 }
1931 
1932 // A work method used by the foreground collector to do
1933 // a mark-sweep-compact.
1934 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1935   GenCollectedHeap* gch = GenCollectedHeap::heap();
1936   TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1937   if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1938     gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1939       "collections passed to foreground collector", _full_gcs_since_conc_gc);
1940   }
1941 
1942   // Sample collection interval time and reset for collection pause.
1943   if (UseAdaptiveSizePolicy) {
1944     size_policy()->msc_collection_begin();
1945   }
1946 
1947   // Temporarily widen the span of the weak reference processing to
1948   // the entire heap.
1949   MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1950   ReferenceProcessorSpanMutator x(ref_processor(), new_span);
1951 
1952   // Temporarily, clear the "is_alive_non_header" field of the
1953   // reference processor.
1954   ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
1955 
1956   // Temporarily make reference _processing_ single threaded (non-MT).
1957   ReferenceProcessorMTProcMutator z(ref_processor(), false);
1958 
1959   // Temporarily make refs discovery atomic
1960   ReferenceProcessorAtomicMutator w(ref_processor(), true);
1961 
1962   ref_processor()->set_enqueuing_is_done(false);
1963   ref_processor()->enable_discovery();
1964   // If an asynchronous collection finishes, the _modUnionTable is
1965   // all clear.  If we are assuming the collection from an asynchronous
1966   // collection, clear the _modUnionTable.
1967   assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1968     "_modUnionTable should be clear if the baton was not passed");
1969   _modUnionTable.clear_all();
1970 
1971   // We must adjust the allocation statistics being maintained
1972   // in the free list space. We do so by reading and clearing
1973   // the sweep timer and updating the block flux rate estimates below.
1974   assert(_sweep_timer.is_active(), "We should never see the timer inactive");
1975   _sweep_timer.stop();
1976   // Note that we do not use this sample to update the _sweep_estimate.
1977   _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
1978                                           _sweep_estimate.padded_average());
1979 
1980   GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1981     ref_processor(), clear_all_soft_refs);
1982   #ifdef ASSERT
1983     CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1984     size_t free_size = cms_space->free();
1985     assert(free_size ==
1986            pointer_delta(cms_space->end(), cms_space->compaction_top())
1987            * HeapWordSize,
1988       "All the free space should be compacted into one chunk at top");
1989     assert(cms_space->dictionary()->totalChunkSize(
1990                                       debug_only(cms_space->freelistLock())) == 0 ||
1991            cms_space->totalSizeInIndexedFreeLists() == 0,
1992       "All the free space should be in a single chunk");
1993     size_t num = cms_space->totalCount();
1994     assert((free_size == 0 && num == 0) ||
1995            (free_size > 0  && (num == 1 || num == 2)),
1996          "There should be at most 2 free chunks after compaction");
1997   #endif // ASSERT
1998   _collectorState = Resetting;
1999   assert(_restart_addr == NULL,
2000          "Should have been NULL'd before baton was passed");
2001   reset(false /* == !asynch */);
2002   _cmsGen->reset_after_compaction();
2003   _concurrent_cycles_since_last_unload = 0;
2004 
2005   if (verifying() && !should_unload_classes()) {
2006     perm_gen_verify_bit_map()->clear_all();
2007   }
2008 
2009   // Clear any data recorded in the PLAB chunk arrays.
2010   if (_survivor_plab_array != NULL) {
2011     reset_survivor_plab_arrays();
2012   }
2013 
2014   // Adjust the per-size allocation stats for the next epoch.
2015   _cmsGen->cmsSpace()->endSweepFLCensus(sweepCount() /* fake */);
2016   // Restart the "sweep timer" for next epoch.
2017   _sweep_timer.reset();
2018   _sweep_timer.start();
2019 
2020   // Sample collection pause time and reset for collection interval.
2021   if (UseAdaptiveSizePolicy) {
2022     size_policy()->msc_collection_end(gch->gc_cause());
2023   }
2024 
2025   // For a mark-sweep-compact, compute_new_size() will be called
2026   // in the heap's do_collection() method.
2027 }
2028 
2029 // A work method used by the foreground collector to do
2030 // a mark-sweep, after taking over from a possibly on-going
2031 // concurrent mark-sweep collection.
2032 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2033   CollectorState first_state, bool should_start_over) {
2034   if (PrintGC && Verbose) {
2035     gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2036       "collector with count %d",
2037       _full_gcs_since_conc_gc);
2038   }
2039   switch (_collectorState) {
2040     case Idling:
2041       if (first_state == Idling || should_start_over) {
2042         // The background GC was not active, or should
2043         // restarted from scratch;  start the cycle.
2044         _collectorState = InitialMarking;
2045       }
2046       // If first_state was not Idling, then a background GC
2047       // was in progress and has now finished.  No need to do it
2048       // again.  Leave the state as Idling.
2049       break;
2050     case Precleaning:
2051       // In the foreground case don't do the precleaning since
2052       // it is not done concurrently and there is extra work
2053       // required.
2054       _collectorState = FinalMarking;
2055   }
2056   if (PrintGCDetails &&
2057       (_collectorState > Idling ||
2058        !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2059     gclog_or_tty->print(" (concurrent mode failure)");
2060   }
2061   collect_in_foreground(clear_all_soft_refs);
2062 
2063   // For a mark-sweep, compute_new_size() will be called
2064   // in the heap's do_collection() method.
2065 }
2066 
2067 
2068 void CMSCollector::getFreelistLocks() const {
2069   // Get locks for all free lists in all generations that this
2070   // collector is responsible for
2071   _cmsGen->freelistLock()->lock_without_safepoint_check();
2072   _permGen->freelistLock()->lock_without_safepoint_check();
2073 }
2074 
2075 void CMSCollector::releaseFreelistLocks() const {
2076   // Release locks for all free lists in all generations that this
2077   // collector is responsible for
2078   _cmsGen->freelistLock()->unlock();
2079   _permGen->freelistLock()->unlock();
2080 }
2081 
2082 bool CMSCollector::haveFreelistLocks() const {
2083   // Check locks for all free lists in all generations that this
2084   // collector is responsible for
2085   assert_lock_strong(_cmsGen->freelistLock());
2086   assert_lock_strong(_permGen->freelistLock());
2087   PRODUCT_ONLY(ShouldNotReachHere());
2088   return true;
2089 }
2090 
2091 // A utility class that is used by the CMS collector to
2092 // temporarily "release" the foreground collector from its
2093 // usual obligation to wait for the background collector to
2094 // complete an ongoing phase before proceeding.
2095 class ReleaseForegroundGC: public StackObj {
2096  private:
2097   CMSCollector* _c;
2098  public:
2099   ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2100     assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2101     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2102     // allow a potentially blocked foreground collector to proceed
2103     _c->_foregroundGCShouldWait = false;
2104     if (_c->_foregroundGCIsActive) {
2105       CGC_lock->notify();
2106     }
2107     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2108            "Possible deadlock");
2109   }
2110 
2111   ~ReleaseForegroundGC() {
2112     assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2113     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2114     _c->_foregroundGCShouldWait = true;
2115   }
2116 };
2117 
2118 // There are separate collect_in_background and collect_in_foreground because of
2119 // the different locking requirements of the background collector and the
2120 // foreground collector.  There was originally an attempt to share
2121 // one "collect" method between the background collector and the foreground
2122 // collector but the if-then-else required made it cleaner to have
2123 // separate methods.
2124 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2125   assert(Thread::current()->is_ConcurrentGC_thread(),
2126     "A CMS asynchronous collection is only allowed on a CMS thread.");
2127 
2128   GenCollectedHeap* gch = GenCollectedHeap::heap();
2129   {
2130     bool safepoint_check = Mutex::_no_safepoint_check_flag;
2131     MutexLockerEx hl(Heap_lock, safepoint_check);
2132     FreelistLocker fll(this);
2133     MutexLockerEx x(CGC_lock, safepoint_check);
2134     if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2135       // The foreground collector is active or we're
2136       // not using asynchronous collections.  Skip this
2137       // background collection.
2138       assert(!_foregroundGCShouldWait, "Should be clear");
2139       return;
2140     } else {
2141       assert(_collectorState == Idling, "Should be idling before start.");
2142       _collectorState = InitialMarking;
2143       // Reset the expansion cause, now that we are about to begin
2144       // a new cycle.
2145       clear_expansion_cause();
2146     }
2147     // Decide if we want to enable class unloading as part of the
2148     // ensuing concurrent GC cycle.
2149     update_should_unload_classes();
2150     _full_gc_requested = false;           // acks all outstanding full gc requests
2151     // Signal that we are about to start a collection
2152     gch->increment_total_full_collections();  // ... starting a collection cycle
2153     _collection_count_start = gch->total_full_collections();
2154   }
2155 
2156   // Used for PrintGC
2157   size_t prev_used;
2158   if (PrintGC && Verbose) {
2159     prev_used = _cmsGen->used(); // XXXPERM
2160   }
2161 
2162   // The change of the collection state is normally done at this level;
2163   // the exceptions are phases that are executed while the world is
2164   // stopped.  For those phases the change of state is done while the
2165   // world is stopped.  For baton passing purposes this allows the
2166   // background collector to finish the phase and change state atomically.
2167   // The foreground collector cannot wait on a phase that is done
2168   // while the world is stopped because the foreground collector already
2169   // has the world stopped and would deadlock.
2170   while (_collectorState != Idling) {
2171     if (TraceCMSState) {
2172       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2173         Thread::current(), _collectorState);
2174     }
2175     // The foreground collector
2176     //   holds the Heap_lock throughout its collection.
2177     //   holds the CMS token (but not the lock)
2178     //     except while it is waiting for the background collector to yield.
2179     //
2180     // The foreground collector should be blocked (not for long)
2181     //   if the background collector is about to start a phase
2182     //   executed with world stopped.  If the background
2183     //   collector has already started such a phase, the
2184     //   foreground collector is blocked waiting for the
2185     //   Heap_lock.  The stop-world phases (InitialMarking and FinalMarking)
2186     //   are executed in the VM thread.
2187     //
2188     // The locking order is
2189     //   PendingListLock (PLL)  -- if applicable (FinalMarking)
2190     //   Heap_lock  (both this & PLL locked in VM_CMS_Operation::prologue())
2191     //   CMS token  (claimed in
2192     //                stop_world_and_do() -->
2193     //                  safepoint_synchronize() -->
2194     //                    CMSThread::synchronize())
2195 
2196     {
2197       // Check if the FG collector wants us to yield.
2198       CMSTokenSync x(true); // is cms thread
2199       if (waitForForegroundGC()) {
2200         // We yielded to a foreground GC, nothing more to be
2201         // done this round.
2202         assert(_foregroundGCShouldWait == false, "We set it to false in "
2203                "waitForForegroundGC()");
2204         if (TraceCMSState) {
2205           gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2206             " exiting collection CMS state %d",
2207             Thread::current(), _collectorState);
2208         }
2209         return;
2210       } else {
2211         // The background collector can run but check to see if the
2212         // foreground collector has done a collection while the
2213         // background collector was waiting to get the CGC_lock
2214         // above.  If yes, break so that _foregroundGCShouldWait
2215         // is cleared before returning.
2216         if (_collectorState == Idling) {
2217           break;
2218         }
2219       }
2220     }
2221 
2222     assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2223       "should be waiting");
2224 
2225     switch (_collectorState) {
2226       case InitialMarking:
2227         {
2228           ReleaseForegroundGC x(this);
2229           stats().record_cms_begin();
2230 
2231           VM_CMS_Initial_Mark initial_mark_op(this);
2232           VMThread::execute(&initial_mark_op);
2233         }
2234         // The collector state may be any legal state at this point
2235         // since the background collector may have yielded to the
2236         // foreground collector.
2237         break;
2238       case Marking:
2239         // initial marking in checkpointRootsInitialWork has been completed
2240         if (markFromRoots(true)) { // we were successful
2241           assert(_collectorState == Precleaning, "Collector state should "
2242             "have changed");
2243         } else {
2244           assert(_foregroundGCIsActive, "Internal state inconsistency");
2245         }
2246         break;
2247       case Precleaning:
2248         if (UseAdaptiveSizePolicy) {
2249           size_policy()->concurrent_precleaning_begin();
2250         }
2251         // marking from roots in markFromRoots has been completed
2252         preclean();
2253         if (UseAdaptiveSizePolicy) {
2254           size_policy()->concurrent_precleaning_end();
2255         }
2256         assert(_collectorState == AbortablePreclean ||
2257                _collectorState == FinalMarking,
2258                "Collector state should have changed");
2259         break;
2260       case AbortablePreclean:
2261         if (UseAdaptiveSizePolicy) {
2262         size_policy()->concurrent_phases_resume();
2263         }
2264         abortable_preclean();
2265         if (UseAdaptiveSizePolicy) {
2266           size_policy()->concurrent_precleaning_end();
2267         }
2268         assert(_collectorState == FinalMarking, "Collector state should "
2269           "have changed");
2270         break;
2271       case FinalMarking:
2272         {
2273           ReleaseForegroundGC x(this);
2274 
2275           VM_CMS_Final_Remark final_remark_op(this);
2276           VMThread::execute(&final_remark_op);
2277           }
2278         assert(_foregroundGCShouldWait, "block post-condition");
2279         break;
2280       case Sweeping:
2281         if (UseAdaptiveSizePolicy) {
2282           size_policy()->concurrent_sweeping_begin();
2283         }
2284         // final marking in checkpointRootsFinal has been completed
2285         sweep(true);
2286         assert(_collectorState == Resizing, "Collector state change "
2287           "to Resizing must be done under the free_list_lock");
2288         _full_gcs_since_conc_gc = 0;
2289 
2290         // Stop the timers for adaptive size policy for the concurrent phases
2291         if (UseAdaptiveSizePolicy) {
2292           size_policy()->concurrent_sweeping_end();
2293           size_policy()->concurrent_phases_end(gch->gc_cause(),
2294                                              gch->prev_gen(_cmsGen)->capacity(),
2295                                              _cmsGen->free());
2296         }
2297 
2298       case Resizing: {
2299         // Sweeping has been completed...
2300         // At this point the background collection has completed.
2301         // Don't move the call to compute_new_size() down
2302         // into code that might be executed if the background
2303         // collection was preempted.
2304         {
2305           ReleaseForegroundGC x(this);   // unblock FG collection
2306           MutexLockerEx       y(Heap_lock, Mutex::_no_safepoint_check_flag);
2307           CMSTokenSync        z(true);   // not strictly needed.
2308           if (_collectorState == Resizing) {
2309             compute_new_size();
2310             _collectorState = Resetting;
2311           } else {
2312             assert(_collectorState == Idling, "The state should only change"
2313                    " because the foreground collector has finished the collection");
2314           }
2315         }
2316         break;
2317       }
2318       case Resetting:
2319         // CMS heap resizing has been completed
2320         reset(true);
2321         assert(_collectorState == Idling, "Collector state should "
2322           "have changed");
2323         stats().record_cms_end();
2324         // Don't move the concurrent_phases_end() and compute_new_size()
2325         // calls to here because a preempted background collection
2326         // has it's state set to "Resetting".
2327         break;
2328       case Idling:
2329       default:
2330         ShouldNotReachHere();
2331         break;
2332     }
2333     if (TraceCMSState) {
2334       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
2335         Thread::current(), _collectorState);
2336     }
2337     assert(_foregroundGCShouldWait, "block post-condition");
2338   }
2339 
2340   // Should this be in gc_epilogue?
2341   collector_policy()->counters()->update_counters();
2342 
2343   {
2344     // Clear _foregroundGCShouldWait and, in the event that the
2345     // foreground collector is waiting, notify it, before
2346     // returning.
2347     MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2348     _foregroundGCShouldWait = false;
2349     if (_foregroundGCIsActive) {
2350       CGC_lock->notify();
2351     }
2352     assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2353            "Possible deadlock");
2354   }
2355   if (TraceCMSState) {
2356     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2357       " exiting collection CMS state %d",
2358       Thread::current(), _collectorState);
2359   }
2360   if (PrintGC && Verbose) {
2361     _cmsGen->print_heap_change(prev_used);
2362   }
2363 }
2364 
2365 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2366   assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2367          "Foreground collector should be waiting, not executing");
2368   assert(Thread::current()->is_VM_thread(), "A foreground collection"
2369     "may only be done by the VM Thread with the world stopped");
2370   assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2371          "VM thread should have CMS token");
2372 
2373   NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2374     true, gclog_or_tty);)
2375   if (UseAdaptiveSizePolicy) {
2376     size_policy()->ms_collection_begin();
2377   }
2378   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2379 
2380   HandleMark hm;  // Discard invalid handles created during verification
2381 
2382   if (VerifyBeforeGC &&
2383       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2384     Universe::verify(true);
2385   }
2386 
2387   bool init_mark_was_synchronous = false; // until proven otherwise
2388   while (_collectorState != Idling) {
2389     if (TraceCMSState) {
2390       gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2391         Thread::current(), _collectorState);
2392     }
2393     switch (_collectorState) {
2394       case InitialMarking:
2395         init_mark_was_synchronous = true;  // fact to be exploited in re-mark
2396         checkpointRootsInitial(false);
2397         assert(_collectorState == Marking, "Collector state should have changed"
2398           " within checkpointRootsInitial()");
2399         break;
2400       case Marking:
2401         // initial marking in checkpointRootsInitialWork has been completed
2402         if (VerifyDuringGC &&
2403             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2404           gclog_or_tty->print("Verify before initial mark: ");
2405           Universe::verify(true);
2406         }
2407         {
2408           bool res = markFromRoots(false);
2409           assert(res && _collectorState == FinalMarking, "Collector state should "
2410             "have changed");
2411           break;
2412         }
2413       case FinalMarking:
2414         if (VerifyDuringGC &&
2415             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2416           gclog_or_tty->print("Verify before re-mark: ");
2417           Universe::verify(true);
2418         }
2419         checkpointRootsFinal(false, clear_all_soft_refs,
2420                              init_mark_was_synchronous);
2421         assert(_collectorState == Sweeping, "Collector state should not "
2422           "have changed within checkpointRootsFinal()");
2423         break;
2424       case Sweeping:
2425         // final marking in checkpointRootsFinal has been completed
2426         if (VerifyDuringGC &&
2427             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2428           gclog_or_tty->print("Verify before sweep: ");
2429           Universe::verify(true);
2430         }
2431         sweep(false);
2432         assert(_collectorState == Resizing, "Incorrect state");
2433         break;
2434       case Resizing: {
2435         // Sweeping has been completed; the actual resize in this case
2436         // is done separately; nothing to be done in this state.
2437         _collectorState = Resetting;
2438         break;
2439       }
2440       case Resetting:
2441         // The heap has been resized.
2442         if (VerifyDuringGC &&
2443             GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2444           gclog_or_tty->print("Verify before reset: ");
2445           Universe::verify(true);
2446         }
2447         reset(false);
2448         assert(_collectorState == Idling, "Collector state should "
2449           "have changed");
2450         break;
2451       case Precleaning:
2452       case AbortablePreclean:
2453         // Elide the preclean phase
2454         _collectorState = FinalMarking;
2455         break;
2456       default:
2457         ShouldNotReachHere();
2458     }
2459     if (TraceCMSState) {
2460       gclog_or_tty->print_cr("  Thread " INTPTR_FORMAT " done - next CMS state %d",
2461         Thread::current(), _collectorState);
2462     }
2463   }
2464 
2465   if (UseAdaptiveSizePolicy) {
2466     GenCollectedHeap* gch = GenCollectedHeap::heap();
2467     size_policy()->ms_collection_end(gch->gc_cause());
2468   }
2469 
2470   if (VerifyAfterGC &&
2471       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2472     Universe::verify(true);
2473   }
2474   if (TraceCMSState) {
2475     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2476       " exiting collection CMS state %d",
2477       Thread::current(), _collectorState);
2478   }
2479 }
2480 
2481 bool CMSCollector::waitForForegroundGC() {
2482   bool res = false;
2483   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2484          "CMS thread should have CMS token");
2485   // Block the foreground collector until the
2486   // background collectors decides whether to
2487   // yield.
2488   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2489   _foregroundGCShouldWait = true;
2490   if (_foregroundGCIsActive) {
2491     // The background collector yields to the
2492     // foreground collector and returns a value
2493     // indicating that it has yielded.  The foreground
2494     // collector can proceed.
2495     res = true;
2496     _foregroundGCShouldWait = false;
2497     ConcurrentMarkSweepThread::clear_CMS_flag(
2498       ConcurrentMarkSweepThread::CMS_cms_has_token);
2499     ConcurrentMarkSweepThread::set_CMS_flag(
2500       ConcurrentMarkSweepThread::CMS_cms_wants_token);
2501     // Get a possibly blocked foreground thread going
2502     CGC_lock->notify();
2503     if (TraceCMSState) {
2504       gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2505         Thread::current(), _collectorState);
2506     }
2507     while (_foregroundGCIsActive) {
2508       CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2509     }
2510     ConcurrentMarkSweepThread::set_CMS_flag(
2511       ConcurrentMarkSweepThread::CMS_cms_has_token);
2512     ConcurrentMarkSweepThread::clear_CMS_flag(
2513       ConcurrentMarkSweepThread::CMS_cms_wants_token);
2514   }
2515   if (TraceCMSState) {
2516     gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2517       Thread::current(), _collectorState);
2518   }
2519   return res;
2520 }
2521 
2522 // Because of the need to lock the free lists and other structures in
2523 // the collector, common to all the generations that the collector is
2524 // collecting, we need the gc_prologues of individual CMS generations
2525 // delegate to their collector. It may have been simpler had the
2526 // current infrastructure allowed one to call a prologue on a
2527 // collector. In the absence of that we have the generation's
2528 // prologue delegate to the collector, which delegates back
2529 // some "local" work to a worker method in the individual generations
2530 // that it's responsible for collecting, while itself doing any
2531 // work common to all generations it's responsible for. A similar
2532 // comment applies to the  gc_epilogue()'s.
2533 // The role of the varaible _between_prologue_and_epilogue is to
2534 // enforce the invocation protocol.
2535 void CMSCollector::gc_prologue(bool full) {
2536   // Call gc_prologue_work() for each CMSGen and PermGen that
2537   // we are responsible for.
2538 
2539   // The following locking discipline assumes that we are only called
2540   // when the world is stopped.
2541   assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2542 
2543   // The CMSCollector prologue must call the gc_prologues for the
2544   // "generations" (including PermGen if any) that it's responsible
2545   // for.
2546 
2547   assert(   Thread::current()->is_VM_thread()
2548          || (   CMSScavengeBeforeRemark
2549              && Thread::current()->is_ConcurrentGC_thread()),
2550          "Incorrect thread type for prologue execution");
2551 
2552   if (_between_prologue_and_epilogue) {
2553     // We have already been invoked; this is a gc_prologue delegation
2554     // from yet another CMS generation that we are responsible for, just
2555     // ignore it since all relevant work has already been done.
2556     return;
2557   }
2558 
2559   // set a bit saying prologue has been called; cleared in epilogue
2560   _between_prologue_and_epilogue = true;
2561   // Claim locks for common data structures, then call gc_prologue_work()
2562   // for each CMSGen and PermGen that we are responsible for.
2563 
2564   getFreelistLocks();   // gets free list locks on constituent spaces
2565   bitMapLock()->lock_without_safepoint_check();
2566 
2567   // Should call gc_prologue_work() for all cms gens we are responsible for
2568   bool registerClosure =    _collectorState >= Marking
2569                          && _collectorState < Sweeping;
2570   ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar
2571                                                : &_modUnionClosure;
2572   _cmsGen->gc_prologue_work(full, registerClosure, muc);
2573   _permGen->gc_prologue_work(full, registerClosure, muc);
2574 
2575   if (!full) {
2576     stats().record_gc0_begin();
2577   }
2578 }
2579 
2580 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2581   // Delegate to CMScollector which knows how to coordinate between
2582   // this and any other CMS generations that it is responsible for
2583   // collecting.
2584   collector()->gc_prologue(full);
2585 }
2586 
2587 // This is a "private" interface for use by this generation's CMSCollector.
2588 // Not to be called directly by any other entity (for instance,
2589 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2590 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2591   bool registerClosure, ModUnionClosure* modUnionClosure) {
2592   assert(!incremental_collection_failed(), "Shouldn't be set yet");
2593   assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2594     "Should be NULL");
2595   if (registerClosure) {
2596     cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2597   }
2598   cmsSpace()->gc_prologue();
2599   // Clear stat counters
2600   NOT_PRODUCT(
2601     assert(_numObjectsPromoted == 0, "check");
2602     assert(_numWordsPromoted   == 0, "check");
2603     if (Verbose && PrintGC) {
2604       gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2605                           SIZE_FORMAT" bytes concurrently",
2606       _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2607     }
2608     _numObjectsAllocated = 0;
2609     _numWordsAllocated   = 0;
2610   )
2611 }
2612 
2613 void CMSCollector::gc_epilogue(bool full) {
2614   // The following locking discipline assumes that we are only called
2615   // when the world is stopped.
2616   assert(SafepointSynchronize::is_at_safepoint(),
2617          "world is stopped assumption");
2618 
2619   // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2620   // if linear allocation blocks need to be appropriately marked to allow the
2621   // the blocks to be parsable. We also check here whether we need to nudge the
2622   // CMS collector thread to start a new cycle (if it's not already active).
2623   assert(   Thread::current()->is_VM_thread()
2624          || (   CMSScavengeBeforeRemark
2625              && Thread::current()->is_ConcurrentGC_thread()),
2626          "Incorrect thread type for epilogue execution");
2627 
2628   if (!_between_prologue_and_epilogue) {
2629     // We have already been invoked; this is a gc_epilogue delegation
2630     // from yet another CMS generation that we are responsible for, just
2631     // ignore it since all relevant work has already been done.
2632     return;
2633   }
2634   assert(haveFreelistLocks(), "must have freelist locks");
2635   assert_lock_strong(bitMapLock());
2636 
2637   _cmsGen->gc_epilogue_work(full);
2638   _permGen->gc_epilogue_work(full);
2639 
2640   if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2641     // in case sampling was not already enabled, enable it
2642     _start_sampling = true;
2643   }
2644   // reset _eden_chunk_array so sampling starts afresh
2645   _eden_chunk_index = 0;
2646 
2647   size_t cms_used   = _cmsGen->cmsSpace()->used();
2648   size_t perm_used  = _permGen->cmsSpace()->used();
2649 
2650   // update performance counters - this uses a special version of
2651   // update_counters() that allows the utilization to be passed as a
2652   // parameter, avoiding multiple calls to used().
2653   //
2654   _cmsGen->update_counters(cms_used);
2655   _permGen->update_counters(perm_used);
2656 
2657   if (CMSIncrementalMode) {
2658     icms_update_allocation_limits();
2659   }
2660 
2661   bitMapLock()->unlock();
2662   releaseFreelistLocks();
2663 
2664   _between_prologue_and_epilogue = false;  // ready for next cycle
2665 }
2666 
2667 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2668   collector()->gc_epilogue(full);
2669 
2670   // Also reset promotion tracking in par gc thread states.
2671   if (ParallelGCThreads > 0) {
2672     for (uint i = 0; i < ParallelGCThreads; i++) {
2673       _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2674     }
2675   }
2676 }
2677 
2678 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2679   assert(!incremental_collection_failed(), "Should have been cleared");
2680   cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2681   cmsSpace()->gc_epilogue();
2682     // Print stat counters
2683   NOT_PRODUCT(
2684     assert(_numObjectsAllocated == 0, "check");
2685     assert(_numWordsAllocated == 0, "check");
2686     if (Verbose && PrintGC) {
2687       gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2688                           SIZE_FORMAT" bytes",
2689                  _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2690     }
2691     _numObjectsPromoted = 0;
2692     _numWordsPromoted   = 0;
2693   )
2694 
2695   if (PrintGC && Verbose) {
2696     // Call down the chain in contiguous_available needs the freelistLock
2697     // so print this out before releasing the freeListLock.
2698     gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2699                         contiguous_available());
2700   }
2701 }
2702 
2703 #ifndef PRODUCT
2704 bool CMSCollector::have_cms_token() {
2705   Thread* thr = Thread::current();
2706   if (thr->is_VM_thread()) {
2707     return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2708   } else if (thr->is_ConcurrentGC_thread()) {
2709     return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2710   } else if (thr->is_GC_task_thread()) {
2711     return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2712            ParGCRareEvent_lock->owned_by_self();
2713   }
2714   return false;
2715 }
2716 #endif
2717 
2718 // Check reachability of the given heap address in CMS generation,
2719 // treating all other generations as roots.
2720 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2721   // We could "guarantee" below, rather than assert, but i'll
2722   // leave these as "asserts" so that an adventurous debugger
2723   // could try this in the product build provided some subset of
2724   // the conditions were met, provided they were intersted in the
2725   // results and knew that the computation below wouldn't interfere
2726   // with other concurrent computations mutating the structures
2727   // being read or written.
2728   assert(SafepointSynchronize::is_at_safepoint(),
2729          "Else mutations in object graph will make answer suspect");
2730   assert(have_cms_token(), "Should hold cms token");
2731   assert(haveFreelistLocks(), "must hold free list locks");
2732   assert_lock_strong(bitMapLock());
2733 
2734   // Clear the marking bit map array before starting, but, just
2735   // for kicks, first report if the given address is already marked
2736   gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2737                 _markBitMap.isMarked(addr) ? "" : " not");
2738 
2739   if (verify_after_remark()) {
2740     MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2741     bool result = verification_mark_bm()->isMarked(addr);
2742     gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2743                            result ? "IS" : "is NOT");
2744     return result;
2745   } else {
2746     gclog_or_tty->print_cr("Could not compute result");
2747     return false;
2748   }
2749 }
2750 
2751 ////////////////////////////////////////////////////////
2752 // CMS Verification Support
2753 ////////////////////////////////////////////////////////
2754 // Following the remark phase, the following invariant
2755 // should hold -- each object in the CMS heap which is
2756 // marked in markBitMap() should be marked in the verification_mark_bm().
2757 
2758 class VerifyMarkedClosure: public BitMapClosure {
2759   CMSBitMap* _marks;
2760   bool       _failed;
2761 
2762  public:
2763   VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2764 
2765   bool do_bit(size_t offset) {
2766     HeapWord* addr = _marks->offsetToHeapWord(offset);
2767     if (!_marks->isMarked(addr)) {
2768       oop(addr)->print();
2769       gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2770       _failed = true;
2771     }
2772     return true;
2773   }
2774 
2775   bool failed() { return _failed; }
2776 };
2777 
2778 bool CMSCollector::verify_after_remark() {
2779   gclog_or_tty->print(" [Verifying CMS Marking... ");
2780   MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2781   static bool init = false;
2782 
2783   assert(SafepointSynchronize::is_at_safepoint(),
2784          "Else mutations in object graph will make answer suspect");
2785   assert(have_cms_token(),
2786          "Else there may be mutual interference in use of "
2787          " verification data structures");
2788   assert(_collectorState > Marking && _collectorState <= Sweeping,
2789          "Else marking info checked here may be obsolete");
2790   assert(haveFreelistLocks(), "must hold free list locks");
2791   assert_lock_strong(bitMapLock());
2792 
2793 
2794   // Allocate marking bit map if not already allocated
2795   if (!init) { // first time
2796     if (!verification_mark_bm()->allocate(_span)) {
2797       return false;
2798     }
2799     init = true;
2800   }
2801 
2802   assert(verification_mark_stack()->isEmpty(), "Should be empty");
2803 
2804   // Turn off refs discovery -- so we will be tracing through refs.
2805   // This is as intended, because by this time
2806   // GC must already have cleared any refs that need to be cleared,
2807   // and traced those that need to be marked; moreover,
2808   // the marking done here is not going to intefere in any
2809   // way with the marking information used by GC.
2810   NoRefDiscovery no_discovery(ref_processor());
2811 
2812   COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2813 
2814   // Clear any marks from a previous round
2815   verification_mark_bm()->clear_all();
2816   assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2817   assert(overflow_list_is_empty(), "overflow list should be empty");
2818 
2819   GenCollectedHeap* gch = GenCollectedHeap::heap();
2820   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
2821   // Update the saved marks which may affect the root scans.
2822   gch->save_marks();
2823 
2824   if (CMSRemarkVerifyVariant == 1) {
2825     // In this first variant of verification, we complete
2826     // all marking, then check if the new marks-verctor is
2827     // a subset of the CMS marks-vector.
2828     verify_after_remark_work_1();
2829   } else if (CMSRemarkVerifyVariant == 2) {
2830     // In this second variant of verification, we flag an error
2831     // (i.e. an object reachable in the new marks-vector not reachable
2832     // in the CMS marks-vector) immediately, also indicating the
2833     // identify of an object (A) that references the unmarked object (B) --
2834     // presumably, a mutation to A failed to be picked up by preclean/remark?
2835     verify_after_remark_work_2();
2836   } else {
2837     warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2838             CMSRemarkVerifyVariant);
2839   }
2840   gclog_or_tty->print(" done] ");
2841   return true;
2842 }
2843 
2844 void CMSCollector::verify_after_remark_work_1() {
2845   ResourceMark rm;
2846   HandleMark  hm;
2847   GenCollectedHeap* gch = GenCollectedHeap::heap();
2848 
2849   // Mark from roots one level into CMS
2850   MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */);
2851   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2852 
2853   gch->gen_process_strong_roots(_cmsGen->level(),
2854                                 true,   // younger gens are roots
2855                                 true,   // collecting perm gen
2856                                 SharedHeap::ScanningOption(roots_scanning_options()),
2857                                 NULL, &notOlder);
2858 
2859   // Now mark from the roots
2860   assert(_revisitStack.isEmpty(), "Should be empty");
2861   MarkFromRootsClosure markFromRootsClosure(this, _span,
2862     verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2863     false /* don't yield */, true /* verifying */);
2864   assert(_restart_addr == NULL, "Expected pre-condition");
2865   verification_mark_bm()->iterate(&markFromRootsClosure);
2866   while (_restart_addr != NULL) {
2867     // Deal with stack overflow: by restarting at the indicated
2868     // address.
2869     HeapWord* ra = _restart_addr;
2870     markFromRootsClosure.reset(ra);
2871     _restart_addr = NULL;
2872     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2873   }
2874   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2875   verify_work_stacks_empty();
2876   // Should reset the revisit stack above, since no class tree
2877   // surgery is forthcoming.
2878   _revisitStack.reset(); // throwing away all contents
2879 
2880   // Marking completed -- now verify that each bit marked in
2881   // verification_mark_bm() is also marked in markBitMap(); flag all
2882   // errors by printing corresponding objects.
2883   VerifyMarkedClosure vcl(markBitMap());
2884   verification_mark_bm()->iterate(&vcl);
2885   if (vcl.failed()) {
2886     gclog_or_tty->print("Verification failed");
2887     Universe::heap()->print();
2888     fatal(" ... aborting");
2889   }
2890 }
2891 
2892 void CMSCollector::verify_after_remark_work_2() {
2893   ResourceMark rm;
2894   HandleMark  hm;
2895   GenCollectedHeap* gch = GenCollectedHeap::heap();
2896 
2897   // Mark from roots one level into CMS
2898   MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2899                                      markBitMap(), true /* nmethods */);
2900   gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2901   gch->gen_process_strong_roots(_cmsGen->level(),
2902                                 true,   // younger gens are roots
2903                                 true,   // collecting perm gen
2904                                 SharedHeap::ScanningOption(roots_scanning_options()),
2905                                 NULL, &notOlder);
2906 
2907   // Now mark from the roots
2908   assert(_revisitStack.isEmpty(), "Should be empty");
2909   MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2910     verification_mark_bm(), markBitMap(), verification_mark_stack());
2911   assert(_restart_addr == NULL, "Expected pre-condition");
2912   verification_mark_bm()->iterate(&markFromRootsClosure);
2913   while (_restart_addr != NULL) {
2914     // Deal with stack overflow: by restarting at the indicated
2915     // address.
2916     HeapWord* ra = _restart_addr;
2917     markFromRootsClosure.reset(ra);
2918     _restart_addr = NULL;
2919     verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2920   }
2921   assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2922   verify_work_stacks_empty();
2923   // Should reset the revisit stack above, since no class tree
2924   // surgery is forthcoming.
2925   _revisitStack.reset(); // throwing away all contents
2926 
2927   // Marking completed -- now verify that each bit marked in
2928   // verification_mark_bm() is also marked in markBitMap(); flag all
2929   // errors by printing corresponding objects.
2930   VerifyMarkedClosure vcl(markBitMap());
2931   verification_mark_bm()->iterate(&vcl);
2932   assert(!vcl.failed(), "Else verification above should not have succeeded");
2933 }
2934 
2935 void ConcurrentMarkSweepGeneration::save_marks() {
2936   // delegate to CMS space
2937   cmsSpace()->save_marks();
2938   for (uint i = 0; i < ParallelGCThreads; i++) {
2939     _par_gc_thread_states[i]->promo.startTrackingPromotions();
2940   }
2941 }
2942 
2943 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2944   return cmsSpace()->no_allocs_since_save_marks();
2945 }
2946 
2947 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix)    \
2948                                                                 \
2949 void ConcurrentMarkSweepGeneration::                            \
2950 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) {   \
2951   cl->set_generation(this);                                     \
2952   cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl);      \
2953   cl->reset_generation();                                       \
2954   save_marks();                                                 \
2955 }
2956 
2957 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2958 
2959 void
2960 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
2961 {
2962   // Not currently implemented; need to do the following. -- ysr.
2963   // dld -- I think that is used for some sort of allocation profiler.  So it
2964   // really means the objects allocated by the mutator since the last
2965   // GC.  We could potentially implement this cheaply by recording only
2966   // the direct allocations in a side data structure.
2967   //
2968   // I think we probably ought not to be required to support these
2969   // iterations at any arbitrary point; I think there ought to be some
2970   // call to enable/disable allocation profiling in a generation/space,
2971   // and the iterator ought to return the objects allocated in the
2972   // gen/space since the enable call, or the last iterator call (which
2973   // will probably be at a GC.)  That way, for gens like CM&S that would
2974   // require some extra data structure to support this, we only pay the
2975   // cost when it's in use...
2976   cmsSpace()->object_iterate_since_last_GC(blk);
2977 }
2978 
2979 void
2980 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
2981   cl->set_generation(this);
2982   younger_refs_in_space_iterate(_cmsSpace, cl);
2983   cl->reset_generation();
2984 }
2985 
2986 void
2987 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
2988   if (freelistLock()->owned_by_self()) {
2989     Generation::oop_iterate(mr, cl);
2990   } else {
2991     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2992     Generation::oop_iterate(mr, cl);
2993   }
2994 }
2995 
2996 void
2997 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
2998   if (freelistLock()->owned_by_self()) {
2999     Generation::oop_iterate(cl);
3000   } else {
3001     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3002     Generation::oop_iterate(cl);
3003   }
3004 }
3005 
3006 void
3007 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3008   if (freelistLock()->owned_by_self()) {
3009     Generation::object_iterate(cl);
3010   } else {
3011     MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3012     Generation::object_iterate(cl);
3013   }
3014 }
3015 
3016 void
3017 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3018 }
3019 
3020 void
3021 ConcurrentMarkSweepGeneration::post_compact() {
3022 }
3023 
3024 void
3025 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3026   // Fix the linear allocation blocks to look like free blocks.
3027 
3028   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3029   // are not called when the heap is verified during universe initialization and
3030   // at vm shutdown.
3031   if (freelistLock()->owned_by_self()) {
3032     cmsSpace()->prepare_for_verify();
3033   } else {
3034     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3035     cmsSpace()->prepare_for_verify();
3036   }
3037 }
3038 
3039 void
3040 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3041   // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3042   // are not called when the heap is verified during universe initialization and
3043   // at vm shutdown.
3044   if (freelistLock()->owned_by_self()) {
3045     cmsSpace()->verify(false /* ignored */);
3046   } else {
3047     MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3048     cmsSpace()->verify(false /* ignored */);
3049   }
3050 }
3051 
3052 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3053   _cmsGen->verify(allow_dirty);
3054   _permGen->verify(allow_dirty);
3055 }
3056 
3057 #ifndef PRODUCT
3058 bool CMSCollector::overflow_list_is_empty() const {
3059   assert(_num_par_pushes >= 0, "Inconsistency");
3060   if (_overflow_list == NULL) {
3061     assert(_num_par_pushes == 0, "Inconsistency");
3062   }
3063   return _overflow_list == NULL;
3064 }
3065 
3066 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3067 // merely consolidate assertion checks that appear to occur together frequently.
3068 void CMSCollector::verify_work_stacks_empty() const {
3069   assert(_markStack.isEmpty(), "Marking stack should be empty");
3070   assert(overflow_list_is_empty(), "Overflow list should be empty");
3071 }
3072 
3073 void CMSCollector::verify_overflow_empty() const {
3074   assert(overflow_list_is_empty(), "Overflow list should be empty");
3075   assert(no_preserved_marks(), "No preserved marks");
3076 }
3077 #endif // PRODUCT
3078 
3079 // Decide if we want to enable class unloading as part of the
3080 // ensuing concurrent GC cycle. We will collect the perm gen and
3081 // unload classes if it's the case that:
3082 // (1) an explicit gc request has been made and the flag
3083 //     ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3084 // (2) (a) class unloading is enabled at the command line, and
3085 //     (b) (i)   perm gen threshold has been crossed, or
3086 //         (ii)  old gen is getting really full, or
3087 //         (iii) the previous N CMS collections did not collect the
3088 //               perm gen
3089 // NOTE: Provided there is no change in the state of the heap between
3090 // calls to this method, it should have idempotent results. Moreover,
3091 // its results should be monotonically increasing (i.e. going from 0 to 1,
3092 // but not 1 to 0) between successive calls between which the heap was
3093 // not collected. For the implementation below, it must thus rely on
3094 // the property that concurrent_cycles_since_last_unload()
3095 // will not decrease unless a collection cycle happened and that
3096 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3097 // themselves also monotonic in that sense. See check_monotonicity()
3098 // below.
3099 bool CMSCollector::update_should_unload_classes() {
3100   _should_unload_classes = false;
3101   // Condition 1 above
3102   if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3103     _should_unload_classes = true;
3104   } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3105     // Disjuncts 2.b.(i,ii,iii) above
3106     _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3107                               CMSClassUnloadingMaxInterval)
3108                            || _permGen->should_concurrent_collect()
3109                            || _cmsGen->is_too_full();
3110   }
3111   return _should_unload_classes;
3112 }
3113 
3114 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3115   bool res = should_concurrent_collect();
3116   res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3117   return res;
3118 }
3119 
3120 void CMSCollector::setup_cms_unloading_and_verification_state() {
3121   const  bool should_verify =    VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3122                              || VerifyBeforeExit;
3123   const  int  rso           =    SharedHeap::SO_Symbols | SharedHeap::SO_Strings
3124                              |   SharedHeap::SO_CodeCache;
3125 
3126   if (should_unload_classes()) {   // Should unload classes this cycle
3127     remove_root_scanning_option(rso);  // Shrink the root set appropriately
3128     set_verifying(should_verify);    // Set verification state for this cycle
3129     return;                            // Nothing else needs to be done at this time
3130   }
3131 
3132   // Not unloading classes this cycle
3133   assert(!should_unload_classes(), "Inconsitency!");
3134   if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3135     // We were not verifying, or we _were_ unloading classes in the last cycle,
3136     // AND some verification options are enabled this cycle; in this case,
3137     // we must make sure that the deadness map is allocated if not already so,
3138     // and cleared (if already allocated previously --
3139     // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3140     if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3141       if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3142         warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3143                 "permanent generation verification disabled");
3144         return;  // Note that we leave verification disabled, so we'll retry this
3145                  // allocation next cycle. We _could_ remember this failure
3146                  // and skip further attempts and permanently disable verification
3147                  // attempts if that is considered more desirable.
3148       }
3149       assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3150               "_perm_gen_ver_bit_map inconsistency?");
3151     } else {
3152       perm_gen_verify_bit_map()->clear_all();
3153     }
3154     // Include symbols, strings and code cache elements to prevent their resurrection.
3155     add_root_scanning_option(rso);
3156     set_verifying(true);
3157   } else if (verifying() && !should_verify) {
3158     // We were verifying, but some verification flags got disabled.
3159     set_verifying(false);
3160     // Exclude symbols, strings and code cache elements from root scanning to
3161     // reduce IM and RM pauses.
3162     remove_root_scanning_option(rso);
3163   }
3164 }
3165 
3166 
3167 #ifndef PRODUCT
3168 HeapWord* CMSCollector::block_start(const void* p) const {
3169   const HeapWord* addr = (HeapWord*)p;
3170   if (_span.contains(p)) {
3171     if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3172       return _cmsGen->cmsSpace()->block_start(p);
3173     } else {
3174       assert(_permGen->cmsSpace()->is_in_reserved(addr),
3175              "Inconsistent _span?");
3176       return _permGen->cmsSpace()->block_start(p);
3177     }
3178   }
3179   return NULL;
3180 }
3181 #endif
3182 
3183 HeapWord*
3184 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3185                                                    bool   tlab,
3186                                                    bool   parallel) {
3187   assert(!tlab, "Can't deal with TLAB allocation");
3188   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3189   expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3190     CMSExpansionCause::_satisfy_allocation);
3191   if (GCExpandToAllocateDelayMillis > 0) {
3192     os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3193   }
3194   return have_lock_and_allocate(word_size, tlab);
3195 }
3196 
3197 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3198 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3199 // to CardGeneration and share it...
3200 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3201   return CardGeneration::expand(bytes, expand_bytes);
3202 }
3203 
3204 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3205   CMSExpansionCause::Cause cause)
3206 {
3207 
3208   bool success = expand(bytes, expand_bytes);
3209 
3210   // remember why we expanded; this information is used
3211   // by shouldConcurrentCollect() when making decisions on whether to start
3212   // a new CMS cycle.
3213   if (success) {
3214     set_expansion_cause(cause);
3215     if (PrintGCDetails && Verbose) {
3216       gclog_or_tty->print_cr("Expanded CMS gen for %s",
3217         CMSExpansionCause::to_string(cause));
3218     }
3219   }
3220 }
3221 
3222 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3223   HeapWord* res = NULL;
3224   MutexLocker x(ParGCRareEvent_lock);
3225   while (true) {
3226     // Expansion by some other thread might make alloc OK now:
3227     res = ps->lab.alloc(word_sz);
3228     if (res != NULL) return res;
3229     // If there's not enough expansion space available, give up.
3230     if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3231       return NULL;
3232     }
3233     // Otherwise, we try expansion.
3234     expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3235       CMSExpansionCause::_allocate_par_lab);
3236     // Now go around the loop and try alloc again;
3237     // A competing par_promote might beat us to the expansion space,
3238     // so we may go around the loop again if promotion fails agaion.
3239     if (GCExpandToAllocateDelayMillis > 0) {
3240       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3241     }
3242   }
3243 }
3244 
3245 
3246 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3247   PromotionInfo* promo) {
3248   MutexLocker x(ParGCRareEvent_lock);
3249   size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3250   while (true) {
3251     // Expansion by some other thread might make alloc OK now:
3252     if (promo->ensure_spooling_space()) {
3253       assert(promo->has_spooling_space(),
3254              "Post-condition of successful ensure_spooling_space()");
3255       return true;
3256     }
3257     // If there's not enough expansion space available, give up.
3258     if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3259       return false;
3260     }
3261     // Otherwise, we try expansion.
3262     expand(refill_size_bytes, MinHeapDeltaBytes,
3263       CMSExpansionCause::_allocate_par_spooling_space);
3264     // Now go around the loop and try alloc again;
3265     // A competing allocation might beat us to the expansion space,
3266     // so we may go around the loop again if allocation fails again.
3267     if (GCExpandToAllocateDelayMillis > 0) {
3268       os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3269     }
3270   }
3271 }
3272 
3273 
3274 
3275 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3276   assert_locked_or_safepoint(Heap_lock);
3277   size_t size = ReservedSpace::page_align_size_down(bytes);
3278   if (size > 0) {
3279     shrink_by(size);
3280   }
3281 }
3282 
3283 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3284   assert_locked_or_safepoint(Heap_lock);
3285   bool result = _virtual_space.expand_by(bytes);
3286   if (result) {
3287     HeapWord* old_end = _cmsSpace->end();
3288     size_t new_word_size =
3289       heap_word_size(_virtual_space.committed_size());
3290     MemRegion mr(_cmsSpace->bottom(), new_word_size);
3291     _bts->resize(new_word_size);  // resize the block offset shared array
3292     Universe::heap()->barrier_set()->resize_covered_region(mr);
3293     // Hmmmm... why doesn't CFLS::set_end verify locking?
3294     // This is quite ugly; FIX ME XXX
3295     _cmsSpace->assert_locked();
3296     _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3297 
3298     // update the space and generation capacity counters
3299     if (UsePerfData) {
3300       _space_counters->update_capacity();
3301       _gen_counters->update_all();
3302     }
3303 
3304     if (Verbose && PrintGC) {
3305       size_t new_mem_size = _virtual_space.committed_size();
3306       size_t old_mem_size = new_mem_size - bytes;
3307       gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3308                     name(), old_mem_size/K, bytes/K, new_mem_size/K);
3309     }
3310   }
3311   return result;
3312 }
3313 
3314 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3315   assert_locked_or_safepoint(Heap_lock);
3316   bool success = true;
3317   const size_t remaining_bytes = _virtual_space.uncommitted_size();
3318   if (remaining_bytes > 0) {
3319     success = grow_by(remaining_bytes);
3320     DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3321   }
3322   return success;
3323 }
3324 
3325 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3326   assert_locked_or_safepoint(Heap_lock);
3327   assert_lock_strong(freelistLock());
3328   // XXX Fix when compaction is implemented.
3329   warning("Shrinking of CMS not yet implemented");
3330   return;
3331 }
3332 
3333 
3334 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3335 // phases.
3336 class CMSPhaseAccounting: public StackObj {
3337  public:
3338   CMSPhaseAccounting(CMSCollector *collector,
3339                      const char *phase,
3340                      bool print_cr = true);
3341   ~CMSPhaseAccounting();
3342 
3343  private:
3344   CMSCollector *_collector;
3345   const char *_phase;
3346   elapsedTimer _wallclock;
3347   bool _print_cr;
3348 
3349  public:
3350   // Not MT-safe; so do not pass around these StackObj's
3351   // where they may be accessed by other threads.
3352   jlong wallclock_millis() {
3353     assert(_wallclock.is_active(), "Wall clock should not stop");
3354     _wallclock.stop();  // to record time
3355     jlong ret = _wallclock.milliseconds();
3356     _wallclock.start(); // restart
3357     return ret;
3358   }
3359 };
3360 
3361 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3362                                        const char *phase,
3363                                        bool print_cr) :
3364   _collector(collector), _phase(phase), _print_cr(print_cr) {
3365 
3366   if (PrintCMSStatistics != 0) {
3367     _collector->resetYields();
3368   }
3369   if (PrintGCDetails && PrintGCTimeStamps) {
3370     gclog_or_tty->date_stamp(PrintGCDateStamps);
3371     gclog_or_tty->stamp();
3372     gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3373       _collector->cmsGen()->short_name(), _phase);
3374   }
3375   _collector->resetTimer();
3376   _wallclock.start();
3377   _collector->startTimer();
3378 }
3379 
3380 CMSPhaseAccounting::~CMSPhaseAccounting() {
3381   assert(_wallclock.is_active(), "Wall clock should not have stopped");
3382   _collector->stopTimer();
3383   _wallclock.stop();
3384   if (PrintGCDetails) {
3385     gclog_or_tty->date_stamp(PrintGCDateStamps);
3386     if (PrintGCTimeStamps) {
3387       gclog_or_tty->stamp();
3388       gclog_or_tty->print(": ");
3389     }
3390     gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3391                  _collector->cmsGen()->short_name(),
3392                  _phase, _collector->timerValue(), _wallclock.seconds());
3393     if (_print_cr) {
3394       gclog_or_tty->print_cr("");
3395     }
3396     if (PrintCMSStatistics != 0) {
3397       gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3398                     _collector->yields());
3399     }
3400   }
3401 }
3402 
3403 // CMS work
3404 
3405 // Checkpoint the roots into this generation from outside
3406 // this generation. [Note this initial checkpoint need only
3407 // be approximate -- we'll do a catch up phase subsequently.]
3408 void CMSCollector::checkpointRootsInitial(bool asynch) {
3409   assert(_collectorState == InitialMarking, "Wrong collector state");
3410   check_correct_thread_executing();
3411   ReferenceProcessor* rp = ref_processor();
3412   SpecializationStats::clear();
3413   assert(_restart_addr == NULL, "Control point invariant");
3414   if (asynch) {
3415     // acquire locks for subsequent manipulations
3416     MutexLockerEx x(bitMapLock(),
3417                     Mutex::_no_safepoint_check_flag);
3418     checkpointRootsInitialWork(asynch);
3419     rp->verify_no_references_recorded();
3420     rp->enable_discovery(); // enable ("weak") refs discovery
3421     _collectorState = Marking;
3422   } else {
3423     // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3424     // which recognizes if we are a CMS generation, and doesn't try to turn on
3425     // discovery; verify that they aren't meddling.
3426     assert(!rp->discovery_is_atomic(),
3427            "incorrect setting of discovery predicate");
3428     assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3429            "ref discovery for this generation kind");
3430     // already have locks
3431     checkpointRootsInitialWork(asynch);
3432     rp->enable_discovery(); // now enable ("weak") refs discovery
3433     _collectorState = Marking;
3434   }
3435   SpecializationStats::print();
3436 }
3437 
3438 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3439   assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3440   assert(_collectorState == InitialMarking, "just checking");
3441 
3442   // If there has not been a GC[n-1] since last GC[n] cycle completed,
3443   // precede our marking with a collection of all
3444   // younger generations to keep floating garbage to a minimum.
3445   // XXX: we won't do this for now -- it's an optimization to be done later.
3446 
3447   // already have locks
3448   assert_lock_strong(bitMapLock());
3449   assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3450 
3451   // Setup the verification and class unloading state for this
3452   // CMS collection cycle.
3453   setup_cms_unloading_and_verification_state();
3454 
3455   NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3456     PrintGCDetails && Verbose, true, gclog_or_tty);)
3457   if (UseAdaptiveSizePolicy) {
3458     size_policy()->checkpoint_roots_initial_begin();
3459   }
3460 
3461   // Reset all the PLAB chunk arrays if necessary.
3462   if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3463     reset_survivor_plab_arrays();
3464   }
3465 
3466   ResourceMark rm;
3467   HandleMark  hm;
3468 
3469   FalseClosure falseClosure;
3470   // In the case of a synchronous collection, we will elide the
3471   // remark step, so it's important to catch all the nmethod oops
3472   // in this step; hence the last argument to the constrcutor below.
3473   MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */);
3474   GenCollectedHeap* gch = GenCollectedHeap::heap();
3475 
3476   verify_work_stacks_empty();
3477   verify_overflow_empty();
3478 
3479   gch->ensure_parsability(false);  // fill TLABs, but no need to retire them
3480   // Update the saved marks which may affect the root scans.
3481   gch->save_marks();
3482 
3483   // weak reference processing has not started yet.
3484   ref_processor()->set_enqueuing_is_done(false);
3485 
3486   {
3487     COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3488     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3489     gch->gen_process_strong_roots(_cmsGen->level(),
3490                                   true,   // younger gens are roots
3491                                   true,   // collecting perm gen
3492                                   SharedHeap::ScanningOption(roots_scanning_options()),
3493                                   NULL, &notOlder);
3494   }
3495 
3496   // Clear mod-union table; it will be dirtied in the prologue of
3497   // CMS generation per each younger generation collection.
3498 
3499   assert(_modUnionTable.isAllClear(),
3500        "Was cleared in most recent final checkpoint phase"
3501        " or no bits are set in the gc_prologue before the start of the next "
3502        "subsequent marking phase.");
3503 
3504   // Temporarily disabled, since pre/post-consumption closures don't
3505   // care about precleaned cards
3506   #if 0
3507   {
3508     MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3509                              (HeapWord*)_virtual_space.high());
3510     _ct->ct_bs()->preclean_dirty_cards(mr);
3511   }
3512   #endif
3513 
3514   // Save the end of the used_region of the constituent generations
3515   // to be used to limit the extent of sweep in each generation.
3516   save_sweep_limits();
3517   if (UseAdaptiveSizePolicy) {
3518     size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3519   }
3520   verify_overflow_empty();
3521 }
3522 
3523 bool CMSCollector::markFromRoots(bool asynch) {
3524   // we might be tempted to assert that:
3525   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3526   //        "inconsistent argument?");
3527   // However that wouldn't be right, because it's possible that
3528   // a safepoint is indeed in progress as a younger generation
3529   // stop-the-world GC happens even as we mark in this generation.
3530   assert(_collectorState == Marking, "inconsistent state?");
3531   check_correct_thread_executing();
3532   verify_overflow_empty();
3533 
3534   bool res;
3535   if (asynch) {
3536 
3537     // Start the timers for adaptive size policy for the concurrent phases
3538     // Do it here so that the foreground MS can use the concurrent
3539     // timer since a foreground MS might has the sweep done concurrently
3540     // or STW.
3541     if (UseAdaptiveSizePolicy) {
3542       size_policy()->concurrent_marking_begin();
3543     }
3544 
3545     // Weak ref discovery note: We may be discovering weak
3546     // refs in this generation concurrent (but interleaved) with
3547     // weak ref discovery by a younger generation collector.
3548 
3549     CMSTokenSyncWithLocks ts(true, bitMapLock());
3550     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3551     CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3552     res = markFromRootsWork(asynch);
3553     if (res) {
3554       _collectorState = Precleaning;
3555     } else { // We failed and a foreground collection wants to take over
3556       assert(_foregroundGCIsActive, "internal state inconsistency");
3557       assert(_restart_addr == NULL,  "foreground will restart from scratch");
3558       if (PrintGCDetails) {
3559         gclog_or_tty->print_cr("bailing out to foreground collection");
3560       }
3561     }
3562     if (UseAdaptiveSizePolicy) {
3563       size_policy()->concurrent_marking_end();
3564     }
3565   } else {
3566     assert(SafepointSynchronize::is_at_safepoint(),
3567            "inconsistent with asynch == false");
3568     if (UseAdaptiveSizePolicy) {
3569       size_policy()->ms_collection_marking_begin();
3570     }
3571     // already have locks
3572     res = markFromRootsWork(asynch);
3573     _collectorState = FinalMarking;
3574     if (UseAdaptiveSizePolicy) {
3575       GenCollectedHeap* gch = GenCollectedHeap::heap();
3576       size_policy()->ms_collection_marking_end(gch->gc_cause());
3577     }
3578   }
3579   verify_overflow_empty();
3580   return res;
3581 }
3582 
3583 bool CMSCollector::markFromRootsWork(bool asynch) {
3584   // iterate over marked bits in bit map, doing a full scan and mark
3585   // from these roots using the following algorithm:
3586   // . if oop is to the right of the current scan pointer,
3587   //   mark corresponding bit (we'll process it later)
3588   // . else (oop is to left of current scan pointer)
3589   //   push oop on marking stack
3590   // . drain the marking stack
3591 
3592   // Note that when we do a marking step we need to hold the
3593   // bit map lock -- recall that direct allocation (by mutators)
3594   // and promotion (by younger generation collectors) is also
3595   // marking the bit map. [the so-called allocate live policy.]
3596   // Because the implementation of bit map marking is not
3597   // robust wrt simultaneous marking of bits in the same word,
3598   // we need to make sure that there is no such interference
3599   // between concurrent such updates.
3600 
3601   // already have locks
3602   assert_lock_strong(bitMapLock());
3603 
3604   // Clear the revisit stack, just in case there are any
3605   // obsolete contents from a short-circuited previous CMS cycle.
3606   _revisitStack.reset();
3607   verify_work_stacks_empty();
3608   verify_overflow_empty();
3609   assert(_revisitStack.isEmpty(), "tabula rasa");
3610 
3611   bool result = false;
3612   if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) {
3613     result = do_marking_mt(asynch);
3614   } else {
3615     result = do_marking_st(asynch);
3616   }
3617   return result;
3618 }
3619 
3620 // Forward decl
3621 class CMSConcMarkingTask;
3622 
3623 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3624   CMSCollector*       _collector;
3625   CMSConcMarkingTask* _task;
3626   bool _yield;
3627  protected:
3628   virtual void yield();
3629  public:
3630   // "n_threads" is the number of threads to be terminated.
3631   // "queue_set" is a set of work queues of other threads.
3632   // "collector" is the CMS collector associated with this task terminator.
3633   // "yield" indicates whether we need the gang as a whole to yield.
3634   CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
3635                            CMSCollector* collector, bool yield) :
3636     ParallelTaskTerminator(n_threads, queue_set),
3637     _collector(collector),
3638     _yield(yield) { }
3639 
3640   void set_task(CMSConcMarkingTask* task) {
3641     _task = task;
3642   }
3643 };
3644 
3645 // MT Concurrent Marking Task
3646 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3647   CMSCollector* _collector;
3648   YieldingFlexibleWorkGang* _workers;        // the whole gang
3649   int           _n_workers;                  // requested/desired # workers
3650   bool          _asynch;
3651   bool          _result;
3652   CompactibleFreeListSpace*  _cms_space;
3653   CompactibleFreeListSpace* _perm_space;
3654   HeapWord*     _global_finger;
3655   HeapWord*     _restart_addr;
3656 
3657   //  Exposed here for yielding support
3658   Mutex* const _bit_map_lock;
3659 
3660   // The per thread work queues, available here for stealing
3661   OopTaskQueueSet*  _task_queues;
3662   CMSConcMarkingTerminator _term;
3663 
3664  public:
3665   CMSConcMarkingTask(CMSCollector* collector,
3666                  CompactibleFreeListSpace* cms_space,
3667                  CompactibleFreeListSpace* perm_space,
3668                  bool asynch, int n_workers,
3669                  YieldingFlexibleWorkGang* workers,
3670                  OopTaskQueueSet* task_queues):
3671     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3672     _collector(collector),
3673     _cms_space(cms_space),
3674     _perm_space(perm_space),
3675     _asynch(asynch), _n_workers(n_workers), _result(true),
3676     _workers(workers), _task_queues(task_queues),
3677     _term(n_workers, task_queues, _collector, asynch),
3678     _bit_map_lock(collector->bitMapLock())
3679   {
3680     assert(n_workers <= workers->total_workers(),
3681            "Else termination won't work correctly today"); // XXX FIX ME!
3682     _requested_size = n_workers;
3683     _term.set_task(this);
3684     assert(_cms_space->bottom() < _perm_space->bottom(),
3685            "Finger incorrectly initialized below");
3686     _restart_addr = _global_finger = _cms_space->bottom();
3687   }
3688 
3689 
3690   OopTaskQueueSet* task_queues()  { return _task_queues; }
3691 
3692   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3693 
3694   HeapWord** global_finger_addr() { return &_global_finger; }
3695 
3696   CMSConcMarkingTerminator* terminator() { return &_term; }
3697 
3698   void work(int i);
3699 
3700   virtual void coordinator_yield();  // stuff done by coordinator
3701   bool result() { return _result; }
3702 
3703   void reset(HeapWord* ra) {
3704     assert(_global_finger >= _cms_space->end(),  "Postcondition of ::work(i)");
3705     assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
3706     assert(ra             <  _perm_space->end(), "ra too large");
3707     _restart_addr = _global_finger = ra;
3708     _term.reset_for_reuse();
3709   }
3710 
3711   static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3712                                            OopTaskQueue* work_q);
3713 
3714  private:
3715   void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3716   void do_work_steal(int i);
3717   void bump_global_finger(HeapWord* f);
3718 };
3719 
3720 void CMSConcMarkingTerminator::yield() {
3721   if (ConcurrentMarkSweepThread::should_yield() &&
3722       !_collector->foregroundGCIsActive() &&
3723       _yield) {
3724     _task->yield();
3725   } else {
3726     ParallelTaskTerminator::yield();
3727   }
3728 }
3729 
3730 ////////////////////////////////////////////////////////////////
3731 // Concurrent Marking Algorithm Sketch
3732 ////////////////////////////////////////////////////////////////
3733 // Until all tasks exhausted (both spaces):
3734 // -- claim next available chunk
3735 // -- bump global finger via CAS
3736 // -- find first object that starts in this chunk
3737 //    and start scanning bitmap from that position
3738 // -- scan marked objects for oops
3739 // -- CAS-mark target, and if successful:
3740 //    . if target oop is above global finger (volatile read)
3741 //      nothing to do
3742 //    . if target oop is in chunk and above local finger
3743 //        then nothing to do
3744 //    . else push on work-queue
3745 // -- Deal with possible overflow issues:
3746 //    . local work-queue overflow causes stuff to be pushed on
3747 //      global (common) overflow queue
3748 //    . always first empty local work queue
3749 //    . then get a batch of oops from global work queue if any
3750 //    . then do work stealing
3751 // -- When all tasks claimed (both spaces)
3752 //    and local work queue empty,
3753 //    then in a loop do:
3754 //    . check global overflow stack; steal a batch of oops and trace
3755 //    . try to steal from other threads oif GOS is empty
3756 //    . if neither is available, offer termination
3757 // -- Terminate and return result
3758 //
3759 void CMSConcMarkingTask::work(int i) {
3760   elapsedTimer _timer;
3761   ResourceMark rm;
3762   HandleMark hm;
3763 
3764   DEBUG_ONLY(_collector->verify_overflow_empty();)
3765 
3766   // Before we begin work, our work queue should be empty
3767   assert(work_queue(i)->size() == 0, "Expected to be empty");
3768   // Scan the bitmap covering _cms_space, tracing through grey objects.
3769   _timer.start();
3770   do_scan_and_mark(i, _cms_space);
3771   _timer.stop();
3772   if (PrintCMSStatistics != 0) {
3773     gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3774       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3775   }
3776 
3777   // ... do the same for the _perm_space
3778   _timer.reset();
3779   _timer.start();
3780   do_scan_and_mark(i, _perm_space);
3781   _timer.stop();
3782   if (PrintCMSStatistics != 0) {
3783     gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3784       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3785   }
3786 
3787   // ... do work stealing
3788   _timer.reset();
3789   _timer.start();
3790   do_work_steal(i);
3791   _timer.stop();
3792   if (PrintCMSStatistics != 0) {
3793     gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3794       i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3795   }
3796   assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3797   assert(work_queue(i)->size() == 0, "Should have been emptied");
3798   // Note that under the current task protocol, the
3799   // following assertion is true even of the spaces
3800   // expanded since the completion of the concurrent
3801   // marking. XXX This will likely change under a strict
3802   // ABORT semantics.
3803   assert(_global_finger >  _cms_space->end() &&
3804          _global_finger >= _perm_space->end(),
3805          "All tasks have been completed");
3806   DEBUG_ONLY(_collector->verify_overflow_empty();)
3807 }
3808 
3809 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3810   HeapWord* read = _global_finger;
3811   HeapWord* cur  = read;
3812   while (f > read) {
3813     cur = read;
3814     read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3815     if (cur == read) {
3816       // our cas succeeded
3817       assert(_global_finger >= f, "protocol consistency");
3818       break;
3819     }
3820   }
3821 }
3822 
3823 // This is really inefficient, and should be redone by
3824 // using (not yet available) block-read and -write interfaces to the
3825 // stack and the work_queue. XXX FIX ME !!!
3826 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3827                                                       OopTaskQueue* work_q) {
3828   // Fast lock-free check
3829   if (ovflw_stk->length() == 0) {
3830     return false;
3831   }
3832   assert(work_q->size() == 0, "Shouldn't steal");
3833   MutexLockerEx ml(ovflw_stk->par_lock(),
3834                    Mutex::_no_safepoint_check_flag);
3835   // Grab up to 1/4 the size of the work queue
3836   size_t num = MIN2((size_t)work_q->max_elems()/4,
3837                     (size_t)ParGCDesiredObjsFromOverflowList);
3838   num = MIN2(num, ovflw_stk->length());
3839   for (int i = (int) num; i > 0; i--) {
3840     oop cur = ovflw_stk->pop();
3841     assert(cur != NULL, "Counted wrong?");
3842     work_q->push(cur);
3843   }
3844   return num > 0;
3845 }
3846 
3847 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3848   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3849   int n_tasks = pst->n_tasks();
3850   // We allow that there may be no tasks to do here because
3851   // we are restarting after a stack overflow.
3852   assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3853   int nth_task = 0;
3854 
3855   HeapWord* aligned_start = sp->bottom();
3856   if (sp->used_region().contains(_restart_addr)) {
3857     // Align down to a card boundary for the start of 0th task
3858     // for this space.
3859     aligned_start =
3860       (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3861                                  CardTableModRefBS::card_size);
3862   }
3863 
3864   size_t chunk_size = sp->marking_task_size();
3865   while (!pst->is_task_claimed(/* reference */ nth_task)) {
3866     // Having claimed the nth task in this space,
3867     // compute the chunk that it corresponds to:
3868     MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3869                                aligned_start + (nth_task+1)*chunk_size);
3870     // Try and bump the global finger via a CAS;
3871     // note that we need to do the global finger bump
3872     // _before_ taking the intersection below, because
3873     // the task corresponding to that region will be
3874     // deemed done even if the used_region() expands
3875     // because of allocation -- as it almost certainly will
3876     // during start-up while the threads yield in the
3877     // closure below.
3878     HeapWord* finger = span.end();
3879     bump_global_finger(finger);   // atomically
3880     // There are null tasks here corresponding to chunks
3881     // beyond the "top" address of the space.
3882     span = span.intersection(sp->used_region());
3883     if (!span.is_empty()) {  // Non-null task
3884       HeapWord* prev_obj;
3885       assert(!span.contains(_restart_addr) || nth_task == 0,
3886              "Inconsistency");
3887       if (nth_task == 0) {
3888         // For the 0th task, we'll not need to compute a block_start.
3889         if (span.contains(_restart_addr)) {
3890           // In the case of a restart because of stack overflow,
3891           // we might additionally skip a chunk prefix.
3892           prev_obj = _restart_addr;
3893         } else {
3894           prev_obj = span.start();
3895         }
3896       } else {
3897         // We want to skip the first object because
3898         // the protocol is to scan any object in its entirety
3899         // that _starts_ in this span; a fortiori, any
3900         // object starting in an earlier span is scanned
3901         // as part of an earlier claimed task.
3902         // Below we use the "careful" version of block_start
3903         // so we do not try to navigate uninitialized objects.
3904         prev_obj = sp->block_start_careful(span.start());
3905         // Below we use a variant of block_size that uses the
3906         // Printezis bits to avoid waiting for allocated
3907         // objects to become initialized/parsable.
3908         while (prev_obj < span.start()) {
3909           size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3910           if (sz > 0) {
3911             prev_obj += sz;
3912           } else {
3913             // In this case we may end up doing a bit of redundant
3914             // scanning, but that appears unavoidable, short of
3915             // locking the free list locks; see bug 6324141.
3916             break;
3917           }
3918         }
3919       }
3920       if (prev_obj < span.end()) {
3921         MemRegion my_span = MemRegion(prev_obj, span.end());
3922         // Do the marking work within a non-empty span --
3923         // the last argument to the constructor indicates whether the
3924         // iteration should be incremental with periodic yields.
3925         Par_MarkFromRootsClosure cl(this, _collector, my_span,
3926                                     &_collector->_markBitMap,
3927                                     work_queue(i),
3928                                     &_collector->_markStack,
3929                                     &_collector->_revisitStack,
3930                                     _asynch);
3931         _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3932       } // else nothing to do for this task
3933     }   // else nothing to do for this task
3934   }
3935   // We'd be tempted to assert here that since there are no
3936   // more tasks left to claim in this space, the global_finger
3937   // must exceed space->top() and a fortiori space->end(). However,
3938   // that would not quite be correct because the bumping of
3939   // global_finger occurs strictly after the claiming of a task,
3940   // so by the time we reach here the global finger may not yet
3941   // have been bumped up by the thread that claimed the last
3942   // task.
3943   pst->all_tasks_completed();
3944 }
3945 
3946 class Par_ConcMarkingClosure: public OopClosure {
3947  private:
3948   CMSCollector* _collector;
3949   MemRegion     _span;
3950   CMSBitMap*    _bit_map;
3951   CMSMarkStack* _overflow_stack;
3952   CMSMarkStack* _revisit_stack;     // XXXXXX Check proper use
3953   OopTaskQueue* _work_queue;
3954  protected:
3955   DO_OOP_WORK_DEFN
3956  public:
3957   Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
3958                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3959     _collector(collector),
3960     _span(_collector->_span),
3961     _work_queue(work_queue),
3962     _bit_map(bit_map),
3963     _overflow_stack(overflow_stack) { }   // need to initialize revisit stack etc.
3964   virtual void do_oop(oop* p);
3965   virtual void do_oop(narrowOop* p);
3966   void trim_queue(size_t max);
3967   void handle_stack_overflow(HeapWord* lost);
3968 };
3969 
3970 // Grey object scanning during work stealing phase --
3971 // the salient assumption here is that any references
3972 // that are in these stolen objects being scanned must
3973 // already have been initialized (else they would not have
3974 // been published), so we do not need to check for
3975 // uninitialized objects before pushing here.
3976 void Par_ConcMarkingClosure::do_oop(oop obj) {
3977   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
3978   HeapWord* addr = (HeapWord*)obj;
3979   // Check if oop points into the CMS generation
3980   // and is not marked
3981   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3982     // a white object ...
3983     // If we manage to "claim" the object, by being the
3984     // first thread to mark it, then we push it on our
3985     // marking stack
3986     if (_bit_map->par_mark(addr)) {     // ... now grey
3987       // push on work queue (grey set)
3988       bool simulate_overflow = false;
3989       NOT_PRODUCT(
3990         if (CMSMarkStackOverflowALot &&
3991             _collector->simulate_overflow()) {
3992           // simulate a stack overflow
3993           simulate_overflow = true;
3994         }
3995       )
3996       if (simulate_overflow ||
3997           !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3998         // stack overflow
3999         if (PrintCMSStatistics != 0) {
4000           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4001                                  SIZE_FORMAT, _overflow_stack->capacity());
4002         }
4003         // We cannot assert that the overflow stack is full because
4004         // it may have been emptied since.
4005         assert(simulate_overflow ||
4006                _work_queue->size() == _work_queue->max_elems(),
4007               "Else push should have succeeded");
4008         handle_stack_overflow(addr);
4009       }
4010     } // Else, some other thread got there first
4011   }
4012 }
4013 
4014 void Par_ConcMarkingClosure::do_oop(oop* p)       { Par_ConcMarkingClosure::do_oop_work(p); }
4015 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4016 
4017 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4018   while (_work_queue->size() > max) {
4019     oop new_oop;
4020     if (_work_queue->pop_local(new_oop)) {
4021       assert(new_oop->is_oop(), "Should be an oop");
4022       assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4023       assert(_span.contains((HeapWord*)new_oop), "Not in span");
4024       assert(new_oop->is_parsable(), "Should be parsable");
4025       new_oop->oop_iterate(this);  // do_oop() above
4026     }
4027   }
4028 }
4029 
4030 // Upon stack overflow, we discard (part of) the stack,
4031 // remembering the least address amongst those discarded
4032 // in CMSCollector's _restart_address.
4033 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4034   // We need to do this under a mutex to prevent other
4035   // workers from interfering with the work done below.
4036   MutexLockerEx ml(_overflow_stack->par_lock(),
4037                    Mutex::_no_safepoint_check_flag);
4038   // Remember the least grey address discarded
4039   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4040   _collector->lower_restart_addr(ra);
4041   _overflow_stack->reset();  // discard stack contents
4042   _overflow_stack->expand(); // expand the stack if possible
4043 }
4044 
4045 
4046 void CMSConcMarkingTask::do_work_steal(int i) {
4047   OopTaskQueue* work_q = work_queue(i);
4048   oop obj_to_scan;
4049   CMSBitMap* bm = &(_collector->_markBitMap);
4050   CMSMarkStack* ovflw = &(_collector->_markStack);
4051   int* seed = _collector->hash_seed(i);
4052   Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw);
4053   while (true) {
4054     cl.trim_queue(0);
4055     assert(work_q->size() == 0, "Should have been emptied above");
4056     if (get_work_from_overflow_stack(ovflw, work_q)) {
4057       // Can't assert below because the work obtained from the
4058       // overflow stack may already have been stolen from us.
4059       // assert(work_q->size() > 0, "Work from overflow stack");
4060       continue;
4061     } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4062       assert(obj_to_scan->is_oop(), "Should be an oop");
4063       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4064       obj_to_scan->oop_iterate(&cl);
4065     } else if (terminator()->offer_termination()) {
4066       assert(work_q->size() == 0, "Impossible!");
4067       break;
4068     }
4069   }
4070 }
4071 
4072 // This is run by the CMS (coordinator) thread.
4073 void CMSConcMarkingTask::coordinator_yield() {
4074   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4075          "CMS thread should hold CMS token");
4076 
4077   // First give up the locks, then yield, then re-lock
4078   // We should probably use a constructor/destructor idiom to
4079   // do this unlock/lock or modify the MutexUnlocker class to
4080   // serve our purpose. XXX
4081   assert_lock_strong(_bit_map_lock);
4082   _bit_map_lock->unlock();
4083   ConcurrentMarkSweepThread::desynchronize(true);
4084   ConcurrentMarkSweepThread::acknowledge_yield_request();
4085   _collector->stopTimer();
4086   if (PrintCMSStatistics != 0) {
4087     _collector->incrementYields();
4088   }
4089   _collector->icms_wait();
4090 
4091   // It is possible for whichever thread initiated the yield request
4092   // not to get a chance to wake up and take the bitmap lock between
4093   // this thread releasing it and reacquiring it. So, while the
4094   // should_yield() flag is on, let's sleep for a bit to give the
4095   // other thread a chance to wake up. The limit imposed on the number
4096   // of iterations is defensive, to avoid any unforseen circumstances
4097   // putting us into an infinite loop. Since it's always been this
4098   // (coordinator_yield()) method that was observed to cause the
4099   // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4100   // which is by default non-zero. For the other seven methods that
4101   // also perform the yield operation, as are using a different
4102   // parameter (CMSYieldSleepCount) which is by default zero. This way we
4103   // can enable the sleeping for those methods too, if necessary.
4104   // See 6442774.
4105   //
4106   // We really need to reconsider the synchronization between the GC
4107   // thread and the yield-requesting threads in the future and we
4108   // should really use wait/notify, which is the recommended
4109   // way of doing this type of interaction. Additionally, we should
4110   // consolidate the eight methods that do the yield operation and they
4111   // are almost identical into one for better maintenability and
4112   // readability. See 6445193.
4113   //
4114   // Tony 2006.06.29
4115   for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4116                    ConcurrentMarkSweepThread::should_yield() &&
4117                    !CMSCollector::foregroundGCIsActive(); ++i) {
4118     os::sleep(Thread::current(), 1, false);
4119     ConcurrentMarkSweepThread::acknowledge_yield_request();
4120   }
4121 
4122   ConcurrentMarkSweepThread::synchronize(true);
4123   _bit_map_lock->lock_without_safepoint_check();
4124   _collector->startTimer();
4125 }
4126 
4127 bool CMSCollector::do_marking_mt(bool asynch) {
4128   assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition");
4129   // In the future this would be determined ergonomically, based
4130   // on #cpu's, # active mutator threads (and load), and mutation rate.
4131   int num_workers = ParallelCMSThreads;
4132 
4133   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
4134   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4135 
4136   CMSConcMarkingTask tsk(this, cms_space, perm_space,
4137                          asynch, num_workers /* number requested XXX */,
4138                          conc_workers(), task_queues());
4139 
4140   // Since the actual number of workers we get may be different
4141   // from the number we requested above, do we need to do anything different
4142   // below? In particular, may be we need to subclass the SequantialSubTasksDone
4143   // class?? XXX
4144   cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4145   perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4146 
4147   // Refs discovery is already non-atomic.
4148   assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4149   // Mutate the Refs discovery so it is MT during the
4150   // multi-threaded marking phase.
4151   ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4152 
4153   conc_workers()->start_task(&tsk);
4154   while (tsk.yielded()) {
4155     tsk.coordinator_yield();
4156     conc_workers()->continue_task(&tsk);
4157   }
4158   // If the task was aborted, _restart_addr will be non-NULL
4159   assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4160   while (_restart_addr != NULL) {
4161     // XXX For now we do not make use of ABORTED state and have not
4162     // yet implemented the right abort semantics (even in the original
4163     // single-threaded CMS case). That needs some more investigation
4164     // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4165     assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4166     // If _restart_addr is non-NULL, a marking stack overflow
4167     // occured; we need to do a fresh marking iteration from the
4168     // indicated restart address.
4169     if (_foregroundGCIsActive && asynch) {
4170       // We may be running into repeated stack overflows, having
4171       // reached the limit of the stack size, while making very
4172       // slow forward progress. It may be best to bail out and
4173       // let the foreground collector do its job.
4174       // Clear _restart_addr, so that foreground GC
4175       // works from scratch. This avoids the headache of
4176       // a "rescan" which would otherwise be needed because
4177       // of the dirty mod union table & card table.
4178       _restart_addr = NULL;
4179       return false;
4180     }
4181     // Adjust the task to restart from _restart_addr
4182     tsk.reset(_restart_addr);
4183     cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4184                   _restart_addr);
4185     perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4186                   _restart_addr);
4187     _restart_addr = NULL;
4188     // Get the workers going again
4189     conc_workers()->start_task(&tsk);
4190     while (tsk.yielded()) {
4191       tsk.coordinator_yield();
4192       conc_workers()->continue_task(&tsk);
4193     }
4194   }
4195   assert(tsk.completed(), "Inconsistency");
4196   assert(tsk.result() == true, "Inconsistency");
4197   return true;
4198 }
4199 
4200 bool CMSCollector::do_marking_st(bool asynch) {
4201   ResourceMark rm;
4202   HandleMark   hm;
4203 
4204   MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4205     &_markStack, &_revisitStack, CMSYield && asynch);
4206   // the last argument to iterate indicates whether the iteration
4207   // should be incremental with periodic yields.
4208   _markBitMap.iterate(&markFromRootsClosure);
4209   // If _restart_addr is non-NULL, a marking stack overflow
4210   // occured; we need to do a fresh iteration from the
4211   // indicated restart address.
4212   while (_restart_addr != NULL) {
4213     if (_foregroundGCIsActive && asynch) {
4214       // We may be running into repeated stack overflows, having
4215       // reached the limit of the stack size, while making very
4216       // slow forward progress. It may be best to bail out and
4217       // let the foreground collector do its job.
4218       // Clear _restart_addr, so that foreground GC
4219       // works from scratch. This avoids the headache of
4220       // a "rescan" which would otherwise be needed because
4221       // of the dirty mod union table & card table.
4222       _restart_addr = NULL;
4223       return false;  // indicating failure to complete marking
4224     }
4225     // Deal with stack overflow:
4226     // we restart marking from _restart_addr
4227     HeapWord* ra = _restart_addr;
4228     markFromRootsClosure.reset(ra);
4229     _restart_addr = NULL;
4230     _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4231   }
4232   return true;
4233 }
4234 
4235 void CMSCollector::preclean() {
4236   check_correct_thread_executing();
4237   assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4238   verify_work_stacks_empty();
4239   verify_overflow_empty();
4240   _abort_preclean = false;
4241   if (CMSPrecleaningEnabled) {
4242     _eden_chunk_index = 0;
4243     size_t used = get_eden_used();
4244     size_t capacity = get_eden_capacity();
4245     // Don't start sampling unless we will get sufficiently
4246     // many samples.
4247     if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4248                 * CMSScheduleRemarkEdenPenetration)) {
4249       _start_sampling = true;
4250     } else {
4251       _start_sampling = false;
4252     }
4253     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4254     CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4255     preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4256   }
4257   CMSTokenSync x(true); // is cms thread
4258   if (CMSPrecleaningEnabled) {
4259     sample_eden();
4260     _collectorState = AbortablePreclean;
4261   } else {
4262     _collectorState = FinalMarking;
4263   }
4264   verify_work_stacks_empty();
4265   verify_overflow_empty();
4266 }
4267 
4268 // Try and schedule the remark such that young gen
4269 // occupancy is CMSScheduleRemarkEdenPenetration %.
4270 void CMSCollector::abortable_preclean() {
4271   check_correct_thread_executing();
4272   assert(CMSPrecleaningEnabled,  "Inconsistent control state");
4273   assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4274 
4275   // If Eden's current occupancy is below this threshold,
4276   // immediately schedule the remark; else preclean
4277   // past the next scavenge in an effort to
4278   // schedule the pause as described avove. By choosing
4279   // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4280   // we will never do an actual abortable preclean cycle.
4281   if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4282     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4283     CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4284     // We need more smarts in the abortable preclean
4285     // loop below to deal with cases where allocation
4286     // in young gen is very very slow, and our precleaning
4287     // is running a losing race against a horde of
4288     // mutators intent on flooding us with CMS updates
4289     // (dirty cards).
4290     // One, admittedly dumb, strategy is to give up
4291     // after a certain number of abortable precleaning loops
4292     // or after a certain maximum time. We want to make
4293     // this smarter in the next iteration.
4294     // XXX FIX ME!!! YSR
4295     size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4296     while (!(should_abort_preclean() ||
4297              ConcurrentMarkSweepThread::should_terminate())) {
4298       workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4299       cumworkdone += workdone;
4300       loops++;
4301       // Voluntarily terminate abortable preclean phase if we have
4302       // been at it for too long.
4303       if ((CMSMaxAbortablePrecleanLoops != 0) &&
4304           loops >= CMSMaxAbortablePrecleanLoops) {
4305         if (PrintGCDetails) {
4306           gclog_or_tty->print(" CMS: abort preclean due to loops ");
4307         }
4308         break;
4309       }
4310       if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4311         if (PrintGCDetails) {
4312           gclog_or_tty->print(" CMS: abort preclean due to time ");
4313         }
4314         break;
4315       }
4316       // If we are doing little work each iteration, we should
4317       // take a short break.
4318       if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4319         // Sleep for some time, waiting for work to accumulate
4320         stopTimer();
4321         cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4322         startTimer();
4323         waited++;
4324       }
4325     }
4326     if (PrintCMSStatistics > 0) {
4327       gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4328                           loops, waited, cumworkdone);
4329     }
4330   }
4331   CMSTokenSync x(true); // is cms thread
4332   if (_collectorState != Idling) {
4333     assert(_collectorState == AbortablePreclean,
4334            "Spontaneous state transition?");
4335     _collectorState = FinalMarking;
4336   } // Else, a foreground collection completed this CMS cycle.
4337   return;
4338 }
4339 
4340 // Respond to an Eden sampling opportunity
4341 void CMSCollector::sample_eden() {
4342   // Make sure a young gc cannot sneak in between our
4343   // reading and recording of a sample.
4344   assert(Thread::current()->is_ConcurrentGC_thread(),
4345          "Only the cms thread may collect Eden samples");
4346   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4347          "Should collect samples while holding CMS token");
4348   if (!_start_sampling) {
4349     return;
4350   }
4351   if (_eden_chunk_array) {
4352     if (_eden_chunk_index < _eden_chunk_capacity) {
4353       _eden_chunk_array[_eden_chunk_index] = *_top_addr;   // take sample
4354       assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4355              "Unexpected state of Eden");
4356       // We'd like to check that what we just sampled is an oop-start address;
4357       // however, we cannot do that here since the object may not yet have been
4358       // initialized. So we'll instead do the check when we _use_ this sample
4359       // later.
4360       if (_eden_chunk_index == 0 ||
4361           (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4362                          _eden_chunk_array[_eden_chunk_index-1])
4363            >= CMSSamplingGrain)) {
4364         _eden_chunk_index++;  // commit sample
4365       }
4366     }
4367   }
4368   if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4369     size_t used = get_eden_used();
4370     size_t capacity = get_eden_capacity();
4371     assert(used <= capacity, "Unexpected state of Eden");
4372     if (used >  (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4373       _abort_preclean = true;
4374     }
4375   }
4376 }
4377 
4378 
4379 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4380   assert(_collectorState == Precleaning ||
4381          _collectorState == AbortablePreclean, "incorrect state");
4382   ResourceMark rm;
4383   HandleMark   hm;
4384   // Do one pass of scrubbing the discovered reference lists
4385   // to remove any reference objects with strongly-reachable
4386   // referents.
4387   if (clean_refs) {
4388     ReferenceProcessor* rp = ref_processor();
4389     CMSPrecleanRefsYieldClosure yield_cl(this);
4390     assert(rp->span().equals(_span), "Spans should be equal");
4391     CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4392                                    &_markStack, true /* preclean */);
4393     CMSDrainMarkingStackClosure complete_trace(this,
4394                                    _span, &_markBitMap, &_markStack,
4395                                    &keep_alive, true /* preclean */);
4396 
4397     // We don't want this step to interfere with a young
4398     // collection because we don't want to take CPU
4399     // or memory bandwidth away from the young GC threads
4400     // (which may be as many as there are CPUs).
4401     // Note that we don't need to protect ourselves from
4402     // interference with mutators because they can't
4403     // manipulate the discovered reference lists nor affect
4404     // the computed reachability of the referents, the
4405     // only properties manipulated by the precleaning
4406     // of these reference lists.
4407     stopTimer();
4408     CMSTokenSyncWithLocks x(true /* is cms thread */,
4409                             bitMapLock());
4410     startTimer();
4411     sample_eden();
4412     // The following will yield to allow foreground
4413     // collection to proceed promptly. XXX YSR:
4414     // The code in this method may need further
4415     // tweaking for better performance and some restructuring
4416     // for cleaner interfaces.
4417     rp->preclean_discovered_references(
4418           rp->is_alive_non_header(), &keep_alive, &complete_trace,
4419           &yield_cl);
4420   }
4421 
4422   if (clean_survivor) {  // preclean the active survivor space(s)
4423     assert(_young_gen->kind() == Generation::DefNew ||
4424            _young_gen->kind() == Generation::ParNew ||
4425            _young_gen->kind() == Generation::ASParNew,
4426          "incorrect type for cast");
4427     DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4428     PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4429                              &_markBitMap, &_modUnionTable,
4430                              &_markStack, &_revisitStack,
4431                              true /* precleaning phase */);
4432     stopTimer();
4433     CMSTokenSyncWithLocks ts(true /* is cms thread */,
4434                              bitMapLock());
4435     startTimer();
4436     unsigned int before_count =
4437       GenCollectedHeap::heap()->total_collections();
4438     SurvivorSpacePrecleanClosure
4439       sss_cl(this, _span, &_markBitMap, &_markStack,
4440              &pam_cl, before_count, CMSYield);
4441     dng->from()->object_iterate_careful(&sss_cl);
4442     dng->to()->object_iterate_careful(&sss_cl);
4443   }
4444   MarkRefsIntoAndScanClosure
4445     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4446              &_markStack, &_revisitStack, this, CMSYield,
4447              true /* precleaning phase */);
4448   // CAUTION: The following closure has persistent state that may need to
4449   // be reset upon a decrease in the sequence of addresses it
4450   // processes.
4451   ScanMarkedObjectsAgainCarefullyClosure
4452     smoac_cl(this, _span,
4453       &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4454 
4455   // Preclean dirty cards in ModUnionTable and CardTable using
4456   // appropriate convergence criterion;
4457   // repeat CMSPrecleanIter times unless we find that
4458   // we are losing.
4459   assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4460   assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4461          "Bad convergence multiplier");
4462   assert(CMSPrecleanThreshold >= 100,
4463          "Unreasonably low CMSPrecleanThreshold");
4464 
4465   size_t numIter, cumNumCards, lastNumCards, curNumCards;
4466   for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4467        numIter < CMSPrecleanIter;
4468        numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4469     curNumCards  = preclean_mod_union_table(_cmsGen, &smoac_cl);
4470     if (CMSPermGenPrecleaningEnabled) {
4471       curNumCards  += preclean_mod_union_table(_permGen, &smoac_cl);
4472     }
4473     if (Verbose && PrintGCDetails) {
4474       gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4475     }
4476     // Either there are very few dirty cards, so re-mark
4477     // pause will be small anyway, or our pre-cleaning isn't
4478     // that much faster than the rate at which cards are being
4479     // dirtied, so we might as well stop and re-mark since
4480     // precleaning won't improve our re-mark time by much.
4481     if (curNumCards <= CMSPrecleanThreshold ||
4482         (numIter > 0 &&
4483          (curNumCards * CMSPrecleanDenominator >
4484          lastNumCards * CMSPrecleanNumerator))) {
4485       numIter++;
4486       cumNumCards += curNumCards;
4487       break;
4488     }
4489   }
4490   curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4491   if (CMSPermGenPrecleaningEnabled) {
4492     curNumCards += preclean_card_table(_permGen, &smoac_cl);
4493   }
4494   cumNumCards += curNumCards;
4495   if (PrintGCDetails && PrintCMSStatistics != 0) {
4496     gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4497                   curNumCards, cumNumCards, numIter);
4498   }
4499   return cumNumCards;   // as a measure of useful work done
4500 }
4501 
4502 // PRECLEANING NOTES:
4503 // Precleaning involves:
4504 // . reading the bits of the modUnionTable and clearing the set bits.
4505 // . For the cards corresponding to the set bits, we scan the
4506 //   objects on those cards. This means we need the free_list_lock
4507 //   so that we can safely iterate over the CMS space when scanning
4508 //   for oops.
4509 // . When we scan the objects, we'll be both reading and setting
4510 //   marks in the marking bit map, so we'll need the marking bit map.
4511 // . For protecting _collector_state transitions, we take the CGC_lock.
4512 //   Note that any races in the reading of of card table entries by the
4513 //   CMS thread on the one hand and the clearing of those entries by the
4514 //   VM thread or the setting of those entries by the mutator threads on the
4515 //   other are quite benign. However, for efficiency it makes sense to keep
4516 //   the VM thread from racing with the CMS thread while the latter is
4517 //   dirty card info to the modUnionTable. We therefore also use the
4518 //   CGC_lock to protect the reading of the card table and the mod union
4519 //   table by the CM thread.
4520 // . We run concurrently with mutator updates, so scanning
4521 //   needs to be done carefully  -- we should not try to scan
4522 //   potentially uninitialized objects.
4523 //
4524 // Locking strategy: While holding the CGC_lock, we scan over and
4525 // reset a maximal dirty range of the mod union / card tables, then lock
4526 // the free_list_lock and bitmap lock to do a full marking, then
4527 // release these locks; and repeat the cycle. This allows for a
4528 // certain amount of fairness in the sharing of these locks between
4529 // the CMS collector on the one hand, and the VM thread and the
4530 // mutators on the other.
4531 
4532 // NOTE: preclean_mod_union_table() and preclean_card_table()
4533 // further below are largely identical; if you need to modify
4534 // one of these methods, please check the other method too.
4535 
4536 size_t CMSCollector::preclean_mod_union_table(
4537   ConcurrentMarkSweepGeneration* gen,
4538   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4539   verify_work_stacks_empty();
4540   verify_overflow_empty();
4541 
4542   // strategy: starting with the first card, accumulate contiguous
4543   // ranges of dirty cards; clear these cards, then scan the region
4544   // covered by these cards.
4545 
4546   // Since all of the MUT is committed ahead, we can just use
4547   // that, in case the generations expand while we are precleaning.
4548   // It might also be fine to just use the committed part of the
4549   // generation, but we might potentially miss cards when the
4550   // generation is rapidly expanding while we are in the midst
4551   // of precleaning.
4552   HeapWord* startAddr = gen->reserved().start();
4553   HeapWord* endAddr   = gen->reserved().end();
4554 
4555   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
4556 
4557   size_t numDirtyCards, cumNumDirtyCards;
4558   HeapWord *nextAddr, *lastAddr;
4559   for (cumNumDirtyCards = numDirtyCards = 0,
4560        nextAddr = lastAddr = startAddr;
4561        nextAddr < endAddr;
4562        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4563 
4564     ResourceMark rm;
4565     HandleMark   hm;
4566 
4567     MemRegion dirtyRegion;
4568     {
4569       stopTimer();
4570       CMSTokenSync ts(true);
4571       startTimer();
4572       sample_eden();
4573       // Get dirty region starting at nextOffset (inclusive),
4574       // simultaneously clearing it.
4575       dirtyRegion =
4576         _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4577       assert(dirtyRegion.start() >= nextAddr,
4578              "returned region inconsistent?");
4579     }
4580     // Remember where the next search should begin.
4581     // The returned region (if non-empty) is a right open interval,
4582     // so lastOffset is obtained from the right end of that
4583     // interval.
4584     lastAddr = dirtyRegion.end();
4585     // Should do something more transparent and less hacky XXX
4586     numDirtyCards =
4587       _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4588 
4589     // We'll scan the cards in the dirty region (with periodic
4590     // yields for foreground GC as needed).
4591     if (!dirtyRegion.is_empty()) {
4592       assert(numDirtyCards > 0, "consistency check");
4593       HeapWord* stop_point = NULL;
4594       {
4595         stopTimer();
4596         CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4597                                  bitMapLock());
4598         startTimer();
4599         verify_work_stacks_empty();
4600         verify_overflow_empty();
4601         sample_eden();
4602         stop_point =
4603           gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4604       }
4605       if (stop_point != NULL) {
4606         // The careful iteration stopped early either because it found an
4607         // uninitialized object, or because we were in the midst of an
4608         // "abortable preclean", which should now be aborted. Redirty
4609         // the bits corresponding to the partially-scanned or unscanned
4610         // cards. We'll either restart at the next block boundary or
4611         // abort the preclean.
4612         assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4613                (_collectorState == AbortablePreclean && should_abort_preclean()),
4614                "Unparsable objects should only be in perm gen.");
4615 
4616         stopTimer();
4617         CMSTokenSyncWithLocks ts(true, bitMapLock());
4618         startTimer();
4619         _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4620         if (should_abort_preclean()) {
4621           break; // out of preclean loop
4622         } else {
4623           // Compute the next address at which preclean should pick up;
4624           // might need bitMapLock in order to read P-bits.
4625           lastAddr = next_card_start_after_block(stop_point);
4626         }
4627       }
4628     } else {
4629       assert(lastAddr == endAddr, "consistency check");
4630       assert(numDirtyCards == 0, "consistency check");
4631       break;
4632     }
4633   }
4634   verify_work_stacks_empty();
4635   verify_overflow_empty();
4636   return cumNumDirtyCards;
4637 }
4638 
4639 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4640 // below are largely identical; if you need to modify
4641 // one of these methods, please check the other method too.
4642 
4643 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4644   ScanMarkedObjectsAgainCarefullyClosure* cl) {
4645   // strategy: it's similar to precleamModUnionTable above, in that
4646   // we accumulate contiguous ranges of dirty cards, mark these cards
4647   // precleaned, then scan the region covered by these cards.
4648   HeapWord* endAddr   = (HeapWord*)(gen->_virtual_space.high());
4649   HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4650 
4651   cl->setFreelistLock(gen->freelistLock());   // needed for yielding
4652 
4653   size_t numDirtyCards, cumNumDirtyCards;
4654   HeapWord *lastAddr, *nextAddr;
4655 
4656   for (cumNumDirtyCards = numDirtyCards = 0,
4657        nextAddr = lastAddr = startAddr;
4658        nextAddr < endAddr;
4659        nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4660 
4661     ResourceMark rm;
4662     HandleMark   hm;
4663 
4664     MemRegion dirtyRegion;
4665     {
4666       // See comments in "Precleaning notes" above on why we
4667       // do this locking. XXX Could the locking overheads be
4668       // too high when dirty cards are sparse? [I don't think so.]
4669       stopTimer();
4670       CMSTokenSync x(true); // is cms thread
4671       startTimer();
4672       sample_eden();
4673       // Get and clear dirty region from card table
4674       dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4675                                     MemRegion(nextAddr, endAddr),
4676                                     true,
4677                                     CardTableModRefBS::precleaned_card_val());
4678 
4679       assert(dirtyRegion.start() >= nextAddr,
4680              "returned region inconsistent?");
4681     }
4682     lastAddr = dirtyRegion.end();
4683     numDirtyCards =
4684       dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4685 
4686     if (!dirtyRegion.is_empty()) {
4687       stopTimer();
4688       CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4689       startTimer();
4690       sample_eden();
4691       verify_work_stacks_empty();
4692       verify_overflow_empty();
4693       HeapWord* stop_point =
4694         gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4695       if (stop_point != NULL) {
4696         // The careful iteration stopped early because it found an
4697         // uninitialized object.  Redirty the bits corresponding to the
4698         // partially-scanned or unscanned cards, and start again at the
4699         // next block boundary.
4700         assert(CMSPermGenPrecleaningEnabled ||
4701                (_collectorState == AbortablePreclean && should_abort_preclean()),
4702                "Unparsable objects should only be in perm gen.");
4703         _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4704         if (should_abort_preclean()) {
4705           break; // out of preclean loop
4706         } else {
4707           // Compute the next address at which preclean should pick up.
4708           lastAddr = next_card_start_after_block(stop_point);
4709         }
4710       }
4711     } else {
4712       break;
4713     }
4714   }
4715   verify_work_stacks_empty();
4716   verify_overflow_empty();
4717   return cumNumDirtyCards;
4718 }
4719 
4720 void CMSCollector::checkpointRootsFinal(bool asynch,
4721   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4722   assert(_collectorState == FinalMarking, "incorrect state transition?");
4723   check_correct_thread_executing();
4724   // world is stopped at this checkpoint
4725   assert(SafepointSynchronize::is_at_safepoint(),
4726          "world should be stopped");
4727   verify_work_stacks_empty();
4728   verify_overflow_empty();
4729 
4730   SpecializationStats::clear();
4731   if (PrintGCDetails) {
4732     gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4733                         _young_gen->used() / K,
4734                         _young_gen->capacity() / K);
4735   }
4736   if (asynch) {
4737     if (CMSScavengeBeforeRemark) {
4738       GenCollectedHeap* gch = GenCollectedHeap::heap();
4739       // Temporarily set flag to false, GCH->do_collection will
4740       // expect it to be false and set to true
4741       FlagSetting fl(gch->_is_gc_active, false);
4742       NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4743         PrintGCDetails && Verbose, true, gclog_or_tty);)
4744       int level = _cmsGen->level() - 1;
4745       if (level >= 0) {
4746         gch->do_collection(true,        // full (i.e. force, see below)
4747                            false,       // !clear_all_soft_refs
4748                            0,           // size
4749                            false,       // is_tlab
4750                            level        // max_level
4751                           );
4752       }
4753     }
4754     FreelistLocker x(this);
4755     MutexLockerEx y(bitMapLock(),
4756                     Mutex::_no_safepoint_check_flag);
4757     assert(!init_mark_was_synchronous, "but that's impossible!");
4758     checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4759   } else {
4760     // already have all the locks
4761     checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4762                              init_mark_was_synchronous);
4763   }
4764   verify_work_stacks_empty();
4765   verify_overflow_empty();
4766   SpecializationStats::print();
4767 }
4768 
4769 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4770   bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4771 
4772   NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4773 
4774   assert(haveFreelistLocks(), "must have free list locks");
4775   assert_lock_strong(bitMapLock());
4776 
4777   if (UseAdaptiveSizePolicy) {
4778     size_policy()->checkpoint_roots_final_begin();
4779   }
4780 
4781   ResourceMark rm;
4782   HandleMark   hm;
4783 
4784   GenCollectedHeap* gch = GenCollectedHeap::heap();
4785 
4786   if (should_unload_classes()) {
4787     CodeCache::gc_prologue();
4788   }
4789   assert(haveFreelistLocks(), "must have free list locks");
4790   assert_lock_strong(bitMapLock());
4791 
4792   if (!init_mark_was_synchronous) {
4793     // We might assume that we need not fill TLAB's when
4794     // CMSScavengeBeforeRemark is set, because we may have just done
4795     // a scavenge which would have filled all TLAB's -- and besides
4796     // Eden would be empty. This however may not always be the case --
4797     // for instance although we asked for a scavenge, it may not have
4798     // happened because of a JNI critical section. We probably need
4799     // a policy for deciding whether we can in that case wait until
4800     // the critical section releases and then do the remark following
4801     // the scavenge, and skip it here. In the absence of that policy,
4802     // or of an indication of whether the scavenge did indeed occur,
4803     // we cannot rely on TLAB's having been filled and must do
4804     // so here just in case a scavenge did not happen.
4805     gch->ensure_parsability(false);  // fill TLAB's, but no need to retire them
4806     // Update the saved marks which may affect the root scans.
4807     gch->save_marks();
4808 
4809     {
4810       COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4811 
4812       // Note on the role of the mod union table:
4813       // Since the marker in "markFromRoots" marks concurrently with
4814       // mutators, it is possible for some reachable objects not to have been
4815       // scanned. For instance, an only reference to an object A was
4816       // placed in object B after the marker scanned B. Unless B is rescanned,
4817       // A would be collected. Such updates to references in marked objects
4818       // are detected via the mod union table which is the set of all cards
4819       // dirtied since the first checkpoint in this GC cycle and prior to
4820       // the most recent young generation GC, minus those cleaned up by the
4821       // concurrent precleaning.
4822       if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
4823         TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4824         do_remark_parallel();
4825       } else {
4826         TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4827                     gclog_or_tty);
4828         do_remark_non_parallel();
4829       }
4830     }
4831   } else {
4832     assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4833     // The initial mark was stop-world, so there's no rescanning to
4834     // do; go straight on to the next step below.
4835   }
4836   verify_work_stacks_empty();
4837   verify_overflow_empty();
4838 
4839   {
4840     NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4841     refProcessingWork(asynch, clear_all_soft_refs);
4842   }
4843   verify_work_stacks_empty();
4844   verify_overflow_empty();
4845 
4846   if (should_unload_classes()) {
4847     CodeCache::gc_epilogue();
4848   }
4849 
4850   // If we encountered any (marking stack / work queue) overflow
4851   // events during the current CMS cycle, take appropriate
4852   // remedial measures, where possible, so as to try and avoid
4853   // recurrence of that condition.
4854   assert(_markStack.isEmpty(), "No grey objects");
4855   size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4856                      _ser_kac_ovflw        + _ser_kac_preclean_ovflw;
4857   if (ser_ovflw > 0) {
4858     if (PrintCMSStatistics != 0) {
4859       gclog_or_tty->print_cr("Marking stack overflow (benign) "
4860         "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
4861         ", kac_preclean="SIZE_FORMAT")",
4862         _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4863         _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4864     }
4865     _markStack.expand();
4866     _ser_pmc_remark_ovflw = 0;
4867     _ser_pmc_preclean_ovflw = 0;
4868     _ser_kac_preclean_ovflw = 0;
4869     _ser_kac_ovflw = 0;
4870   }
4871   if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4872     if (PrintCMSStatistics != 0) {
4873       gclog_or_tty->print_cr("Work queue overflow (benign) "
4874         "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4875         _par_pmc_remark_ovflw, _par_kac_ovflw);
4876     }
4877     _par_pmc_remark_ovflw = 0;
4878     _par_kac_ovflw = 0;
4879   }
4880   if (PrintCMSStatistics != 0) {
4881      if (_markStack._hit_limit > 0) {
4882        gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4883                               _markStack._hit_limit);
4884      }
4885      if (_markStack._failed_double > 0) {
4886        gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4887                               " current capacity "SIZE_FORMAT,
4888                               _markStack._failed_double,
4889                               _markStack.capacity());
4890      }
4891   }
4892   _markStack._hit_limit = 0;
4893   _markStack._failed_double = 0;
4894 
4895   if ((VerifyAfterGC || VerifyDuringGC) &&
4896       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4897     verify_after_remark();
4898   }
4899 
4900   // Change under the freelistLocks.
4901   _collectorState = Sweeping;
4902   // Call isAllClear() under bitMapLock
4903   assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
4904     " final marking");
4905   if (UseAdaptiveSizePolicy) {
4906     size_policy()->checkpoint_roots_final_end(gch->gc_cause());
4907   }
4908 }
4909 
4910 // Parallel remark task
4911 class CMSParRemarkTask: public AbstractGangTask {
4912   CMSCollector* _collector;
4913   WorkGang*     _workers;
4914   int           _n_workers;
4915   CompactibleFreeListSpace* _cms_space;
4916   CompactibleFreeListSpace* _perm_space;
4917 
4918   // The per-thread work queues, available here for stealing.
4919   OopTaskQueueSet*       _task_queues;
4920   ParallelTaskTerminator _term;
4921 
4922  public:
4923   CMSParRemarkTask(CMSCollector* collector,
4924                    CompactibleFreeListSpace* cms_space,
4925                    CompactibleFreeListSpace* perm_space,
4926                    int n_workers, WorkGang* workers,
4927                    OopTaskQueueSet* task_queues):
4928     AbstractGangTask("Rescan roots and grey objects in parallel"),
4929     _collector(collector),
4930     _cms_space(cms_space), _perm_space(perm_space),
4931     _n_workers(n_workers),
4932     _workers(workers),
4933     _task_queues(task_queues),
4934     _term(workers->total_workers(), task_queues) { }
4935 
4936   OopTaskQueueSet* task_queues() { return _task_queues; }
4937 
4938   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4939 
4940   ParallelTaskTerminator* terminator() { return &_term; }
4941 
4942   void work(int i);
4943 
4944  private:
4945   // Work method in support of parallel rescan ... of young gen spaces
4946   void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
4947                              ContiguousSpace* space,
4948                              HeapWord** chunk_array, size_t chunk_top);
4949 
4950   // ... of  dirty cards in old space
4951   void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4952                                   Par_MarkRefsIntoAndScanClosure* cl);
4953 
4954   // ... work stealing for the above
4955   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4956 };
4957 
4958 void CMSParRemarkTask::work(int i) {
4959   elapsedTimer _timer;
4960   ResourceMark rm;
4961   HandleMark   hm;
4962 
4963   // ---------- rescan from roots --------------
4964   _timer.start();
4965   GenCollectedHeap* gch = GenCollectedHeap::heap();
4966   Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4967     _collector->_span, _collector->ref_processor(),
4968     &(_collector->_markBitMap),
4969     work_queue(i), &(_collector->_revisitStack));
4970 
4971   // Rescan young gen roots first since these are likely
4972   // coarsely partitioned and may, on that account, constitute
4973   // the critical path; thus, it's best to start off that
4974   // work first.
4975   // ---------- young gen roots --------------
4976   {
4977     DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
4978     EdenSpace* eden_space = dng->eden();
4979     ContiguousSpace* from_space = dng->from();
4980     ContiguousSpace* to_space   = dng->to();
4981 
4982     HeapWord** eca = _collector->_eden_chunk_array;
4983     size_t     ect = _collector->_eden_chunk_index;
4984     HeapWord** sca = _collector->_survivor_chunk_array;
4985     size_t     sct = _collector->_survivor_chunk_index;
4986 
4987     assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4988     assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4989 
4990     do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
4991     do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
4992     do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
4993 
4994     _timer.stop();
4995     if (PrintCMSStatistics != 0) {
4996       gclog_or_tty->print_cr(
4997         "Finished young gen rescan work in %dth thread: %3.3f sec",
4998         i, _timer.seconds());
4999     }
5000   }
5001 
5002   // ---------- remaining roots --------------
5003   _timer.reset();
5004   _timer.start();
5005   gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5006                                 false,     // yg was scanned above
5007                                 true,      // collecting perm gen
5008                                 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5009                                 NULL, &par_mrias_cl);
5010   _timer.stop();
5011   if (PrintCMSStatistics != 0) {
5012     gclog_or_tty->print_cr(
5013       "Finished remaining root rescan work in %dth thread: %3.3f sec",
5014       i, _timer.seconds());
5015   }
5016 
5017   // ---------- rescan dirty cards ------------
5018   _timer.reset();
5019   _timer.start();
5020 
5021   // Do the rescan tasks for each of the two spaces
5022   // (cms_space and perm_space) in turn.
5023   do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
5024   do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
5025   _timer.stop();
5026   if (PrintCMSStatistics != 0) {
5027     gclog_or_tty->print_cr(
5028       "Finished dirty card rescan work in %dth thread: %3.3f sec",
5029       i, _timer.seconds());
5030   }
5031 
5032   // ---------- steal work from other threads ...
5033   // ---------- ... and drain overflow list.
5034   _timer.reset();
5035   _timer.start();
5036   do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5037   _timer.stop();
5038   if (PrintCMSStatistics != 0) {
5039     gclog_or_tty->print_cr(
5040       "Finished work stealing in %dth thread: %3.3f sec",
5041       i, _timer.seconds());
5042   }
5043 }
5044 
5045 void
5046 CMSParRemarkTask::do_young_space_rescan(int i,
5047   Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5048   HeapWord** chunk_array, size_t chunk_top) {
5049   // Until all tasks completed:
5050   // . claim an unclaimed task
5051   // . compute region boundaries corresponding to task claimed
5052   //   using chunk_array
5053   // . par_oop_iterate(cl) over that region
5054 
5055   ResourceMark rm;
5056   HandleMark   hm;
5057 
5058   SequentialSubTasksDone* pst = space->par_seq_tasks();
5059   assert(pst->valid(), "Uninitialized use?");
5060 
5061   int nth_task = 0;
5062   int n_tasks  = pst->n_tasks();
5063 
5064   HeapWord *start, *end;
5065   while (!pst->is_task_claimed(/* reference */ nth_task)) {
5066     // We claimed task # nth_task; compute its boundaries.
5067     if (chunk_top == 0) {  // no samples were taken
5068       assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5069       start = space->bottom();
5070       end   = space->top();
5071     } else if (nth_task == 0) {
5072       start = space->bottom();
5073       end   = chunk_array[nth_task];
5074     } else if (nth_task < (jint)chunk_top) {
5075       assert(nth_task >= 1, "Control point invariant");
5076       start = chunk_array[nth_task - 1];
5077       end   = chunk_array[nth_task];
5078     } else {
5079       assert(nth_task == (jint)chunk_top, "Control point invariant");
5080       start = chunk_array[chunk_top - 1];
5081       end   = space->top();
5082     }
5083     MemRegion mr(start, end);
5084     // Verify that mr is in space
5085     assert(mr.is_empty() || space->used_region().contains(mr),
5086            "Should be in space");
5087     // Verify that "start" is an object boundary
5088     assert(mr.is_empty() || oop(mr.start())->is_oop(),
5089            "Should be an oop");
5090     space->par_oop_iterate(mr, cl);
5091   }
5092   pst->all_tasks_completed();
5093 }
5094 
5095 void
5096 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5097   CompactibleFreeListSpace* sp, int i,
5098   Par_MarkRefsIntoAndScanClosure* cl) {
5099   // Until all tasks completed:
5100   // . claim an unclaimed task
5101   // . compute region boundaries corresponding to task claimed
5102   // . transfer dirty bits ct->mut for that region
5103   // . apply rescanclosure to dirty mut bits for that region
5104 
5105   ResourceMark rm;
5106   HandleMark   hm;
5107 
5108   OopTaskQueue* work_q = work_queue(i);
5109   ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5110   // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5111   // CAUTION: This closure has state that persists across calls to
5112   // the work method dirty_range_iterate_clear() in that it has
5113   // imbedded in it a (subtype of) UpwardsObjectClosure. The
5114   // use of that state in the imbedded UpwardsObjectClosure instance
5115   // assumes that the cards are always iterated (even if in parallel
5116   // by several threads) in monotonically increasing order per each
5117   // thread. This is true of the implementation below which picks
5118   // card ranges (chunks) in monotonically increasing order globally
5119   // and, a-fortiori, in monotonically increasing order per thread
5120   // (the latter order being a subsequence of the former).
5121   // If the work code below is ever reorganized into a more chaotic
5122   // work-partitioning form than the current "sequential tasks"
5123   // paradigm, the use of that persistent state will have to be
5124   // revisited and modified appropriately. See also related
5125   // bug 4756801 work on which should examine this code to make
5126   // sure that the changes there do not run counter to the
5127   // assumptions made here and necessary for correctness and
5128   // efficiency. Note also that this code might yield inefficient
5129   // behaviour in the case of very large objects that span one or
5130   // more work chunks. Such objects would potentially be scanned
5131   // several times redundantly. Work on 4756801 should try and
5132   // address that performance anomaly if at all possible. XXX
5133   MemRegion  full_span  = _collector->_span;
5134   CMSBitMap* bm    = &(_collector->_markBitMap);     // shared
5135   CMSMarkStack* rs = &(_collector->_revisitStack);   // shared
5136   MarkFromDirtyCardsClosure
5137     greyRescanClosure(_collector, full_span, // entire span of interest
5138                       sp, bm, work_q, rs, cl);
5139 
5140   SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5141   assert(pst->valid(), "Uninitialized use?");
5142   int nth_task = 0;
5143   const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5144   MemRegion span = sp->used_region();
5145   HeapWord* start_addr = span.start();
5146   HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5147                                            alignment);
5148   const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5149   assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5150          start_addr, "Check alignment");
5151   assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5152          chunk_size, "Check alignment");
5153 
5154   while (!pst->is_task_claimed(/* reference */ nth_task)) {
5155     // Having claimed the nth_task, compute corresponding mem-region,
5156     // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5157     // The alignment restriction ensures that we do not need any
5158     // synchronization with other gang-workers while setting or
5159     // clearing bits in thus chunk of the MUT.
5160     MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5161                                     start_addr + (nth_task+1)*chunk_size);
5162     // The last chunk's end might be way beyond end of the
5163     // used region. In that case pull back appropriately.
5164     if (this_span.end() > end_addr) {
5165       this_span.set_end(end_addr);
5166       assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5167     }
5168     // Iterate over the dirty cards covering this chunk, marking them
5169     // precleaned, and setting the corresponding bits in the mod union
5170     // table. Since we have been careful to partition at Card and MUT-word
5171     // boundaries no synchronization is needed between parallel threads.
5172     _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5173                                                  &modUnionClosure);
5174 
5175     // Having transferred these marks into the modUnionTable,
5176     // rescan the marked objects on the dirty cards in the modUnionTable.
5177     // Even if this is at a synchronous collection, the initial marking
5178     // may have been done during an asynchronous collection so there
5179     // may be dirty bits in the mod-union table.
5180     _collector->_modUnionTable.dirty_range_iterate_clear(
5181                   this_span, &greyRescanClosure);
5182     _collector->_modUnionTable.verifyNoOneBitsInRange(
5183                                  this_span.start(),
5184                                  this_span.end());
5185   }
5186   pst->all_tasks_completed();  // declare that i am done
5187 }
5188 
5189 // . see if we can share work_queues with ParNew? XXX
5190 void
5191 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5192                                 int* seed) {
5193   OopTaskQueue* work_q = work_queue(i);
5194   NOT_PRODUCT(int num_steals = 0;)
5195   oop obj_to_scan;
5196   CMSBitMap* bm = &(_collector->_markBitMap);
5197   size_t num_from_overflow_list =
5198            MIN2((size_t)work_q->max_elems()/4,
5199                 (size_t)ParGCDesiredObjsFromOverflowList);
5200 
5201   while (true) {
5202     // Completely finish any left over work from (an) earlier round(s)
5203     cl->trim_queue(0);
5204     // Now check if there's any work in the overflow list
5205     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5206                                                 work_q)) {
5207       // found something in global overflow list;
5208       // not yet ready to go stealing work from others.
5209       // We'd like to assert(work_q->size() != 0, ...)
5210       // because we just took work from the overflow list,
5211       // but of course we can't since all of that could have
5212       // been already stolen from us.
5213       // "He giveth and He taketh away."
5214       continue;
5215     }
5216     // Verify that we have no work before we resort to stealing
5217     assert(work_q->size() == 0, "Have work, shouldn't steal");
5218     // Try to steal from other queues that have work
5219     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5220       NOT_PRODUCT(num_steals++;)
5221       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5222       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5223       // Do scanning work
5224       obj_to_scan->oop_iterate(cl);
5225       // Loop around, finish this work, and try to steal some more
5226     } else if (terminator()->offer_termination()) {
5227         break;  // nirvana from the infinite cycle
5228     }
5229   }
5230   NOT_PRODUCT(
5231     if (PrintCMSStatistics != 0) {
5232       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5233     }
5234   )
5235   assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5236          "Else our work is not yet done");
5237 }
5238 
5239 // Return a thread-local PLAB recording array, as appropriate.
5240 void* CMSCollector::get_data_recorder(int thr_num) {
5241   if (_survivor_plab_array != NULL &&
5242       (CMSPLABRecordAlways ||
5243        (_collectorState > Marking && _collectorState < FinalMarking))) {
5244     assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5245     ChunkArray* ca = &_survivor_plab_array[thr_num];
5246     ca->reset();   // clear it so that fresh data is recorded
5247     return (void*) ca;
5248   } else {
5249     return NULL;
5250   }
5251 }
5252 
5253 // Reset all the thread-local PLAB recording arrays
5254 void CMSCollector::reset_survivor_plab_arrays() {
5255   for (uint i = 0; i < ParallelGCThreads; i++) {
5256     _survivor_plab_array[i].reset();
5257   }
5258 }
5259 
5260 // Merge the per-thread plab arrays into the global survivor chunk
5261 // array which will provide the partitioning of the survivor space
5262 // for CMS rescan.
5263 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
5264   assert(_survivor_plab_array  != NULL, "Error");
5265   assert(_survivor_chunk_array != NULL, "Error");
5266   assert(_collectorState == FinalMarking, "Error");
5267   for (uint j = 0; j < ParallelGCThreads; j++) {
5268     _cursor[j] = 0;
5269   }
5270   HeapWord* top = surv->top();
5271   size_t i;
5272   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
5273     HeapWord* min_val = top;          // Higher than any PLAB address
5274     uint      min_tid = 0;            // position of min_val this round
5275     for (uint j = 0; j < ParallelGCThreads; j++) {
5276       ChunkArray* cur_sca = &_survivor_plab_array[j];
5277       if (_cursor[j] == cur_sca->end()) {
5278         continue;
5279       }
5280       assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5281       HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5282       assert(surv->used_region().contains(cur_val), "Out of bounds value");
5283       if (cur_val < min_val) {
5284         min_tid = j;
5285         min_val = cur_val;
5286       } else {
5287         assert(cur_val < top, "All recorded addresses should be less");
5288       }
5289     }
5290     // At this point min_val and min_tid are respectively
5291     // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5292     // and the thread (j) that witnesses that address.
5293     // We record this address in the _survivor_chunk_array[i]
5294     // and increment _cursor[min_tid] prior to the next round i.
5295     if (min_val == top) {
5296       break;
5297     }
5298     _survivor_chunk_array[i] = min_val;
5299     _cursor[min_tid]++;
5300   }
5301   // We are all done; record the size of the _survivor_chunk_array
5302   _survivor_chunk_index = i; // exclusive: [0, i)
5303   if (PrintCMSStatistics > 0) {
5304     gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5305   }
5306   // Verify that we used up all the recorded entries
5307   #ifdef ASSERT
5308     size_t total = 0;
5309     for (uint j = 0; j < ParallelGCThreads; j++) {
5310       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5311       total += _cursor[j];
5312     }
5313     assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5314     // Check that the merged array is in sorted order
5315     if (total > 0) {
5316       for (size_t i = 0; i < total - 1; i++) {
5317         if (PrintCMSStatistics > 0) {
5318           gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5319                               i, _survivor_chunk_array[i]);
5320         }
5321         assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5322                "Not sorted");
5323       }
5324     }
5325   #endif // ASSERT
5326 }
5327 
5328 // Set up the space's par_seq_tasks structure for work claiming
5329 // for parallel rescan of young gen.
5330 // See ParRescanTask where this is currently used.
5331 void
5332 CMSCollector::
5333 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5334   assert(n_threads > 0, "Unexpected n_threads argument");
5335   DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5336 
5337   // Eden space
5338   {
5339     SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5340     assert(!pst->valid(), "Clobbering existing data?");
5341     // Each valid entry in [0, _eden_chunk_index) represents a task.
5342     size_t n_tasks = _eden_chunk_index + 1;
5343     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5344     pst->set_par_threads(n_threads);
5345     pst->set_n_tasks((int)n_tasks);
5346   }
5347 
5348   // Merge the survivor plab arrays into _survivor_chunk_array
5349   if (_survivor_plab_array != NULL) {
5350     merge_survivor_plab_arrays(dng->from());
5351   } else {
5352     assert(_survivor_chunk_index == 0, "Error");
5353   }
5354 
5355   // To space
5356   {
5357     SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5358     assert(!pst->valid(), "Clobbering existing data?");
5359     pst->set_par_threads(n_threads);
5360     pst->set_n_tasks(1);
5361     assert(pst->valid(), "Error");
5362   }
5363 
5364   // From space
5365   {
5366     SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5367     assert(!pst->valid(), "Clobbering existing data?");
5368     size_t n_tasks = _survivor_chunk_index + 1;
5369     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5370     pst->set_par_threads(n_threads);
5371     pst->set_n_tasks((int)n_tasks);
5372     assert(pst->valid(), "Error");
5373   }
5374 }
5375 
5376 // Parallel version of remark
5377 void CMSCollector::do_remark_parallel() {
5378   GenCollectedHeap* gch = GenCollectedHeap::heap();
5379   WorkGang* workers = gch->workers();
5380   assert(workers != NULL, "Need parallel worker threads.");
5381   int n_workers = workers->total_workers();
5382   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
5383   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5384 
5385   CMSParRemarkTask tsk(this,
5386     cms_space, perm_space,
5387     n_workers, workers, task_queues());
5388 
5389   // Set up for parallel process_strong_roots work.
5390   gch->set_par_threads(n_workers);
5391   gch->change_strong_roots_parity();
5392   // We won't be iterating over the cards in the card table updating
5393   // the younger_gen cards, so we shouldn't call the following else
5394   // the verification code as well as subsequent younger_refs_iterate
5395   // code would get confused. XXX
5396   // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5397 
5398   // The young gen rescan work will not be done as part of
5399   // process_strong_roots (which currently doesn't knw how to
5400   // parallelize such a scan), but rather will be broken up into
5401   // a set of parallel tasks (via the sampling that the [abortable]
5402   // preclean phase did of EdenSpace, plus the [two] tasks of
5403   // scanning the [two] survivor spaces. Further fine-grain
5404   // parallelization of the scanning of the survivor spaces
5405   // themselves, and of precleaning of the younger gen itself
5406   // is deferred to the future.
5407   initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5408 
5409   // The dirty card rescan work is broken up into a "sequence"
5410   // of parallel tasks (per constituent space) that are dynamically
5411   // claimed by the parallel threads.
5412   cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5413   perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5414 
5415   // It turns out that even when we're using 1 thread, doing the work in a
5416   // separate thread causes wide variance in run times.  We can't help this
5417   // in the multi-threaded case, but we special-case n=1 here to get
5418   // repeatable measurements of the 1-thread overhead of the parallel code.
5419   if (n_workers > 1) {
5420     // Make refs discovery MT-safe
5421     ReferenceProcessorMTMutator mt(ref_processor(), true);
5422     workers->run_task(&tsk);
5423   } else {
5424     tsk.work(0);
5425   }
5426   gch->set_par_threads(0);  // 0 ==> non-parallel.
5427   // restore, single-threaded for now, any preserved marks
5428   // as a result of work_q overflow
5429   restore_preserved_marks_if_any();
5430 }
5431 
5432 // Non-parallel version of remark
5433 void CMSCollector::do_remark_non_parallel() {
5434   ResourceMark rm;
5435   HandleMark   hm;
5436   GenCollectedHeap* gch = GenCollectedHeap::heap();
5437   MarkRefsIntoAndScanClosure
5438     mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5439              &_markStack, &_revisitStack, this,
5440              false /* should_yield */, false /* not precleaning */);
5441   MarkFromDirtyCardsClosure
5442     markFromDirtyCardsClosure(this, _span,
5443                               NULL,  // space is set further below
5444                               &_markBitMap, &_markStack, &_revisitStack,
5445                               &mrias_cl);
5446   {
5447     TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5448     // Iterate over the dirty cards, setting the corresponding bits in the
5449     // mod union table.
5450     {
5451       ModUnionClosure modUnionClosure(&_modUnionTable);
5452       _ct->ct_bs()->dirty_card_iterate(
5453                       _cmsGen->used_region(),
5454                       &modUnionClosure);
5455       _ct->ct_bs()->dirty_card_iterate(
5456                       _permGen->used_region(),
5457                       &modUnionClosure);
5458     }
5459     // Having transferred these marks into the modUnionTable, we just need
5460     // to rescan the marked objects on the dirty cards in the modUnionTable.
5461     // The initial marking may have been done during an asynchronous
5462     // collection so there may be dirty bits in the mod-union table.
5463     const int alignment =
5464       CardTableModRefBS::card_size * BitsPerWord;
5465     {
5466       // ... First handle dirty cards in CMS gen
5467       markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5468       MemRegion ur = _cmsGen->used_region();
5469       HeapWord* lb = ur.start();
5470       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5471       MemRegion cms_span(lb, ub);
5472       _modUnionTable.dirty_range_iterate_clear(cms_span,
5473                                                &markFromDirtyCardsClosure);
5474       verify_work_stacks_empty();
5475       if (PrintCMSStatistics != 0) {
5476         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5477           markFromDirtyCardsClosure.num_dirty_cards());
5478       }
5479     }
5480     {
5481       // .. and then repeat for dirty cards in perm gen
5482       markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5483       MemRegion ur = _permGen->used_region();
5484       HeapWord* lb = ur.start();
5485       HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5486       MemRegion perm_span(lb, ub);
5487       _modUnionTable.dirty_range_iterate_clear(perm_span,
5488                                                &markFromDirtyCardsClosure);
5489       verify_work_stacks_empty();
5490       if (PrintCMSStatistics != 0) {
5491         gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5492           markFromDirtyCardsClosure.num_dirty_cards());
5493       }
5494     }
5495   }
5496   if (VerifyDuringGC &&
5497       GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5498     HandleMark hm;  // Discard invalid handles created during verification
5499     Universe::verify(true);
5500   }
5501   {
5502     TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5503 
5504     verify_work_stacks_empty();
5505 
5506     gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5507     gch->gen_process_strong_roots(_cmsGen->level(),
5508                                   true,  // younger gens as roots
5509                                   true,  // collecting perm gen
5510                                   SharedHeap::ScanningOption(roots_scanning_options()),
5511                                   NULL, &mrias_cl);
5512   }
5513   verify_work_stacks_empty();
5514   // Restore evacuated mark words, if any, used for overflow list links
5515   if (!CMSOverflowEarlyRestoration) {
5516     restore_preserved_marks_if_any();
5517   }
5518   verify_overflow_empty();
5519 }
5520 
5521 ////////////////////////////////////////////////////////
5522 // Parallel Reference Processing Task Proxy Class
5523 ////////////////////////////////////////////////////////
5524 class CMSRefProcTaskProxy: public AbstractGangTask {
5525   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5526   CMSCollector*          _collector;
5527   CMSBitMap*             _mark_bit_map;
5528   const MemRegion        _span;
5529   OopTaskQueueSet*       _task_queues;
5530   ParallelTaskTerminator _term;
5531   ProcessTask&           _task;
5532 
5533 public:
5534   CMSRefProcTaskProxy(ProcessTask&     task,
5535                       CMSCollector*    collector,
5536                       const MemRegion& span,
5537                       CMSBitMap*       mark_bit_map,
5538                       int              total_workers,
5539                       OopTaskQueueSet* task_queues):
5540     AbstractGangTask("Process referents by policy in parallel"),
5541     _task(task),
5542     _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
5543     _task_queues(task_queues),
5544     _term(total_workers, task_queues)
5545     {
5546       assert(_collector->_span.equals(_span) && !_span.is_empty(),
5547              "Inconsistency in _span");
5548     }
5549 
5550   OopTaskQueueSet* task_queues() { return _task_queues; }
5551 
5552   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5553 
5554   ParallelTaskTerminator* terminator() { return &_term; }
5555 
5556   void do_work_steal(int i,
5557                      CMSParDrainMarkingStackClosure* drain,
5558                      CMSParKeepAliveClosure* keep_alive,
5559                      int* seed);
5560 
5561   virtual void work(int i);
5562 };
5563 
5564 void CMSRefProcTaskProxy::work(int i) {
5565   assert(_collector->_span.equals(_span), "Inconsistency in _span");
5566   CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5567                                         _mark_bit_map, work_queue(i));
5568   CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5569                                                  _mark_bit_map, work_queue(i));
5570   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5571   _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5572   if (_task.marks_oops_alive()) {
5573     do_work_steal(i, &par_drain_stack, &par_keep_alive,
5574                   _collector->hash_seed(i));
5575   }
5576   assert(work_queue(i)->size() == 0, "work_queue should be empty");
5577   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5578 }
5579 
5580 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5581   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5582   EnqueueTask& _task;
5583 
5584 public:
5585   CMSRefEnqueueTaskProxy(EnqueueTask& task)
5586     : AbstractGangTask("Enqueue reference objects in parallel"),
5587       _task(task)
5588   { }
5589 
5590   virtual void work(int i)
5591   {
5592     _task.work(i);
5593   }
5594 };
5595 
5596 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5597   MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5598    _collector(collector),
5599    _span(span),
5600    _bit_map(bit_map),
5601    _work_queue(work_queue),
5602    _mark_and_push(collector, span, bit_map, work_queue),
5603    _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5604                         (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5605 { }
5606 
5607 // . see if we can share work_queues with ParNew? XXX
5608 void CMSRefProcTaskProxy::do_work_steal(int i,
5609   CMSParDrainMarkingStackClosure* drain,
5610   CMSParKeepAliveClosure* keep_alive,
5611   int* seed) {
5612   OopTaskQueue* work_q = work_queue(i);
5613   NOT_PRODUCT(int num_steals = 0;)
5614   oop obj_to_scan;
5615   size_t num_from_overflow_list =
5616            MIN2((size_t)work_q->max_elems()/4,
5617                 (size_t)ParGCDesiredObjsFromOverflowList);
5618 
5619   while (true) {
5620     // Completely finish any left over work from (an) earlier round(s)
5621     drain->trim_queue(0);
5622     // Now check if there's any work in the overflow list
5623     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5624                                                 work_q)) {
5625       // Found something in global overflow list;
5626       // not yet ready to go stealing work from others.
5627       // We'd like to assert(work_q->size() != 0, ...)
5628       // because we just took work from the overflow list,
5629       // but of course we can't, since all of that might have
5630       // been already stolen from us.
5631       continue;
5632     }
5633     // Verify that we have no work before we resort to stealing
5634     assert(work_q->size() == 0, "Have work, shouldn't steal");
5635     // Try to steal from other queues that have work
5636     if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5637       NOT_PRODUCT(num_steals++;)
5638       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5639       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5640       // Do scanning work
5641       obj_to_scan->oop_iterate(keep_alive);
5642       // Loop around, finish this work, and try to steal some more
5643     } else if (terminator()->offer_termination()) {
5644       break;  // nirvana from the infinite cycle
5645     }
5646   }
5647   NOT_PRODUCT(
5648     if (PrintCMSStatistics != 0) {
5649       gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5650     }
5651   )
5652 }
5653 
5654 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5655 {
5656   GenCollectedHeap* gch = GenCollectedHeap::heap();
5657   WorkGang* workers = gch->workers();
5658   assert(workers != NULL, "Need parallel worker threads.");
5659   int n_workers = workers->total_workers();
5660   CMSRefProcTaskProxy rp_task(task, &_collector,
5661                               _collector.ref_processor()->span(),
5662                               _collector.markBitMap(),
5663                               n_workers, _collector.task_queues());
5664   workers->run_task(&rp_task);
5665 }
5666 
5667 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5668 {
5669 
5670   GenCollectedHeap* gch = GenCollectedHeap::heap();
5671   WorkGang* workers = gch->workers();
5672   assert(workers != NULL, "Need parallel worker threads.");
5673   CMSRefEnqueueTaskProxy enq_task(task);
5674   workers->run_task(&enq_task);
5675 }
5676 
5677 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5678 
5679   ResourceMark rm;
5680   HandleMark   hm;
5681   ReferencePolicy* soft_ref_policy;
5682 
5683   assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete");
5684   // Process weak references.
5685   if (clear_all_soft_refs) {
5686     soft_ref_policy = new AlwaysClearPolicy();
5687   } else {
5688 #ifdef COMPILER2
5689     soft_ref_policy = new LRUMaxHeapPolicy();
5690 #else
5691     soft_ref_policy = new LRUCurrentHeapPolicy();
5692 #endif // COMPILER2
5693   }
5694   verify_work_stacks_empty();
5695 
5696   ReferenceProcessor* rp = ref_processor();
5697   assert(rp->span().equals(_span), "Spans should be equal");
5698   CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5699                                           &_markStack, false /* !preclean */);
5700   CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5701                                 _span, &_markBitMap, &_markStack,
5702                                 &cmsKeepAliveClosure, false /* !preclean */);
5703   {
5704     TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5705     if (rp->processing_is_mt()) {
5706       CMSRefProcTaskExecutor task_executor(*this);
5707       rp->process_discovered_references(soft_ref_policy,
5708                                         &_is_alive_closure,
5709                                         &cmsKeepAliveClosure,
5710                                         &cmsDrainMarkingStackClosure,
5711                                         &task_executor);
5712     } else {
5713       rp->process_discovered_references(soft_ref_policy,
5714                                         &_is_alive_closure,
5715                                         &cmsKeepAliveClosure,
5716                                         &cmsDrainMarkingStackClosure,
5717                                         NULL);
5718     }
5719     verify_work_stacks_empty();
5720   }
5721 
5722   if (should_unload_classes()) {
5723     {
5724       TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5725 
5726       // Follow SystemDictionary roots and unload classes
5727       bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5728 
5729       // Follow CodeCache roots and unload any methods marked for unloading
5730       CodeCache::do_unloading(&_is_alive_closure,
5731                               &cmsKeepAliveClosure,
5732                               purged_class);
5733 
5734       cmsDrainMarkingStackClosure.do_void();
5735       verify_work_stacks_empty();
5736 
5737       // Update subklass/sibling/implementor links in KlassKlass descendants
5738       assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5739       oop k;
5740       while ((k = _revisitStack.pop()) != NULL) {
5741         ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5742                        &_is_alive_closure,
5743                        &cmsKeepAliveClosure);
5744       }
5745       assert(!ClassUnloading ||
5746              (_markStack.isEmpty() && overflow_list_is_empty()),
5747              "Should not have found new reachable objects");
5748       assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5749       cmsDrainMarkingStackClosure.do_void();
5750       verify_work_stacks_empty();
5751     }
5752 
5753     {
5754       TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5755       // Now clean up stale oops in SymbolTable and StringTable
5756       SymbolTable::unlink(&_is_alive_closure);
5757       StringTable::unlink(&_is_alive_closure);
5758     }
5759   }
5760 
5761   verify_work_stacks_empty();
5762   // Restore any preserved marks as a result of mark stack or
5763   // work queue overflow
5764   restore_preserved_marks_if_any();  // done single-threaded for now
5765 
5766   rp->set_enqueuing_is_done(true);
5767   if (rp->processing_is_mt()) {
5768     CMSRefProcTaskExecutor task_executor(*this);
5769     rp->enqueue_discovered_references(&task_executor);
5770   } else {
5771     rp->enqueue_discovered_references(NULL);
5772   }
5773   rp->verify_no_references_recorded();
5774   assert(!rp->discovery_enabled(), "should have been disabled");
5775 
5776   // JVMTI object tagging is based on JNI weak refs. If any of these
5777   // refs were cleared then JVMTI needs to update its maps and
5778   // maybe post ObjectFrees to agents.
5779   JvmtiExport::cms_ref_processing_epilogue();
5780 }
5781 
5782 #ifndef PRODUCT
5783 void CMSCollector::check_correct_thread_executing() {
5784   Thread* t = Thread::current();
5785   // Only the VM thread or the CMS thread should be here.
5786   assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5787          "Unexpected thread type");
5788   // If this is the vm thread, the foreground process
5789   // should not be waiting.  Note that _foregroundGCIsActive is
5790   // true while the foreground collector is waiting.
5791   if (_foregroundGCShouldWait) {
5792     // We cannot be the VM thread
5793     assert(t->is_ConcurrentGC_thread(),
5794            "Should be CMS thread");
5795   } else {
5796     // We can be the CMS thread only if we are in a stop-world
5797     // phase of CMS collection.
5798     if (t->is_ConcurrentGC_thread()) {
5799       assert(_collectorState == InitialMarking ||
5800              _collectorState == FinalMarking,
5801              "Should be a stop-world phase");
5802       // The CMS thread should be holding the CMS_token.
5803       assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5804              "Potential interference with concurrently "
5805              "executing VM thread");
5806     }
5807   }
5808 }
5809 #endif
5810 
5811 void CMSCollector::sweep(bool asynch) {
5812   assert(_collectorState == Sweeping, "just checking");
5813   check_correct_thread_executing();
5814   verify_work_stacks_empty();
5815   verify_overflow_empty();
5816   incrementSweepCount();
5817   _sweep_timer.stop();
5818   _sweep_estimate.sample(_sweep_timer.seconds());
5819   size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5820 
5821   // PermGen verification support: If perm gen sweeping is disabled in
5822   // this cycle, we preserve the perm gen object "deadness" information
5823   // in the perm_gen_verify_bit_map. In order to do that we traverse
5824   // all blocks in perm gen and mark all dead objects.
5825   if (verifying() && !should_unload_classes()) {
5826     assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
5827            "Should have already been allocated");
5828     MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
5829                                markBitMap(), perm_gen_verify_bit_map());
5830     if (asynch) {
5831       CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5832                                bitMapLock());
5833       _permGen->cmsSpace()->blk_iterate(&mdo);
5834     } else {
5835       // In the case of synchronous sweep, we already have
5836       // the requisite locks/tokens.
5837       _permGen->cmsSpace()->blk_iterate(&mdo);
5838     }
5839   }
5840 
5841   if (asynch) {
5842     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5843     CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5844     // First sweep the old gen then the perm gen
5845     {
5846       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5847                                bitMapLock());
5848       sweepWork(_cmsGen, asynch);
5849     }
5850 
5851     // Now repeat for perm gen
5852     if (should_unload_classes()) {
5853       CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5854                              bitMapLock());
5855       sweepWork(_permGen, asynch);
5856     }
5857 
5858     // Update Universe::_heap_*_at_gc figures.
5859     // We need all the free list locks to make the abstract state
5860     // transition from Sweeping to Resetting. See detailed note
5861     // further below.
5862     {
5863       CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5864                                _permGen->freelistLock());
5865       // Update heap occupancy information which is used as
5866       // input to soft ref clearing policy at the next gc.
5867       Universe::update_heap_info_at_gc();
5868       _collectorState = Resizing;
5869     }
5870   } else {
5871     // already have needed locks
5872     sweepWork(_cmsGen,  asynch);
5873 
5874     if (should_unload_classes()) {
5875       sweepWork(_permGen, asynch);
5876     }
5877     // Update heap occupancy information which is used as
5878     // input to soft ref clearing policy at the next gc.
5879     Universe::update_heap_info_at_gc();
5880     _collectorState = Resizing;
5881   }
5882   verify_work_stacks_empty();
5883   verify_overflow_empty();
5884 
5885   _sweep_timer.reset();
5886   _sweep_timer.start();
5887 
5888   update_time_of_last_gc(os::javaTimeMillis());
5889 
5890   // NOTE on abstract state transitions:
5891   // Mutators allocate-live and/or mark the mod-union table dirty
5892   // based on the state of the collection.  The former is done in
5893   // the interval [Marking, Sweeping] and the latter in the interval
5894   // [Marking, Sweeping).  Thus the transitions into the Marking state
5895   // and out of the Sweeping state must be synchronously visible
5896   // globally to the mutators.
5897   // The transition into the Marking state happens with the world
5898   // stopped so the mutators will globally see it.  Sweeping is
5899   // done asynchronously by the background collector so the transition
5900   // from the Sweeping state to the Resizing state must be done
5901   // under the freelistLock (as is the check for whether to
5902   // allocate-live and whether to dirty the mod-union table).
5903   assert(_collectorState == Resizing, "Change of collector state to"
5904     " Resizing must be done under the freelistLocks (plural)");
5905 
5906   // Now that sweeping has been completed, if the GCH's
5907   // incremental_collection_will_fail flag is set, clear it,
5908   // thus inviting a younger gen collection to promote into
5909   // this generation. If such a promotion may still fail,
5910   // the flag will be set again when a young collection is
5911   // attempted.
5912   // I think the incremental_collection_will_fail flag's use
5913   // is specific to a 2 generation collection policy, so i'll
5914   // assert that that's the configuration we are operating within.
5915   // The use of the flag can and should be generalized appropriately
5916   // in the future to deal with a general n-generation system.
5917 
5918   GenCollectedHeap* gch = GenCollectedHeap::heap();
5919   assert(gch->collector_policy()->is_two_generation_policy(),
5920          "Resetting of incremental_collection_will_fail flag"
5921          " may be incorrect otherwise");
5922   gch->clear_incremental_collection_will_fail();
5923   gch->update_full_collections_completed(_collection_count_start);
5924 }
5925 
5926 // FIX ME!!! Looks like this belongs in CFLSpace, with
5927 // CMSGen merely delegating to it.
5928 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5929   double nearLargestPercent = 0.999;
5930   HeapWord*  minAddr        = _cmsSpace->bottom();
5931   HeapWord*  largestAddr    =
5932     (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
5933   if (largestAddr == 0) {
5934     // The dictionary appears to be empty.  In this case
5935     // try to coalesce at the end of the heap.
5936     largestAddr = _cmsSpace->end();
5937   }
5938   size_t largestOffset     = pointer_delta(largestAddr, minAddr);
5939   size_t nearLargestOffset =
5940     (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5941   _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5942 }
5943 
5944 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5945   return addr >= _cmsSpace->nearLargestChunk();
5946 }
5947 
5948 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5949   return _cmsSpace->find_chunk_at_end();
5950 }
5951 
5952 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5953                                                     bool full) {
5954   // The next lower level has been collected.  Gather any statistics
5955   // that are of interest at this point.
5956   if (!full && (current_level + 1) == level()) {
5957     // Gather statistics on the young generation collection.
5958     collector()->stats().record_gc0_end(used());
5959   }
5960 }
5961 
5962 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
5963   GenCollectedHeap* gch = GenCollectedHeap::heap();
5964   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
5965     "Wrong type of heap");
5966   CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
5967     gch->gen_policy()->size_policy();
5968   assert(sp->is_gc_cms_adaptive_size_policy(),
5969     "Wrong type of size policy");
5970   return sp;
5971 }
5972 
5973 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
5974   if (PrintGCDetails && Verbose) {
5975     gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
5976   }
5977   _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
5978   _debug_collection_type =
5979     (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
5980   if (PrintGCDetails && Verbose) {
5981     gclog_or_tty->print_cr("to %d ", _debug_collection_type);
5982   }
5983 }
5984 
5985 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
5986   bool asynch) {
5987   // We iterate over the space(s) underlying this generation,
5988   // checking the mark bit map to see if the bits corresponding
5989   // to specific blocks are marked or not. Blocks that are
5990   // marked are live and are not swept up. All remaining blocks
5991   // are swept up, with coalescing on-the-fly as we sweep up
5992   // contiguous free and/or garbage blocks:
5993   // We need to ensure that the sweeper synchronizes with allocators
5994   // and stop-the-world collectors. In particular, the following
5995   // locks are used:
5996   // . CMS token: if this is held, a stop the world collection cannot occur
5997   // . freelistLock: if this is held no allocation can occur from this
5998   //                 generation by another thread
5999   // . bitMapLock: if this is held, no other thread can access or update
6000   //
6001 
6002   // Note that we need to hold the freelistLock if we use
6003   // block iterate below; else the iterator might go awry if
6004   // a mutator (or promotion) causes block contents to change
6005   // (for instance if the allocator divvies up a block).
6006   // If we hold the free list lock, for all practical purposes
6007   // young generation GC's can't occur (they'll usually need to
6008   // promote), so we might as well prevent all young generation
6009   // GC's while we do a sweeping step. For the same reason, we might
6010   // as well take the bit map lock for the entire duration
6011 
6012   // check that we hold the requisite locks
6013   assert(have_cms_token(), "Should hold cms token");
6014   assert(   (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6015          || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6016         "Should possess CMS token to sweep");
6017   assert_lock_strong(gen->freelistLock());
6018   assert_lock_strong(bitMapLock());
6019 
6020   assert(!_sweep_timer.is_active(), "Was switched off in an outer context");
6021   gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
6022                                       _sweep_estimate.padded_average());
6023   gen->setNearLargestChunk();
6024 
6025   {
6026     SweepClosure sweepClosure(this, gen, &_markBitMap,
6027                             CMSYield && asynch);
6028     gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6029     // We need to free-up/coalesce garbage/blocks from a
6030     // co-terminal free run. This is done in the SweepClosure
6031     // destructor; so, do not remove this scope, else the
6032     // end-of-sweep-census below will be off by a little bit.
6033   }
6034   gen->cmsSpace()->sweep_completed();
6035   gen->cmsSpace()->endSweepFLCensus(sweepCount());
6036   if (should_unload_classes()) {                // unloaded classes this cycle,
6037     _concurrent_cycles_since_last_unload = 0;   // ... reset count
6038   } else {                                      // did not unload classes,
6039     _concurrent_cycles_since_last_unload++;     // ... increment count
6040   }
6041 }
6042 
6043 // Reset CMS data structures (for now just the marking bit map)
6044 // preparatory for the next cycle.
6045 void CMSCollector::reset(bool asynch) {
6046   GenCollectedHeap* gch = GenCollectedHeap::heap();
6047   CMSAdaptiveSizePolicy* sp = size_policy();
6048   AdaptiveSizePolicyOutput(sp, gch->total_collections());
6049   if (asynch) {
6050     CMSTokenSyncWithLocks ts(true, bitMapLock());
6051 
6052     // If the state is not "Resetting", the foreground  thread
6053     // has done a collection and the resetting.
6054     if (_collectorState != Resetting) {
6055       assert(_collectorState == Idling, "The state should only change"
6056         " because the foreground collector has finished the collection");
6057       return;
6058     }
6059 
6060     // Clear the mark bitmap (no grey objects to start with)
6061     // for the next cycle.
6062     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6063     CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6064 
6065     HeapWord* curAddr = _markBitMap.startWord();
6066     while (curAddr < _markBitMap.endWord()) {
6067       size_t remaining  = pointer_delta(_markBitMap.endWord(), curAddr);
6068       MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6069       _markBitMap.clear_large_range(chunk);
6070       if (ConcurrentMarkSweepThread::should_yield() &&
6071           !foregroundGCIsActive() &&
6072           CMSYield) {
6073         assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6074                "CMS thread should hold CMS token");
6075         assert_lock_strong(bitMapLock());
6076         bitMapLock()->unlock();
6077         ConcurrentMarkSweepThread::desynchronize(true);
6078         ConcurrentMarkSweepThread::acknowledge_yield_request();
6079         stopTimer();
6080         if (PrintCMSStatistics != 0) {
6081           incrementYields();
6082         }
6083         icms_wait();
6084 
6085         // See the comment in coordinator_yield()
6086         for (unsigned i = 0; i < CMSYieldSleepCount &&
6087                          ConcurrentMarkSweepThread::should_yield() &&
6088                          !CMSCollector::foregroundGCIsActive(); ++i) {
6089           os::sleep(Thread::current(), 1, false);
6090           ConcurrentMarkSweepThread::acknowledge_yield_request();
6091         }
6092 
6093         ConcurrentMarkSweepThread::synchronize(true);
6094         bitMapLock()->lock_without_safepoint_check();
6095         startTimer();
6096       }
6097       curAddr = chunk.end();
6098     }
6099     _collectorState = Idling;
6100   } else {
6101     // already have the lock
6102     assert(_collectorState == Resetting, "just checking");
6103     assert_lock_strong(bitMapLock());
6104     _markBitMap.clear_all();
6105     _collectorState = Idling;
6106   }
6107 
6108   // Stop incremental mode after a cycle completes, so that any future cycles
6109   // are triggered by allocation.
6110   stop_icms();
6111 
6112   NOT_PRODUCT(
6113     if (RotateCMSCollectionTypes) {
6114       _cmsGen->rotate_debug_collection_type();
6115     }
6116   )
6117 }
6118 
6119 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6120   gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6121   TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6122   TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6123   TraceCollectorStats tcs(counters());
6124 
6125   switch (op) {
6126     case CMS_op_checkpointRootsInitial: {
6127       checkpointRootsInitial(true);       // asynch
6128       if (PrintGC) {
6129         _cmsGen->printOccupancy("initial-mark");
6130       }
6131       break;
6132     }
6133     case CMS_op_checkpointRootsFinal: {
6134       checkpointRootsFinal(true,    // asynch
6135                            false,   // !clear_all_soft_refs
6136                            false);  // !init_mark_was_synchronous
6137       if (PrintGC) {
6138         _cmsGen->printOccupancy("remark");
6139       }
6140       break;
6141     }
6142     default:
6143       fatal("No such CMS_op");
6144   }
6145 }
6146 
6147 #ifndef PRODUCT
6148 size_t const CMSCollector::skip_header_HeapWords() {
6149   return FreeChunk::header_size();
6150 }
6151 
6152 // Try and collect here conditions that should hold when
6153 // CMS thread is exiting. The idea is that the foreground GC
6154 // thread should not be blocked if it wants to terminate
6155 // the CMS thread and yet continue to run the VM for a while
6156 // after that.
6157 void CMSCollector::verify_ok_to_terminate() const {
6158   assert(Thread::current()->is_ConcurrentGC_thread(),
6159          "should be called by CMS thread");
6160   assert(!_foregroundGCShouldWait, "should be false");
6161   // We could check here that all the various low-level locks
6162   // are not held by the CMS thread, but that is overkill; see
6163   // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6164   // is checked.
6165 }
6166 #endif
6167 
6168 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6169   assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6170          "missing Printezis mark?");
6171   HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6172   size_t size = pointer_delta(nextOneAddr + 1, addr);
6173   assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6174          "alignment problem");
6175   assert(size >= 3, "Necessary for Printezis marks to work");
6176   return size;
6177 }
6178 
6179 // A variant of the above (block_size_using_printezis_bits()) except
6180 // that we return 0 if the P-bits are not yet set.
6181 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6182   if (_markBitMap.isMarked(addr)) {
6183     assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
6184     HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6185     size_t size = pointer_delta(nextOneAddr + 1, addr);
6186     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6187            "alignment problem");
6188     assert(size >= 3, "Necessary for Printezis marks to work");
6189     return size;
6190   } else {
6191     assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
6192     return 0;
6193   }
6194 }
6195 
6196 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6197   size_t sz = 0;
6198   oop p = (oop)addr;
6199   if (p->klass_or_null() != NULL && p->is_parsable()) {
6200     sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6201   } else {
6202     sz = block_size_using_printezis_bits(addr);
6203   }
6204   assert(sz > 0, "size must be nonzero");
6205   HeapWord* next_block = addr + sz;
6206   HeapWord* next_card  = (HeapWord*)round_to((uintptr_t)next_block,
6207                                              CardTableModRefBS::card_size);
6208   assert(round_down((uintptr_t)addr,      CardTableModRefBS::card_size) <
6209          round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6210          "must be different cards");
6211   return next_card;
6212 }
6213 
6214 
6215 // CMS Bit Map Wrapper /////////////////////////////////////////
6216 
6217 // Construct a CMS bit map infrastructure, but don't create the
6218 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6219 // further below.
6220 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6221   _bm(),
6222   _shifter(shifter),
6223   _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6224 {
6225   _bmStartWord = 0;
6226   _bmWordSize  = 0;
6227 }
6228 
6229 bool CMSBitMap::allocate(MemRegion mr) {
6230   _bmStartWord = mr.start();
6231   _bmWordSize  = mr.word_size();
6232   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6233                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6234   if (!brs.is_reserved()) {
6235     warning("CMS bit map allocation failure");
6236     return false;
6237   }
6238   // For now we'll just commit all of the bit map up fromt.
6239   // Later on we'll try to be more parsimonious with swap.
6240   if (!_virtual_space.initialize(brs, brs.size())) {
6241     warning("CMS bit map backing store failure");
6242     return false;
6243   }
6244   assert(_virtual_space.committed_size() == brs.size(),
6245          "didn't reserve backing store for all of CMS bit map?");
6246   _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6247   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6248          _bmWordSize, "inconsistency in bit map sizing");
6249   _bm.set_size(_bmWordSize >> _shifter);
6250 
6251   // bm.clear(); // can we rely on getting zero'd memory? verify below
6252   assert(isAllClear(),
6253          "Expected zero'd memory from ReservedSpace constructor");
6254   assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6255          "consistency check");
6256   return true;
6257 }
6258 
6259 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6260   HeapWord *next_addr, *end_addr, *last_addr;
6261   assert_locked();
6262   assert(covers(mr), "out-of-range error");
6263   // XXX assert that start and end are appropriately aligned
6264   for (next_addr = mr.start(), end_addr = mr.end();
6265        next_addr < end_addr; next_addr = last_addr) {
6266     MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6267     last_addr = dirty_region.end();
6268     if (!dirty_region.is_empty()) {
6269       cl->do_MemRegion(dirty_region);
6270     } else {
6271       assert(last_addr == end_addr, "program logic");
6272       return;
6273     }
6274   }
6275 }
6276 
6277 #ifndef PRODUCT
6278 void CMSBitMap::assert_locked() const {
6279   CMSLockVerifier::assert_locked(lock());
6280 }
6281 
6282 bool CMSBitMap::covers(MemRegion mr) const {
6283   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6284   assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6285          "size inconsistency");
6286   return (mr.start() >= _bmStartWord) &&
6287          (mr.end()   <= endWord());
6288 }
6289 
6290 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6291     return (start >= _bmStartWord && (start + size) <= endWord());
6292 }
6293 
6294 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6295   // verify that there are no 1 bits in the interval [left, right)
6296   FalseBitMapClosure falseBitMapClosure;
6297   iterate(&falseBitMapClosure, left, right);
6298 }
6299 
6300 void CMSBitMap::region_invariant(MemRegion mr)
6301 {
6302   assert_locked();
6303   // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6304   assert(!mr.is_empty(), "unexpected empty region");
6305   assert(covers(mr), "mr should be covered by bit map");
6306   // convert address range into offset range
6307   size_t start_ofs = heapWordToOffset(mr.start());
6308   // Make sure that end() is appropriately aligned
6309   assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6310                         (1 << (_shifter+LogHeapWordSize))),
6311          "Misaligned mr.end()");
6312   size_t end_ofs   = heapWordToOffset(mr.end());
6313   assert(end_ofs > start_ofs, "Should mark at least one bit");
6314 }
6315 
6316 #endif
6317 
6318 bool CMSMarkStack::allocate(size_t size) {
6319   // allocate a stack of the requisite depth
6320   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6321                    size * sizeof(oop)));
6322   if (!rs.is_reserved()) {
6323     warning("CMSMarkStack allocation failure");
6324     return false;
6325   }
6326   if (!_virtual_space.initialize(rs, rs.size())) {
6327     warning("CMSMarkStack backing store failure");
6328     return false;
6329   }
6330   assert(_virtual_space.committed_size() == rs.size(),
6331          "didn't reserve backing store for all of CMS stack?");
6332   _base = (oop*)(_virtual_space.low());
6333   _index = 0;
6334   _capacity = size;
6335   NOT_PRODUCT(_max_depth = 0);
6336   return true;
6337 }
6338 
6339 // XXX FIX ME !!! In the MT case we come in here holding a
6340 // leaf lock. For printing we need to take a further lock
6341 // which has lower rank. We need to recallibrate the two
6342 // lock-ranks involved in order to be able to rpint the
6343 // messages below. (Or defer the printing to the caller.
6344 // For now we take the expedient path of just disabling the
6345 // messages for the problematic case.)
6346 void CMSMarkStack::expand() {
6347   assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted");
6348   if (_capacity == CMSMarkStackSizeMax) {
6349     if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6350       // We print a warning message only once per CMS cycle.
6351       gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6352     }
6353     return;
6354   }
6355   // Double capacity if possible
6356   size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax);
6357   // Do not give up existing stack until we have managed to
6358   // get the double capacity that we desired.
6359   ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6360                    new_capacity * sizeof(oop)));
6361   if (rs.is_reserved()) {
6362     // Release the backing store associated with old stack
6363     _virtual_space.release();
6364     // Reinitialize virtual space for new stack
6365     if (!_virtual_space.initialize(rs, rs.size())) {
6366       fatal("Not enough swap for expanded marking stack");
6367     }
6368     _base = (oop*)(_virtual_space.low());
6369     _index = 0;
6370     _capacity = new_capacity;
6371   } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6372     // Failed to double capacity, continue;
6373     // we print a detail message only once per CMS cycle.
6374     gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6375             SIZE_FORMAT"K",
6376             _capacity / K, new_capacity / K);
6377   }
6378 }
6379 
6380 
6381 // Closures
6382 // XXX: there seems to be a lot of code  duplication here;
6383 // should refactor and consolidate common code.
6384 
6385 // This closure is used to mark refs into the CMS generation in
6386 // the CMS bit map. Called at the first checkpoint. This closure
6387 // assumes that we do not need to re-mark dirty cards; if the CMS
6388 // generation on which this is used is not an oldest (modulo perm gen)
6389 // generation then this will lose younger_gen cards!
6390 
6391 MarkRefsIntoClosure::MarkRefsIntoClosure(
6392   MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods):
6393     _span(span),
6394     _bitMap(bitMap),
6395     _should_do_nmethods(should_do_nmethods)
6396 {
6397     assert(_ref_processor == NULL, "deliberately left NULL");
6398     assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6399 }
6400 
6401 void MarkRefsIntoClosure::do_oop(oop obj) {
6402   // if p points into _span, then mark corresponding bit in _markBitMap
6403   assert(obj->is_oop(), "expected an oop");
6404   HeapWord* addr = (HeapWord*)obj;
6405   if (_span.contains(addr)) {
6406     // this should be made more efficient
6407     _bitMap->mark(addr);
6408   }
6409 }
6410 
6411 void MarkRefsIntoClosure::do_oop(oop* p)       { MarkRefsIntoClosure::do_oop_work(p); }
6412 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6413 
6414 // A variant of the above, used for CMS marking verification.
6415 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6416   MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6417   bool should_do_nmethods):
6418     _span(span),
6419     _verification_bm(verification_bm),
6420     _cms_bm(cms_bm),
6421     _should_do_nmethods(should_do_nmethods) {
6422     assert(_ref_processor == NULL, "deliberately left NULL");
6423     assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6424 }
6425 
6426 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6427   // if p points into _span, then mark corresponding bit in _markBitMap
6428   assert(obj->is_oop(), "expected an oop");
6429   HeapWord* addr = (HeapWord*)obj;
6430   if (_span.contains(addr)) {
6431     _verification_bm->mark(addr);
6432     if (!_cms_bm->isMarked(addr)) {
6433       oop(addr)->print();
6434       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6435       fatal("... aborting");
6436     }
6437   }
6438 }
6439 
6440 void MarkRefsIntoVerifyClosure::do_oop(oop* p)       { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6441 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6442 
6443 //////////////////////////////////////////////////
6444 // MarkRefsIntoAndScanClosure
6445 //////////////////////////////////////////////////
6446 
6447 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6448                                                        ReferenceProcessor* rp,
6449                                                        CMSBitMap* bit_map,
6450                                                        CMSBitMap* mod_union_table,
6451                                                        CMSMarkStack*  mark_stack,
6452                                                        CMSMarkStack*  revisit_stack,
6453                                                        CMSCollector* collector,
6454                                                        bool should_yield,
6455                                                        bool concurrent_precleaning):
6456   _collector(collector),
6457   _span(span),
6458   _bit_map(bit_map),
6459   _mark_stack(mark_stack),
6460   _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6461                       mark_stack, revisit_stack, concurrent_precleaning),
6462   _yield(should_yield),
6463   _concurrent_precleaning(concurrent_precleaning),
6464   _freelistLock(NULL)
6465 {
6466   _ref_processor = rp;
6467   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6468 }
6469 
6470 // This closure is used to mark refs into the CMS generation at the
6471 // second (final) checkpoint, and to scan and transitively follow
6472 // the unmarked oops. It is also used during the concurrent precleaning
6473 // phase while scanning objects on dirty cards in the CMS generation.
6474 // The marks are made in the marking bit map and the marking stack is
6475 // used for keeping the (newly) grey objects during the scan.
6476 // The parallel version (Par_...) appears further below.
6477 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6478   if (obj != NULL) {
6479     assert(obj->is_oop(), "expected an oop");
6480     HeapWord* addr = (HeapWord*)obj;
6481     assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6482     assert(_collector->overflow_list_is_empty(),
6483            "overflow list should be empty");
6484     if (_span.contains(addr) &&
6485         !_bit_map->isMarked(addr)) {
6486       // mark bit map (object is now grey)
6487       _bit_map->mark(addr);
6488       // push on marking stack (stack should be empty), and drain the
6489       // stack by applying this closure to the oops in the oops popped
6490       // from the stack (i.e. blacken the grey objects)
6491       bool res = _mark_stack->push(obj);
6492       assert(res, "Should have space to push on empty stack");
6493       do {
6494         oop new_oop = _mark_stack->pop();
6495         assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6496         assert(new_oop->is_parsable(), "Found unparsable oop");
6497         assert(_bit_map->isMarked((HeapWord*)new_oop),
6498                "only grey objects on this stack");
6499         // iterate over the oops in this oop, marking and pushing
6500         // the ones in CMS heap (i.e. in _span).
6501         new_oop->oop_iterate(&_pushAndMarkClosure);
6502         // check if it's time to yield
6503         do_yield_check();
6504       } while (!_mark_stack->isEmpty() ||
6505                (!_concurrent_precleaning && take_from_overflow_list()));
6506         // if marking stack is empty, and we are not doing this
6507         // during precleaning, then check the overflow list
6508     }
6509     assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6510     assert(_collector->overflow_list_is_empty(),
6511            "overflow list was drained above");
6512     // We could restore evacuated mark words, if any, used for
6513     // overflow list links here because the overflow list is
6514     // provably empty here. That would reduce the maximum
6515     // size requirements for preserved_{oop,mark}_stack.
6516     // But we'll just postpone it until we are all done
6517     // so we can just stream through.
6518     if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6519       _collector->restore_preserved_marks_if_any();
6520       assert(_collector->no_preserved_marks(), "No preserved marks");
6521     }
6522     assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6523            "All preserved marks should have been restored above");
6524   }
6525 }
6526 
6527 void MarkRefsIntoAndScanClosure::do_oop(oop* p)       { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6528 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6529 
6530 void MarkRefsIntoAndScanClosure::do_yield_work() {
6531   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6532          "CMS thread should hold CMS token");
6533   assert_lock_strong(_freelistLock);
6534   assert_lock_strong(_bit_map->lock());
6535   // relinquish the free_list_lock and bitMaplock()
6536   _bit_map->lock()->unlock();
6537   _freelistLock->unlock();
6538   ConcurrentMarkSweepThread::desynchronize(true);
6539   ConcurrentMarkSweepThread::acknowledge_yield_request();
6540   _collector->stopTimer();
6541   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6542   if (PrintCMSStatistics != 0) {
6543     _collector->incrementYields();
6544   }
6545   _collector->icms_wait();
6546 
6547   // See the comment in coordinator_yield()
6548   for (unsigned i = 0;
6549        i < CMSYieldSleepCount &&
6550        ConcurrentMarkSweepThread::should_yield() &&
6551        !CMSCollector::foregroundGCIsActive();
6552        ++i) {
6553     os::sleep(Thread::current(), 1, false);
6554     ConcurrentMarkSweepThread::acknowledge_yield_request();
6555   }
6556 
6557   ConcurrentMarkSweepThread::synchronize(true);
6558   _freelistLock->lock_without_safepoint_check();
6559   _bit_map->lock()->lock_without_safepoint_check();
6560   _collector->startTimer();
6561 }
6562 
6563 ///////////////////////////////////////////////////////////
6564 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6565 //                                 MarkRefsIntoAndScanClosure
6566 ///////////////////////////////////////////////////////////
6567 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6568   CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6569   CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack*  revisit_stack):
6570   _span(span),
6571   _bit_map(bit_map),
6572   _work_queue(work_queue),
6573   _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6574                        (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6575   _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6576                           revisit_stack)
6577 {
6578   _ref_processor = rp;
6579   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6580 }
6581 
6582 // This closure is used to mark refs into the CMS generation at the
6583 // second (final) checkpoint, and to scan and transitively follow
6584 // the unmarked oops. The marks are made in the marking bit map and
6585 // the work_queue is used for keeping the (newly) grey objects during
6586 // the scan phase whence they are also available for stealing by parallel
6587 // threads. Since the marking bit map is shared, updates are
6588 // synchronized (via CAS).
6589 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6590   if (obj != NULL) {
6591     // Ignore mark word because this could be an already marked oop
6592     // that may be chained at the end of the overflow list.
6593     assert(obj->is_oop(true), "expected an oop");
6594     HeapWord* addr = (HeapWord*)obj;
6595     if (_span.contains(addr) &&
6596         !_bit_map->isMarked(addr)) {
6597       // mark bit map (object will become grey):
6598       // It is possible for several threads to be
6599       // trying to "claim" this object concurrently;
6600       // the unique thread that succeeds in marking the
6601       // object first will do the subsequent push on
6602       // to the work queue (or overflow list).
6603       if (_bit_map->par_mark(addr)) {
6604         // push on work_queue (which may not be empty), and trim the
6605         // queue to an appropriate length by applying this closure to
6606         // the oops in the oops popped from the stack (i.e. blacken the
6607         // grey objects)
6608         bool res = _work_queue->push(obj);
6609         assert(res, "Low water mark should be less than capacity?");
6610         trim_queue(_low_water_mark);
6611       } // Else, another thread claimed the object
6612     }
6613   }
6614 }
6615 
6616 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p)       { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6617 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6618 
6619 // This closure is used to rescan the marked objects on the dirty cards
6620 // in the mod union table and the card table proper.
6621 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6622   oop p, MemRegion mr) {
6623 
6624   size_t size = 0;
6625   HeapWord* addr = (HeapWord*)p;
6626   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6627   assert(_span.contains(addr), "we are scanning the CMS generation");
6628   // check if it's time to yield
6629   if (do_yield_check()) {
6630     // We yielded for some foreground stop-world work,
6631     // and we have been asked to abort this ongoing preclean cycle.
6632     return 0;
6633   }
6634   if (_bitMap->isMarked(addr)) {
6635     // it's marked; is it potentially uninitialized?
6636     if (p->klass_or_null() != NULL) {
6637       if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) {
6638         // Signal precleaning to redirty the card since
6639         // the klass pointer is already installed.
6640         assert(size == 0, "Initial value");
6641       } else {
6642         assert(p->is_parsable(), "must be parsable.");
6643         // an initialized object; ignore mark word in verification below
6644         // since we are running concurrent with mutators
6645         assert(p->is_oop(true), "should be an oop");
6646         if (p->is_objArray()) {
6647           // objArrays are precisely marked; restrict scanning
6648           // to dirty cards only.
6649           size = CompactibleFreeListSpace::adjustObjectSize(
6650                    p->oop_iterate(_scanningClosure, mr));
6651         } else {
6652           // A non-array may have been imprecisely marked; we need
6653           // to scan object in its entirety.
6654           size = CompactibleFreeListSpace::adjustObjectSize(
6655                    p->oop_iterate(_scanningClosure));
6656         }
6657         #ifdef DEBUG
6658           size_t direct_size =
6659             CompactibleFreeListSpace::adjustObjectSize(p->size());
6660           assert(size == direct_size, "Inconsistency in size");
6661           assert(size >= 3, "Necessary for Printezis marks to work");
6662           if (!_bitMap->isMarked(addr+1)) {
6663             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6664           } else {
6665             _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6666             assert(_bitMap->isMarked(addr+size-1),
6667                    "inconsistent Printezis mark");
6668           }
6669         #endif // DEBUG
6670       }
6671     } else {
6672       // an unitialized object
6673       assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6674       HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6675       size = pointer_delta(nextOneAddr + 1, addr);
6676       assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6677              "alignment problem");
6678       // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6679       // will dirty the card when the klass pointer is installed in the
6680       // object (signalling the completion of initialization).
6681     }
6682   } else {
6683     // Either a not yet marked object or an uninitialized object
6684     if (p->klass_or_null() == NULL || !p->is_parsable()) {
6685       // An uninitialized object, skip to the next card, since
6686       // we may not be able to read its P-bits yet.
6687       assert(size == 0, "Initial value");
6688     } else {
6689       // An object not (yet) reached by marking: we merely need to
6690       // compute its size so as to go look at the next block.
6691       assert(p->is_oop(true), "should be an oop");
6692       size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6693     }
6694   }
6695   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6696   return size;
6697 }
6698 
6699 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6700   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6701          "CMS thread should hold CMS token");
6702   assert_lock_strong(_freelistLock);
6703   assert_lock_strong(_bitMap->lock());
6704   // relinquish the free_list_lock and bitMaplock()
6705   _bitMap->lock()->unlock();
6706   _freelistLock->unlock();
6707   ConcurrentMarkSweepThread::desynchronize(true);
6708   ConcurrentMarkSweepThread::acknowledge_yield_request();
6709   _collector->stopTimer();
6710   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6711   if (PrintCMSStatistics != 0) {
6712     _collector->incrementYields();
6713   }
6714   _collector->icms_wait();
6715 
6716   // See the comment in coordinator_yield()
6717   for (unsigned i = 0; i < CMSYieldSleepCount &&
6718                    ConcurrentMarkSweepThread::should_yield() &&
6719                    !CMSCollector::foregroundGCIsActive(); ++i) {
6720     os::sleep(Thread::current(), 1, false);
6721     ConcurrentMarkSweepThread::acknowledge_yield_request();
6722   }
6723 
6724   ConcurrentMarkSweepThread::synchronize(true);
6725   _freelistLock->lock_without_safepoint_check();
6726   _bitMap->lock()->lock_without_safepoint_check();
6727   _collector->startTimer();
6728 }
6729 
6730 
6731 //////////////////////////////////////////////////////////////////
6732 // SurvivorSpacePrecleanClosure
6733 //////////////////////////////////////////////////////////////////
6734 // This (single-threaded) closure is used to preclean the oops in
6735 // the survivor spaces.
6736 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6737 
6738   HeapWord* addr = (HeapWord*)p;
6739   DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6740   assert(!_span.contains(addr), "we are scanning the survivor spaces");
6741   assert(p->klass_or_null() != NULL, "object should be initializd");
6742   assert(p->is_parsable(), "must be parsable.");
6743   // an initialized object; ignore mark word in verification below
6744   // since we are running concurrent with mutators
6745   assert(p->is_oop(true), "should be an oop");
6746   // Note that we do not yield while we iterate over
6747   // the interior oops of p, pushing the relevant ones
6748   // on our marking stack.
6749   size_t size = p->oop_iterate(_scanning_closure);
6750   do_yield_check();
6751   // Observe that below, we do not abandon the preclean
6752   // phase as soon as we should; rather we empty the
6753   // marking stack before returning. This is to satisfy
6754   // some existing assertions. In general, it may be a
6755   // good idea to abort immediately and complete the marking
6756   // from the grey objects at a later time.
6757   while (!_mark_stack->isEmpty()) {
6758     oop new_oop = _mark_stack->pop();
6759     assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6760     assert(new_oop->is_parsable(), "Found unparsable oop");
6761     assert(_bit_map->isMarked((HeapWord*)new_oop),
6762            "only grey objects on this stack");
6763     // iterate over the oops in this oop, marking and pushing
6764     // the ones in CMS heap (i.e. in _span).
6765     new_oop->oop_iterate(_scanning_closure);
6766     // check if it's time to yield
6767     do_yield_check();
6768   }
6769   unsigned int after_count =
6770     GenCollectedHeap::heap()->total_collections();
6771   bool abort = (_before_count != after_count) ||
6772                _collector->should_abort_preclean();
6773   return abort ? 0 : size;
6774 }
6775 
6776 void SurvivorSpacePrecleanClosure::do_yield_work() {
6777   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6778          "CMS thread should hold CMS token");
6779   assert_lock_strong(_bit_map->lock());
6780   // Relinquish the bit map lock
6781   _bit_map->lock()->unlock();
6782   ConcurrentMarkSweepThread::desynchronize(true);
6783   ConcurrentMarkSweepThread::acknowledge_yield_request();
6784   _collector->stopTimer();
6785   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6786   if (PrintCMSStatistics != 0) {
6787     _collector->incrementYields();
6788   }
6789   _collector->icms_wait();
6790 
6791   // See the comment in coordinator_yield()
6792   for (unsigned i = 0; i < CMSYieldSleepCount &&
6793                        ConcurrentMarkSweepThread::should_yield() &&
6794                        !CMSCollector::foregroundGCIsActive(); ++i) {
6795     os::sleep(Thread::current(), 1, false);
6796     ConcurrentMarkSweepThread::acknowledge_yield_request();
6797   }
6798 
6799   ConcurrentMarkSweepThread::synchronize(true);
6800   _bit_map->lock()->lock_without_safepoint_check();
6801   _collector->startTimer();
6802 }
6803 
6804 // This closure is used to rescan the marked objects on the dirty cards
6805 // in the mod union table and the card table proper. In the parallel
6806 // case, although the bitMap is shared, we do a single read so the
6807 // isMarked() query is "safe".
6808 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6809   // Ignore mark word because we are running concurrent with mutators
6810   assert(p->is_oop_or_null(true), "expected an oop or null");
6811   HeapWord* addr = (HeapWord*)p;
6812   assert(_span.contains(addr), "we are scanning the CMS generation");
6813   bool is_obj_array = false;
6814   #ifdef DEBUG
6815     if (!_parallel) {
6816       assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6817       assert(_collector->overflow_list_is_empty(),
6818              "overflow list should be empty");
6819 
6820     }
6821   #endif // DEBUG
6822   if (_bit_map->isMarked(addr)) {
6823     // Obj arrays are precisely marked, non-arrays are not;
6824     // so we scan objArrays precisely and non-arrays in their
6825     // entirety.
6826     if (p->is_objArray()) {
6827       is_obj_array = true;
6828       if (_parallel) {
6829         p->oop_iterate(_par_scan_closure, mr);
6830       } else {
6831         p->oop_iterate(_scan_closure, mr);
6832       }
6833     } else {
6834       if (_parallel) {
6835         p->oop_iterate(_par_scan_closure);
6836       } else {
6837         p->oop_iterate(_scan_closure);
6838       }
6839     }
6840   }
6841   #ifdef DEBUG
6842     if (!_parallel) {
6843       assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6844       assert(_collector->overflow_list_is_empty(),
6845              "overflow list should be empty");
6846 
6847     }
6848   #endif // DEBUG
6849   return is_obj_array;
6850 }
6851 
6852 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6853                         MemRegion span,
6854                         CMSBitMap* bitMap, CMSMarkStack*  markStack,
6855                         CMSMarkStack*  revisitStack,
6856                         bool should_yield, bool verifying):
6857   _collector(collector),
6858   _span(span),
6859   _bitMap(bitMap),
6860   _mut(&collector->_modUnionTable),
6861   _markStack(markStack),
6862   _revisitStack(revisitStack),
6863   _yield(should_yield),
6864   _skipBits(0)
6865 {
6866   assert(_markStack->isEmpty(), "stack should be empty");
6867   _finger = _bitMap->startWord();
6868   _threshold = _finger;
6869   assert(_collector->_restart_addr == NULL, "Sanity check");
6870   assert(_span.contains(_finger), "Out of bounds _finger?");
6871   DEBUG_ONLY(_verifying = verifying;)
6872 }
6873 
6874 void MarkFromRootsClosure::reset(HeapWord* addr) {
6875   assert(_markStack->isEmpty(), "would cause duplicates on stack");
6876   assert(_span.contains(addr), "Out of bounds _finger?");
6877   _finger = addr;
6878   _threshold = (HeapWord*)round_to(
6879                  (intptr_t)_finger, CardTableModRefBS::card_size);
6880 }
6881 
6882 // Should revisit to see if this should be restructured for
6883 // greater efficiency.
6884 bool MarkFromRootsClosure::do_bit(size_t offset) {
6885   if (_skipBits > 0) {
6886     _skipBits--;
6887     return true;
6888   }
6889   // convert offset into a HeapWord*
6890   HeapWord* addr = _bitMap->startWord() + offset;
6891   assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6892          "address out of range");
6893   assert(_bitMap->isMarked(addr), "tautology");
6894   if (_bitMap->isMarked(addr+1)) {
6895     // this is an allocated but not yet initialized object
6896     assert(_skipBits == 0, "tautology");
6897     _skipBits = 2;  // skip next two marked bits ("Printezis-marks")
6898     oop p = oop(addr);
6899     if (p->klass_or_null() == NULL || !p->is_parsable()) {
6900       DEBUG_ONLY(if (!_verifying) {)
6901         // We re-dirty the cards on which this object lies and increase
6902         // the _threshold so that we'll come back to scan this object
6903         // during the preclean or remark phase. (CMSCleanOnEnter)
6904         if (CMSCleanOnEnter) {
6905           size_t sz = _collector->block_size_using_printezis_bits(addr);
6906           HeapWord* end_card_addr   = (HeapWord*)round_to(
6907                                          (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6908           MemRegion redirty_range = MemRegion(addr, end_card_addr);
6909           assert(!redirty_range.is_empty(), "Arithmetical tautology");
6910           // Bump _threshold to end_card_addr; note that
6911           // _threshold cannot possibly exceed end_card_addr, anyhow.
6912           // This prevents future clearing of the card as the scan proceeds
6913           // to the right.
6914           assert(_threshold <= end_card_addr,
6915                  "Because we are just scanning into this object");
6916           if (_threshold < end_card_addr) {
6917             _threshold = end_card_addr;
6918           }
6919           if (p->klass_or_null() != NULL) {
6920             // Redirty the range of cards...
6921             _mut->mark_range(redirty_range);
6922           } // ...else the setting of klass will dirty the card anyway.
6923         }
6924       DEBUG_ONLY(})
6925       return true;
6926     }
6927   }
6928   scanOopsInOop(addr);
6929   return true;
6930 }
6931 
6932 // We take a break if we've been at this for a while,
6933 // so as to avoid monopolizing the locks involved.
6934 void MarkFromRootsClosure::do_yield_work() {
6935   // First give up the locks, then yield, then re-lock
6936   // We should probably use a constructor/destructor idiom to
6937   // do this unlock/lock or modify the MutexUnlocker class to
6938   // serve our purpose. XXX
6939   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6940          "CMS thread should hold CMS token");
6941   assert_lock_strong(_bitMap->lock());
6942   _bitMap->lock()->unlock();
6943   ConcurrentMarkSweepThread::desynchronize(true);
6944   ConcurrentMarkSweepThread::acknowledge_yield_request();
6945   _collector->stopTimer();
6946   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6947   if (PrintCMSStatistics != 0) {
6948     _collector->incrementYields();
6949   }
6950   _collector->icms_wait();
6951 
6952   // See the comment in coordinator_yield()
6953   for (unsigned i = 0; i < CMSYieldSleepCount &&
6954                        ConcurrentMarkSweepThread::should_yield() &&
6955                        !CMSCollector::foregroundGCIsActive(); ++i) {
6956     os::sleep(Thread::current(), 1, false);
6957     ConcurrentMarkSweepThread::acknowledge_yield_request();
6958   }
6959 
6960   ConcurrentMarkSweepThread::synchronize(true);
6961   _bitMap->lock()->lock_without_safepoint_check();
6962   _collector->startTimer();
6963 }
6964 
6965 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6966   assert(_bitMap->isMarked(ptr), "expected bit to be set");
6967   assert(_markStack->isEmpty(),
6968          "should drain stack to limit stack usage");
6969   // convert ptr to an oop preparatory to scanning
6970   oop obj = oop(ptr);
6971   // Ignore mark word in verification below, since we
6972   // may be running concurrent with mutators.
6973   assert(obj->is_oop(true), "should be an oop");
6974   assert(_finger <= ptr, "_finger runneth ahead");
6975   // advance the finger to right end of this object
6976   _finger = ptr + obj->size();
6977   assert(_finger > ptr, "we just incremented it above");
6978   // On large heaps, it may take us some time to get through
6979   // the marking phase (especially if running iCMS). During
6980   // this time it's possible that a lot of mutations have
6981   // accumulated in the card table and the mod union table --
6982   // these mutation records are redundant until we have
6983   // actually traced into the corresponding card.
6984   // Here, we check whether advancing the finger would make
6985   // us cross into a new card, and if so clear corresponding
6986   // cards in the MUT (preclean them in the card-table in the
6987   // future).
6988 
6989   DEBUG_ONLY(if (!_verifying) {)
6990     // The clean-on-enter optimization is disabled by default,
6991     // until we fix 6178663.
6992     if (CMSCleanOnEnter && (_finger > _threshold)) {
6993       // [_threshold, _finger) represents the interval
6994       // of cards to be cleared  in MUT (or precleaned in card table).
6995       // The set of cards to be cleared is all those that overlap
6996       // with the interval [_threshold, _finger); note that
6997       // _threshold is always kept card-aligned but _finger isn't
6998       // always card-aligned.
6999       HeapWord* old_threshold = _threshold;
7000       assert(old_threshold == (HeapWord*)round_to(
7001               (intptr_t)old_threshold, CardTableModRefBS::card_size),
7002              "_threshold should always be card-aligned");
7003       _threshold = (HeapWord*)round_to(
7004                      (intptr_t)_finger, CardTableModRefBS::card_size);
7005       MemRegion mr(old_threshold, _threshold);
7006       assert(!mr.is_empty(), "Control point invariant");
7007       assert(_span.contains(mr), "Should clear within span");
7008       // XXX When _finger crosses from old gen into perm gen
7009       // we may be doing unnecessary cleaning; do better in the
7010       // future by detecting that condition and clearing fewer
7011       // MUT/CT entries.
7012       _mut->clear_range(mr);
7013     }
7014   DEBUG_ONLY(})
7015 
7016   // Note: the finger doesn't advance while we drain
7017   // the stack below.
7018   PushOrMarkClosure pushOrMarkClosure(_collector,
7019                                       _span, _bitMap, _markStack,
7020                                       _revisitStack,
7021                                       _finger, this);
7022   bool res = _markStack->push(obj);
7023   assert(res, "Empty non-zero size stack should have space for single push");
7024   while (!_markStack->isEmpty()) {
7025     oop new_oop = _markStack->pop();
7026     // Skip verifying header mark word below because we are
7027     // running concurrent with mutators.
7028     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7029     // now scan this oop's oops
7030     new_oop->oop_iterate(&pushOrMarkClosure);
7031     do_yield_check();
7032   }
7033   assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7034 }
7035 
7036 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7037                        CMSCollector* collector, MemRegion span,
7038                        CMSBitMap* bit_map,
7039                        OopTaskQueue* work_queue,
7040                        CMSMarkStack*  overflow_stack,
7041                        CMSMarkStack*  revisit_stack,
7042                        bool should_yield):
7043   _collector(collector),
7044   _whole_span(collector->_span),
7045   _span(span),
7046   _bit_map(bit_map),
7047   _mut(&collector->_modUnionTable),
7048   _work_queue(work_queue),
7049   _overflow_stack(overflow_stack),
7050   _revisit_stack(revisit_stack),
7051   _yield(should_yield),
7052   _skip_bits(0),
7053   _task(task)
7054 {
7055   assert(_work_queue->size() == 0, "work_queue should be empty");
7056   _finger = span.start();
7057   _threshold = _finger;     // XXX Defer clear-on-enter optimization for now
7058   assert(_span.contains(_finger), "Out of bounds _finger?");
7059 }
7060 
7061 // Should revisit to see if this should be restructured for
7062 // greater efficiency.
7063 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7064   if (_skip_bits > 0) {
7065     _skip_bits--;
7066     return true;
7067   }
7068   // convert offset into a HeapWord*
7069   HeapWord* addr = _bit_map->startWord() + offset;
7070   assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7071          "address out of range");
7072   assert(_bit_map->isMarked(addr), "tautology");
7073   if (_bit_map->isMarked(addr+1)) {
7074     // this is an allocated object that might not yet be initialized
7075     assert(_skip_bits == 0, "tautology");
7076     _skip_bits = 2;  // skip next two marked bits ("Printezis-marks")
7077     oop p = oop(addr);
7078     if (p->klass_or_null() == NULL || !p->is_parsable()) {
7079       // in the case of Clean-on-Enter optimization, redirty card
7080       // and avoid clearing card by increasing  the threshold.
7081       return true;
7082     }
7083   }
7084   scan_oops_in_oop(addr);
7085   return true;
7086 }
7087 
7088 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7089   assert(_bit_map->isMarked(ptr), "expected bit to be set");
7090   // Should we assert that our work queue is empty or
7091   // below some drain limit?
7092   assert(_work_queue->size() == 0,
7093          "should drain stack to limit stack usage");
7094   // convert ptr to an oop preparatory to scanning
7095   oop obj = oop(ptr);
7096   // Ignore mark word in verification below, since we
7097   // may be running concurrent with mutators.
7098   assert(obj->is_oop(true), "should be an oop");
7099   assert(_finger <= ptr, "_finger runneth ahead");
7100   // advance the finger to right end of this object
7101   _finger = ptr + obj->size();
7102   assert(_finger > ptr, "we just incremented it above");
7103   // On large heaps, it may take us some time to get through
7104   // the marking phase (especially if running iCMS). During
7105   // this time it's possible that a lot of mutations have
7106   // accumulated in the card table and the mod union table --
7107   // these mutation records are redundant until we have
7108   // actually traced into the corresponding card.
7109   // Here, we check whether advancing the finger would make
7110   // us cross into a new card, and if so clear corresponding
7111   // cards in the MUT (preclean them in the card-table in the
7112   // future).
7113 
7114   // The clean-on-enter optimization is disabled by default,
7115   // until we fix 6178663.
7116   if (CMSCleanOnEnter && (_finger > _threshold)) {
7117     // [_threshold, _finger) represents the interval
7118     // of cards to be cleared  in MUT (or precleaned in card table).
7119     // The set of cards to be cleared is all those that overlap
7120     // with the interval [_threshold, _finger); note that
7121     // _threshold is always kept card-aligned but _finger isn't
7122     // always card-aligned.
7123     HeapWord* old_threshold = _threshold;
7124     assert(old_threshold == (HeapWord*)round_to(
7125             (intptr_t)old_threshold, CardTableModRefBS::card_size),
7126            "_threshold should always be card-aligned");
7127     _threshold = (HeapWord*)round_to(
7128                    (intptr_t)_finger, CardTableModRefBS::card_size);
7129     MemRegion mr(old_threshold, _threshold);
7130     assert(!mr.is_empty(), "Control point invariant");
7131     assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7132     // XXX When _finger crosses from old gen into perm gen
7133     // we may be doing unnecessary cleaning; do better in the
7134     // future by detecting that condition and clearing fewer
7135     // MUT/CT entries.
7136     _mut->clear_range(mr);
7137   }
7138 
7139   // Note: the local finger doesn't advance while we drain
7140   // the stack below, but the global finger sure can and will.
7141   HeapWord** gfa = _task->global_finger_addr();
7142   Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7143                                       _span, _bit_map,
7144                                       _work_queue,
7145                                       _overflow_stack,
7146                                       _revisit_stack,
7147                                       _finger,
7148                                       gfa, this);
7149   bool res = _work_queue->push(obj);   // overflow could occur here
7150   assert(res, "Will hold once we use workqueues");
7151   while (true) {
7152     oop new_oop;
7153     if (!_work_queue->pop_local(new_oop)) {
7154       // We emptied our work_queue; check if there's stuff that can
7155       // be gotten from the overflow stack.
7156       if (CMSConcMarkingTask::get_work_from_overflow_stack(
7157             _overflow_stack, _work_queue)) {
7158         do_yield_check();
7159         continue;
7160       } else {  // done
7161         break;
7162       }
7163     }
7164     // Skip verifying header mark word below because we are
7165     // running concurrent with mutators.
7166     assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7167     // now scan this oop's oops
7168     new_oop->oop_iterate(&pushOrMarkClosure);
7169     do_yield_check();
7170   }
7171   assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7172 }
7173 
7174 // Yield in response to a request from VM Thread or
7175 // from mutators.
7176 void Par_MarkFromRootsClosure::do_yield_work() {
7177   assert(_task != NULL, "sanity");
7178   _task->yield();
7179 }
7180 
7181 // A variant of the above used for verifying CMS marking work.
7182 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7183                         MemRegion span,
7184                         CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7185                         CMSMarkStack*  mark_stack):
7186   _collector(collector),
7187   _span(span),
7188   _verification_bm(verification_bm),
7189   _cms_bm(cms_bm),
7190   _mark_stack(mark_stack),
7191   _pam_verify_closure(collector, span, verification_bm, cms_bm,
7192                       mark_stack)
7193 {
7194   assert(_mark_stack->isEmpty(), "stack should be empty");
7195   _finger = _verification_bm->startWord();
7196   assert(_collector->_restart_addr == NULL, "Sanity check");
7197   assert(_span.contains(_finger), "Out of bounds _finger?");
7198 }
7199 
7200 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7201   assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7202   assert(_span.contains(addr), "Out of bounds _finger?");
7203   _finger = addr;
7204 }
7205 
7206 // Should revisit to see if this should be restructured for
7207 // greater efficiency.
7208 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7209   // convert offset into a HeapWord*
7210   HeapWord* addr = _verification_bm->startWord() + offset;
7211   assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7212          "address out of range");
7213   assert(_verification_bm->isMarked(addr), "tautology");
7214   assert(_cms_bm->isMarked(addr), "tautology");
7215 
7216   assert(_mark_stack->isEmpty(),
7217          "should drain stack to limit stack usage");
7218   // convert addr to an oop preparatory to scanning
7219   oop obj = oop(addr);
7220   assert(obj->is_oop(), "should be an oop");
7221   assert(_finger <= addr, "_finger runneth ahead");
7222   // advance the finger to right end of this object
7223   _finger = addr + obj->size();
7224   assert(_finger > addr, "we just incremented it above");
7225   // Note: the finger doesn't advance while we drain
7226   // the stack below.
7227   bool res = _mark_stack->push(obj);
7228   assert(res, "Empty non-zero size stack should have space for single push");
7229   while (!_mark_stack->isEmpty()) {
7230     oop new_oop = _mark_stack->pop();
7231     assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7232     // now scan this oop's oops
7233     new_oop->oop_iterate(&_pam_verify_closure);
7234   }
7235   assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7236   return true;
7237 }
7238 
7239 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7240   CMSCollector* collector, MemRegion span,
7241   CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7242   CMSMarkStack*  mark_stack):
7243   OopClosure(collector->ref_processor()),
7244   _collector(collector),
7245   _span(span),
7246   _verification_bm(verification_bm),
7247   _cms_bm(cms_bm),
7248   _mark_stack(mark_stack)
7249 { }
7250 
7251 void PushAndMarkVerifyClosure::do_oop(oop* p)       { PushAndMarkVerifyClosure::do_oop_work(p); }
7252 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7253 
7254 // Upon stack overflow, we discard (part of) the stack,
7255 // remembering the least address amongst those discarded
7256 // in CMSCollector's _restart_address.
7257 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7258   // Remember the least grey address discarded
7259   HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7260   _collector->lower_restart_addr(ra);
7261   _mark_stack->reset();  // discard stack contents
7262   _mark_stack->expand(); // expand the stack if possible
7263 }
7264 
7265 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7266   assert(obj->is_oop_or_null(), "expected an oop or NULL");
7267   HeapWord* addr = (HeapWord*)obj;
7268   if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7269     // Oop lies in _span and isn't yet grey or black
7270     _verification_bm->mark(addr);            // now grey
7271     if (!_cms_bm->isMarked(addr)) {
7272       oop(addr)->print();
7273       gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7274                              addr);
7275       fatal("... aborting");
7276     }
7277 
7278     if (!_mark_stack->push(obj)) { // stack overflow
7279       if (PrintCMSStatistics != 0) {
7280         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7281                                SIZE_FORMAT, _mark_stack->capacity());
7282       }
7283       assert(_mark_stack->isFull(), "Else push should have succeeded");
7284       handle_stack_overflow(addr);
7285     }
7286     // anything including and to the right of _finger
7287     // will be scanned as we iterate over the remainder of the
7288     // bit map
7289   }
7290 }
7291 
7292 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7293                      MemRegion span,
7294                      CMSBitMap* bitMap, CMSMarkStack*  markStack,
7295                      CMSMarkStack*  revisitStack,
7296                      HeapWord* finger, MarkFromRootsClosure* parent) :
7297   OopClosure(collector->ref_processor()),
7298   _collector(collector),
7299   _span(span),
7300   _bitMap(bitMap),
7301   _markStack(markStack),
7302   _revisitStack(revisitStack),
7303   _finger(finger),
7304   _parent(parent),
7305   _should_remember_klasses(collector->should_unload_classes())
7306 { }
7307 
7308 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7309                      MemRegion span,
7310                      CMSBitMap* bit_map,
7311                      OopTaskQueue* work_queue,
7312                      CMSMarkStack*  overflow_stack,
7313                      CMSMarkStack*  revisit_stack,
7314                      HeapWord* finger,
7315                      HeapWord** global_finger_addr,
7316                      Par_MarkFromRootsClosure* parent) :
7317   OopClosure(collector->ref_processor()),
7318   _collector(collector),
7319   _whole_span(collector->_span),
7320   _span(span),
7321   _bit_map(bit_map),
7322   _work_queue(work_queue),
7323   _overflow_stack(overflow_stack),
7324   _revisit_stack(revisit_stack),
7325   _finger(finger),
7326   _global_finger_addr(global_finger_addr),
7327   _parent(parent),
7328   _should_remember_klasses(collector->should_unload_classes())
7329 { }
7330 
7331 // Assumes thread-safe access by callers, who are
7332 // responsible for mutual exclusion.
7333 void CMSCollector::lower_restart_addr(HeapWord* low) {
7334   assert(_span.contains(low), "Out of bounds addr");
7335   if (_restart_addr == NULL) {
7336     _restart_addr = low;
7337   } else {
7338     _restart_addr = MIN2(_restart_addr, low);
7339   }
7340 }
7341 
7342 // Upon stack overflow, we discard (part of) the stack,
7343 // remembering the least address amongst those discarded
7344 // in CMSCollector's _restart_address.
7345 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7346   // Remember the least grey address discarded
7347   HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7348   _collector->lower_restart_addr(ra);
7349   _markStack->reset();  // discard stack contents
7350   _markStack->expand(); // expand the stack if possible
7351 }
7352 
7353 // Upon stack overflow, we discard (part of) the stack,
7354 // remembering the least address amongst those discarded
7355 // in CMSCollector's _restart_address.
7356 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7357   // We need to do this under a mutex to prevent other
7358   // workers from interfering with the work done below.
7359   MutexLockerEx ml(_overflow_stack->par_lock(),
7360                    Mutex::_no_safepoint_check_flag);
7361   // Remember the least grey address discarded
7362   HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7363   _collector->lower_restart_addr(ra);
7364   _overflow_stack->reset();  // discard stack contents
7365   _overflow_stack->expand(); // expand the stack if possible
7366 }
7367 
7368 void PushOrMarkClosure::do_oop(oop obj) {
7369   // Ignore mark word because we are running concurrent with mutators.
7370   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7371   HeapWord* addr = (HeapWord*)obj;
7372   if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7373     // Oop lies in _span and isn't yet grey or black
7374     _bitMap->mark(addr);            // now grey
7375     if (addr < _finger) {
7376       // the bit map iteration has already either passed, or
7377       // sampled, this bit in the bit map; we'll need to
7378       // use the marking stack to scan this oop's oops.
7379       bool simulate_overflow = false;
7380       NOT_PRODUCT(
7381         if (CMSMarkStackOverflowALot &&
7382             _collector->simulate_overflow()) {
7383           // simulate a stack overflow
7384           simulate_overflow = true;
7385         }
7386       )
7387       if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7388         if (PrintCMSStatistics != 0) {
7389           gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7390                                  SIZE_FORMAT, _markStack->capacity());
7391         }
7392         assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7393         handle_stack_overflow(addr);
7394       }
7395     }
7396     // anything including and to the right of _finger
7397     // will be scanned as we iterate over the remainder of the
7398     // bit map
7399     do_yield_check();
7400   }
7401 }
7402 
7403 void PushOrMarkClosure::do_oop(oop* p)       { PushOrMarkClosure::do_oop_work(p); }
7404 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7405 
7406 void Par_PushOrMarkClosure::do_oop(oop obj) {
7407   // Ignore mark word because we are running concurrent with mutators.
7408   assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7409   HeapWord* addr = (HeapWord*)obj;
7410   if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7411     // Oop lies in _span and isn't yet grey or black
7412     // We read the global_finger (volatile read) strictly after marking oop
7413     bool res = _bit_map->par_mark(addr);    // now grey
7414     volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7415     // Should we push this marked oop on our stack?
7416     // -- if someone else marked it, nothing to do
7417     // -- if target oop is above global finger nothing to do
7418     // -- if target oop is in chunk and above local finger
7419     //      then nothing to do
7420     // -- else push on work queue
7421     if (   !res       // someone else marked it, they will deal with it
7422         || (addr >= *gfa)  // will be scanned in a later task
7423         || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7424       return;
7425     }
7426     // the bit map iteration has already either passed, or
7427     // sampled, this bit in the bit map; we'll need to
7428     // use the marking stack to scan this oop's oops.
7429     bool simulate_overflow = false;
7430     NOT_PRODUCT(
7431       if (CMSMarkStackOverflowALot &&
7432           _collector->simulate_overflow()) {
7433         // simulate a stack overflow
7434         simulate_overflow = true;
7435       }
7436     )
7437     if (simulate_overflow ||
7438         !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7439       // stack overflow
7440       if (PrintCMSStatistics != 0) {
7441         gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7442                                SIZE_FORMAT, _overflow_stack->capacity());
7443       }
7444       // We cannot assert that the overflow stack is full because
7445       // it may have been emptied since.
7446       assert(simulate_overflow ||
7447              _work_queue->size() == _work_queue->max_elems(),
7448             "Else push should have succeeded");
7449       handle_stack_overflow(addr);
7450     }
7451     do_yield_check();
7452   }
7453 }
7454 
7455 void Par_PushOrMarkClosure::do_oop(oop* p)       { Par_PushOrMarkClosure::do_oop_work(p); }
7456 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7457 
7458 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7459                                        MemRegion span,
7460                                        ReferenceProcessor* rp,
7461                                        CMSBitMap* bit_map,
7462                                        CMSBitMap* mod_union_table,
7463                                        CMSMarkStack*  mark_stack,
7464                                        CMSMarkStack*  revisit_stack,
7465                                        bool           concurrent_precleaning):
7466   OopClosure(rp),
7467   _collector(collector),
7468   _span(span),
7469   _bit_map(bit_map),
7470   _mod_union_table(mod_union_table),
7471   _mark_stack(mark_stack),
7472   _revisit_stack(revisit_stack),
7473   _concurrent_precleaning(concurrent_precleaning),
7474   _should_remember_klasses(collector->should_unload_classes())
7475 {
7476   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7477 }
7478 
7479 // Grey object rescan during pre-cleaning and second checkpoint phases --
7480 // the non-parallel version (the parallel version appears further below.)
7481 void PushAndMarkClosure::do_oop(oop obj) {
7482   // Ignore mark word verification. If during concurrent precleaning,
7483   // the object monitor may be locked. If during the checkpoint
7484   // phases, the object may already have been reached by a  different
7485   // path and may be at the end of the global overflow list (so
7486   // the mark word may be NULL).
7487   assert(obj->is_oop_or_null(true /* ignore mark word */),
7488          "expected an oop or NULL");
7489   HeapWord* addr = (HeapWord*)obj;
7490   // Check if oop points into the CMS generation
7491   // and is not marked
7492   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7493     // a white object ...
7494     _bit_map->mark(addr);         // ... now grey
7495     // push on the marking stack (grey set)
7496     bool simulate_overflow = false;
7497     NOT_PRODUCT(
7498       if (CMSMarkStackOverflowALot &&
7499           _collector->simulate_overflow()) {
7500         // simulate a stack overflow
7501         simulate_overflow = true;
7502       }
7503     )
7504     if (simulate_overflow || !_mark_stack->push(obj)) {
7505       if (_concurrent_precleaning) {
7506          // During precleaning we can just dirty the appropriate card(s)
7507          // in the mod union table, thus ensuring that the object remains
7508          // in the grey set  and continue. In the case of object arrays
7509          // we need to dirty all of the cards that the object spans,
7510          // since the rescan of object arrays will be limited to the
7511          // dirty cards.
7512          // Note that no one can be intefering with us in this action
7513          // of dirtying the mod union table, so no locking or atomics
7514          // are required.
7515          if (obj->is_objArray()) {
7516            size_t sz = obj->size();
7517            HeapWord* end_card_addr = (HeapWord*)round_to(
7518                                         (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7519            MemRegion redirty_range = MemRegion(addr, end_card_addr);
7520            assert(!redirty_range.is_empty(), "Arithmetical tautology");
7521            _mod_union_table->mark_range(redirty_range);
7522          } else {
7523            _mod_union_table->mark(addr);
7524          }
7525          _collector->_ser_pmc_preclean_ovflw++;
7526       } else {
7527          // During the remark phase, we need to remember this oop
7528          // in the overflow list.
7529          _collector->push_on_overflow_list(obj);
7530          _collector->_ser_pmc_remark_ovflw++;
7531       }
7532     }
7533   }
7534 }
7535 
7536 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7537                                                MemRegion span,
7538                                                ReferenceProcessor* rp,
7539                                                CMSBitMap* bit_map,
7540                                                OopTaskQueue* work_queue,
7541                                                CMSMarkStack* revisit_stack):
7542   OopClosure(rp),
7543   _collector(collector),
7544   _span(span),
7545   _bit_map(bit_map),
7546   _work_queue(work_queue),
7547   _revisit_stack(revisit_stack),
7548   _should_remember_klasses(collector->should_unload_classes())
7549 {
7550   assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7551 }
7552 
7553 void PushAndMarkClosure::do_oop(oop* p)       { PushAndMarkClosure::do_oop_work(p); }
7554 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7555 
7556 // Grey object rescan during second checkpoint phase --
7557 // the parallel version.
7558 void Par_PushAndMarkClosure::do_oop(oop obj) {
7559   // In the assert below, we ignore the mark word because
7560   // this oop may point to an already visited object that is
7561   // on the overflow stack (in which case the mark word has
7562   // been hijacked for chaining into the overflow stack --
7563   // if this is the last object in the overflow stack then
7564   // its mark word will be NULL). Because this object may
7565   // have been subsequently popped off the global overflow
7566   // stack, and the mark word possibly restored to the prototypical
7567   // value, by the time we get to examined this failing assert in
7568   // the debugger, is_oop_or_null(false) may subsequently start
7569   // to hold.
7570   assert(obj->is_oop_or_null(true),
7571          "expected an oop or NULL");
7572   HeapWord* addr = (HeapWord*)obj;
7573   // Check if oop points into the CMS generation
7574   // and is not marked
7575   if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7576     // a white object ...
7577     // If we manage to "claim" the object, by being the
7578     // first thread to mark it, then we push it on our
7579     // marking stack
7580     if (_bit_map->par_mark(addr)) {     // ... now grey
7581       // push on work queue (grey set)
7582       bool simulate_overflow = false;
7583       NOT_PRODUCT(
7584         if (CMSMarkStackOverflowALot &&
7585             _collector->par_simulate_overflow()) {
7586           // simulate a stack overflow
7587           simulate_overflow = true;
7588         }
7589       )
7590       if (simulate_overflow || !_work_queue->push(obj)) {
7591         _collector->par_push_on_overflow_list(obj);
7592         _collector->_par_pmc_remark_ovflw++; //  imprecise OK: no need to CAS
7593       }
7594     } // Else, some other thread got there first
7595   }
7596 }
7597 
7598 void Par_PushAndMarkClosure::do_oop(oop* p)       { Par_PushAndMarkClosure::do_oop_work(p); }
7599 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7600 
7601 void PushAndMarkClosure::remember_klass(Klass* k) {
7602   if (!_revisit_stack->push(oop(k))) {
7603     fatal("Revisit stack overflowed in PushAndMarkClosure");
7604   }
7605 }
7606 
7607 void Par_PushAndMarkClosure::remember_klass(Klass* k) {
7608   if (!_revisit_stack->par_push(oop(k))) {
7609     fatal("Revist stack overflowed in Par_PushAndMarkClosure");
7610   }
7611 }
7612 
7613 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7614   Mutex* bml = _collector->bitMapLock();
7615   assert_lock_strong(bml);
7616   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7617          "CMS thread should hold CMS token");
7618 
7619   bml->unlock();
7620   ConcurrentMarkSweepThread::desynchronize(true);
7621 
7622   ConcurrentMarkSweepThread::acknowledge_yield_request();
7623 
7624   _collector->stopTimer();
7625   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7626   if (PrintCMSStatistics != 0) {
7627     _collector->incrementYields();
7628   }
7629   _collector->icms_wait();
7630 
7631   // See the comment in coordinator_yield()
7632   for (unsigned i = 0; i < CMSYieldSleepCount &&
7633                        ConcurrentMarkSweepThread::should_yield() &&
7634                        !CMSCollector::foregroundGCIsActive(); ++i) {
7635     os::sleep(Thread::current(), 1, false);
7636     ConcurrentMarkSweepThread::acknowledge_yield_request();
7637   }
7638 
7639   ConcurrentMarkSweepThread::synchronize(true);
7640   bml->lock();
7641 
7642   _collector->startTimer();
7643 }
7644 
7645 bool CMSPrecleanRefsYieldClosure::should_return() {
7646   if (ConcurrentMarkSweepThread::should_yield()) {
7647     do_yield_work();
7648   }
7649   return _collector->foregroundGCIsActive();
7650 }
7651 
7652 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7653   assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7654          "mr should be aligned to start at a card boundary");
7655   // We'd like to assert:
7656   // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7657   //        "mr should be a range of cards");
7658   // However, that would be too strong in one case -- the last
7659   // partition ends at _unallocated_block which, in general, can be
7660   // an arbitrary boundary, not necessarily card aligned.
7661   if (PrintCMSStatistics != 0) {
7662     _num_dirty_cards +=
7663          mr.word_size()/CardTableModRefBS::card_size_in_words;
7664   }
7665   _space->object_iterate_mem(mr, &_scan_cl);
7666 }
7667 
7668 SweepClosure::SweepClosure(CMSCollector* collector,
7669                            ConcurrentMarkSweepGeneration* g,
7670                            CMSBitMap* bitMap, bool should_yield) :
7671   _collector(collector),
7672   _g(g),
7673   _sp(g->cmsSpace()),
7674   _limit(_sp->sweep_limit()),
7675   _freelistLock(_sp->freelistLock()),
7676   _bitMap(bitMap),
7677   _yield(should_yield),
7678   _inFreeRange(false),           // No free range at beginning of sweep
7679   _freeRangeInFreeLists(false),  // No free range at beginning of sweep
7680   _lastFreeRangeCoalesced(false),
7681   _freeFinger(g->used_region().start())
7682 {
7683   NOT_PRODUCT(
7684     _numObjectsFreed = 0;
7685     _numWordsFreed   = 0;
7686     _numObjectsLive = 0;
7687     _numWordsLive = 0;
7688     _numObjectsAlreadyFree = 0;
7689     _numWordsAlreadyFree = 0;
7690     _last_fc = NULL;
7691 
7692     _sp->initializeIndexedFreeListArrayReturnedBytes();
7693     _sp->dictionary()->initializeDictReturnedBytes();
7694   )
7695   assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7696          "sweep _limit out of bounds");
7697   if (CMSTraceSweeper) {
7698     gclog_or_tty->print("\n====================\nStarting new sweep\n");
7699   }
7700 }
7701 
7702 // We need this destructor to reclaim any space at the end
7703 // of the space, which do_blk below may not have added back to
7704 // the free lists. [basically dealing with the "fringe effect"]
7705 SweepClosure::~SweepClosure() {
7706   assert_lock_strong(_freelistLock);
7707   // this should be treated as the end of a free run if any
7708   // The current free range should be returned to the free lists
7709   // as one coalesced chunk.
7710   if (inFreeRange()) {
7711     flushCurFreeChunk(freeFinger(),
7712       pointer_delta(_limit, freeFinger()));
7713     assert(freeFinger() < _limit, "the finger pointeth off base");
7714     if (CMSTraceSweeper) {
7715       gclog_or_tty->print("destructor:");
7716       gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
7717                  "[coalesced:"SIZE_FORMAT"]\n",
7718                  freeFinger(), pointer_delta(_limit, freeFinger()),
7719                  lastFreeRangeCoalesced());
7720     }
7721   }
7722   NOT_PRODUCT(
7723     if (Verbose && PrintGC) {
7724       gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7725                           SIZE_FORMAT " bytes",
7726                  _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7727       gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects,  "
7728                              SIZE_FORMAT" bytes  "
7729         "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7730         _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7731         _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7732       size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7733         sizeof(HeapWord);
7734       gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7735 
7736       if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7737         size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7738         size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7739         size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7740         gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7741         gclog_or_tty->print("   Indexed List Returned "SIZE_FORMAT" bytes",
7742           indexListReturnedBytes);
7743         gclog_or_tty->print_cr("        Dictionary Returned "SIZE_FORMAT" bytes",
7744           dictReturnedBytes);
7745       }
7746     }
7747   )
7748   // Now, in debug mode, just null out the sweep_limit
7749   NOT_PRODUCT(_sp->clear_sweep_limit();)
7750   if (CMSTraceSweeper) {
7751     gclog_or_tty->print("end of sweep\n================\n");
7752   }
7753 }
7754 
7755 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7756     bool freeRangeInFreeLists) {
7757   if (CMSTraceSweeper) {
7758     gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
7759                freeFinger, _sp->block_size(freeFinger),
7760                freeRangeInFreeLists);
7761   }
7762   assert(!inFreeRange(), "Trampling existing free range");
7763   set_inFreeRange(true);
7764   set_lastFreeRangeCoalesced(false);
7765 
7766   set_freeFinger(freeFinger);
7767   set_freeRangeInFreeLists(freeRangeInFreeLists);
7768   if (CMSTestInFreeList) {
7769     if (freeRangeInFreeLists) {
7770       FreeChunk* fc = (FreeChunk*) freeFinger;
7771       assert(fc->isFree(), "A chunk on the free list should be free.");
7772       assert(fc->size() > 0, "Free range should have a size");
7773       assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7774     }
7775   }
7776 }
7777 
7778 // Note that the sweeper runs concurrently with mutators. Thus,
7779 // it is possible for direct allocation in this generation to happen
7780 // in the middle of the sweep. Note that the sweeper also coalesces
7781 // contiguous free blocks. Thus, unless the sweeper and the allocator
7782 // synchronize appropriately freshly allocated blocks may get swept up.
7783 // This is accomplished by the sweeper locking the free lists while
7784 // it is sweeping. Thus blocks that are determined to be free are
7785 // indeed free. There is however one additional complication:
7786 // blocks that have been allocated since the final checkpoint and
7787 // mark, will not have been marked and so would be treated as
7788 // unreachable and swept up. To prevent this, the allocator marks
7789 // the bit map when allocating during the sweep phase. This leads,
7790 // however, to a further complication -- objects may have been allocated
7791 // but not yet initialized -- in the sense that the header isn't yet
7792 // installed. The sweeper can not then determine the size of the block
7793 // in order to skip over it. To deal with this case, we use a technique
7794 // (due to Printezis) to encode such uninitialized block sizes in the
7795 // bit map. Since the bit map uses a bit per every HeapWord, but the
7796 // CMS generation has a minimum object size of 3 HeapWords, it follows
7797 // that "normal marks" won't be adjacent in the bit map (there will
7798 // always be at least two 0 bits between successive 1 bits). We make use
7799 // of these "unused" bits to represent uninitialized blocks -- the bit
7800 // corresponding to the start of the uninitialized object and the next
7801 // bit are both set. Finally, a 1 bit marks the end of the object that
7802 // started with the two consecutive 1 bits to indicate its potentially
7803 // uninitialized state.
7804 
7805 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7806   FreeChunk* fc = (FreeChunk*)addr;
7807   size_t res;
7808 
7809   // check if we are done sweepinrg
7810   if (addr == _limit) { // we have swept up to the limit, do nothing more
7811     assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7812            "sweep _limit out of bounds");
7813     // help the closure application finish
7814     return pointer_delta(_sp->end(), _limit);
7815   }
7816   assert(addr <= _limit, "sweep invariant");
7817 
7818   // check if we should yield
7819   do_yield_check(addr);
7820   if (fc->isFree()) {
7821     // Chunk that is already free
7822     res = fc->size();
7823     doAlreadyFreeChunk(fc);
7824     debug_only(_sp->verifyFreeLists());
7825     assert(res == fc->size(), "Don't expect the size to change");
7826     NOT_PRODUCT(
7827       _numObjectsAlreadyFree++;
7828       _numWordsAlreadyFree += res;
7829     )
7830     NOT_PRODUCT(_last_fc = fc;)
7831   } else if (!_bitMap->isMarked(addr)) {
7832     // Chunk is fresh garbage
7833     res = doGarbageChunk(fc);
7834     debug_only(_sp->verifyFreeLists());
7835     NOT_PRODUCT(
7836       _numObjectsFreed++;
7837       _numWordsFreed += res;
7838     )
7839   } else {
7840     // Chunk that is alive.
7841     res = doLiveChunk(fc);
7842     debug_only(_sp->verifyFreeLists());
7843     NOT_PRODUCT(
7844         _numObjectsLive++;
7845         _numWordsLive += res;
7846     )
7847   }
7848   return res;
7849 }
7850 
7851 // For the smart allocation, record following
7852 //  split deaths - a free chunk is removed from its free list because
7853 //      it is being split into two or more chunks.
7854 //  split birth - a free chunk is being added to its free list because
7855 //      a larger free chunk has been split and resulted in this free chunk.
7856 //  coal death - a free chunk is being removed from its free list because
7857 //      it is being coalesced into a large free chunk.
7858 //  coal birth - a free chunk is being added to its free list because
7859 //      it was created when two or more free chunks where coalesced into
7860 //      this free chunk.
7861 //
7862 // These statistics are used to determine the desired number of free
7863 // chunks of a given size.  The desired number is chosen to be relative
7864 // to the end of a CMS sweep.  The desired number at the end of a sweep
7865 // is the
7866 //      count-at-end-of-previous-sweep (an amount that was enough)
7867 //              - count-at-beginning-of-current-sweep  (the excess)
7868 //              + split-births  (gains in this size during interval)
7869 //              - split-deaths  (demands on this size during interval)
7870 // where the interval is from the end of one sweep to the end of the
7871 // next.
7872 //
7873 // When sweeping the sweeper maintains an accumulated chunk which is
7874 // the chunk that is made up of chunks that have been coalesced.  That
7875 // will be termed the left-hand chunk.  A new chunk of garbage that
7876 // is being considered for coalescing will be referred to as the
7877 // right-hand chunk.
7878 //
7879 // When making a decision on whether to coalesce a right-hand chunk with
7880 // the current left-hand chunk, the current count vs. the desired count
7881 // of the left-hand chunk is considered.  Also if the right-hand chunk
7882 // is near the large chunk at the end of the heap (see
7883 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7884 // left-hand chunk is coalesced.
7885 //
7886 // When making a decision about whether to split a chunk, the desired count
7887 // vs. the current count of the candidate to be split is also considered.
7888 // If the candidate is underpopulated (currently fewer chunks than desired)
7889 // a chunk of an overpopulated (currently more chunks than desired) size may
7890 // be chosen.  The "hint" associated with a free list, if non-null, points
7891 // to a free list which may be overpopulated.
7892 //
7893 
7894 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
7895   size_t size = fc->size();
7896   // Chunks that cannot be coalesced are not in the
7897   // free lists.
7898   if (CMSTestInFreeList && !fc->cantCoalesce()) {
7899     assert(_sp->verifyChunkInFreeLists(fc),
7900       "free chunk should be in free lists");
7901   }
7902   // a chunk that is already free, should not have been
7903   // marked in the bit map
7904   HeapWord* addr = (HeapWord*) fc;
7905   assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7906   // Verify that the bit map has no bits marked between
7907   // addr and purported end of this block.
7908   _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7909 
7910   // Some chunks cannot be coalesced in under any circumstances.
7911   // See the definition of cantCoalesce().
7912   if (!fc->cantCoalesce()) {
7913     // This chunk can potentially be coalesced.
7914     if (_sp->adaptive_freelists()) {
7915       // All the work is done in
7916       doPostIsFreeOrGarbageChunk(fc, size);
7917     } else {  // Not adaptive free lists
7918       // this is a free chunk that can potentially be coalesced by the sweeper;
7919       if (!inFreeRange()) {
7920         // if the next chunk is a free block that can't be coalesced
7921         // it doesn't make sense to remove this chunk from the free lists
7922         FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7923         assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
7924         if ((HeapWord*)nextChunk < _limit  &&    // there's a next chunk...
7925             nextChunk->isFree()    &&            // which is free...
7926             nextChunk->cantCoalesce()) {         // ... but cant be coalesced
7927           // nothing to do
7928         } else {
7929           // Potentially the start of a new free range:
7930           // Don't eagerly remove it from the free lists.
7931           // No need to remove it if it will just be put
7932           // back again.  (Also from a pragmatic point of view
7933           // if it is a free block in a region that is beyond
7934           // any allocated blocks, an assertion will fail)
7935           // Remember the start of a free run.
7936           initialize_free_range(addr, true);
7937           // end - can coalesce with next chunk
7938         }
7939       } else {
7940         // the midst of a free range, we are coalescing
7941         debug_only(record_free_block_coalesced(fc);)
7942         if (CMSTraceSweeper) {
7943           gclog_or_tty->print("  -- pick up free block 0x%x (%d)\n", fc, size);
7944         }
7945         // remove it from the free lists
7946         _sp->removeFreeChunkFromFreeLists(fc);
7947         set_lastFreeRangeCoalesced(true);
7948         // If the chunk is being coalesced and the current free range is
7949         // in the free lists, remove the current free range so that it
7950         // will be returned to the free lists in its entirety - all
7951         // the coalesced pieces included.
7952         if (freeRangeInFreeLists()) {
7953           FreeChunk* ffc = (FreeChunk*) freeFinger();
7954           assert(ffc->size() == pointer_delta(addr, freeFinger()),
7955             "Size of free range is inconsistent with chunk size.");
7956           if (CMSTestInFreeList) {
7957             assert(_sp->verifyChunkInFreeLists(ffc),
7958               "free range is not in free lists");
7959           }
7960           _sp->removeFreeChunkFromFreeLists(ffc);
7961           set_freeRangeInFreeLists(false);
7962         }
7963       }
7964     }
7965   } else {
7966     // Code path common to both original and adaptive free lists.
7967 
7968     // cant coalesce with previous block; this should be treated
7969     // as the end of a free run if any
7970     if (inFreeRange()) {
7971       // we kicked some butt; time to pick up the garbage
7972       assert(freeFinger() < addr, "the finger pointeth off base");
7973       flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
7974     }
7975     // else, nothing to do, just continue
7976   }
7977 }
7978 
7979 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
7980   // This is a chunk of garbage.  It is not in any free list.
7981   // Add it to a free list or let it possibly be coalesced into
7982   // a larger chunk.
7983   HeapWord* addr = (HeapWord*) fc;
7984   size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7985 
7986   if (_sp->adaptive_freelists()) {
7987     // Verify that the bit map has no bits marked between
7988     // addr and purported end of just dead object.
7989     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7990 
7991     doPostIsFreeOrGarbageChunk(fc, size);
7992   } else {
7993     if (!inFreeRange()) {
7994       // start of a new free range
7995       assert(size > 0, "A free range should have a size");
7996       initialize_free_range(addr, false);
7997 
7998     } else {
7999       // this will be swept up when we hit the end of the
8000       // free range
8001       if (CMSTraceSweeper) {
8002         gclog_or_tty->print("  -- pick up garbage 0x%x (%d) \n", fc, size);
8003       }
8004       // If the chunk is being coalesced and the current free range is
8005       // in the free lists, remove the current free range so that it
8006       // will be returned to the free lists in its entirety - all
8007       // the coalesced pieces included.
8008       if (freeRangeInFreeLists()) {
8009         FreeChunk* ffc = (FreeChunk*)freeFinger();
8010         assert(ffc->size() == pointer_delta(addr, freeFinger()),
8011           "Size of free range is inconsistent with chunk size.");
8012         if (CMSTestInFreeList) {
8013           assert(_sp->verifyChunkInFreeLists(ffc),
8014             "free range is not in free lists");
8015         }
8016         _sp->removeFreeChunkFromFreeLists(ffc);
8017         set_freeRangeInFreeLists(false);
8018       }
8019       set_lastFreeRangeCoalesced(true);
8020     }
8021     // this will be swept up when we hit the end of the free range
8022 
8023     // Verify that the bit map has no bits marked between
8024     // addr and purported end of just dead object.
8025     _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8026   }
8027   return size;
8028 }
8029 
8030 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
8031   HeapWord* addr = (HeapWord*) fc;
8032   // The sweeper has just found a live object. Return any accumulated
8033   // left hand chunk to the free lists.
8034   if (inFreeRange()) {
8035     if (_sp->adaptive_freelists()) {
8036       flushCurFreeChunk(freeFinger(),
8037                         pointer_delta(addr, freeFinger()));
8038     } else { // not adaptive freelists
8039       set_inFreeRange(false);
8040       // Add the free range back to the free list if it is not already
8041       // there.
8042       if (!freeRangeInFreeLists()) {
8043         assert(freeFinger() < addr, "the finger pointeth off base");
8044         if (CMSTraceSweeper) {
8045           gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
8046             "[coalesced:%d]\n",
8047             freeFinger(), pointer_delta(addr, freeFinger()),
8048             lastFreeRangeCoalesced());
8049         }
8050         _sp->addChunkAndRepairOffsetTable(freeFinger(),
8051           pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
8052       }
8053     }
8054   }
8055 
8056   // Common code path for original and adaptive free lists.
8057 
8058   // this object is live: we'd normally expect this to be
8059   // an oop, and like to assert the following:
8060   // assert(oop(addr)->is_oop(), "live block should be an oop");
8061   // However, as we commented above, this may be an object whose
8062   // header hasn't yet been initialized.
8063   size_t size;
8064   assert(_bitMap->isMarked(addr), "Tautology for this control point");
8065   if (_bitMap->isMarked(addr + 1)) {
8066     // Determine the size from the bit map, rather than trying to
8067     // compute it from the object header.
8068     HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8069     size = pointer_delta(nextOneAddr + 1, addr);
8070     assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8071            "alignment problem");
8072 
8073     #ifdef DEBUG
8074       if (oop(addr)->klass_or_null() != NULL &&
8075           (   !_collector->should_unload_classes()
8076            || oop(addr)->is_parsable())) {
8077         // Ignore mark word because we are running concurrent with mutators
8078         assert(oop(addr)->is_oop(true), "live block should be an oop");
8079         assert(size ==
8080                CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8081                "P-mark and computed size do not agree");
8082       }
8083     #endif
8084 
8085   } else {
8086     // This should be an initialized object that's alive.
8087     assert(oop(addr)->klass_or_null() != NULL &&
8088            (!_collector->should_unload_classes()
8089             || oop(addr)->is_parsable()),
8090            "Should be an initialized object");
8091     // Ignore mark word because we are running concurrent with mutators
8092     assert(oop(addr)->is_oop(true), "live block should be an oop");
8093     // Verify that the bit map has no bits marked between
8094     // addr and purported end of this block.
8095     size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8096     assert(size >= 3, "Necessary for Printezis marks to work");
8097     assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8098     DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8099   }
8100   return size;
8101 }
8102 
8103 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
8104                                             size_t chunkSize) {
8105   // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
8106   // scheme.
8107   bool fcInFreeLists = fc->isFree();
8108   assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8109   assert((HeapWord*)fc <= _limit, "sweep invariant");
8110   if (CMSTestInFreeList && fcInFreeLists) {
8111     assert(_sp->verifyChunkInFreeLists(fc),
8112       "free chunk is not in free lists");
8113   }
8114 
8115 
8116   if (CMSTraceSweeper) {
8117     gclog_or_tty->print_cr("  -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8118   }
8119 
8120   HeapWord* addr = (HeapWord*) fc;
8121 
8122   bool coalesce;
8123   size_t left  = pointer_delta(addr, freeFinger());
8124   size_t right = chunkSize;
8125   switch (FLSCoalescePolicy) {
8126     // numeric value forms a coalition aggressiveness metric
8127     case 0:  { // never coalesce
8128       coalesce = false;
8129       break;
8130     }
8131     case 1: { // coalesce if left & right chunks on overpopulated lists
8132       coalesce = _sp->coalOverPopulated(left) &&
8133                  _sp->coalOverPopulated(right);
8134       break;
8135     }
8136     case 2: { // coalesce if left chunk on overpopulated list (default)
8137       coalesce = _sp->coalOverPopulated(left);
8138       break;
8139     }
8140     case 3: { // coalesce if left OR right chunk on overpopulated list
8141       coalesce = _sp->coalOverPopulated(left) ||
8142                  _sp->coalOverPopulated(right);
8143       break;
8144     }
8145     case 4: { // always coalesce
8146       coalesce = true;
8147       break;
8148     }
8149     default:
8150      ShouldNotReachHere();
8151   }
8152 
8153   // Should the current free range be coalesced?
8154   // If the chunk is in a free range and either we decided to coalesce above
8155   // or the chunk is near the large block at the end of the heap
8156   // (isNearLargestChunk() returns true), then coalesce this chunk.
8157   bool doCoalesce = inFreeRange() &&
8158     (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8159   if (doCoalesce) {
8160     // Coalesce the current free range on the left with the new
8161     // chunk on the right.  If either is on a free list,
8162     // it must be removed from the list and stashed in the closure.
8163     if (freeRangeInFreeLists()) {
8164       FreeChunk* ffc = (FreeChunk*)freeFinger();
8165       assert(ffc->size() == pointer_delta(addr, freeFinger()),
8166         "Size of free range is inconsistent with chunk size.");
8167       if (CMSTestInFreeList) {
8168         assert(_sp->verifyChunkInFreeLists(ffc),
8169           "Chunk is not in free lists");
8170       }
8171       _sp->coalDeath(ffc->size());
8172       _sp->removeFreeChunkFromFreeLists(ffc);
8173       set_freeRangeInFreeLists(false);
8174     }
8175     if (fcInFreeLists) {
8176       _sp->coalDeath(chunkSize);
8177       assert(fc->size() == chunkSize,
8178         "The chunk has the wrong size or is not in the free lists");
8179       _sp->removeFreeChunkFromFreeLists(fc);
8180     }
8181     set_lastFreeRangeCoalesced(true);
8182   } else {  // not in a free range and/or should not coalesce
8183     // Return the current free range and start a new one.
8184     if (inFreeRange()) {
8185       // In a free range but cannot coalesce with the right hand chunk.
8186       // Put the current free range into the free lists.
8187       flushCurFreeChunk(freeFinger(),
8188         pointer_delta(addr, freeFinger()));
8189     }
8190     // Set up for new free range.  Pass along whether the right hand
8191     // chunk is in the free lists.
8192     initialize_free_range((HeapWord*)fc, fcInFreeLists);
8193   }
8194 }
8195 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
8196   assert(inFreeRange(), "Should only be called if currently in a free range.");
8197   assert(size > 0,
8198     "A zero sized chunk cannot be added to the free lists.");
8199   if (!freeRangeInFreeLists()) {
8200     if(CMSTestInFreeList) {
8201       FreeChunk* fc = (FreeChunk*) chunk;
8202       fc->setSize(size);
8203       assert(!_sp->verifyChunkInFreeLists(fc),
8204         "chunk should not be in free lists yet");
8205     }
8206     if (CMSTraceSweeper) {
8207       gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8208                     chunk, size);
8209     }
8210     // A new free range is going to be starting.  The current
8211     // free range has not been added to the free lists yet or
8212     // was removed so add it back.
8213     // If the current free range was coalesced, then the death
8214     // of the free range was recorded.  Record a birth now.
8215     if (lastFreeRangeCoalesced()) {
8216       _sp->coalBirth(size);
8217     }
8218     _sp->addChunkAndRepairOffsetTable(chunk, size,
8219             lastFreeRangeCoalesced());
8220   }
8221   set_inFreeRange(false);
8222   set_freeRangeInFreeLists(false);
8223 }
8224 
8225 // We take a break if we've been at this for a while,
8226 // so as to avoid monopolizing the locks involved.
8227 void SweepClosure::do_yield_work(HeapWord* addr) {
8228   // Return current free chunk being used for coalescing (if any)
8229   // to the appropriate freelist.  After yielding, the next
8230   // free block encountered will start a coalescing range of
8231   // free blocks.  If the next free block is adjacent to the
8232   // chunk just flushed, they will need to wait for the next
8233   // sweep to be coalesced.
8234   if (inFreeRange()) {
8235     flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
8236   }
8237 
8238   // First give up the locks, then yield, then re-lock.
8239   // We should probably use a constructor/destructor idiom to
8240   // do this unlock/lock or modify the MutexUnlocker class to
8241   // serve our purpose. XXX
8242   assert_lock_strong(_bitMap->lock());
8243   assert_lock_strong(_freelistLock);
8244   assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8245          "CMS thread should hold CMS token");
8246   _bitMap->lock()->unlock();
8247   _freelistLock->unlock();
8248   ConcurrentMarkSweepThread::desynchronize(true);
8249   ConcurrentMarkSweepThread::acknowledge_yield_request();
8250   _collector->stopTimer();
8251   GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8252   if (PrintCMSStatistics != 0) {
8253     _collector->incrementYields();
8254   }
8255   _collector->icms_wait();
8256 
8257   // See the comment in coordinator_yield()
8258   for (unsigned i = 0; i < CMSYieldSleepCount &&
8259                        ConcurrentMarkSweepThread::should_yield() &&
8260                        !CMSCollector::foregroundGCIsActive(); ++i) {
8261     os::sleep(Thread::current(), 1, false);
8262     ConcurrentMarkSweepThread::acknowledge_yield_request();
8263   }
8264 
8265   ConcurrentMarkSweepThread::synchronize(true);
8266   _freelistLock->lock();
8267   _bitMap->lock()->lock_without_safepoint_check();
8268   _collector->startTimer();
8269 }
8270 
8271 #ifndef PRODUCT
8272 // This is actually very useful in a product build if it can
8273 // be called from the debugger.  Compile it into the product
8274 // as needed.
8275 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8276   return debug_cms_space->verifyChunkInFreeLists(fc);
8277 }
8278 
8279 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8280   if (CMSTraceSweeper) {
8281     gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8282   }
8283 }
8284 #endif
8285 
8286 // CMSIsAliveClosure
8287 bool CMSIsAliveClosure::do_object_b(oop obj) {
8288   HeapWord* addr = (HeapWord*)obj;
8289   return addr != NULL &&
8290          (!_span.contains(addr) || _bit_map->isMarked(addr));
8291 }
8292 
8293 // CMSKeepAliveClosure: the serial version
8294 void CMSKeepAliveClosure::do_oop(oop obj) {
8295   HeapWord* addr = (HeapWord*)obj;
8296   if (_span.contains(addr) &&
8297       !_bit_map->isMarked(addr)) {
8298     _bit_map->mark(addr);
8299     bool simulate_overflow = false;
8300     NOT_PRODUCT(
8301       if (CMSMarkStackOverflowALot &&
8302           _collector->simulate_overflow()) {
8303         // simulate a stack overflow
8304         simulate_overflow = true;
8305       }
8306     )
8307     if (simulate_overflow || !_mark_stack->push(obj)) {
8308       if (_concurrent_precleaning) {
8309         // We dirty the overflown object and let the remark
8310         // phase deal with it.
8311         assert(_collector->overflow_list_is_empty(), "Error");
8312         // In the case of object arrays, we need to dirty all of
8313         // the cards that the object spans. No locking or atomics
8314         // are needed since no one else can be mutating the mod union
8315         // table.
8316         if (obj->is_objArray()) {
8317           size_t sz = obj->size();
8318           HeapWord* end_card_addr =
8319             (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8320           MemRegion redirty_range = MemRegion(addr, end_card_addr);
8321           assert(!redirty_range.is_empty(), "Arithmetical tautology");
8322           _collector->_modUnionTable.mark_range(redirty_range);
8323         } else {
8324           _collector->_modUnionTable.mark(addr);
8325         }
8326         _collector->_ser_kac_preclean_ovflw++;
8327       } else {
8328         _collector->push_on_overflow_list(obj);
8329         _collector->_ser_kac_ovflw++;
8330       }
8331     }
8332   }
8333 }
8334 
8335 void CMSKeepAliveClosure::do_oop(oop* p)       { CMSKeepAliveClosure::do_oop_work(p); }
8336 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8337 
8338 // CMSParKeepAliveClosure: a parallel version of the above.
8339 // The work queues are private to each closure (thread),
8340 // but (may be) available for stealing by other threads.
8341 void CMSParKeepAliveClosure::do_oop(oop obj) {
8342   HeapWord* addr = (HeapWord*)obj;
8343   if (_span.contains(addr) &&
8344       !_bit_map->isMarked(addr)) {
8345     // In general, during recursive tracing, several threads
8346     // may be concurrently getting here; the first one to
8347     // "tag" it, claims it.
8348     if (_bit_map->par_mark(addr)) {
8349       bool res = _work_queue->push(obj);
8350       assert(res, "Low water mark should be much less than capacity");
8351       // Do a recursive trim in the hope that this will keep
8352       // stack usage lower, but leave some oops for potential stealers
8353       trim_queue(_low_water_mark);
8354     } // Else, another thread got there first
8355   }
8356 }
8357 
8358 void CMSParKeepAliveClosure::do_oop(oop* p)       { CMSParKeepAliveClosure::do_oop_work(p); }
8359 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8360 
8361 void CMSParKeepAliveClosure::trim_queue(uint max) {
8362   while (_work_queue->size() > max) {
8363     oop new_oop;
8364     if (_work_queue->pop_local(new_oop)) {
8365       assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8366       assert(_bit_map->isMarked((HeapWord*)new_oop),
8367              "no white objects on this stack!");
8368       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8369       // iterate over the oops in this oop, marking and pushing
8370       // the ones in CMS heap (i.e. in _span).
8371       new_oop->oop_iterate(&_mark_and_push);
8372     }
8373   }
8374 }
8375 
8376 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8377   HeapWord* addr = (HeapWord*)obj;
8378   if (_span.contains(addr) &&
8379       !_bit_map->isMarked(addr)) {
8380     if (_bit_map->par_mark(addr)) {
8381       bool simulate_overflow = false;
8382       NOT_PRODUCT(
8383         if (CMSMarkStackOverflowALot &&
8384             _collector->par_simulate_overflow()) {
8385           // simulate a stack overflow
8386           simulate_overflow = true;
8387         }
8388       )
8389       if (simulate_overflow || !_work_queue->push(obj)) {
8390         _collector->par_push_on_overflow_list(obj);
8391         _collector->_par_kac_ovflw++;
8392       }
8393     } // Else another thread got there already
8394   }
8395 }
8396 
8397 void CMSInnerParMarkAndPushClosure::do_oop(oop* p)       { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8398 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8399 
8400 //////////////////////////////////////////////////////////////////
8401 //  CMSExpansionCause                /////////////////////////////
8402 //////////////////////////////////////////////////////////////////
8403 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8404   switch (cause) {
8405     case _no_expansion:
8406       return "No expansion";
8407     case _satisfy_free_ratio:
8408       return "Free ratio";
8409     case _satisfy_promotion:
8410       return "Satisfy promotion";
8411     case _satisfy_allocation:
8412       return "allocation";
8413     case _allocate_par_lab:
8414       return "Par LAB";
8415     case _allocate_par_spooling_space:
8416       return "Par Spooling Space";
8417     case _adaptive_size_policy:
8418       return "Ergonomics";
8419     default:
8420       return "unknown";
8421   }
8422 }
8423 
8424 void CMSDrainMarkingStackClosure::do_void() {
8425   // the max number to take from overflow list at a time
8426   const size_t num = _mark_stack->capacity()/4;
8427   assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8428          "Overflow list should be NULL during concurrent phases");
8429   while (!_mark_stack->isEmpty() ||
8430          // if stack is empty, check the overflow list
8431          _collector->take_from_overflow_list(num, _mark_stack)) {
8432     oop obj = _mark_stack->pop();
8433     HeapWord* addr = (HeapWord*)obj;
8434     assert(_span.contains(addr), "Should be within span");
8435     assert(_bit_map->isMarked(addr), "Should be marked");
8436     assert(obj->is_oop(), "Should be an oop");
8437     obj->oop_iterate(_keep_alive);
8438   }
8439 }
8440 
8441 void CMSParDrainMarkingStackClosure::do_void() {
8442   // drain queue
8443   trim_queue(0);
8444 }
8445 
8446 // Trim our work_queue so its length is below max at return
8447 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8448   while (_work_queue->size() > max) {
8449     oop new_oop;
8450     if (_work_queue->pop_local(new_oop)) {
8451       assert(new_oop->is_oop(), "Expected an oop");
8452       assert(_bit_map->isMarked((HeapWord*)new_oop),
8453              "no white objects on this stack!");
8454       assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8455       // iterate over the oops in this oop, marking and pushing
8456       // the ones in CMS heap (i.e. in _span).
8457       new_oop->oop_iterate(&_mark_and_push);
8458     }
8459   }
8460 }
8461 
8462 ////////////////////////////////////////////////////////////////////
8463 // Support for Marking Stack Overflow list handling and related code
8464 ////////////////////////////////////////////////////////////////////
8465 // Much of the following code is similar in shape and spirit to the
8466 // code used in ParNewGC. We should try and share that code
8467 // as much as possible in the future.
8468 
8469 #ifndef PRODUCT
8470 // Debugging support for CMSStackOverflowALot
8471 
8472 // It's OK to call this multi-threaded;  the worst thing
8473 // that can happen is that we'll get a bunch of closely
8474 // spaced simulated oveflows, but that's OK, in fact
8475 // probably good as it would exercise the overflow code
8476 // under contention.
8477 bool CMSCollector::simulate_overflow() {
8478   if (_overflow_counter-- <= 0) { // just being defensive
8479     _overflow_counter = CMSMarkStackOverflowInterval;
8480     return true;
8481   } else {
8482     return false;
8483   }
8484 }
8485 
8486 bool CMSCollector::par_simulate_overflow() {
8487   return simulate_overflow();
8488 }
8489 #endif
8490 
8491 // Single-threaded
8492 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8493   assert(stack->isEmpty(), "Expected precondition");
8494   assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8495   size_t i = num;
8496   oop  cur = _overflow_list;
8497   const markOop proto = markOopDesc::prototype();
8498   NOT_PRODUCT(size_t n = 0;)
8499   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8500     next = oop(cur->mark());
8501     cur->set_mark(proto);   // until proven otherwise
8502     assert(cur->is_oop(), "Should be an oop");
8503     bool res = stack->push(cur);
8504     assert(res, "Bit off more than can chew?");
8505     NOT_PRODUCT(n++;)
8506   }
8507   _overflow_list = cur;
8508 #ifndef PRODUCT
8509   assert(_num_par_pushes >= n, "Too many pops?");
8510   _num_par_pushes -=n;
8511 #endif
8512   return !stack->isEmpty();
8513 }
8514 
8515 // Multi-threaded; use CAS to break off a prefix
8516 bool CMSCollector::par_take_from_overflow_list(size_t num,
8517                                                OopTaskQueue* work_q) {
8518   assert(work_q->size() == 0, "That's the current policy");
8519   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8520   if (_overflow_list == NULL) {
8521     return false;
8522   }
8523   // Grab the entire list; we'll put back a suffix
8524   oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
8525   if (prefix == NULL) {  // someone grabbed it before we did ...
8526     // ... we could spin for a short while, but for now we don't
8527     return false;
8528   }
8529   size_t i = num;
8530   oop cur = prefix;
8531   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8532   if (cur->mark() != NULL) {
8533     oop suffix_head = cur->mark(); // suffix will be put back on global list
8534     cur->set_mark(NULL);           // break off suffix
8535     // Find tail of suffix so we can prepend suffix to global list
8536     for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8537     oop suffix_tail = cur;
8538     assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8539            "Tautology");
8540     oop observed_overflow_list = _overflow_list;
8541     do {
8542       cur = observed_overflow_list;
8543       suffix_tail->set_mark(markOop(cur));
8544       observed_overflow_list =
8545         (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
8546     } while (cur != observed_overflow_list);
8547   }
8548 
8549   // Push the prefix elements on work_q
8550   assert(prefix != NULL, "control point invariant");
8551   const markOop proto = markOopDesc::prototype();
8552   oop next;
8553   NOT_PRODUCT(size_t n = 0;)
8554   for (cur = prefix; cur != NULL; cur = next) {
8555     next = oop(cur->mark());
8556     cur->set_mark(proto);   // until proven otherwise
8557     assert(cur->is_oop(), "Should be an oop");
8558     bool res = work_q->push(cur);
8559     assert(res, "Bit off more than we can chew?");
8560     NOT_PRODUCT(n++;)
8561   }
8562 #ifndef PRODUCT
8563   assert(_num_par_pushes >= n, "Too many pops?");
8564   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8565 #endif
8566   return true;
8567 }
8568 
8569 // Single-threaded
8570 void CMSCollector::push_on_overflow_list(oop p) {
8571   NOT_PRODUCT(_num_par_pushes++;)
8572   assert(p->is_oop(), "Not an oop");
8573   preserve_mark_if_necessary(p);
8574   p->set_mark((markOop)_overflow_list);
8575   _overflow_list = p;
8576 }
8577 
8578 // Multi-threaded; use CAS to prepend to overflow list
8579 void CMSCollector::par_push_on_overflow_list(oop p) {
8580   NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8581   assert(p->is_oop(), "Not an oop");
8582   par_preserve_mark_if_necessary(p);
8583   oop observed_overflow_list = _overflow_list;
8584   oop cur_overflow_list;
8585   do {
8586     cur_overflow_list = observed_overflow_list;
8587     p->set_mark(markOop(cur_overflow_list));
8588     observed_overflow_list =
8589       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8590   } while (cur_overflow_list != observed_overflow_list);
8591 }
8592 
8593 // Single threaded
8594 // General Note on GrowableArray: pushes may silently fail
8595 // because we are (temporarily) out of C-heap for expanding
8596 // the stack. The problem is quite ubiquitous and affects
8597 // a lot of code in the JVM. The prudent thing for GrowableArray
8598 // to do (for now) is to exit with an error. However, that may
8599 // be too draconian in some cases because the caller may be
8600 // able to recover without much harm. For suych cases, we
8601 // should probably introduce a "soft_push" method which returns
8602 // an indication of success or failure with the assumption that
8603 // the caller may be able to recover from a failure; code in
8604 // the VM can then be changed, incrementally, to deal with such
8605 // failures where possible, thus, incrementally hardening the VM
8606 // in such low resource situations.
8607 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8608   int PreserveMarkStackSize = 128;
8609 
8610   if (_preserved_oop_stack == NULL) {
8611     assert(_preserved_mark_stack == NULL,
8612            "bijection with preserved_oop_stack");
8613     // Allocate the stacks
8614     _preserved_oop_stack  = new (ResourceObj::C_HEAP)
8615       GrowableArray<oop>(PreserveMarkStackSize, true);
8616     _preserved_mark_stack = new (ResourceObj::C_HEAP)
8617       GrowableArray<markOop>(PreserveMarkStackSize, true);
8618     if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
8619       vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
8620                             "Preserved Mark/Oop Stack for CMS (C-heap)");
8621     }
8622   }
8623   _preserved_oop_stack->push(p);
8624   _preserved_mark_stack->push(m);
8625   assert(m == p->mark(), "Mark word changed");
8626   assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
8627          "bijection");
8628 }
8629 
8630 // Single threaded
8631 void CMSCollector::preserve_mark_if_necessary(oop p) {
8632   markOop m = p->mark();
8633   if (m->must_be_preserved(p)) {
8634     preserve_mark_work(p, m);
8635   }
8636 }
8637 
8638 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8639   markOop m = p->mark();
8640   if (m->must_be_preserved(p)) {
8641     MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8642     // Even though we read the mark word without holding
8643     // the lock, we are assured that it will not change
8644     // because we "own" this oop, so no other thread can
8645     // be trying to push it on the overflow list; see
8646     // the assertion in preserve_mark_work() that checks
8647     // that m == p->mark().
8648     preserve_mark_work(p, m);
8649   }
8650 }
8651 
8652 // We should be able to do this multi-threaded,
8653 // a chunk of stack being a task (this is
8654 // correct because each oop only ever appears
8655 // once in the overflow list. However, it's
8656 // not very easy to completely overlap this with
8657 // other operations, so will generally not be done
8658 // until all work's been completed. Because we
8659 // expect the preserved oop stack (set) to be small,
8660 // it's probably fine to do this single-threaded.
8661 // We can explore cleverer concurrent/overlapped/parallel
8662 // processing of preserved marks if we feel the
8663 // need for this in the future. Stack overflow should
8664 // be so rare in practice and, when it happens, its
8665 // effect on performance so great that this will
8666 // likely just be in the noise anyway.
8667 void CMSCollector::restore_preserved_marks_if_any() {
8668   if (_preserved_oop_stack == NULL) {
8669     assert(_preserved_mark_stack == NULL,
8670            "bijection with preserved_oop_stack");
8671     return;
8672   }
8673 
8674   assert(SafepointSynchronize::is_at_safepoint(),
8675          "world should be stopped");
8676   assert(Thread::current()->is_ConcurrentGC_thread() ||
8677          Thread::current()->is_VM_thread(),
8678          "should be single-threaded");
8679 
8680   int length = _preserved_oop_stack->length();
8681   assert(_preserved_mark_stack->length() == length, "bijection");
8682   for (int i = 0; i < length; i++) {
8683     oop p = _preserved_oop_stack->at(i);
8684     assert(p->is_oop(), "Should be an oop");
8685     assert(_span.contains(p), "oop should be in _span");
8686     assert(p->mark() == markOopDesc::prototype(),
8687            "Set when taken from overflow list");
8688     markOop m = _preserved_mark_stack->at(i);
8689     p->set_mark(m);
8690   }
8691   _preserved_mark_stack->clear();
8692   _preserved_oop_stack->clear();
8693   assert(_preserved_mark_stack->is_empty() &&
8694          _preserved_oop_stack->is_empty(),
8695          "stacks were cleared above");
8696 }
8697 
8698 #ifndef PRODUCT
8699 bool CMSCollector::no_preserved_marks() const {
8700   return (   (   _preserved_mark_stack == NULL
8701               && _preserved_oop_stack == NULL)
8702           || (   _preserved_mark_stack->is_empty()
8703               && _preserved_oop_stack->is_empty()));
8704 }
8705 #endif
8706 
8707 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8708 {
8709   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8710   CMSAdaptiveSizePolicy* size_policy =
8711     (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8712   assert(size_policy->is_gc_cms_adaptive_size_policy(),
8713     "Wrong type for size policy");
8714   return size_policy;
8715 }
8716 
8717 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
8718                                            size_t desired_promo_size) {
8719   if (cur_promo_size < desired_promo_size) {
8720     size_t expand_bytes = desired_promo_size - cur_promo_size;
8721     if (PrintAdaptiveSizePolicy && Verbose) {
8722       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8723         "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
8724         expand_bytes);
8725     }
8726     expand(expand_bytes,
8727            MinHeapDeltaBytes,
8728            CMSExpansionCause::_adaptive_size_policy);
8729   } else if (desired_promo_size < cur_promo_size) {
8730     size_t shrink_bytes = cur_promo_size - desired_promo_size;
8731     if (PrintAdaptiveSizePolicy && Verbose) {
8732       gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8733         "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
8734         shrink_bytes);
8735     }
8736     shrink(shrink_bytes);
8737   }
8738 }
8739 
8740 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
8741   GenCollectedHeap* gch = GenCollectedHeap::heap();
8742   CMSGCAdaptivePolicyCounters* counters =
8743     (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
8744   assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
8745     "Wrong kind of counters");
8746   return counters;
8747 }
8748 
8749 
8750 void ASConcurrentMarkSweepGeneration::update_counters() {
8751   if (UsePerfData) {
8752     _space_counters->update_all();
8753     _gen_counters->update_all();
8754     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8755     GenCollectedHeap* gch = GenCollectedHeap::heap();
8756     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8757     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8758       "Wrong gc statistics type");
8759     counters->update_counters(gc_stats_l);
8760   }
8761 }
8762 
8763 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
8764   if (UsePerfData) {
8765     _space_counters->update_used(used);
8766     _space_counters->update_capacity();
8767     _gen_counters->update_all();
8768 
8769     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8770     GenCollectedHeap* gch = GenCollectedHeap::heap();
8771     CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8772     assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8773       "Wrong gc statistics type");
8774     counters->update_counters(gc_stats_l);
8775   }
8776 }
8777 
8778 // The desired expansion delta is computed so that:
8779 // . desired free percentage or greater is used
8780 void ASConcurrentMarkSweepGeneration::compute_new_size() {
8781   assert_locked_or_safepoint(Heap_lock);
8782 
8783   GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8784 
8785   // If incremental collection failed, we just want to expand
8786   // to the limit.
8787   if (incremental_collection_failed()) {
8788     clear_incremental_collection_failed();
8789     grow_to_reserved();
8790     return;
8791   }
8792 
8793   assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
8794 
8795   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
8796     "Wrong type of heap");
8797   int prev_level = level() - 1;
8798   assert(prev_level >= 0, "The cms generation is the lowest generation");
8799   Generation* prev_gen = gch->get_gen(prev_level);
8800   assert(prev_gen->kind() == Generation::ASParNew,
8801     "Wrong type of young generation");
8802   ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
8803   size_t cur_eden = younger_gen->eden()->capacity();
8804   CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
8805   size_t cur_promo = free();
8806   size_policy->compute_tenured_generation_free_space(cur_promo,
8807                                                        max_available(),
8808                                                        cur_eden);
8809   resize(cur_promo, size_policy->promo_size());
8810 
8811   // Record the new size of the space in the cms generation
8812   // that is available for promotions.  This is temporary.
8813   // It should be the desired promo size.
8814   size_policy->avg_cms_promo()->sample(free());
8815   size_policy->avg_old_live()->sample(used());
8816 
8817   if (UsePerfData) {
8818     CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8819     counters->update_cms_capacity_counter(capacity());
8820   }
8821 }
8822 
8823 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
8824   assert_locked_or_safepoint(Heap_lock);
8825   assert_lock_strong(freelistLock());
8826   HeapWord* old_end = _cmsSpace->end();
8827   HeapWord* unallocated_start = _cmsSpace->unallocated_block();
8828   assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
8829   FreeChunk* chunk_at_end = find_chunk_at_end();
8830   if (chunk_at_end == NULL) {
8831     // No room to shrink
8832     if (PrintGCDetails && Verbose) {
8833       gclog_or_tty->print_cr("No room to shrink: old_end  "
8834         PTR_FORMAT "  unallocated_start  " PTR_FORMAT
8835         " chunk_at_end  " PTR_FORMAT,
8836         old_end, unallocated_start, chunk_at_end);
8837     }
8838     return;
8839   } else {
8840 
8841     // Find the chunk at the end of the space and determine
8842     // how much it can be shrunk.
8843     size_t shrinkable_size_in_bytes = chunk_at_end->size();
8844     size_t aligned_shrinkable_size_in_bytes =
8845       align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
8846     assert(unallocated_start <= chunk_at_end->end(),
8847       "Inconsistent chunk at end of space");
8848     size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
8849     size_t word_size_before = heap_word_size(_virtual_space.committed_size());
8850 
8851     // Shrink the underlying space
8852     _virtual_space.shrink_by(bytes);
8853     if (PrintGCDetails && Verbose) {
8854       gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
8855         " desired_bytes " SIZE_FORMAT
8856         " shrinkable_size_in_bytes " SIZE_FORMAT
8857         " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
8858         "  bytes  " SIZE_FORMAT,
8859         desired_bytes, shrinkable_size_in_bytes,
8860         aligned_shrinkable_size_in_bytes, bytes);
8861       gclog_or_tty->print_cr("          old_end  " SIZE_FORMAT
8862         "  unallocated_start  " SIZE_FORMAT,
8863         old_end, unallocated_start);
8864     }
8865 
8866     // If the space did shrink (shrinking is not guaranteed),
8867     // shrink the chunk at the end by the appropriate amount.
8868     if (((HeapWord*)_virtual_space.high()) < old_end) {
8869       size_t new_word_size =
8870         heap_word_size(_virtual_space.committed_size());
8871 
8872       // Have to remove the chunk from the dictionary because it is changing
8873       // size and might be someplace elsewhere in the dictionary.
8874 
8875       // Get the chunk at end, shrink it, and put it
8876       // back.
8877       _cmsSpace->removeChunkFromDictionary(chunk_at_end);
8878       size_t word_size_change = word_size_before - new_word_size;
8879       size_t chunk_at_end_old_size = chunk_at_end->size();
8880       assert(chunk_at_end_old_size >= word_size_change,
8881         "Shrink is too large");
8882       chunk_at_end->setSize(chunk_at_end_old_size -
8883                           word_size_change);
8884       _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
8885         word_size_change);
8886 
8887       _cmsSpace->returnChunkToDictionary(chunk_at_end);
8888 
8889       MemRegion mr(_cmsSpace->bottom(), new_word_size);
8890       _bts->resize(new_word_size);  // resize the block offset shared array
8891       Universe::heap()->barrier_set()->resize_covered_region(mr);
8892       _cmsSpace->assert_locked();
8893       _cmsSpace->set_end((HeapWord*)_virtual_space.high());
8894 
8895       NOT_PRODUCT(_cmsSpace->dictionary()->verify());
8896 
8897       // update the space and generation capacity counters
8898       if (UsePerfData) {
8899         _space_counters->update_capacity();
8900         _gen_counters->update_all();
8901       }
8902 
8903       if (Verbose && PrintGCDetails) {
8904         size_t new_mem_size = _virtual_space.committed_size();
8905         size_t old_mem_size = new_mem_size + bytes;
8906         gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
8907                       name(), old_mem_size/K, bytes/K, new_mem_size/K);
8908       }
8909     }
8910 
8911     assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
8912       "Inconsistency at end of space");
8913     assert(chunk_at_end->end() == _cmsSpace->end(),
8914       "Shrinking is inconsistent");
8915     return;
8916   }
8917 }
8918 
8919 // Transfer some number of overflown objects to usual marking
8920 // stack. Return true if some objects were transferred.
8921 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8922   size_t num = MIN2((size_t)_mark_stack->capacity()/4,
8923                     (size_t)ParGCDesiredObjsFromOverflowList);
8924 
8925   bool res = _collector->take_from_overflow_list(num, _mark_stack);
8926   assert(_collector->overflow_list_is_empty() || res,
8927          "If list is not empty, we should have taken something");
8928   assert(!res || !_mark_stack->isEmpty(),
8929          "If we took something, it should now be on our stack");
8930   return res;
8931 }
8932 
8933 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8934   size_t res = _sp->block_size_no_stall(addr, _collector);
8935   assert(res != 0, "Should always be able to compute a size");
8936   if (_sp->block_is_obj(addr)) {
8937     if (_live_bit_map->isMarked(addr)) {
8938       // It can't have been dead in a previous cycle
8939       guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8940     } else {
8941       _dead_bit_map->mark(addr);      // mark the dead object
8942     }
8943   }
8944   return res;
8945 }